From 26cec1aafb07cb73ba941d088f2c33c6c407c8c6 Mon Sep 17 00:00:00 2001 From: Kunal Pathak Date: Mon, 17 May 2021 00:00:20 -0700 Subject: [PATCH 1/7] refactor lsra --- src/coreclr/jit/CMakeLists.txt | 2 + src/coreclr/jit/jitconfigvalues.h | 1 + src/coreclr/jit/lsra.cpp | 14821 ++++++++++++++-------------- src/coreclr/jit/lsra.h | 164 +- 4 files changed, 7641 insertions(+), 7347 deletions(-) diff --git a/src/coreclr/jit/CMakeLists.txt b/src/coreclr/jit/CMakeLists.txt index 1d8e8645f5046..6fd2a686cc8e6 100644 --- a/src/coreclr/jit/CMakeLists.txt +++ b/src/coreclr/jit/CMakeLists.txt @@ -213,6 +213,8 @@ if (CLR_CMAKE_TARGET_WIN32) loopcloningopts.h lower.h lsra_reftypes.h + lsra_stats.h + lsra_score.h lsra.h namedintrinsiclist.h objectalloc.h diff --git a/src/coreclr/jit/jitconfigvalues.h b/src/coreclr/jit/jitconfigvalues.h index 9634442323cc1..9dc27e581263c 100644 --- a/src/coreclr/jit/jitconfigvalues.h +++ b/src/coreclr/jit/jitconfigvalues.h @@ -29,6 +29,7 @@ CONFIG_INTEGER(DisplayLoopHoistStats, W("JitLoopHoistStats"), 0) // Display JIT CONFIG_INTEGER(DisplayLsraStats, W("JitLsraStats"), 0) // Display JIT Linear Scan Register Allocator statistics // if set to 1. If set to "2", display the stats in csv format. // Recommended to use with JitStdOutFile flag. +CONFIG_STRING(JitLsraOrdering, W("JitLsraOrdering")) // LSRA heuristics ordering CONFIG_INTEGER(DumpJittedMethods, W("DumpJittedMethods"), 0) // Prints all jitted methods to the console CONFIG_INTEGER(EnablePCRelAddr, W("JitEnablePCRelAddr"), 1) // Whether absolute addr be encoded as PC-rel offset by // RyuJIT where possible diff --git a/src/coreclr/jit/lsra.cpp b/src/coreclr/jit/lsra.cpp index 779e2419e1c4c..d946d9a90e4ff 100644 --- a/src/coreclr/jit/lsra.cpp +++ b/src/coreclr/jit/lsra.cpp @@ -626,6 +626,7 @@ LinearScan::LinearScan(Compiler* theCompiler) , refPositions(theCompiler->getAllocator(CMK_LSRA_RefPosition)) , listNodePool(theCompiler) { + regSelector = new (theCompiler, CMK_LSRA) RegisterSelection(this); firstColdLoc = MaxLocation; #ifdef DEBUG @@ -2756,4331 +2757,4529 @@ bool LinearScan::isMatchingConstant(RegRecord* physRegRecord, RefPosition* refPo regNumber LinearScan::allocateReg(Interval* currentInterval, RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)) { - regNumber foundReg = REG_NA; -#ifdef DEBUG - *registerScore = NONE; -#endif - - RegisterType regType = getRegisterType(currentInterval, refPosition); - - LsraLocation currentLocation = refPosition->nodeLocation; - RefPosition* nextRefPos = refPosition->nextRefPosition; - regMaskTP candidates = refPosition->registerAssignment; - regMaskTP preferences = currentInterval->registerPreferences; - - if (RefTypeIsDef(refPosition->refType)) + regMaskTP foundRegBit = regSelector->select(currentInterval, refPosition DEBUG_ARG(registerScore)); + regNumber foundReg = genRegNumFromMask(foundRegBit); + RegRecord* availablePhysRegRecord = getRegisterRecord(foundReg); + Interval* assignedInterval = availablePhysRegRecord->assignedInterval; + if ((assignedInterval != currentInterval) && + isAssigned(availablePhysRegRecord ARM_ARG(getRegisterType(currentInterval, refPosition)))) { - if (currentInterval->hasConflictingDefUse) + if (regSelector->isSpilling()) { - resolveConflictingDefAndUse(currentInterval, refPosition); - candidates = refPosition->registerAssignment; + // We're spilling. + CLANG_FORMAT_COMMENT_ANCHOR; + +#ifdef TARGET_ARM + if (currentInterval->registerType == TYP_DOUBLE) + { + assert(genIsValidDoubleReg(availablePhysRegRecord->regNum)); + unassignDoublePhysReg(availablePhysRegRecord); + } + else if (assignedInterval->registerType == TYP_DOUBLE) + { + // Make sure we spill both halves of the double register. + assert(genIsValidDoubleReg(assignedInterval->assignedReg->regNum)); + unassignPhysReg(assignedInterval->assignedReg, assignedInterval->recentRefPosition); + } + else +#endif + { + unassignPhysReg(availablePhysRegRecord, assignedInterval->recentRefPosition); + } } - // Otherwise, check for the case of a fixed-reg def of a reg that will be killed before the - // use, or interferes at the point of use (which shouldn't happen, but Lower doesn't mark - // the contained nodes as interfering). - // Note that we may have a ParamDef RefPosition that is marked isFixedRegRef, but which - // has had its registerAssignment changed to no longer be a single register. - else if (refPosition->isFixedRegRef && nextRefPos != nullptr && RefTypeIsUse(nextRefPos->refType) && - !nextRefPos->isFixedRegRef && genMaxOneBit(refPosition->registerAssignment)) + else { - regNumber defReg = refPosition->assignedReg(); - RegRecord* defRegRecord = getRegisterRecord(defReg); - - RefPosition* currFixedRegRefPosition = defRegRecord->recentRefPosition; - assert(currFixedRegRefPosition != nullptr && - currFixedRegRefPosition->nodeLocation == refPosition->nodeLocation); - - // If there is another fixed reference to this register before the use, change the candidates - // on this RefPosition to include that of nextRefPos. - RefPosition* nextFixedRegRefPosition = defRegRecord->getNextRefPosition(); - if (nextFixedRegRefPosition != nullptr && - nextFixedRegRefPosition->nodeLocation <= nextRefPos->getRefEndLocation()) + // If we considered this "unassigned" because this interval's lifetime ends before + // the next ref, remember it. + // For historical reasons (due to former short-circuiting of this case), if we're reassigning + // the current interval to a previous assignment, we don't remember the previous interval. + // Note that we need to compute this condition before calling unassignPhysReg, which wil reset + // assignedInterval->physReg. + bool wasAssigned = regSelector->foundUnassignedReg() && (assignedInterval != nullptr) && + (assignedInterval->physReg == foundReg); + unassignPhysReg(availablePhysRegRecord ARM_ARG(currentInterval->registerType)); + if (regSelector->isMatchingConstant()) { - candidates |= nextRefPos->registerAssignment; - if (preferences == refPosition->registerAssignment) - { - preferences = candidates; - } + assert(assignedInterval->isConstant); + refPosition->treeNode->SetReuseRegVal(); + } + else if (wasAssigned) + { + updatePreviousInterval(availablePhysRegRecord, assignedInterval, assignedInterval->registerType); + } + else + { + assert(!regSelector->isConstAvailable()); } } } + assignPhysReg(availablePhysRegRecord, currentInterval); + refPosition->registerAssignment = foundRegBit; + return foundReg; +} - preferences &= candidates; - if (preferences == RBM_NONE) +//------------------------------------------------------------------------ +// canSpillReg: Determine whether we can spill physRegRecord +// +// Arguments: +// physRegRecord - reg to spill +// refLocation - Location of RefPosition where this register will be spilled +// +// Return Value: +// True - if we can spill physRegRecord +// False - otherwise +// +bool LinearScan::canSpillReg(RegRecord* physRegRecord, LsraLocation refLocation) +{ + assert(physRegRecord->assignedInterval != nullptr); + RefPosition* recentAssignedRef = physRegRecord->assignedInterval->recentRefPosition; + + if (recentAssignedRef != nullptr) { - preferences = candidates; + // We can't spill a register that's active at the current location. + // We should already have determined this with isRegBusy before calling this method. + assert(!isRefPositionActive(recentAssignedRef, refLocation)); + return true; } + // recentAssignedRef can only be null if this is a parameter that has not yet been + // moved to a register (or stack), in which case we can't spill it yet. + assert(physRegRecord->assignedInterval->getLocalVar(compiler)->lvIsParam); + return false; +} -#ifdef DEBUG - candidates = stressLimitRegs(refPosition, candidates); -#endif - assert(candidates != RBM_NONE); +//------------------------------------------------------------------------ +// getSpillWeight: Get the weight associated with spilling the given register +// +// Arguments: +// physRegRecord - reg to spill +// +// Return Value: +// The weight associated with the location at which we will spill. +// +// Note: This helper is designed to be used only from allocateReg() and getDoubleSpillWeight() +// +float LinearScan::getSpillWeight(RegRecord* physRegRecord) +{ + assert(physRegRecord->assignedInterval != nullptr); + RefPosition* recentAssignedRef = physRegRecord->assignedInterval->recentRefPosition; + float weight = BB_ZERO_WEIGHT; - Interval* relatedInterval = currentInterval->relatedInterval; - if (currentInterval->isSpecialPutArg) - { - // This is not actually a preference, it's merely to track the lclVar that this - // "specialPutArg" is using. - relatedInterval = nullptr; - } - Interval* nextRelatedInterval = relatedInterval; - Interval* finalRelatedInterval = relatedInterval; - Interval* rangeEndInterval = relatedInterval; - regMaskTP relatedPreferences = (relatedInterval == nullptr) ? RBM_NONE : relatedInterval->getCurrentPreferences(); - LsraLocation rangeEndLocation = refPosition->getRangeEndLocation(); - LsraLocation relatedLastLocation = rangeEndLocation; + // We shouldn't call this method if there is no recentAssignedRef. + assert(recentAssignedRef != nullptr); + // We shouldn't call this method if the register is active at this location. + assert(!isRefPositionActive(recentAssignedRef, currentLoc)); + weight = getWeight(recentAssignedRef); + return weight; +} - bool preferCalleeSave = currentInterval->preferCalleeSave; +#ifdef TARGET_ARM +//------------------------------------------------------------------------ +// canSpillDoubleReg: Determine whether we can spill physRegRecord +// +// Arguments: +// physRegRecord - reg to spill (must be a valid double register) +// refLocation - Location of RefPosition where this register will be spilled +// +// Return Value: +// True - if we can spill physRegRecord +// False - otherwise +// +bool LinearScan::canSpillDoubleReg(RegRecord* physRegRecord, LsraLocation refLocation) +{ + assert(genIsValidDoubleReg(physRegRecord->regNum)); + RegRecord* physRegRecord2 = getSecondHalfRegRec(physRegRecord); - bool avoidByteRegs = false; -#ifdef TARGET_X86 - if ((relatedPreferences & ~RBM_BYTE_REGS) != RBM_NONE) + if ((physRegRecord->assignedInterval != nullptr) && !canSpillReg(physRegRecord, refLocation)) { - avoidByteRegs = true; + return false; } -#endif - - // Follow the chain of related intervals, as long as: - // - The next reference is a def. We don't want to use the relatedInterval for preferencing if its next reference - // is not a new definition (as it either is or will become live). - // - The next (def) reference is downstream. Otherwise we could iterate indefinitely because the preferences can be - // circular. - // - The intersection of preferenced registers is non-empty. - // - while (nextRelatedInterval != nullptr) + if ((physRegRecord2->assignedInterval != nullptr) && !canSpillReg(physRegRecord2, refLocation)) { - RefPosition* nextRelatedRefPosition = nextRelatedInterval->getNextRefPosition(); + return false; + } + return true; +} - // Only use the relatedInterval for preferencing if the related interval's next reference - // is a new definition. - if ((nextRelatedRefPosition != nullptr) && RefTypeIsDef(nextRelatedRefPosition->refType)) - { - finalRelatedInterval = nextRelatedInterval; - nextRelatedInterval = nullptr; +//------------------------------------------------------------------------ +// unassignDoublePhysReg: unassign a double register (pair) +// +// Arguments: +// doubleRegRecord - reg to unassign +// +// Note: +// The given RegRecord must be a valid (even numbered) double register. +// +void LinearScan::unassignDoublePhysReg(RegRecord* doubleRegRecord) +{ + assert(genIsValidDoubleReg(doubleRegRecord->regNum)); - // First, get the preferences for this interval - regMaskTP thisRelatedPreferences = finalRelatedInterval->getCurrentPreferences(); - // Now, determine if they are compatible and update the relatedPreferences that we'll consider. - regMaskTP newRelatedPreferences = thisRelatedPreferences & relatedPreferences; - if (newRelatedPreferences != RBM_NONE && (!avoidByteRegs || thisRelatedPreferences != RBM_BYTE_REGS)) - { - // TODO-CQ: The following isFree() check doesn't account for the possibility that there's an - // assignedInterval whose recentRefPosition was delayFree. It also fails to account for - // the TYP_DOUBLE case on ARM. It would be better to replace the call to isFree with - // isRegAvailable(genRegNumFromMask(newRelatedPreferences), regType)), but this is retained - // to achieve zero diffs. - // - bool thisIsSingleReg = isSingleRegister(newRelatedPreferences); - if (!thisIsSingleReg || (finalRelatedInterval->isLocalVar && - isFree(getRegisterRecord(genRegNumFromMask(newRelatedPreferences))))) - { - relatedPreferences = newRelatedPreferences; - // If this Interval has a downstream def without a single-register preference, continue to iterate. - if (nextRelatedRefPosition->nodeLocation > rangeEndLocation) - { - preferCalleeSave = (preferCalleeSave || finalRelatedInterval->preferCalleeSave); - rangeEndLocation = nextRelatedRefPosition->getRangeEndLocation(); - rangeEndInterval = finalRelatedInterval; - nextRelatedInterval = finalRelatedInterval->relatedInterval; - } - } - } + RegRecord* doubleRegRecordLo = doubleRegRecord; + RegRecord* doubleRegRecordHi = getSecondHalfRegRec(doubleRegRecordLo); + // For a double register, we has following four cases. + // Case 1: doubleRegRecLo is assigned to TYP_DOUBLE interval + // Case 2: doubleRegRecLo and doubleRegRecHi are assigned to different TYP_FLOAT intervals + // Case 3: doubelRegRecLo is assgined to TYP_FLOAT interval and doubleRegRecHi is nullptr + // Case 4: doubleRegRecordLo is nullptr, and doubleRegRecordHi is assigned to a TYP_FLOAT interval + if (doubleRegRecordLo->assignedInterval != nullptr) + { + if (doubleRegRecordLo->assignedInterval->registerType == TYP_DOUBLE) + { + // Case 1: doubleRegRecLo is assigned to TYP_DOUBLE interval + unassignPhysReg(doubleRegRecordLo, doubleRegRecordLo->assignedInterval->recentRefPosition); } else { - if (nextRelatedInterval == relatedInterval) + // Case 2: doubleRegRecLo and doubleRegRecHi are assigned to different TYP_FLOAT intervals + // Case 3: doubelRegRecLo is assgined to TYP_FLOAT interval and doubleRegRecHi is nullptr + assert(doubleRegRecordLo->assignedInterval->registerType == TYP_FLOAT); + unassignPhysReg(doubleRegRecordLo, doubleRegRecordLo->assignedInterval->recentRefPosition); + + if (doubleRegRecordHi != nullptr) { - relatedInterval = nullptr; - relatedPreferences = RBM_NONE; + if (doubleRegRecordHi->assignedInterval != nullptr) + { + assert(doubleRegRecordHi->assignedInterval->registerType == TYP_FLOAT); + unassignPhysReg(doubleRegRecordHi, doubleRegRecordHi->assignedInterval->recentRefPosition); + } } - nextRelatedInterval = nullptr; } } - - // For floating point, we want to be less aggressive about using callee-save registers. - // So in that case, we just need to ensure that the current RefPosition is covered. - RefPosition* rangeEndRefPosition; - RefPosition* lastRefPosition = currentInterval->lastRefPosition; - if (useFloatReg(currentInterval->registerType)) - { - rangeEndRefPosition = refPosition; - preferCalleeSave = currentInterval->preferCalleeSave; - } - else if (currentInterval->isWriteThru && refPosition->spillAfter) - { - // This is treated as a last use of the register, as there is an upcoming EH boundary. - rangeEndRefPosition = refPosition; - } else { - rangeEndRefPosition = refPosition->getRangeEndRef(); - // If we have a chain of related intervals, and a finalRelatedInterval that - // is not currently occupying a register, and whose lifetime begins after this one, - // we want to try to select a register that will cover its lifetime. - if ((rangeEndInterval != nullptr) && (rangeEndInterval->assignedReg == nullptr) && - !rangeEndInterval->isWriteThru && - (rangeEndInterval->getNextRefLocation() >= rangeEndRefPosition->nodeLocation)) - { - lastRefPosition = rangeEndInterval->lastRefPosition; - } - } - if ((relatedInterval != nullptr) && !relatedInterval->isWriteThru) - { - relatedLastLocation = relatedInterval->lastRefPosition->nodeLocation; + // Case 4: doubleRegRecordLo is nullptr, and doubleRegRecordHi is assigned to a TYP_FLOAT interval + assert(doubleRegRecordHi->assignedInterval != nullptr); + assert(doubleRegRecordHi->assignedInterval->registerType == TYP_FLOAT); + unassignPhysReg(doubleRegRecordHi, doubleRegRecordHi->assignedInterval->recentRefPosition); } +} - regMaskTP callerCalleePrefs; - if (preferCalleeSave) +#endif // TARGET_ARM + +//------------------------------------------------------------------------ +// isRefPositionActive: Determine whether a given RefPosition is active at the given location +// +// Arguments: +// refPosition - the RefPosition of interest +// refLocation - the LsraLocation at which we want to know if it is active +// +// Return Value: +// True - if this RefPosition occurs at the given location, OR +// if it occurs at the previous location and is marked delayRegFree. +// False - otherwise +// +bool LinearScan::isRefPositionActive(RefPosition* refPosition, LsraLocation refLocation) +{ + return (refPosition->nodeLocation == refLocation || + ((refPosition->nodeLocation + 1 == refLocation) && refPosition->delayRegFree)); +} + +//------------------------------------------------------------------------ +// isSpillCandidate: Determine if a register is a spill candidate for a given RefPosition. +// +// Arguments: +// current The interval for the current allocation +// refPosition The RefPosition of the current Interval for which a register is being allocated +// physRegRecord The RegRecord for the register we're considering for spill +// +// Return Value: +// True iff the given register can be spilled to accommodate the given RefPosition. +// +bool LinearScan::isSpillCandidate(Interval* current, RefPosition* refPosition, RegRecord* physRegRecord) +{ + regMaskTP candidateBit = genRegMask(physRegRecord->regNum); + LsraLocation refLocation = refPosition->nodeLocation; + // We shouldn't be calling this if we haven't already determined that the register is not + // busy until the next kill. + assert(!isRegBusy(physRegRecord->regNum, current->registerType)); + // We should already have determined that the register isn't actively in use. + assert(!isRegInUse(physRegRecord->regNum, current->registerType)); + // We shouldn't be calling this if 'refPosition' is a fixed reference to this register. + assert(!refPosition->isFixedRefOfRegMask(candidateBit)); + // We shouldn't be calling this if there is a fixed reference at the same location + // (and it's not due to this reference), as checked above. + assert(!conflictingFixedRegReference(physRegRecord->regNum, refPosition)); + + bool canSpill; +#ifdef TARGET_ARM + if (current->registerType == TYP_DOUBLE) { - regMaskTP calleeSaveCandidates = calleeSaveRegs(currentInterval->registerType); - if (currentInterval->isWriteThru) - { - // We'll only prefer a callee-save register if it's already been used. - regMaskTP unusedCalleeSaves = calleeSaveCandidates & ~(compiler->codeGen->regSet.rsGetModifiedRegsMask()); - callerCalleePrefs = calleeSaveCandidates & ~unusedCalleeSaves; - preferences &= ~unusedCalleeSaves; - } - else - { - callerCalleePrefs = calleeSaveCandidates; - } + canSpill = canSpillDoubleReg(physRegRecord, refLocation); } else +#endif // TARGET_ARM { - callerCalleePrefs = callerSaveRegs(currentInterval->registerType); + canSpill = canSpillReg(physRegRecord, refLocation); + } + if (!canSpill) + { + return false; } + return true; +} - // If this has a delayed use (due to being used in a rmw position of a - // non-commutative operator), its endLocation is delayed until the "def" - // position, which is one location past the use (getRefEndLocation() takes care of this). - rangeEndLocation = rangeEndRefPosition->getRefEndLocation(); - LsraLocation lastLocation = lastRefPosition->getRefEndLocation(); - RegRecord* prevRegRec = currentInterval->assignedReg; +// Grab a register to use to copy and then immediately use. +// This is called only for localVar intervals that already have a register +// assignment that is not compatible with the current RefPosition. +// This is not like regular assignment, because we don't want to change +// any preferences or existing register assignments. +// Prefer a free register that's got the earliest next use. +// Otherwise, spill something with the farthest next use +// +regNumber LinearScan::assignCopyReg(RefPosition* refPosition) +{ + Interval* currentInterval = refPosition->getInterval(); + assert(currentInterval != nullptr); + assert(currentInterval->isActive); - //------------------------------------------------------------------------- - // Register Selection + // Save the relatedInterval, if any, so that it doesn't get modified during allocation. + Interval* savedRelatedInterval = currentInterval->relatedInterval; + currentInterval->relatedInterval = nullptr; + + // We don't want really want to change the default assignment, + // so 1) pretend this isn't active, and 2) remember the old reg + regNumber oldPhysReg = currentInterval->physReg; + RegRecord* oldRegRecord = currentInterval->assignedReg; + assert(oldRegRecord->regNum == oldPhysReg); + currentInterval->isActive = false; - RegRecord* availablePhysRegRecord = nullptr; + // We *must* allocate a register, and it will be a copyReg. Set that field now, so that + // refPosition->RegOptional() will return false. + refPosition->copyReg = true; - // These are used in the post-selection updates, and must be set for any selection. - regMaskTP freeCandidates = RBM_NONE; - regMaskTP matchingConstants = RBM_NONE; - regMaskTP unassignedSet = RBM_NONE; + RegisterScore registerScore = NONE; + regNumber allocatedReg = allocateReg(currentInterval, refPosition DEBUG_ARG(®isterScore)); + assert(allocatedReg != REG_NA); - // We'll set this to short-circuit remaining heuristics when we have a single candidate. - bool found = false; + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_COPY_REG, currentInterval, allocatedReg, nullptr, registerScore)); - // Is this a fixedReg? - regMaskTP fixedRegMask = RBM_NONE; - if (refPosition->isFixedRegRef) + // Now restore the old info + currentInterval->relatedInterval = savedRelatedInterval; + currentInterval->physReg = oldPhysReg; + currentInterval->assignedReg = oldRegRecord; + currentInterval->isActive = true; + + return allocatedReg; +} + +//------------------------------------------------------------------------ +// isAssigned: This is the function to check if the given RegRecord has an assignedInterval. +// +// Arguments: +// regRec - The RegRecord to check that it is assigned. +// newRegType - There are elements to judge according to the upcoming register type. +// +// Return Value: +// Returns true if the given RegRecord has an assignedInterval. +// +bool LinearScan::isAssigned(RegRecord* regRec ARM_ARG(RegisterType newRegType)) +{ + if (regRec->assignedInterval != nullptr) { - assert(genMaxOneBit(refPosition->registerAssignment)); - fixedRegMask = refPosition->registerAssignment; - if (candidates == refPosition->registerAssignment) - { - found = true; - if (nextIntervalRef[genRegNumFromMask(candidates)] > lastLocation) - { - unassignedSet = candidates; - } - } + return true; } - - // Eliminate candidates that are in-use or busy. - if (!found) +#ifdef TARGET_ARM + if (newRegType == TYP_DOUBLE) { - regMaskTP busyRegs = regsBusyUntilKill | regsInUseThisLocation; - candidates &= ~busyRegs; + RegRecord* otherRegRecord = getSecondHalfRegRec(regRec); - // Also eliminate as busy any register with a conflicting fixed reference at this or - // the next location. - // Note that this will eliminate the fixedReg, if any, but we'll add it back below. - regMaskTP checkConflictMask = candidates & fixedRegs; - while (checkConflictMask != RBM_NONE) + if (otherRegRecord->assignedInterval != nullptr) { - regMaskTP checkConflictBit = genFindLowestBit(checkConflictMask); - checkConflictMask &= ~checkConflictBit; - regNumber checkConflictReg = genRegNumFromMask(checkConflictBit); - LsraLocation checkConflictLocation = nextFixedRef[checkConflictReg]; - - if ((checkConflictLocation == currentLocation) || - (refPosition->delayRegFree && (checkConflictLocation == (currentLocation + 1)))) - { - candidates &= ~checkConflictBit; - } + return true; } - candidates |= fixedRegMask; - found = isSingleRegister(candidates); } +#endif + return false; +} - // By chance, is prevRegRec already holding this interval, as a copyReg or having - // been restored as inactive after a kill? - // NOTE: this is not currently considered one of the selection criteria - it always wins - // if it is the assignedInterval of 'prevRegRec'. - regMaskTP prevRegBit = RBM_NONE; - if (!found && (prevRegRec != nullptr)) +//------------------------------------------------------------------------ +// checkAndAssignInterval: Check if the interval is already assigned and +// if it is then unassign the physical record +// and set the assignedInterval to 'interval' +// +// Arguments: +// regRec - The RegRecord of interest +// interval - The Interval that we're going to assign to 'regRec' +// +void LinearScan::checkAndAssignInterval(RegRecord* regRec, Interval* interval) +{ + Interval* assignedInterval = regRec->assignedInterval; + if (assignedInterval != nullptr && assignedInterval != interval) { - prevRegBit = genRegMask(prevRegRec->regNum); - if ((prevRegRec->assignedInterval == currentInterval) && ((candidates & prevRegBit) != RBM_NONE)) + // This is allocated to another interval. Either it is inactive, or it was allocated as a + // copyReg and is therefore not the "assignedReg" of the other interval. In the latter case, + // we simply unassign it - in the former case we need to set the physReg on the interval to + // REG_NA to indicate that it is no longer in that register. + // The lack of checking for this case resulted in an assert in the retail version of System.dll, + // in method SerialStream.GetDcbFlag. + // Note that we can't check for the copyReg case, because we may have seen a more recent + // RefPosition for the Interval that was NOT a copyReg. + if (assignedInterval->assignedReg == regRec) { - candidates = prevRegBit; - found = true; -#ifdef DEBUG - *registerScore = THIS_ASSIGNED; -#endif + assert(assignedInterval->isActive == false); + assignedInterval->physReg = REG_NA; } + unassignPhysReg(regRec->regNum); } - - if (!found && (candidates == RBM_NONE)) +#ifdef TARGET_ARM + // If 'interval' and 'assignedInterval' were both TYP_DOUBLE, then we have unassigned 'assignedInterval' + // from both halves. Otherwise, if 'interval' is TYP_DOUBLE, we now need to unassign the other half. + if ((interval->registerType == TYP_DOUBLE) && + ((assignedInterval == nullptr) || (assignedInterval->registerType == TYP_FLOAT))) { - assert(refPosition->RegOptional()); - currentInterval->assignedReg = nullptr; - return REG_NA; + RegRecord* otherRegRecord = getSecondHalfRegRec(regRec); + assignedInterval = otherRegRecord->assignedInterval; + if (assignedInterval != nullptr && assignedInterval != interval) + { + if (assignedInterval->assignedReg == otherRegRecord) + { + assert(assignedInterval->isActive == false); + assignedInterval->physReg = REG_NA; + } + unassignPhysReg(otherRegRecord->regNum); + } } +#endif - // TODO-Cleanup: Previously, the "reverseSelect" stress mode reversed the order of the heuristics. - // It needs to be re-engineered with this refactoring. - // In non-debug builds, this will simply get optimized away - bool reverseSelect = false; -#ifdef DEBUG - reverseSelect = doReverseSelect(); -#endif // DEBUG - - // At this point we have determined all the candidates that CAN be considered, - // and we'll begin to apply the heuristics in order. - registerSelector selector; - selector.candidates = candidates; - selector.score = 0; -#ifdef TARGET_ARM - selector.regType = regType; -#endif // TARGET_ARM - - freeCandidates = getFreeCandidates(candidates, regType); + updateAssignedInterval(regRec, interval, interval->registerType); +} - // Apply the FREE heuristic. - if (freeCandidates == RBM_NONE) - { - // We won't spill if this refPosition is not an actual ref. - if (!refPosition->IsActualRef()) - { - currentInterval->assignedReg = nullptr; - return REG_NA; - } - } - else if (!found) +// Assign the given physical register interval to the given interval +void LinearScan::assignPhysReg(RegRecord* regRec, Interval* interval) +{ + regMaskTP assignedRegMask = genRegMask(regRec->regNum); + compiler->codeGen->regSet.rsSetRegsModified(assignedRegMask DEBUGARG(true)); + + interval->assignedReg = regRec; + checkAndAssignInterval(regRec, interval); + + interval->physReg = regRec->regNum; + interval->isActive = true; + if (interval->isLocalVar) { - found = selector.applySelection(FREE, freeCandidates DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_FREE, refPosition->bbNum)); + // Prefer this register for future references + interval->updateRegisterPreferences(assignedRegMask); } +} - // Apply the CONST_AVAILABLE (matching constant) heuristic. Only applies if we have freeCandidates. - // Note that we always need to define the 'matchingConstants' set. - if (freeCandidates != RBM_NONE) +//------------------------------------------------------------------------ +// setIntervalAsSplit: Set this Interval as being split +// +// Arguments: +// interval - The Interval which is being split +// +// Return Value: +// None. +// +// Notes: +// The given Interval will be marked as split, and it will be added to the +// set of splitOrSpilledVars. +// +// Assumptions: +// "interval" must be a lclVar interval, as tree temps are never split. +// This is asserted in the call to getVarIndex(). +// +void LinearScan::setIntervalAsSplit(Interval* interval) +{ + if (interval->isLocalVar) { - if (currentInterval->isConstant && RefTypeIsDef(refPosition->refType)) + unsigned varIndex = interval->getVarIndex(compiler); + if (!interval->isSplit) + { + VarSetOps::AddElemD(compiler, splitOrSpilledVars, varIndex); + } + else { - matchingConstants = getMatchingConstants(selector.candidates, currentInterval, refPosition); - found = selector.applySelection(CONST_AVAILABLE, matchingConstants DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_CONST_AVAILABLE, refPosition->bbNum)); + assert(VarSetOps::IsMember(compiler, splitOrSpilledVars, varIndex)); } } + interval->isSplit = true; +} - // Apply the THIS_ASSIGNED heuristic. Only applies if we have freeCandidates. - if (!found && (prevRegRec != nullptr) && (freeCandidates != RBM_NONE)) +//------------------------------------------------------------------------ +// setIntervalAsSpilled: Set this Interval as being spilled +// +// Arguments: +// interval - The Interval which is being spilled +// +// Return Value: +// None. +// +// Notes: +// The given Interval will be marked as spilled, and it will be added +// to the set of splitOrSpilledVars. +// +void LinearScan::setIntervalAsSpilled(Interval* interval) +{ +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + if (interval->isUpperVector) { - found = - selector.applySelection(THIS_ASSIGNED, freeCandidates & preferences & prevRegBit DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_THIS_ASSIGNED, refPosition->bbNum)); + assert(interval->relatedInterval->isLocalVar); + interval->isSpilled = true; + // Now we need to mark the local as spilled also, even if the lower half is never spilled, + // as this will use the upper part of its home location. + interval = interval->relatedInterval; + // We'll now mark this as spilled, so it changes the spillCost. + RefPosition* recentRefPos = interval->recentRefPosition; + if (!interval->isSpilled && interval->isActive && (recentRefPos != nullptr)) + { + VarSetOps::AddElemD(compiler, splitOrSpilledVars, interval->getVarIndex(compiler)); + interval->isSpilled = true; + regNumber reg = interval->physReg; + spillCost[reg] = getSpillWeight(getRegisterRecord(reg)); + } } - - // Compute the sets for COVERS, OWN_PREFERENCE, COVERS_RELATED, COVERS_FULL and UNASSIGNED together, - // as they all require similar computation. - regMaskTP coversSet = RBM_NONE; - regMaskTP preferenceSet = RBM_NONE; - regMaskTP coversRelatedSet = RBM_NONE; - regMaskTP coversFullSet = RBM_NONE; - if (freeCandidates != RBM_NONE) +#endif + if (interval->isLocalVar) { - preferenceSet = (selector.candidates & preferences); - regMaskTP coversCandidates = (preferenceSet == RBM_NONE) ? selector.candidates : preferenceSet; - for (; coversCandidates != RBM_NONE;) + unsigned varIndex = interval->getVarIndex(compiler); + if (!interval->isSpilled) { - regMaskTP coversCandidateBit = genFindLowestBit(coversCandidates); - coversCandidates &= ~coversCandidateBit; - regNumber coversCandidateRegNum = genRegNumFromMask(coversCandidateBit); - - // If we have a single candidate we don't need to compute the preference-related sets, but we - // do need to compute the unassignedSet. - if (!found) - { - // Find the next RefPosition of the register. - LsraLocation nextIntervalLocation = getNextIntervalRef(coversCandidateRegNum, regType); - LsraLocation nextPhysRefLocation = getNextFixedRef(coversCandidateRegNum, regType); - LsraLocation coversCandidateLocation = Min(nextPhysRefLocation, nextIntervalLocation); - - // If the nextPhysRefLocation is a fixedRef for the rangeEndRefPosition, increment it so that - // we don't think it isn't covering the live range. - // This doesn't handle the case where earlier RefPositions for this Interval are also - // FixedRefs of this regNum, but at least those are only interesting in the case where those - // are "local last uses" of the Interval - otherwise the liveRange would interfere with the reg. - if (coversCandidateLocation == rangeEndLocation && - rangeEndRefPosition->isFixedRefOfReg(coversCandidateRegNum)) - { - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_INCREMENT_RANGE_END, currentInterval)); - coversCandidateLocation++; - } - if (coversCandidateLocation > rangeEndLocation) - { - coversSet |= coversCandidateBit; - } - if ((coversCandidateBit & relatedPreferences) != RBM_NONE) - { - if (coversCandidateLocation > relatedInterval->lastRefPosition->nodeLocation) - { - coversRelatedSet |= coversCandidateBit; - } - } - else if (coversCandidateBit == refPosition->registerAssignment) - { - // If we had a fixed-reg def of a reg that will be killed before the use, prefer it to any other - // registers with the same score. (Note that we haven't changed the original registerAssignment - // on the RefPosition). - // Overload the RELATED_PREFERENCE value. - // TODO-CQ: Consider if this should be split out. - coversRelatedSet |= coversCandidateBit; - } - // Does this cover the full range of the interval? - if (coversCandidateLocation > lastLocation) - { - coversFullSet |= coversCandidateBit; - } - } - // The register is considered unassigned if it has no assignedInterval, OR - // if its next reference is beyond the range of this interval. - if (nextIntervalRef[coversCandidateRegNum] > lastLocation) - { - unassignedSet |= coversCandidateBit; - } + VarSetOps::AddElemD(compiler, splitOrSpilledVars, varIndex); + } + else + { + assert(VarSetOps::IsMember(compiler, splitOrSpilledVars, varIndex)); } } + interval->isSpilled = true; +} - // Apply the COVERS heuristic. - if (!found) - { - found = selector.applySelection(COVERS, coversSet & preferenceSet DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_COVERS, refPosition->bbNum)); - } +//------------------------------------------------------------------------ +// spill: Spill the "interval" starting from "fromRefPosition" (upto "toRefPosition") +// +// Arguments: +// interval - The interval that contains the RefPosition to be spilled +// fromRefPosition - The RefPosition at which the Interval is to be spilled +// toRefPosition - The RefPosition at which it must be reloaded (debug only arg) +// +// Return Value: +// None. +// +// Assumptions: +// fromRefPosition and toRefPosition must not be null +// +void LinearScan::spillInterval(Interval* interval, RefPosition* fromRefPosition DEBUGARG(RefPosition* toRefPosition)) +{ + assert(fromRefPosition != nullptr && toRefPosition != nullptr); + assert(fromRefPosition->getInterval() == interval && toRefPosition->getInterval() == interval); + assert(fromRefPosition->nextRefPosition == toRefPosition); - // Apply the OWN_PREFERENCE heuristic. - // Note that 'preferenceSet' already includes only freeCandidates. - if (!found) + if (!fromRefPosition->lastUse) { - assert((preferenceSet & freeCandidates) == preferenceSet); - found = selector.applySelection(OWN_PREFERENCE, preferenceSet DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_OWN_PREFERENCE, refPosition->bbNum)); + // If not allocated a register, Lcl var def/use ref positions even if reg optional + // should be marked as spillAfter. Note that if it is a WriteThru interval, the value is always + // written to the stack, but the WriteThru indicates that the register is no longer live. + if (fromRefPosition->RegOptional() && !(interval->isLocalVar && fromRefPosition->IsActualRef())) + { + fromRefPosition->registerAssignment = RBM_NONE; + } + else + { + fromRefPosition->spillAfter = true; + } } + assert(toRefPosition != nullptr); - // Apply the COVERS_RELATED heuristic. - if (!found) +#ifdef DEBUG + if (VERBOSE) { - assert((coversRelatedSet & freeCandidates) == coversRelatedSet); - found = selector.applySelection(COVERS_RELATED, coversRelatedSet DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_COVERS_RELATED, refPosition->bbNum)); + dumpLsraAllocationEvent(LSRA_EVENT_SPILL, interval); } +#endif // DEBUG - // Apply the RELATED_PREFERENCE heuristic. - if (!found) - { - found = - selector.applySelection(RELATED_PREFERENCE, relatedPreferences & freeCandidates DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_RELATED_PREFERENCE, refPosition->bbNum)); - } + INTRACK_STATS(updateLsraStat(STAT_SPILL, fromRefPosition->bbNum)); - // Apply the CALLER_CALLEE heuristic. - if (!found) - { - found = selector.applySelection(CALLER_CALLEE, callerCalleePrefs & freeCandidates DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_CALLER_CALLEE, refPosition->bbNum)); - } + interval->isActive = false; + setIntervalAsSpilled(interval); - // Apply the UNASSIGNED heuristic. - if (!found) + // If fromRefPosition occurs before the beginning of this block, mark this as living in the stack + // on entry to this block. + if (fromRefPosition->nodeLocation <= curBBStartLocation) { - found = selector.applySelection(UNASSIGNED, unassignedSet DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_UNASSIGNED, refPosition->bbNum)); + // This must be a lclVar interval + assert(interval->isLocalVar); + setInVarRegForBB(curBBNum, interval->varNum, REG_STK); } +} - // Apply the COVERS_FULL heuristic. - if (!found) - { - assert((coversFullSet & freeCandidates) == coversFullSet); - found = selector.applySelection(COVERS_FULL, coversFullSet DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_COVERS_FULL, refPosition->bbNum)); - } +//------------------------------------------------------------------------ +// unassignPhysRegNoSpill: Unassign the given physical register record from +// an active interval, without spilling. +// +// Arguments: +// regRec - the RegRecord to be unassigned +// +// Return Value: +// None. +// +// Assumptions: +// The assignedInterval must not be null, and must be active. +// +// Notes: +// This method is used to unassign a register when an interval needs to be moved to a +// different register, but not (yet) spilled. - // Apply the BEST_FIT heuristic. Only applies if we have freeCandidates. - if (!found && (freeCandidates != RBM_NONE)) - { - regMaskTP bestFitSet = RBM_NONE; - // If the best score includes COVERS_FULL, pick the one that's killed soonest. - // If none cover the full range, the BEST_FIT is the one that's killed later. - bool earliestIsBest = ((selector.score & COVERS_FULL) != 0); - LsraLocation bestFitLocation = earliestIsBest ? MaxLocation : MinLocation; - for (regMaskTP bestFitCandidates = selector.candidates; bestFitCandidates != RBM_NONE;) - { - regMaskTP bestFitCandidateBit = genFindLowestBit(bestFitCandidates); - bestFitCandidates &= ~bestFitCandidateBit; - regNumber bestFitCandidateRegNum = genRegNumFromMask(bestFitCandidateBit); +void LinearScan::unassignPhysRegNoSpill(RegRecord* regRec) +{ + Interval* assignedInterval = regRec->assignedInterval; + assert(assignedInterval != nullptr && assignedInterval->isActive); + assignedInterval->isActive = false; + unassignPhysReg(regRec, nullptr); + assignedInterval->isActive = true; +} - // Find the next RefPosition of the register. - LsraLocation nextIntervalLocation = getNextIntervalRef(bestFitCandidateRegNum, regType); - LsraLocation nextPhysRefLocation = getNextFixedRef(bestFitCandidateRegNum, regType); - nextPhysRefLocation = Min(nextPhysRefLocation, nextIntervalLocation); - // If the nextPhysRefLocation is a fixedRef for the rangeEndRefPosition, increment it so that - // we don't think it isn't covering the live range. - // This doesn't handle the case where earlier RefPositions for this Interval are also - // FixedRefs of this regNum, but at least those are only interesting in the case where those - // are "local last uses" of the Interval - otherwise the liveRange would interfere with the reg. - // TODO: This duplicates code in an earlier loop, and is basically here to duplicate previous - // behavior; see if we can avoid this. - if (nextPhysRefLocation == rangeEndLocation && rangeEndRefPosition->isFixedRefOfReg(bestFitCandidateRegNum)) - { - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_INCREMENT_RANGE_END, currentInterval)); - nextPhysRefLocation++; - } +//------------------------------------------------------------------------ +// checkAndClearInterval: Clear the assignedInterval for the given +// physical register record +// +// Arguments: +// regRec - the physical RegRecord to be unassigned +// spillRefPosition - The RefPosition at which the assignedInterval is to be spilled +// or nullptr if we aren't spilling +// +// Return Value: +// None. +// +// Assumptions: +// see unassignPhysReg +// +void LinearScan::checkAndClearInterval(RegRecord* regRec, RefPosition* spillRefPosition) +{ + Interval* assignedInterval = regRec->assignedInterval; + assert(assignedInterval != nullptr); + regNumber thisRegNum = regRec->regNum; - if (nextPhysRefLocation == bestFitLocation) - { - bestFitSet |= bestFitCandidateBit; - } - else - { - bool isBetter = false; - if (nextPhysRefLocation > lastLocation) - { - // This covers the full range; favor it if the other doesn't, or if it's a closer match. - if ((bestFitLocation <= lastLocation) || (nextPhysRefLocation < bestFitLocation)) - { - isBetter = true; - } - } - else - { - // This doesn't cover the full range; favor it if the other doesn't either, but this ends later. - if ((bestFitLocation <= lastLocation) && (nextPhysRefLocation > bestFitLocation)) - { - isBetter = true; - } - } - if (isBetter) - { - bestFitSet = bestFitCandidateBit; - bestFitLocation = nextPhysRefLocation; - } - } + if (spillRefPosition == nullptr) + { + // Note that we can't assert for the copyReg case + // + if (assignedInterval->physReg == thisRegNum) + { + assert(assignedInterval->isActive == false); } - assert(bestFitSet != RBM_NONE); - found = selector.applySelection(BEST_FIT, bestFitSet DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_BEST_FIT, refPosition->bbNum)); } - - // Apply the IS_PREV_REG heuristic. TODO: Check if Only applies if we have freeCandidates. - // Oddly, the previous heuristics only considered this if it covered the range. - if ((prevRegRec != nullptr) && ((selector.score & COVERS_FULL) != 0)) + else { - found = selector.applySingleRegSelection(IS_PREV_REG, prevRegBit DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_IS_PREV_REG, refPosition->bbNum)); + assert(spillRefPosition->getInterval() == assignedInterval); } - // Apply the REG_ORDER heuristic. Only applies if we have freeCandidates. - if (!found && (freeCandidates != RBM_NONE)) + updateAssignedInterval(regRec, nullptr, assignedInterval->registerType); +} + +//------------------------------------------------------------------------ +// unassignPhysReg: Unassign the given physical register record, and spill the +// assignedInterval at the given spillRefPosition, if any. +// +// Arguments: +// regRec - The RegRecord to be unassigned +// newRegType - The RegisterType of interval that would be assigned +// +// Return Value: +// None. +// +// Notes: +// On ARM architecture, Intervals have to be unassigned considering +// with the register type of interval that would be assigned. +// +void LinearScan::unassignPhysReg(RegRecord* regRec ARM_ARG(RegisterType newRegType)) +{ + RegRecord* regRecToUnassign = regRec; +#ifdef TARGET_ARM + RegRecord* anotherRegRec = nullptr; + + if ((regRecToUnassign->assignedInterval != nullptr) && + (regRecToUnassign->assignedInterval->registerType == TYP_DOUBLE)) { - // This will always result in a single candidate. That is, it is the tie-breaker - // for free candidates, and doesn't make sense as anything other than the last - // heuristic for free registers. - unsigned lowestRegOrder = UINT_MAX; - regMaskTP lowestRegOrderBit = RBM_NONE; - for (regMaskTP regOrderCandidates = selector.candidates; regOrderCandidates != RBM_NONE;) + // If the register type of interval(being unassigned or new) is TYP_DOUBLE, + // It should have to be valid double register (even register) + if (!genIsValidDoubleReg(regRecToUnassign->regNum)) { - regMaskTP regOrderCandidateBit = genFindLowestBit(regOrderCandidates); - regOrderCandidates &= ~regOrderCandidateBit; - regNumber regOrderCandidateRegNum = genRegNumFromMask(regOrderCandidateBit); - unsigned thisRegOrder = getRegisterRecord(regOrderCandidateRegNum)->regOrder; - if (thisRegOrder < lowestRegOrder) - { - lowestRegOrder = thisRegOrder; - lowestRegOrderBit = regOrderCandidateBit; - } + regRecToUnassign = findAnotherHalfRegRec(regRec); + } + } + else + { + if (newRegType == TYP_DOUBLE) + { + anotherRegRec = getSecondHalfRegRec(regRecToUnassign); } - assert(lowestRegOrderBit != RBM_NONE); - found = selector.applySingleRegSelection(REG_ORDER, lowestRegOrderBit DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_REG_ORDER, refPosition->bbNum)); } +#endif - // The set of registers with the lowest spill weight. - regMaskTP lowestCostSpillSet = RBM_NONE; - // Apply the SPILL_COST heuristic and eliminate regs that can't be spilled. - if (!found) + if (regRecToUnassign->assignedInterval != nullptr) + { + unassignPhysReg(regRecToUnassign, regRecToUnassign->assignedInterval->recentRefPosition); + } +#ifdef TARGET_ARM + if ((anotherRegRec != nullptr) && (anotherRegRec->assignedInterval != nullptr)) { - // The spill weight for 'refPosition' (the one we're allocating now). - float thisSpillWeight = getWeight(refPosition); - // The spill weight for the best candidate we've found so far. - float bestSpillWeight = FloatingPointUtils::infinite_float(); + unassignPhysReg(anotherRegRec, anotherRegRec->assignedInterval->recentRefPosition); + } +#endif +} - for (regMaskTP spillCandidates = selector.candidates; spillCandidates != RBM_NONE;) - { - regMaskTP spillCandidateBit = genFindLowestBit(spillCandidates); - spillCandidates &= ~spillCandidateBit; - regNumber spillCandidateRegNum = genRegNumFromMask(spillCandidateBit); - RegRecord* spillCandidateRegRecord = &physRegs[spillCandidateRegNum]; +//------------------------------------------------------------------------ +// unassignPhysReg: Unassign the given physical register record, and spill the +// assignedInterval at the given spillRefPosition, if any. +// +// Arguments: +// regRec - the RegRecord to be unassigned +// spillRefPosition - The RefPosition at which the assignedInterval is to be spilled +// +// Return Value: +// None. +// +// Assumptions: +// The assignedInterval must not be null. +// If spillRefPosition is null, the assignedInterval must be inactive, or not currently +// assigned to this register (e.g. this is a copyReg for that Interval). +// Otherwise, spillRefPosition must be associated with the assignedInterval. +// +void LinearScan::unassignPhysReg(RegRecord* regRec, RefPosition* spillRefPosition) +{ + Interval* assignedInterval = regRec->assignedInterval; + assert(assignedInterval != nullptr); + assert(spillRefPosition == nullptr || spillRefPosition->getInterval() == assignedInterval); + regNumber thisRegNum = regRec->regNum; - // Can and should the interval in this register be spilled for this one, - // if we don't find a better alternative? - if ((getNextIntervalRef(spillCandidateRegNum, regType) == currentLocation) && - !spillCandidateRegRecord->assignedInterval->getNextRefPosition()->RegOptional()) - { - continue; - } - if (!isSpillCandidate(currentInterval, refPosition, spillCandidateRegRecord)) - { - continue; - } + // Is assignedInterval actually still assigned to this register? + bool intervalIsAssigned = (assignedInterval->physReg == thisRegNum); + regNumber regToUnassign = thisRegNum; - float currentSpillWeight = spillCost[spillCandidateRegNum]; #ifdef TARGET_ARM - if (currentInterval->registerType == TYP_DOUBLE) - { - currentSpillWeight = max(currentSpillWeight, spillCost[REG_NEXT(spillCandidateRegNum)]); - } -#endif - if (currentSpillWeight < bestSpillWeight) - { - bestSpillWeight = currentSpillWeight; - lowestCostSpillSet = spillCandidateBit; - } - else if (currentSpillWeight == bestSpillWeight) - { - lowestCostSpillSet |= spillCandidateBit; - } + RegRecord* anotherRegRec = nullptr; + + // Prepare second half RegRecord of a double register for TYP_DOUBLE + if (assignedInterval->registerType == TYP_DOUBLE) + { + assert(isFloatRegType(regRec->registerType)); + RegRecord* doubleRegRec; + if (genIsValidDoubleReg(thisRegNum)) + { + anotherRegRec = getSecondHalfRegRec(regRec); + doubleRegRec = regRec; + } + else + { + regToUnassign = REG_PREV(thisRegNum); + anotherRegRec = getRegisterRecord(regToUnassign); + doubleRegRec = anotherRegRec; } - // We won't spill if this refPosition is RegOptional() and we have no candidates - // with a lower spill cost. - if ((bestSpillWeight >= thisSpillWeight) && refPosition->RegOptional()) + // Both RegRecords should have been assigned to the same interval. + assert(assignedInterval == anotherRegRec->assignedInterval); + if (!intervalIsAssigned && (assignedInterval->physReg == anotherRegRec->regNum)) { - currentInterval->assignedReg = nullptr; - return REG_NA; + intervalIsAssigned = true; } + + clearNextIntervalRef(regToUnassign, TYP_DOUBLE); + clearSpillCost(regToUnassign, TYP_DOUBLE); + checkAndClearInterval(doubleRegRec, spillRefPosition); + + // Both RegRecords should have been unassigned together. + assert(regRec->assignedInterval == nullptr); + assert(anotherRegRec->assignedInterval == nullptr); + } + else +#endif // TARGET_ARM + { + clearNextIntervalRef(thisRegNum, assignedInterval->registerType); + clearSpillCost(thisRegNum, assignedInterval->registerType); + checkAndClearInterval(regRec, spillRefPosition); } + makeRegAvailable(regToUnassign, assignedInterval->registerType); - if (!found) + RefPosition* nextRefPosition = nullptr; + if (spillRefPosition != nullptr) { - // We must have at least one with the lowest spill cost. - assert(lowestCostSpillSet != RBM_NONE); - found = selector.applySelection(SPILL_COST, lowestCostSpillSet DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_SPILL_COST, refPosition->bbNum)); + nextRefPosition = spillRefPosition->nextRefPosition; } - // Apply the FAR_NEXT_REF heuristic. - if (!found) + if (!intervalIsAssigned && assignedInterval->physReg != REG_NA) { - LsraLocation farthestLocation = MinLocation; - regMaskTP farthestSet = RBM_NONE; - for (regMaskTP farthestCandidates = selector.candidates; farthestCandidates != RBM_NONE;) - { - regMaskTP farthestCandidateBit = genFindLowestBit(farthestCandidates); - farthestCandidates &= ~farthestCandidateBit; - regNumber farthestCandidateRegNum = genRegNumFromMask(farthestCandidateBit); + // This must have been a temporary copy reg, but we can't assert that because there + // may have been intervening RefPositions that were not copyRegs. - // Find the next RefPosition of the register. - LsraLocation nextIntervalLocation = - getNextIntervalRef(farthestCandidateRegNum, currentInterval->registerType); - LsraLocation nextPhysRefLocation = Min(nextFixedRef[farthestCandidateRegNum], nextIntervalLocation); - if (nextPhysRefLocation == farthestLocation) - { - farthestSet |= farthestCandidateBit; - } - else if (nextPhysRefLocation > farthestLocation) - { - farthestSet = farthestCandidateBit; - farthestLocation = nextPhysRefLocation; - } - } - // We must have at least one with the lowest spill cost. - assert(farthestSet != RBM_NONE); - found = selector.applySelection(FAR_NEXT_REF, farthestSet DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_FAR_NEXT_REF, refPosition->bbNum)); + // reg->assignedInterval has already been set to nullptr by checkAndClearInterval() + assert(regRec->assignedInterval == nullptr); + return; } - // Apply the PREV_REG_OPT heuristic. - if (!found) + // regNumber victimAssignedReg = assignedInterval->physReg; + assignedInterval->physReg = REG_NA; + + bool spill = assignedInterval->isActive && nextRefPosition != nullptr; + if (spill) { - regMaskTP prevRegOptSet = RBM_NONE; - for (regMaskTP prevRegOptCandidates = selector.candidates; prevRegOptCandidates != RBM_NONE;) - { - regMaskTP prevRegOptCandidateBit = genFindLowestBit(prevRegOptCandidates); - prevRegOptCandidates &= ~prevRegOptCandidateBit; - regNumber prevRegOptCandidateRegNum = genRegNumFromMask(prevRegOptCandidateBit); - Interval* assignedInterval = physRegs[prevRegOptCandidateRegNum].assignedInterval; - bool foundPrevRegOptReg = true; -#ifdef DEBUG - bool hasAssignedInterval = false; -#endif + // If this is an active interval, it must have a recentRefPosition, + // otherwise it would not be active + assert(spillRefPosition != nullptr); - if ((assignedInterval != nullptr) && (assignedInterval->recentRefPosition != nullptr)) - { - foundPrevRegOptReg &= - (assignedInterval->recentRefPosition->reload && assignedInterval->recentRefPosition->RegOptional()); -#ifdef DEBUG - hasAssignedInterval = true; -#endif - } -#ifndef TARGET_ARM - else +#if 0 + // TODO-CQ: Enable this and insert an explicit GT_COPY (otherwise there's no way to communicate + // to codegen that we want the copyReg to be the new home location). + // If the last reference was a copyReg, and we're spilling the register + // it was copied from, then make the copyReg the new primary location + // if possible + if (spillRefPosition->copyReg) + { + regNumber copyFromRegNum = victimAssignedReg; + regNumber copyRegNum = genRegNumFromMask(spillRefPosition->registerAssignment); + if (copyFromRegNum == thisRegNum && + getRegisterRecord(copyRegNum)->assignedInterval == assignedInterval) { - foundPrevRegOptReg = false; + assert(copyRegNum != thisRegNum); + assignedInterval->physReg = copyRegNum; + assignedInterval->assignedReg = this->getRegisterRecord(copyRegNum); + return; } -#endif - -#ifdef TARGET_ARM - // If current interval is TYP_DOUBLE, verify if the other half register matches the heuristics. - // We have three cases: - // 1. One of the register of the pair have an assigned interval: Check if that register's refPosition - // matches the heuristics. If yes, add it to the set. - // 2. Both registers of the pair have an assigned interval: Conservatively "and" conditions for - // heuristics of their corresponding refPositions. If both register's heuristic matches, add them - // to the set. TODO-CQ-ARM: We may implement a better condition later. - // 3. None of the register have an assigned interval: Skip adding register and assert. - if (currentInterval->registerType == TYP_DOUBLE) + } +#endif // 0 +#ifdef DEBUG + // With JitStressRegs == 0x80 (LSRA_EXTEND_LIFETIMES), we may have a RefPosition + // that is not marked lastUse even though the treeNode is a lastUse. In that case + // we must not mark it for spill because the register will have been immediately freed + // after use. While we could conceivably add special handling for this case in codegen, + // it would be messy and undesirably cause the "bleeding" of LSRA stress modes outside + // of LSRA. + if (extendLifetimes() && assignedInterval->isLocalVar && RefTypeIsUse(spillRefPosition->refType) && + spillRefPosition->treeNode != nullptr && + spillRefPosition->treeNode->AsLclVar()->IsLastUse(spillRefPosition->multiRegIdx)) + { + dumpLsraAllocationEvent(LSRA_EVENT_SPILL_EXTENDED_LIFETIME, assignedInterval); + assignedInterval->isActive = false; + spill = false; + // If the spillRefPosition occurs before the beginning of this block, it will have + // been marked as living in this register on entry to this block, but we now need + // to mark this as living on the stack. + if (spillRefPosition->nodeLocation <= curBBStartLocation) { - regNumber anotherHalfRegNum = findAnotherHalfRegNum(prevRegOptCandidateRegNum); - assignedInterval = physRegs[anotherHalfRegNum].assignedInterval; - if ((assignedInterval != nullptr) && (assignedInterval->recentRefPosition != nullptr)) + setInVarRegForBB(curBBNum, assignedInterval->varNum, REG_STK); + if (spillRefPosition->nextRefPosition != nullptr) { - if (assignedInterval->recentRefPosition->reload && - assignedInterval->recentRefPosition->RegOptional()) - { - foundPrevRegOptReg &= (assignedInterval->recentRefPosition->reload && - assignedInterval->recentRefPosition->RegOptional()); - } -#ifdef DEBUG - hasAssignedInterval = true; -#endif + setIntervalAsSpilled(assignedInterval); } } -#endif - - if (foundPrevRegOptReg) - { - // TODO-Cleanup: Previously, we always used the highest regNum with a previous regOptional - // RefPosition, which is not really consistent with the way other selection criteria are - // applied. should probably be: prevRegOptSet |= prevRegOptCandidateBit; - prevRegOptSet = prevRegOptCandidateBit; - } - -#ifdef DEBUG - // The assigned should be non-null, and should have a recentRefPosition, however since - // this is a heuristic, we don't want a fatal error, so we just assert (not noway_assert). - if (!hasAssignedInterval) + else { - assert(!"Spill candidate has no assignedInterval recentRefPosition"); + // Otherwise, we need to mark spillRefPosition as lastUse, or the interval + // will remain active beyond its allocated range during the resolution phase. + spillRefPosition->lastUse = true; } -#endif } - found = selector.applySelection(PREV_REG_OPT, prevRegOptSet DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_PREV_REG_OPT, refPosition->bbNum)); + else +#endif // DEBUG + { + spillInterval(assignedInterval, spillRefPosition DEBUGARG(nextRefPosition)); + } } - - // Apply the REG_NUM heuristic. - if (!found) + // Maintain the association with the interval, if it has more references. + // Or, if we "remembered" an interval assigned to this register, restore it. + if (nextRefPosition != nullptr) { - found = - selector.applySingleRegSelection(REG_NUM, genFindLowestBit(selector.candidates) DEBUG_ARG(registerScore)); - INTRACK_STATS_IF(found, updateLsraStat(STAT_REG_NUM, refPosition->bbNum)); + assignedInterval->assignedReg = regRec; } - - assert(found && isSingleRegister(selector.candidates)); - regMaskTP foundRegBit = selector.candidates; - foundReg = genRegNumFromMask(foundRegBit); - availablePhysRegRecord = getRegisterRecord(foundReg); - Interval* assignedInterval = availablePhysRegRecord->assignedInterval; - if ((assignedInterval != currentInterval) && isAssigned(availablePhysRegRecord ARM_ARG(regType))) + else if (canRestorePreviousInterval(regRec, assignedInterval)) { - if ((foundRegBit & freeCandidates) == RBM_NONE) + regRec->assignedInterval = regRec->previousInterval; + regRec->previousInterval = nullptr; + if (regRec->assignedInterval->physReg != thisRegNum) { - // We're spilling. - CLANG_FORMAT_COMMENT_ANCHOR; + clearNextIntervalRef(thisRegNum, regRec->assignedInterval->registerType); + } + else + { + updateNextIntervalRef(thisRegNum, regRec->assignedInterval); + } #ifdef TARGET_ARM - if (currentInterval->registerType == TYP_DOUBLE) - { - assert(genIsValidDoubleReg(availablePhysRegRecord->regNum)); - unassignDoublePhysReg(availablePhysRegRecord); - } - else if (assignedInterval->registerType == TYP_DOUBLE) - { - // Make sure we spill both halves of the double register. - assert(genIsValidDoubleReg(assignedInterval->assignedReg->regNum)); - unassignPhysReg(assignedInterval->assignedReg, assignedInterval->recentRefPosition); - } - else -#endif - { - unassignPhysReg(availablePhysRegRecord, assignedInterval->recentRefPosition); - } + // Note: + // We can not use updateAssignedInterval() and updatePreviousInterval() here, + // because regRec may not be a even-numbered float register. + + // Update second half RegRecord of a double register for TYP_DOUBLE + if (regRec->assignedInterval->registerType == TYP_DOUBLE) + { + RegRecord* anotherHalfRegRec = findAnotherHalfRegRec(regRec); + + anotherHalfRegRec->assignedInterval = regRec->assignedInterval; + anotherHalfRegRec->previousInterval = nullptr; + } +#endif // TARGET_ARM + +#ifdef DEBUG + if (spill) + { + dumpLsraAllocationEvent(LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL_AFTER_SPILL, regRec->assignedInterval, + thisRegNum); } else { - // If we considered this "unassigned" because this interval's lifetime ends before - // the next ref, remember it. - // For historical reasons (due to former short-circuiting of this case), if we're reassigning - // the current interval to a previous assignment, we don't remember the previous interval. - // Note that we need to compute this condition before calling unassignPhysReg, which wil reset - // assignedInterval->physReg. - bool wasThisAssigned = ((prevRegBit & preferences) == foundRegBit); - bool wasAssigned = (((foundRegBit & unassignedSet) != RBM_NONE) && !wasThisAssigned && - (assignedInterval != nullptr) && (assignedInterval->physReg == foundReg)); - unassignPhysReg(availablePhysRegRecord ARM_ARG(currentInterval->registerType)); - if ((matchingConstants & foundRegBit) != RBM_NONE) - { - assert(assignedInterval->isConstant); - refPosition->treeNode->SetReuseRegVal(); - } - else if (wasAssigned) - { - updatePreviousInterval(availablePhysRegRecord, assignedInterval, assignedInterval->registerType); - } - else - { - assert((selector.score & CONST_AVAILABLE) == 0); - } + dumpLsraAllocationEvent(LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL, regRec->assignedInterval, thisRegNum); } +#endif // DEBUG + } + else + { + updateAssignedInterval(regRec, nullptr, assignedInterval->registerType); + updatePreviousInterval(regRec, nullptr, assignedInterval->registerType); } - assignPhysReg(availablePhysRegRecord, currentInterval); - refPosition->registerAssignment = foundRegBit; - return foundReg; } //------------------------------------------------------------------------ -// canSpillReg: Determine whether we can spill physRegRecord +// spillGCRefs: Spill any GC-type intervals that are currently in registers. // // Arguments: -// physRegRecord - reg to spill -// refLocation - Location of RefPosition where this register will be spilled +// killRefPosition - The RefPosition for the kill // // Return Value: -// True - if we can spill physRegRecord -// False - otherwise +// None. // -bool LinearScan::canSpillReg(RegRecord* physRegRecord, LsraLocation refLocation) +// Notes: +// This is used to ensure that we have no live GC refs in registers at an +// unmanaged call. +// +void LinearScan::spillGCRefs(RefPosition* killRefPosition) { - assert(physRegRecord->assignedInterval != nullptr); - RefPosition* recentAssignedRef = physRegRecord->assignedInterval->recentRefPosition; - - if (recentAssignedRef != nullptr) + // For each physical register that can hold a GC type, + // if it is occupied by an interval of a GC type, spill that interval. + regMaskTP candidateRegs = killRefPosition->registerAssignment; + INDEBUG(bool killedRegs = false); + while (candidateRegs != RBM_NONE) { - // We can't spill a register that's active at the current location. - // We should already have determined this with isRegBusy before calling this method. - assert(!isRefPositionActive(recentAssignedRef, refLocation)); - return true; + regMaskTP nextRegBit = genFindLowestBit(candidateRegs); + candidateRegs &= ~nextRegBit; + regNumber nextReg = genRegNumFromMask(nextRegBit); + RegRecord* regRecord = getRegisterRecord(nextReg); + Interval* assignedInterval = regRecord->assignedInterval; + if (assignedInterval == nullptr || (assignedInterval->isActive == false)) + { + continue; + } + bool needsKill = varTypeIsGC(assignedInterval->registerType); + if (!needsKill) + { + // The importer will assign a GC type to the rhs of an assignment if the lhs type is a GC type, + // even if the rhs is not. See the CEE_STLOC* case in impImportBlockCode(). As a result, + // we can have a 'GT_LCL_VAR' node with a GC type, when the lclVar itself is an integer type. + // The emitter will mark this register as holding a GC type. Therfore we must spill this value. + // This was exposed on Arm32 with EH write-thru. + if ((assignedInterval->recentRefPosition != nullptr) && + (assignedInterval->recentRefPosition->treeNode != nullptr)) + { + needsKill = varTypeIsGC(assignedInterval->recentRefPosition->treeNode); + } + } + if (needsKill) + { + INDEBUG(killedRegs = true); + unassignPhysReg(regRecord, assignedInterval->recentRefPosition); + makeRegAvailable(nextReg, assignedInterval->registerType); + } } - // recentAssignedRef can only be null if this is a parameter that has not yet been - // moved to a register (or stack), in which case we can't spill it yet. - assert(physRegRecord->assignedInterval->getLocalVar(compiler)->lvIsParam); - return false; + INDEBUG(dumpLsraAllocationEvent(killedRegs ? LSRA_EVENT_DONE_KILL_GC_REFS : LSRA_EVENT_NO_GC_KILLS, nullptr, REG_NA, + nullptr)); } //------------------------------------------------------------------------ -// getSpillWeight: Get the weight associated with spilling the given register +// processBlockEndAllocation: Update var locations after 'currentBlock' has been allocated // // Arguments: -// physRegRecord - reg to spill +// currentBlock - the BasicBlock we have just finished allocating registers for // // Return Value: -// The weight associated with the location at which we will spill. -// -// Note: This helper is designed to be used only from allocateReg() and getDoubleSpillWeight() +// None // -float LinearScan::getSpillWeight(RegRecord* physRegRecord) +// Notes: +// Calls processBlockEndLocations() to set the outVarToRegMap, then gets the next block, +// and sets the inVarToRegMap appropriately. + +void LinearScan::processBlockEndAllocation(BasicBlock* currentBlock) { - assert(physRegRecord->assignedInterval != nullptr); - RefPosition* recentAssignedRef = physRegRecord->assignedInterval->recentRefPosition; - float weight = BB_ZERO_WEIGHT; + assert(currentBlock != nullptr); + if (enregisterLocalVars) + { + processBlockEndLocations(currentBlock); + } + markBlockVisited(currentBlock); - // We shouldn't call this method if there is no recentAssignedRef. - assert(recentAssignedRef != nullptr); - // We shouldn't call this method if the register is active at this location. - assert(!isRefPositionActive(recentAssignedRef, currentLoc)); - weight = getWeight(recentAssignedRef); - return weight; + // Get the next block to allocate. + // When the last block in the method has successors, there will be a final "RefTypeBB" to + // ensure that we get the varToRegMap set appropriately, but in that case we don't need + // to worry about "nextBlock". + BasicBlock* nextBlock = getNextBlock(); + if (nextBlock != nullptr) + { + processBlockStartLocations(nextBlock); + } } -#ifdef TARGET_ARM //------------------------------------------------------------------------ -// canSpillDoubleReg: Determine whether we can spill physRegRecord +// rotateBlockStartLocation: When in the LSRA_BLOCK_BOUNDARY_ROTATE stress mode, attempt to +// "rotate" the register assignment for a localVar to the next higher +// register that is available. // // Arguments: -// physRegRecord - reg to spill (must be a valid double register) -// refLocation - Location of RefPosition where this register will be spilled +// interval - the Interval for the variable whose register is getting rotated +// targetReg - its register assignment from the predecessor block being used for live-in +// availableRegs - registers available for use // // Return Value: -// True - if we can spill physRegRecord -// False - otherwise -// -bool LinearScan::canSpillDoubleReg(RegRecord* physRegRecord, LsraLocation refLocation) -{ - assert(genIsValidDoubleReg(physRegRecord->regNum)); - RegRecord* physRegRecord2 = getSecondHalfRegRec(physRegRecord); +// The new register to use. - if ((physRegRecord->assignedInterval != nullptr) && !canSpillReg(physRegRecord, refLocation)) - { - return false; - } - if ((physRegRecord2->assignedInterval != nullptr) && !canSpillReg(physRegRecord2, refLocation)) +#ifdef DEBUG +regNumber LinearScan::rotateBlockStartLocation(Interval* interval, regNumber targetReg, regMaskTP availableRegs) +{ + if (targetReg != REG_STK && getLsraBlockBoundaryLocations() == LSRA_BLOCK_BOUNDARY_ROTATE) { - return false; + // If we're rotating the register locations at block boundaries, try to use + // the next higher register number of the appropriate register type. + regMaskTP candidateRegs = allRegs(interval->registerType) & availableRegs; + regNumber firstReg = REG_NA; + regNumber newReg = REG_NA; + while (candidateRegs != RBM_NONE) + { + regMaskTP nextRegBit = genFindLowestBit(candidateRegs); + candidateRegs &= ~nextRegBit; + regNumber nextReg = genRegNumFromMask(nextRegBit); + if (nextReg > targetReg) + { + newReg = nextReg; + break; + } + else if (firstReg == REG_NA) + { + firstReg = nextReg; + } + } + if (newReg == REG_NA) + { + assert(firstReg != REG_NA); + newReg = firstReg; + } + targetReg = newReg; } - return true; + return targetReg; } +#endif // DEBUG -//------------------------------------------------------------------------ -// unassignDoublePhysReg: unassign a double register (pair) +#ifdef TARGET_ARM +//-------------------------------------------------------------------------------------- +// isSecondHalfReg: Test if recRec is second half of double register +// which is assigned to an interval. // // Arguments: -// doubleRegRecord - reg to unassign +// regRec - a register to be tested +// interval - an interval which is assigned to some register // -// Note: -// The given RegRecord must be a valid (even numbered) double register. +// Assumptions: +// None // -void LinearScan::unassignDoublePhysReg(RegRecord* doubleRegRecord) +// Return Value: +// True only if regRec is second half of assignedReg in interval +// +bool LinearScan::isSecondHalfReg(RegRecord* regRec, Interval* interval) { - assert(genIsValidDoubleReg(doubleRegRecord->regNum)); + RegRecord* assignedReg = interval->assignedReg; - RegRecord* doubleRegRecordLo = doubleRegRecord; - RegRecord* doubleRegRecordHi = getSecondHalfRegRec(doubleRegRecordLo); - // For a double register, we has following four cases. - // Case 1: doubleRegRecLo is assigned to TYP_DOUBLE interval - // Case 2: doubleRegRecLo and doubleRegRecHi are assigned to different TYP_FLOAT intervals - // Case 3: doubelRegRecLo is assgined to TYP_FLOAT interval and doubleRegRecHi is nullptr - // Case 4: doubleRegRecordLo is nullptr, and doubleRegRecordHi is assigned to a TYP_FLOAT interval - if (doubleRegRecordLo->assignedInterval != nullptr) + if (assignedReg != nullptr && interval->registerType == TYP_DOUBLE) { - if (doubleRegRecordLo->assignedInterval->registerType == TYP_DOUBLE) - { - // Case 1: doubleRegRecLo is assigned to TYP_DOUBLE interval - unassignPhysReg(doubleRegRecordLo, doubleRegRecordLo->assignedInterval->recentRefPosition); - } - else - { - // Case 2: doubleRegRecLo and doubleRegRecHi are assigned to different TYP_FLOAT intervals - // Case 3: doubelRegRecLo is assgined to TYP_FLOAT interval and doubleRegRecHi is nullptr - assert(doubleRegRecordLo->assignedInterval->registerType == TYP_FLOAT); - unassignPhysReg(doubleRegRecordLo, doubleRegRecordLo->assignedInterval->recentRefPosition); + // interval should have been allocated to a valid double register + assert(genIsValidDoubleReg(assignedReg->regNum)); - if (doubleRegRecordHi != nullptr) - { - if (doubleRegRecordHi->assignedInterval != nullptr) - { - assert(doubleRegRecordHi->assignedInterval->registerType == TYP_FLOAT); - unassignPhysReg(doubleRegRecordHi, doubleRegRecordHi->assignedInterval->recentRefPosition); - } - } - } - } - else - { - // Case 4: doubleRegRecordLo is nullptr, and doubleRegRecordHi is assigned to a TYP_FLOAT interval - assert(doubleRegRecordHi->assignedInterval != nullptr); - assert(doubleRegRecordHi->assignedInterval->registerType == TYP_FLOAT); - unassignPhysReg(doubleRegRecordHi, doubleRegRecordHi->assignedInterval->recentRefPosition); + // Find a second half RegRecord of double register + regNumber firstRegNum = assignedReg->regNum; + regNumber secondRegNum = REG_NEXT(firstRegNum); + + assert(genIsValidFloatReg(secondRegNum) && !genIsValidDoubleReg(secondRegNum)); + + RegRecord* secondRegRec = getRegisterRecord(secondRegNum); + + return secondRegRec == regRec; } + + return false; } -#endif // TARGET_ARM +//------------------------------------------------------------------------------------------ +// getSecondHalfRegRec: Get the second (odd) half of an ARM32 double register +// +// Arguments: +// regRec - A float RegRecord +// +// Assumptions: +// regRec must be a valid double register (i.e. even) +// +// Return Value: +// The RegRecord for the second half of the double register +// +RegRecord* LinearScan::getSecondHalfRegRec(RegRecord* regRec) +{ + regNumber secondHalfRegNum; + RegRecord* secondHalfRegRec; -//------------------------------------------------------------------------ -// isRefPositionActive: Determine whether a given RefPosition is active at the given location + assert(genIsValidDoubleReg(regRec->regNum)); + + secondHalfRegNum = REG_NEXT(regRec->regNum); + secondHalfRegRec = getRegisterRecord(secondHalfRegNum); + + return secondHalfRegRec; +} +//------------------------------------------------------------------------------------------ +// findAnotherHalfRegRec: Find another half RegRecord which forms same ARM32 double register // // Arguments: -// refPosition - the RefPosition of interest -// refLocation - the LsraLocation at which we want to know if it is active +// regRec - A float RegRecord +// +// Assumptions: +// None // // Return Value: -// True - if this RefPosition occurs at the given location, OR -// if it occurs at the previous location and is marked delayRegFree. -// False - otherwise +// A RegRecord which forms same double register with regRec // -bool LinearScan::isRefPositionActive(RefPosition* refPosition, LsraLocation refLocation) +RegRecord* LinearScan::findAnotherHalfRegRec(RegRecord* regRec) { - return (refPosition->nodeLocation == refLocation || - ((refPosition->nodeLocation + 1 == refLocation) && refPosition->delayRegFree)); + regNumber anotherHalfRegNum = findAnotherHalfRegNum(regRec->regNum); + return getRegisterRecord(anotherHalfRegNum); } - -//------------------------------------------------------------------------ -// isSpillCandidate: Determine if a register is a spill candidate for a given RefPosition. +//------------------------------------------------------------------------------------------ +// findAnotherHalfRegNum: Find another half register's number which forms same ARM32 double register // // Arguments: -// current The interval for the current allocation -// refPosition The RefPosition of the current Interval for which a register is being allocated -// physRegRecord The RegRecord for the register we're considering for spill +// regNumber - A float regNumber +// +// Assumptions: +// None // // Return Value: -// True iff the given register can be spilled to accommodate the given RefPosition. +// A register number which forms same double register with regNum. // -bool LinearScan::isSpillCandidate(Interval* current, RefPosition* refPosition, RegRecord* physRegRecord) +regNumber LinearScan::findAnotherHalfRegNum(regNumber regNum) { - regMaskTP candidateBit = genRegMask(physRegRecord->regNum); - LsraLocation refLocation = refPosition->nodeLocation; - // We shouldn't be calling this if we haven't already determined that the register is not - // busy until the next kill. - assert(!isRegBusy(physRegRecord->regNum, current->registerType)); - // We should already have determined that the register isn't actively in use. - assert(!isRegInUse(physRegRecord->regNum, current->registerType)); - // We shouldn't be calling this if 'refPosition' is a fixed reference to this register. - assert(!refPosition->isFixedRefOfRegMask(candidateBit)); - // We shouldn't be calling this if there is a fixed reference at the same location - // (and it's not due to this reference), as checked above. - assert(!conflictingFixedRegReference(physRegRecord->regNum, refPosition)); + regNumber anotherHalfRegNum; - bool canSpill; -#ifdef TARGET_ARM - if (current->registerType == TYP_DOUBLE) + assert(genIsValidFloatReg(regNum)); + + // Find another half register for TYP_DOUBLE interval, + // following same logic in canRestorePreviousInterval(). + if (genIsValidDoubleReg(regNum)) { - canSpill = canSpillDoubleReg(physRegRecord, refLocation); + anotherHalfRegNum = REG_NEXT(regNum); + assert(!genIsValidDoubleReg(anotherHalfRegNum)); } else -#endif // TARGET_ARM - { - canSpill = canSpillReg(physRegRecord, refLocation); - } - if (!canSpill) { - return false; + anotherHalfRegNum = REG_PREV(regNum); + assert(genIsValidDoubleReg(anotherHalfRegNum)); } - return true; -} - -// Grab a register to use to copy and then immediately use. -// This is called only for localVar intervals that already have a register -// assignment that is not compatible with the current RefPosition. -// This is not like regular assignment, because we don't want to change -// any preferences or existing register assignments. -// Prefer a free register that's got the earliest next use. -// Otherwise, spill something with the farthest next use -// -regNumber LinearScan::assignCopyReg(RefPosition* refPosition) -{ - Interval* currentInterval = refPosition->getInterval(); - assert(currentInterval != nullptr); - assert(currentInterval->isActive); - - // Save the relatedInterval, if any, so that it doesn't get modified during allocation. - Interval* savedRelatedInterval = currentInterval->relatedInterval; - currentInterval->relatedInterval = nullptr; - - // We don't want really want to change the default assignment, - // so 1) pretend this isn't active, and 2) remember the old reg - regNumber oldPhysReg = currentInterval->physReg; - RegRecord* oldRegRecord = currentInterval->assignedReg; - assert(oldRegRecord->regNum == oldPhysReg); - currentInterval->isActive = false; - - // We *must* allocate a register, and it will be a copyReg. Set that field now, so that - // refPosition->RegOptional() will return false. - refPosition->copyReg = true; - - RegisterScore registerScore = NONE; - regNumber allocatedReg = allocateReg(currentInterval, refPosition DEBUG_ARG(®isterScore)); - assert(allocatedReg != REG_NA); - - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_COPY_REG, currentInterval, allocatedReg, nullptr, registerScore)); - - // Now restore the old info - currentInterval->relatedInterval = savedRelatedInterval; - currentInterval->physReg = oldPhysReg; - currentInterval->assignedReg = oldRegRecord; - currentInterval->isActive = true; - return allocatedReg; + return anotherHalfRegNum; } +#endif -//------------------------------------------------------------------------ -// isAssigned: This is the function to check if the given RegRecord has an assignedInterval. +//-------------------------------------------------------------------------------------- +// canRestorePreviousInterval: Test if we can restore previous interval // // Arguments: -// regRec - The RegRecord to check that it is assigned. -// newRegType - There are elements to judge according to the upcoming register type. +// regRec - a register which contains previous interval to be restored +// assignedInterval - an interval just unassigned +// +// Assumptions: +// None // // Return Value: -// Returns true if the given RegRecord has an assignedInterval. +// True only if previous interval of regRec can be restored // -bool LinearScan::isAssigned(RegRecord* regRec ARM_ARG(RegisterType newRegType)) +bool LinearScan::canRestorePreviousInterval(RegRecord* regRec, Interval* assignedInterval) { - if (regRec->assignedInterval != nullptr) + bool retVal = + (regRec->previousInterval != nullptr && regRec->previousInterval != assignedInterval && + regRec->previousInterval->assignedReg == regRec && regRec->previousInterval->getNextRefPosition() != nullptr); + +#ifdef TARGET_ARM + if (retVal && regRec->previousInterval->registerType == TYP_DOUBLE) { - return true; + RegRecord* anotherHalfRegRec = findAnotherHalfRegRec(regRec); + + retVal = retVal && anotherHalfRegRec->assignedInterval == nullptr; } +#endif + + return retVal; +} + +bool LinearScan::isAssignedToInterval(Interval* interval, RegRecord* regRec) +{ + bool isAssigned = (interval->assignedReg == regRec); #ifdef TARGET_ARM - if (newRegType == TYP_DOUBLE) + isAssigned |= isSecondHalfReg(regRec, interval); +#endif + return isAssigned; +} + +void LinearScan::unassignIntervalBlockStart(RegRecord* regRecord, VarToRegMap inVarToRegMap) +{ + // Is there another interval currently assigned to this register? If so unassign it. + Interval* assignedInterval = regRecord->assignedInterval; + if (assignedInterval != nullptr) { - RegRecord* otherRegRecord = getSecondHalfRegRec(regRec); + if (isAssignedToInterval(assignedInterval, regRecord)) + { + // Only localVars, constants or vector upper halves should be assigned to registers at block boundaries. + if (!assignedInterval->isLocalVar) + { + assert(assignedInterval->isConstant || assignedInterval->IsUpperVector()); + // Don't need to update the VarToRegMap. + inVarToRegMap = nullptr; + } - if (otherRegRecord->assignedInterval != nullptr) + regNumber assignedRegNum = assignedInterval->assignedReg->regNum; + + // If the interval is active, it will be set to active when we reach its new + // register assignment (which we must not yet have done, or it wouldn't still be + // assigned to this register). + assignedInterval->isActive = false; + unassignPhysReg(assignedInterval->assignedReg, nullptr); + if ((inVarToRegMap != nullptr) && inVarToRegMap[assignedInterval->getVarIndex(compiler)] == assignedRegNum) + { + inVarToRegMap[assignedInterval->getVarIndex(compiler)] = REG_STK; + } + } + else { - return true; + // This interval is no longer assigned to this register. + updateAssignedInterval(regRecord, nullptr, assignedInterval->registerType); } } -#endif - return false; } //------------------------------------------------------------------------ -// checkAndAssignInterval: Check if the interval is already assigned and -// if it is then unassign the physical record -// and set the assignedInterval to 'interval' +// processBlockStartLocations: Update var locations on entry to 'currentBlock' and clear constant +// registers. // // Arguments: -// regRec - The RegRecord of interest -// interval - The Interval that we're going to assign to 'regRec' +// currentBlock - the BasicBlock we are about to allocate registers for +// allocationPass - true if we are currently allocating registers (versus writing them back) // -void LinearScan::checkAndAssignInterval(RegRecord* regRec, Interval* interval) -{ - Interval* assignedInterval = regRec->assignedInterval; - if (assignedInterval != nullptr && assignedInterval != interval) - { - // This is allocated to another interval. Either it is inactive, or it was allocated as a - // copyReg and is therefore not the "assignedReg" of the other interval. In the latter case, - // we simply unassign it - in the former case we need to set the physReg on the interval to - // REG_NA to indicate that it is no longer in that register. - // The lack of checking for this case resulted in an assert in the retail version of System.dll, - // in method SerialStream.GetDcbFlag. - // Note that we can't check for the copyReg case, because we may have seen a more recent - // RefPosition for the Interval that was NOT a copyReg. - if (assignedInterval->assignedReg == regRec) - { - assert(assignedInterval->isActive == false); - assignedInterval->physReg = REG_NA; - } - unassignPhysReg(regRec->regNum); - } -#ifdef TARGET_ARM - // If 'interval' and 'assignedInterval' were both TYP_DOUBLE, then we have unassigned 'assignedInterval' - // from both halves. Otherwise, if 'interval' is TYP_DOUBLE, we now need to unassign the other half. - if ((interval->registerType == TYP_DOUBLE) && - ((assignedInterval == nullptr) || (assignedInterval->registerType == TYP_FLOAT))) +// Return Value: +// None +// +// Notes: +// During the allocation pass, we use the outVarToRegMap of the selected predecessor to +// determine the lclVar locations for the inVarToRegMap. +// During the resolution (write-back) pass, we only modify the inVarToRegMap in cases where +// a lclVar was spilled after the block had been completed. +void LinearScan::processBlockStartLocations(BasicBlock* currentBlock) +{ + // If we have no register candidates we should only call this method during allocation. + + assert(enregisterLocalVars || !allocationPassComplete); + + if (!enregisterLocalVars) { - RegRecord* otherRegRecord = getSecondHalfRegRec(regRec); - assignedInterval = otherRegRecord->assignedInterval; - if (assignedInterval != nullptr && assignedInterval != interval) + // Just clear any constant registers and return. + resetAvailableRegs(); + for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) { - if (assignedInterval->assignedReg == otherRegRecord) + RegRecord* physRegRecord = getRegisterRecord(reg); + Interval* assignedInterval = physRegRecord->assignedInterval; + clearNextIntervalRef(reg, physRegRecord->registerType); + clearSpillCost(reg, physRegRecord->registerType); + if (assignedInterval != nullptr) { - assert(assignedInterval->isActive == false); - assignedInterval->physReg = REG_NA; + assert(assignedInterval->isConstant); + physRegRecord->assignedInterval = nullptr; } - unassignPhysReg(otherRegRecord->regNum); } + return; } -#endif - - updateAssignedInterval(regRec, interval, interval->registerType); -} - -// Assign the given physical register interval to the given interval -void LinearScan::assignPhysReg(RegRecord* regRec, Interval* interval) -{ - regMaskTP assignedRegMask = genRegMask(regRec->regNum); - compiler->codeGen->regSet.rsSetRegsModified(assignedRegMask DEBUGARG(true)); - - interval->assignedReg = regRec; - checkAndAssignInterval(regRec, interval); - interval->physReg = regRec->regNum; - interval->isActive = true; - if (interval->isLocalVar) - { - // Prefer this register for future references - interval->updateRegisterPreferences(assignedRegMask); - } -} + unsigned predBBNum = blockInfo[currentBlock->bbNum].predBBNum; + VarToRegMap predVarToRegMap = getOutVarToRegMap(predBBNum); + VarToRegMap inVarToRegMap = getInVarToRegMap(currentBlock->bbNum); -//------------------------------------------------------------------------ -// setIntervalAsSplit: Set this Interval as being split -// -// Arguments: -// interval - The Interval which is being split -// -// Return Value: -// None. -// -// Notes: -// The given Interval will be marked as split, and it will be added to the -// set of splitOrSpilledVars. -// -// Assumptions: -// "interval" must be a lclVar interval, as tree temps are never split. -// This is asserted in the call to getVarIndex(). -// -void LinearScan::setIntervalAsSplit(Interval* interval) -{ - if (interval->isLocalVar) + // If this block enters an exception region, all incoming vars are on the stack. + if (predBBNum == 0) { - unsigned varIndex = interval->getVarIndex(compiler); - if (!interval->isSplit) - { - VarSetOps::AddElemD(compiler, splitOrSpilledVars, varIndex); - } - else +#if DEBUG + if (blockInfo[currentBlock->bbNum].hasEHBoundaryIn || !allocationPassComplete) { - assert(VarSetOps::IsMember(compiler, splitOrSpilledVars, varIndex)); + // This should still be in its initialized empty state. + for (unsigned varIndex = 0; varIndex < compiler->lvaTrackedCount; varIndex++) + { + // In the case where we're extending lifetimes for stress, we are intentionally modeling variables + // as live when they really aren't to create extra register pressure & constraints. + // However, this means that non-EH-vars will be live into EH regions. We can and should ignore the + // locations of these. Note that they aren't reported to codegen anyway. + if (!getLsraExtendLifeTimes() || VarSetOps::IsMember(compiler, currentBlock->bbLiveIn, varIndex)) + { + assert(inVarToRegMap[varIndex] == REG_STK); + } + } } +#endif // DEBUG + predVarToRegMap = inVarToRegMap; } - interval->isSplit = true; -} -//------------------------------------------------------------------------ -// setIntervalAsSpilled: Set this Interval as being spilled -// -// Arguments: -// interval - The Interval which is being spilled -// -// Return Value: -// None. -// -// Notes: -// The given Interval will be marked as spilled, and it will be added -// to the set of splitOrSpilledVars. -// -void LinearScan::setIntervalAsSpilled(Interval* interval) -{ -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - if (interval->isUpperVector) + VarSetOps::AssignNoCopy(compiler, currentLiveVars, + VarSetOps::Intersection(compiler, registerCandidateVars, currentBlock->bbLiveIn)); +#ifdef DEBUG + if (getLsraExtendLifeTimes()) { - assert(interval->relatedInterval->isLocalVar); - interval->isSpilled = true; - // Now we need to mark the local as spilled also, even if the lower half is never spilled, - // as this will use the upper part of its home location. - interval = interval->relatedInterval; - // We'll now mark this as spilled, so it changes the spillCost. - RefPosition* recentRefPos = interval->recentRefPosition; - if (!interval->isSpilled && interval->isActive && (recentRefPos != nullptr)) - { - VarSetOps::AddElemD(compiler, splitOrSpilledVars, interval->getVarIndex(compiler)); - interval->isSpilled = true; - regNumber reg = interval->physReg; - spillCost[reg] = getSpillWeight(getRegisterRecord(reg)); - } + VarSetOps::AssignNoCopy(compiler, currentLiveVars, registerCandidateVars); } -#endif - if (interval->isLocalVar) + // If we are rotating register assignments at block boundaries, we want to make the + // inactive registers available for the rotation. + regMaskTP inactiveRegs = RBM_NONE; +#endif // DEBUG + regMaskTP liveRegs = RBM_NONE; + VarSetOps::Iter iter(compiler, currentLiveVars); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex)) { - unsigned varIndex = interval->getVarIndex(compiler); - if (!interval->isSpilled) - { - VarSetOps::AddElemD(compiler, splitOrSpilledVars, varIndex); - } - else + if (!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate) { - assert(VarSetOps::IsMember(compiler, splitOrSpilledVars, varIndex)); + continue; } - } - interval->isSpilled = true; -} + regNumber targetReg; + Interval* interval = getIntervalForLocalVar(varIndex); + RefPosition* nextRefPosition = interval->getNextRefPosition(); + assert((nextRefPosition != nullptr) || (interval->isWriteThru)); -//------------------------------------------------------------------------ -// spill: Spill the "interval" starting from "fromRefPosition" (upto "toRefPosition") -// -// Arguments: -// interval - The interval that contains the RefPosition to be spilled -// fromRefPosition - The RefPosition at which the Interval is to be spilled -// toRefPosition - The RefPosition at which it must be reloaded (debug only arg) -// -// Return Value: -// None. -// -// Assumptions: -// fromRefPosition and toRefPosition must not be null -// -void LinearScan::spillInterval(Interval* interval, RefPosition* fromRefPosition DEBUGARG(RefPosition* toRefPosition)) -{ - assert(fromRefPosition != nullptr && toRefPosition != nullptr); - assert(fromRefPosition->getInterval() == interval && toRefPosition->getInterval() == interval); - assert(fromRefPosition->nextRefPosition == toRefPosition); + bool leaveOnStack = false; - if (!fromRefPosition->lastUse) - { - // If not allocated a register, Lcl var def/use ref positions even if reg optional - // should be marked as spillAfter. Note that if it is a WriteThru interval, the value is always - // written to the stack, but the WriteThru indicates that the register is no longer live. - if (fromRefPosition->RegOptional() && !(interval->isLocalVar && fromRefPosition->IsActualRef())) - { - fromRefPosition->registerAssignment = RBM_NONE; - } - else + // Special handling for variables live in/out of exception handlers. + if (interval->isWriteThru) { - fromRefPosition->spillAfter = true; + // There are 3 cases where we will leave writethru lclVars on the stack: + // 1) There is no predecessor. + // 2) It is conservatively or artificially live - that is, it has no next use, + // so there is no place for codegen to record that the register is no longer occupied. + // 3) This block has a predecessor with an outgoing EH edge. We won't be able to add "join" + // resolution to load the EH var into a register along that edge, so it must be on stack. + if ((predBBNum == 0) || (nextRefPosition == nullptr) || (RefTypeIsDef(nextRefPosition->refType)) || + blockInfo[currentBlock->bbNum].hasEHPred) + { + leaveOnStack = true; + } } - } - assert(toRefPosition != nullptr); + if (!allocationPassComplete) + { + targetReg = getVarReg(predVarToRegMap, varIndex); + if (leaveOnStack) + { + targetReg = REG_STK; + } #ifdef DEBUG - if (VERBOSE) - { - dumpLsraAllocationEvent(LSRA_EVENT_SPILL, interval); - } + regNumber newTargetReg = rotateBlockStartLocation(interval, targetReg, (~liveRegs | inactiveRegs)); + if (newTargetReg != targetReg) + { + targetReg = newTargetReg; + setIntervalAsSplit(interval); + } #endif // DEBUG + setVarReg(inVarToRegMap, varIndex, targetReg); + } + else // allocationPassComplete (i.e. resolution/write-back pass) + { + targetReg = getVarReg(inVarToRegMap, varIndex); + // There are four cases that we need to consider during the resolution pass: + // 1. This variable had a register allocated initially, and it was not spilled in the RefPosition + // that feeds this block. In this case, both targetReg and predVarToRegMap[varIndex] will be targetReg. + // 2. This variable had not been spilled prior to the end of predBB, but was later spilled, so + // predVarToRegMap[varIndex] will be REG_STK, but targetReg is its former allocated value. + // In this case, we will normally change it to REG_STK. We will update its "spilled" status when we + // encounter it in resolveLocalRef(). + // 2a. If the next RefPosition is marked as a copyReg, we need to retain the allocated register. This is + // because the copyReg RefPosition will not have recorded the "home" register, yet downstream + // RefPositions rely on the correct "home" register. + // 3. This variable was spilled before we reached the end of predBB. In this case, both targetReg and + // predVarToRegMap[varIndex] will be REG_STK, and the next RefPosition will have been marked + // as reload during allocation time if necessary (note that by the time we actually reach the next + // RefPosition, we may be using a different predecessor, at which it is still in a register). + // 4. This variable was spilled during the allocation of this block, so targetReg is REG_STK + // (because we set inVarToRegMap at the time we spilled it), but predVarToRegMap[varIndex] + // is not REG_STK. We retain the REG_STK value in the inVarToRegMap. + if (targetReg != REG_STK) + { + if (getVarReg(predVarToRegMap, varIndex) != REG_STK) + { + // Case #1 above. + assert(getVarReg(predVarToRegMap, varIndex) == targetReg || + getLsraBlockBoundaryLocations() == LSRA_BLOCK_BOUNDARY_ROTATE); + } + else if (!nextRefPosition->copyReg) + { + // case #2 above. + setVarReg(inVarToRegMap, varIndex, REG_STK); + targetReg = REG_STK; + } + // Else case 2a. - retain targetReg. + } + // Else case #3 or #4, we retain targetReg and nothing further to do or assert. + } - INTRACK_STATS(updateLsraStat(STAT_SPILL, fromRefPosition->bbNum)); - - interval->isActive = false; - setIntervalAsSpilled(interval); - - // If fromRefPosition occurs before the beginning of this block, mark this as living in the stack - // on entry to this block. - if (fromRefPosition->nodeLocation <= curBBStartLocation) + if (interval->physReg == targetReg) + { + if (interval->isActive) + { + assert(targetReg != REG_STK); + assert(interval->assignedReg != nullptr && interval->assignedReg->regNum == targetReg && + interval->assignedReg->assignedInterval == interval); + liveRegs |= getRegMask(targetReg, interval->registerType); + continue; + } + } + else if (interval->physReg != REG_NA) + { + // This can happen if we are using the locations from a basic block other than the + // immediately preceding one - where the variable was in a different location. + if ((targetReg != REG_STK) || leaveOnStack) + { + // Unassign it from the register (it may get a new register below). + if (interval->assignedReg != nullptr && interval->assignedReg->assignedInterval == interval) + { + interval->isActive = false; + unassignPhysReg(getRegisterRecord(interval->physReg), nullptr); + } + else + { + // This interval was live in this register the last time we saw a reference to it, + // but has since been displaced. + interval->physReg = REG_NA; + } + } + else if (!allocationPassComplete) + { + // Keep the register assignment - if another var has it, it will get unassigned. + // Otherwise, resolution will fix it up later, and it will be more + // likely to match other assignments this way. + targetReg = interval->physReg; + interval->isActive = true; + liveRegs |= getRegMask(targetReg, interval->registerType); + INDEBUG(inactiveRegs |= genRegMask(targetReg)); + setVarReg(inVarToRegMap, varIndex, targetReg); + } + else + { + interval->physReg = REG_NA; + } + } + if (targetReg != REG_STK) + { + RegRecord* targetRegRecord = getRegisterRecord(targetReg); + liveRegs |= getRegMask(targetReg, interval->registerType); + if (!allocationPassComplete) + { + updateNextIntervalRef(targetReg, interval); + updateSpillCost(targetReg, interval); + } + if (!interval->isActive) + { + interval->isActive = true; + interval->physReg = targetReg; + interval->assignedReg = targetRegRecord; + } + if (targetRegRecord->assignedInterval != interval) + { +#ifdef TARGET_ARM + // If this is a TYP_DOUBLE interval, and the assigned interval is either null or is TYP_FLOAT, + // we also need to unassign the other half of the register. + // Note that if the assigned interval is TYP_DOUBLE, it will be unassigned below. + if ((interval->registerType == TYP_DOUBLE) && + ((targetRegRecord->assignedInterval == nullptr) || + (targetRegRecord->assignedInterval->registerType == TYP_FLOAT))) + { + assert(genIsValidDoubleReg(targetReg)); + unassignIntervalBlockStart(getSecondHalfRegRec(targetRegRecord), + allocationPassComplete ? nullptr : inVarToRegMap); + } + + // If this is a TYP_FLOAT interval, and the assigned interval was TYP_DOUBLE, we also + // need to update the liveRegs to specify that the other half is not live anymore. + // As mentioned above, for TYP_DOUBLE, the other half will be unassigned further below. + if ((interval->registerType == TYP_FLOAT) && + ((targetRegRecord->assignedInterval != nullptr) && + (targetRegRecord->assignedInterval->registerType == TYP_DOUBLE))) + { + RegRecord* anotherHalfRegRec = findAnotherHalfRegRec(targetRegRecord); + + // Use TYP_FLOAT to get the regmask of just the half reg. + liveRegs &= ~getRegMask(anotherHalfRegRec->regNum, TYP_FLOAT); + } + +#endif // TARGET_ARM + unassignIntervalBlockStart(targetRegRecord, allocationPassComplete ? nullptr : inVarToRegMap); + assignPhysReg(targetRegRecord, interval); + } + if (interval->recentRefPosition != nullptr && !interval->recentRefPosition->copyReg && + interval->recentRefPosition->registerAssignment != genRegMask(targetReg)) + { + interval->getNextRefPosition()->outOfOrder = true; + } + } + } + + // Unassign any registers that are no longer live, and set register state, if allocating. + if (!allocationPassComplete) { - // This must be a lclVar interval - assert(interval->isLocalVar); - setInVarRegForBB(curBBNum, interval->varNum, REG_STK); + resetRegState(); + setRegsInUse(liveRegs); + } + for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) + { + RegRecord* physRegRecord = getRegisterRecord(reg); + if ((liveRegs & genRegMask(reg)) == 0) + { + makeRegAvailable(reg, physRegRecord->registerType); + Interval* assignedInterval = physRegRecord->assignedInterval; + + if (assignedInterval != nullptr) + { + assert(assignedInterval->isLocalVar || assignedInterval->isConstant || + assignedInterval->IsUpperVector()); + + if (!assignedInterval->isConstant && assignedInterval->assignedReg == physRegRecord) + { + assignedInterval->isActive = false; + if (assignedInterval->getNextRefPosition() == nullptr) + { + unassignPhysReg(physRegRecord, nullptr); + } + if (!assignedInterval->IsUpperVector()) + { + inVarToRegMap[assignedInterval->getVarIndex(compiler)] = REG_STK; + } + } + else + { + // This interval may still be active, but was in another register in an + // intervening block. + updateAssignedInterval(physRegRecord, nullptr, assignedInterval->registerType); + } + +#ifdef TARGET_ARM + // unassignPhysReg, above, may have restored a 'previousInterval', in which case we need to + // get the value of 'physRegRecord->assignedInterval' rather than using 'assignedInterval'. + if (physRegRecord->assignedInterval != nullptr) + { + assignedInterval = physRegRecord->assignedInterval; + } + if (assignedInterval->registerType == TYP_DOUBLE) + { + // Skip next float register, because we already addressed a double register + assert(genIsValidDoubleReg(reg)); + reg = REG_NEXT(reg); + makeRegAvailable(reg, physRegRecord->registerType); + } +#endif // TARGET_ARM + } + } +#ifdef TARGET_ARM + else + { + Interval* assignedInterval = physRegRecord->assignedInterval; + + if (assignedInterval != nullptr && assignedInterval->registerType == TYP_DOUBLE) + { + // Skip next float register, because we already addressed a double register + assert(genIsValidDoubleReg(reg)); + reg = REG_NEXT(reg); + } + } +#endif // TARGET_ARM } } //------------------------------------------------------------------------ -// unassignPhysRegNoSpill: Unassign the given physical register record from -// an active interval, without spilling. +// processBlockEndLocations: Record the variables occupying registers after completing the current block. // // Arguments: -// regRec - the RegRecord to be unassigned +// currentBlock - the block we have just completed. // // Return Value: -// None. -// -// Assumptions: -// The assignedInterval must not be null, and must be active. +// None // // Notes: -// This method is used to unassign a register when an interval needs to be moved to a -// different register, but not (yet) spilled. +// This must be called both during the allocation and resolution (write-back) phases. +// This is because we need to have the outVarToRegMap locations in order to set the locations +// at successor blocks during allocation time, but if lclVars are spilled after a block has been +// completed, we need to record the REG_STK location for those variables at resolution time. -void LinearScan::unassignPhysRegNoSpill(RegRecord* regRec) +void LinearScan::processBlockEndLocations(BasicBlock* currentBlock) { - Interval* assignedInterval = regRec->assignedInterval; - assert(assignedInterval != nullptr && assignedInterval->isActive); - assignedInterval->isActive = false; - unassignPhysReg(regRec, nullptr); - assignedInterval->isActive = true; -} + assert(currentBlock != nullptr && currentBlock->bbNum == curBBNum); + VarToRegMap outVarToRegMap = getOutVarToRegMap(curBBNum); -//------------------------------------------------------------------------ -// checkAndClearInterval: Clear the assignedInterval for the given -// physical register record -// -// Arguments: -// regRec - the physical RegRecord to be unassigned -// spillRefPosition - The RefPosition at which the assignedInterval is to be spilled -// or nullptr if we aren't spilling -// -// Return Value: -// None. -// -// Assumptions: -// see unassignPhysReg -// -void LinearScan::checkAndClearInterval(RegRecord* regRec, RefPosition* spillRefPosition) -{ - Interval* assignedInterval = regRec->assignedInterval; - assert(assignedInterval != nullptr); - regNumber thisRegNum = regRec->regNum; - - if (spillRefPosition == nullptr) + VarSetOps::AssignNoCopy(compiler, currentLiveVars, + VarSetOps::Intersection(compiler, registerCandidateVars, currentBlock->bbLiveOut)); +#ifdef DEBUG + if (getLsraExtendLifeTimes()) { - // Note that we can't assert for the copyReg case - // - if (assignedInterval->physReg == thisRegNum) + VarSetOps::Assign(compiler, currentLiveVars, registerCandidateVars); + } +#endif // DEBUG + VarSetOps::Iter iter(compiler, currentLiveVars); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex)) + { + Interval* interval = getIntervalForLocalVar(varIndex); + if (interval->isActive) { - assert(assignedInterval->isActive == false); + assert(interval->physReg != REG_NA && interval->physReg != REG_STK); + setVarReg(outVarToRegMap, varIndex, interval->physReg); + } + else + { + outVarToRegMap[varIndex] = REG_STK; } +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + // Ensure that we have no partially-spilled large vector locals. + assert(!Compiler::varTypeNeedsPartialCalleeSave(interval->registerType) || !interval->isPartiallySpilled); +#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE } - else + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_END_BB)); +} + +#ifdef DEBUG +void LinearScan::dumpRefPositions(const char* str) +{ + printf("------------\n"); + printf("REFPOSITIONS %s: \n", str); + printf("------------\n"); + for (RefPosition& refPos : refPositions) { - assert(spillRefPosition->getInterval() == assignedInterval); + refPos.dump(); } - - updateAssignedInterval(regRec, nullptr, assignedInterval->registerType); } +#endif // DEBUG //------------------------------------------------------------------------ -// unassignPhysReg: Unassign the given physical register record, and spill the -// assignedInterval at the given spillRefPosition, if any. +// LinearScan::makeRegisterInactive: Make the interval currently assigned to +// a register inactive. // // Arguments: -// regRec - The RegRecord to be unassigned -// newRegType - The RegisterType of interval that would be assigned +// physRegRecord - the RegRecord for the register // // Return Value: // None. // // Notes: -// On ARM architecture, Intervals have to be unassigned considering -// with the register type of interval that would be assigned. +// It may be that the RegRecord has already been freed, e.g. due to a kill, +// or it may be that the register was a copyReg, so is not the assigned register +// of the Interval currently occupying the register, in which case this method has no effect. // -void LinearScan::unassignPhysReg(RegRecord* regRec ARM_ARG(RegisterType newRegType)) +void LinearScan::makeRegisterInactive(RegRecord* physRegRecord) { - RegRecord* regRecToUnassign = regRec; -#ifdef TARGET_ARM - RegRecord* anotherRegRec = nullptr; - - if ((regRecToUnassign->assignedInterval != nullptr) && - (regRecToUnassign->assignedInterval->registerType == TYP_DOUBLE)) - { - // If the register type of interval(being unassigned or new) is TYP_DOUBLE, - // It should have to be valid double register (even register) - if (!genIsValidDoubleReg(regRecToUnassign->regNum)) - { - regRecToUnassign = findAnotherHalfRegRec(regRec); - } - } - else + Interval* assignedInterval = physRegRecord->assignedInterval; + // It may have already been freed by a "Kill" + if ((assignedInterval != nullptr) && (assignedInterval->physReg == physRegRecord->regNum)) { - if (newRegType == TYP_DOUBLE) + assignedInterval->isActive = false; + if (assignedInterval->isConstant) { - anotherRegRec = getSecondHalfRegRec(regRecToUnassign); + clearNextIntervalRef(physRegRecord->regNum, assignedInterval->registerType); } } -#endif - - if (regRecToUnassign->assignedInterval != nullptr) - { - unassignPhysReg(regRecToUnassign, regRecToUnassign->assignedInterval->recentRefPosition); - } -#ifdef TARGET_ARM - if ((anotherRegRec != nullptr) && (anotherRegRec->assignedInterval != nullptr)) - { - unassignPhysReg(anotherRegRec, anotherRegRec->assignedInterval->recentRefPosition); - } -#endif } //------------------------------------------------------------------------ -// unassignPhysReg: Unassign the given physical register record, and spill the -// assignedInterval at the given spillRefPosition, if any. +// LinearScan::freeRegister: Make a register available for use // // Arguments: -// regRec - the RegRecord to be unassigned -// spillRefPosition - The RefPosition at which the assignedInterval is to be spilled +// physRegRecord - the RegRecord for the register to be freed. // // Return Value: // None. // // Assumptions: -// The assignedInterval must not be null. -// If spillRefPosition is null, the assignedInterval must be inactive, or not currently -// assigned to this register (e.g. this is a copyReg for that Interval). -// Otherwise, spillRefPosition must be associated with the assignedInterval. +// None. +// It may be that the RegRecord has already been freed, e.g. due to a kill, +// in which case this method has no effect. // -void LinearScan::unassignPhysReg(RegRecord* regRec, RefPosition* spillRefPosition) +// Notes: +// If there is currently an Interval assigned to this register, and it has +// more references (i.e. this is a local last-use, but more uses and/or +// defs remain), it will remain assigned to the physRegRecord. However, since +// it is marked inactive, the register will be available, albeit less desirable +// to allocate. +// +void LinearScan::freeRegister(RegRecord* physRegRecord) { - Interval* assignedInterval = regRec->assignedInterval; - assert(assignedInterval != nullptr); - assert(spillRefPosition == nullptr || spillRefPosition->getInterval() == assignedInterval); - regNumber thisRegNum = regRec->regNum; - - // Is assignedInterval actually still assigned to this register? - bool intervalIsAssigned = (assignedInterval->physReg == thisRegNum); - regNumber regToUnassign = thisRegNum; - -#ifdef TARGET_ARM - RegRecord* anotherRegRec = nullptr; + Interval* assignedInterval = physRegRecord->assignedInterval; + makeRegAvailable(physRegRecord->regNum, physRegRecord->registerType); + clearSpillCost(physRegRecord->regNum, physRegRecord->registerType); + makeRegisterInactive(physRegRecord); - // Prepare second half RegRecord of a double register for TYP_DOUBLE - if (assignedInterval->registerType == TYP_DOUBLE) + if (assignedInterval != nullptr) { - assert(isFloatRegType(regRec->registerType)); - RegRecord* doubleRegRec; - if (genIsValidDoubleReg(thisRegNum)) - { - anotherRegRec = getSecondHalfRegRec(regRec); - doubleRegRec = regRec; - } - else + // TODO: Under the following conditions we should be just putting it in regsToMakeInactive + // not regsToFree. + // + // We don't unassign in the following conditions: + // - If this is a constant node, that we may encounter again, OR + // - If its recent RefPosition is not a last-use and its next RefPosition is non-null. + // - If there are no more RefPositions, or the next + // one is a def. Note that the latter condition doesn't actually ensure that + // there aren't subsequent uses that could be reached by a value in the assigned + // register, but is merely a heuristic to avoid tying up the register (or using + // it when it's non-optimal). A better alternative would be to use SSA, so that + // we wouldn't unnecessarily link separate live ranges to the same register. + // + RefPosition* nextRefPosition = assignedInterval->getNextRefPosition(); + if (!assignedInterval->isConstant && (nextRefPosition == nullptr || RefTypeIsDef(nextRefPosition->refType))) { - regToUnassign = REG_PREV(thisRegNum); - anotherRegRec = getRegisterRecord(regToUnassign); - doubleRegRec = anotherRegRec; +#ifdef TARGET_ARM + assert((assignedInterval->registerType != TYP_DOUBLE) || genIsValidDoubleReg(physRegRecord->regNum)); +#endif // TARGET_ARM + unassignPhysReg(physRegRecord, nullptr); } + } +} - // Both RegRecords should have been assigned to the same interval. - assert(assignedInterval == anotherRegRec->assignedInterval); - if (!intervalIsAssigned && (assignedInterval->physReg == anotherRegRec->regNum)) +//------------------------------------------------------------------------ +// LinearScan::freeRegisters: Free the registers in 'regsToFree' +// +// Arguments: +// regsToFree - the mask of registers to free +// +void LinearScan::freeRegisters(regMaskTP regsToFree) +{ + if (regsToFree == RBM_NONE) + { + return; + } + + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_FREE_REGS)); + makeRegsAvailable(regsToFree); + while (regsToFree != RBM_NONE) + { + regMaskTP nextRegBit = genFindLowestBit(regsToFree); + regsToFree &= ~nextRegBit; + regNumber nextReg = genRegNumFromMask(nextRegBit); + RegRecord* regRecord = getRegisterRecord(nextReg); +#ifdef TARGET_ARM + if (regRecord->assignedInterval != nullptr && (regRecord->assignedInterval->registerType == TYP_DOUBLE)) { - intervalIsAssigned = true; + assert(genIsValidDoubleReg(nextReg)); + regsToFree &= ~(nextRegBit << 1); } +#endif + freeRegister(regRecord); + } +} - clearNextIntervalRef(regToUnassign, TYP_DOUBLE); - clearSpillCost(regToUnassign, TYP_DOUBLE); - checkAndClearInterval(doubleRegRec, spillRefPosition); +//------------------------------------------------------------------------ +// LinearScan::allocateRegisters: Perform the actual register allocation by iterating over +// all of the previously constructed Intervals +// +void LinearScan::allocateRegisters() +{ + JITDUMP("*************** In LinearScan::allocateRegisters()\n"); + DBEXEC(VERBOSE, lsraDumpIntervals("before allocateRegisters")); - // Both RegRecords should have been unassigned together. - assert(regRec->assignedInterval == nullptr); - assert(anotherRegRec->assignedInterval == nullptr); - } - else -#endif // TARGET_ARM + // at start, nothing is active except for register args + for (Interval& interval : intervals) { - clearNextIntervalRef(thisRegNum, assignedInterval->registerType); - clearSpillCost(thisRegNum, assignedInterval->registerType); - checkAndClearInterval(regRec, spillRefPosition); + Interval* currentInterval = &interval; + currentInterval->recentRefPosition = nullptr; + currentInterval->isActive = false; + if (currentInterval->isLocalVar) + { + LclVarDsc* varDsc = currentInterval->getLocalVar(compiler); + if (varDsc->lvIsRegArg && currentInterval->firstRefPosition != nullptr) + { + currentInterval->isActive = true; + } + } } - makeRegAvailable(regToUnassign, assignedInterval->registerType); - RefPosition* nextRefPosition = nullptr; - if (spillRefPosition != nullptr) + if (enregisterLocalVars) { - nextRefPosition = spillRefPosition->nextRefPosition; +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + VarSetOps::Iter largeVectorVarsIter(compiler, largeVectorVars); + unsigned largeVectorVarIndex = 0; + while (largeVectorVarsIter.NextElem(&largeVectorVarIndex)) + { + Interval* lclVarInterval = getIntervalForLocalVar(largeVectorVarIndex); + lclVarInterval->isPartiallySpilled = false; + } +#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE } - if (!intervalIsAssigned && assignedInterval->physReg != REG_NA) + resetRegState(); + for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) { - // This must have been a temporary copy reg, but we can't assert that because there - // may have been intervening RefPositions that were not copyRegs. - - // reg->assignedInterval has already been set to nullptr by checkAndClearInterval() - assert(regRec->assignedInterval == nullptr); - return; - } - - // regNumber victimAssignedReg = assignedInterval->physReg; - assignedInterval->physReg = REG_NA; - - bool spill = assignedInterval->isActive && nextRefPosition != nullptr; - if (spill) - { - // If this is an active interval, it must have a recentRefPosition, - // otherwise it would not be active - assert(spillRefPosition != nullptr); + RegRecord* physRegRecord = getRegisterRecord(reg); + physRegRecord->recentRefPosition = nullptr; + updateNextFixedRef(physRegRecord, physRegRecord->firstRefPosition); -#if 0 - // TODO-CQ: Enable this and insert an explicit GT_COPY (otherwise there's no way to communicate - // to codegen that we want the copyReg to be the new home location). - // If the last reference was a copyReg, and we're spilling the register - // it was copied from, then make the copyReg the new primary location - // if possible - if (spillRefPosition->copyReg) - { - regNumber copyFromRegNum = victimAssignedReg; - regNumber copyRegNum = genRegNumFromMask(spillRefPosition->registerAssignment); - if (copyFromRegNum == thisRegNum && - getRegisterRecord(copyRegNum)->assignedInterval == assignedInterval) - { - assert(copyRegNum != thisRegNum); - assignedInterval->physReg = copyRegNum; - assignedInterval->assignedReg = this->getRegisterRecord(copyRegNum); - return; - } - } -#endif // 0 -#ifdef DEBUG - // With JitStressRegs == 0x80 (LSRA_EXTEND_LIFETIMES), we may have a RefPosition - // that is not marked lastUse even though the treeNode is a lastUse. In that case - // we must not mark it for spill because the register will have been immediately freed - // after use. While we could conceivably add special handling for this case in codegen, - // it would be messy and undesirably cause the "bleeding" of LSRA stress modes outside - // of LSRA. - if (extendLifetimes() && assignedInterval->isLocalVar && RefTypeIsUse(spillRefPosition->refType) && - spillRefPosition->treeNode != nullptr && - spillRefPosition->treeNode->AsLclVar()->IsLastUse(spillRefPosition->multiRegIdx)) + // Is this an incoming arg register? (Note that we don't, currently, consider reassigning + // an incoming arg register as having spill cost.) + Interval* interval = physRegRecord->assignedInterval; + if (interval != nullptr) { - dumpLsraAllocationEvent(LSRA_EVENT_SPILL_EXTENDED_LIFETIME, assignedInterval); - assignedInterval->isActive = false; - spill = false; - // If the spillRefPosition occurs before the beginning of this block, it will have - // been marked as living in this register on entry to this block, but we now need - // to mark this as living on the stack. - if (spillRefPosition->nodeLocation <= curBBStartLocation) - { - setInVarRegForBB(curBBNum, assignedInterval->varNum, REG_STK); - if (spillRefPosition->nextRefPosition != nullptr) - { - setIntervalAsSpilled(assignedInterval); - } - } - else +#ifdef TARGET_ARM + if ((interval->registerType != TYP_DOUBLE) || genIsValidDoubleReg(reg)) +#endif // TARGET_ARM { - // Otherwise, we need to mark spillRefPosition as lastUse, or the interval - // will remain active beyond its allocated range during the resolution phase. - spillRefPosition->lastUse = true; + updateNextIntervalRef(reg, interval); + updateSpillCost(reg, interval); + setRegInUse(reg, interval->registerType); + INDEBUG(registersToDump |= getRegMask(reg, interval->registerType)); } } else -#endif // DEBUG { - spillInterval(assignedInterval, spillRefPosition DEBUGARG(nextRefPosition)); + clearNextIntervalRef(reg, physRegRecord->registerType); + clearSpillCost(reg, physRegRecord->registerType); } } - // Maintain the association with the interval, if it has more references. - // Or, if we "remembered" an interval assigned to this register, restore it. - if (nextRefPosition != nullptr) + +#ifdef DEBUG + if (VERBOSE) { - assignedInterval->assignedReg = regRec; + dumpRefPositions("BEFORE ALLOCATION"); + dumpVarRefPositions("BEFORE ALLOCATION"); + + printf("\n\nAllocating Registers\n" + "--------------------\n"); + // Start with a small set of commonly used registers, so that we don't keep having to print a new title. + // Include all the arg regs, as they may already have values assigned to them. + registersToDump = LsraLimitSmallIntSet | LsraLimitSmallFPSet | RBM_ARG_REGS; + dumpRegRecordHeader(); + // Now print an empty "RefPosition", since we complete the dump of the regs at the beginning of the loop. + printf(indentFormat, ""); } - else if (canRestorePreviousInterval(regRec, assignedInterval)) +#endif // DEBUG + + BasicBlock* currentBlock = nullptr; + + LsraLocation prevLocation = MinLocation; + regMaskTP regsToFree = RBM_NONE; + regMaskTP delayRegsToFree = RBM_NONE; + regMaskTP regsToMakeInactive = RBM_NONE; + regMaskTP delayRegsToMakeInactive = RBM_NONE; + regMaskTP copyRegsToFree = RBM_NONE; + regsInUseThisLocation = RBM_NONE; + regsInUseNextLocation = RBM_NONE; + + // This is the most recent RefPosition for which a register was allocated + // - currently only used for DEBUG but maintained in non-debug, for clarity of code + // (and will be optimized away because in non-debug spillAlways() unconditionally returns false) + RefPosition* lastAllocatedRefPosition = nullptr; + + bool handledBlockEnd = false; + + for (RefPosition& refPositionIterator : refPositions) { - regRec->assignedInterval = regRec->previousInterval; - regRec->previousInterval = nullptr; - if (regRec->assignedInterval->physReg != thisRegNum) - { - clearNextIntervalRef(thisRegNum, regRec->assignedInterval->registerType); - } - else + RefPosition* currentRefPosition = &refPositionIterator; + RefPosition* nextRefPosition = currentRefPosition->nextRefPosition; + + // TODO: Can we combine this with the freeing of registers below? It might + // mess with the dump, since this was previously being done before the call below + // to dumpRegRecords. + regMaskTP tempRegsToMakeInactive = (regsToMakeInactive | delayRegsToMakeInactive); + while (tempRegsToMakeInactive != RBM_NONE) { - updateNextIntervalRef(thisRegNum, regRec->assignedInterval); + regMaskTP nextRegBit = genFindLowestBit(tempRegsToMakeInactive); + tempRegsToMakeInactive &= ~nextRegBit; + regNumber nextReg = genRegNumFromMask(nextRegBit); + RegRecord* regRecord = getRegisterRecord(nextReg); + clearSpillCost(regRecord->regNum, regRecord->registerType); + makeRegisterInactive(regRecord); } - -#ifdef TARGET_ARM - // Note: - // We can not use updateAssignedInterval() and updatePreviousInterval() here, - // because regRec may not be a even-numbered float register. - - // Update second half RegRecord of a double register for TYP_DOUBLE - if (regRec->assignedInterval->registerType == TYP_DOUBLE) + if (currentRefPosition->nodeLocation > prevLocation) { - RegRecord* anotherHalfRegRec = findAnotherHalfRegRec(regRec); - - anotherHalfRegRec->assignedInterval = regRec->assignedInterval; - anotherHalfRegRec->previousInterval = nullptr; + makeRegsAvailable(regsToMakeInactive); + // TODO: Clean this up. We need to make the delayRegs inactive as well, but don't want + // to mark them as free yet. + regsToMakeInactive |= delayRegsToMakeInactive; + regsToMakeInactive = delayRegsToMakeInactive; + delayRegsToMakeInactive = RBM_NONE; } -#endif // TARGET_ARM #ifdef DEBUG - if (spill) - { - dumpLsraAllocationEvent(LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL_AFTER_SPILL, regRec->assignedInterval, - thisRegNum); - } - else + // Set the activeRefPosition to null until we're done with any boundary handling. + activeRefPosition = nullptr; + if (VERBOSE) { - dumpLsraAllocationEvent(LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL, regRec->assignedInterval, thisRegNum); + // We're really dumping the RegRecords "after" the previous RefPosition, but it's more convenient + // to do this here, since there are a number of "continue"s in this loop. + dumpRegRecords(); } #endif // DEBUG - } - else - { - updateAssignedInterval(regRec, nullptr, assignedInterval->registerType); - updatePreviousInterval(regRec, nullptr, assignedInterval->registerType); - } -} -//------------------------------------------------------------------------ -// spillGCRefs: Spill any GC-type intervals that are currently in registers. -// -// Arguments: -// killRefPosition - The RefPosition for the kill -// -// Return Value: -// None. -// -// Notes: -// This is used to ensure that we have no live GC refs in registers at an -// unmanaged call. -// -void LinearScan::spillGCRefs(RefPosition* killRefPosition) -{ - // For each physical register that can hold a GC type, - // if it is occupied by an interval of a GC type, spill that interval. - regMaskTP candidateRegs = killRefPosition->registerAssignment; - INDEBUG(bool killedRegs = false); - while (candidateRegs != RBM_NONE) - { - regMaskTP nextRegBit = genFindLowestBit(candidateRegs); - candidateRegs &= ~nextRegBit; - regNumber nextReg = genRegNumFromMask(nextRegBit); - RegRecord* regRecord = getRegisterRecord(nextReg); - Interval* assignedInterval = regRecord->assignedInterval; - if (assignedInterval == nullptr || (assignedInterval->isActive == false)) + // This is the previousRefPosition of the current Referent, if any + RefPosition* previousRefPosition = nullptr; + + Interval* currentInterval = nullptr; + Referenceable* currentReferent = nullptr; + RefType refType = currentRefPosition->refType; + + currentReferent = currentRefPosition->referent; + + if (spillAlways() && lastAllocatedRefPosition != nullptr && !lastAllocatedRefPosition->IsPhysRegRef() && + !lastAllocatedRefPosition->getInterval()->isInternal && + (RefTypeIsDef(lastAllocatedRefPosition->refType) || lastAllocatedRefPosition->getInterval()->isLocalVar)) { - continue; + assert(lastAllocatedRefPosition->registerAssignment != RBM_NONE); + RegRecord* regRecord = lastAllocatedRefPosition->getInterval()->assignedReg; + unassignPhysReg(regRecord, lastAllocatedRefPosition); + // Now set lastAllocatedRefPosition to null, so that we don't try to spill it again + lastAllocatedRefPosition = nullptr; } - bool needsKill = varTypeIsGC(assignedInterval->registerType); - if (!needsKill) + + // We wait to free any registers until we've completed all the + // uses for the current node. + // This avoids reusing registers too soon. + // We free before the last true def (after all the uses & internal + // registers), and then again at the beginning of the next node. + // This is made easier by assigning two LsraLocations per node - one + // for all the uses, internal registers & all but the last def, and + // another for the final def (if any). + + LsraLocation currentLocation = currentRefPosition->nodeLocation; + + // Free at a new location. + if (currentLocation > prevLocation) { - // The importer will assign a GC type to the rhs of an assignment if the lhs type is a GC type, - // even if the rhs is not. See the CEE_STLOC* case in impImportBlockCode(). As a result, - // we can have a 'GT_LCL_VAR' node with a GC type, when the lclVar itself is an integer type. - // The emitter will mark this register as holding a GC type. Therfore we must spill this value. - // This was exposed on Arm32 with EH write-thru. - if ((assignedInterval->recentRefPosition != nullptr) && - (assignedInterval->recentRefPosition->treeNode != nullptr)) + // CopyRegs are simply made available - we don't want to make the associated interval inactive. + makeRegsAvailable(copyRegsToFree); + copyRegsToFree = RBM_NONE; + regsInUseThisLocation = regsInUseNextLocation; + regsInUseNextLocation = RBM_NONE; + if ((regsToFree | delayRegsToFree) != RBM_NONE) { - needsKill = varTypeIsGC(assignedInterval->recentRefPosition->treeNode); + freeRegisters(regsToFree); + if ((currentLocation > (prevLocation + 1)) && (delayRegsToFree != RBM_NONE)) + { + // We should never see a delayReg that is delayed until a Location that has no RefPosition + // (that would be the RefPosition that it was supposed to interfere with). + assert(!"Found a delayRegFree associated with Location with no reference"); + // However, to be cautious for the Release build case, we will free them. + freeRegisters(delayRegsToFree); + delayRegsToFree = RBM_NONE; + regsInUseThisLocation = RBM_NONE; + } + regsToFree = delayRegsToFree; + delayRegsToFree = RBM_NONE; + +#ifdef DEBUG + // Validate the current state just after we've freed the registers. This ensures that any pending + // freed registers will have had their state updated to reflect the intervals they were holding. + for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) + { + regMaskTP regMask = genRegMask(reg); + // If this isn't available or if it's still waiting to be freed (i.e. it was in + // delayRegsToFree and so now it's in regsToFree), then skip it. + if ((regMask & (availableIntRegs | availableFloatRegs) & ~regsToFree) == RBM_NONE) + { + continue; + } + RegRecord* physRegRecord = getRegisterRecord(reg); + Interval* assignedInterval = physRegRecord->assignedInterval; + if (assignedInterval != nullptr) + { + bool isAssignedReg = (assignedInterval->physReg == reg); + RefPosition* recentRefPosition = assignedInterval->recentRefPosition; + // If we have a copyReg or a moveReg, we might have assigned this register to an Interval, + // but that isn't considered its assignedReg. + if (recentRefPosition != nullptr) + { + if (recentRefPosition->refType == RefTypeExpUse) + { + // We don't update anything on these, as they're just placeholders to extend the + // lifetime. + continue; + } + // For copyReg or moveReg, we don't have anything further to assert. + if (recentRefPosition->copyReg || recentRefPosition->moveReg) + { + continue; + } + assert(assignedInterval->isConstant == isRegConstant(reg, assignedInterval->registerType)); + if (assignedInterval->isActive) + { + // If this is not the register most recently allocated, it must be from a copyReg, + // or it was placed there by the inVarToRegMap. In either case it must be a lclVar. + + if (!isAssignedToInterval(assignedInterval, physRegRecord)) + { + assert(assignedInterval->isLocalVar); + // We'd like to assert that this was either set by the inVarToRegMap, or by + // a copyReg, but we can't traverse backward to check for a copyReg, because + // we only have recentRefPosition, and there may be a previous RefPosition + // at the same Location with a copyReg. + } + if (isAssignedReg) + { + assert(nextIntervalRef[reg] == assignedInterval->getNextRefLocation()); + assert(!isRegAvailable(reg, assignedInterval->registerType)); + assert((recentRefPosition == nullptr) || + (spillCost[reg] == getSpillWeight(physRegRecord))); + } + else + { + assert((nextIntervalRef[reg] == MaxLocation) || + isRegBusy(reg, assignedInterval->registerType)); + } + } + else + { + if ((assignedInterval->physReg == reg) && !assignedInterval->isConstant) + { + assert(nextIntervalRef[reg] == assignedInterval->getNextRefLocation()); + } + else + { + assert(nextIntervalRef[reg] == MaxLocation); + assert(isRegAvailable(reg, assignedInterval->registerType)); + assert(spillCost[reg] == 0); + } + } + } + } + else + { + assert(isRegAvailable(reg, physRegRecord->registerType)); + assert(!isRegConstant(reg, physRegRecord->registerType)); + assert(nextIntervalRef[reg] == MaxLocation); + assert(spillCost[reg] == 0); + } + LsraLocation thisNextFixedRef = physRegRecord->getNextRefLocation(); + assert(nextFixedRef[reg] == thisNextFixedRef); +#ifdef TARGET_ARM + // If this is occupied by a double interval, skip the corresponding float reg. + if ((assignedInterval != nullptr) && (assignedInterval->registerType == TYP_DOUBLE)) + { + reg = REG_NEXT(reg); + } +#endif + } +#endif // DEBUG } } - if (needsKill) + prevLocation = currentLocation; + + // get previous refposition, then current refpos is the new previous + if (currentReferent != nullptr) { - INDEBUG(killedRegs = true); - unassignPhysReg(regRecord, assignedInterval->recentRefPosition); - makeRegAvailable(nextReg, assignedInterval->registerType); + previousRefPosition = currentReferent->recentRefPosition; + currentReferent->recentRefPosition = currentRefPosition; + } + else + { + assert((refType == RefTypeBB) || (refType == RefTypeKillGCRefs)); } - } - INDEBUG(dumpLsraAllocationEvent(killedRegs ? LSRA_EVENT_DONE_KILL_GC_REFS : LSRA_EVENT_NO_GC_KILLS, nullptr, REG_NA, - nullptr)); -} - -//------------------------------------------------------------------------ -// processBlockEndAllocation: Update var locations after 'currentBlock' has been allocated -// -// Arguments: -// currentBlock - the BasicBlock we have just finished allocating registers for -// -// Return Value: -// None -// -// Notes: -// Calls processBlockEndLocations() to set the outVarToRegMap, then gets the next block, -// and sets the inVarToRegMap appropriately. -void LinearScan::processBlockEndAllocation(BasicBlock* currentBlock) -{ - assert(currentBlock != nullptr); - if (enregisterLocalVars) - { - processBlockEndLocations(currentBlock); - } - markBlockVisited(currentBlock); +#ifdef DEBUG + activeRefPosition = currentRefPosition; - // Get the next block to allocate. - // When the last block in the method has successors, there will be a final "RefTypeBB" to - // ensure that we get the varToRegMap set appropriately, but in that case we don't need - // to worry about "nextBlock". - BasicBlock* nextBlock = getNextBlock(); - if (nextBlock != nullptr) - { - processBlockStartLocations(nextBlock); - } -} + // For the purposes of register resolution, we handle the DummyDefs before + // the block boundary - so the RefTypeBB is after all the DummyDefs. + // However, for the purposes of allocation, we want to handle the block + // boundary first, so that we can free any registers occupied by lclVars + // that aren't live in the next block and make them available for the + // DummyDefs. -//------------------------------------------------------------------------ -// rotateBlockStartLocation: When in the LSRA_BLOCK_BOUNDARY_ROTATE stress mode, attempt to -// "rotate" the register assignment for a localVar to the next higher -// register that is available. -// -// Arguments: -// interval - the Interval for the variable whose register is getting rotated -// targetReg - its register assignment from the predecessor block being used for live-in -// availableRegs - registers available for use -// -// Return Value: -// The new register to use. + // If we've already handled the BlockEnd, but now we're seeing the RefTypeBB, + // dump it now. + if ((refType == RefTypeBB) && handledBlockEnd) + { + dumpNewBlock(currentBlock, currentRefPosition->nodeLocation); + } +#endif // DEBUG -#ifdef DEBUG -regNumber LinearScan::rotateBlockStartLocation(Interval* interval, regNumber targetReg, regMaskTP availableRegs) -{ - if (targetReg != REG_STK && getLsraBlockBoundaryLocations() == LSRA_BLOCK_BOUNDARY_ROTATE) - { - // If we're rotating the register locations at block boundaries, try to use - // the next higher register number of the appropriate register type. - regMaskTP candidateRegs = allRegs(interval->registerType) & availableRegs; - regNumber firstReg = REG_NA; - regNumber newReg = REG_NA; - while (candidateRegs != RBM_NONE) + if (!handledBlockEnd && (refType == RefTypeBB || refType == RefTypeDummyDef)) { - regMaskTP nextRegBit = genFindLowestBit(candidateRegs); - candidateRegs &= ~nextRegBit; - regNumber nextReg = genRegNumFromMask(nextRegBit); - if (nextReg > targetReg) + // Free any delayed regs (now in regsToFree) before processing the block boundary + freeRegisters(regsToFree); + regsToFree = RBM_NONE; + regsInUseThisLocation = RBM_NONE; + regsInUseNextLocation = RBM_NONE; + handledBlockEnd = true; + curBBStartLocation = currentRefPosition->nodeLocation; + if (currentBlock == nullptr) { - newReg = nextReg; - break; + currentBlock = startBlockSequence(); + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_START_BB, nullptr, REG_NA, compiler->fgFirstBB)); } - else if (firstReg == REG_NA) + else { - firstReg = nextReg; + processBlockEndAllocation(currentBlock); + currentBlock = moveToNextBlock(); + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_START_BB, nullptr, REG_NA, currentBlock)); } } - if (newReg == REG_NA) + + if (refType == RefTypeBB) { - assert(firstReg != REG_NA); - newReg = firstReg; + handledBlockEnd = false; + continue; } - targetReg = newReg; - } - return targetReg; -} -#endif // DEBUG -#ifdef TARGET_ARM -//-------------------------------------------------------------------------------------- -// isSecondHalfReg: Test if recRec is second half of double register -// which is assigned to an interval. -// -// Arguments: -// regRec - a register to be tested -// interval - an interval which is assigned to some register -// -// Assumptions: -// None -// -// Return Value: -// True only if regRec is second half of assignedReg in interval -// -bool LinearScan::isSecondHalfReg(RegRecord* regRec, Interval* interval) -{ - RegRecord* assignedReg = interval->assignedReg; + if (refType == RefTypeKillGCRefs) + { + spillGCRefs(currentRefPosition); + continue; + } - if (assignedReg != nullptr && interval->registerType == TYP_DOUBLE) - { - // interval should have been allocated to a valid double register - assert(genIsValidDoubleReg(assignedReg->regNum)); - - // Find a second half RegRecord of double register - regNumber firstRegNum = assignedReg->regNum; - regNumber secondRegNum = REG_NEXT(firstRegNum); - - assert(genIsValidFloatReg(secondRegNum) && !genIsValidDoubleReg(secondRegNum)); - - RegRecord* secondRegRec = getRegisterRecord(secondRegNum); - - return secondRegRec == regRec; - } - - return false; -} - -//------------------------------------------------------------------------------------------ -// getSecondHalfRegRec: Get the second (odd) half of an ARM32 double register -// -// Arguments: -// regRec - A float RegRecord -// -// Assumptions: -// regRec must be a valid double register (i.e. even) -// -// Return Value: -// The RegRecord for the second half of the double register -// -RegRecord* LinearScan::getSecondHalfRegRec(RegRecord* regRec) -{ - regNumber secondHalfRegNum; - RegRecord* secondHalfRegRec; - - assert(genIsValidDoubleReg(regRec->regNum)); - - secondHalfRegNum = REG_NEXT(regRec->regNum); - secondHalfRegRec = getRegisterRecord(secondHalfRegNum); - - return secondHalfRegRec; -} -//------------------------------------------------------------------------------------------ -// findAnotherHalfRegRec: Find another half RegRecord which forms same ARM32 double register -// -// Arguments: -// regRec - A float RegRecord -// -// Assumptions: -// None -// -// Return Value: -// A RegRecord which forms same double register with regRec -// -RegRecord* LinearScan::findAnotherHalfRegRec(RegRecord* regRec) -{ - regNumber anotherHalfRegNum = findAnotherHalfRegNum(regRec->regNum); - return getRegisterRecord(anotherHalfRegNum); -} -//------------------------------------------------------------------------------------------ -// findAnotherHalfRegNum: Find another half register's number which forms same ARM32 double register -// -// Arguments: -// regNumber - A float regNumber -// -// Assumptions: -// None -// -// Return Value: -// A register number which forms same double register with regNum. -// -regNumber LinearScan::findAnotherHalfRegNum(regNumber regNum) -{ - regNumber anotherHalfRegNum; - - assert(genIsValidFloatReg(regNum)); - - // Find another half register for TYP_DOUBLE interval, - // following same logic in canRestorePreviousInterval(). - if (genIsValidDoubleReg(regNum)) - { - anotherHalfRegNum = REG_NEXT(regNum); - assert(!genIsValidDoubleReg(anotherHalfRegNum)); - } - else - { - anotherHalfRegNum = REG_PREV(regNum); - assert(genIsValidDoubleReg(anotherHalfRegNum)); - } - - return anotherHalfRegNum; -} -#endif - -//-------------------------------------------------------------------------------------- -// canRestorePreviousInterval: Test if we can restore previous interval -// -// Arguments: -// regRec - a register which contains previous interval to be restored -// assignedInterval - an interval just unassigned -// -// Assumptions: -// None -// -// Return Value: -// True only if previous interval of regRec can be restored -// -bool LinearScan::canRestorePreviousInterval(RegRecord* regRec, Interval* assignedInterval) -{ - bool retVal = - (regRec->previousInterval != nullptr && regRec->previousInterval != assignedInterval && - regRec->previousInterval->assignedReg == regRec && regRec->previousInterval->getNextRefPosition() != nullptr); - -#ifdef TARGET_ARM - if (retVal && regRec->previousInterval->registerType == TYP_DOUBLE) - { - RegRecord* anotherHalfRegRec = findAnotherHalfRegRec(regRec); + if (currentRefPosition->isPhysRegRef) + { + RegRecord* regRecord = currentRefPosition->getReg(); + Interval* assignedInterval = regRecord->assignedInterval; - retVal = retVal && anotherHalfRegRec->assignedInterval == nullptr; - } -#endif + updateNextFixedRef(regRecord, currentRefPosition->nextRefPosition); - return retVal; -} + // If this is a FixedReg, disassociate any inactive constant interval from this register. + // Otherwise, do nothing. + if (refType == RefTypeFixedReg) + { + if (assignedInterval != nullptr && !assignedInterval->isActive && assignedInterval->isConstant) + { + clearConstantReg(regRecord->regNum, assignedInterval->registerType); + regRecord->assignedInterval = nullptr; + spillCost[regRecord->regNum] = 0; -bool LinearScan::isAssignedToInterval(Interval* interval, RegRecord* regRec) -{ - bool isAssigned = (interval->assignedReg == regRec); #ifdef TARGET_ARM - isAssigned |= isSecondHalfReg(regRec, interval); -#endif - return isAssigned; -} - -void LinearScan::unassignIntervalBlockStart(RegRecord* regRecord, VarToRegMap inVarToRegMap) -{ - // Is there another interval currently assigned to this register? If so unassign it. - Interval* assignedInterval = regRecord->assignedInterval; - if (assignedInterval != nullptr) - { - if (isAssignedToInterval(assignedInterval, regRecord)) - { - // Only localVars, constants or vector upper halves should be assigned to registers at block boundaries. - if (!assignedInterval->isLocalVar) + // Update overlapping floating point register for TYP_DOUBLE + if (assignedInterval->registerType == TYP_DOUBLE) + { + RegRecord* otherRegRecord = findAnotherHalfRegRec(regRecord); + assert(otherRegRecord->assignedInterval == assignedInterval); + otherRegRecord->assignedInterval = nullptr; + spillCost[otherRegRecord->regNum] = 0; + } +#endif // TARGET_ARM + } + regsInUseThisLocation |= currentRefPosition->registerAssignment; + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_FIXED_REG, nullptr, currentRefPosition->assignedReg())); + continue; + } + if (refType == RefTypeKill) { - assert(assignedInterval->isConstant || assignedInterval->IsUpperVector()); - // Don't need to update the VarToRegMap. - inVarToRegMap = nullptr; + if (assignedInterval != nullptr) + { + unassignPhysReg(regRecord, assignedInterval->recentRefPosition); + clearConstantReg(regRecord->regNum, assignedInterval->registerType); + makeRegAvailable(regRecord->regNum, assignedInterval->registerType); + } + clearRegBusyUntilKill(regRecord->regNum); + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum)); + continue; } + } - regNumber assignedRegNum = assignedInterval->assignedReg->regNum; + // If this is an exposed use, do nothing - this is merely a placeholder to attempt to + // ensure that a register is allocated for the full lifetime. The resolution logic + // will take care of moving to the appropriate register if needed. - // If the interval is active, it will be set to active when we reach its new - // register assignment (which we must not yet have done, or it wouldn't still be - // assigned to this register). - assignedInterval->isActive = false; - unassignPhysReg(assignedInterval->assignedReg, nullptr); - if ((inVarToRegMap != nullptr) && inVarToRegMap[assignedInterval->getVarIndex(compiler)] == assignedRegNum) + if (refType == RefTypeExpUse) + { + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_EXP_USE)); + currentInterval = currentRefPosition->getInterval(); + if (currentInterval->physReg != REG_NA) { - inVarToRegMap[assignedInterval->getVarIndex(compiler)] = REG_STK; + updateNextIntervalRef(currentInterval->physReg, currentInterval); } + continue; } - else - { - // This interval is no longer assigned to this register. - updateAssignedInterval(regRecord, nullptr, assignedInterval->registerType); - } - } -} -//------------------------------------------------------------------------ -// processBlockStartLocations: Update var locations on entry to 'currentBlock' and clear constant -// registers. -// -// Arguments: -// currentBlock - the BasicBlock we are about to allocate registers for -// allocationPass - true if we are currently allocating registers (versus writing them back) -// -// Return Value: -// None -// -// Notes: -// During the allocation pass, we use the outVarToRegMap of the selected predecessor to -// determine the lclVar locations for the inVarToRegMap. -// During the resolution (write-back) pass, we only modify the inVarToRegMap in cases where -// a lclVar was spilled after the block had been completed. -void LinearScan::processBlockStartLocations(BasicBlock* currentBlock) -{ - // If we have no register candidates we should only call this method during allocation. + regNumber assignedRegister = REG_NA; - assert(enregisterLocalVars || !allocationPassComplete); + assert(currentRefPosition->isIntervalRef()); + currentInterval = currentRefPosition->getInterval(); + assert(currentInterval != nullptr); + assignedRegister = currentInterval->physReg; - if (!enregisterLocalVars) - { - // Just clear any constant registers and return. - resetAvailableRegs(); - for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) + // Identify the special cases where we decide up-front not to allocate + bool allocate = true; + bool didDump = false; + + if (refType == RefTypeParamDef || refType == RefTypeZeroInit) { - RegRecord* physRegRecord = getRegisterRecord(reg); - Interval* assignedInterval = physRegRecord->assignedInterval; - clearNextIntervalRef(reg, physRegRecord->registerType); - clearSpillCost(reg, physRegRecord->registerType); - if (assignedInterval != nullptr) + if (nextRefPosition == nullptr) { - assert(assignedInterval->isConstant); - physRegRecord->assignedInterval = nullptr; + // If it has no actual references, mark it as "lastUse"; since they're not actually part + // of any flow they won't have been marked during dataflow. Otherwise, if we allocate a + // register we won't unassign it. + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_ZERO_REF, currentInterval)); + currentRefPosition->lastUse = true; } - } - return; - } - - unsigned predBBNum = blockInfo[currentBlock->bbNum].predBBNum; - VarToRegMap predVarToRegMap = getOutVarToRegMap(predBBNum); - VarToRegMap inVarToRegMap = getInVarToRegMap(currentBlock->bbNum); - - // If this block enters an exception region, all incoming vars are on the stack. - if (predBBNum == 0) - { -#if DEBUG - if (blockInfo[currentBlock->bbNum].hasEHBoundaryIn || !allocationPassComplete) - { - // This should still be in its initialized empty state. - for (unsigned varIndex = 0; varIndex < compiler->lvaTrackedCount; varIndex++) + LclVarDsc* varDsc = currentInterval->getLocalVar(compiler); + assert(varDsc != nullptr); + assert(!blockInfo[compiler->fgFirstBB->bbNum].hasEHBoundaryIn || currentInterval->isWriteThru); + if (blockInfo[compiler->fgFirstBB->bbNum].hasEHBoundaryIn || + blockInfo[compiler->fgFirstBB->bbNum].hasEHPred) { - // In the case where we're extending lifetimes for stress, we are intentionally modeling variables - // as live when they really aren't to create extra register pressure & constraints. - // However, this means that non-EH-vars will be live into EH regions. We can and should ignore the - // locations of these. Note that they aren't reported to codegen anyway. - if (!getLsraExtendLifeTimes() || VarSetOps::IsMember(compiler, currentBlock->bbLiveIn, varIndex)) - { - assert(inVarToRegMap[varIndex] == REG_STK); - } + allocate = false; } - } -#endif // DEBUG - predVarToRegMap = inVarToRegMap; - } - - VarSetOps::AssignNoCopy(compiler, currentLiveVars, - VarSetOps::Intersection(compiler, registerCandidateVars, currentBlock->bbLiveIn)); -#ifdef DEBUG - if (getLsraExtendLifeTimes()) - { - VarSetOps::AssignNoCopy(compiler, currentLiveVars, registerCandidateVars); - } - // If we are rotating register assignments at block boundaries, we want to make the - // inactive registers available for the rotation. - regMaskTP inactiveRegs = RBM_NONE; -#endif // DEBUG - regMaskTP liveRegs = RBM_NONE; - VarSetOps::Iter iter(compiler, currentLiveVars); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex)) - { - if (!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate) - { - continue; - } - regNumber targetReg; - Interval* interval = getIntervalForLocalVar(varIndex); - RefPosition* nextRefPosition = interval->getNextRefPosition(); - assert((nextRefPosition != nullptr) || (interval->isWriteThru)); - - bool leaveOnStack = false; - - // Special handling for variables live in/out of exception handlers. - if (interval->isWriteThru) - { - // There are 3 cases where we will leave writethru lclVars on the stack: - // 1) There is no predecessor. - // 2) It is conservatively or artificially live - that is, it has no next use, - // so there is no place for codegen to record that the register is no longer occupied. - // 3) This block has a predecessor with an outgoing EH edge. We won't be able to add "join" - // resolution to load the EH var into a register along that edge, so it must be on stack. - if ((predBBNum == 0) || (nextRefPosition == nullptr) || (RefTypeIsDef(nextRefPosition->refType)) || - blockInfo[currentBlock->bbNum].hasEHPred) + else if (refType == RefTypeParamDef && (varDsc->lvRefCntWtd() <= BB_UNITY_WEIGHT) && + (!currentRefPosition->lastUse || (currentInterval->physReg == REG_STK))) { - leaveOnStack = true; + // If this is a low ref-count parameter, and either it is used (def is not the last use) or it's + // passed on the stack, don't allocate a register. + // Note that if this is an unused register parameter we don't want to set allocate to false because that + // will cause us to allocate stack space to spill it. + allocate = false; } - } - - if (!allocationPassComplete) - { - targetReg = getVarReg(predVarToRegMap, varIndex); - if (leaveOnStack) + else if ((currentInterval->physReg == REG_STK) && nextRefPosition->treeNode->OperIs(GT_BITCAST)) { - targetReg = REG_STK; + // In the case of ABI mismatches, avoid allocating a register only to have to immediately move + // it to a different register file. + allocate = false; } -#ifdef DEBUG - regNumber newTargetReg = rotateBlockStartLocation(interval, targetReg, (~liveRegs | inactiveRegs)); - if (newTargetReg != targetReg) + if (!allocate) { - targetReg = newTargetReg; - setIntervalAsSplit(interval); + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_ENTRY_REG_ALLOCATED, currentInterval)); + didDump = true; + setIntervalAsSpilled(currentInterval); + if (assignedRegister != REG_NA) + { + clearNextIntervalRef(assignedRegister, currentInterval->registerType); + clearSpillCost(assignedRegister, currentInterval->registerType); + makeRegAvailable(assignedRegister, currentInterval->registerType); + } } -#endif // DEBUG - setVarReg(inVarToRegMap, varIndex, targetReg); } - else // allocationPassComplete (i.e. resolution/write-back pass) +#ifdef FEATURE_SIMD +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + else if (currentInterval->isUpperVector) { - targetReg = getVarReg(inVarToRegMap, varIndex); - // There are four cases that we need to consider during the resolution pass: - // 1. This variable had a register allocated initially, and it was not spilled in the RefPosition - // that feeds this block. In this case, both targetReg and predVarToRegMap[varIndex] will be targetReg. - // 2. This variable had not been spilled prior to the end of predBB, but was later spilled, so - // predVarToRegMap[varIndex] will be REG_STK, but targetReg is its former allocated value. - // In this case, we will normally change it to REG_STK. We will update its "spilled" status when we - // encounter it in resolveLocalRef(). - // 2a. If the next RefPosition is marked as a copyReg, we need to retain the allocated register. This is - // because the copyReg RefPosition will not have recorded the "home" register, yet downstream - // RefPositions rely on the correct "home" register. - // 3. This variable was spilled before we reached the end of predBB. In this case, both targetReg and - // predVarToRegMap[varIndex] will be REG_STK, and the next RefPosition will have been marked - // as reload during allocation time if necessary (note that by the time we actually reach the next - // RefPosition, we may be using a different predecessor, at which it is still in a register). - // 4. This variable was spilled during the allocation of this block, so targetReg is REG_STK - // (because we set inVarToRegMap at the time we spilled it), but predVarToRegMap[varIndex] - // is not REG_STK. We retain the REG_STK value in the inVarToRegMap. - if (targetReg != REG_STK) + // This is a save or restore of the upper half of a large vector lclVar. + Interval* lclVarInterval = currentInterval->relatedInterval; + assert(lclVarInterval->isLocalVar); + if (refType == RefTypeUpperVectorSave) { - if (getVarReg(predVarToRegMap, varIndex) != REG_STK) + if ((lclVarInterval->physReg == REG_NA) || + (lclVarInterval->isPartiallySpilled && (currentInterval->physReg == REG_STK))) { - // Case #1 above. - assert(getVarReg(predVarToRegMap, varIndex) == targetReg || - getLsraBlockBoundaryLocations() == LSRA_BLOCK_BOUNDARY_ROTATE); + allocate = false; } - else if (!nextRefPosition->copyReg) + else { - // case #2 above. - setVarReg(inVarToRegMap, varIndex, REG_STK); - targetReg = REG_STK; + lclVarInterval->isPartiallySpilled = true; } - // Else case 2a. - retain targetReg. - } - // Else case #3 or #4, we retain targetReg and nothing further to do or assert. - } - - if (interval->physReg == targetReg) - { - if (interval->isActive) - { - assert(targetReg != REG_STK); - assert(interval->assignedReg != nullptr && interval->assignedReg->regNum == targetReg && - interval->assignedReg->assignedInterval == interval); - liveRegs |= getRegMask(targetReg, interval->registerType); - continue; } - } - else if (interval->physReg != REG_NA) - { - // This can happen if we are using the locations from a basic block other than the - // immediately preceding one - where the variable was in a different location. - if ((targetReg != REG_STK) || leaveOnStack) + else if (refType == RefTypeUpperVectorRestore) { - // Unassign it from the register (it may get a new register below). - if (interval->assignedReg != nullptr && interval->assignedReg->assignedInterval == interval) + assert(currentInterval->isUpperVector); + if (lclVarInterval->isPartiallySpilled) { - interval->isActive = false; - unassignPhysReg(getRegisterRecord(interval->physReg), nullptr); + lclVarInterval->isPartiallySpilled = false; } else { - // This interval was live in this register the last time we saw a reference to it, - // but has since been displaced. - interval->physReg = REG_NA; + allocate = false; } } - else if (!allocationPassComplete) - { - // Keep the register assignment - if another var has it, it will get unassigned. - // Otherwise, resolution will fix it up later, and it will be more - // likely to match other assignments this way. - targetReg = interval->physReg; - interval->isActive = true; - liveRegs |= getRegMask(targetReg, interval->registerType); - INDEBUG(inactiveRegs |= genRegMask(targetReg)); - setVarReg(inVarToRegMap, varIndex, targetReg); - } - else + } + else if (refType == RefTypeUpperVectorSave) + { + assert(!currentInterval->isLocalVar); + // Note that this case looks a lot like the case below, but in this case we need to spill + // at the previous RefPosition. + // We may want to consider allocating two callee-save registers for this case, but it happens rarely + // enough that it may not warrant the additional complexity. + if (assignedRegister != REG_NA) { - interval->physReg = REG_NA; + unassignPhysReg(getRegisterRecord(assignedRegister), currentInterval->firstRefPosition); + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, currentInterval)); } + currentRefPosition->registerAssignment = RBM_NONE; + continue; } - if (targetReg != REG_STK) +#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE +#endif // FEATURE_SIMD + + if (allocate == false) { - RegRecord* targetRegRecord = getRegisterRecord(targetReg); - liveRegs |= getRegMask(targetReg, interval->registerType); - if (!allocationPassComplete) + if (assignedRegister != REG_NA) { - updateNextIntervalRef(targetReg, interval); - updateSpillCost(targetReg, interval); + unassignPhysReg(getRegisterRecord(assignedRegister), currentRefPosition); } - if (!interval->isActive) + else if (!didDump) { - interval->isActive = true; - interval->physReg = targetReg; - interval->assignedReg = targetRegRecord; + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, currentInterval)); + didDump = true; } - if (targetRegRecord->assignedInterval != interval) + currentRefPosition->registerAssignment = RBM_NONE; + continue; + } + + if (currentInterval->isSpecialPutArg) + { + assert(!currentInterval->isLocalVar); + Interval* srcInterval = currentInterval->relatedInterval; + assert(srcInterval != nullptr && srcInterval->isLocalVar); + if (refType == RefTypeDef) { -#ifdef TARGET_ARM - // If this is a TYP_DOUBLE interval, and the assigned interval is either null or is TYP_FLOAT, - // we also need to unassign the other half of the register. - // Note that if the assigned interval is TYP_DOUBLE, it will be unassigned below. - if ((interval->registerType == TYP_DOUBLE) && - ((targetRegRecord->assignedInterval == nullptr) || - (targetRegRecord->assignedInterval->registerType == TYP_FLOAT))) - { - assert(genIsValidDoubleReg(targetReg)); - unassignIntervalBlockStart(getSecondHalfRegRec(targetRegRecord), - allocationPassComplete ? nullptr : inVarToRegMap); - } + assert(srcInterval->recentRefPosition->nodeLocation == currentLocation - 1); + RegRecord* physRegRecord = srcInterval->assignedReg; - // If this is a TYP_FLOAT interval, and the assigned interval was TYP_DOUBLE, we also - // need to update the liveRegs to specify that the other half is not live anymore. - // As mentioned above, for TYP_DOUBLE, the other half will be unassigned further below. - if ((interval->registerType == TYP_FLOAT) && - ((targetRegRecord->assignedInterval != nullptr) && - (targetRegRecord->assignedInterval->registerType == TYP_DOUBLE))) - { - RegRecord* anotherHalfRegRec = findAnotherHalfRegRec(targetRegRecord); + // For a putarg_reg to be special, its next use location has to be the same + // as fixed reg's next kill location. Otherwise, if source lcl var's next use + // is after the kill of fixed reg but before putarg_reg's next use, fixed reg's + // kill would lead to spill of source but not the putarg_reg if it were treated + // as special. + if (srcInterval->isActive && + genRegMask(srcInterval->physReg) == currentRefPosition->registerAssignment && + currentInterval->getNextRefLocation() == nextFixedRef[srcInterval->physReg]) + { + assert(physRegRecord->regNum == srcInterval->physReg); - // Use TYP_FLOAT to get the regmask of just the half reg. - liveRegs &= ~getRegMask(anotherHalfRegRec->regNum, TYP_FLOAT); + // Special putarg_reg acts as a pass-thru since both source lcl var + // and putarg_reg have the same register allocated. Physical reg + // record of reg continue to point to source lcl var's interval + // instead of to putarg_reg's interval. So if a spill of reg + // allocated to source lcl var happens, to reallocate to another + // tree node, before its use at call node it will lead to spill of + // lcl var instead of putarg_reg since physical reg record is pointing + // to lcl var's interval. As a result, arg reg would get trashed leading + // to bad codegen. The assumption here is that source lcl var of a + // special putarg_reg doesn't get spilled and re-allocated prior to + // its use at the call node. This is ensured by marking physical reg + // record as busy until next kill. + setRegBusyUntilKill(srcInterval->physReg, srcInterval->registerType); + } + else + { + currentInterval->isSpecialPutArg = false; } - -#endif // TARGET_ARM - unassignIntervalBlockStart(targetRegRecord, allocationPassComplete ? nullptr : inVarToRegMap); - assignPhysReg(targetRegRecord, interval); } - if (interval->recentRefPosition != nullptr && !interval->recentRefPosition->copyReg && - interval->recentRefPosition->registerAssignment != genRegMask(targetReg)) + // If this is still a SpecialPutArg, continue; + if (currentInterval->isSpecialPutArg) { - interval->getNextRefPosition()->outOfOrder = true; + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_SPECIAL_PUTARG, currentInterval, + currentRefPosition->assignedReg())); + continue; } } - } - // Unassign any registers that are no longer live, and set register state, if allocating. - if (!allocationPassComplete) - { - resetRegState(); - setRegsInUse(liveRegs); - } - for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) - { - RegRecord* physRegRecord = getRegisterRecord(reg); - if ((liveRegs & genRegMask(reg)) == 0) + if (assignedRegister == REG_NA && RefTypeIsUse(refType)) { - makeRegAvailable(reg, physRegRecord->registerType); - Interval* assignedInterval = physRegRecord->assignedInterval; + currentRefPosition->reload = true; + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_RELOAD, currentInterval, assignedRegister)); + } - if (assignedInterval != nullptr) + regMaskTP assignedRegBit = RBM_NONE; + bool isInRegister = false; + if (assignedRegister != REG_NA) + { + isInRegister = true; + assignedRegBit = genRegMask(assignedRegister); + if (!currentInterval->isActive) { - assert(assignedInterval->isLocalVar || assignedInterval->isConstant || - assignedInterval->IsUpperVector()); + // If this is a use, it must have started the block on the stack, but the register + // was available for use so we kept the association. + if (RefTypeIsUse(refType)) + { + assert(enregisterLocalVars); + assert(inVarToRegMaps[curBBNum][currentInterval->getVarIndex(compiler)] == REG_STK && + previousRefPosition->nodeLocation <= curBBStartLocation); + isInRegister = false; + } + else + { + currentInterval->isActive = true; + setRegInUse(assignedRegister, currentInterval->registerType); + updateSpillCost(assignedRegister, currentInterval); + } + updateNextIntervalRef(assignedRegister, currentInterval); + } + assert(currentInterval->assignedReg != nullptr && + currentInterval->assignedReg->regNum == assignedRegister && + currentInterval->assignedReg->assignedInterval == currentInterval); + } - if (!assignedInterval->isConstant && assignedInterval->assignedReg == physRegRecord) + if (previousRefPosition != nullptr) + { + assert(previousRefPosition->nextRefPosition == currentRefPosition); + assert(assignedRegister == REG_NA || assignedRegBit == previousRefPosition->registerAssignment || + currentRefPosition->outOfOrder || previousRefPosition->copyReg || + previousRefPosition->refType == RefTypeExpUse || currentRefPosition->refType == RefTypeDummyDef); + } + else if (assignedRegister != REG_NA) + { + // Handle the case where this is a preassigned register (i.e. parameter). + // We don't want to actually use the preassigned register if it's not + // going to cover the lifetime - but we had to preallocate it to ensure + // that it remained live. + // TODO-CQ: At some point we may want to refine the analysis here, in case + // it might be beneficial to keep it in this reg for PART of the lifetime + if (currentInterval->isLocalVar) + { + regMaskTP preferences = currentInterval->registerPreferences; + bool keepAssignment = true; + bool matchesPreferences = (preferences & genRegMask(assignedRegister)) != RBM_NONE; + + // Will the assigned register cover the lifetime? If not, does it at least + // meet the preferences for the next RefPosition? + LsraLocation nextPhysRegLocation = nextFixedRef[assignedRegister]; + if (nextPhysRegLocation <= currentInterval->lastRefPosition->nodeLocation) { - assignedInterval->isActive = false; - if (assignedInterval->getNextRefPosition() == nullptr) + // Check to see if the existing assignment matches the preferences (e.g. callee save registers) + // and ensure that the next use of this localVar does not occur after the nextPhysRegRefPos + // There must be a next RefPosition, because we know that the Interval extends beyond the + // nextPhysRegRefPos. + assert(nextRefPosition != nullptr); + if (!matchesPreferences || nextPhysRegLocation < nextRefPosition->nodeLocation) { - unassignPhysReg(physRegRecord, nullptr); + keepAssignment = false; } - if (!assignedInterval->IsUpperVector()) + else if ((nextRefPosition->registerAssignment != assignedRegBit) && + (nextPhysRegLocation <= nextRefPosition->getRefEndLocation())) { - inVarToRegMap[assignedInterval->getVarIndex(compiler)] = REG_STK; + keepAssignment = false; } } - else + else if (refType == RefTypeParamDef && !matchesPreferences) { - // This interval may still be active, but was in another register in an - // intervening block. - updateAssignedInterval(physRegRecord, nullptr, assignedInterval->registerType); + // Don't use the register, even if available, if it doesn't match the preferences. + // Note that this case is only for ParamDefs, for which we haven't yet taken preferences + // into account (we've just automatically got the initial location). In other cases, + // we would already have put it in a preferenced register, if it was available. + // TODO-CQ: Consider expanding this to check availability - that would duplicate + // code here, but otherwise we may wind up in this register anyway. + keepAssignment = false; } -#ifdef TARGET_ARM - // unassignPhysReg, above, may have restored a 'previousInterval', in which case we need to - // get the value of 'physRegRecord->assignedInterval' rather than using 'assignedInterval'. - if (physRegRecord->assignedInterval != nullptr) - { - assignedInterval = physRegRecord->assignedInterval; - } - if (assignedInterval->registerType == TYP_DOUBLE) + if (keepAssignment == false) { - // Skip next float register, because we already addressed a double register - assert(genIsValidDoubleReg(reg)); - reg = REG_NEXT(reg); - makeRegAvailable(reg, physRegRecord->registerType); + RegRecord* physRegRecord = getRegisterRecord(currentInterval->physReg); + currentRefPosition->registerAssignment = allRegs(currentInterval->registerType); + currentRefPosition->isFixedRegRef = false; + unassignPhysRegNoSpill(physRegRecord); + + // If the preferences are currently set to just this register, reset them to allRegs + // of the appropriate type (just as we just reset the registerAssignment for this + // RefPosition. + // Otherwise, simply remove this register from the preferences, if it's there. + + if (currentInterval->registerPreferences == assignedRegBit) + { + currentInterval->registerPreferences = currentRefPosition->registerAssignment; + } + else + { + currentInterval->registerPreferences &= ~assignedRegBit; + } + + assignedRegister = REG_NA; + assignedRegBit = RBM_NONE; } -#endif // TARGET_ARM } } -#ifdef TARGET_ARM - else - { - Interval* assignedInterval = physRegRecord->assignedInterval; - if (assignedInterval != nullptr && assignedInterval->registerType == TYP_DOUBLE) + if (assignedRegister != REG_NA) + { + RegRecord* physRegRecord = getRegisterRecord(assignedRegister); + assert((assignedRegBit == currentRefPosition->registerAssignment) || + (physRegRecord->assignedInterval == currentInterval) || + !isRegInUse(assignedRegister, currentInterval->registerType)); + if (conflictingFixedRegReference(assignedRegister, currentRefPosition)) { - // Skip next float register, because we already addressed a double register - assert(genIsValidDoubleReg(reg)); - reg = REG_NEXT(reg); + // We may have already reassigned the register to the conflicting reference. + // If not, we need to unassign this interval. + if (physRegRecord->assignedInterval == currentInterval) + { + unassignPhysRegNoSpill(physRegRecord); + physRegRecord->assignedInterval = nullptr; + clearConstantReg(assignedRegister, currentInterval->registerType); + } + currentRefPosition->moveReg = true; + assignedRegister = REG_NA; + currentRefPosition->registerAssignment &= ~assignedRegBit; + setIntervalAsSplit(currentInterval); + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_MOVE_REG, currentInterval, assignedRegister)); } - } -#endif // TARGET_ARM - } -} - -//------------------------------------------------------------------------ -// processBlockEndLocations: Record the variables occupying registers after completing the current block. -// -// Arguments: -// currentBlock - the block we have just completed. -// -// Return Value: -// None -// -// Notes: -// This must be called both during the allocation and resolution (write-back) phases. -// This is because we need to have the outVarToRegMap locations in order to set the locations -// at successor blocks during allocation time, but if lclVars are spilled after a block has been -// completed, we need to record the REG_STK location for those variables at resolution time. - -void LinearScan::processBlockEndLocations(BasicBlock* currentBlock) -{ - assert(currentBlock != nullptr && currentBlock->bbNum == curBBNum); - VarToRegMap outVarToRegMap = getOutVarToRegMap(curBBNum); - - VarSetOps::AssignNoCopy(compiler, currentLiveVars, - VarSetOps::Intersection(compiler, registerCandidateVars, currentBlock->bbLiveOut)); -#ifdef DEBUG - if (getLsraExtendLifeTimes()) - { - VarSetOps::Assign(compiler, currentLiveVars, registerCandidateVars); - } -#endif // DEBUG - VarSetOps::Iter iter(compiler, currentLiveVars); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex)) - { - Interval* interval = getIntervalForLocalVar(varIndex); - if (interval->isActive) - { - assert(interval->physReg != REG_NA && interval->physReg != REG_STK); - setVarReg(outVarToRegMap, varIndex, interval->physReg); - } - else - { - outVarToRegMap[varIndex] = REG_STK; - } -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - // Ensure that we have no partially-spilled large vector locals. - assert(!Compiler::varTypeNeedsPartialCalleeSave(interval->registerType) || !interval->isPartiallySpilled); -#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE - } - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_END_BB)); -} - -#ifdef DEBUG -void LinearScan::dumpRefPositions(const char* str) -{ - printf("------------\n"); - printf("REFPOSITIONS %s: \n", str); - printf("------------\n"); - for (RefPosition& refPos : refPositions) - { - refPos.dump(); - } -} -#endif // DEBUG + else if ((genRegMask(assignedRegister) & currentRefPosition->registerAssignment) != 0) + { + currentRefPosition->registerAssignment = assignedRegBit; + if (!currentInterval->isActive) + { + // If we've got an exposed use at the top of a block, the + // interval might not have been active. Otherwise if it's a use, + // the interval must be active. + if (refType == RefTypeDummyDef) + { + currentInterval->isActive = true; + assert(getRegisterRecord(assignedRegister)->assignedInterval == currentInterval); + } + else + { + currentRefPosition->reload = true; + } + } + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, currentInterval, assignedRegister)); + } + else + { + // It's already in a register, but not one we need. + if (!RefTypeIsDef(currentRefPosition->refType)) + { + regNumber copyReg = assignCopyReg(currentRefPosition); + lastAllocatedRefPosition = currentRefPosition; + bool unassign = false; + if (currentInterval->isWriteThru) + { + if (currentRefPosition->refType == RefTypeDef) + { + currentRefPosition->writeThru = true; + } + if (!currentRefPosition->lastUse) + { + if (currentRefPosition->spillAfter) + { + unassign = true; + } + } + } + regMaskTP copyRegMask = getRegMask(copyReg, currentInterval->registerType); + regMaskTP assignedRegMask = getRegMask(assignedRegister, currentInterval->registerType); + regsInUseThisLocation |= copyRegMask | assignedRegMask; + if (currentRefPosition->lastUse) + { + if (currentRefPosition->delayRegFree) + { + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE_DELAYED, currentInterval, + assignedRegister)); + delayRegsToFree |= copyRegMask | assignedRegMask; + regsInUseNextLocation |= copyRegMask | assignedRegMask; + } + else + { + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE, currentInterval, assignedRegister)); + regsToFree |= copyRegMask | assignedRegMask; + } + } + else + { + copyRegsToFree |= copyRegMask; + if (currentRefPosition->delayRegFree) + { + regsInUseNextLocation |= copyRegMask | assignedRegMask; + } + } -//------------------------------------------------------------------------ -// LinearScan::makeRegisterInactive: Make the interval currently assigned to -// a register inactive. -// -// Arguments: -// physRegRecord - the RegRecord for the register -// -// Return Value: -// None. -// -// Notes: -// It may be that the RegRecord has already been freed, e.g. due to a kill, -// or it may be that the register was a copyReg, so is not the assigned register -// of the Interval currently occupying the register, in which case this method has no effect. -// -void LinearScan::makeRegisterInactive(RegRecord* physRegRecord) -{ - Interval* assignedInterval = physRegRecord->assignedInterval; - // It may have already been freed by a "Kill" - if ((assignedInterval != nullptr) && (assignedInterval->physReg == physRegRecord->regNum)) - { - assignedInterval->isActive = false; - if (assignedInterval->isConstant) - { - clearNextIntervalRef(physRegRecord->regNum, assignedInterval->registerType); + // If this is a tree temp (non-localVar) interval, we will need an explicit move. + // Note: In theory a moveReg should cause the Interval to now have the new reg as its + // assigned register. However, that's not currently how this works. + // If we ever actually move lclVar intervals instead of copying, this will need to change. + if (!currentInterval->isLocalVar) + { + currentRefPosition->moveReg = true; + currentRefPosition->copyReg = false; + } + clearNextIntervalRef(copyReg, currentInterval->registerType); + clearSpillCost(copyReg, currentInterval->registerType); + updateNextIntervalRef(assignedRegister, currentInterval); + updateSpillCost(assignedRegister, currentInterval); + continue; + } + else + { + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NEEDS_NEW_REG, nullptr, assignedRegister)); + regsToFree |= getRegMask(assignedRegister, currentInterval->registerType); + // We want a new register, but we don't want this to be considered a spill. + assignedRegister = REG_NA; + if (physRegRecord->assignedInterval == currentInterval) + { + unassignPhysRegNoSpill(physRegRecord); + } + } + } } - } -} -//------------------------------------------------------------------------ -// LinearScan::freeRegister: Make a register available for use -// -// Arguments: -// physRegRecord - the RegRecord for the register to be freed. -// -// Return Value: -// None. -// -// Assumptions: -// None. -// It may be that the RegRecord has already been freed, e.g. due to a kill, -// in which case this method has no effect. -// -// Notes: -// If there is currently an Interval assigned to this register, and it has -// more references (i.e. this is a local last-use, but more uses and/or -// defs remain), it will remain assigned to the physRegRecord. However, since -// it is marked inactive, the register will be available, albeit less desirable -// to allocate. -// -void LinearScan::freeRegister(RegRecord* physRegRecord) -{ - Interval* assignedInterval = physRegRecord->assignedInterval; - makeRegAvailable(physRegRecord->regNum, physRegRecord->registerType); - clearSpillCost(physRegRecord->regNum, physRegRecord->registerType); - makeRegisterInactive(physRegRecord); - - if (assignedInterval != nullptr) - { - // TODO: Under the following conditions we should be just putting it in regsToMakeInactive - // not regsToFree. - // - // We don't unassign in the following conditions: - // - If this is a constant node, that we may encounter again, OR - // - If its recent RefPosition is not a last-use and its next RefPosition is non-null. - // - If there are no more RefPositions, or the next - // one is a def. Note that the latter condition doesn't actually ensure that - // there aren't subsequent uses that could be reached by a value in the assigned - // register, but is merely a heuristic to avoid tying up the register (or using - // it when it's non-optimal). A better alternative would be to use SSA, so that - // we wouldn't unnecessarily link separate live ranges to the same register. - // - RefPosition* nextRefPosition = assignedInterval->getNextRefPosition(); - if (!assignedInterval->isConstant && (nextRefPosition == nullptr || RefTypeIsDef(nextRefPosition->refType))) + if (assignedRegister == REG_NA) { -#ifdef TARGET_ARM - assert((assignedInterval->registerType != TYP_DOUBLE) || genIsValidDoubleReg(physRegRecord->regNum)); -#endif // TARGET_ARM - unassignPhysReg(physRegRecord, nullptr); - } - } -} - -//------------------------------------------------------------------------ -// LinearScan::freeRegisters: Free the registers in 'regsToFree' -// -// Arguments: -// regsToFree - the mask of registers to free -// -void LinearScan::freeRegisters(regMaskTP regsToFree) -{ - if (regsToFree == RBM_NONE) - { - return; - } + if (currentRefPosition->RegOptional()) + { + // We can avoid allocating a register if it is a last use requiring a reload. + if (currentRefPosition->lastUse && currentRefPosition->reload) + { + allocate = false; + } + else if (currentInterval->isWriteThru) + { + // Don't allocate if the next reference is in a cold block. + if (nextRefPosition == nullptr || (nextRefPosition->nodeLocation >= firstColdLoc)) + { + allocate = false; + } + } - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_FREE_REGS)); - makeRegsAvailable(regsToFree); - while (regsToFree != RBM_NONE) - { - regMaskTP nextRegBit = genFindLowestBit(regsToFree); - regsToFree &= ~nextRegBit; - regNumber nextReg = genRegNumFromMask(nextRegBit); - RegRecord* regRecord = getRegisterRecord(nextReg); -#ifdef TARGET_ARM - if (regRecord->assignedInterval != nullptr && (regRecord->assignedInterval->registerType == TYP_DOUBLE)) - { - assert(genIsValidDoubleReg(nextReg)); - regsToFree &= ~(nextRegBit << 1); - } +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE && defined(TARGET_XARCH) + // We can also avoid allocating a register (in fact we don't want to) if we have + // an UpperVectorRestore on xarch where the value is on the stack. + if ((currentRefPosition->refType == RefTypeUpperVectorRestore) && (currentInterval->physReg == REG_NA)) + { + assert(currentRefPosition->regOptional); + allocate = false; + } #endif - freeRegister(regRecord); - } -} -//------------------------------------------------------------------------ -// LinearScan::allocateRegisters: Perform the actual register allocation by iterating over -// all of the previously constructed Intervals -// -void LinearScan::allocateRegisters() -{ - JITDUMP("*************** In LinearScan::allocateRegisters()\n"); - DBEXEC(VERBOSE, lsraDumpIntervals("before allocateRegisters")); +#ifdef DEBUG + // Under stress mode, don't allocate registers to RegOptional RefPositions. + if (allocate && regOptionalNoAlloc()) + { + allocate = false; + } +#endif + } - // at start, nothing is active except for register args - for (Interval& interval : intervals) - { - Interval* currentInterval = &interval; - currentInterval->recentRefPosition = nullptr; - currentInterval->isActive = false; - if (currentInterval->isLocalVar) - { - LclVarDsc* varDsc = currentInterval->getLocalVar(compiler); - if (varDsc->lvIsRegArg && currentInterval->firstRefPosition != nullptr) + RegisterScore registerScore = NONE; + if (allocate) { - currentInterval->isActive = true; + // Allocate a register, if we must, or if it is profitable to do so. + // If we have a fixed reg requirement, and the interval is inactive in another register, + // unassign that register. + if (currentRefPosition->isFixedRegRef && !currentInterval->isActive && + (currentInterval->assignedReg != nullptr) && + (currentInterval->assignedReg->assignedInterval == currentInterval) && + (genRegMask(currentInterval->assignedReg->regNum) != currentRefPosition->registerAssignment)) + { + unassignPhysReg(currentInterval->assignedReg, nullptr); + } + assignedRegister = allocateReg(currentInterval, currentRefPosition DEBUG_ARG(®isterScore)); } - } - } - - if (enregisterLocalVars) - { -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - VarSetOps::Iter largeVectorVarsIter(compiler, largeVectorVars); - unsigned largeVectorVarIndex = 0; - while (largeVectorVarsIter.NextElem(&largeVectorVarIndex)) - { - Interval* lclVarInterval = getIntervalForLocalVar(largeVectorVarIndex); - lclVarInterval->isPartiallySpilled = false; - } -#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE - } - - resetRegState(); - for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) - { - RegRecord* physRegRecord = getRegisterRecord(reg); - physRegRecord->recentRefPosition = nullptr; - updateNextFixedRef(physRegRecord, physRegRecord->firstRefPosition); - // Is this an incoming arg register? (Note that we don't, currently, consider reassigning - // an incoming arg register as having spill cost.) - Interval* interval = physRegRecord->assignedInterval; - if (interval != nullptr) - { -#ifdef TARGET_ARM - if ((interval->registerType != TYP_DOUBLE) || genIsValidDoubleReg(reg)) -#endif // TARGET_ARM + // If no register was found, this RefPosition must not require a register. + if (assignedRegister == REG_NA) { - updateNextIntervalRef(reg, interval); - updateSpillCost(reg, interval); - setRegInUse(reg, interval->registerType); - INDEBUG(registersToDump |= getRegMask(reg, interval->registerType)); + assert(currentRefPosition->RegOptional()); + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, currentInterval)); + currentRefPosition->registerAssignment = RBM_NONE; + currentRefPosition->reload = false; + currentInterval->isActive = false; + setIntervalAsSpilled(currentInterval); } - } - else - { - clearNextIntervalRef(reg, physRegRecord->registerType); - clearSpillCost(reg, physRegRecord->registerType); - } - } - #ifdef DEBUG - if (VERBOSE) - { - dumpRefPositions("BEFORE ALLOCATION"); - dumpVarRefPositions("BEFORE ALLOCATION"); - - printf("\n\nAllocating Registers\n" - "--------------------\n"); - // Start with a small set of commonly used registers, so that we don't keep having to print a new title. - // Include all the arg regs, as they may already have values assigned to them. - registersToDump = LsraLimitSmallIntSet | LsraLimitSmallFPSet | RBM_ARG_REGS; - dumpRegRecordHeader(); - // Now print an empty "RefPosition", since we complete the dump of the regs at the beginning of the loop. - printf(indentFormat, ""); - } + else + { + if (VERBOSE) + { + if (currentInterval->isConstant && (currentRefPosition->treeNode != nullptr) && + currentRefPosition->treeNode->IsReuseRegVal()) + { + dumpLsraAllocationEvent(LSRA_EVENT_REUSE_REG, currentInterval, assignedRegister, currentBlock, + registerScore); + } + else + { + dumpLsraAllocationEvent(LSRA_EVENT_ALLOC_REG, currentInterval, assignedRegister, currentBlock, + registerScore); + } + } + } #endif // DEBUG - BasicBlock* currentBlock = nullptr; - - LsraLocation prevLocation = MinLocation; - regMaskTP regsToFree = RBM_NONE; - regMaskTP delayRegsToFree = RBM_NONE; - regMaskTP regsToMakeInactive = RBM_NONE; - regMaskTP delayRegsToMakeInactive = RBM_NONE; - regMaskTP copyRegsToFree = RBM_NONE; - regsInUseThisLocation = RBM_NONE; - regsInUseNextLocation = RBM_NONE; - - // This is the most recent RefPosition for which a register was allocated - // - currently only used for DEBUG but maintained in non-debug, for clarity of code - // (and will be optimized away because in non-debug spillAlways() unconditionally returns false) - RefPosition* lastAllocatedRefPosition = nullptr; - - bool handledBlockEnd = false; - - for (RefPosition& refPositionIterator : refPositions) - { - RefPosition* currentRefPosition = &refPositionIterator; - RefPosition* nextRefPosition = currentRefPosition->nextRefPosition; + if (refType == RefTypeDummyDef && assignedRegister != REG_NA) + { + setInVarRegForBB(curBBNum, currentInterval->varNum, assignedRegister); + } - // TODO: Can we combine this with the freeing of registers below? It might - // mess with the dump, since this was previously being done before the call below - // to dumpRegRecords. - regMaskTP tempRegsToMakeInactive = (regsToMakeInactive | delayRegsToMakeInactive); - while (tempRegsToMakeInactive != RBM_NONE) - { - regMaskTP nextRegBit = genFindLowestBit(tempRegsToMakeInactive); - tempRegsToMakeInactive &= ~nextRegBit; - regNumber nextReg = genRegNumFromMask(nextRegBit); - RegRecord* regRecord = getRegisterRecord(nextReg); - clearSpillCost(regRecord->regNum, regRecord->registerType); - makeRegisterInactive(regRecord); - } - if (currentRefPosition->nodeLocation > prevLocation) - { - makeRegsAvailable(regsToMakeInactive); - // TODO: Clean this up. We need to make the delayRegs inactive as well, but don't want - // to mark them as free yet. - regsToMakeInactive |= delayRegsToMakeInactive; - regsToMakeInactive = delayRegsToMakeInactive; - delayRegsToMakeInactive = RBM_NONE; + // If we allocated a register, and this is a use of a spilled value, + // it should have been marked for reload above. + if (assignedRegister != REG_NA && RefTypeIsUse(refType) && !isInRegister) + { + assert(currentRefPosition->reload); + } } -#ifdef DEBUG - // Set the activeRefPosition to null until we're done with any boundary handling. - activeRefPosition = nullptr; - if (VERBOSE) + // If we allocated a register, record it + if (assignedRegister != REG_NA) { - // We're really dumping the RegRecords "after" the previous RefPosition, but it's more convenient - // to do this here, since there are a number of "continue"s in this loop. - dumpRegRecords(); - } -#endif // DEBUG - - // This is the previousRefPosition of the current Referent, if any - RefPosition* previousRefPosition = nullptr; + assignedRegBit = genRegMask(assignedRegister); + regMaskTP regMask = getRegMask(assignedRegister, currentInterval->registerType); + regsInUseThisLocation |= regMask; + if (currentRefPosition->delayRegFree) + { + regsInUseNextLocation |= regMask; + } + currentRefPosition->registerAssignment = assignedRegBit; - Interval* currentInterval = nullptr; - Referenceable* currentReferent = nullptr; - RefType refType = currentRefPosition->refType; + currentInterval->physReg = assignedRegister; + regsToFree &= ~regMask; // we'll set it again later if it's dead - currentReferent = currentRefPosition->referent; + // If this interval is dead, free the register. + // The interval could be dead if this is a user variable, or if the + // node is being evaluated for side effects, or a call whose result + // is not used, etc. + // If this is an UpperVector we'll neither free it nor preference it + // (it will be freed when it is used). + bool unassign = false; + if (!currentInterval->IsUpperVector()) + { + if (currentInterval->isWriteThru) + { + if (currentRefPosition->refType == RefTypeDef) + { + currentRefPosition->writeThru = true; + } + if (!currentRefPosition->lastUse) + { + if (currentRefPosition->spillAfter) + { + unassign = true; + } + } + } + if (currentRefPosition->lastUse || currentRefPosition->nextRefPosition == nullptr) + { + assert(currentRefPosition->isIntervalRef()); + // If this isn't a final use, we'll mark the register as available, but keep the association. + if ((refType != RefTypeExpUse) && (currentRefPosition->nextRefPosition == nullptr)) + { + unassign = true; + } + else + { + if (currentRefPosition->delayRegFree) + { + delayRegsToMakeInactive |= regMask; + } + else + { + regsToMakeInactive |= regMask; + } + // TODO-Cleanup: this makes things consistent with previous, and will enable preferences + // to be propagated, but it seems less than ideal. + currentInterval->isActive = false; + } + // Update the register preferences for the relatedInterval, if this is 'preferencedToDef'. + // Don't propagate to subsequent relatedIntervals; that will happen as they are allocated, and we + // don't know yet whether the register will be retained. + if (currentInterval->relatedInterval != nullptr) + { + currentInterval->relatedInterval->updateRegisterPreferences(assignedRegBit); + } + } - if (spillAlways() && lastAllocatedRefPosition != nullptr && !lastAllocatedRefPosition->IsPhysRegRef() && - !lastAllocatedRefPosition->getInterval()->isInternal && - (RefTypeIsDef(lastAllocatedRefPosition->refType) || lastAllocatedRefPosition->getInterval()->isLocalVar)) - { - assert(lastAllocatedRefPosition->registerAssignment != RBM_NONE); - RegRecord* regRecord = lastAllocatedRefPosition->getInterval()->assignedReg; - unassignPhysReg(regRecord, lastAllocatedRefPosition); - // Now set lastAllocatedRefPosition to null, so that we don't try to spill it again - lastAllocatedRefPosition = nullptr; - } + if (unassign) + { + if (currentRefPosition->delayRegFree) + { + delayRegsToFree |= regMask; - // We wait to free any registers until we've completed all the - // uses for the current node. - // This avoids reusing registers too soon. - // We free before the last true def (after all the uses & internal - // registers), and then again at the beginning of the next node. - // This is made easier by assigning two LsraLocations per node - one - // for all the uses, internal registers & all but the last def, and - // another for the final def (if any). + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE_DELAYED)); + } + else + { + regsToFree |= regMask; - LsraLocation currentLocation = currentRefPosition->nodeLocation; + INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE)); + } + } + } + if (!unassign) + { + updateNextIntervalRef(assignedRegister, currentInterval); + updateSpillCost(assignedRegister, currentInterval); + } + } + lastAllocatedRefPosition = currentRefPosition; + } - // Free at a new location. - if (currentLocation > prevLocation) +#ifdef JIT32_GCENCODER + // For the JIT32_GCENCODER, when lvaKeepAliveAndReportThis is true, we must either keep the "this" pointer + // in the same register for the entire method, or keep it on the stack. Rather than imposing this constraint + // as we allocate, we will force all refs to the stack if it is split or spilled. + if (enregisterLocalVars && compiler->lvaKeepAliveAndReportThis()) + { + LclVarDsc* thisVarDsc = compiler->lvaGetDesc(compiler->info.compThisArg); + if (thisVarDsc->lvLRACandidate) { - // CopyRegs are simply made available - we don't want to make the associated interval inactive. - makeRegsAvailable(copyRegsToFree); - copyRegsToFree = RBM_NONE; - regsInUseThisLocation = regsInUseNextLocation; - regsInUseNextLocation = RBM_NONE; - if ((regsToFree | delayRegsToFree) != RBM_NONE) + Interval* interval = getIntervalForLocalVar(thisVarDsc->lvVarIndex); + if (interval->isSplit) { - freeRegisters(regsToFree); - if ((currentLocation > (prevLocation + 1)) && (delayRegsToFree != RBM_NONE)) - { - // We should never see a delayReg that is delayed until a Location that has no RefPosition - // (that would be the RefPosition that it was supposed to interfere with). - assert(!"Found a delayRegFree associated with Location with no reference"); - // However, to be cautious for the Release build case, we will free them. - freeRegisters(delayRegsToFree); - delayRegsToFree = RBM_NONE; - regsInUseThisLocation = RBM_NONE; - } - regsToFree = delayRegsToFree; - delayRegsToFree = RBM_NONE; - -#ifdef DEBUG - // Validate the current state just after we've freed the registers. This ensures that any pending - // freed registers will have had their state updated to reflect the intervals they were holding. - for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) + // We'll have to spill this. + setIntervalAsSpilled(interval); + } + if (interval->isSpilled) + { + unsigned prevBBNum = 0; + for (RefPosition* ref = interval->firstRefPosition; ref != nullptr; ref = ref->nextRefPosition) { - regMaskTP regMask = genRegMask(reg); - // If this isn't available or if it's still waiting to be freed (i.e. it was in - // delayRegsToFree and so now it's in regsToFree), then skip it. - if ((regMask & (availableIntRegs | availableFloatRegs) & ~regsToFree) == RBM_NONE) + // For the resolution phase, we need to ensure that any block with exposed uses has the + // incoming reg for 'this' as REG_STK. + if (RefTypeIsUse(ref->refType) && (ref->bbNum != prevBBNum)) { - continue; + VarToRegMap inVarToRegMap = getInVarToRegMap(ref->bbNum); + setVarReg(inVarToRegMap, thisVarDsc->lvVarIndex, REG_STK); } - RegRecord* physRegRecord = getRegisterRecord(reg); - Interval* assignedInterval = physRegRecord->assignedInterval; - if (assignedInterval != nullptr) + if (ref->RegOptional()) { - bool isAssignedReg = (assignedInterval->physReg == reg); - RefPosition* recentRefPosition = assignedInterval->recentRefPosition; - // If we have a copyReg or a moveReg, we might have assigned this register to an Interval, - // but that isn't considered its assignedReg. - if (recentRefPosition != nullptr) - { - if (recentRefPosition->refType == RefTypeExpUse) - { - // We don't update anything on these, as they're just placeholders to extend the - // lifetime. - continue; - } - // For copyReg or moveReg, we don't have anything further to assert. - if (recentRefPosition->copyReg || recentRefPosition->moveReg) - { - continue; - } - assert(assignedInterval->isConstant == isRegConstant(reg, assignedInterval->registerType)); - if (assignedInterval->isActive) + ref->registerAssignment = RBM_NONE; + ref->reload = false; + ref->spillAfter = false; + } + switch (ref->refType) + { + case RefTypeDef: + if (ref->registerAssignment != RBM_NONE) { - // If this is not the register most recently allocated, it must be from a copyReg, - // or it was placed there by the inVarToRegMap. In either case it must be a lclVar. - - if (!isAssignedToInterval(assignedInterval, physRegRecord)) - { - assert(assignedInterval->isLocalVar); - // We'd like to assert that this was either set by the inVarToRegMap, or by - // a copyReg, but we can't traverse backward to check for a copyReg, because - // we only have recentRefPosition, and there may be a previous RefPosition - // at the same Location with a copyReg. - } - if (isAssignedReg) - { - assert(nextIntervalRef[reg] == assignedInterval->getNextRefLocation()); - assert(!isRegAvailable(reg, assignedInterval->registerType)); - assert((recentRefPosition == nullptr) || - (spillCost[reg] == getSpillWeight(physRegRecord))); - } - else - { - assert((nextIntervalRef[reg] == MaxLocation) || - isRegBusy(reg, assignedInterval->registerType)); - } + ref->spillAfter = true; } - else + break; + case RefTypeUse: + if (ref->registerAssignment != RBM_NONE) { - if ((assignedInterval->physReg == reg) && !assignedInterval->isConstant) - { - assert(nextIntervalRef[reg] == assignedInterval->getNextRefLocation()); - } - else - { - assert(nextIntervalRef[reg] == MaxLocation); - assert(isRegAvailable(reg, assignedInterval->registerType)); - assert(spillCost[reg] == 0); - } + ref->reload = true; + ref->spillAfter = true; + ref->copyReg = false; + ref->moveReg = false; } - } - } - else - { - assert(isRegAvailable(reg, physRegRecord->registerType)); - assert(!isRegConstant(reg, physRegRecord->registerType)); - assert(nextIntervalRef[reg] == MaxLocation); - assert(spillCost[reg] == 0); - } - LsraLocation thisNextFixedRef = physRegRecord->getNextRefLocation(); - assert(nextFixedRef[reg] == thisNextFixedRef); -#ifdef TARGET_ARM - // If this is occupied by a double interval, skip the corresponding float reg. - if ((assignedInterval != nullptr) && (assignedInterval->registerType == TYP_DOUBLE)) - { - reg = REG_NEXT(reg); + break; + default: + break; } -#endif + prevBBNum = ref->bbNum; } -#endif // DEBUG } } - prevLocation = currentLocation; + } +#endif // JIT32_GCENCODER - // get previous refposition, then current refpos is the new previous - if (currentReferent != nullptr) - { - previousRefPosition = currentReferent->recentRefPosition; - currentReferent->recentRefPosition = currentRefPosition; - } - else + // Free registers to clear associated intervals for resolution phase + CLANG_FORMAT_COMMENT_ANCHOR; + +#ifdef DEBUG + if (getLsraExtendLifeTimes()) + { + // If we have extended lifetimes, we need to make sure all the registers are freed. + for (size_t regNumIndex = 0; regNumIndex <= REG_FP_LAST; regNumIndex++) { - assert((refType == RefTypeBB) || (refType == RefTypeKillGCRefs)); + RegRecord& regRecord = physRegs[regNumIndex]; + Interval* interval = regRecord.assignedInterval; + if (interval != nullptr) + { + interval->isActive = false; + unassignPhysReg(®Record, nullptr); + } } + } + else +#endif // DEBUG + { + freeRegisters(regsToFree | delayRegsToFree); + } #ifdef DEBUG - activeRefPosition = currentRefPosition; + if (VERBOSE) + { + // Dump the RegRecords after the last RefPosition is handled. + dumpRegRecords(); + printf("\n"); - // For the purposes of register resolution, we handle the DummyDefs before - // the block boundary - so the RefTypeBB is after all the DummyDefs. - // However, for the purposes of allocation, we want to handle the block - // boundary first, so that we can free any registers occupied by lclVars - // that aren't live in the next block and make them available for the - // DummyDefs. + dumpRefPositions("AFTER ALLOCATION"); + dumpVarRefPositions("AFTER ALLOCATION"); - // If we've already handled the BlockEnd, but now we're seeing the RefTypeBB, - // dump it now. - if ((refType == RefTypeBB) && handledBlockEnd) - { - dumpNewBlock(currentBlock, currentRefPosition->nodeLocation); - } -#endif // DEBUG + // Dump the intervals that remain active + printf("Active intervals at end of allocation:\n"); - if (!handledBlockEnd && (refType == RefTypeBB || refType == RefTypeDummyDef)) + // We COULD just reuse the intervalIter from above, but ArrayListIterator doesn't + // provide a Reset function (!) - we'll probably replace this so don't bother + // adding it + + for (Interval& interval : intervals) { - // Free any delayed regs (now in regsToFree) before processing the block boundary - freeRegisters(regsToFree); - regsToFree = RBM_NONE; - regsInUseThisLocation = RBM_NONE; - regsInUseNextLocation = RBM_NONE; - handledBlockEnd = true; - curBBStartLocation = currentRefPosition->nodeLocation; - if (currentBlock == nullptr) - { - currentBlock = startBlockSequence(); - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_START_BB, nullptr, REG_NA, compiler->fgFirstBB)); - } - else + if (interval.isActive) { - processBlockEndAllocation(currentBlock); - currentBlock = moveToNextBlock(); - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_START_BB, nullptr, REG_NA, currentBlock)); + printf("Active "); + interval.dump(); } } - if (refType == RefTypeBB) - { - handledBlockEnd = false; - continue; - } + printf("\n"); + } +#endif // DEBUG +} - if (refType == RefTypeKillGCRefs) +//----------------------------------------------------------------------------- +// updateAssignedInterval: Update assigned interval of register. +// +// Arguments: +// reg - register to be updated +// interval - interval to be assigned +// regType - register type +// +// Return Value: +// None +// +// Assumptions: +// For ARM32, when "regType" is TYP_DOUBLE, "reg" should be a even-numbered +// float register, i.e. lower half of double register. +// +// Note: +// For ARM32, two float registers consisting a double register are updated +// together when "regType" is TYP_DOUBLE. +// +void LinearScan::updateAssignedInterval(RegRecord* reg, Interval* interval, RegisterType regType) +{ +#ifdef TARGET_ARM + // Update overlapping floating point register for TYP_DOUBLE. + Interval* oldAssignedInterval = reg->assignedInterval; + regNumber doubleReg = REG_NA; + if (regType == TYP_DOUBLE) + { + doubleReg = reg->regNum; + RegRecord* anotherHalfReg = getSecondHalfRegRec(reg); + anotherHalfReg->assignedInterval = interval; + } + else if ((oldAssignedInterval != nullptr) && (oldAssignedInterval->registerType == TYP_DOUBLE)) + { + RegRecord* anotherHalfReg = findAnotherHalfRegRec(reg); + doubleReg = genIsValidDoubleReg(reg->regNum) ? reg->regNum : anotherHalfReg->regNum; + anotherHalfReg->assignedInterval = nullptr; + } + if (doubleReg != REG_NA) + { + clearNextIntervalRef(doubleReg, TYP_DOUBLE); + clearSpillCost(doubleReg, TYP_DOUBLE); + clearConstantReg(doubleReg, TYP_DOUBLE); + } +#endif + reg->assignedInterval = interval; + if (interval != nullptr) + { + setRegInUse(reg->regNum, interval->registerType); + if (interval->isConstant) { - spillGCRefs(currentRefPosition); - continue; + setConstantReg(reg->regNum, interval->registerType); } + else + { + clearConstantReg(reg->regNum, interval->registerType); + } + updateNextIntervalRef(reg->regNum, interval); + updateSpillCost(reg->regNum, interval); + } + else + { + clearNextIntervalRef(reg->regNum, reg->registerType); + clearSpillCost(reg->regNum, reg->registerType); + } +} - if (currentRefPosition->isPhysRegRef) - { - RegRecord* regRecord = currentRefPosition->getReg(); - Interval* assignedInterval = regRecord->assignedInterval; - - updateNextFixedRef(regRecord, currentRefPosition->nextRefPosition); - - // If this is a FixedReg, disassociate any inactive constant interval from this register. - // Otherwise, do nothing. - if (refType == RefTypeFixedReg) - { - if (assignedInterval != nullptr && !assignedInterval->isActive && assignedInterval->isConstant) - { - clearConstantReg(regRecord->regNum, assignedInterval->registerType); - regRecord->assignedInterval = nullptr; - spillCost[regRecord->regNum] = 0; +//----------------------------------------------------------------------------- +// updatePreviousInterval: Update previous interval of register. +// +// Arguments: +// reg - register to be updated +// interval - interval to be assigned +// regType - register type +// +// Return Value: +// None +// +// Assumptions: +// For ARM32, when "regType" is TYP_DOUBLE, "reg" should be a even-numbered +// float register, i.e. lower half of double register. +// +// Note: +// For ARM32, two float registers consisting a double register are updated +// together when "regType" is TYP_DOUBLE. +// +void LinearScan::updatePreviousInterval(RegRecord* reg, Interval* interval, RegisterType regType) +{ + reg->previousInterval = interval; #ifdef TARGET_ARM - // Update overlapping floating point register for TYP_DOUBLE - if (assignedInterval->registerType == TYP_DOUBLE) - { - RegRecord* otherRegRecord = findAnotherHalfRegRec(regRecord); - assert(otherRegRecord->assignedInterval == assignedInterval); - otherRegRecord->assignedInterval = nullptr; - spillCost[otherRegRecord->regNum] = 0; - } -#endif // TARGET_ARM - } - regsInUseThisLocation |= currentRefPosition->registerAssignment; - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_FIXED_REG, nullptr, currentRefPosition->assignedReg())); - continue; - } - if (refType == RefTypeKill) - { - if (assignedInterval != nullptr) - { - unassignPhysReg(regRecord, assignedInterval->recentRefPosition); - clearConstantReg(regRecord->regNum, assignedInterval->registerType); - makeRegAvailable(regRecord->regNum, assignedInterval->registerType); - } - clearRegBusyUntilKill(regRecord->regNum); - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum)); - continue; - } - } + // Update overlapping floating point register for TYP_DOUBLE + if (regType == TYP_DOUBLE) + { + RegRecord* anotherHalfReg = findAnotherHalfRegRec(reg); - // If this is an exposed use, do nothing - this is merely a placeholder to attempt to - // ensure that a register is allocated for the full lifetime. The resolution logic - // will take care of moving to the appropriate register if needed. + anotherHalfReg->previousInterval = interval; + } +#endif +} - if (refType == RefTypeExpUse) - { - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_EXP_USE)); - currentInterval = currentRefPosition->getInterval(); - if (currentInterval->physReg != REG_NA) - { - updateNextIntervalRef(currentInterval->physReg, currentInterval); - } - continue; - } +//----------------------------------------------------------------------------- +// writeLocalReg: Write the register assignment for a GT_LCL_VAR node. +// +// Arguments: +// lclNode - The GT_LCL_VAR node +// varNum - The variable number for the register +// reg - The assigned register +// +// Return Value: +// None +// +// Note: +// For a multireg node, 'varNum' will be the field local for the given register. +// +void LinearScan::writeLocalReg(GenTreeLclVar* lclNode, unsigned varNum, regNumber reg) +{ + assert((lclNode->GetLclNum() == varNum) == !lclNode->IsMultiReg()); + if (lclNode->GetLclNum() == varNum) + { + lclNode->SetRegNum(reg); + } + else + { + assert(compiler->lvaEnregMultiRegVars); + LclVarDsc* parentVarDsc = compiler->lvaGetDesc(lclNode->GetLclNum()); + assert(parentVarDsc->lvPromoted); + unsigned regIndex = varNum - parentVarDsc->lvFieldLclStart; + assert(regIndex < MAX_MULTIREG_COUNT); + lclNode->SetRegNumByIdx(reg, regIndex); + } +} - regNumber assignedRegister = REG_NA; +//----------------------------------------------------------------------------- +// LinearScan::resolveLocalRef +// Description: +// Update the graph for a local reference. +// Also, track the register (if any) that is currently occupied. +// Arguments: +// treeNode: The lclVar that's being resolved +// currentRefPosition: the RefPosition associated with the treeNode +// +// Details: +// This method is called for each local reference, during the resolveRegisters +// phase of LSRA. It is responsible for keeping the following in sync: +// - varDsc->GetRegNum() (and GetOtherReg()) contain the unique register location. +// If it is not in the same register through its lifetime, it is set to REG_STK. +// - interval->physReg is set to the assigned register +// (i.e. at the code location which is currently being handled by resolveRegisters()) +// - interval->isActive is true iff the interval is live and occupying a register +// - interval->isSpilled should have already been set to true if the interval is EVER spilled +// - interval->isSplit is set to true if the interval does not occupy the same +// register throughout the method +// - RegRecord->assignedInterval points to the interval which currently occupies +// the register +// - For each lclVar node: +// - GetRegNum()/gtRegPair is set to the currently allocated register(s). +// - GTF_SPILLED is set on a use if it must be reloaded prior to use. +// - GTF_SPILL is set if it must be spilled after use. +// +// A copyReg is an ugly case where the variable must be in a specific (fixed) register, +// but it currently resides elsewhere. The register allocator must track the use of the +// fixed register, but it marks the lclVar node with the register it currently lives in +// and the code generator does the necessary move. +// +// Before beginning, the varDsc for each parameter must be set to its initial location. +// +// NICE: Consider tracking whether an Interval is always in the same location (register/stack) +// in which case it will require no resolution. +// +void LinearScan::resolveLocalRef(BasicBlock* block, GenTreeLclVar* treeNode, RefPosition* currentRefPosition) +{ + assert((block == nullptr) == (treeNode == nullptr)); + assert(enregisterLocalVars); - assert(currentRefPosition->isIntervalRef()); - currentInterval = currentRefPosition->getInterval(); - assert(currentInterval != nullptr); - assignedRegister = currentInterval->physReg; + // Is this a tracked local? Or just a register allocated for loading + // a non-tracked one? + Interval* interval = currentRefPosition->getInterval(); + assert(interval->isLocalVar); - // Identify the special cases where we decide up-front not to allocate - bool allocate = true; - bool didDump = false; + interval->recentRefPosition = currentRefPosition; + LclVarDsc* varDsc = interval->getLocalVar(compiler); - if (refType == RefTypeParamDef || refType == RefTypeZeroInit) - { - if (nextRefPosition == nullptr) - { - // If it has no actual references, mark it as "lastUse"; since they're not actually part - // of any flow they won't have been marked during dataflow. Otherwise, if we allocate a - // register we won't unassign it. - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_ZERO_REF, currentInterval)); - currentRefPosition->lastUse = true; - } - LclVarDsc* varDsc = currentInterval->getLocalVar(compiler); - assert(varDsc != nullptr); - assert(!blockInfo[compiler->fgFirstBB->bbNum].hasEHBoundaryIn || currentInterval->isWriteThru); - if (blockInfo[compiler->fgFirstBB->bbNum].hasEHBoundaryIn || - blockInfo[compiler->fgFirstBB->bbNum].hasEHPred) - { - allocate = false; - } - else if (refType == RefTypeParamDef && (varDsc->lvRefCntWtd() <= BB_UNITY_WEIGHT) && - (!currentRefPosition->lastUse || (currentInterval->physReg == REG_STK))) - { - // If this is a low ref-count parameter, and either it is used (def is not the last use) or it's - // passed on the stack, don't allocate a register. - // Note that if this is an unused register parameter we don't want to set allocate to false because that - // will cause us to allocate stack space to spill it. - allocate = false; - } - else if ((currentInterval->physReg == REG_STK) && nextRefPosition->treeNode->OperIs(GT_BITCAST)) - { - // In the case of ABI mismatches, avoid allocating a register only to have to immediately move - // it to a different register file. - allocate = false; - } - if (!allocate) - { - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_ENTRY_REG_ALLOCATED, currentInterval)); - didDump = true; - setIntervalAsSpilled(currentInterval); - if (assignedRegister != REG_NA) - { - clearNextIntervalRef(assignedRegister, currentInterval->registerType); - clearSpillCost(assignedRegister, currentInterval->registerType); - makeRegAvailable(assignedRegister, currentInterval->registerType); - } - } - } -#ifdef FEATURE_SIMD -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - else if (currentInterval->isUpperVector) - { - // This is a save or restore of the upper half of a large vector lclVar. - Interval* lclVarInterval = currentInterval->relatedInterval; - assert(lclVarInterval->isLocalVar); - if (refType == RefTypeUpperVectorSave) - { - if ((lclVarInterval->physReg == REG_NA) || - (lclVarInterval->isPartiallySpilled && (currentInterval->physReg == REG_STK))) - { - allocate = false; - } - else - { - lclVarInterval->isPartiallySpilled = true; - } - } - else if (refType == RefTypeUpperVectorRestore) - { - assert(currentInterval->isUpperVector); - if (lclVarInterval->isPartiallySpilled) - { - lclVarInterval->isPartiallySpilled = false; - } - else - { - allocate = false; - } - } + // NOTE: we set the LastUse flag here unless we are extending lifetimes, in which case we write + // this bit in checkLastUses. This is a bit of a hack, but is necessary because codegen requires + // accurate last use info that is not reflected in the lastUse bit on ref positions when we are extending + // lifetimes. See also the comments in checkLastUses. + if ((treeNode != nullptr) && !extendLifetimes()) + { + if (currentRefPosition->lastUse) + { + treeNode->SetLastUse(currentRefPosition->getMultiRegIdx()); } - else if (refType == RefTypeUpperVectorSave) + else { - assert(!currentInterval->isLocalVar); - // Note that this case looks a lot like the case below, but in this case we need to spill - // at the previous RefPosition. - // We may want to consider allocating two callee-save registers for this case, but it happens rarely - // enough that it may not warrant the additional complexity. - if (assignedRegister != REG_NA) - { - unassignPhysReg(getRegisterRecord(assignedRegister), currentInterval->firstRefPosition); - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, currentInterval)); - } - currentRefPosition->registerAssignment = RBM_NONE; - continue; + treeNode->ClearLastUse(currentRefPosition->getMultiRegIdx()); } -#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE -#endif // FEATURE_SIMD - if (allocate == false) + if ((currentRefPosition->registerAssignment != RBM_NONE) && (interval->physReg == REG_NA) && + currentRefPosition->RegOptional() && currentRefPosition->lastUse && + (currentRefPosition->refType == RefTypeUse)) { - if (assignedRegister != REG_NA) - { - unassignPhysReg(getRegisterRecord(assignedRegister), currentRefPosition); - } - else if (!didDump) - { - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, currentInterval)); - didDump = true; - } + // This can happen if the incoming location for the block was changed from a register to the stack + // during resolution. In this case we're better off making it contained. + assert(inVarToRegMaps[curBBNum][varDsc->lvVarIndex] == REG_STK); currentRefPosition->registerAssignment = RBM_NONE; - continue; + writeLocalReg(treeNode->AsLclVar(), interval->varNum, REG_NA); } + } - if (currentInterval->isSpecialPutArg) + if (currentRefPosition->registerAssignment == RBM_NONE) + { + assert(currentRefPosition->RegOptional()); + assert(interval->isSpilled); + + varDsc->SetRegNum(REG_STK); + if (interval->assignedReg != nullptr && interval->assignedReg->assignedInterval == interval) { - assert(!currentInterval->isLocalVar); - Interval* srcInterval = currentInterval->relatedInterval; - assert(srcInterval != nullptr && srcInterval->isLocalVar); - if (refType == RefTypeDef) - { - assert(srcInterval->recentRefPosition->nodeLocation == currentLocation - 1); - RegRecord* physRegRecord = srcInterval->assignedReg; + updateAssignedInterval(interval->assignedReg, nullptr, interval->registerType); + } + interval->assignedReg = nullptr; + interval->physReg = REG_NA; + interval->isActive = false; - // For a putarg_reg to be special, its next use location has to be the same - // as fixed reg's next kill location. Otherwise, if source lcl var's next use - // is after the kill of fixed reg but before putarg_reg's next use, fixed reg's - // kill would lead to spill of source but not the putarg_reg if it were treated - // as special. - if (srcInterval->isActive && - genRegMask(srcInterval->physReg) == currentRefPosition->registerAssignment && - currentInterval->getNextRefLocation() == nextFixedRef[srcInterval->physReg]) - { - assert(physRegRecord->regNum == srcInterval->physReg); + // Set this as contained if it is not a multi-reg (we could potentially mark it s contained + // if all uses are from spill, but that adds complexity. + if ((currentRefPosition->refType == RefTypeUse) && !treeNode->IsMultiReg()) + { + assert(treeNode != nullptr); + treeNode->SetContained(); + } - // Special putarg_reg acts as a pass-thru since both source lcl var - // and putarg_reg have the same register allocated. Physical reg - // record of reg continue to point to source lcl var's interval - // instead of to putarg_reg's interval. So if a spill of reg - // allocated to source lcl var happens, to reallocate to another - // tree node, before its use at call node it will lead to spill of - // lcl var instead of putarg_reg since physical reg record is pointing - // to lcl var's interval. As a result, arg reg would get trashed leading - // to bad codegen. The assumption here is that source lcl var of a - // special putarg_reg doesn't get spilled and re-allocated prior to - // its use at the call node. This is ensured by marking physical reg - // record as busy until next kill. - setRegBusyUntilKill(srcInterval->physReg, srcInterval->registerType); - } - else - { - currentInterval->isSpecialPutArg = false; - } - } - // If this is still a SpecialPutArg, continue; - if (currentInterval->isSpecialPutArg) + return; + } + + // In most cases, assigned and home registers will be the same + // The exception is the copyReg case, where we've assigned a register + // for a specific purpose, but will be keeping the register assignment + regNumber assignedReg = currentRefPosition->assignedReg(); + regNumber homeReg = assignedReg; + + // Undo any previous association with a physical register, UNLESS this + // is a copyReg + if (!currentRefPosition->copyReg) + { + regNumber oldAssignedReg = interval->physReg; + if (oldAssignedReg != REG_NA && assignedReg != oldAssignedReg) + { + RegRecord* oldRegRecord = getRegisterRecord(oldAssignedReg); + if (oldRegRecord->assignedInterval == interval) { - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_SPECIAL_PUTARG, currentInterval, - currentRefPosition->assignedReg())); - continue; + updateAssignedInterval(oldRegRecord, nullptr, interval->registerType); } } + } - if (assignedRegister == REG_NA && RefTypeIsUse(refType)) + if (currentRefPosition->refType == RefTypeUse && !currentRefPosition->reload) + { + // Was this spilled after our predecessor was scheduled? + if (interval->physReg == REG_NA) { + assert(inVarToRegMaps[curBBNum][varDsc->lvVarIndex] == REG_STK); currentRefPosition->reload = true; - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_RELOAD, currentInterval, assignedRegister)); } + } - regMaskTP assignedRegBit = RBM_NONE; - bool isInRegister = false; - if (assignedRegister != REG_NA) + bool reload = currentRefPosition->reload; + bool spillAfter = currentRefPosition->spillAfter; + bool writeThru = currentRefPosition->writeThru; + + // In the reload case we either: + // - Set the register to REG_STK if it will be referenced only from the home location, or + // - Set the register to the assigned register and set GTF_SPILLED if it must be loaded into a register. + if (reload) + { + assert(currentRefPosition->refType != RefTypeDef); + assert(interval->isSpilled); + varDsc->SetRegNum(REG_STK); + if (!spillAfter) { - isInRegister = true; - assignedRegBit = genRegMask(assignedRegister); - if (!currentInterval->isActive) + interval->physReg = assignedReg; + } + + // If there is no treeNode, this must be a RefTypeExpUse, in + // which case we did the reload already + if (treeNode != nullptr) + { + treeNode->gtFlags |= GTF_SPILLED; + if (treeNode->IsMultiReg()) { - // If this is a use, it must have started the block on the stack, but the register - // was available for use so we kept the association. - if (RefTypeIsUse(refType)) + treeNode->SetRegSpillFlagByIdx(GTF_SPILLED, currentRefPosition->getMultiRegIdx()); + } + if (spillAfter) + { + if (currentRefPosition->RegOptional()) { - assert(enregisterLocalVars); - assert(inVarToRegMaps[curBBNum][currentInterval->getVarIndex(compiler)] == REG_STK && - previousRefPosition->nodeLocation <= curBBStartLocation); - isInRegister = false; + // This is a use of lclVar that is flagged as reg-optional + // by lower/codegen and marked for both reload and spillAfter. + // In this case we can avoid unnecessary reload and spill + // by setting reg on lclVar to REG_STK and reg on tree node + // to REG_NA. Codegen will generate the code by considering + // it as a contained memory operand. + // + // Note that varDsc->GetRegNum() is already to REG_STK above. + interval->physReg = REG_NA; + writeLocalReg(treeNode->AsLclVar(), interval->varNum, REG_NA); + treeNode->gtFlags &= ~GTF_SPILLED; + treeNode->SetContained(); + // We don't support RegOptional for multi-reg localvars. + assert(!treeNode->IsMultiReg()); } else { - currentInterval->isActive = true; - setRegInUse(assignedRegister, currentInterval->registerType); - updateSpillCost(assignedRegister, currentInterval); + treeNode->gtFlags |= GTF_SPILL; + if (treeNode->IsMultiReg()) + { + treeNode->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); + } } - updateNextIntervalRef(assignedRegister, currentInterval); } - assert(currentInterval->assignedReg != nullptr && - currentInterval->assignedReg->regNum == assignedRegister && - currentInterval->assignedReg->assignedInterval == currentInterval); } + else + { + assert(currentRefPosition->refType == RefTypeExpUse); + } + } + else if (spillAfter && !RefTypeIsUse(currentRefPosition->refType) && + (!treeNode->IsMultiReg() || treeNode->gtGetOp1()->IsMultiRegNode())) + { + // In the case of a pure def, don't bother spilling - just assign it to the + // stack. However, we need to remember that it was spilled. + // We can't do this in the case of a multi-reg node with a non-multireg source as + // we need the register to extract into. - if (previousRefPosition != nullptr) + assert(interval->isSpilled); + varDsc->SetRegNum(REG_STK); + interval->physReg = REG_NA; + if (treeNode != nullptr) { - assert(previousRefPosition->nextRefPosition == currentRefPosition); - assert(assignedRegister == REG_NA || assignedRegBit == previousRefPosition->registerAssignment || - currentRefPosition->outOfOrder || previousRefPosition->copyReg || - previousRefPosition->refType == RefTypeExpUse || currentRefPosition->refType == RefTypeDummyDef); + writeLocalReg(treeNode->AsLclVar(), interval->varNum, REG_NA); } - else if (assignedRegister != REG_NA) + } + else // Not reload and Not pure-def that's spillAfter + { + if (currentRefPosition->copyReg || currentRefPosition->moveReg) { - // Handle the case where this is a preassigned register (i.e. parameter). - // We don't want to actually use the preassigned register if it's not - // going to cover the lifetime - but we had to preallocate it to ensure - // that it remained live. - // TODO-CQ: At some point we may want to refine the analysis here, in case - // it might be beneficial to keep it in this reg for PART of the lifetime - if (currentInterval->isLocalVar) + // For a copyReg or moveReg, we have two cases: + // - In the first case, we have a fixedReg - i.e. a register which the code + // generator is constrained to use. + // The code generator will generate the appropriate move to meet the requirement. + // - In the second case, we were forced to use a different register because of + // interference (or JitStressRegs). + // In this case, we generate a GT_COPY. + // In either case, we annotate the treeNode with the register in which the value + // currently lives. For moveReg, the homeReg is the new register (as assigned above). + // But for copyReg, the homeReg remains unchanged. + + assert(treeNode != nullptr); + writeLocalReg(treeNode->AsLclVar(), interval->varNum, interval->physReg); + + if (currentRefPosition->copyReg) + { + homeReg = interval->physReg; + } + else + { + assert(interval->isSplit); + interval->physReg = assignedReg; + } + + if (!currentRefPosition->isFixedRegRef || currentRefPosition->moveReg) { - regMaskTP preferences = currentInterval->registerPreferences; - bool keepAssignment = true; - bool matchesPreferences = (preferences & genRegMask(assignedRegister)) != RBM_NONE; + // This is the second case, where we need to generate a copy + insertCopyOrReload(block, treeNode, currentRefPosition->getMultiRegIdx(), currentRefPosition); + } + } + else + { + interval->physReg = assignedReg; - // Will the assigned register cover the lifetime? If not, does it at least - // meet the preferences for the next RefPosition? - LsraLocation nextPhysRegLocation = nextFixedRef[assignedRegister]; - if (nextPhysRegLocation <= currentInterval->lastRefPosition->nodeLocation) + if (!interval->isSpilled && !interval->isSplit) + { + if (varDsc->GetRegNum() != REG_STK) { - // Check to see if the existing assignment matches the preferences (e.g. callee save registers) - // and ensure that the next use of this localVar does not occur after the nextPhysRegRefPos - // There must be a next RefPosition, because we know that the Interval extends beyond the - // nextPhysRegRefPos. - assert(nextRefPosition != nullptr); - if (!matchesPreferences || nextPhysRegLocation < nextRefPosition->nodeLocation) - { - keepAssignment = false; - } - else if ((nextRefPosition->registerAssignment != assignedRegBit) && - (nextPhysRegLocation <= nextRefPosition->getRefEndLocation())) + // If the register assignments don't match, then this interval is split. + if (varDsc->GetRegNum() != assignedReg) { - keepAssignment = false; + setIntervalAsSplit(interval); + varDsc->SetRegNum(REG_STK); } } - else if (refType == RefTypeParamDef && !matchesPreferences) - { - // Don't use the register, even if available, if it doesn't match the preferences. - // Note that this case is only for ParamDefs, for which we haven't yet taken preferences - // into account (we've just automatically got the initial location). In other cases, - // we would already have put it in a preferenced register, if it was available. - // TODO-CQ: Consider expanding this to check availability - that would duplicate - // code here, but otherwise we may wind up in this register anyway. - keepAssignment = false; - } - - if (keepAssignment == false) + else { - RegRecord* physRegRecord = getRegisterRecord(currentInterval->physReg); - currentRefPosition->registerAssignment = allRegs(currentInterval->registerType); - currentRefPosition->isFixedRegRef = false; - unassignPhysRegNoSpill(physRegRecord); - - // If the preferences are currently set to just this register, reset them to allRegs - // of the appropriate type (just as we just reset the registerAssignment for this - // RefPosition. - // Otherwise, simply remove this register from the preferences, if it's there. - - if (currentInterval->registerPreferences == assignedRegBit) - { - currentInterval->registerPreferences = currentRefPosition->registerAssignment; - } - else - { - currentInterval->registerPreferences &= ~assignedRegBit; - } - - assignedRegister = REG_NA; - assignedRegBit = RBM_NONE; + varDsc->SetRegNum(assignedReg); } } } - - if (assignedRegister != REG_NA) + if (spillAfter) { - RegRecord* physRegRecord = getRegisterRecord(assignedRegister); - assert((assignedRegBit == currentRefPosition->registerAssignment) || - (physRegRecord->assignedInterval == currentInterval) || - !isRegInUse(assignedRegister, currentInterval->registerType)); - if (conflictingFixedRegReference(assignedRegister, currentRefPosition)) + if (treeNode != nullptr) { - // We may have already reassigned the register to the conflicting reference. - // If not, we need to unassign this interval. - if (physRegRecord->assignedInterval == currentInterval) + treeNode->gtFlags |= GTF_SPILL; + if (treeNode->IsMultiReg()) { - unassignPhysRegNoSpill(physRegRecord); - physRegRecord->assignedInterval = nullptr; - clearConstantReg(assignedRegister, currentInterval->registerType); + treeNode->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); } - currentRefPosition->moveReg = true; - assignedRegister = REG_NA; - currentRefPosition->registerAssignment &= ~assignedRegBit; - setIntervalAsSplit(currentInterval); - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_MOVE_REG, currentInterval, assignedRegister)); } - else if ((genRegMask(assignedRegister) & currentRefPosition->registerAssignment) != 0) + assert(interval->isSpilled); + interval->physReg = REG_NA; + varDsc->SetRegNum(REG_STK); + } + if (writeThru && (treeNode != nullptr)) + { + // This is a def of a write-thru EH var (only defs are marked 'writeThru'). + treeNode->gtFlags |= GTF_SPILL; + // We also mark writeThru defs that are not last-use with GTF_SPILLED to indicate that they are conceptually + // spilled and immediately "reloaded", i.e. the register remains live. + // Note that we can have a "last use" write that has no exposed uses in the standard + // (non-eh) control flow, but that may be used on an exception path. Hence the need + // to retain these defs, and to ensure that they write. + if (!currentRefPosition->lastUse) { - currentRefPosition->registerAssignment = assignedRegBit; - if (!currentInterval->isActive) + treeNode->gtFlags |= GTF_SPILLED; + if (treeNode->IsMultiReg()) { - // If we've got an exposed use at the top of a block, the - // interval might not have been active. Otherwise if it's a use, - // the interval must be active. - if (refType == RefTypeDummyDef) - { - currentInterval->isActive = true; - assert(getRegisterRecord(assignedRegister)->assignedInterval == currentInterval); - } - else - { - currentRefPosition->reload = true; - } + treeNode->SetRegSpillFlagByIdx(GTF_SPILLED, currentRefPosition->getMultiRegIdx()); } - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, currentInterval, assignedRegister)); } - else - { - // It's already in a register, but not one we need. - if (!RefTypeIsDef(currentRefPosition->refType)) - { - regNumber copyReg = assignCopyReg(currentRefPosition); - lastAllocatedRefPosition = currentRefPosition; - bool unassign = false; - if (currentInterval->isWriteThru) - { - if (currentRefPosition->refType == RefTypeDef) - { - currentRefPosition->writeThru = true; - } - if (!currentRefPosition->lastUse) - { - if (currentRefPosition->spillAfter) - { - unassign = true; - } - } - } - regMaskTP copyRegMask = getRegMask(copyReg, currentInterval->registerType); - regMaskTP assignedRegMask = getRegMask(assignedRegister, currentInterval->registerType); - regsInUseThisLocation |= copyRegMask | assignedRegMask; - if (currentRefPosition->lastUse) - { - if (currentRefPosition->delayRegFree) - { - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE_DELAYED, currentInterval, - assignedRegister)); - delayRegsToFree |= copyRegMask | assignedRegMask; - regsInUseNextLocation |= copyRegMask | assignedRegMask; - } - else - { - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE, currentInterval, assignedRegister)); - regsToFree |= copyRegMask | assignedRegMask; - } - } - else - { - copyRegsToFree |= copyRegMask; - if (currentRefPosition->delayRegFree) - { - regsInUseNextLocation |= copyRegMask | assignedRegMask; - } - } + } + } + + // Update the physRegRecord for the register, so that we know what vars are in + // regs at the block boundaries + RegRecord* physRegRecord = getRegisterRecord(homeReg); + if (spillAfter || currentRefPosition->lastUse) + { + interval->isActive = false; + interval->assignedReg = nullptr; + interval->physReg = REG_NA; + + updateAssignedInterval(physRegRecord, nullptr, interval->registerType); + } + else + { + interval->isActive = true; + interval->assignedReg = physRegRecord; + + updateAssignedInterval(physRegRecord, interval, interval->registerType); + } +} + +void LinearScan::writeRegisters(RefPosition* currentRefPosition, GenTree* tree) +{ + lsraAssignRegToTree(tree, currentRefPosition->assignedReg(), currentRefPosition->getMultiRegIdx()); +} + +//------------------------------------------------------------------------ +// insertCopyOrReload: Insert a copy in the case where a tree node value must be moved +// to a different register at the point of use (GT_COPY), or it is reloaded to a different register +// than the one it was spilled from (GT_RELOAD). +// +// Arguments: +// block - basic block in which GT_COPY/GT_RELOAD is inserted. +// tree - This is the node to copy or reload. +// Insert copy or reload node between this node and its parent. +// multiRegIdx - register position of tree node for which copy or reload is needed. +// refPosition - The RefPosition at which copy or reload will take place. +// +// Notes: +// The GT_COPY or GT_RELOAD will be inserted in the proper spot in execution order where the reload is to occur. +// +// For example, for this tree (numbers are execution order, lower is earlier and higher is later): +// +// +---------+----------+ +// | GT_ADD (3) | +// +---------+----------+ +// | +// / '\' +// / '\' +// / '\' +// +-------------------+ +----------------------+ +// | x (1) | "tree" | y (2) | +// +-------------------+ +----------------------+ +// +// generate this tree: +// +// +---------+----------+ +// | GT_ADD (4) | +// +---------+----------+ +// | +// / '\' +// / '\' +// / '\' +// +-------------------+ +----------------------+ +// | GT_RELOAD (3) | | y (2) | +// +-------------------+ +----------------------+ +// | +// +-------------------+ +// | x (1) | "tree" +// +-------------------+ +// +// Note in particular that the GT_RELOAD node gets inserted in execution order immediately before the parent of "tree", +// which seems a bit weird since normally a node's parent (in this case, the parent of "x", GT_RELOAD in the "after" +// picture) immediately follows all of its children (that is, normally the execution ordering is postorder). +// The ordering must be this weird "out of normal order" way because the "x" node is being spilled, probably +// because the expression in the tree represented above by "y" has high register requirements. We don't want +// to reload immediately, of course. So we put GT_RELOAD where the reload should actually happen. +// +// Note that GT_RELOAD is required when we reload to a different register than the one we spilled to. It can also be +// used if we reload to the same register. Normally, though, in that case we just mark the node with GTF_SPILLED, +// and the unspilling code automatically reuses the same register, and does the reload when it notices that flag +// when considering a node's operands. +// +void LinearScan::insertCopyOrReload(BasicBlock* block, GenTree* tree, unsigned multiRegIdx, RefPosition* refPosition) +{ + LIR::Range& blockRange = LIR::AsRange(block); - // If this is a tree temp (non-localVar) interval, we will need an explicit move. - // Note: In theory a moveReg should cause the Interval to now have the new reg as its - // assigned register. However, that's not currently how this works. - // If we ever actually move lclVar intervals instead of copying, this will need to change. - if (!currentInterval->isLocalVar) - { - currentRefPosition->moveReg = true; - currentRefPosition->copyReg = false; - } - clearNextIntervalRef(copyReg, currentInterval->registerType); - clearSpillCost(copyReg, currentInterval->registerType); - updateNextIntervalRef(assignedRegister, currentInterval); - updateSpillCost(assignedRegister, currentInterval); - continue; - } - else - { - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NEEDS_NEW_REG, nullptr, assignedRegister)); - regsToFree |= getRegMask(assignedRegister, currentInterval->registerType); - // We want a new register, but we don't want this to be considered a spill. - assignedRegister = REG_NA; - if (physRegRecord->assignedInterval == currentInterval) - { - unassignPhysRegNoSpill(physRegRecord); - } - } - } - } + LIR::Use treeUse; + bool foundUse = blockRange.TryGetUse(tree, &treeUse); + assert(foundUse); - if (assignedRegister == REG_NA) + GenTree* parent = treeUse.User(); + + genTreeOps oper; + if (refPosition->reload) + { + oper = GT_RELOAD; + } + else + { + oper = GT_COPY; + + INTRACK_STATS(updateLsraStat(STAT_COPY_REG, block->bbNum)); + } + + // If the parent is a reload/copy node, then tree must be a multi-reg node + // that has already had one of its registers spilled. + // It is possible that one of its RefTypeDef positions got spilled and the next + // use of it requires it to be in a different register. + // + // In this case set the i'th position reg of reload/copy node to the reg allocated + // for copy/reload refPosition. Essentially a copy/reload node will have a reg + // for each multi-reg position of its child. If there is a valid reg in i'th + // position of GT_COPY or GT_RELOAD node then the corresponding result of its + // child needs to be copied or reloaded to that reg. + if (parent->IsCopyOrReload()) + { + noway_assert(parent->OperGet() == oper); + noway_assert(tree->IsMultiRegNode()); + GenTreeCopyOrReload* copyOrReload = parent->AsCopyOrReload(); + noway_assert(copyOrReload->GetRegNumByIdx(multiRegIdx) == REG_NA); + copyOrReload->SetRegNumByIdx(refPosition->assignedReg(), multiRegIdx); + } + else + { + // Create the new node, with "tree" as its only child. + var_types treeType = tree->TypeGet(); + + GenTreeCopyOrReload* newNode = new (compiler, oper) GenTreeCopyOrReload(oper, treeType, tree); + assert(refPosition->registerAssignment != RBM_NONE); + SetLsraAdded(newNode); + newNode->SetRegNumByIdx(refPosition->assignedReg(), multiRegIdx); + if (refPosition->copyReg) { - if (currentRefPosition->RegOptional()) - { - // We can avoid allocating a register if it is a last use requiring a reload. - if (currentRefPosition->lastUse && currentRefPosition->reload) - { - allocate = false; - } - else if (currentInterval->isWriteThru) - { - // Don't allocate if the next reference is in a cold block. - if (nextRefPosition == nullptr || (nextRefPosition->nodeLocation >= firstColdLoc)) - { - allocate = false; - } - } + // This is a TEMPORARY copy + assert(isCandidateLocalRef(tree) || tree->IsMultiRegLclVar()); + newNode->SetLastUse(multiRegIdx); + } -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE && defined(TARGET_XARCH) - // We can also avoid allocating a register (in fact we don't want to) if we have - // an UpperVectorRestore on xarch where the value is on the stack. - if ((currentRefPosition->refType == RefTypeUpperVectorRestore) && (currentInterval->physReg == REG_NA)) - { - assert(currentRefPosition->regOptional); - allocate = false; - } -#endif + // Insert the copy/reload after the spilled node and replace the use of the original node with a use + // of the copy/reload. + blockRange.InsertAfter(tree, newNode); + treeUse.ReplaceWith(compiler, newNode); + } +} -#ifdef DEBUG - // Under stress mode, don't allocate registers to RegOptional RefPositions. - if (allocate && regOptionalNoAlloc()) - { - allocate = false; - } +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE +//------------------------------------------------------------------------ +// insertUpperVectorSave: Insert code to save the upper half of a vector that lives +// in a callee-save register at the point of a kill (the upper half is +// not preserved). +// +// Arguments: +// tree - This is the node before which we will insert the Save. +// It will be a call or some node that turns into a call. +// refPosition - The RefTypeUpperVectorSave RefPosition. +// upperInterval - The Interval for the upper half of the large vector lclVar. +// block - the BasicBlock containing the call. +// +void LinearScan::insertUpperVectorSave(GenTree* tree, + RefPosition* refPosition, + Interval* upperVectorInterval, + BasicBlock* block) +{ + JITDUMP("Inserting UpperVectorSave for RP #%d before %d.%s:\n", refPosition->rpNum, tree->gtTreeID, + GenTree::OpName(tree->gtOper)); + Interval* lclVarInterval = upperVectorInterval->relatedInterval; + assert(lclVarInterval->isLocalVar == true); + assert(refPosition->getInterval() == upperVectorInterval); + regNumber lclVarReg = lclVarInterval->physReg; + if (lclVarReg == REG_NA) + { + return; + } + + LclVarDsc* varDsc = compiler->lvaTable + lclVarInterval->varNum; + assert(Compiler::varTypeNeedsPartialCalleeSave(varDsc->lvType)); + + // On Arm64, we must always have a register to save the upper half, + // while on x86 we can spill directly to memory. + regNumber spillReg = refPosition->assignedReg(); +#ifdef TARGET_ARM64 + bool spillToMem = refPosition->spillAfter; + assert(spillReg != REG_NA); +#else + bool spillToMem = (spillReg == REG_NA); + assert(!refPosition->spillAfter); #endif - } - RegisterScore registerScore = NONE; - if (allocate) - { - // Allocate a register, if we must, or if it is profitable to do so. - // If we have a fixed reg requirement, and the interval is inactive in another register, - // unassign that register. - if (currentRefPosition->isFixedRegRef && !currentInterval->isActive && - (currentInterval->assignedReg != nullptr) && - (currentInterval->assignedReg->assignedInterval == currentInterval) && - (genRegMask(currentInterval->assignedReg->regNum) != currentRefPosition->registerAssignment)) - { - unassignPhysReg(currentInterval->assignedReg, nullptr); - } - assignedRegister = allocateReg(currentInterval, currentRefPosition DEBUG_ARG(®isterScore)); - } + LIR::Range& blockRange = LIR::AsRange(block); - // If no register was found, this RefPosition must not require a register. - if (assignedRegister == REG_NA) - { - assert(currentRefPosition->RegOptional()); - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, currentInterval)); - currentRefPosition->registerAssignment = RBM_NONE; - currentRefPosition->reload = false; - currentInterval->isActive = false; - setIntervalAsSpilled(currentInterval); - } -#ifdef DEBUG - else - { - if (VERBOSE) - { - if (currentInterval->isConstant && (currentRefPosition->treeNode != nullptr) && - currentRefPosition->treeNode->IsReuseRegVal()) - { - dumpLsraAllocationEvent(LSRA_EVENT_REUSE_REG, currentInterval, assignedRegister, currentBlock, - registerScore); - } - else - { - dumpLsraAllocationEvent(LSRA_EVENT_ALLOC_REG, currentInterval, assignedRegister, currentBlock, - registerScore); - } - } - } -#endif // DEBUG + // Insert the save before the call. - if (refType == RefTypeDummyDef && assignedRegister != REG_NA) - { - setInVarRegForBB(curBBNum, currentInterval->varNum, assignedRegister); - } + GenTree* saveLcl = compiler->gtNewLclvNode(lclVarInterval->varNum, varDsc->lvType); + saveLcl->SetRegNum(lclVarReg); + SetLsraAdded(saveLcl); - // If we allocated a register, and this is a use of a spilled value, - // it should have been marked for reload above. - if (assignedRegister != REG_NA && RefTypeIsUse(refType) && !isInRegister) - { - assert(currentRefPosition->reload); - } - } + GenTreeSIMD* simdNode = + new (compiler, GT_SIMD) GenTreeSIMD(LargeVectorSaveType, saveLcl, nullptr, SIMDIntrinsicUpperSave, + varDsc->GetSimdBaseJitType(), genTypeSize(varDsc->lvType)); + + if (simdNode->GetSimdBaseJitType() == CORINFO_TYPE_UNDEF) + { + // There are a few scenarios where we can get a LCL_VAR which + // doesn't know the underlying baseType. In that scenario, we + // will just lie and say it is a float. Codegen doesn't actually + // care what the type is but this avoids an assert that would + // otherwise be fired from the more general checks that happen. + simdNode->SetSimdBaseJitType(CORINFO_TYPE_FLOAT); + } + + SetLsraAdded(simdNode); + simdNode->SetRegNum(spillReg); + if (spillToMem) + { + simdNode->gtFlags |= GTF_SPILL; + upperVectorInterval->physReg = REG_NA; + } + else + { + assert((genRegMask(spillReg) & RBM_FLT_CALLEE_SAVED) != RBM_NONE); + upperVectorInterval->physReg = spillReg; + } + + blockRange.InsertBefore(tree, LIR::SeqTree(compiler, simdNode)); + DISPTREE(simdNode); + JITDUMP("\n"); +} + +//------------------------------------------------------------------------ +// insertUpperVectorRestore: Insert code to restore the upper half of a vector that has been partially spilled. +// +// Arguments: +// tree - This is the node for which we will insert the Restore. +// If non-null, it will be a use of the large vector lclVar. +// If null, the Restore will be added to the end of the block. +// upperVectorInterval - The Interval for the upper vector for the lclVar. +// block - the BasicBlock into which we will be inserting the code. +// +// Notes: +// In the case where 'tree' is non-null, we will insert the restore just prior to +// its use, in order to ensure the proper ordering. +// +void LinearScan::insertUpperVectorRestore(GenTree* tree, + RefPosition* refPosition, + Interval* upperVectorInterval, + BasicBlock* block) +{ + JITDUMP("Adding UpperVectorRestore for RP #%d ", refPosition->rpNum); + Interval* lclVarInterval = upperVectorInterval->relatedInterval; + assert(lclVarInterval->isLocalVar == true); + regNumber lclVarReg = lclVarInterval->physReg; - // If we allocated a register, record it - if (assignedRegister != REG_NA) - { - assignedRegBit = genRegMask(assignedRegister); - regMaskTP regMask = getRegMask(assignedRegister, currentInterval->registerType); - regsInUseThisLocation |= regMask; - if (currentRefPosition->delayRegFree) - { - regsInUseNextLocation |= regMask; - } - currentRefPosition->registerAssignment = assignedRegBit; + // We should not call this method if the lclVar is not in a register (we should have simply marked the entire + // lclVar as spilled). + assert(lclVarReg != REG_NA); + LclVarDsc* varDsc = compiler->lvaTable + lclVarInterval->varNum; + assert(Compiler::varTypeNeedsPartialCalleeSave(varDsc->lvType)); - currentInterval->physReg = assignedRegister; - regsToFree &= ~regMask; // we'll set it again later if it's dead + GenTree* restoreLcl = nullptr; + restoreLcl = compiler->gtNewLclvNode(lclVarInterval->varNum, varDsc->lvType); + restoreLcl->SetRegNum(lclVarReg); + SetLsraAdded(restoreLcl); - // If this interval is dead, free the register. - // The interval could be dead if this is a user variable, or if the - // node is being evaluated for side effects, or a call whose result - // is not used, etc. - // If this is an UpperVector we'll neither free it nor preference it - // (it will be freed when it is used). - bool unassign = false; - if (!currentInterval->IsUpperVector()) - { - if (currentInterval->isWriteThru) - { - if (currentRefPosition->refType == RefTypeDef) - { - currentRefPosition->writeThru = true; - } - if (!currentRefPosition->lastUse) - { - if (currentRefPosition->spillAfter) - { - unassign = true; - } - } - } - if (currentRefPosition->lastUse || currentRefPosition->nextRefPosition == nullptr) - { - assert(currentRefPosition->isIntervalRef()); - // If this isn't a final use, we'll mark the register as available, but keep the association. - if ((refType != RefTypeExpUse) && (currentRefPosition->nextRefPosition == nullptr)) - { - unassign = true; - } - else - { - if (currentRefPosition->delayRegFree) - { - delayRegsToMakeInactive |= regMask; - } - else - { - regsToMakeInactive |= regMask; - } - // TODO-Cleanup: this makes things consistent with previous, and will enable preferences - // to be propagated, but it seems less than ideal. - currentInterval->isActive = false; - } - // Update the register preferences for the relatedInterval, if this is 'preferencedToDef'. - // Don't propagate to subsequent relatedIntervals; that will happen as they are allocated, and we - // don't know yet whether the register will be retained. - if (currentInterval->relatedInterval != nullptr) - { - currentInterval->relatedInterval->updateRegisterPreferences(assignedRegBit); - } - } + GenTreeSIMD* simdNode = + new (compiler, GT_SIMD) GenTreeSIMD(varDsc->lvType, restoreLcl, nullptr, SIMDIntrinsicUpperRestore, + varDsc->GetSimdBaseJitType(), genTypeSize(varDsc->lvType)); - if (unassign) - { - if (currentRefPosition->delayRegFree) - { - delayRegsToFree |= regMask; + if (simdNode->GetSimdBaseJitType() == CORINFO_TYPE_UNDEF) + { + // There are a few scenarios where we can get a LCL_VAR which + // doesn't know the underlying baseType. In that scenario, we + // will just lie and say it is a float. Codegen doesn't actually + // care what the type is but this avoids an assert that would + // otherwise be fired from the more general checks that happen. + simdNode->SetSimdBaseJitType(CORINFO_TYPE_FLOAT); + } - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE_DELAYED)); - } - else - { - regsToFree |= regMask; + regNumber restoreReg = upperVectorInterval->physReg; + SetLsraAdded(simdNode); - INDEBUG(dumpLsraAllocationEvent(LSRA_EVENT_LAST_USE)); - } - } - } - if (!unassign) - { - updateNextIntervalRef(assignedRegister, currentInterval); - updateSpillCost(assignedRegister, currentInterval); - } - } - lastAllocatedRefPosition = currentRefPosition; + if (restoreReg == REG_NA) + { + // We need a stack location for this. + assert(lclVarInterval->isSpilled); +#ifdef TARGET_AMD64 + assert(refPosition->assignedReg() == REG_NA); + simdNode->gtFlags |= GTF_NOREG_AT_USE; +#else + simdNode->gtFlags |= GTF_SPILLED; + assert(refPosition->assignedReg() != REG_NA); + restoreReg = refPosition->assignedReg(); +#endif } + simdNode->SetRegNum(restoreReg); -#ifdef JIT32_GCENCODER - // For the JIT32_GCENCODER, when lvaKeepAliveAndReportThis is true, we must either keep the "this" pointer - // in the same register for the entire method, or keep it on the stack. Rather than imposing this constraint - // as we allocate, we will force all refs to the stack if it is split or spilled. - if (enregisterLocalVars && compiler->lvaKeepAliveAndReportThis()) + LIR::Range& blockRange = LIR::AsRange(block); + JITDUMP("Adding UpperVectorRestore "); + if (tree != nullptr) { - LclVarDsc* thisVarDsc = compiler->lvaGetDesc(compiler->info.compThisArg); - if (thisVarDsc->lvLRACandidate) + JITDUMP("before %d.%s:\n", tree->gtTreeID, GenTree::OpName(tree->gtOper)); + LIR::Use treeUse; + bool foundUse = blockRange.TryGetUse(tree, &treeUse); + assert(foundUse); + // We need to insert the restore prior to the use, not (necessarily) immediately after the lclVar. + blockRange.InsertBefore(treeUse.User(), LIR::SeqTree(compiler, simdNode)); + } + else + { + JITDUMP("at end of " FMT_BB ":\n", block->bbNum); + if (block->bbJumpKind == BBJ_COND || block->bbJumpKind == BBJ_SWITCH) { - Interval* interval = getIntervalForLocalVar(thisVarDsc->lvVarIndex); - if (interval->isSplit) - { - // We'll have to spill this. - setIntervalAsSpilled(interval); - } - if (interval->isSpilled) - { - unsigned prevBBNum = 0; - for (RefPosition* ref = interval->firstRefPosition; ref != nullptr; ref = ref->nextRefPosition) - { - // For the resolution phase, we need to ensure that any block with exposed uses has the - // incoming reg for 'this' as REG_STK. - if (RefTypeIsUse(ref->refType) && (ref->bbNum != prevBBNum)) - { - VarToRegMap inVarToRegMap = getInVarToRegMap(ref->bbNum); - setVarReg(inVarToRegMap, thisVarDsc->lvVarIndex, REG_STK); - } - if (ref->RegOptional()) - { - ref->registerAssignment = RBM_NONE; - ref->reload = false; - ref->spillAfter = false; - } - switch (ref->refType) - { - case RefTypeDef: - if (ref->registerAssignment != RBM_NONE) - { - ref->spillAfter = true; - } - break; - case RefTypeUse: - if (ref->registerAssignment != RBM_NONE) - { - ref->reload = true; - ref->spillAfter = true; - ref->copyReg = false; - ref->moveReg = false; - } - break; - default: - break; - } - prevBBNum = ref->bbNum; - } - } + noway_assert(!blockRange.IsEmpty()); + + GenTree* branch = blockRange.LastNode(); + assert(branch->OperIsConditionalJump() || branch->OperGet() == GT_SWITCH_TABLE || + branch->OperGet() == GT_SWITCH); + + blockRange.InsertBefore(branch, LIR::SeqTree(compiler, simdNode)); + } + else + { + assert(block->bbJumpKind == BBJ_NONE || block->bbJumpKind == BBJ_ALWAYS); + blockRange.InsertAtEnd(LIR::SeqTree(compiler, simdNode)); } } -#endif // JIT32_GCENCODER + DISPTREE(simdNode); + JITDUMP("\n"); +} +#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE - // Free registers to clear associated intervals for resolution phase - CLANG_FORMAT_COMMENT_ANCHOR; +//------------------------------------------------------------------------ +// initMaxSpill: Initializes the LinearScan members used to track the max number +// of concurrent spills. This is needed so that we can set the +// fields in Compiler, so that the code generator, in turn can +// allocate the right number of spill locations. +// +// Arguments: +// None. +// +// Return Value: +// None. +// +// Assumptions: +// This is called before any calls to updateMaxSpill(). + +void LinearScan::initMaxSpill() +{ + needDoubleTmpForFPCall = false; + needFloatTmpForFPCall = false; + for (int i = 0; i < TYP_COUNT; i++) + { + maxSpill[i] = 0; + currentSpill[i] = 0; + } +} -#ifdef DEBUG - if (getLsraExtendLifeTimes()) +//------------------------------------------------------------------------ +// recordMaxSpill: Sets the fields in Compiler for the max number of concurrent spills. +// (See the comment on initMaxSpill.) +// +// Arguments: +// None. +// +// Return Value: +// None. +// +// Assumptions: +// This is called after updateMaxSpill() has been called for all "real" +// RefPositions. + +void LinearScan::recordMaxSpill() +{ + // Note: due to the temp normalization process (see tmpNormalizeType) + // only a few types should actually be seen here. + JITDUMP("Recording the maximum number of concurrent spills:\n"); +#ifdef TARGET_X86 + var_types returnType = RegSet::tmpNormalizeType(compiler->info.compRetType); + if (needDoubleTmpForFPCall || (returnType == TYP_DOUBLE)) { - // If we have extended lifetimes, we need to make sure all the registers are freed. - for (size_t regNumIndex = 0; regNumIndex <= REG_FP_LAST; regNumIndex++) - { - RegRecord& regRecord = physRegs[regNumIndex]; - Interval* interval = regRecord.assignedInterval; - if (interval != nullptr) - { - interval->isActive = false; - unassignPhysReg(®Record, nullptr); - } - } + JITDUMP("Adding a spill temp for moving a double call/return value between xmm reg and x87 stack.\n"); + maxSpill[TYP_DOUBLE] += 1; } - else -#endif // DEBUG + if (needFloatTmpForFPCall || (returnType == TYP_FLOAT)) { - freeRegisters(regsToFree | delayRegsToFree); + JITDUMP("Adding a spill temp for moving a float call/return value between xmm reg and x87 stack.\n"); + maxSpill[TYP_FLOAT] += 1; } - -#ifdef DEBUG - if (VERBOSE) +#endif // TARGET_X86 + for (int i = 0; i < TYP_COUNT; i++) { - // Dump the RegRecords after the last RefPosition is handled. - dumpRegRecords(); - printf("\n"); - - dumpRefPositions("AFTER ALLOCATION"); - dumpVarRefPositions("AFTER ALLOCATION"); - - // Dump the intervals that remain active - printf("Active intervals at end of allocation:\n"); - - // We COULD just reuse the intervalIter from above, but ArrayListIterator doesn't - // provide a Reset function (!) - we'll probably replace this so don't bother - // adding it - - for (Interval& interval : intervals) + if (var_types(i) != RegSet::tmpNormalizeType(var_types(i))) { - if (interval.isActive) - { - printf("Active "); - interval.dump(); - } + // Only normalized types should have anything in the maxSpill array. + // We assume here that if type 'i' does not normalize to itself, then + // nothing else normalizes to 'i', either. + assert(maxSpill[i] == 0); + } + if (maxSpill[i] != 0) + { + JITDUMP(" %s: %d\n", varTypeName(var_types(i)), maxSpill[i]); + compiler->codeGen->regSet.tmpPreAllocateTemps(var_types(i), maxSpill[i]); } - - printf("\n"); } -#endif // DEBUG + JITDUMP("\n"); } -//----------------------------------------------------------------------------- -// updateAssignedInterval: Update assigned interval of register. +//------------------------------------------------------------------------ +// updateMaxSpill: Update the maximum number of concurrent spills // // Arguments: -// reg - register to be updated -// interval - interval to be assigned -// regType - register type +// refPosition - the current RefPosition being handled // // Return Value: -// None +// None. // // Assumptions: -// For ARM32, when "regType" is TYP_DOUBLE, "reg" should be a even-numbered -// float register, i.e. lower half of double register. -// -// Note: -// For ARM32, two float registers consisting a double register are updated -// together when "regType" is TYP_DOUBLE. +// The RefPosition has an associated interval (getInterval() will +// otherwise assert). // -void LinearScan::updateAssignedInterval(RegRecord* reg, Interval* interval, RegisterType regType) +// Notes: +// This is called for each "real" RefPosition during the writeback +// phase of LSRA. It keeps track of how many concurrently-live +// spills there are, and the largest number seen so far. + +void LinearScan::updateMaxSpill(RefPosition* refPosition) { -#ifdef TARGET_ARM - // Update overlapping floating point register for TYP_DOUBLE. - Interval* oldAssignedInterval = reg->assignedInterval; - regNumber doubleReg = REG_NA; - if (regType == TYP_DOUBLE) + RefType refType = refPosition->refType; + +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + if ((refType == RefTypeUpperVectorSave) || (refType == RefTypeUpperVectorRestore)) { - doubleReg = reg->regNum; - RegRecord* anotherHalfReg = getSecondHalfRegRec(reg); - anotherHalfReg->assignedInterval = interval; + Interval* interval = refPosition->getInterval(); + // If this is not an 'upperVector', it must be a tree temp that has been already + // (fully) spilled. + if (!interval->isUpperVector) + { + assert(interval->firstRefPosition->spillAfter); + } + else + { + // The UpperVector RefPositions spill to the localVar's home location. + Interval* lclVarInterval = interval->relatedInterval; + assert(lclVarInterval->isSpilled || (!refPosition->spillAfter && !refPosition->reload)); + } + return; } - else if ((oldAssignedInterval != nullptr) && (oldAssignedInterval->registerType == TYP_DOUBLE)) +#endif // !FEATURE_PARTIAL_SIMD_CALLEE_SAVE + if (refPosition->spillAfter || refPosition->reload || + (refPosition->RegOptional() && refPosition->assignedReg() == REG_NA)) { - RegRecord* anotherHalfReg = findAnotherHalfRegRec(reg); - doubleReg = genIsValidDoubleReg(reg->regNum) ? reg->regNum : anotherHalfReg->regNum; - anotherHalfReg->assignedInterval = nullptr; + Interval* interval = refPosition->getInterval(); + if (!interval->isLocalVar) + { + GenTree* treeNode = refPosition->treeNode; + if (treeNode == nullptr) + { + assert(RefTypeIsUse(refType)); + treeNode = interval->firstRefPosition->treeNode; + } + assert(treeNode != nullptr); + + // The tmp allocation logic 'normalizes' types to a small number of + // types that need distinct stack locations from each other. + // Those types are currently gc refs, byrefs, <= 4 byte non-GC items, + // 8-byte non-GC items, and 16-byte or 32-byte SIMD vectors. + // LSRA is agnostic to those choices but needs + // to know what they are here. + var_types type; + if (!treeNode->IsMultiRegNode()) + { + type = getDefType(treeNode); + } + else + { + type = treeNode->GetRegTypeByIndex(refPosition->getMultiRegIdx()); + } + + type = RegSet::tmpNormalizeType(type); + + if (refPosition->spillAfter && !refPosition->reload) + { + currentSpill[type]++; + if (currentSpill[type] > maxSpill[type]) + { + maxSpill[type] = currentSpill[type]; + } + } + else if (refPosition->reload) + { + assert(currentSpill[type] > 0); + currentSpill[type]--; + } + else if (refPosition->RegOptional() && refPosition->assignedReg() == REG_NA) + { + // A spill temp not getting reloaded into a reg because it is + // marked as allocate if profitable and getting used from its + // memory location. To properly account max spill for typ we + // decrement spill count. + assert(RefTypeIsUse(refType)); + assert(currentSpill[type] > 0); + currentSpill[type]--; + } + JITDUMP(" Max spill for %s is %d\n", varTypeName(type), maxSpill[type]); + } } - if (doubleReg != REG_NA) +} + +// This is the final phase of register allocation. It writes the register assignments to +// the tree, and performs resolution across joins and backedges. +// +void LinearScan::resolveRegisters() +{ + // Iterate over the tree and the RefPositions in lockstep + // - annotate the tree with register assignments by setting GetRegNum() or gtRegPair (for longs) + // on the tree node + // - track globally-live var locations + // - add resolution points at split/merge/critical points as needed + + // Need to use the same traversal order as the one that assigns the location numbers. + + // Dummy RefPositions have been added at any split, join or critical edge, at the + // point where resolution may be required. These are located: + // - for a split, at the top of the non-adjacent block + // - for a join, at the bottom of the non-adjacent joining block + // - for a critical edge, at the top of the target block of each critical + // edge. + // Note that a target block may have multiple incoming critical or split edges + // + // These RefPositions record the expected location of the Interval at that point. + // At each branch, we identify the location of each liveOut interval, and check + // against the RefPositions at the target. + + BasicBlock* block; + LsraLocation currentLocation = MinLocation; + + // Clear register assignments - these will be reestablished as lclVar defs (including RefTypeParamDefs) + // are encountered. + if (enregisterLocalVars) { - clearNextIntervalRef(doubleReg, TYP_DOUBLE); - clearSpillCost(doubleReg, TYP_DOUBLE); - clearConstantReg(doubleReg, TYP_DOUBLE); + for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) + { + RegRecord* physRegRecord = getRegisterRecord(reg); + Interval* assignedInterval = physRegRecord->assignedInterval; + if (assignedInterval != nullptr) + { + assignedInterval->assignedReg = nullptr; + assignedInterval->physReg = REG_NA; + } + physRegRecord->assignedInterval = nullptr; + physRegRecord->recentRefPosition = nullptr; + } + + // Clear "recentRefPosition" for lclVar intervals + for (unsigned varIndex = 0; varIndex < compiler->lvaTrackedCount; varIndex++) + { + if (localVarIntervals[varIndex] != nullptr) + { + localVarIntervals[varIndex]->recentRefPosition = nullptr; + localVarIntervals[varIndex]->isActive = false; + } + else + { + assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); + } + } } -#endif - reg->assignedInterval = interval; - if (interval != nullptr) + + // handle incoming arguments and special temps + RefPositionIterator refPosIterator = refPositions.begin(); + RefPosition* currentRefPosition = &refPosIterator; + + if (enregisterLocalVars) { - setRegInUse(reg->regNum, interval->registerType); - if (interval->isConstant) - { - setConstantReg(reg->regNum, interval->registerType); - } - else + VarToRegMap entryVarToRegMap = inVarToRegMaps[compiler->fgFirstBB->bbNum]; + for (; refPosIterator != refPositions.end() && + (currentRefPosition->refType == RefTypeParamDef || currentRefPosition->refType == RefTypeZeroInit); + ++refPosIterator, currentRefPosition = &refPosIterator) { - clearConstantReg(reg->regNum, interval->registerType); + Interval* interval = currentRefPosition->getInterval(); + assert(interval != nullptr && interval->isLocalVar); + resolveLocalRef(nullptr, nullptr, currentRefPosition); + regNumber reg = REG_STK; + int varIndex = interval->getVarIndex(compiler); + + if (!currentRefPosition->spillAfter && currentRefPosition->registerAssignment != RBM_NONE) + { + reg = currentRefPosition->assignedReg(); + } + else + { + reg = REG_STK; + interval->isActive = false; + } + setVarReg(entryVarToRegMap, varIndex, reg); } - updateNextIntervalRef(reg->regNum, interval); - updateSpillCost(reg->regNum, interval); } else { - clearNextIntervalRef(reg->regNum, reg->registerType); - clearSpillCost(reg->regNum, reg->registerType); + assert(refPosIterator == refPositions.end() || + (refPosIterator->refType != RefTypeParamDef && refPosIterator->refType != RefTypeZeroInit)); } -} - -//----------------------------------------------------------------------------- -// updatePreviousInterval: Update previous interval of register. -// -// Arguments: -// reg - register to be updated -// interval - interval to be assigned -// regType - register type -// -// Return Value: -// None -// -// Assumptions: -// For ARM32, when "regType" is TYP_DOUBLE, "reg" should be a even-numbered -// float register, i.e. lower half of double register. -// -// Note: -// For ARM32, two float registers consisting a double register are updated -// together when "regType" is TYP_DOUBLE. -// -void LinearScan::updatePreviousInterval(RegRecord* reg, Interval* interval, RegisterType regType) -{ - reg->previousInterval = interval; -#ifdef TARGET_ARM - // Update overlapping floating point register for TYP_DOUBLE - if (regType == TYP_DOUBLE) + // write back assignments + for (block = startBlockSequence(); block != nullptr; block = moveToNextBlock()) { - RegRecord* anotherHalfReg = findAnotherHalfRegRec(reg); + assert(curBBNum == block->bbNum); - anotherHalfReg->previousInterval = interval; - } -#endif -} + if (enregisterLocalVars) + { + // Record the var locations at the start of this block. + // (If it's fgFirstBB, we've already done that above, see entryVarToRegMap) -//----------------------------------------------------------------------------- -// writeLocalReg: Write the register assignment for a GT_LCL_VAR node. -// -// Arguments: -// lclNode - The GT_LCL_VAR node -// varNum - The variable number for the register -// reg - The assigned register -// -// Return Value: -// None -// -// Note: -// For a multireg node, 'varNum' will be the field local for the given register. -// -void LinearScan::writeLocalReg(GenTreeLclVar* lclNode, unsigned varNum, regNumber reg) -{ - assert((lclNode->GetLclNum() == varNum) == !lclNode->IsMultiReg()); - if (lclNode->GetLclNum() == varNum) - { - lclNode->SetRegNum(reg); - } - else - { - assert(compiler->lvaEnregMultiRegVars); - LclVarDsc* parentVarDsc = compiler->lvaGetDesc(lclNode->GetLclNum()); - assert(parentVarDsc->lvPromoted); - unsigned regIndex = varNum - parentVarDsc->lvFieldLclStart; - assert(regIndex < MAX_MULTIREG_COUNT); - lclNode->SetRegNumByIdx(reg, regIndex); - } -} + curBBStartLocation = currentRefPosition->nodeLocation; + if (block != compiler->fgFirstBB) + { + processBlockStartLocations(block); + } -//----------------------------------------------------------------------------- -// LinearScan::resolveLocalRef -// Description: -// Update the graph for a local reference. -// Also, track the register (if any) that is currently occupied. -// Arguments: -// treeNode: The lclVar that's being resolved -// currentRefPosition: the RefPosition associated with the treeNode -// -// Details: -// This method is called for each local reference, during the resolveRegisters -// phase of LSRA. It is responsible for keeping the following in sync: -// - varDsc->GetRegNum() (and GetOtherReg()) contain the unique register location. -// If it is not in the same register through its lifetime, it is set to REG_STK. -// - interval->physReg is set to the assigned register -// (i.e. at the code location which is currently being handled by resolveRegisters()) -// - interval->isActive is true iff the interval is live and occupying a register -// - interval->isSpilled should have already been set to true if the interval is EVER spilled -// - interval->isSplit is set to true if the interval does not occupy the same -// register throughout the method -// - RegRecord->assignedInterval points to the interval which currently occupies -// the register -// - For each lclVar node: -// - GetRegNum()/gtRegPair is set to the currently allocated register(s). -// - GTF_SPILLED is set on a use if it must be reloaded prior to use. -// - GTF_SPILL is set if it must be spilled after use. -// -// A copyReg is an ugly case where the variable must be in a specific (fixed) register, -// but it currently resides elsewhere. The register allocator must track the use of the -// fixed register, but it marks the lclVar node with the register it currently lives in -// and the code generator does the necessary move. -// -// Before beginning, the varDsc for each parameter must be set to its initial location. -// -// NICE: Consider tracking whether an Interval is always in the same location (register/stack) -// in which case it will require no resolution. -// -void LinearScan::resolveLocalRef(BasicBlock* block, GenTreeLclVar* treeNode, RefPosition* currentRefPosition) -{ - assert((block == nullptr) == (treeNode == nullptr)); - assert(enregisterLocalVars); + // Handle the DummyDefs, updating the incoming var location. + for (; refPosIterator != refPositions.end() && currentRefPosition->refType == RefTypeDummyDef; + ++refPosIterator, currentRefPosition = &refPosIterator) + { + assert(currentRefPosition->isIntervalRef()); + // Don't mark dummy defs as reload + currentRefPosition->reload = false; + resolveLocalRef(nullptr, nullptr, currentRefPosition); + regNumber reg; + if (currentRefPosition->registerAssignment != RBM_NONE) + { + reg = currentRefPosition->assignedReg(); + } + else + { + reg = REG_STK; + currentRefPosition->getInterval()->isActive = false; + } + setInVarRegForBB(curBBNum, currentRefPosition->getInterval()->varNum, reg); + } + } - // Is this a tracked local? Or just a register allocated for loading - // a non-tracked one? - Interval* interval = currentRefPosition->getInterval(); - assert(interval->isLocalVar); + // The next RefPosition should be for the block. Move past it. + assert(refPosIterator != refPositions.end()); + assert(currentRefPosition->refType == RefTypeBB); + ++refPosIterator; + currentRefPosition = &refPosIterator; + + // Handle the RefPositions for the block + for (; refPosIterator != refPositions.end() && currentRefPosition->refType != RefTypeBB && + currentRefPosition->refType != RefTypeDummyDef; + ++refPosIterator, currentRefPosition = &refPosIterator) + { + currentLocation = currentRefPosition->nodeLocation; + + // Ensure that the spill & copy info is valid. + // First, if it's reload, it must not be copyReg or moveReg + assert(!currentRefPosition->reload || (!currentRefPosition->copyReg && !currentRefPosition->moveReg)); + // If it's copyReg it must not be moveReg, and vice-versa + assert(!currentRefPosition->copyReg || !currentRefPosition->moveReg); + + switch (currentRefPosition->refType) + { +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + case RefTypeUpperVectorSave: + case RefTypeUpperVectorRestore: +#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE + case RefTypeUse: + case RefTypeDef: + // These are the ones we're interested in + break; + case RefTypeKill: + case RefTypeFixedReg: + // These require no handling at resolution time + assert(currentRefPosition->referent != nullptr); + currentRefPosition->referent->recentRefPosition = currentRefPosition; + continue; + case RefTypeExpUse: + // Ignore the ExpUse cases - a RefTypeExpUse would only exist if the + // variable is dead at the entry to the next block. So we'll mark + // it as in its current location and resolution will take care of any + // mismatch. + assert(getNextBlock() == nullptr || + !VarSetOps::IsMember(compiler, getNextBlock()->bbLiveIn, + currentRefPosition->getInterval()->getVarIndex(compiler))); + currentRefPosition->referent->recentRefPosition = currentRefPosition; + continue; + case RefTypeKillGCRefs: + // No action to take at resolution time, and no interval to update recentRefPosition for. + continue; + case RefTypeDummyDef: + case RefTypeParamDef: + case RefTypeZeroInit: + // Should have handled all of these already + default: + unreached(); + break; + } + updateMaxSpill(currentRefPosition); + GenTree* treeNode = currentRefPosition->treeNode; + +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + if (currentRefPosition->refType == RefTypeUpperVectorSave) + { + // The treeNode is a call or something that might become one. + noway_assert(treeNode != nullptr); + // If the associated interval is an UpperVector, this must be a RefPosition for a LargeVectorType + // LocalVar. + // Otherwise, this is a non-lclVar interval that has been spilled, and we don't need to do anything. + Interval* interval = currentRefPosition->getInterval(); + if (interval->isUpperVector) + { + Interval* localVarInterval = interval->relatedInterval; + if ((localVarInterval->physReg != REG_NA) && !localVarInterval->isPartiallySpilled) + { + // If the localVar is in a register, it must be in a register that is not trashed by + // the current node (otherwise it would have already been spilled). + assert((genRegMask(localVarInterval->physReg) & getKillSetForNode(treeNode)) == RBM_NONE); + // If we have allocated a register to spill it to, we will use that; otherwise, we will spill it + // to the stack. We can use as a temp register any non-arg caller-save register. + currentRefPosition->referent->recentRefPosition = currentRefPosition; + insertUpperVectorSave(treeNode, currentRefPosition, currentRefPosition->getInterval(), block); + localVarInterval->isPartiallySpilled = true; + } + } + else + { + // This is a non-lclVar interval that must have been spilled. + assert(!currentRefPosition->getInterval()->isLocalVar); + assert(currentRefPosition->getInterval()->firstRefPosition->spillAfter); + } + continue; + } + else if (currentRefPosition->refType == RefTypeUpperVectorRestore) + { + // Since we don't do partial restores of tree temp intervals, this must be an upperVector. + Interval* interval = currentRefPosition->getInterval(); + Interval* localVarInterval = interval->relatedInterval; + assert(interval->isUpperVector && (localVarInterval != nullptr)); + if (localVarInterval->physReg != REG_NA) + { + assert(localVarInterval->isPartiallySpilled); + assert((localVarInterval->assignedReg != nullptr) && + (localVarInterval->assignedReg->regNum == localVarInterval->physReg) && + (localVarInterval->assignedReg->assignedInterval == localVarInterval)); + insertUpperVectorRestore(treeNode, currentRefPosition, interval, block); + } + localVarInterval->isPartiallySpilled = false; + } +#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE - interval->recentRefPosition = currentRefPosition; - LclVarDsc* varDsc = interval->getLocalVar(compiler); + // Most uses won't actually need to be recorded (they're on the def). + // In those cases, treeNode will be nullptr. + if (treeNode == nullptr) + { + // This is either a use, a dead def, or a field of a struct + Interval* interval = currentRefPosition->getInterval(); + assert(currentRefPosition->refType == RefTypeUse || + currentRefPosition->registerAssignment == RBM_NONE || interval->isStructField || + interval->IsUpperVector()); - // NOTE: we set the LastUse flag here unless we are extending lifetimes, in which case we write - // this bit in checkLastUses. This is a bit of a hack, but is necessary because codegen requires - // accurate last use info that is not reflected in the lastUse bit on ref positions when we are extending - // lifetimes. See also the comments in checkLastUses. - if ((treeNode != nullptr) && !extendLifetimes()) - { - if (currentRefPosition->lastUse) - { - treeNode->SetLastUse(currentRefPosition->getMultiRegIdx()); - } - else - { - treeNode->ClearLastUse(currentRefPosition->getMultiRegIdx()); - } + // TODO-Review: Need to handle the case where any of the struct fields + // are reloaded/spilled at this use + assert(!interval->isStructField || + (currentRefPosition->reload == false && currentRefPosition->spillAfter == false)); - if ((currentRefPosition->registerAssignment != RBM_NONE) && (interval->physReg == REG_NA) && - currentRefPosition->RegOptional() && currentRefPosition->lastUse && - (currentRefPosition->refType == RefTypeUse)) - { - // This can happen if the incoming location for the block was changed from a register to the stack - // during resolution. In this case we're better off making it contained. - assert(inVarToRegMaps[curBBNum][varDsc->lvVarIndex] == REG_STK); - currentRefPosition->registerAssignment = RBM_NONE; - writeLocalReg(treeNode->AsLclVar(), interval->varNum, REG_NA); - } - } + if (interval->isLocalVar && !interval->isStructField) + { + LclVarDsc* varDsc = interval->getLocalVar(compiler); - if (currentRefPosition->registerAssignment == RBM_NONE) - { - assert(currentRefPosition->RegOptional()); - assert(interval->isSpilled); + // This must be a dead definition. We need to mark the lclVar + // so that it's not considered a candidate for lvRegister, as + // this dead def will have to go to the stack. + assert(currentRefPosition->refType == RefTypeDef); + varDsc->SetRegNum(REG_STK); + } + continue; + } - varDsc->SetRegNum(REG_STK); - if (interval->assignedReg != nullptr && interval->assignedReg->assignedInterval == interval) - { - updateAssignedInterval(interval->assignedReg, nullptr, interval->registerType); - } - interval->assignedReg = nullptr; - interval->physReg = REG_NA; - interval->isActive = false; + assert(currentRefPosition->isIntervalRef()); + if (currentRefPosition->getInterval()->isInternal) + { + treeNode->gtRsvdRegs |= currentRefPosition->registerAssignment; + } + else + { + writeRegisters(currentRefPosition, treeNode); - // Set this as contained if it is not a multi-reg (we could potentially mark it s contained - // if all uses are from spill, but that adds complexity. - if ((currentRefPosition->refType == RefTypeUse) && !treeNode->IsMultiReg()) - { - assert(treeNode != nullptr); - treeNode->SetContained(); - } + if (treeNode->OperIs(GT_LCL_VAR, GT_STORE_LCL_VAR) && currentRefPosition->getInterval()->isLocalVar) + { + resolveLocalRef(block, treeNode->AsLclVar(), currentRefPosition); + } - return; - } + // Mark spill locations on temps + // (local vars are handled in resolveLocalRef, above) + // Note that the tree node will be changed from GTF_SPILL to GTF_SPILLED + // in codegen, taking care of the "reload" case for temps + else if (currentRefPosition->spillAfter || (currentRefPosition->nextRefPosition != nullptr && + currentRefPosition->nextRefPosition->moveReg)) + { + if (treeNode != nullptr) + { + if (currentRefPosition->spillAfter) + { + treeNode->gtFlags |= GTF_SPILL; - // In most cases, assigned and home registers will be the same - // The exception is the copyReg case, where we've assigned a register - // for a specific purpose, but will be keeping the register assignment - regNumber assignedReg = currentRefPosition->assignedReg(); - regNumber homeReg = assignedReg; + // If this is a constant interval that is reusing a pre-existing value, we actually need + // to generate the value at this point in order to spill it. + if (treeNode->IsReuseRegVal()) + { + treeNode->ResetReuseRegVal(); + } - // Undo any previous association with a physical register, UNLESS this - // is a copyReg - if (!currentRefPosition->copyReg) - { - regNumber oldAssignedReg = interval->physReg; - if (oldAssignedReg != REG_NA && assignedReg != oldAssignedReg) - { - RegRecord* oldRegRecord = getRegisterRecord(oldAssignedReg); - if (oldRegRecord->assignedInterval == interval) - { - updateAssignedInterval(oldRegRecord, nullptr, interval->registerType); - } - } - } + // In case of multi-reg node, also set spill flag on the + // register specified by multi-reg index of current RefPosition. + // Note that the spill flag on treeNode indicates that one or + // more its allocated registers are in that state. + if (treeNode->IsMultiRegCall()) + { + GenTreeCall* call = treeNode->AsCall(); + call->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); + } +#if FEATURE_ARG_SPLIT + else if (treeNode->OperIsPutArgSplit()) + { + GenTreePutArgSplit* splitArg = treeNode->AsPutArgSplit(); + splitArg->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); + } +#ifdef TARGET_ARM + else if (treeNode->OperIsMultiRegOp()) + { + GenTreeMultiRegOp* multiReg = treeNode->AsMultiRegOp(); + multiReg->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); + } +#endif // TARGET_ARM +#endif // FEATURE_ARG_SPLIT + } - if (currentRefPosition->refType == RefTypeUse && !currentRefPosition->reload) - { - // Was this spilled after our predecessor was scheduled? - if (interval->physReg == REG_NA) - { - assert(inVarToRegMaps[curBBNum][varDsc->lvVarIndex] == REG_STK); - currentRefPosition->reload = true; - } - } + // If the value is reloaded or moved to a different register, we need to insert + // a node to hold the register to which it should be reloaded + RefPosition* nextRefPosition = currentRefPosition->nextRefPosition; + noway_assert(nextRefPosition != nullptr); + if (INDEBUG(alwaysInsertReload() ||) + nextRefPosition->assignedReg() != currentRefPosition->assignedReg()) + { +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + // Note that we asserted above that this is an Interval RefPosition. + Interval* currentInterval = currentRefPosition->getInterval(); + if (!currentInterval->isUpperVector && nextRefPosition->refType == RefTypeUpperVectorSave) + { + // The currentRefPosition is a spill of a tree temp. + // These have no associated Restore, as we always spill if the vector is + // in a register when this is encountered. + // The nextRefPosition we're interested in (where we may need to insert a + // reload or flag as GTF_NOREG_AT_USE) is the subsequent RefPosition. + assert(!currentInterval->isLocalVar); + nextRefPosition = nextRefPosition->nextRefPosition; + assert(nextRefPosition->refType != RefTypeUpperVectorSave); + } + // UpperVector intervals may have unique assignments at each reference. + if (!currentInterval->isUpperVector) +#endif + { + if (nextRefPosition->assignedReg() != REG_NA) + { + insertCopyOrReload(block, treeNode, currentRefPosition->getMultiRegIdx(), + nextRefPosition); + } + else + { + assert(nextRefPosition->RegOptional()); - bool reload = currentRefPosition->reload; - bool spillAfter = currentRefPosition->spillAfter; - bool writeThru = currentRefPosition->writeThru; + // In case of tree temps, if def is spilled and use didn't + // get a register, set a flag on tree node to be treated as + // contained at the point of its use. + if (currentRefPosition->spillAfter && currentRefPosition->refType == RefTypeDef && + nextRefPosition->refType == RefTypeUse) + { + assert(nextRefPosition->treeNode == nullptr); + treeNode->gtFlags |= GTF_NOREG_AT_USE; + } + } + } + } + } - // In the reload case we either: - // - Set the register to REG_STK if it will be referenced only from the home location, or - // - Set the register to the assigned register and set GTF_SPILLED if it must be loaded into a register. - if (reload) - { - assert(currentRefPosition->refType != RefTypeDef); - assert(interval->isSpilled); - varDsc->SetRegNum(REG_STK); - if (!spillAfter) + // We should never have to "spill after" a temp use, since + // they're single use + else + { + unreached(); + } + } + } + } + + if (enregisterLocalVars) { - interval->physReg = assignedReg; + processBlockEndLocations(block); } + } - // If there is no treeNode, this must be a RefTypeExpUse, in - // which case we did the reload already - if (treeNode != nullptr) + if (enregisterLocalVars) + { +#ifdef DEBUG + if (VERBOSE) { - treeNode->gtFlags |= GTF_SPILLED; - if (treeNode->IsMultiReg()) - { - treeNode->SetRegSpillFlagByIdx(GTF_SPILLED, currentRefPosition->getMultiRegIdx()); - } - if (spillAfter) + printf("-----------------------\n"); + printf("RESOLVING BB BOUNDARIES\n"); + printf("-----------------------\n"); + + printf("Resolution Candidates: "); + dumpConvertedVarSet(compiler, resolutionCandidateVars); + printf("\n"); + printf("Has %sCritical Edges\n\n", hasCriticalEdges ? "" : "No "); + + printf("Prior to Resolution\n"); + foreach_block(compiler, block) { - if (currentRefPosition->RegOptional()) + printf("\n" FMT_BB, block->bbNum); + if (block->hasEHBoundaryIn()) { - // This is a use of lclVar that is flagged as reg-optional - // by lower/codegen and marked for both reload and spillAfter. - // In this case we can avoid unnecessary reload and spill - // by setting reg on lclVar to REG_STK and reg on tree node - // to REG_NA. Codegen will generate the code by considering - // it as a contained memory operand. - // - // Note that varDsc->GetRegNum() is already to REG_STK above. - interval->physReg = REG_NA; - writeLocalReg(treeNode->AsLclVar(), interval->varNum, REG_NA); - treeNode->gtFlags &= ~GTF_SPILLED; - treeNode->SetContained(); - // We don't support RegOptional for multi-reg localvars. - assert(!treeNode->IsMultiReg()); + JITDUMP(" EH flow in"); } - else + if (block->hasEHBoundaryOut()) { - treeNode->gtFlags |= GTF_SPILL; - if (treeNode->IsMultiReg()) - { - treeNode->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); - } + JITDUMP(" EH flow out"); } + + printf("\nuse def in out\n"); + dumpConvertedVarSet(compiler, block->bbVarUse); + printf("\n"); + dumpConvertedVarSet(compiler, block->bbVarDef); + printf("\n"); + dumpConvertedVarSet(compiler, block->bbLiveIn); + printf("\n"); + dumpConvertedVarSet(compiler, block->bbLiveOut); + printf("\n"); + + dumpInVarToRegMap(block); + dumpOutVarToRegMap(block); } - } - else - { - assert(currentRefPosition->refType == RefTypeExpUse); - } - } - else if (spillAfter && !RefTypeIsUse(currentRefPosition->refType) && - (!treeNode->IsMultiReg() || treeNode->gtGetOp1()->IsMultiRegNode())) - { - // In the case of a pure def, don't bother spilling - just assign it to the - // stack. However, we need to remember that it was spilled. - // We can't do this in the case of a multi-reg node with a non-multireg source as - // we need the register to extract into. - assert(interval->isSpilled); - varDsc->SetRegNum(REG_STK); - interval->physReg = REG_NA; - if (treeNode != nullptr) - { - writeLocalReg(treeNode->AsLclVar(), interval->varNum, REG_NA); + printf("\n\n"); } - } - else // Not reload and Not pure-def that's spillAfter - { - if (currentRefPosition->copyReg || currentRefPosition->moveReg) - { - // For a copyReg or moveReg, we have two cases: - // - In the first case, we have a fixedReg - i.e. a register which the code - // generator is constrained to use. - // The code generator will generate the appropriate move to meet the requirement. - // - In the second case, we were forced to use a different register because of - // interference (or JitStressRegs). - // In this case, we generate a GT_COPY. - // In either case, we annotate the treeNode with the register in which the value - // currently lives. For moveReg, the homeReg is the new register (as assigned above). - // But for copyReg, the homeReg remains unchanged. +#endif // DEBUG - assert(treeNode != nullptr); - writeLocalReg(treeNode->AsLclVar(), interval->varNum, interval->physReg); + resolveEdges(); - if (currentRefPosition->copyReg) + // Verify register assignments on variables + unsigned lclNum; + LclVarDsc* varDsc; + for (lclNum = 0, varDsc = compiler->lvaTable; lclNum < compiler->lvaCount; lclNum++, varDsc++) + { + if (!isCandidateVar(varDsc)) { - homeReg = interval->physReg; + varDsc->SetRegNum(REG_STK); } else { - assert(interval->isSplit); - interval->physReg = assignedReg; - } + Interval* interval = getIntervalForLocalVar(varDsc->lvVarIndex); - if (!currentRefPosition->isFixedRegRef || currentRefPosition->moveReg) - { - // This is the second case, where we need to generate a copy - insertCopyOrReload(block, treeNode, currentRefPosition->getMultiRegIdx(), currentRefPosition); - } - } - else - { - interval->physReg = assignedReg; + // Determine initial position for parameters - if (!interval->isSpilled && !interval->isSplit) - { - if (varDsc->GetRegNum() != REG_STK) + if (varDsc->lvIsParam) { - // If the register assignments don't match, then this interval is split. - if (varDsc->GetRegNum() != assignedReg) + regMaskTP initialRegMask = interval->firstRefPosition->registerAssignment; + regNumber initialReg = (initialRegMask == RBM_NONE || interval->firstRefPosition->spillAfter) + ? REG_STK + : genRegNumFromMask(initialRegMask); + +#ifdef TARGET_ARM + if (varTypeIsMultiReg(varDsc)) { - setIntervalAsSplit(interval); - varDsc->SetRegNum(REG_STK); + // TODO-ARM-NYI: Map the hi/lo intervals back to lvRegNum and GetOtherReg() (these should NYI + // before this) + assert(!"Multi-reg types not yet supported"); } + else +#endif // TARGET_ARM + { + varDsc->SetArgInitReg(initialReg); + JITDUMP(" Set V%02u argument initial register to %s\n", lclNum, getRegName(initialReg)); + } + + // Stack args that are part of dependently-promoted structs should never be register candidates (see + // LinearScan::isRegCandidate). + assert(varDsc->lvIsRegArg || !compiler->lvaIsFieldOfDependentlyPromotedStruct(varDsc)); } - else - { - varDsc->SetRegNum(assignedReg); - } - } - } - if (spillAfter) - { - if (treeNode != nullptr) - { - treeNode->gtFlags |= GTF_SPILL; - if (treeNode->IsMultiReg()) + + // If lvRegNum is REG_STK, that means that either no register + // was assigned, or (more likely) that the same register was not + // used for all references. In that case, codegen gets the register + // from the tree node. + if (varDsc->GetRegNum() == REG_STK || interval->isSpilled || interval->isSplit) { - treeNode->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); + // For codegen purposes, we'll set lvRegNum to whatever register + // it's currently in as we go. + // However, we never mark an interval as lvRegister if it has either been spilled + // or split. + varDsc->lvRegister = false; + + // Skip any dead defs or exposed uses + // (first use exposed will only occur when there is no explicit initialization) + RefPosition* firstRefPosition = interval->firstRefPosition; + while ((firstRefPosition != nullptr) && (firstRefPosition->refType == RefTypeExpUse)) + { + firstRefPosition = firstRefPosition->nextRefPosition; + } + if (firstRefPosition == nullptr) + { + // Dead interval + varDsc->lvLRACandidate = false; + if (varDsc->lvRefCnt() == 0) + { + varDsc->lvOnFrame = false; + } + else + { + // We may encounter cases where a lclVar actually has no references, but + // a non-zero refCnt. For safety (in case this is some "hidden" lclVar that we're + // not correctly recognizing), we'll mark those as needing a stack location. + // TODO-Cleanup: Make this an assert if/when we correct the refCnt + // updating. + varDsc->lvOnFrame = true; + } + } + else + { + // If the interval was not spilled, it doesn't need a stack location. + if (!interval->isSpilled) + { + varDsc->lvOnFrame = false; + } + if (firstRefPosition->registerAssignment == RBM_NONE || firstRefPosition->spillAfter) + { + // Either this RefPosition is spilled, or regOptional or it is not a "real" def or use + assert( + firstRefPosition->spillAfter || firstRefPosition->RegOptional() || + (firstRefPosition->refType != RefTypeDef && firstRefPosition->refType != RefTypeUse)); + varDsc->SetRegNum(REG_STK); + } + else + { + varDsc->SetRegNum(firstRefPosition->assignedReg()); + } + } } - } - assert(interval->isSpilled); - interval->physReg = REG_NA; - varDsc->SetRegNum(REG_STK); - } - if (writeThru && (treeNode != nullptr)) - { - // This is a def of a write-thru EH var (only defs are marked 'writeThru'). - treeNode->gtFlags |= GTF_SPILL; - // We also mark writeThru defs that are not last-use with GTF_SPILLED to indicate that they are conceptually - // spilled and immediately "reloaded", i.e. the register remains live. - // Note that we can have a "last use" write that has no exposed uses in the standard - // (non-eh) control flow, but that may be used on an exception path. Hence the need - // to retain these defs, and to ensure that they write. - if (!currentRefPosition->lastUse) - { - treeNode->gtFlags |= GTF_SPILLED; - if (treeNode->IsMultiReg()) + else { - treeNode->SetRegSpillFlagByIdx(GTF_SPILLED, currentRefPosition->getMultiRegIdx()); + { + varDsc->lvRegister = true; + varDsc->lvOnFrame = false; + } +#ifdef DEBUG + regMaskTP registerAssignment = genRegMask(varDsc->GetRegNum()); + assert(!interval->isSpilled && !interval->isSplit); + RefPosition* refPosition = interval->firstRefPosition; + assert(refPosition != nullptr); + + while (refPosition != nullptr) + { + // All RefPositions must match, except for dead definitions, + // copyReg/moveReg and RefTypeExpUse positions + if (refPosition->registerAssignment != RBM_NONE && !refPosition->copyReg && + !refPosition->moveReg && refPosition->refType != RefTypeExpUse) + { + assert(refPosition->registerAssignment == registerAssignment); + } + refPosition = refPosition->nextRefPosition; + } +#endif // DEBUG } } } } - // Update the physRegRecord for the register, so that we know what vars are in - // regs at the block boundaries - RegRecord* physRegRecord = getRegisterRecord(homeReg); - if (spillAfter || currentRefPosition->lastUse) +#ifdef DEBUG + if (VERBOSE) { - interval->isActive = false; - interval->assignedReg = nullptr; - interval->physReg = REG_NA; - - updateAssignedInterval(physRegRecord, nullptr, interval->registerType); + printf("Trees after linear scan register allocator (LSRA)\n"); + compiler->fgDispBasicBlocks(true); } - else - { - interval->isActive = true; - interval->assignedReg = physRegRecord; - updateAssignedInterval(physRegRecord, interval, interval->registerType); - } -} + verifyFinalAllocation(); +#endif // DEBUG -void LinearScan::writeRegisters(RefPosition* currentRefPosition, GenTree* tree) -{ - lsraAssignRegToTree(tree, currentRefPosition->assignedReg(), currentRefPosition->getMultiRegIdx()); + compiler->raMarkStkVars(); + recordMaxSpill(); + + // TODO-CQ: Review this comment and address as needed. + // Change all unused promoted non-argument struct locals to a non-GC type (in this case TYP_INT) + // so that the gc tracking logic and lvMustInit logic will ignore them. + // Extract the code that does this from raAssignVars, and call it here. + // PRECONDITIONS: Ensure that lvPromoted is set on promoted structs, if and + // only if it is promoted on all paths. + // Call might be something like: + // compiler->BashUnusedStructLocals(); } +// //------------------------------------------------------------------------ -// insertCopyOrReload: Insert a copy in the case where a tree node value must be moved -// to a different register at the point of use (GT_COPY), or it is reloaded to a different register -// than the one it was spilled from (GT_RELOAD). +// insertMove: Insert a move of a lclVar with the given lclNum into the given block. // // Arguments: -// block - basic block in which GT_COPY/GT_RELOAD is inserted. -// tree - This is the node to copy or reload. -// Insert copy or reload node between this node and its parent. -// multiRegIdx - register position of tree node for which copy or reload is needed. -// refPosition - The RefPosition at which copy or reload will take place. -// -// Notes: -// The GT_COPY or GT_RELOAD will be inserted in the proper spot in execution order where the reload is to occur. -// -// For example, for this tree (numbers are execution order, lower is earlier and higher is later): -// -// +---------+----------+ -// | GT_ADD (3) | -// +---------+----------+ -// | -// / '\' -// / '\' -// / '\' -// +-------------------+ +----------------------+ -// | x (1) | "tree" | y (2) | -// +-------------------+ +----------------------+ -// -// generate this tree: -// -// +---------+----------+ -// | GT_ADD (4) | -// +---------+----------+ -// | -// / '\' -// / '\' -// / '\' -// +-------------------+ +----------------------+ -// | GT_RELOAD (3) | | y (2) | -// +-------------------+ +----------------------+ -// | -// +-------------------+ -// | x (1) | "tree" -// +-------------------+ -// -// Note in particular that the GT_RELOAD node gets inserted in execution order immediately before the parent of "tree", -// which seems a bit weird since normally a node's parent (in this case, the parent of "x", GT_RELOAD in the "after" -// picture) immediately follows all of its children (that is, normally the execution ordering is postorder). -// The ordering must be this weird "out of normal order" way because the "x" node is being spilled, probably -// because the expression in the tree represented above by "y" has high register requirements. We don't want -// to reload immediately, of course. So we put GT_RELOAD where the reload should actually happen. +// block - the BasicBlock into which the move will be inserted. +// insertionPoint - the instruction before which to insert the move +// lclNum - the lclNum of the var to be moved +// fromReg - the register from which the var is moving +// toReg - the register to which the var is moving // -// Note that GT_RELOAD is required when we reload to a different register than the one we spilled to. It can also be -// used if we reload to the same register. Normally, though, in that case we just mark the node with GTF_SPILLED, -// and the unspilling code automatically reuses the same register, and does the reload when it notices that flag -// when considering a node's operands. +// Return Value: +// None. // -void LinearScan::insertCopyOrReload(BasicBlock* block, GenTree* tree, unsigned multiRegIdx, RefPosition* refPosition) -{ - LIR::Range& blockRange = LIR::AsRange(block); - - LIR::Use treeUse; - bool foundUse = blockRange.TryGetUse(tree, &treeUse); - assert(foundUse); +// Notes: +// If insertionPoint is non-NULL, insert before that instruction; +// otherwise, insert "near" the end (prior to the branch, if any). +// If fromReg or toReg is REG_STK, then move from/to memory, respectively. - GenTree* parent = treeUse.User(); +void LinearScan::insertMove( + BasicBlock* block, GenTree* insertionPoint, unsigned lclNum, regNumber fromReg, regNumber toReg) +{ + LclVarDsc* varDsc = compiler->lvaTable + lclNum; + // the lclVar must be a register candidate + assert(isRegCandidate(varDsc)); + // One or both MUST be a register + assert(fromReg != REG_STK || toReg != REG_STK); + // They must not be the same register. + assert(fromReg != toReg); - genTreeOps oper; - if (refPosition->reload) - { - oper = GT_RELOAD; - } - else - { - oper = GT_COPY; + // This var can't be marked lvRegister now + varDsc->SetRegNum(REG_STK); - INTRACK_STATS(updateLsraStat(STAT_COPY_REG, block->bbNum)); - } + GenTree* src = compiler->gtNewLclvNode(lclNum, varDsc->TypeGet()); + SetLsraAdded(src); - // If the parent is a reload/copy node, then tree must be a multi-reg node - // that has already had one of its registers spilled. - // It is possible that one of its RefTypeDef positions got spilled and the next - // use of it requires it to be in a different register. + // There are three cases we need to handle: + // - We are loading a lclVar from the stack. + // - We are storing a lclVar to the stack. + // - We are copying a lclVar between registers. // - // In this case set the i'th position reg of reload/copy node to the reg allocated - // for copy/reload refPosition. Essentially a copy/reload node will have a reg - // for each multi-reg position of its child. If there is a valid reg in i'th - // position of GT_COPY or GT_RELOAD node then the corresponding result of its - // child needs to be copied or reloaded to that reg. - if (parent->IsCopyOrReload()) - { - noway_assert(parent->OperGet() == oper); - noway_assert(tree->IsMultiRegNode()); - GenTreeCopyOrReload* copyOrReload = parent->AsCopyOrReload(); - noway_assert(copyOrReload->GetRegNumByIdx(multiRegIdx) == REG_NA); - copyOrReload->SetRegNumByIdx(refPosition->assignedReg(), multiRegIdx); - } - else - { - // Create the new node, with "tree" as its only child. - var_types treeType = tree->TypeGet(); - - GenTreeCopyOrReload* newNode = new (compiler, oper) GenTreeCopyOrReload(oper, treeType, tree); - assert(refPosition->registerAssignment != RBM_NONE); - SetLsraAdded(newNode); - newNode->SetRegNumByIdx(refPosition->assignedReg(), multiRegIdx); - if (refPosition->copyReg) - { - // This is a TEMPORARY copy - assert(isCandidateLocalRef(tree) || tree->IsMultiRegLclVar()); - newNode->SetLastUse(multiRegIdx); - } - - // Insert the copy/reload after the spilled node and replace the use of the original node with a use - // of the copy/reload. - blockRange.InsertAfter(tree, newNode); - treeUse.ReplaceWith(compiler, newNode); - } -} + // In the first and second cases, the lclVar node will be marked with GTF_SPILLED and GTF_SPILL, respectively. + // It is up to the code generator to ensure that any necessary normalization is done when loading or storing the + // lclVar's value. + // + // In the third case, we generate GT_COPY(GT_LCL_VAR) and type each node with the normalized type of the lclVar. + // This is safe because a lclVar is always normalized once it is in a register. -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE -//------------------------------------------------------------------------ -// insertUpperVectorSave: Insert code to save the upper half of a vector that lives -// in a callee-save register at the point of a kill (the upper half is -// not preserved). -// -// Arguments: -// tree - This is the node before which we will insert the Save. -// It will be a call or some node that turns into a call. -// refPosition - The RefTypeUpperVectorSave RefPosition. -// upperInterval - The Interval for the upper half of the large vector lclVar. -// block - the BasicBlock containing the call. -// -void LinearScan::insertUpperVectorSave(GenTree* tree, - RefPosition* refPosition, - Interval* upperVectorInterval, - BasicBlock* block) -{ - JITDUMP("Inserting UpperVectorSave for RP #%d before %d.%s:\n", refPosition->rpNum, tree->gtTreeID, - GenTree::OpName(tree->gtOper)); - Interval* lclVarInterval = upperVectorInterval->relatedInterval; - assert(lclVarInterval->isLocalVar == true); - assert(refPosition->getInterval() == upperVectorInterval); - regNumber lclVarReg = lclVarInterval->physReg; - if (lclVarReg == REG_NA) + GenTree* dst = src; + if (fromReg == REG_STK) { - return; + src->gtFlags |= GTF_SPILLED; + src->SetRegNum(toReg); } - - LclVarDsc* varDsc = compiler->lvaTable + lclVarInterval->varNum; - assert(Compiler::varTypeNeedsPartialCalleeSave(varDsc->lvType)); - - // On Arm64, we must always have a register to save the upper half, - // while on x86 we can spill directly to memory. - regNumber spillReg = refPosition->assignedReg(); -#ifdef TARGET_ARM64 - bool spillToMem = refPosition->spillAfter; - assert(spillReg != REG_NA); -#else - bool spillToMem = (spillReg == REG_NA); - assert(!refPosition->spillAfter); -#endif - - LIR::Range& blockRange = LIR::AsRange(block); - - // Insert the save before the call. - - GenTree* saveLcl = compiler->gtNewLclvNode(lclVarInterval->varNum, varDsc->lvType); - saveLcl->SetRegNum(lclVarReg); - SetLsraAdded(saveLcl); - - GenTreeSIMD* simdNode = - new (compiler, GT_SIMD) GenTreeSIMD(LargeVectorSaveType, saveLcl, nullptr, SIMDIntrinsicUpperSave, - varDsc->GetSimdBaseJitType(), genTypeSize(varDsc->lvType)); - - if (simdNode->GetSimdBaseJitType() == CORINFO_TYPE_UNDEF) + else if (toReg == REG_STK) { - // There are a few scenarios where we can get a LCL_VAR which - // doesn't know the underlying baseType. In that scenario, we - // will just lie and say it is a float. Codegen doesn't actually - // care what the type is but this avoids an assert that would - // otherwise be fired from the more general checks that happen. - simdNode->SetSimdBaseJitType(CORINFO_TYPE_FLOAT); + src->gtFlags |= GTF_SPILL; + src->SetRegNum(fromReg); } + else + { + var_types movType = varDsc->GetRegisterType(); + src->gtType = movType; - SetLsraAdded(simdNode); - simdNode->SetRegNum(spillReg); - if (spillToMem) + dst = new (compiler, GT_COPY) GenTreeCopyOrReload(GT_COPY, movType, src); + // This is the new home of the lclVar - indicate that by clearing the GTF_VAR_DEATH flag. + // Note that if src is itself a lastUse, this will have no effect. + dst->gtFlags &= ~(GTF_VAR_DEATH); + src->SetRegNum(fromReg); + dst->SetRegNum(toReg); + SetLsraAdded(dst); + } + dst->SetUnusedValue(); + + LIR::Range treeRange = LIR::SeqTree(compiler, dst); + LIR::Range& blockRange = LIR::AsRange(block); + + if (insertionPoint != nullptr) { - simdNode->gtFlags |= GTF_SPILL; - upperVectorInterval->physReg = REG_NA; + blockRange.InsertBefore(insertionPoint, std::move(treeRange)); } else { - assert((genRegMask(spillReg) & RBM_FLT_CALLEE_SAVED) != RBM_NONE); - upperVectorInterval->physReg = spillReg; - } + // Put the copy at the bottom + GenTree* lastNode = blockRange.LastNode(); + if (block->bbJumpKind == BBJ_COND || block->bbJumpKind == BBJ_SWITCH) + { + noway_assert(!blockRange.IsEmpty()); - blockRange.InsertBefore(tree, LIR::SeqTree(compiler, simdNode)); - DISPTREE(simdNode); - JITDUMP("\n"); + GenTree* branch = lastNode; + assert(branch->OperIsConditionalJump() || branch->OperGet() == GT_SWITCH_TABLE || + branch->OperGet() == GT_SWITCH); + + blockRange.InsertBefore(branch, std::move(treeRange)); + } + else + { + // These block kinds don't have a branch at the end. + assert((lastNode == nullptr) || (!lastNode->OperIsConditionalJump() && + !lastNode->OperIs(GT_SWITCH_TABLE, GT_SWITCH, GT_RETURN, GT_RETFILT))); + blockRange.InsertAtEnd(std::move(treeRange)); + } + } } -//------------------------------------------------------------------------ -// insertUpperVectorRestore: Insert code to restore the upper half of a vector that has been partially spilled. -// -// Arguments: -// tree - This is the node for which we will insert the Restore. -// If non-null, it will be a use of the large vector lclVar. -// If null, the Restore will be added to the end of the block. -// upperVectorInterval - The Interval for the upper vector for the lclVar. -// block - the BasicBlock into which we will be inserting the code. -// -// Notes: -// In the case where 'tree' is non-null, we will insert the restore just prior to -// its use, in order to ensure the proper ordering. -// -void LinearScan::insertUpperVectorRestore(GenTree* tree, - RefPosition* refPosition, - Interval* upperVectorInterval, - BasicBlock* block) +void LinearScan::insertSwap( + BasicBlock* block, GenTree* insertionPoint, unsigned lclNum1, regNumber reg1, unsigned lclNum2, regNumber reg2) { - JITDUMP("Adding UpperVectorRestore for RP #%d ", refPosition->rpNum); - Interval* lclVarInterval = upperVectorInterval->relatedInterval; - assert(lclVarInterval->isLocalVar == true); - regNumber lclVarReg = lclVarInterval->physReg; - - // We should not call this method if the lclVar is not in a register (we should have simply marked the entire - // lclVar as spilled). - assert(lclVarReg != REG_NA); - LclVarDsc* varDsc = compiler->lvaTable + lclVarInterval->varNum; - assert(Compiler::varTypeNeedsPartialCalleeSave(varDsc->lvType)); +#ifdef DEBUG + if (VERBOSE) + { + const char* insertionPointString = "top"; + if (insertionPoint == nullptr) + { + insertionPointString = "bottom"; + } + printf(" " FMT_BB " %s: swap V%02u in %s with V%02u in %s\n", block->bbNum, insertionPointString, lclNum1, + getRegName(reg1), lclNum2, getRegName(reg2)); + } +#endif // DEBUG - GenTree* restoreLcl = nullptr; - restoreLcl = compiler->gtNewLclvNode(lclVarInterval->varNum, varDsc->lvType); - restoreLcl->SetRegNum(lclVarReg); - SetLsraAdded(restoreLcl); + LclVarDsc* varDsc1 = compiler->lvaTable + lclNum1; + LclVarDsc* varDsc2 = compiler->lvaTable + lclNum2; + assert(reg1 != REG_STK && reg1 != REG_NA && reg2 != REG_STK && reg2 != REG_NA); - GenTreeSIMD* simdNode = - new (compiler, GT_SIMD) GenTreeSIMD(varDsc->lvType, restoreLcl, nullptr, SIMDIntrinsicUpperRestore, - varDsc->GetSimdBaseJitType(), genTypeSize(varDsc->lvType)); + GenTree* lcl1 = compiler->gtNewLclvNode(lclNum1, varDsc1->TypeGet()); + lcl1->SetRegNum(reg1); + SetLsraAdded(lcl1); - if (simdNode->GetSimdBaseJitType() == CORINFO_TYPE_UNDEF) - { - // There are a few scenarios where we can get a LCL_VAR which - // doesn't know the underlying baseType. In that scenario, we - // will just lie and say it is a float. Codegen doesn't actually - // care what the type is but this avoids an assert that would - // otherwise be fired from the more general checks that happen. - simdNode->SetSimdBaseJitType(CORINFO_TYPE_FLOAT); - } + GenTree* lcl2 = compiler->gtNewLclvNode(lclNum2, varDsc2->TypeGet()); + lcl2->SetRegNum(reg2); + SetLsraAdded(lcl2); - regNumber restoreReg = upperVectorInterval->physReg; - SetLsraAdded(simdNode); + GenTree* swap = compiler->gtNewOperNode(GT_SWAP, TYP_VOID, lcl1, lcl2); + swap->SetRegNum(REG_NA); + SetLsraAdded(swap); - if (restoreReg == REG_NA) - { - // We need a stack location for this. - assert(lclVarInterval->isSpilled); -#ifdef TARGET_AMD64 - assert(refPosition->assignedReg() == REG_NA); - simdNode->gtFlags |= GTF_NOREG_AT_USE; -#else - simdNode->gtFlags |= GTF_SPILLED; - assert(refPosition->assignedReg() != REG_NA); - restoreReg = refPosition->assignedReg(); -#endif - } - simdNode->SetRegNum(restoreReg); + lcl1->gtNext = lcl2; + lcl2->gtPrev = lcl1; + lcl2->gtNext = swap; + swap->gtPrev = lcl2; + LIR::Range swapRange = LIR::SeqTree(compiler, swap); LIR::Range& blockRange = LIR::AsRange(block); - JITDUMP("Adding UpperVectorRestore "); - if (tree != nullptr) + + if (insertionPoint != nullptr) { - JITDUMP("before %d.%s:\n", tree->gtTreeID, GenTree::OpName(tree->gtOper)); - LIR::Use treeUse; - bool foundUse = blockRange.TryGetUse(tree, &treeUse); - assert(foundUse); - // We need to insert the restore prior to the use, not (necessarily) immediately after the lclVar. - blockRange.InsertBefore(treeUse.User(), LIR::SeqTree(compiler, simdNode)); + blockRange.InsertBefore(insertionPoint, std::move(swapRange)); } else { - JITDUMP("at end of " FMT_BB ":\n", block->bbNum); + // Put the copy at the bottom + // If there's a branch, make an embedded statement that executes just prior to the branch if (block->bbJumpKind == BBJ_COND || block->bbJumpKind == BBJ_SWITCH) { noway_assert(!blockRange.IsEmpty()); @@ -7089,4436 +7288,4440 @@ void LinearScan::insertUpperVectorRestore(GenTree* tree, assert(branch->OperIsConditionalJump() || branch->OperGet() == GT_SWITCH_TABLE || branch->OperGet() == GT_SWITCH); - blockRange.InsertBefore(branch, LIR::SeqTree(compiler, simdNode)); + blockRange.InsertBefore(branch, std::move(swapRange)); } else { assert(block->bbJumpKind == BBJ_NONE || block->bbJumpKind == BBJ_ALWAYS); - blockRange.InsertAtEnd(LIR::SeqTree(compiler, simdNode)); + blockRange.InsertAtEnd(std::move(swapRange)); } } - DISPTREE(simdNode); - JITDUMP("\n"); } -#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE //------------------------------------------------------------------------ -// initMaxSpill: Initializes the LinearScan members used to track the max number -// of concurrent spills. This is needed so that we can set the -// fields in Compiler, so that the code generator, in turn can -// allocate the right number of spill locations. +// getTempRegForResolution: Get a free register to use for resolution code. // // Arguments: -// None. +// fromBlock - The "from" block on the edge being resolved. +// toBlock - The "to"block on the edge +// type - the type of register required // // Return Value: -// None. +// Returns a register that is free on the given edge, or REG_NA if none is available. // -// Assumptions: -// This is called before any calls to updateMaxSpill(). +// Notes: +// It is up to the caller to check the return value, and to determine whether a register is +// available, and to handle that case appropriately. +// It is also up to the caller to cache the return value, as this is not cheap to compute. -void LinearScan::initMaxSpill() +regNumber LinearScan::getTempRegForResolution(BasicBlock* fromBlock, BasicBlock* toBlock, var_types type) { - needDoubleTmpForFPCall = false; - needFloatTmpForFPCall = false; - for (int i = 0; i < TYP_COUNT; i++) + // TODO-Throughput: This would be much more efficient if we add RegToVarMaps instead of VarToRegMaps + // and they would be more space-efficient as well. + VarToRegMap fromVarToRegMap = getOutVarToRegMap(fromBlock->bbNum); + VarToRegMap toVarToRegMap = getInVarToRegMap(toBlock->bbNum); + +#ifdef TARGET_ARM + regMaskTP freeRegs; + if (type == TYP_DOUBLE) { - maxSpill[i] = 0; - currentSpill[i] = 0; + // We have to consider all float registers for TYP_DOUBLE + freeRegs = allRegs(TYP_FLOAT); } -} + else + { + freeRegs = allRegs(type); + } +#else // !TARGET_ARM + regMaskTP freeRegs = allRegs(type); +#endif // !TARGET_ARM -//------------------------------------------------------------------------ -// recordMaxSpill: Sets the fields in Compiler for the max number of concurrent spills. -// (See the comment on initMaxSpill.) -// -// Arguments: -// None. -// -// Return Value: -// None. -// -// Assumptions: -// This is called after updateMaxSpill() has been called for all "real" -// RefPositions. +#ifdef DEBUG + if (getStressLimitRegs() == LSRA_LIMIT_SMALL_SET) + { + return REG_NA; + } +#endif // DEBUG + INDEBUG(freeRegs = stressLimitRegs(nullptr, freeRegs)); -void LinearScan::recordMaxSpill() -{ - // Note: due to the temp normalization process (see tmpNormalizeType) - // only a few types should actually be seen here. - JITDUMP("Recording the maximum number of concurrent spills:\n"); -#ifdef TARGET_X86 - var_types returnType = RegSet::tmpNormalizeType(compiler->info.compRetType); - if (needDoubleTmpForFPCall || (returnType == TYP_DOUBLE)) + // We are only interested in the variables that are live-in to the "to" block. + VarSetOps::Iter iter(compiler, toBlock->bbLiveIn); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex) && freeRegs != RBM_NONE) + { + regNumber fromReg = getVarReg(fromVarToRegMap, varIndex); + regNumber toReg = getVarReg(toVarToRegMap, varIndex); + assert(fromReg != REG_NA && toReg != REG_NA); + if (fromReg != REG_STK) + { + freeRegs &= ~genRegMask(fromReg, getIntervalForLocalVar(varIndex)->registerType); + } + if (toReg != REG_STK) + { + freeRegs &= ~genRegMask(toReg, getIntervalForLocalVar(varIndex)->registerType); + } + } + +#ifdef TARGET_ARM + if (type == TYP_DOUBLE) { - JITDUMP("Adding a spill temp for moving a double call/return value between xmm reg and x87 stack.\n"); - maxSpill[TYP_DOUBLE] += 1; + // Exclude any doubles for which the odd half isn't in freeRegs. + freeRegs = freeRegs & ((freeRegs << 1) & RBM_ALLDOUBLE); } - if (needFloatTmpForFPCall || (returnType == TYP_FLOAT)) +#endif + + if (freeRegs == RBM_NONE) { - JITDUMP("Adding a spill temp for moving a float call/return value between xmm reg and x87 stack.\n"); - maxSpill[TYP_FLOAT] += 1; + return REG_NA; } -#endif // TARGET_X86 - for (int i = 0; i < TYP_COUNT; i++) + else { - if (var_types(i) != RegSet::tmpNormalizeType(var_types(i))) - { - // Only normalized types should have anything in the maxSpill array. - // We assume here that if type 'i' does not normalize to itself, then - // nothing else normalizes to 'i', either. - assert(maxSpill[i] == 0); - } - if (maxSpill[i] != 0) - { - JITDUMP(" %s: %d\n", varTypeName(var_types(i)), maxSpill[i]); - compiler->codeGen->regSet.tmpPreAllocateTemps(var_types(i), maxSpill[i]); - } + regNumber tempReg = genRegNumFromMask(genFindLowestBit(freeRegs)); + return tempReg; } - JITDUMP("\n"); } +#ifdef TARGET_ARM //------------------------------------------------------------------------ -// updateMaxSpill: Update the maximum number of concurrent spills +// addResolutionForDouble: Add resolution move(s) for TYP_DOUBLE interval +// and update location. // // Arguments: -// refPosition - the current RefPosition being handled +// block - the BasicBlock into which the move will be inserted. +// insertionPoint - the instruction before which to insert the move +// sourceIntervals - maintains sourceIntervals[reg] which each 'reg' is associated with +// location - maintains location[reg] which is the location of the var that was originally in 'reg'. +// toReg - the register to which the var is moving +// fromReg - the register from which the var is moving +// resolveType - the type of resolution to be performed // // Return Value: // None. // -// Assumptions: -// The RefPosition has an associated interval (getInterval() will -// otherwise assert). -// // Notes: -// This is called for each "real" RefPosition during the writeback -// phase of LSRA. It keeps track of how many concurrently-live -// spills there are, and the largest number seen so far. - -void LinearScan::updateMaxSpill(RefPosition* refPosition) +// It inserts at least one move and updates incoming parameter 'location'. +// +void LinearScan::addResolutionForDouble(BasicBlock* block, + GenTree* insertionPoint, + Interval** sourceIntervals, + regNumberSmall* location, + regNumber toReg, + regNumber fromReg, + ResolveType resolveType) { - RefType refType = refPosition->refType; + regNumber secondHalfTargetReg = REG_NEXT(fromReg); + Interval* intervalToBeMoved1 = sourceIntervals[fromReg]; + Interval* intervalToBeMoved2 = sourceIntervals[secondHalfTargetReg]; -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - if ((refType == RefTypeUpperVectorSave) || (refType == RefTypeUpperVectorRestore)) + assert(!(intervalToBeMoved1 == nullptr && intervalToBeMoved2 == nullptr)); + + if (intervalToBeMoved1 != nullptr) { - Interval* interval = refPosition->getInterval(); - // If this is not an 'upperVector', it must be a tree temp that has been already - // (fully) spilled. - if (!interval->isUpperVector) + if (intervalToBeMoved1->registerType == TYP_DOUBLE) { - assert(interval->firstRefPosition->spillAfter); + // TYP_DOUBLE interval occupies a double register, i.e. two float registers. + assert(intervalToBeMoved2 == nullptr); + assert(genIsValidDoubleReg(toReg)); } else { - // The UpperVector RefPositions spill to the localVar's home location. - Interval* lclVarInterval = interval->relatedInterval; - assert(lclVarInterval->isSpilled || (!refPosition->spillAfter && !refPosition->reload)); + // TYP_FLOAT interval occupies 1st half of double register, i.e. 1st float register + assert(genIsValidFloatReg(toReg)); } - return; + addResolution(block, insertionPoint, intervalToBeMoved1, toReg, fromReg); + JITDUMP(" (%s)\n", resolveTypeName[resolveType]); + location[fromReg] = (regNumberSmall)toReg; } -#endif // !FEATURE_PARTIAL_SIMD_CALLEE_SAVE - if (refPosition->spillAfter || refPosition->reload || - (refPosition->RegOptional() && refPosition->assignedReg() == REG_NA)) - { - Interval* interval = refPosition->getInterval(); - if (!interval->isLocalVar) - { - GenTree* treeNode = refPosition->treeNode; - if (treeNode == nullptr) - { - assert(RefTypeIsUse(refType)); - treeNode = interval->firstRefPosition->treeNode; - } - assert(treeNode != nullptr); - - // The tmp allocation logic 'normalizes' types to a small number of - // types that need distinct stack locations from each other. - // Those types are currently gc refs, byrefs, <= 4 byte non-GC items, - // 8-byte non-GC items, and 16-byte or 32-byte SIMD vectors. - // LSRA is agnostic to those choices but needs - // to know what they are here. - var_types type; - if (!treeNode->IsMultiRegNode()) - { - type = getDefType(treeNode); - } - else - { - type = treeNode->GetRegTypeByIndex(refPosition->getMultiRegIdx()); - } - type = RegSet::tmpNormalizeType(type); + if (intervalToBeMoved2 != nullptr) + { + // TYP_FLOAT interval occupies 2nd half of double register. + assert(intervalToBeMoved2->registerType == TYP_FLOAT); + regNumber secondHalfTempReg = REG_NEXT(toReg); - if (refPosition->spillAfter && !refPosition->reload) - { - currentSpill[type]++; - if (currentSpill[type] > maxSpill[type]) - { - maxSpill[type] = currentSpill[type]; - } - } - else if (refPosition->reload) - { - assert(currentSpill[type] > 0); - currentSpill[type]--; - } - else if (refPosition->RegOptional() && refPosition->assignedReg() == REG_NA) - { - // A spill temp not getting reloaded into a reg because it is - // marked as allocate if profitable and getting used from its - // memory location. To properly account max spill for typ we - // decrement spill count. - assert(RefTypeIsUse(refType)); - assert(currentSpill[type] > 0); - currentSpill[type]--; - } - JITDUMP(" Max spill for %s is %d\n", varTypeName(type), maxSpill[type]); - } + addResolution(block, insertionPoint, intervalToBeMoved2, secondHalfTempReg, secondHalfTargetReg); + JITDUMP(" (%s)\n", resolveTypeName[resolveType]); + location[secondHalfTargetReg] = (regNumberSmall)secondHalfTempReg; } + + return; } +#endif // TARGET_ARM -// This is the final phase of register allocation. It writes the register assignments to -// the tree, and performs resolution across joins and backedges. +//------------------------------------------------------------------------ +// addResolution: Add a resolution move of the given interval // -void LinearScan::resolveRegisters() +// Arguments: +// block - the BasicBlock into which the move will be inserted. +// insertionPoint - the instruction before which to insert the move +// interval - the interval of the var to be moved +// toReg - the register to which the var is moving +// fromReg - the register from which the var is moving +// +// Return Value: +// None. +// +// Notes: +// For joins, we insert at the bottom (indicated by an insertionPoint +// of nullptr), while for splits we insert at the top. +// This is because for joins 'block' is a pred of the join, while for splits it is a succ. +// For critical edges, this function may be called twice - once to move from +// the source (fromReg), if any, to the stack, in which case toReg will be +// REG_STK, and we insert at the bottom (leave insertionPoint as nullptr). +// The next time, we want to move from the stack to the destination (toReg), +// in which case fromReg will be REG_STK, and we insert at the top. + +void LinearScan::addResolution( + BasicBlock* block, GenTree* insertionPoint, Interval* interval, regNumber toReg, regNumber fromReg) { - // Iterate over the tree and the RefPositions in lockstep - // - annotate the tree with register assignments by setting GetRegNum() or gtRegPair (for longs) - // on the tree node - // - track globally-live var locations - // - add resolution points at split/merge/critical points as needed +#ifdef DEBUG + const char* insertionPointString; + if (insertionPoint == nullptr) + { + // We can't add resolution to a register at the bottom of a block that has an EHBoundaryOut, + // except in the case of the "EH Dummy" resolution from the stack. + assert((block->bbNum > bbNumMaxBeforeResolution) || (fromReg == REG_STK) || + !blockInfo[block->bbNum].hasEHBoundaryOut); + insertionPointString = "bottom"; + } + else + { + // We can't add resolution at the top of a block that has an EHBoundaryIn, + // except in the case of the "EH Dummy" resolution to the stack. + assert((block->bbNum > bbNumMaxBeforeResolution) || (toReg == REG_STK) || + !blockInfo[block->bbNum].hasEHBoundaryIn); + insertionPointString = "top"; + } - // Need to use the same traversal order as the one that assigns the location numbers. + // We should never add resolution move inside BBCallAlwaysPairTail. + noway_assert(!block->isBBCallAlwaysPairTail()); - // Dummy RefPositions have been added at any split, join or critical edge, at the - // point where resolution may be required. These are located: - // - for a split, at the top of the non-adjacent block - // - for a join, at the bottom of the non-adjacent joining block - // - for a critical edge, at the top of the target block of each critical - // edge. - // Note that a target block may have multiple incoming critical or split edges - // - // These RefPositions record the expected location of the Interval at that point. - // At each branch, we identify the location of each liveOut interval, and check - // against the RefPositions at the target. +#endif // DEBUG - BasicBlock* block; - LsraLocation currentLocation = MinLocation; + JITDUMP(" " FMT_BB " %s: move V%02u from ", block->bbNum, insertionPointString, interval->varNum); + JITDUMP("%s to %s", getRegName(fromReg), getRegName(toReg)); - // Clear register assignments - these will be reestablished as lclVar defs (including RefTypeParamDefs) - // are encountered. - if (enregisterLocalVars) + insertMove(block, insertionPoint, interval->varNum, fromReg, toReg); + if (fromReg == REG_STK || toReg == REG_STK) { - for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) - { - RegRecord* physRegRecord = getRegisterRecord(reg); - Interval* assignedInterval = physRegRecord->assignedInterval; - if (assignedInterval != nullptr) - { - assignedInterval->assignedReg = nullptr; - assignedInterval->physReg = REG_NA; - } - physRegRecord->assignedInterval = nullptr; - physRegRecord->recentRefPosition = nullptr; - } + assert(interval->isSpilled); + } + else + { + // We should have already marked this as spilled or split. + assert((interval->isSpilled) || (interval->isSplit)); + } + + INTRACK_STATS(updateLsraStat(STAT_RESOLUTION_MOV, block->bbNum)); +} + +//------------------------------------------------------------------------ +// handleOutgoingCriticalEdges: Performs the necessary resolution on all critical edges that feed out of 'block' +// +// Arguments: +// block - the block with outgoing critical edges. +// +// Return Value: +// None.. +// +// Notes: +// For all outgoing critical edges (i.e. any successor of this block which is +// a join edge), if there are any conflicts, split the edge by adding a new block, +// and generate the resolution code into that block. - // Clear "recentRefPosition" for lclVar intervals - for (unsigned varIndex = 0; varIndex < compiler->lvaTrackedCount; varIndex++) - { - if (localVarIntervals[varIndex] != nullptr) - { - localVarIntervals[varIndex]->recentRefPosition = nullptr; - localVarIntervals[varIndex]->isActive = false; - } - else - { - assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); - } - } +void LinearScan::handleOutgoingCriticalEdges(BasicBlock* block) +{ + VARSET_TP outResolutionSet(VarSetOps::Intersection(compiler, block->bbLiveOut, resolutionCandidateVars)); + if (VarSetOps::IsEmpty(compiler, outResolutionSet)) + { + return; } + VARSET_TP sameResolutionSet(VarSetOps::MakeEmpty(compiler)); + VARSET_TP diffResolutionSet(VarSetOps::MakeEmpty(compiler)); - // handle incoming arguments and special temps - RefPositionIterator refPosIterator = refPositions.begin(); - RefPosition* currentRefPosition = &refPosIterator; + // Get the outVarToRegMap for this block + VarToRegMap outVarToRegMap = getOutVarToRegMap(block->bbNum); + unsigned succCount = block->NumSucc(compiler); + assert(succCount > 1); - if (enregisterLocalVars) + // First, determine the live regs at the end of this block so that we know what regs are + // available to copy into. + // Note that for this purpose we use the full live-out set, because we must ensure that + // even the registers that remain the same across the edge are preserved correctly. + regMaskTP liveOutRegs = RBM_NONE; + VarSetOps::Iter liveOutIter(compiler, block->bbLiveOut); + unsigned liveOutVarIndex = 0; + while (liveOutIter.NextElem(&liveOutVarIndex)) { - VarToRegMap entryVarToRegMap = inVarToRegMaps[compiler->fgFirstBB->bbNum]; - for (; refPosIterator != refPositions.end() && - (currentRefPosition->refType == RefTypeParamDef || currentRefPosition->refType == RefTypeZeroInit); - ++refPosIterator, currentRefPosition = &refPosIterator) + regNumber fromReg = getVarReg(outVarToRegMap, liveOutVarIndex); + if (fromReg != REG_STK) { - Interval* interval = currentRefPosition->getInterval(); - assert(interval != nullptr && interval->isLocalVar); - resolveLocalRef(nullptr, nullptr, currentRefPosition); - regNumber reg = REG_STK; - int varIndex = interval->getVarIndex(compiler); - - if (!currentRefPosition->spillAfter && currentRefPosition->registerAssignment != RBM_NONE) - { - reg = currentRefPosition->assignedReg(); - } - else - { - reg = REG_STK; - interval->isActive = false; - } - setVarReg(entryVarToRegMap, varIndex, reg); + regMaskTP fromRegMask = genRegMask(fromReg, getIntervalForLocalVar(liveOutVarIndex)->registerType); + liveOutRegs |= fromRegMask; } } - else + + // Next, if this blocks ends with a switch table, or for Arm64, ends with JCMP instruction, + // make sure to not copy into the registers that are consumed at the end of this block. + // + // Note: Only switches and JCMP (for Arm4) have input regs (and so can be fed by copies), so those + // are the only block-ending branches that need special handling. + regMaskTP consumedRegs = RBM_NONE; + if (block->bbJumpKind == BBJ_SWITCH) { - assert(refPosIterator == refPositions.end() || - (refPosIterator->refType != RefTypeParamDef && refPosIterator->refType != RefTypeZeroInit)); + // At this point, Lowering has transformed any non-switch-table blocks into + // cascading ifs. + GenTree* switchTable = LIR::AsRange(block).LastNode(); + assert(switchTable != nullptr && switchTable->OperGet() == GT_SWITCH_TABLE); + + consumedRegs = switchTable->gtRsvdRegs; + GenTree* op1 = switchTable->gtGetOp1(); + GenTree* op2 = switchTable->gtGetOp2(); + noway_assert(op1 != nullptr && op2 != nullptr); + assert(op1->GetRegNum() != REG_NA && op2->GetRegNum() != REG_NA); + // No floating point values, so no need to worry about the register type + // (i.e. for ARM32, where we used the genRegMask overload with a type). + assert(varTypeIsIntegralOrI(op1) && varTypeIsIntegralOrI(op2)); + consumedRegs |= genRegMask(op1->GetRegNum()); + consumedRegs |= genRegMask(op2->GetRegNum()); } - // write back assignments - for (block = startBlockSequence(); block != nullptr; block = moveToNextBlock()) +#ifdef TARGET_ARM64 + // Next, if this blocks ends with a JCMP, we have to make sure: + // 1. Not to copy into the register that JCMP uses + // e.g. JCMP w21, BRANCH + // 2. Not to copy into the source of JCMP's operand before it is consumed + // e.g. Should not use w0 since it will contain wrong value after resolution + // call METHOD + // ; mov w0, w19 <-- should not resolve in w0 here. + // mov w21, w0 + // JCMP w21, BRANCH + // 3. Not to modify the local variable it must consume + + // Note: GT_COPY has special handling in codegen and its generation is merged with the + // node that consumes its result. So both, the input and output regs of GT_COPY must be + // excluded from the set available for resolution. + LclVarDsc* jcmpLocalVarDsc = nullptr; + if (block->bbJumpKind == BBJ_COND) { - assert(curBBNum == block->bbNum); + GenTree* lastNode = LIR::AsRange(block).LastNode(); - if (enregisterLocalVars) + if (lastNode->OperIs(GT_JCMP)) { - // Record the var locations at the start of this block. - // (If it's fgFirstBB, we've already done that above, see entryVarToRegMap) + GenTree* op1 = lastNode->gtGetOp1(); + consumedRegs |= genRegMask(op1->GetRegNum()); - curBBStartLocation = currentRefPosition->nodeLocation; - if (block != compiler->fgFirstBB) + if (op1->OperIs(GT_COPY)) { - processBlockStartLocations(block); + GenTree* srcOp1 = op1->gtGetOp1(); + consumedRegs |= genRegMask(srcOp1->GetRegNum()); } - // Handle the DummyDefs, updating the incoming var location. - for (; refPosIterator != refPositions.end() && currentRefPosition->refType == RefTypeDummyDef; - ++refPosIterator, currentRefPosition = &refPosIterator) + if (op1->IsLocal()) { - assert(currentRefPosition->isIntervalRef()); - // Don't mark dummy defs as reload - currentRefPosition->reload = false; - resolveLocalRef(nullptr, nullptr, currentRefPosition); - regNumber reg; - if (currentRefPosition->registerAssignment != RBM_NONE) - { - reg = currentRefPosition->assignedReg(); - } - else - { - reg = REG_STK; - currentRefPosition->getInterval()->isActive = false; - } - setInVarRegForBB(curBBNum, currentRefPosition->getInterval()->varNum, reg); + GenTreeLclVarCommon* lcl = op1->AsLclVarCommon(); + jcmpLocalVarDsc = &compiler->lvaTable[lcl->GetLclNum()]; } } + } +#endif - // The next RefPosition should be for the block. Move past it. - assert(refPosIterator != refPositions.end()); - assert(currentRefPosition->refType == RefTypeBB); - ++refPosIterator; - currentRefPosition = &refPosIterator; - - // Handle the RefPositions for the block - for (; refPosIterator != refPositions.end() && currentRefPosition->refType != RefTypeBB && - currentRefPosition->refType != RefTypeDummyDef; - ++refPosIterator, currentRefPosition = &refPosIterator) - { - currentLocation = currentRefPosition->nodeLocation; - - // Ensure that the spill & copy info is valid. - // First, if it's reload, it must not be copyReg or moveReg - assert(!currentRefPosition->reload || (!currentRefPosition->copyReg && !currentRefPosition->moveReg)); - // If it's copyReg it must not be moveReg, and vice-versa - assert(!currentRefPosition->copyReg || !currentRefPosition->moveReg); + VarToRegMap sameVarToRegMap = sharedCriticalVarToRegMap; + regMaskTP sameWriteRegs = RBM_NONE; + regMaskTP diffReadRegs = RBM_NONE; - switch (currentRefPosition->refType) - { -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - case RefTypeUpperVectorSave: - case RefTypeUpperVectorRestore: -#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE - case RefTypeUse: - case RefTypeDef: - // These are the ones we're interested in - break; - case RefTypeKill: - case RefTypeFixedReg: - // These require no handling at resolution time - assert(currentRefPosition->referent != nullptr); - currentRefPosition->referent->recentRefPosition = currentRefPosition; - continue; - case RefTypeExpUse: - // Ignore the ExpUse cases - a RefTypeExpUse would only exist if the - // variable is dead at the entry to the next block. So we'll mark - // it as in its current location and resolution will take care of any - // mismatch. - assert(getNextBlock() == nullptr || - !VarSetOps::IsMember(compiler, getNextBlock()->bbLiveIn, - currentRefPosition->getInterval()->getVarIndex(compiler))); - currentRefPosition->referent->recentRefPosition = currentRefPosition; - continue; - case RefTypeKillGCRefs: - // No action to take at resolution time, and no interval to update recentRefPosition for. - continue; - case RefTypeDummyDef: - case RefTypeParamDef: - case RefTypeZeroInit: - // Should have handled all of these already - default: - unreached(); - break; - } - updateMaxSpill(currentRefPosition); - GenTree* treeNode = currentRefPosition->treeNode; + // For each var that may require resolution, classify them as: + // - in the same register at the end of this block and at each target (no resolution needed) + // - in different registers at different targets (resolve separately): + // diffResolutionSet + // - in the same register at each target at which it's live, but different from the end of + // this block. We may be able to resolve these as if it is "join", but only if they do not + // write to any registers that are read by those in the diffResolutionSet: + // sameResolutionSet -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - if (currentRefPosition->refType == RefTypeUpperVectorSave) - { - // The treeNode is a call or something that might become one. - noway_assert(treeNode != nullptr); - // If the associated interval is an UpperVector, this must be a RefPosition for a LargeVectorType - // LocalVar. - // Otherwise, this is a non-lclVar interval that has been spilled, and we don't need to do anything. - Interval* interval = currentRefPosition->getInterval(); - if (interval->isUpperVector) - { - Interval* localVarInterval = interval->relatedInterval; - if ((localVarInterval->physReg != REG_NA) && !localVarInterval->isPartiallySpilled) - { - // If the localVar is in a register, it must be in a register that is not trashed by - // the current node (otherwise it would have already been spilled). - assert((genRegMask(localVarInterval->physReg) & getKillSetForNode(treeNode)) == RBM_NONE); - // If we have allocated a register to spill it to, we will use that; otherwise, we will spill it - // to the stack. We can use as a temp register any non-arg caller-save register. - currentRefPosition->referent->recentRefPosition = currentRefPosition; - insertUpperVectorSave(treeNode, currentRefPosition, currentRefPosition->getInterval(), block); - localVarInterval->isPartiallySpilled = true; - } - } - else - { - // This is a non-lclVar interval that must have been spilled. - assert(!currentRefPosition->getInterval()->isLocalVar); - assert(currentRefPosition->getInterval()->firstRefPosition->spillAfter); - } + VarSetOps::Iter outResolutionSetIter(compiler, outResolutionSet); + unsigned outResolutionSetVarIndex = 0; + while (outResolutionSetIter.NextElem(&outResolutionSetVarIndex)) + { + regNumber fromReg = getVarReg(outVarToRegMap, outResolutionSetVarIndex); + bool maybeSameLivePaths = false; + bool liveOnlyAtSplitEdge = true; + regNumber sameToReg = REG_NA; + for (unsigned succIndex = 0; succIndex < succCount; succIndex++) + { + BasicBlock* succBlock = block->GetSucc(succIndex, compiler); + if (!VarSetOps::IsMember(compiler, succBlock->bbLiveIn, outResolutionSetVarIndex)) + { + maybeSameLivePaths = true; continue; } - else if (currentRefPosition->refType == RefTypeUpperVectorRestore) + else if (liveOnlyAtSplitEdge) { - // Since we don't do partial restores of tree temp intervals, this must be an upperVector. - Interval* interval = currentRefPosition->getInterval(); - Interval* localVarInterval = interval->relatedInterval; - assert(interval->isUpperVector && (localVarInterval != nullptr)); - if (localVarInterval->physReg != REG_NA) - { - assert(localVarInterval->isPartiallySpilled); - assert((localVarInterval->assignedReg != nullptr) && - (localVarInterval->assignedReg->regNum == localVarInterval->physReg) && - (localVarInterval->assignedReg->assignedInterval == localVarInterval)); - insertUpperVectorRestore(treeNode, currentRefPosition, interval, block); - } - localVarInterval->isPartiallySpilled = false; + // Is the var live only at those target blocks which are connected by a split edge to this block + liveOnlyAtSplitEdge = ((succBlock->bbPreds->flNext == nullptr) && (succBlock != compiler->fgFirstBB)); } -#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE - // Most uses won't actually need to be recorded (they're on the def). - // In those cases, treeNode will be nullptr. - if (treeNode == nullptr) + regNumber toReg = getVarReg(getInVarToRegMap(succBlock->bbNum), outResolutionSetVarIndex); + if (sameToReg == REG_NA) + { + sameToReg = toReg; + continue; + } + if (toReg == sameToReg) { - // This is either a use, a dead def, or a field of a struct - Interval* interval = currentRefPosition->getInterval(); - assert(currentRefPosition->refType == RefTypeUse || - currentRefPosition->registerAssignment == RBM_NONE || interval->isStructField || - interval->IsUpperVector()); - - // TODO-Review: Need to handle the case where any of the struct fields - // are reloaded/spilled at this use - assert(!interval->isStructField || - (currentRefPosition->reload == false && currentRefPosition->spillAfter == false)); - - if (interval->isLocalVar && !interval->isStructField) - { - LclVarDsc* varDsc = interval->getLocalVar(compiler); - - // This must be a dead definition. We need to mark the lclVar - // so that it's not considered a candidate for lvRegister, as - // this dead def will have to go to the stack. - assert(currentRefPosition->refType == RefTypeDef); - varDsc->SetRegNum(REG_STK); - } continue; } + sameToReg = REG_NA; + break; + } - assert(currentRefPosition->isIntervalRef()); - if (currentRefPosition->getInterval()->isInternal) + // Check for the cases where we can't write to a register. + // We only need to check for these cases if sameToReg is an actual register (not REG_STK). + if (sameToReg != REG_NA && sameToReg != REG_STK) + { + // If there's a path on which this var isn't live, it may use the original value in sameToReg. + // In this case, sameToReg will be in the liveOutRegs of this block. + // Similarly, if sameToReg is in sameWriteRegs, it has already been used (i.e. for a lclVar that's + // live only at another target), and we can't copy another lclVar into that reg in this block. + regMaskTP sameToRegMask = + genRegMask(sameToReg, getIntervalForLocalVar(outResolutionSetVarIndex)->registerType); + if (maybeSameLivePaths && + (((sameToRegMask & liveOutRegs) != RBM_NONE) || ((sameToRegMask & sameWriteRegs) != RBM_NONE))) { - treeNode->gtRsvdRegs |= currentRefPosition->registerAssignment; + sameToReg = REG_NA; } - else + // If this register is busy because it is used by a switch table at the end of the block + // (or for Arm64, it is consumed by JCMP), we can't do the copy in this block since we can't + // insert it after the switch (or for Arm64, can't insert and overwrite the operand/source + // of operand of JCMP). + if ((sameToRegMask & consumedRegs) != RBM_NONE) { - writeRegisters(currentRefPosition, treeNode); + sameToReg = REG_NA; + } - if (treeNode->OperIs(GT_LCL_VAR, GT_STORE_LCL_VAR) && currentRefPosition->getInterval()->isLocalVar) - { - resolveLocalRef(block, treeNode->AsLclVar(), currentRefPosition); - } +#ifdef TARGET_ARM64 + if (jcmpLocalVarDsc && (jcmpLocalVarDsc->lvVarIndex == outResolutionSetVarIndex)) + { + sameToReg = REG_NA; + } +#endif - // Mark spill locations on temps - // (local vars are handled in resolveLocalRef, above) - // Note that the tree node will be changed from GTF_SPILL to GTF_SPILLED - // in codegen, taking care of the "reload" case for temps - else if (currentRefPosition->spillAfter || (currentRefPosition->nextRefPosition != nullptr && - currentRefPosition->nextRefPosition->moveReg)) - { - if (treeNode != nullptr) - { - if (currentRefPosition->spillAfter) - { - treeNode->gtFlags |= GTF_SPILL; + // If the var is live only at those blocks connected by a split edge and not live-in at some of the + // target blocks, we will resolve it the same way as if it were in diffResolutionSet and resolution + // will be deferred to the handling of split edges, which means copy will only be at those target(s). + // + // Another way to achieve similar resolution for vars live only at split edges is by removing them + // from consideration up-front but it requires that we traverse those edges anyway to account for + // the registers that must note be overwritten. + if (liveOnlyAtSplitEdge && maybeSameLivePaths) + { + sameToReg = REG_NA; + } + } - // If this is a constant interval that is reusing a pre-existing value, we actually need - // to generate the value at this point in order to spill it. - if (treeNode->IsReuseRegVal()) - { - treeNode->ResetReuseRegVal(); - } + if (sameToReg == REG_NA) + { + VarSetOps::AddElemD(compiler, diffResolutionSet, outResolutionSetVarIndex); + if (fromReg != REG_STK) + { + diffReadRegs |= genRegMask(fromReg, getIntervalForLocalVar(outResolutionSetVarIndex)->registerType); + } + } + else if (sameToReg != fromReg) + { + VarSetOps::AddElemD(compiler, sameResolutionSet, outResolutionSetVarIndex); + setVarReg(sameVarToRegMap, outResolutionSetVarIndex, sameToReg); + if (sameToReg != REG_STK) + { + sameWriteRegs |= genRegMask(sameToReg, getIntervalForLocalVar(outResolutionSetVarIndex)->registerType); + } + } + } - // In case of multi-reg node, also set spill flag on the - // register specified by multi-reg index of current RefPosition. - // Note that the spill flag on treeNode indicates that one or - // more its allocated registers are in that state. - if (treeNode->IsMultiRegCall()) - { - GenTreeCall* call = treeNode->AsCall(); - call->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); - } -#if FEATURE_ARG_SPLIT - else if (treeNode->OperIsPutArgSplit()) - { - GenTreePutArgSplit* splitArg = treeNode->AsPutArgSplit(); - splitArg->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); - } -#ifdef TARGET_ARM - else if (treeNode->OperIsMultiRegOp()) - { - GenTreeMultiRegOp* multiReg = treeNode->AsMultiRegOp(); - multiReg->SetRegSpillFlagByIdx(GTF_SPILL, currentRefPosition->getMultiRegIdx()); - } -#endif // TARGET_ARM -#endif // FEATURE_ARG_SPLIT - } + if (!VarSetOps::IsEmpty(compiler, sameResolutionSet)) + { + if ((sameWriteRegs & diffReadRegs) != RBM_NONE) + { + // We cannot split the "same" and "diff" regs if the "same" set writes registers + // that must be read by the "diff" set. (Note that when these are done as a "batch" + // we carefully order them to ensure all the input regs are read before they are + // overwritten.) + VarSetOps::UnionD(compiler, diffResolutionSet, sameResolutionSet); + VarSetOps::ClearD(compiler, sameResolutionSet); + } + else + { + // For any vars in the sameResolutionSet, we can simply add the move at the end of "block". + resolveEdge(block, nullptr, ResolveSharedCritical, sameResolutionSet); + } + } + if (!VarSetOps::IsEmpty(compiler, diffResolutionSet)) + { + for (unsigned succIndex = 0; succIndex < succCount; succIndex++) + { + BasicBlock* succBlock = block->GetSucc(succIndex, compiler); - // If the value is reloaded or moved to a different register, we need to insert - // a node to hold the register to which it should be reloaded - RefPosition* nextRefPosition = currentRefPosition->nextRefPosition; - noway_assert(nextRefPosition != nullptr); - if (INDEBUG(alwaysInsertReload() ||) - nextRefPosition->assignedReg() != currentRefPosition->assignedReg()) - { -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - // Note that we asserted above that this is an Interval RefPosition. - Interval* currentInterval = currentRefPosition->getInterval(); - if (!currentInterval->isUpperVector && nextRefPosition->refType == RefTypeUpperVectorSave) - { - // The currentRefPosition is a spill of a tree temp. - // These have no associated Restore, as we always spill if the vector is - // in a register when this is encountered. - // The nextRefPosition we're interested in (where we may need to insert a - // reload or flag as GTF_NOREG_AT_USE) is the subsequent RefPosition. - assert(!currentInterval->isLocalVar); - nextRefPosition = nextRefPosition->nextRefPosition; - assert(nextRefPosition->refType != RefTypeUpperVectorSave); - } - // UpperVector intervals may have unique assignments at each reference. - if (!currentInterval->isUpperVector) -#endif - { - if (nextRefPosition->assignedReg() != REG_NA) - { - insertCopyOrReload(block, treeNode, currentRefPosition->getMultiRegIdx(), - nextRefPosition); - } - else - { - assert(nextRefPosition->RegOptional()); + // Any "diffResolutionSet" resolution for a block with no other predecessors will be handled later + // as split resolution. + if ((succBlock->bbPreds->flNext == nullptr) && (succBlock != compiler->fgFirstBB)) + { + continue; + } - // In case of tree temps, if def is spilled and use didn't - // get a register, set a flag on tree node to be treated as - // contained at the point of its use. - if (currentRefPosition->spillAfter && currentRefPosition->refType == RefTypeDef && - nextRefPosition->refType == RefTypeUse) - { - assert(nextRefPosition->treeNode == nullptr); - treeNode->gtFlags |= GTF_NOREG_AT_USE; - } - } - } - } - } + // Now collect the resolution set for just this edge, if any. + // Check only the vars in diffResolutionSet that are live-in to this successor. + VarToRegMap succInVarToRegMap = getInVarToRegMap(succBlock->bbNum); + VARSET_TP edgeResolutionSet(VarSetOps::Intersection(compiler, diffResolutionSet, succBlock->bbLiveIn)); + VarSetOps::Iter iter(compiler, edgeResolutionSet); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex)) + { + regNumber fromReg = getVarReg(outVarToRegMap, varIndex); + regNumber toReg = getVarReg(succInVarToRegMap, varIndex); - // We should never have to "spill after" a temp use, since - // they're single use - else + if (fromReg == toReg) + { + VarSetOps::RemoveElemD(compiler, edgeResolutionSet, varIndex); + } + } + if (!VarSetOps::IsEmpty(compiler, edgeResolutionSet)) + { + // For EH vars, we can always safely load them from the stack into the target for this block, + // so if we have only EH vars, we'll do that instead of splitting the edge. + if ((compiler->compHndBBtabCount > 0) && VarSetOps::IsSubset(compiler, edgeResolutionSet, exceptVars)) + { + GenTree* insertionPoint = LIR::AsRange(succBlock).FirstNonPhiNode(); + VarSetOps::Iter edgeSetIter(compiler, edgeResolutionSet); + unsigned edgeVarIndex = 0; + while (edgeSetIter.NextElem(&edgeVarIndex)) { - unreached(); + regNumber toReg = getVarReg(succInVarToRegMap, edgeVarIndex); + setVarReg(succInVarToRegMap, edgeVarIndex, REG_STK); + if (toReg != REG_STK) + { + Interval* interval = getIntervalForLocalVar(edgeVarIndex); + assert(interval->isWriteThru); + addResolution(succBlock, insertionPoint, interval, toReg, REG_STK); + JITDUMP(" (EHvar)\n"); + } } } + else + { + resolveEdge(block, succBlock, ResolveCritical, edgeResolutionSet); + } } } + } +} - if (enregisterLocalVars) +//------------------------------------------------------------------------ +// resolveEdges: Perform resolution across basic block edges +// +// Arguments: +// None. +// +// Return Value: +// None. +// +// Notes: +// Traverse the basic blocks. +// - If this block has a single predecessor that is not the immediately +// preceding block, perform any needed 'split' resolution at the beginning of this block +// - Otherwise if this block has critical incoming edges, handle them. +// - If this block has a single successor that has multiple predecesors, perform any needed +// 'join' resolution at the end of this block. +// Note that a block may have both 'split' or 'critical' incoming edge(s) and 'join' outgoing +// edges. + +void LinearScan::resolveEdges() +{ + JITDUMP("RESOLVING EDGES\n"); + + // The resolutionCandidateVars set was initialized with all the lclVars that are live-in to + // any block. We now intersect that set with any lclVars that ever spilled or split. + // If there are no candidates for resoultion, simply return. + + VarSetOps::IntersectionD(compiler, resolutionCandidateVars, splitOrSpilledVars); + if (VarSetOps::IsEmpty(compiler, resolutionCandidateVars)) + { + return; + } + + BasicBlock *block, *prevBlock = nullptr; + + // Handle all the critical edges first. + // We will try to avoid resolution across critical edges in cases where all the critical-edge + // targets of a block have the same home. We will then split the edges only for the + // remaining mismatches. We visit the out-edges, as that allows us to share the moves that are + // common among all the targets. + + if (hasCriticalEdges) + { + foreach_block(compiler, block) { - processBlockEndLocations(block); + if (block->bbNum > bbNumMaxBeforeResolution) + { + // This is a new block added during resolution - we don't need to visit these now. + continue; + } + if (blockInfo[block->bbNum].hasCriticalOutEdge) + { + handleOutgoingCriticalEdges(block); + } + prevBlock = block; } } - if (enregisterLocalVars) + prevBlock = nullptr; + foreach_block(compiler, block) { -#ifdef DEBUG - if (VERBOSE) + if (block->bbNum > bbNumMaxBeforeResolution) { - printf("-----------------------\n"); - printf("RESOLVING BB BOUNDARIES\n"); - printf("-----------------------\n"); + // This is a new block added during resolution - we don't need to visit these now. + continue; + } - printf("Resolution Candidates: "); - dumpConvertedVarSet(compiler, resolutionCandidateVars); - printf("\n"); - printf("Has %sCritical Edges\n\n", hasCriticalEdges ? "" : "No "); + unsigned succCount = block->NumSucc(compiler); + BasicBlock* uniquePredBlock = block->GetUniquePred(compiler); - printf("Prior to Resolution\n"); - foreach_block(compiler, block) + // First, if this block has a single predecessor, + // we may need resolution at the beginning of this block. + // This may be true even if it's the block we used for starting locations, + // if a variable was spilled. + VARSET_TP inResolutionSet(VarSetOps::Intersection(compiler, block->bbLiveIn, resolutionCandidateVars)); + if (!VarSetOps::IsEmpty(compiler, inResolutionSet)) + { + if (uniquePredBlock != nullptr) { - printf("\n" FMT_BB, block->bbNum); - if (block->hasEHBoundaryIn()) - { - JITDUMP(" EH flow in"); - } - if (block->hasEHBoundaryOut()) + // We may have split edges during critical edge resolution, and in the process split + // a non-critical edge as well. + // It is unlikely that we would ever have more than one of these in sequence (indeed, + // I don't think it's possible), but there's no need to assume that it can't. + while (uniquePredBlock->bbNum > bbNumMaxBeforeResolution) { - JITDUMP(" EH flow out"); + uniquePredBlock = uniquePredBlock->GetUniquePred(compiler); + noway_assert(uniquePredBlock != nullptr); } - - printf("\nuse def in out\n"); - dumpConvertedVarSet(compiler, block->bbVarUse); - printf("\n"); - dumpConvertedVarSet(compiler, block->bbVarDef); - printf("\n"); - dumpConvertedVarSet(compiler, block->bbLiveIn); - printf("\n"); - dumpConvertedVarSet(compiler, block->bbLiveOut); - printf("\n"); - - dumpInVarToRegMap(block); - dumpOutVarToRegMap(block); + resolveEdge(uniquePredBlock, block, ResolveSplit, inResolutionSet); } - - printf("\n\n"); } -#endif // DEBUG - resolveEdges(); + // Finally, if this block has a single successor: + // - and that has at least one other predecessor (otherwise we will do the resolution at the + // top of the successor), + // - and that is not the target of a critical edge (otherwise we've already handled it) + // we may need resolution at the end of this block. - // Verify register assignments on variables - unsigned lclNum; - LclVarDsc* varDsc; - for (lclNum = 0, varDsc = compiler->lvaTable; lclNum < compiler->lvaCount; lclNum++, varDsc++) + if (succCount == 1) { - if (!isCandidateVar(varDsc)) + BasicBlock* succBlock = block->GetSucc(0, compiler); + if (succBlock->GetUniquePred(compiler) == nullptr) { - varDsc->SetRegNum(REG_STK); + VARSET_TP outResolutionSet( + VarSetOps::Intersection(compiler, succBlock->bbLiveIn, resolutionCandidateVars)); + if (!VarSetOps::IsEmpty(compiler, outResolutionSet)) + { + resolveEdge(block, succBlock, ResolveJoin, outResolutionSet); + } } - else - { - Interval* interval = getIntervalForLocalVar(varDsc->lvVarIndex); + } + } - // Determine initial position for parameters + // Now, fixup the mapping for any blocks that were adding for edge splitting. + // See the comment prior to the call to fgSplitEdge() in resolveEdge(). + // Note that we could fold this loop in with the checking code below, but that + // would only improve the debug case, and would clutter up the code somewhat. + if (compiler->fgBBNumMax > bbNumMaxBeforeResolution) + { + foreach_block(compiler, block) + { + if (block->bbNum > bbNumMaxBeforeResolution) + { + // There may be multiple blocks inserted when we split. But we must always have exactly + // one path (i.e. all blocks must be single-successor and single-predecessor), + // and only one block along the path may be non-empty. + // Note that we may have a newly-inserted block that is empty, but which connects + // two non-resolution blocks. This happens when an edge is split that requires it. - if (varDsc->lvIsParam) + BasicBlock* succBlock = block; + do { - regMaskTP initialRegMask = interval->firstRefPosition->registerAssignment; - regNumber initialReg = (initialRegMask == RBM_NONE || interval->firstRefPosition->spillAfter) - ? REG_STK - : genRegNumFromMask(initialRegMask); - -#ifdef TARGET_ARM - if (varTypeIsMultiReg(varDsc)) - { - // TODO-ARM-NYI: Map the hi/lo intervals back to lvRegNum and GetOtherReg() (these should NYI - // before this) - assert(!"Multi-reg types not yet supported"); - } - else -#endif // TARGET_ARM - { - varDsc->SetArgInitReg(initialReg); - JITDUMP(" Set V%02u argument initial register to %s\n", lclNum, getRegName(initialReg)); - } - - // Stack args that are part of dependently-promoted structs should never be register candidates (see - // LinearScan::isRegCandidate). - assert(varDsc->lvIsRegArg || !compiler->lvaIsFieldOfDependentlyPromotedStruct(varDsc)); - } + succBlock = succBlock->GetUniqueSucc(); + noway_assert(succBlock != nullptr); + } while ((succBlock->bbNum > bbNumMaxBeforeResolution) && succBlock->isEmpty()); - // If lvRegNum is REG_STK, that means that either no register - // was assigned, or (more likely) that the same register was not - // used for all references. In that case, codegen gets the register - // from the tree node. - if (varDsc->GetRegNum() == REG_STK || interval->isSpilled || interval->isSplit) + BasicBlock* predBlock = block; + do { - // For codegen purposes, we'll set lvRegNum to whatever register - // it's currently in as we go. - // However, we never mark an interval as lvRegister if it has either been spilled - // or split. - varDsc->lvRegister = false; + predBlock = predBlock->GetUniquePred(compiler); + noway_assert(predBlock != nullptr); + } while ((predBlock->bbNum > bbNumMaxBeforeResolution) && predBlock->isEmpty()); - // Skip any dead defs or exposed uses - // (first use exposed will only occur when there is no explicit initialization) - RefPosition* firstRefPosition = interval->firstRefPosition; - while ((firstRefPosition != nullptr) && (firstRefPosition->refType == RefTypeExpUse)) - { - firstRefPosition = firstRefPosition->nextRefPosition; - } - if (firstRefPosition == nullptr) - { - // Dead interval - varDsc->lvLRACandidate = false; - if (varDsc->lvRefCnt() == 0) - { - varDsc->lvOnFrame = false; - } - else - { - // We may encounter cases where a lclVar actually has no references, but - // a non-zero refCnt. For safety (in case this is some "hidden" lclVar that we're - // not correctly recognizing), we'll mark those as needing a stack location. - // TODO-Cleanup: Make this an assert if/when we correct the refCnt - // updating. - varDsc->lvOnFrame = true; - } + unsigned succBBNum = succBlock->bbNum; + unsigned predBBNum = predBlock->bbNum; + if (block->isEmpty()) + { + // For the case of the empty block, find the non-resolution block (succ or pred). + if (predBBNum > bbNumMaxBeforeResolution) + { + assert(succBBNum <= bbNumMaxBeforeResolution); + predBBNum = 0; } else { - // If the interval was not spilled, it doesn't need a stack location. - if (!interval->isSpilled) - { - varDsc->lvOnFrame = false; - } - if (firstRefPosition->registerAssignment == RBM_NONE || firstRefPosition->spillAfter) - { - // Either this RefPosition is spilled, or regOptional or it is not a "real" def or use - assert( - firstRefPosition->spillAfter || firstRefPosition->RegOptional() || - (firstRefPosition->refType != RefTypeDef && firstRefPosition->refType != RefTypeUse)); - varDsc->SetRegNum(REG_STK); - } - else - { - varDsc->SetRegNum(firstRefPosition->assignedReg()); - } + succBBNum = 0; } } else { - { - varDsc->lvRegister = true; - varDsc->lvOnFrame = false; - } -#ifdef DEBUG - regMaskTP registerAssignment = genRegMask(varDsc->GetRegNum()); - assert(!interval->isSpilled && !interval->isSplit); - RefPosition* refPosition = interval->firstRefPosition; - assert(refPosition != nullptr); + assert((succBBNum <= bbNumMaxBeforeResolution) && (predBBNum <= bbNumMaxBeforeResolution)); + } + SplitEdgeInfo info = {predBBNum, succBBNum}; + getSplitBBNumToTargetBBNumMap()->Set(block->bbNum, info); - while (refPosition != nullptr) + // Set both the live-in and live-out to the live-in of the successor (by construction liveness + // doesn't change in a split block). + VarSetOps::Assign(compiler, block->bbLiveIn, succBlock->bbLiveIn); + VarSetOps::Assign(compiler, block->bbLiveOut, succBlock->bbLiveIn); + } + } + } + +#ifdef DEBUG + // Make sure the varToRegMaps match up on all edges. + bool foundMismatch = false; + foreach_block(compiler, block) + { + if (block->isEmpty() && block->bbNum > bbNumMaxBeforeResolution) + { + continue; + } + VarToRegMap toVarToRegMap = getInVarToRegMap(block->bbNum); + for (flowList* pred = block->bbPreds; pred != nullptr; pred = pred->flNext) + { + BasicBlock* predBlock = pred->getBlock(); + VarToRegMap fromVarToRegMap = getOutVarToRegMap(predBlock->bbNum); + VarSetOps::Iter iter(compiler, block->bbLiveIn); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex)) + { + regNumber fromReg = getVarReg(fromVarToRegMap, varIndex); + regNumber toReg = getVarReg(toVarToRegMap, varIndex); + if (fromReg != toReg) + { + Interval* interval = getIntervalForLocalVar(varIndex); + // The fromReg and toReg may not match for a write-thru interval where the toReg is + // REG_STK, since the stack value is always valid for that case (so no move is needed). + if (!interval->isWriteThru || (toReg != REG_STK)) { - // All RefPositions must match, except for dead definitions, - // copyReg/moveReg and RefTypeExpUse positions - if (refPosition->registerAssignment != RBM_NONE && !refPosition->copyReg && - !refPosition->moveReg && refPosition->refType != RefTypeExpUse) + if (!foundMismatch) { - assert(refPosition->registerAssignment == registerAssignment); + foundMismatch = true; + printf("Found mismatched var locations after resolution!\n"); } - refPosition = refPosition->nextRefPosition; + printf(" V%02u: " FMT_BB " to " FMT_BB ": %s to %s\n", interval->varNum, predBlock->bbNum, + block->bbNum, getRegName(fromReg), getRegName(toReg)); } -#endif // DEBUG } } } } - -#ifdef DEBUG - if (VERBOSE) - { - printf("Trees after linear scan register allocator (LSRA)\n"); - compiler->fgDispBasicBlocks(true); - } - - verifyFinalAllocation(); -#endif // DEBUG - - compiler->raMarkStkVars(); - recordMaxSpill(); - - // TODO-CQ: Review this comment and address as needed. - // Change all unused promoted non-argument struct locals to a non-GC type (in this case TYP_INT) - // so that the gc tracking logic and lvMustInit logic will ignore them. - // Extract the code that does this from raAssignVars, and call it here. - // PRECONDITIONS: Ensure that lvPromoted is set on promoted structs, if and - // only if it is promoted on all paths. - // Call might be something like: - // compiler->BashUnusedStructLocals(); + assert(!foundMismatch); +#endif + JITDUMP("\n"); } -// //------------------------------------------------------------------------ -// insertMove: Insert a move of a lclVar with the given lclNum into the given block. +// resolveEdge: Perform the specified type of resolution between two blocks. // // Arguments: -// block - the BasicBlock into which the move will be inserted. -// insertionPoint - the instruction before which to insert the move -// lclNum - the lclNum of the var to be moved -// fromReg - the register from which the var is moving -// toReg - the register to which the var is moving +// fromBlock - the block from which the edge originates +// toBlock - the block at which the edge terminates +// resolveType - the type of resolution to be performed +// liveSet - the set of tracked lclVar indices which may require resolution // // Return Value: // None. // +// Assumptions: +// The caller must have performed the analysis to determine the type of the edge. +// // Notes: -// If insertionPoint is non-NULL, insert before that instruction; -// otherwise, insert "near" the end (prior to the branch, if any). -// If fromReg or toReg is REG_STK, then move from/to memory, respectively. +// This method emits the correctly ordered moves necessary to place variables in the +// correct registers across a Split, Join or Critical edge. +// In order to avoid overwriting register values before they have been moved to their +// new home (register/stack), it first does the register-to-stack moves (to free those +// registers), then the register to register moves, ensuring that the target register +// is free before the move, and then finally the stack to register moves. -void LinearScan::insertMove( - BasicBlock* block, GenTree* insertionPoint, unsigned lclNum, regNumber fromReg, regNumber toReg) +void LinearScan::resolveEdge(BasicBlock* fromBlock, + BasicBlock* toBlock, + ResolveType resolveType, + VARSET_VALARG_TP liveSet) { - LclVarDsc* varDsc = compiler->lvaTable + lclNum; - // the lclVar must be a register candidate - assert(isRegCandidate(varDsc)); - // One or both MUST be a register - assert(fromReg != REG_STK || toReg != REG_STK); - // They must not be the same register. - assert(fromReg != toReg); + VarToRegMap fromVarToRegMap = getOutVarToRegMap(fromBlock->bbNum); + VarToRegMap toVarToRegMap; + if (resolveType == ResolveSharedCritical) + { + toVarToRegMap = sharedCriticalVarToRegMap; + } + else + { + toVarToRegMap = getInVarToRegMap(toBlock->bbNum); + } - // This var can't be marked lvRegister now - varDsc->SetRegNum(REG_STK); + // The block to which we add the resolution moves depends on the resolveType + BasicBlock* block; + switch (resolveType) + { + case ResolveJoin: + case ResolveSharedCritical: + block = fromBlock; + break; + case ResolveSplit: + block = toBlock; + break; + case ResolveCritical: + // fgSplitEdge may add one or two BasicBlocks. It returns the block that splits + // the edge from 'fromBlock' and 'toBlock', but if it inserts that block right after + // a block with a fall-through it will have to create another block to handle that edge. + // These new blocks can be mapped to existing blocks in order to correctly handle + // the calls to recordVarLocationsAtStartOfBB() from codegen. That mapping is handled + // in resolveEdges(), after all the edge resolution has been done (by calling this + // method for each edge). + block = compiler->fgSplitEdge(fromBlock, toBlock); - GenTree* src = compiler->gtNewLclvNode(lclNum, varDsc->TypeGet()); - SetLsraAdded(src); + // Split edges are counted against fromBlock. + INTRACK_STATS(updateLsraStat(STAT_SPLIT_EDGE, fromBlock->bbNum)); + break; + default: + unreached(); + break; + } - // There are three cases we need to handle: - // - We are loading a lclVar from the stack. - // - We are storing a lclVar to the stack. - // - We are copying a lclVar between registers. - // - // In the first and second cases, the lclVar node will be marked with GTF_SPILLED and GTF_SPILL, respectively. - // It is up to the code generator to ensure that any necessary normalization is done when loading or storing the - // lclVar's value. - // - // In the third case, we generate GT_COPY(GT_LCL_VAR) and type each node with the normalized type of the lclVar. - // This is safe because a lclVar is always normalized once it is in a register. +#ifndef TARGET_XARCH + // We record tempregs for beginning and end of each block. + // For amd64/x86 we only need a tempReg for float - we'll use xchg for int. + // TODO-Throughput: It would be better to determine the tempRegs on demand, but the code below + // modifies the varToRegMaps so we don't have all the correct registers at the time + // we need to get the tempReg. + regNumber tempRegInt = + (resolveType == ResolveSharedCritical) ? REG_NA : getTempRegForResolution(fromBlock, toBlock, TYP_INT); +#endif // !TARGET_XARCH + regNumber tempRegFlt = REG_NA; +#ifdef TARGET_ARM + regNumber tempRegDbl = REG_NA; +#endif + if ((compiler->compFloatingPointUsed) && (resolveType != ResolveSharedCritical)) + { +#ifdef TARGET_ARM + // Try to reserve a double register for TYP_DOUBLE and use it for TYP_FLOAT too if available. + tempRegDbl = getTempRegForResolution(fromBlock, toBlock, TYP_DOUBLE); + if (tempRegDbl != REG_NA) + { + tempRegFlt = tempRegDbl; + } + else +#endif // TARGET_ARM + { + tempRegFlt = getTempRegForResolution(fromBlock, toBlock, TYP_FLOAT); + } + } + + regMaskTP targetRegsToDo = RBM_NONE; + regMaskTP targetRegsReady = RBM_NONE; + regMaskTP targetRegsFromStack = RBM_NONE; + + // The following arrays capture the location of the registers as they are moved: + // - location[reg] gives the current location of the var that was originally in 'reg'. + // (Note that a var may be moved more than once.) + // - source[reg] gives the original location of the var that needs to be moved to 'reg'. + // For example, if a var is in rax and needs to be moved to rsi, then we would start with: + // location[rax] == rax + // source[rsi] == rax -- this doesn't change + // Then, if for some reason we need to move it temporary to rbx, we would have: + // location[rax] == rbx + // Once we have completed the move, we will have: + // location[rax] == REG_NA + // This indicates that the var originally in rax is now in its target register. + + regNumberSmall location[REG_COUNT]; + C_ASSERT(sizeof(char) == sizeof(regNumberSmall)); // for memset to work + memset(location, REG_NA, REG_COUNT); + regNumberSmall source[REG_COUNT]; + memset(source, REG_NA, REG_COUNT); + + // What interval is this register associated with? + // (associated with incoming reg) + Interval* sourceIntervals[REG_COUNT]; + memset(&sourceIntervals, 0, sizeof(sourceIntervals)); - GenTree* dst = src; - if (fromReg == REG_STK) - { - src->gtFlags |= GTF_SPILLED; - src->SetRegNum(toReg); - } - else if (toReg == REG_STK) + // Intervals for vars that need to be loaded from the stack + Interval* stackToRegIntervals[REG_COUNT]; + memset(&stackToRegIntervals, 0, sizeof(stackToRegIntervals)); + + // Get the starting insertion point for the "to" resolution + GenTree* insertionPoint = nullptr; + if (resolveType == ResolveSplit || resolveType == ResolveCritical) { - src->gtFlags |= GTF_SPILL; - src->SetRegNum(fromReg); + insertionPoint = LIR::AsRange(block).FirstNonPhiNode(); } - else - { - var_types movType = varDsc->GetRegisterType(); - src->gtType = movType; - dst = new (compiler, GT_COPY) GenTreeCopyOrReload(GT_COPY, movType, src); - // This is the new home of the lclVar - indicate that by clearing the GTF_VAR_DEATH flag. - // Note that if src is itself a lastUse, this will have no effect. - dst->gtFlags &= ~(GTF_VAR_DEATH); - src->SetRegNum(fromReg); - dst->SetRegNum(toReg); - SetLsraAdded(dst); + // If this is an edge between EH regions, we may have "extra" live-out EH vars. + // If we are adding resolution at the end of the block, we need to create "virtual" moves + // for these so that their registers are freed and can be reused. + if ((resolveType == ResolveJoin) && (compiler->compHndBBtabCount > 0)) + { + VARSET_TP extraLiveSet(VarSetOps::Diff(compiler, block->bbLiveOut, toBlock->bbLiveIn)); + VarSetOps::IntersectionD(compiler, extraLiveSet, registerCandidateVars); + VarSetOps::Iter iter(compiler, extraLiveSet); + unsigned extraVarIndex = 0; + while (iter.NextElem(&extraVarIndex)) + { + Interval* interval = getIntervalForLocalVar(extraVarIndex); + assert(interval->isWriteThru); + regNumber fromReg = getVarReg(fromVarToRegMap, extraVarIndex); + if (fromReg != REG_STK) + { + addResolution(block, insertionPoint, interval, REG_STK, fromReg); + JITDUMP(" (EH DUMMY)\n"); + setVarReg(fromVarToRegMap, extraVarIndex, REG_STK); + } + } } - dst->SetUnusedValue(); - LIR::Range treeRange = LIR::SeqTree(compiler, dst); - LIR::Range& blockRange = LIR::AsRange(block); + // First: + // - Perform all moves from reg to stack (no ordering needed on these) + // - For reg to reg moves, record the current location, associating their + // source location with the target register they need to go into + // - For stack to reg moves (done last, no ordering needed between them) + // record the interval associated with the target reg + // TODO-Throughput: We should be looping over the liveIn and liveOut registers, since + // that will scale better than the live variables - if (insertionPoint != nullptr) - { - blockRange.InsertBefore(insertionPoint, std::move(treeRange)); - } - else + VarSetOps::Iter iter(compiler, liveSet); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex)) { - // Put the copy at the bottom - GenTree* lastNode = blockRange.LastNode(); - if (block->bbJumpKind == BBJ_COND || block->bbJumpKind == BBJ_SWITCH) + Interval* interval = getIntervalForLocalVar(varIndex); + regNumber fromReg = getVarReg(fromVarToRegMap, varIndex); + regNumber toReg = getVarReg(toVarToRegMap, varIndex); + if (fromReg == toReg) { - noway_assert(!blockRange.IsEmpty()); + continue; + } + if (interval->isWriteThru && (toReg == REG_STK)) + { + // We don't actually move a writeThru var back to the stack, as its stack value is always valid. + // However, if this is a Join edge (i.e. the move is happening at the bottom of the block), + // and it is a "normal" flow edge, we will go ahead and generate a mov instruction, which will be + // a NOP but will cause the variable to be removed from being live in the register. + if ((resolveType == ResolveSplit) || block->hasEHBoundaryOut()) + { + continue; + } + } + // For Critical edges, the location will not change on either side of the edge, + // since we'll add a new block to do the move. + if (resolveType == ResolveSplit) + { + setVarReg(toVarToRegMap, varIndex, fromReg); + } + else if (resolveType == ResolveJoin || resolveType == ResolveSharedCritical) + { + setVarReg(fromVarToRegMap, varIndex, toReg); + } - GenTree* branch = lastNode; - assert(branch->OperIsConditionalJump() || branch->OperGet() == GT_SWITCH_TABLE || - branch->OperGet() == GT_SWITCH); + assert(fromReg < UCHAR_MAX && toReg < UCHAR_MAX); - blockRange.InsertBefore(branch, std::move(treeRange)); + if (fromReg == REG_STK) + { + stackToRegIntervals[toReg] = interval; + targetRegsFromStack |= genRegMask(toReg); + } + else if (toReg == REG_STK) + { + // Do the reg to stack moves now + addResolution(block, insertionPoint, interval, REG_STK, fromReg); + JITDUMP(" (%s)\n", + (interval->isWriteThru && (toReg == REG_STK)) ? "EH DUMMY" : resolveTypeName[resolveType]); } else { - // These block kinds don't have a branch at the end. - assert((lastNode == nullptr) || (!lastNode->OperIsConditionalJump() && - !lastNode->OperIs(GT_SWITCH_TABLE, GT_SWITCH, GT_RETURN, GT_RETFILT))); - blockRange.InsertAtEnd(std::move(treeRange)); + location[fromReg] = (regNumberSmall)fromReg; + source[toReg] = (regNumberSmall)fromReg; + sourceIntervals[fromReg] = interval; + targetRegsToDo |= genRegMask(toReg); } } -} -void LinearScan::insertSwap( - BasicBlock* block, GenTree* insertionPoint, unsigned lclNum1, regNumber reg1, unsigned lclNum2, regNumber reg2) -{ -#ifdef DEBUG - if (VERBOSE) + // REGISTER to REGISTER MOVES + + // First, find all the ones that are ready to move now + regMaskTP targetCandidates = targetRegsToDo; + while (targetCandidates != RBM_NONE) { - const char* insertionPointString = "top"; - if (insertionPoint == nullptr) + regMaskTP targetRegMask = genFindLowestBit(targetCandidates); + targetCandidates &= ~targetRegMask; + regNumber targetReg = genRegNumFromMask(targetRegMask); + if (location[targetReg] == REG_NA) { - insertionPointString = "bottom"; +#ifdef TARGET_ARM + regNumber sourceReg = (regNumber)source[targetReg]; + Interval* interval = sourceIntervals[sourceReg]; + if (interval->registerType == TYP_DOUBLE) + { + // For ARM32, make sure that both of the float halves of the double register are available. + assert(genIsValidDoubleReg(targetReg)); + regNumber anotherHalfRegNum = REG_NEXT(targetReg); + if (location[anotherHalfRegNum] == REG_NA) + { + targetRegsReady |= targetRegMask; + } + } + else +#endif // TARGET_ARM + { + targetRegsReady |= targetRegMask; + } } - printf(" " FMT_BB " %s: swap V%02u in %s with V%02u in %s\n", block->bbNum, insertionPointString, lclNum1, - getRegName(reg1), lclNum2, getRegName(reg2)); } -#endif // DEBUG - - LclVarDsc* varDsc1 = compiler->lvaTable + lclNum1; - LclVarDsc* varDsc2 = compiler->lvaTable + lclNum2; - assert(reg1 != REG_STK && reg1 != REG_NA && reg2 != REG_STK && reg2 != REG_NA); - - GenTree* lcl1 = compiler->gtNewLclvNode(lclNum1, varDsc1->TypeGet()); - lcl1->SetRegNum(reg1); - SetLsraAdded(lcl1); - - GenTree* lcl2 = compiler->gtNewLclvNode(lclNum2, varDsc2->TypeGet()); - lcl2->SetRegNum(reg2); - SetLsraAdded(lcl2); - - GenTree* swap = compiler->gtNewOperNode(GT_SWAP, TYP_VOID, lcl1, lcl2); - swap->SetRegNum(REG_NA); - SetLsraAdded(swap); - - lcl1->gtNext = lcl2; - lcl2->gtPrev = lcl1; - lcl2->gtNext = swap; - swap->gtPrev = lcl2; - - LIR::Range swapRange = LIR::SeqTree(compiler, swap); - LIR::Range& blockRange = LIR::AsRange(block); - if (insertionPoint != nullptr) - { - blockRange.InsertBefore(insertionPoint, std::move(swapRange)); - } - else + // Perform reg to reg moves + while (targetRegsToDo != RBM_NONE) { - // Put the copy at the bottom - // If there's a branch, make an embedded statement that executes just prior to the branch - if (block->bbJumpKind == BBJ_COND || block->bbJumpKind == BBJ_SWITCH) + while (targetRegsReady != RBM_NONE) + { + regMaskTP targetRegMask = genFindLowestBit(targetRegsReady); + targetRegsToDo &= ~targetRegMask; + targetRegsReady &= ~targetRegMask; + regNumber targetReg = genRegNumFromMask(targetRegMask); + assert(location[targetReg] != targetReg); + assert(targetReg < REG_COUNT); + regNumber sourceReg = (regNumber)source[targetReg]; + assert(sourceReg < REG_COUNT); + regNumber fromReg = (regNumber)location[sourceReg]; + // stack to reg movs should be done last as part of "targetRegsFromStack" + assert(fromReg < REG_STK); + Interval* interval = sourceIntervals[sourceReg]; + assert(interval != nullptr); + addResolution(block, insertionPoint, interval, targetReg, fromReg); + JITDUMP(" (%s)\n", resolveTypeName[resolveType]); + sourceIntervals[sourceReg] = nullptr; + location[sourceReg] = REG_NA; + regMaskTP fromRegMask = genRegMask(fromReg); + + // Do we have a free targetReg? + if (fromReg == sourceReg) + { + if (source[fromReg] != REG_NA && ((targetRegsFromStack & fromRegMask) != fromRegMask)) + { + targetRegsReady |= fromRegMask; +#ifdef TARGET_ARM + if (genIsValidDoubleReg(fromReg)) + { + // Ensure that either: + // - the Interval targeting fromReg is not double, or + // - the other half of the double is free. + Interval* otherInterval = sourceIntervals[source[fromReg]]; + regNumber upperHalfReg = REG_NEXT(fromReg); + if ((otherInterval->registerType == TYP_DOUBLE) && (location[upperHalfReg] != REG_NA)) + { + targetRegsReady &= ~fromRegMask; + } + } + } + else if (genIsValidFloatReg(fromReg) && !genIsValidDoubleReg(fromReg)) + { + // We may have freed up the other half of a double where the lower half + // was already free. + regNumber lowerHalfReg = REG_PREV(fromReg); + regNumber lowerHalfSrcReg = (regNumber)source[lowerHalfReg]; + regNumber lowerHalfSrcLoc = (regNumber)location[lowerHalfReg]; + regMaskTP lowerHalfRegMask = genRegMask(lowerHalfReg); + // Necessary conditions: + // - There is a source register for this reg (lowerHalfSrcReg != REG_NA) + // - It is currently free (lowerHalfSrcLoc == REG_NA) + // - The source interval isn't yet completed (sourceIntervals[lowerHalfSrcReg] != nullptr) + // - It's not in the ready set ((targetRegsReady & lowerHalfRegMask) == + // RBM_NONE) + // - It's not resolved from stack ((targetRegsFromStack & lowerHalfRegMask) != + // lowerHalfRegMask) + if ((lowerHalfSrcReg != REG_NA) && (lowerHalfSrcLoc == REG_NA) && + (sourceIntervals[lowerHalfSrcReg] != nullptr) && + ((targetRegsReady & lowerHalfRegMask) == RBM_NONE) && + ((targetRegsFromStack & lowerHalfRegMask) != lowerHalfRegMask)) + { + // This must be a double interval, otherwise it would be in targetRegsReady, or already + // completed. + assert(sourceIntervals[lowerHalfSrcReg]->registerType == TYP_DOUBLE); + targetRegsReady |= lowerHalfRegMask; + } +#endif // TARGET_ARM + } + } + } + if (targetRegsToDo != RBM_NONE) { - noway_assert(!blockRange.IsEmpty()); + regMaskTP targetRegMask = genFindLowestBit(targetRegsToDo); + regNumber targetReg = genRegNumFromMask(targetRegMask); - GenTree* branch = blockRange.LastNode(); - assert(branch->OperIsConditionalJump() || branch->OperGet() == GT_SWITCH_TABLE || - branch->OperGet() == GT_SWITCH); + // Is it already there due to other moves? + // If not, move it to the temp reg, OR swap it with another register + regNumber sourceReg = (regNumber)source[targetReg]; + regNumber fromReg = (regNumber)location[sourceReg]; + if (targetReg == fromReg) + { + targetRegsToDo &= ~targetRegMask; + } + else + { + regNumber tempReg = REG_NA; + bool useSwap = false; + if (emitter::isFloatReg(targetReg)) + { +#ifdef TARGET_ARM + if (sourceIntervals[fromReg]->registerType == TYP_DOUBLE) + { + // ARM32 requires a double temp register for TYP_DOUBLE. + tempReg = tempRegDbl; + } + else +#endif // TARGET_ARM + tempReg = tempRegFlt; + } +#ifdef TARGET_XARCH + else + { + useSwap = true; + } +#else // !TARGET_XARCH - blockRange.InsertBefore(branch, std::move(swapRange)); - } - else - { - assert(block->bbJumpKind == BBJ_NONE || block->bbJumpKind == BBJ_ALWAYS); - blockRange.InsertAtEnd(std::move(swapRange)); - } - } -} + else + { + tempReg = tempRegInt; + } -//------------------------------------------------------------------------ -// getTempRegForResolution: Get a free register to use for resolution code. -// -// Arguments: -// fromBlock - The "from" block on the edge being resolved. -// toBlock - The "to"block on the edge -// type - the type of register required -// -// Return Value: -// Returns a register that is free on the given edge, or REG_NA if none is available. -// -// Notes: -// It is up to the caller to check the return value, and to determine whether a register is -// available, and to handle that case appropriately. -// It is also up to the caller to cache the return value, as this is not cheap to compute. +#endif // !TARGET_XARCH + if (useSwap || tempReg == REG_NA) + { + // First, we have to figure out the destination register for what's currently in fromReg, + // so that we can find its sourceInterval. + regNumber otherTargetReg = REG_NA; -regNumber LinearScan::getTempRegForResolution(BasicBlock* fromBlock, BasicBlock* toBlock, var_types type) -{ - // TODO-Throughput: This would be much more efficient if we add RegToVarMaps instead of VarToRegMaps - // and they would be more space-efficient as well. - VarToRegMap fromVarToRegMap = getOutVarToRegMap(fromBlock->bbNum); - VarToRegMap toVarToRegMap = getInVarToRegMap(toBlock->bbNum); + // By chance, is fromReg going where it belongs? + if (location[source[fromReg]] == targetReg) + { + otherTargetReg = fromReg; + // If we can swap, we will be done with otherTargetReg as well. + // Otherwise, we'll spill it to the stack and reload it later. + if (useSwap) + { + regMaskTP fromRegMask = genRegMask(fromReg); + targetRegsToDo &= ~fromRegMask; + } + } + else + { + // Look at the remaining registers from targetRegsToDo (which we expect to be relatively + // small at this point) to find out what's currently in targetReg. + regMaskTP mask = targetRegsToDo; + while (mask != RBM_NONE && otherTargetReg == REG_NA) + { + regMaskTP nextRegMask = genFindLowestBit(mask); + regNumber nextReg = genRegNumFromMask(nextRegMask); + mask &= ~nextRegMask; + if (location[source[nextReg]] == targetReg) + { + otherTargetReg = nextReg; + } + } + } + assert(otherTargetReg != REG_NA); -#ifdef TARGET_ARM - regMaskTP freeRegs; - if (type == TYP_DOUBLE) - { - // We have to consider all float registers for TYP_DOUBLE - freeRegs = allRegs(TYP_FLOAT); - } - else - { - freeRegs = allRegs(type); - } -#else // !TARGET_ARM - regMaskTP freeRegs = allRegs(type); -#endif // !TARGET_ARM + if (useSwap) + { + // Generate a "swap" of fromReg and targetReg + insertSwap(block, insertionPoint, sourceIntervals[source[otherTargetReg]]->varNum, targetReg, + sourceIntervals[sourceReg]->varNum, fromReg); + location[sourceReg] = REG_NA; + location[source[otherTargetReg]] = (regNumberSmall)fromReg; -#ifdef DEBUG - if (getStressLimitRegs() == LSRA_LIMIT_SMALL_SET) - { - return REG_NA; - } -#endif // DEBUG - INDEBUG(freeRegs = stressLimitRegs(nullptr, freeRegs)); + INTRACK_STATS(updateLsraStat(STAT_RESOLUTION_MOV, block->bbNum)); + } + else + { + // Spill "targetReg" to the stack and add its eventual target (otherTargetReg) + // to "targetRegsFromStack", which will be handled below. + // NOTE: This condition is very rare. Setting COMPlus_JitStressRegs=0x203 + // has been known to trigger it in JIT SH. - // We are only interested in the variables that are live-in to the "to" block. - VarSetOps::Iter iter(compiler, toBlock->bbLiveIn); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex) && freeRegs != RBM_NONE) - { - regNumber fromReg = getVarReg(fromVarToRegMap, varIndex); - regNumber toReg = getVarReg(toVarToRegMap, varIndex); - assert(fromReg != REG_NA && toReg != REG_NA); - if (fromReg != REG_STK) - { - freeRegs &= ~genRegMask(fromReg, getIntervalForLocalVar(varIndex)->registerType); - } - if (toReg != REG_STK) - { - freeRegs &= ~genRegMask(toReg, getIntervalForLocalVar(varIndex)->registerType); - } - } + // First, spill "otherInterval" from targetReg to the stack. + Interval* otherInterval = sourceIntervals[source[otherTargetReg]]; + setIntervalAsSpilled(otherInterval); + addResolution(block, insertionPoint, otherInterval, REG_STK, targetReg); + JITDUMP(" (%s)\n", resolveTypeName[resolveType]); + location[source[otherTargetReg]] = REG_STK; -#ifdef TARGET_ARM - if (type == TYP_DOUBLE) - { - // Exclude any doubles for which the odd half isn't in freeRegs. - freeRegs = freeRegs & ((freeRegs << 1) & RBM_ALLDOUBLE); - } -#endif + regMaskTP otherTargetRegMask = genRegMask(otherTargetReg); + targetRegsFromStack |= otherTargetRegMask; + stackToRegIntervals[otherTargetReg] = otherInterval; + targetRegsToDo &= ~otherTargetRegMask; - if (freeRegs == RBM_NONE) - { - return REG_NA; - } - else - { - regNumber tempReg = genRegNumFromMask(genFindLowestBit(freeRegs)); - return tempReg; - } -} + // Now, move the interval that is going to targetReg. + addResolution(block, insertionPoint, sourceIntervals[sourceReg], targetReg, fromReg); + JITDUMP(" (%s)\n", resolveTypeName[resolveType]); + location[sourceReg] = REG_NA; + // Add its "fromReg" to "targetRegsReady", only if: + // - It was one of the target register we originally determined. + // - It is not the eventual target (otherTargetReg) because its + // value will be retrieved from STK. + if (source[fromReg] != REG_NA && fromReg != otherTargetReg) + { + regMaskTP fromRegMask = genRegMask(fromReg); + targetRegsReady |= fromRegMask; #ifdef TARGET_ARM -//------------------------------------------------------------------------ -// addResolutionForDouble: Add resolution move(s) for TYP_DOUBLE interval -// and update location. -// -// Arguments: -// block - the BasicBlock into which the move will be inserted. -// insertionPoint - the instruction before which to insert the move -// sourceIntervals - maintains sourceIntervals[reg] which each 'reg' is associated with -// location - maintains location[reg] which is the location of the var that was originally in 'reg'. -// toReg - the register to which the var is moving -// fromReg - the register from which the var is moving -// resolveType - the type of resolution to be performed -// -// Return Value: -// None. -// -// Notes: -// It inserts at least one move and updates incoming parameter 'location'. -// -void LinearScan::addResolutionForDouble(BasicBlock* block, - GenTree* insertionPoint, - Interval** sourceIntervals, - regNumberSmall* location, - regNumber toReg, - regNumber fromReg, - ResolveType resolveType) -{ - regNumber secondHalfTargetReg = REG_NEXT(fromReg); - Interval* intervalToBeMoved1 = sourceIntervals[fromReg]; - Interval* intervalToBeMoved2 = sourceIntervals[secondHalfTargetReg]; + if (genIsValidDoubleReg(fromReg)) + { + // Ensure that either: + // - the Interval targeting fromReg is not double, or + // - the other half of the double is free. + Interval* otherInterval = sourceIntervals[source[fromReg]]; + regNumber upperHalfReg = REG_NEXT(fromReg); + if ((otherInterval->registerType == TYP_DOUBLE) && (location[upperHalfReg] != REG_NA)) + { + targetRegsReady &= ~fromRegMask; + } + } +#endif // TARGET_ARM + } + } + targetRegsToDo &= ~targetRegMask; + } + else + { + compiler->codeGen->regSet.rsSetRegsModified(genRegMask(tempReg) DEBUGARG(true)); +#ifdef TARGET_ARM + if (sourceIntervals[fromReg]->registerType == TYP_DOUBLE) + { + assert(genIsValidDoubleReg(targetReg)); + assert(genIsValidDoubleReg(tempReg)); - assert(!(intervalToBeMoved1 == nullptr && intervalToBeMoved2 == nullptr)); + addResolutionForDouble(block, insertionPoint, sourceIntervals, location, tempReg, targetReg, + resolveType); + } + else +#endif // TARGET_ARM + { + assert(sourceIntervals[targetReg] != nullptr); - if (intervalToBeMoved1 != nullptr) - { - if (intervalToBeMoved1->registerType == TYP_DOUBLE) - { - // TYP_DOUBLE interval occupies a double register, i.e. two float registers. - assert(intervalToBeMoved2 == nullptr); - assert(genIsValidDoubleReg(toReg)); - } - else - { - // TYP_FLOAT interval occupies 1st half of double register, i.e. 1st float register - assert(genIsValidFloatReg(toReg)); + addResolution(block, insertionPoint, sourceIntervals[targetReg], tempReg, targetReg); + JITDUMP(" (%s)\n", resolveTypeName[resolveType]); + location[targetReg] = (regNumberSmall)tempReg; + } + targetRegsReady |= targetRegMask; + } + } } - addResolution(block, insertionPoint, intervalToBeMoved1, toReg, fromReg); - JITDUMP(" (%s)\n", resolveTypeName[resolveType]); - location[fromReg] = (regNumberSmall)toReg; } - if (intervalToBeMoved2 != nullptr) + // Finally, perform stack to reg moves + // All the target regs will be empty at this point + while (targetRegsFromStack != RBM_NONE) { - // TYP_FLOAT interval occupies 2nd half of double register. - assert(intervalToBeMoved2->registerType == TYP_FLOAT); - regNumber secondHalfTempReg = REG_NEXT(toReg); + regMaskTP targetRegMask = genFindLowestBit(targetRegsFromStack); + targetRegsFromStack &= ~targetRegMask; + regNumber targetReg = genRegNumFromMask(targetRegMask); - addResolution(block, insertionPoint, intervalToBeMoved2, secondHalfTempReg, secondHalfTargetReg); + Interval* interval = stackToRegIntervals[targetReg]; + assert(interval != nullptr); + + addResolution(block, insertionPoint, interval, targetReg, REG_STK); JITDUMP(" (%s)\n", resolveTypeName[resolveType]); - location[secondHalfTargetReg] = (regNumberSmall)secondHalfTempReg; } - - return; } -#endif // TARGET_ARM -//------------------------------------------------------------------------ -// addResolution: Add a resolution move of the given interval -// -// Arguments: -// block - the BasicBlock into which the move will be inserted. -// insertionPoint - the instruction before which to insert the move -// interval - the interval of the var to be moved -// toReg - the register to which the var is moving -// fromReg - the register from which the var is moving -// -// Return Value: -// None. -// -// Notes: -// For joins, we insert at the bottom (indicated by an insertionPoint -// of nullptr), while for splits we insert at the top. -// This is because for joins 'block' is a pred of the join, while for splits it is a succ. -// For critical edges, this function may be called twice - once to move from -// the source (fromReg), if any, to the stack, in which case toReg will be -// REG_STK, and we insert at the bottom (leave insertionPoint as nullptr). -// The next time, we want to move from the stack to the destination (toReg), -// in which case fromReg will be REG_STK, and we insert at the top. +#if TRACK_LSRA_STATS -void LinearScan::addResolution( - BasicBlock* block, GenTree* insertionPoint, Interval* interval, regNumber toReg, regNumber fromReg) +const char* LinearScan::getStatName(unsigned stat) { -#ifdef DEBUG - const char* insertionPointString; - if (insertionPoint == nullptr) - { - // We can't add resolution to a register at the bottom of a block that has an EHBoundaryOut, - // except in the case of the "EH Dummy" resolution from the stack. - assert((block->bbNum > bbNumMaxBeforeResolution) || (fromReg == REG_STK) || - !blockInfo[block->bbNum].hasEHBoundaryOut); - insertionPointString = "bottom"; - } - else - { - // We can't add resolution at the top of a block that has an EHBoundaryIn, - // except in the case of the "EH Dummy" resolution to the stack. - assert((block->bbNum > bbNumMaxBeforeResolution) || (toReg == REG_STK) || - !blockInfo[block->bbNum].hasEHBoundaryIn); - insertionPointString = "top"; - } - - // We should never add resolution move inside BBCallAlwaysPairTail. - noway_assert(!block->isBBCallAlwaysPairTail()); + LsraStat lsraStat = (LsraStat)stat; + assert(lsraStat != LsraStat::COUNT); -#endif // DEBUG + static const char* const lsraStatNames[] = { +#define LSRA_STAT_DEF(stat, name) name, +#include "lsra_stats.h" +#undef LSRA_STAT_DEF +#define REG_SEL_DEF(stat, value, shortname) #stat, +#include "lsra_score.h" +#undef REG_SEL_DEF + }; - JITDUMP(" " FMT_BB " %s: move V%02u from ", block->bbNum, insertionPointString, interval->varNum); - JITDUMP("%s to %s", getRegName(fromReg), getRegName(toReg)); + assert(stat < ArrLen(lsraStatNames)); + return lsraStatNames[lsraStat]; +} - insertMove(block, insertionPoint, interval->varNum, fromReg, toReg); - if (fromReg == REG_STK || toReg == REG_STK) - { - assert(interval->isSpilled); - } - else +LsraStat LinearScan::getLsraStatFromScore(RegisterScore registerScore) +{ + switch (registerScore) { - // We should have already marked this as spilled or split. - assert((interval->isSpilled) || (interval->isSplit)); +#define REG_SEL_DEF(stat, value, shortname) \ + case RegisterScore::stat: \ + return LsraStat::STAT_##stat; +#include "lsra_score.h" +#undef REG_SEL_DEF + default: + return LsraStat::STAT_FREE; } - - INTRACK_STATS(updateLsraStat(STAT_RESOLUTION_MOV, block->bbNum)); } -//------------------------------------------------------------------------ -// handleOutgoingCriticalEdges: Performs the necessary resolution on all critical edges that feed out of 'block' +// ---------------------------------------------------------- +// updateLsraStat: Increment LSRA stat counter. // // Arguments: -// block - the block with outgoing critical edges. -// -// Return Value: -// None.. +// stat - LSRA stat enum +// bbNum - Basic block to which LSRA stat needs to be +// associated with. // -// Notes: -// For all outgoing critical edges (i.e. any successor of this block which is -// a join edge), if there are any conflicts, split the edge by adding a new block, -// and generate the resolution code into that block. - -void LinearScan::handleOutgoingCriticalEdges(BasicBlock* block) +void LinearScan::updateLsraStat(LsraStat stat, unsigned bbNum) { - VARSET_TP outResolutionSet(VarSetOps::Intersection(compiler, block->bbLiveOut, resolutionCandidateVars)); - if (VarSetOps::IsEmpty(compiler, outResolutionSet)) + if (bbNum > bbNumMaxBeforeResolution) { + // This is a newly created basic block as part of resolution. + // These blocks contain resolution moves that are already accounted. return; } - VARSET_TP sameResolutionSet(VarSetOps::MakeEmpty(compiler)); - VARSET_TP diffResolutionSet(VarSetOps::MakeEmpty(compiler)); - // Get the outVarToRegMap for this block - VarToRegMap outVarToRegMap = getOutVarToRegMap(block->bbNum); - unsigned succCount = block->NumSucc(compiler); - assert(succCount > 1); + ++(blockInfo[bbNum].stats[(unsigned)stat]); +} - // First, determine the live regs at the end of this block so that we know what regs are - // available to copy into. - // Note that for this purpose we use the full live-out set, because we must ensure that - // even the registers that remain the same across the edge are preserved correctly. - regMaskTP liveOutRegs = RBM_NONE; - VarSetOps::Iter liveOutIter(compiler, block->bbLiveOut); - unsigned liveOutVarIndex = 0; - while (liveOutIter.NextElem(&liveOutVarIndex)) +// ----------------------------------------------------------- +// dumpLsraStats - dumps Lsra stats to given file. +// +// Arguments: +// file - file to which stats are to be written. +// +void LinearScan::dumpLsraStats(FILE* file) +{ + unsigned sumStats[LsraStat::COUNT] = {0}; + BasicBlock::weight_t wtdStats[LsraStat::COUNT] = {0}; + + fprintf(file, "----------\n"); + fprintf(file, "LSRA Stats"); +#ifdef DEBUG + if (!VERBOSE) { - regNumber fromReg = getVarReg(outVarToRegMap, liveOutVarIndex); - if (fromReg != REG_STK) - { - regMaskTP fromRegMask = genRegMask(fromReg, getIntervalForLocalVar(liveOutVarIndex)->registerType); - liveOutRegs |= fromRegMask; - } + fprintf(file, " : %s\n", compiler->info.compFullName); } - - // Next, if this blocks ends with a switch table, or for Arm64, ends with JCMP instruction, - // make sure to not copy into the registers that are consumed at the end of this block. - // - // Note: Only switches and JCMP (for Arm4) have input regs (and so can be fed by copies), so those - // are the only block-ending branches that need special handling. - regMaskTP consumedRegs = RBM_NONE; - if (block->bbJumpKind == BBJ_SWITCH) + else { - // At this point, Lowering has transformed any non-switch-table blocks into - // cascading ifs. - GenTree* switchTable = LIR::AsRange(block).LastNode(); - assert(switchTable != nullptr && switchTable->OperGet() == GT_SWITCH_TABLE); - - consumedRegs = switchTable->gtRsvdRegs; - GenTree* op1 = switchTable->gtGetOp1(); - GenTree* op2 = switchTable->gtGetOp2(); - noway_assert(op1 != nullptr && op2 != nullptr); - assert(op1->GetRegNum() != REG_NA && op2->GetRegNum() != REG_NA); - // No floating point values, so no need to worry about the register type - // (i.e. for ARM32, where we used the genRegMask overload with a type). - assert(varTypeIsIntegralOrI(op1) && varTypeIsIntegralOrI(op2)); - consumedRegs |= genRegMask(op1->GetRegNum()); - consumedRegs |= genRegMask(op2->GetRegNum()); + // In verbose mode no need to print full name + // while printing lsra stats. + fprintf(file, "\n"); } +#else + fprintf(file, " : %s\n", compiler->eeGetMethodFullName(compiler->info.compCompHnd)); +#endif -#ifdef TARGET_ARM64 - // Next, if this blocks ends with a JCMP, we have to make sure: - // 1. Not to copy into the register that JCMP uses - // e.g. JCMP w21, BRANCH - // 2. Not to copy into the source of JCMP's operand before it is consumed - // e.g. Should not use w0 since it will contain wrong value after resolution - // call METHOD - // ; mov w0, w19 <-- should not resolve in w0 here. - // mov w21, w0 - // JCMP w21, BRANCH - // 3. Not to modify the local variable it must consume + fprintf(file, "----------\n"); +#ifdef DEBUG + fprintf(file, "Register selection order: %S\n", JitConfig.JitLsraOrdering()); +#endif + fprintf(file, "Total Tracked Vars: %d\n", compiler->lvaTrackedCount); + fprintf(file, "Total Reg Cand Vars: %d\n", regCandidateVarCount); + fprintf(file, "Total number of Intervals: %d\n", + static_cast((intervals.size() == 0 ? 0 : (intervals.size() - 1)))); + fprintf(file, "Total number of RefPositions: %d\n", static_cast(refPositions.size() - 1)); - // Note: GT_COPY has special handling in codegen and its generation is merged with the - // node that consumes its result. So both, the input and output regs of GT_COPY must be - // excluded from the set available for resolution. - LclVarDsc* jcmpLocalVarDsc = nullptr; - if (block->bbJumpKind == BBJ_COND) + // compute total number of spill temps created + unsigned numSpillTemps = 0; + for (int i = 0; i < TYP_COUNT; i++) + { + numSpillTemps += maxSpill[i]; + } + fprintf(file, "Total Number of spill temps created: %d\n", numSpillTemps); + fprintf(file, "..........\n"); + bool addedBlockHeader = false; + bool anyNonZeroStat = false; + + for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) { - GenTree* lastNode = LIR::AsRange(block).LastNode(); + unsigned lsraStat = blockInfo[0].stats[statIndex]; - if (lastNode->OperIs(GT_JCMP)) + if (lsraStat != 0) { - GenTree* op1 = lastNode->gtGetOp1(); - consumedRegs |= genRegMask(op1->GetRegNum()); - - if (op1->OperIs(GT_COPY)) + if (!addedBlockHeader) { - GenTree* srcOp1 = op1->gtGetOp1(); - consumedRegs |= genRegMask(srcOp1->GetRegNum()); + addedBlockHeader = true; + fprintf(file, FMT_BB " [%8.2f]: ", 0, blockInfo[0].weight); + fprintf(file, "%s = %d", getStatName(statIndex), lsraStat); } - - if (op1->IsLocal()) + else { - GenTreeLclVarCommon* lcl = op1->AsLclVarCommon(); - jcmpLocalVarDsc = &compiler->lvaTable[lcl->GetLclNum()]; + fprintf(file, ", %s = %d", getStatName(statIndex), lsraStat); } + + sumStats[statIndex] += lsraStat; + wtdStats[statIndex] += (lsraStat * blockInfo[0].weight); + anyNonZeroStat = true; } } -#endif - - VarToRegMap sameVarToRegMap = sharedCriticalVarToRegMap; - regMaskTP sameWriteRegs = RBM_NONE; - regMaskTP diffReadRegs = RBM_NONE; - - // For each var that may require resolution, classify them as: - // - in the same register at the end of this block and at each target (no resolution needed) - // - in different registers at different targets (resolve separately): - // diffResolutionSet - // - in the same register at each target at which it's live, but different from the end of - // this block. We may be able to resolve these as if it is "join", but only if they do not - // write to any registers that are read by those in the diffResolutionSet: - // sameResolutionSet + if (anyNonZeroStat) + { + fprintf(file, "\n"); + } - VarSetOps::Iter outResolutionSetIter(compiler, outResolutionSet); - unsigned outResolutionSetVarIndex = 0; - while (outResolutionSetIter.NextElem(&outResolutionSetVarIndex)) + for (BasicBlock* block = compiler->fgFirstBB; block != nullptr; block = block->bbNext) { - regNumber fromReg = getVarReg(outVarToRegMap, outResolutionSetVarIndex); - bool maybeSameLivePaths = false; - bool liveOnlyAtSplitEdge = true; - regNumber sameToReg = REG_NA; - for (unsigned succIndex = 0; succIndex < succCount; succIndex++) + if (block->bbNum > bbNumMaxBeforeResolution) { - BasicBlock* succBlock = block->GetSucc(succIndex, compiler); - if (!VarSetOps::IsMember(compiler, succBlock->bbLiveIn, outResolutionSetVarIndex)) - { - maybeSameLivePaths = true; - continue; - } - else if (liveOnlyAtSplitEdge) - { - // Is the var live only at those target blocks which are connected by a split edge to this block - liveOnlyAtSplitEdge = ((succBlock->bbPreds->flNext == nullptr) && (succBlock != compiler->fgFirstBB)); - } + continue; + } - regNumber toReg = getVarReg(getInVarToRegMap(succBlock->bbNum), outResolutionSetVarIndex); - if (sameToReg == REG_NA) - { - sameToReg = toReg; - continue; - } - if (toReg == sameToReg) + addedBlockHeader = false; + anyNonZeroStat = false; + for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) + { + unsigned lsraStat = blockInfo[block->bbNum].stats[statIndex]; + + if (lsraStat != 0) { - continue; + if (!addedBlockHeader) + { + addedBlockHeader = true; + fprintf(file, FMT_BB " [%8.2f]: ", block->bbNum, block->bbWeight); + fprintf(file, "%s = %d", getStatName(statIndex), lsraStat); + } + else + { + fprintf(file, ", %s = %d", getStatName(statIndex), lsraStat); + } + + sumStats[statIndex] += lsraStat; + wtdStats[statIndex] += (lsraStat * block->bbWeight); + anyNonZeroStat = true; } - sameToReg = REG_NA; - break; } - // Check for the cases where we can't write to a register. - // We only need to check for these cases if sameToReg is an actual register (not REG_STK). - if (sameToReg != REG_NA && sameToReg != REG_STK) + if (anyNonZeroStat) { - // If there's a path on which this var isn't live, it may use the original value in sameToReg. - // In this case, sameToReg will be in the liveOutRegs of this block. - // Similarly, if sameToReg is in sameWriteRegs, it has already been used (i.e. for a lclVar that's - // live only at another target), and we can't copy another lclVar into that reg in this block. - regMaskTP sameToRegMask = - genRegMask(sameToReg, getIntervalForLocalVar(outResolutionSetVarIndex)->registerType); - if (maybeSameLivePaths && - (((sameToRegMask & liveOutRegs) != RBM_NONE) || ((sameToRegMask & sameWriteRegs) != RBM_NONE))) + fprintf(file, "\n"); + } + } + + fprintf(file, "..........\n"); + for (int regSelectI = 0; regSelectI < LsraStat::COUNT; regSelectI++) + { + if (regSelectI == firstRegSelStat) + { + fprintf(file, "..........\n"); + } + // TODO-review: I don't see a point of displaying Stats (SpillCount, etc.) if they are zero. Thoughts? + if ((regSelectI < firstRegSelStat) || (sumStats[regSelectI] != 0)) + { + // Print register selection stats + if (regSelectI >= firstRegSelStat) { - sameToReg = REG_NA; + fprintf(file, "Total %s [#%2d] : %d Weighted: %f\n", getStatName(regSelectI), + (regSelectI - firstRegSelStat + 1), sumStats[regSelectI], wtdStats[regSelectI]); } - // If this register is busy because it is used by a switch table at the end of the block - // (or for Arm64, it is consumed by JCMP), we can't do the copy in this block since we can't - // insert it after the switch (or for Arm64, can't insert and overwrite the operand/source - // of operand of JCMP). - if ((sameToRegMask & consumedRegs) != RBM_NONE) + else { - sameToReg = REG_NA; + fprintf(file, "Total %s : %d Weighted: %f\n", getStatName(regSelectI), sumStats[regSelectI], + wtdStats[regSelectI]); } + } + } + printf("\n"); +} -#ifdef TARGET_ARM64 - if (jcmpLocalVarDsc && (jcmpLocalVarDsc->lvVarIndex == outResolutionSetVarIndex)) - { - sameToReg = REG_NA; - } -#endif +// ----------------------------------------------------------- +// dumpLsraStatsCsvFormat - dumps Lsra stats to given file in csv format. +// +// Arguments: +// file - file to which stats are to be written. +// +void LinearScan::dumpLsraStatsCsv(FILE* file) +{ + unsigned sumStats[LsraStat::COUNT] = {0}; - // If the var is live only at those blocks connected by a split edge and not live-in at some of the - // target blocks, we will resolve it the same way as if it were in diffResolutionSet and resolution - // will be deferred to the handling of split edges, which means copy will only be at those target(s). - // - // Another way to achieve similar resolution for vars live only at split edges is by removing them - // from consideration up-front but it requires that we traverse those edges anyway to account for - // the registers that must note be overwritten. - if (liveOnlyAtSplitEdge && maybeSameLivePaths) - { - sameToReg = REG_NA; - } + // Write the header if the file is empty + if (ftell(file) == 0) + { + // header + fprintf(file, "\"Method Name\""); + for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) + { + fprintf(file, ",\"%s\"", LinearScan::getStatName(statIndex)); } + fprintf(file, ",\"PerfScore\"\n"); + } - if (sameToReg == REG_NA) + // bbNum == 0 + for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) + { + sumStats[statIndex] += blockInfo[0].stats[statIndex]; + } + + // blocks + for (BasicBlock* block = compiler->fgFirstBB; block != nullptr; block = block->bbNext) + { + if (block->bbNum > bbNumMaxBeforeResolution) { - VarSetOps::AddElemD(compiler, diffResolutionSet, outResolutionSetVarIndex); - if (fromReg != REG_STK) - { - diffReadRegs |= genRegMask(fromReg, getIntervalForLocalVar(outResolutionSetVarIndex)->registerType); - } + continue; } - else if (sameToReg != fromReg) + + for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) { - VarSetOps::AddElemD(compiler, sameResolutionSet, outResolutionSetVarIndex); - setVarReg(sameVarToRegMap, outResolutionSetVarIndex, sameToReg); - if (sameToReg != REG_STK) - { - sameWriteRegs |= genRegMask(sameToReg, getIntervalForLocalVar(outResolutionSetVarIndex)->registerType); - } + sumStats[statIndex] += blockInfo[block->bbNum].stats[statIndex]; } } - if (!VarSetOps::IsEmpty(compiler, sameResolutionSet)) + fprintf(file, "\"%s\"", compiler->info.compFullName); + for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) + { + fprintf(file, ",%u", sumStats[statIndex]); + } + fprintf(file, ",%.2f\n", compiler->info.compPerfScore); +} +#endif // TRACK_LSRA_STATS + +#ifdef DEBUG +void dumpRegMask(regMaskTP regs) +{ + if (regs == RBM_ALLINT) + { + printf("[allInt]"); + } + else if (regs == (RBM_ALLINT & ~RBM_FPBASE)) + { + printf("[allIntButFP]"); + } + else if (regs == RBM_ALLFLOAT) + { + printf("[allFloat]"); + } + else if (regs == RBM_ALLDOUBLE) + { + printf("[allDouble]"); + } + else + { + dspRegMask(regs); + } +} + +static const char* getRefTypeName(RefType refType) +{ + switch (refType) + { +#define DEF_REFTYPE(memberName, memberValue, shortName) \ + case memberName: \ + return #memberName; +#include "lsra_reftypes.h" +#undef DEF_REFTYPE + default: + return nullptr; + } +} + +static const char* getRefTypeShortName(RefType refType) +{ + switch (refType) { - if ((sameWriteRegs & diffReadRegs) != RBM_NONE) - { - // We cannot split the "same" and "diff" regs if the "same" set writes registers - // that must be read by the "diff" set. (Note that when these are done as a "batch" - // we carefully order them to ensure all the input regs are read before they are - // overwritten.) - VarSetOps::UnionD(compiler, diffResolutionSet, sameResolutionSet); - VarSetOps::ClearD(compiler, sameResolutionSet); - } - else - { - // For any vars in the sameResolutionSet, we can simply add the move at the end of "block". - resolveEdge(block, nullptr, ResolveSharedCritical, sameResolutionSet); - } +#define DEF_REFTYPE(memberName, memberValue, shortName) \ + case memberName: \ + return shortName; +#include "lsra_reftypes.h" +#undef DEF_REFTYPE + default: + return nullptr; } - if (!VarSetOps::IsEmpty(compiler, diffResolutionSet)) +} + +//------------------------------------------------------------------------ +// getScoreName: Returns the texual name of register score +const char* LinearScan::getScoreName(RegisterScore score) +{ + switch (score) { - for (unsigned succIndex = 0; succIndex < succCount; succIndex++) - { - BasicBlock* succBlock = block->GetSucc(succIndex, compiler); +#define REG_SEL_DEF(stat, value, shortname) \ + case stat: \ + return shortname; +#include "lsra_score.h" +#undef REG_SEL_DEF + default: + return " - "; + } +} - // Any "diffResolutionSet" resolution for a block with no other predecessors will be handled later - // as split resolution. - if ((succBlock->bbPreds->flNext == nullptr) && (succBlock != compiler->fgFirstBB)) - { - continue; - } +void RefPosition::dump() +{ + printf("bbNum); - VARSET_TP edgeResolutionSet(VarSetOps::Intersection(compiler, diffResolutionSet, succBlock->bbLiveIn)); - VarSetOps::Iter iter(compiler, edgeResolutionSet); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex)) - { - regNumber fromReg = getVarReg(outVarToRegMap, varIndex); - regNumber toReg = getVarReg(succInVarToRegMap, varIndex); + printf(" %s ", getRefTypeName(refType)); - if (fromReg == toReg) - { - VarSetOps::RemoveElemD(compiler, edgeResolutionSet, varIndex); - } - } - if (!VarSetOps::IsEmpty(compiler, edgeResolutionSet)) - { - // For EH vars, we can always safely load them from the stack into the target for this block, - // so if we have only EH vars, we'll do that instead of splitting the edge. - if ((compiler->compHndBBtabCount > 0) && VarSetOps::IsSubset(compiler, edgeResolutionSet, exceptVars)) - { - GenTree* insertionPoint = LIR::AsRange(succBlock).FirstNonPhiNode(); - VarSetOps::Iter edgeSetIter(compiler, edgeResolutionSet); - unsigned edgeVarIndex = 0; - while (edgeSetIter.NextElem(&edgeVarIndex)) - { - regNumber toReg = getVarReg(succInVarToRegMap, edgeVarIndex); - setVarReg(succInVarToRegMap, edgeVarIndex, REG_STK); - if (toReg != REG_STK) - { - Interval* interval = getIntervalForLocalVar(edgeVarIndex); - assert(interval->isWriteThru); - addResolution(succBlock, insertionPoint, interval, toReg, REG_STK); - JITDUMP(" (EHvar)\n"); - } - } - } - else - { - resolveEdge(block, succBlock, ResolveCritical, edgeResolutionSet); - } - } + if (this->IsPhysRegRef()) + { + this->getReg()->tinyDump(); + } + else if (getInterval()) + { + this->getInterval()->tinyDump(); + } + if (this->treeNode) + { + printf("%s", treeNode->OpName(treeNode->OperGet())); + if (this->treeNode->IsMultiRegNode()) + { + printf("[%d]", this->multiRegIdx); } } -} + printf(" " FMT_BB " ", this->bbNum); -//------------------------------------------------------------------------ -// resolveEdges: Perform resolution across basic block edges -// -// Arguments: -// None. -// -// Return Value: -// None. -// -// Notes: -// Traverse the basic blocks. -// - If this block has a single predecessor that is not the immediately -// preceding block, perform any needed 'split' resolution at the beginning of this block -// - Otherwise if this block has critical incoming edges, handle them. -// - If this block has a single successor that has multiple predecesors, perform any needed -// 'join' resolution at the end of this block. -// Note that a block may have both 'split' or 'critical' incoming edge(s) and 'join' outgoing -// edges. + printf("regmask="); + dumpRegMask(registerAssignment); -void LinearScan::resolveEdges() -{ - JITDUMP("RESOLVING EDGES\n"); + printf(" minReg=%d", minRegCandidateCount); - // The resolutionCandidateVars set was initialized with all the lclVars that are live-in to - // any block. We now intersect that set with any lclVars that ever spilled or split. - // If there are no candidates for resoultion, simply return. + if (this->lastUse) + { + printf(" last"); + } + if (this->reload) + { + printf(" reload"); + } + if (this->spillAfter) + { + printf(" spillAfter"); + } + if (this->writeThru) + { + printf(" writeThru"); + } + if (this->moveReg) + { + printf(" move"); + } + if (this->copyReg) + { + printf(" copy"); + } + if (this->isFixedRegRef) + { + printf(" fixed"); + } + if (this->isLocalDefUse) + { + printf(" local"); + } + if (this->delayRegFree) + { + printf(" delay"); + } + if (this->outOfOrder) + { + printf(" outOfOrder"); + } - VarSetOps::IntersectionD(compiler, resolutionCandidateVars, splitOrSpilledVars); - if (VarSetOps::IsEmpty(compiler, resolutionCandidateVars)) + if (this->RegOptional()) { - return; + printf(" regOptional"); } + printf(">\n"); +} - BasicBlock *block, *prevBlock = nullptr; +void RegRecord::dump() +{ + tinyDump(); +} - // Handle all the critical edges first. - // We will try to avoid resolution across critical edges in cases where all the critical-edge - // targets of a block have the same home. We will then split the edges only for the - // remaining mismatches. We visit the out-edges, as that allows us to share the moves that are - // common among all the targets. +void Interval::dump() +{ + printf("Interval %2u:", intervalIndex); - if (hasCriticalEdges) + if (isLocalVar) { - foreach_block(compiler, block) - { - if (block->bbNum > bbNumMaxBeforeResolution) - { - // This is a new block added during resolution - we don't need to visit these now. - continue; - } - if (blockInfo[block->bbNum].hasCriticalOutEdge) - { - handleOutgoingCriticalEdges(block); - } - prevBlock = block; - } + printf(" (V%02u)", varNum); + } + else if (IsUpperVector()) + { + assert(relatedInterval != nullptr); + printf(" (U%02u)", relatedInterval->varNum); + } + printf(" %s", varTypeName(registerType)); + if (isInternal) + { + printf(" (INTERNAL)"); + } + if (isSpilled) + { + printf(" (SPILLED)"); + } + if (isSplit) + { + printf(" (SPLIT)"); + } + if (isStructField) + { + printf(" (field)"); + } + if (isPromotedStruct) + { + printf(" (promoted struct)"); + } + if (hasConflictingDefUse) + { + printf(" (def-use conflict)"); + } + if (hasInterferingUses) + { + printf(" (interfering uses)"); + } + if (isSpecialPutArg) + { + printf(" (specialPutArg)"); + } + if (isConstant) + { + printf(" (constant)"); + } + if (isWriteThru) + { + printf(" (writeThru)"); } - prevBlock = nullptr; - foreach_block(compiler, block) + printf(" RefPositions {"); + for (RefPosition* refPosition = this->firstRefPosition; refPosition != nullptr; + refPosition = refPosition->nextRefPosition) { - if (block->bbNum > bbNumMaxBeforeResolution) + printf("#%u@%u", refPosition->rpNum, refPosition->nodeLocation); + if (refPosition->nextRefPosition) { - // This is a new block added during resolution - we don't need to visit these now. - continue; + printf(" "); } + } + printf("}"); - unsigned succCount = block->NumSucc(compiler); - BasicBlock* uniquePredBlock = block->GetUniquePred(compiler); - - // First, if this block has a single predecessor, - // we may need resolution at the beginning of this block. - // This may be true even if it's the block we used for starting locations, - // if a variable was spilled. - VARSET_TP inResolutionSet(VarSetOps::Intersection(compiler, block->bbLiveIn, resolutionCandidateVars)); - if (!VarSetOps::IsEmpty(compiler, inResolutionSet)) - { - if (uniquePredBlock != nullptr) - { - // We may have split edges during critical edge resolution, and in the process split - // a non-critical edge as well. - // It is unlikely that we would ever have more than one of these in sequence (indeed, - // I don't think it's possible), but there's no need to assume that it can't. - while (uniquePredBlock->bbNum > bbNumMaxBeforeResolution) - { - uniquePredBlock = uniquePredBlock->GetUniquePred(compiler); - noway_assert(uniquePredBlock != nullptr); - } - resolveEdge(uniquePredBlock, block, ResolveSplit, inResolutionSet); - } - } + // this is not used (yet?) + // printf(" SpillOffset %d", this->spillOffset); - // Finally, if this block has a single successor: - // - and that has at least one other predecessor (otherwise we will do the resolution at the - // top of the successor), - // - and that is not the target of a critical edge (otherwise we've already handled it) - // we may need resolution at the end of this block. + printf(" physReg:%s", getRegName(physReg)); - if (succCount == 1) - { - BasicBlock* succBlock = block->GetSucc(0, compiler); - if (succBlock->GetUniquePred(compiler) == nullptr) - { - VARSET_TP outResolutionSet( - VarSetOps::Intersection(compiler, succBlock->bbLiveIn, resolutionCandidateVars)); - if (!VarSetOps::IsEmpty(compiler, outResolutionSet)) - { - resolveEdge(block, succBlock, ResolveJoin, outResolutionSet); - } - } - } - } + printf(" Preferences="); + dumpRegMask(this->registerPreferences); - // Now, fixup the mapping for any blocks that were adding for edge splitting. - // See the comment prior to the call to fgSplitEdge() in resolveEdge(). - // Note that we could fold this loop in with the checking code below, but that - // would only improve the debug case, and would clutter up the code somewhat. - if (compiler->fgBBNumMax > bbNumMaxBeforeResolution) + if (relatedInterval) { - foreach_block(compiler, block) - { - if (block->bbNum > bbNumMaxBeforeResolution) - { - // There may be multiple blocks inserted when we split. But we must always have exactly - // one path (i.e. all blocks must be single-successor and single-predecessor), - // and only one block along the path may be non-empty. - // Note that we may have a newly-inserted block that is empty, but which connects - // two non-resolution blocks. This happens when an edge is split that requires it. - - BasicBlock* succBlock = block; - do - { - succBlock = succBlock->GetUniqueSucc(); - noway_assert(succBlock != nullptr); - } while ((succBlock->bbNum > bbNumMaxBeforeResolution) && succBlock->isEmpty()); - - BasicBlock* predBlock = block; - do - { - predBlock = predBlock->GetUniquePred(compiler); - noway_assert(predBlock != nullptr); - } while ((predBlock->bbNum > bbNumMaxBeforeResolution) && predBlock->isEmpty()); + printf(" RelatedInterval "); + relatedInterval->microDump(); + } - unsigned succBBNum = succBlock->bbNum; - unsigned predBBNum = predBlock->bbNum; - if (block->isEmpty()) - { - // For the case of the empty block, find the non-resolution block (succ or pred). - if (predBBNum > bbNumMaxBeforeResolution) - { - assert(succBBNum <= bbNumMaxBeforeResolution); - predBBNum = 0; - } - else - { - succBBNum = 0; - } - } - else - { - assert((succBBNum <= bbNumMaxBeforeResolution) && (predBBNum <= bbNumMaxBeforeResolution)); - } - SplitEdgeInfo info = {predBBNum, succBBNum}; - getSplitBBNumToTargetBBNumMap()->Set(block->bbNum, info); + printf("\n"); +} - // Set both the live-in and live-out to the live-in of the successor (by construction liveness - // doesn't change in a split block). - VarSetOps::Assign(compiler, block->bbLiveIn, succBlock->bbLiveIn); - VarSetOps::Assign(compiler, block->bbLiveOut, succBlock->bbLiveIn); - } - } +// print out very concise representation +void Interval::tinyDump() +{ + printf("varNum); + } + else if (isInternal) + { + printf(" internal"); } + printf("> "); +} -#ifdef DEBUG - // Make sure the varToRegMaps match up on all edges. - bool foundMismatch = false; - foreach_block(compiler, block) +// print out extremely concise representation +void Interval::microDump() +{ + if (isLocalVar) { - if (block->isEmpty() && block->bbNum > bbNumMaxBeforeResolution) - { - continue; - } - VarToRegMap toVarToRegMap = getInVarToRegMap(block->bbNum); - for (flowList* pred = block->bbPreds; pred != nullptr; pred = pred->flNext) - { - BasicBlock* predBlock = pred->getBlock(); - VarToRegMap fromVarToRegMap = getOutVarToRegMap(predBlock->bbNum); - VarSetOps::Iter iter(compiler, block->bbLiveIn); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex)) - { - regNumber fromReg = getVarReg(fromVarToRegMap, varIndex); - regNumber toReg = getVarReg(toVarToRegMap, varIndex); - if (fromReg != toReg) - { - Interval* interval = getIntervalForLocalVar(varIndex); - // The fromReg and toReg may not match for a write-thru interval where the toReg is - // REG_STK, since the stack value is always valid for that case (so no move is needed). - if (!interval->isWriteThru || (toReg != REG_STK)) - { - if (!foundMismatch) - { - foundMismatch = true; - printf("Found mismatched var locations after resolution!\n"); - } - printf(" V%02u: " FMT_BB " to " FMT_BB ": %s to %s\n", interval->varNum, predBlock->bbNum, - block->bbNum, getRegName(fromReg), getRegName(toReg)); - } - } - } - } + printf("", varNum, intervalIndex); + return; } - assert(!foundMismatch); -#endif - JITDUMP("\n"); + else if (IsUpperVector()) + { + assert(relatedInterval != nullptr); + printf(" (U%02u)", relatedInterval->varNum); + } + char intervalTypeChar = 'I'; + if (isInternal) + { + intervalTypeChar = 'T'; + } + printf("<%c%u>", intervalTypeChar, intervalIndex); } -//------------------------------------------------------------------------ -// resolveEdge: Perform the specified type of resolution between two blocks. -// -// Arguments: -// fromBlock - the block from which the edge originates -// toBlock - the block at which the edge terminates -// resolveType - the type of resolution to be performed -// liveSet - the set of tracked lclVar indices which may require resolution -// -// Return Value: -// None. -// -// Assumptions: -// The caller must have performed the analysis to determine the type of the edge. -// -// Notes: -// This method emits the correctly ordered moves necessary to place variables in the -// correct registers across a Split, Join or Critical edge. -// In order to avoid overwriting register values before they have been moved to their -// new home (register/stack), it first does the register-to-stack moves (to free those -// registers), then the register to register moves, ensuring that the target register -// is free before the move, and then finally the stack to register moves. +void RegRecord::tinyDump() +{ + printf(" ", getRegName(regNum)); +} -void LinearScan::resolveEdge(BasicBlock* fromBlock, - BasicBlock* toBlock, - ResolveType resolveType, - VARSET_VALARG_TP liveSet) +void LinearScan::dumpDefList() { - VarToRegMap fromVarToRegMap = getOutVarToRegMap(fromBlock->bbNum); - VarToRegMap toVarToRegMap; - if (resolveType == ResolveSharedCritical) + if (!VERBOSE) { - toVarToRegMap = sharedCriticalVarToRegMap; + return; } - else + JITDUMP("DefList: { "); + bool first = true; + for (RefInfoListNode *listNode = defList.Begin(), *end = defList.End(); listNode != end; + listNode = listNode->Next()) + { + GenTree* node = listNode->treeNode; + JITDUMP("%sN%03u.t%d. %s", first ? "" : "; ", node->gtSeqNum, node->gtTreeID, GenTree::OpName(node->OperGet())); + first = false; + } + JITDUMP(" }\n"); +} + +void LinearScan::lsraDumpIntervals(const char* msg) +{ + printf("\nLinear scan intervals %s:\n", msg); + for (Interval& interval : intervals) { - toVarToRegMap = getInVarToRegMap(toBlock->bbNum); + // only dump something if it has references + // if (interval->firstRefPosition) + interval.dump(); } - // The block to which we add the resolution moves depends on the resolveType - BasicBlock* block; - switch (resolveType) + printf("\n"); +} + +// Dumps a tree node as a destination or source operand, with the style +// of dump dependent on the mode +void LinearScan::lsraGetOperandString(GenTree* tree, + LsraTupleDumpMode mode, + char* operandString, + unsigned operandStringLength) +{ + const char* lastUseChar = ""; + if (tree->OperIsScalarLocal() && ((tree->gtFlags & GTF_VAR_DEATH) != 0)) { - case ResolveJoin: - case ResolveSharedCritical: - block = fromBlock; - break; - case ResolveSplit: - block = toBlock; + lastUseChar = "*"; + } + switch (mode) + { + case LinearScan::LSRA_DUMP_PRE: + case LinearScan::LSRA_DUMP_REFPOS: + _snprintf_s(operandString, operandStringLength, operandStringLength, "t%d%s", tree->gtTreeID, lastUseChar); break; - case ResolveCritical: - // fgSplitEdge may add one or two BasicBlocks. It returns the block that splits - // the edge from 'fromBlock' and 'toBlock', but if it inserts that block right after - // a block with a fall-through it will have to create another block to handle that edge. - // These new blocks can be mapped to existing blocks in order to correctly handle - // the calls to recordVarLocationsAtStartOfBB() from codegen. That mapping is handled - // in resolveEdges(), after all the edge resolution has been done (by calling this - // method for each edge). - block = compiler->fgSplitEdge(fromBlock, toBlock); + case LinearScan::LSRA_DUMP_POST: + { + Compiler* compiler = JitTls::GetCompiler(); - // Split edges are counted against fromBlock. - INTRACK_STATS(updateLsraStat(STAT_SPLIT_EDGE, fromBlock->bbNum)); - break; + if (!tree->gtHasReg()) + { + _snprintf_s(operandString, operandStringLength, operandStringLength, "STK%s", lastUseChar); + } + else + { + regNumber reg = tree->GetRegNum(); + int charCount = _snprintf_s(operandString, operandStringLength, operandStringLength, "%s%s", + getRegName(reg, genIsValidFloatReg(reg)), lastUseChar); + operandString += charCount; + operandStringLength -= charCount; + + if (tree->IsMultiRegNode()) + { + unsigned regCount = tree->IsMultiRegLclVar() + ? compiler->lvaGetDesc(tree->AsLclVar()->GetLclNum())->lvFieldCnt + : tree->GetMultiRegCount(); + for (unsigned regIndex = 1; regIndex < regCount; regIndex++) + { + regNumber reg = tree->GetRegByIndex(regIndex); + charCount = _snprintf_s(operandString, operandStringLength, operandStringLength, ",%s%s", + getRegName(reg, genIsValidFloatReg(reg)), lastUseChar); + operandString += charCount; + operandStringLength -= charCount; + } + } + } + } + break; default: - unreached(); + printf("ERROR: INVALID TUPLE DUMP MODE\n"); break; } +} +void LinearScan::lsraDispNode(GenTree* tree, LsraTupleDumpMode mode, bool hasDest) +{ + Compiler* compiler = JitTls::GetCompiler(); + const unsigned operandStringLength = 6 * MAX_MULTIREG_COUNT + 1; + char operandString[operandStringLength]; + const char* emptyDestOperand = " "; + char spillChar = ' '; -#ifndef TARGET_XARCH - // We record tempregs for beginning and end of each block. - // For amd64/x86 we only need a tempReg for float - we'll use xchg for int. - // TODO-Throughput: It would be better to determine the tempRegs on demand, but the code below - // modifies the varToRegMaps so we don't have all the correct registers at the time - // we need to get the tempReg. - regNumber tempRegInt = - (resolveType == ResolveSharedCritical) ? REG_NA : getTempRegForResolution(fromBlock, toBlock, TYP_INT); -#endif // !TARGET_XARCH - regNumber tempRegFlt = REG_NA; -#ifdef TARGET_ARM - regNumber tempRegDbl = REG_NA; -#endif - if ((compiler->compFloatingPointUsed) && (resolveType != ResolveSharedCritical)) + if (mode == LinearScan::LSRA_DUMP_POST) { -#ifdef TARGET_ARM - // Try to reserve a double register for TYP_DOUBLE and use it for TYP_FLOAT too if available. - tempRegDbl = getTempRegForResolution(fromBlock, toBlock, TYP_DOUBLE); - if (tempRegDbl != REG_NA) + if ((tree->gtFlags & GTF_SPILL) != 0) { - tempRegFlt = tempRegDbl; + spillChar = 'S'; } - else -#endif // TARGET_ARM + if (!hasDest && tree->gtHasReg()) { - tempRegFlt = getTempRegForResolution(fromBlock, toBlock, TYP_FLOAT); + // A node can define a register, but not produce a value for a parent to consume, + // i.e. in the "localDefUse" case. + // There used to be an assert here that we wouldn't spill such a node. + // However, we can have unused lclVars that wind up being the node at which + // it is spilled. This probably indicates a bug, but we don't realy want to + // assert during a dump. + if (spillChar == 'S') + { + spillChar = '$'; + } + else + { + spillChar = '*'; + } + hasDest = true; } } + printf("%c N%03u. ", spillChar, tree->gtSeqNum); - regMaskTP targetRegsToDo = RBM_NONE; - regMaskTP targetRegsReady = RBM_NONE; - regMaskTP targetRegsFromStack = RBM_NONE; - - // The following arrays capture the location of the registers as they are moved: - // - location[reg] gives the current location of the var that was originally in 'reg'. - // (Note that a var may be moved more than once.) - // - source[reg] gives the original location of the var that needs to be moved to 'reg'. - // For example, if a var is in rax and needs to be moved to rsi, then we would start with: - // location[rax] == rax - // source[rsi] == rax -- this doesn't change - // Then, if for some reason we need to move it temporary to rbx, we would have: - // location[rax] == rbx - // Once we have completed the move, we will have: - // location[rax] == REG_NA - // This indicates that the var originally in rax is now in its target register. - - regNumberSmall location[REG_COUNT]; - C_ASSERT(sizeof(char) == sizeof(regNumberSmall)); // for memset to work - memset(location, REG_NA, REG_COUNT); - regNumberSmall source[REG_COUNT]; - memset(source, REG_NA, REG_COUNT); - - // What interval is this register associated with? - // (associated with incoming reg) - Interval* sourceIntervals[REG_COUNT]; - memset(&sourceIntervals, 0, sizeof(sourceIntervals)); - - // Intervals for vars that need to be loaded from the stack - Interval* stackToRegIntervals[REG_COUNT]; - memset(&stackToRegIntervals, 0, sizeof(stackToRegIntervals)); - - // Get the starting insertion point for the "to" resolution - GenTree* insertionPoint = nullptr; - if (resolveType == ResolveSplit || resolveType == ResolveCritical) - { - insertionPoint = LIR::AsRange(block).FirstNonPhiNode(); - } - - // If this is an edge between EH regions, we may have "extra" live-out EH vars. - // If we are adding resolution at the end of the block, we need to create "virtual" moves - // for these so that their registers are freed and can be reused. - if ((resolveType == ResolveJoin) && (compiler->compHndBBtabCount > 0)) + LclVarDsc* varDsc = nullptr; + unsigned varNum = UINT_MAX; + if (tree->IsLocal()) { - VARSET_TP extraLiveSet(VarSetOps::Diff(compiler, block->bbLiveOut, toBlock->bbLiveIn)); - VarSetOps::IntersectionD(compiler, extraLiveSet, registerCandidateVars); - VarSetOps::Iter iter(compiler, extraLiveSet); - unsigned extraVarIndex = 0; - while (iter.NextElem(&extraVarIndex)) + varNum = tree->AsLclVarCommon()->GetLclNum(); + varDsc = &(compiler->lvaTable[varNum]); + if (varDsc->lvLRACandidate) { - Interval* interval = getIntervalForLocalVar(extraVarIndex); - assert(interval->isWriteThru); - regNumber fromReg = getVarReg(fromVarToRegMap, extraVarIndex); - if (fromReg != REG_STK) - { - addResolution(block, insertionPoint, interval, REG_STK, fromReg); - JITDUMP(" (EH DUMMY)\n"); - setVarReg(fromVarToRegMap, extraVarIndex, REG_STK); - } + hasDest = false; } } - - // First: - // - Perform all moves from reg to stack (no ordering needed on these) - // - For reg to reg moves, record the current location, associating their - // source location with the target register they need to go into - // - For stack to reg moves (done last, no ordering needed between them) - // record the interval associated with the target reg - // TODO-Throughput: We should be looping over the liveIn and liveOut registers, since - // that will scale better than the live variables - - VarSetOps::Iter iter(compiler, liveSet); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex)) + if (hasDest) { - Interval* interval = getIntervalForLocalVar(varIndex); - regNumber fromReg = getVarReg(fromVarToRegMap, varIndex); - regNumber toReg = getVarReg(toVarToRegMap, varIndex); - if (fromReg == toReg) + if (mode == LinearScan::LSRA_DUMP_POST && tree->gtFlags & GTF_SPILLED) { - continue; + assert(tree->gtHasReg()); } - if (interval->isWriteThru && (toReg == REG_STK)) + lsraGetOperandString(tree, mode, operandString, operandStringLength); + printf("%-15s =", operandString); + } + else + { + printf("%-15s ", emptyDestOperand); + } + if (varDsc != nullptr) + { + if (varDsc->lvLRACandidate) { - // We don't actually move a writeThru var back to the stack, as its stack value is always valid. - // However, if this is a Join edge (i.e. the move is happening at the bottom of the block), - // and it is a "normal" flow edge, we will go ahead and generate a mov instruction, which will be - // a NOP but will cause the variable to be removed from being live in the register. - if ((resolveType == ResolveSplit) || block->hasEHBoundaryOut()) + if (mode == LSRA_DUMP_REFPOS) { - continue; + printf(" V%02u(L%d)", varNum, getIntervalForLocalVar(varDsc->lvVarIndex)->intervalIndex); + } + else + { + lsraGetOperandString(tree, mode, operandString, operandStringLength); + printf(" V%02u(%s)", varNum, operandString); + if (mode == LinearScan::LSRA_DUMP_POST && tree->gtFlags & GTF_SPILLED) + { + printf("R"); + } } } - // For Critical edges, the location will not change on either side of the edge, - // since we'll add a new block to do the move. - if (resolveType == ResolveSplit) + else { - setVarReg(toVarToRegMap, varIndex, fromReg); + printf(" V%02u MEM", varNum); } - else if (resolveType == ResolveJoin || resolveType == ResolveSharedCritical) + } + else if (tree->OperIs(GT_ASG)) + { + assert(!tree->gtHasReg()); + printf(" asg%s ", GenTree::OpName(tree->OperGet())); + } + else + { + compiler->gtDispNodeName(tree); + if (tree->OperKind() & GTK_LEAF) { - setVarReg(fromVarToRegMap, varIndex, toReg); + compiler->gtDispLeaf(tree, nullptr); } + } +} - assert(fromReg < UCHAR_MAX && toReg < UCHAR_MAX); +//------------------------------------------------------------------------ +// DumpOperandDefs: dumps the registers defined by a node. +// +// Arguments: +// operand - The operand for which to compute a register count. +// +// Returns: +// The number of registers defined by `operand`. +// +void LinearScan::DumpOperandDefs( + GenTree* operand, bool& first, LsraTupleDumpMode mode, char* operandString, const unsigned operandStringLength) +{ + assert(operand != nullptr); + assert(operandString != nullptr); + if (!operand->IsLIR()) + { + return; + } - if (fromReg == REG_STK) - { - stackToRegIntervals[toReg] = interval; - targetRegsFromStack |= genRegMask(toReg); - } - else if (toReg == REG_STK) + int dstCount = ComputeOperandDstCount(operand); + + if (dstCount != 0) + { + // This operand directly produces registers; print it. + if (!first) { - // Do the reg to stack moves now - addResolution(block, insertionPoint, interval, REG_STK, fromReg); - JITDUMP(" (%s)\n", - (interval->isWriteThru && (toReg == REG_STK)) ? "EH DUMMY" : resolveTypeName[resolveType]); + printf(","); } - else + lsraGetOperandString(operand, mode, operandString, operandStringLength); + printf("%s", operandString); + first = false; + } + else if (operand->isContained()) + { + // This is a contained node. Dump the defs produced by its operands. + for (GenTree* op : operand->Operands()) { - location[fromReg] = (regNumberSmall)fromReg; - source[toReg] = (regNumberSmall)fromReg; - sourceIntervals[fromReg] = interval; - targetRegsToDo |= genRegMask(toReg); + DumpOperandDefs(op, first, mode, operandString, operandStringLength); } } +} - // REGISTER to REGISTER MOVES +void LinearScan::TupleStyleDump(LsraTupleDumpMode mode) +{ + BasicBlock* block; + LsraLocation currentLoc = 1; // 0 is the entry + const unsigned operandStringLength = 6 * MAX_MULTIREG_COUNT + 1; + char operandString[operandStringLength]; - // First, find all the ones that are ready to move now - regMaskTP targetCandidates = targetRegsToDo; - while (targetCandidates != RBM_NONE) + // currentRefPosition is not used for LSRA_DUMP_PRE + // We keep separate iterators for defs, so that we can print them + // on the lhs of the dump + RefPositionIterator refPosIterator = refPositions.begin(); + RefPosition* currentRefPosition = &refPosIterator; + + switch (mode) { - regMaskTP targetRegMask = genFindLowestBit(targetCandidates); - targetCandidates &= ~targetRegMask; - regNumber targetReg = genRegNumFromMask(targetRegMask); - if (location[targetReg] == REG_NA) + case LSRA_DUMP_PRE: + printf("TUPLE STYLE DUMP BEFORE LSRA\n"); + break; + case LSRA_DUMP_REFPOS: + printf("TUPLE STYLE DUMP WITH REF POSITIONS\n"); + break; + case LSRA_DUMP_POST: + printf("TUPLE STYLE DUMP WITH REGISTER ASSIGNMENTS\n"); + break; + default: + printf("ERROR: INVALID TUPLE DUMP MODE\n"); + return; + } + + if (mode != LSRA_DUMP_PRE) + { + printf("Incoming Parameters: "); + for (; refPosIterator != refPositions.end() && currentRefPosition->refType != RefTypeBB; + ++refPosIterator, currentRefPosition = &refPosIterator) { -#ifdef TARGET_ARM - regNumber sourceReg = (regNumber)source[targetReg]; - Interval* interval = sourceIntervals[sourceReg]; - if (interval->registerType == TYP_DOUBLE) + Interval* interval = currentRefPosition->getInterval(); + assert(interval != nullptr && interval->isLocalVar); + printf(" V%02d", interval->varNum); + if (mode == LSRA_DUMP_POST) { - // For ARM32, make sure that both of the float halves of the double register are available. - assert(genIsValidDoubleReg(targetReg)); - regNumber anotherHalfRegNum = REG_NEXT(targetReg); - if (location[anotherHalfRegNum] == REG_NA) + regNumber reg; + if (currentRefPosition->registerAssignment == RBM_NONE) { - targetRegsReady |= targetRegMask; + reg = REG_STK; } - } - else -#endif // TARGET_ARM - { - targetRegsReady |= targetRegMask; + else + { + reg = currentRefPosition->assignedReg(); + } + LclVarDsc* varDsc = &(compiler->lvaTable[interval->varNum]); + printf("("); + regNumber assignedReg = varDsc->GetRegNum(); + regNumber argReg = (varDsc->lvIsRegArg) ? varDsc->GetArgReg() : REG_STK; + + assert(reg == assignedReg || varDsc->lvRegister == false); + if (reg != argReg) + { + printf(getRegName(argReg, isFloatRegType(interval->registerType))); + printf("=>"); + } + printf("%s)", getRegName(reg, isFloatRegType(interval->registerType))); } } + printf("\n"); } - // Perform reg to reg moves - while (targetRegsToDo != RBM_NONE) + for (block = startBlockSequence(); block != nullptr; block = moveToNextBlock()) { - while (targetRegsReady != RBM_NONE) - { - regMaskTP targetRegMask = genFindLowestBit(targetRegsReady); - targetRegsToDo &= ~targetRegMask; - targetRegsReady &= ~targetRegMask; - regNumber targetReg = genRegNumFromMask(targetRegMask); - assert(location[targetReg] != targetReg); - assert(targetReg < REG_COUNT); - regNumber sourceReg = (regNumber)source[targetReg]; - assert(sourceReg < REG_COUNT); - regNumber fromReg = (regNumber)location[sourceReg]; - // stack to reg movs should be done last as part of "targetRegsFromStack" - assert(fromReg < REG_STK); - Interval* interval = sourceIntervals[sourceReg]; - assert(interval != nullptr); - addResolution(block, insertionPoint, interval, targetReg, fromReg); - JITDUMP(" (%s)\n", resolveTypeName[resolveType]); - sourceIntervals[sourceReg] = nullptr; - location[sourceReg] = REG_NA; - regMaskTP fromRegMask = genRegMask(fromReg); + currentLoc += 2; - // Do we have a free targetReg? - if (fromReg == sourceReg) + if (mode == LSRA_DUMP_REFPOS) + { + bool printedBlockHeader = false; + // We should find the boundary RefPositions in the order of exposed uses, dummy defs, and the blocks + for (; refPosIterator != refPositions.end() && + (currentRefPosition->refType == RefTypeExpUse || currentRefPosition->refType == RefTypeDummyDef || + (currentRefPosition->refType == RefTypeBB && !printedBlockHeader)); + ++refPosIterator, currentRefPosition = &refPosIterator) { - if (source[fromReg] != REG_NA && ((targetRegsFromStack & fromRegMask) != fromRegMask)) + Interval* interval = nullptr; + if (currentRefPosition->isIntervalRef()) { - targetRegsReady |= fromRegMask; -#ifdef TARGET_ARM - if (genIsValidDoubleReg(fromReg)) - { - // Ensure that either: - // - the Interval targeting fromReg is not double, or - // - the other half of the double is free. - Interval* otherInterval = sourceIntervals[source[fromReg]]; - regNumber upperHalfReg = REG_NEXT(fromReg); - if ((otherInterval->registerType == TYP_DOUBLE) && (location[upperHalfReg] != REG_NA)) - { - targetRegsReady &= ~fromRegMask; - } - } + interval = currentRefPosition->getInterval(); } - else if (genIsValidFloatReg(fromReg) && !genIsValidDoubleReg(fromReg)) + switch (currentRefPosition->refType) { - // We may have freed up the other half of a double where the lower half - // was already free. - regNumber lowerHalfReg = REG_PREV(fromReg); - regNumber lowerHalfSrcReg = (regNumber)source[lowerHalfReg]; - regNumber lowerHalfSrcLoc = (regNumber)location[lowerHalfReg]; - regMaskTP lowerHalfRegMask = genRegMask(lowerHalfReg); - // Necessary conditions: - // - There is a source register for this reg (lowerHalfSrcReg != REG_NA) - // - It is currently free (lowerHalfSrcLoc == REG_NA) - // - The source interval isn't yet completed (sourceIntervals[lowerHalfSrcReg] != nullptr) - // - It's not in the ready set ((targetRegsReady & lowerHalfRegMask) == - // RBM_NONE) - // - It's not resolved from stack ((targetRegsFromStack & lowerHalfRegMask) != - // lowerHalfRegMask) - if ((lowerHalfSrcReg != REG_NA) && (lowerHalfSrcLoc == REG_NA) && - (sourceIntervals[lowerHalfSrcReg] != nullptr) && - ((targetRegsReady & lowerHalfRegMask) == RBM_NONE) && - ((targetRegsFromStack & lowerHalfRegMask) != lowerHalfRegMask)) - { - // This must be a double interval, otherwise it would be in targetRegsReady, or already - // completed. - assert(sourceIntervals[lowerHalfSrcReg]->registerType == TYP_DOUBLE); - targetRegsReady |= lowerHalfRegMask; - } -#endif // TARGET_ARM + case RefTypeExpUse: + assert(interval != nullptr); + assert(interval->isLocalVar); + printf(" Exposed use of V%02u at #%d\n", interval->varNum, currentRefPosition->rpNum); + break; + case RefTypeDummyDef: + assert(interval != nullptr); + assert(interval->isLocalVar); + printf(" Dummy def of V%02u at #%d\n", interval->varNum, currentRefPosition->rpNum); + break; + case RefTypeBB: + block->dspBlockHeader(compiler); + printedBlockHeader = true; + printf("=====\n"); + break; + default: + printf("Unexpected RefPosition type at #%d\n", currentRefPosition->rpNum); + break; } } } - if (targetRegsToDo != RBM_NONE) + else + { + block->dspBlockHeader(compiler); + printf("=====\n"); + } + if (enregisterLocalVars && mode == LSRA_DUMP_POST && block != compiler->fgFirstBB && + block->bbNum <= bbNumMaxBeforeResolution) + { + printf("Predecessor for variable locations: " FMT_BB "\n", blockInfo[block->bbNum].predBBNum); + dumpInVarToRegMap(block); + } + if (block->bbNum > bbNumMaxBeforeResolution) + { + SplitEdgeInfo splitEdgeInfo; + splitBBNumToTargetBBNumMap->Lookup(block->bbNum, &splitEdgeInfo); + assert(splitEdgeInfo.toBBNum <= bbNumMaxBeforeResolution); + assert(splitEdgeInfo.fromBBNum <= bbNumMaxBeforeResolution); + printf("New block introduced for resolution from " FMT_BB " to " FMT_BB "\n", splitEdgeInfo.fromBBNum, + splitEdgeInfo.toBBNum); + } + + for (GenTree* node : LIR::AsRange(block).NonPhiNodes()) { - regMaskTP targetRegMask = genFindLowestBit(targetRegsToDo); - regNumber targetReg = genRegNumFromMask(targetRegMask); + GenTree* tree = node; - // Is it already there due to other moves? - // If not, move it to the temp reg, OR swap it with another register - regNumber sourceReg = (regNumber)source[targetReg]; - regNumber fromReg = (regNumber)location[sourceReg]; - if (targetReg == fromReg) - { - targetRegsToDo &= ~targetRegMask; - } - else + int produce = tree->IsValue() ? ComputeOperandDstCount(tree) : 0; + int consume = ComputeAvailableSrcCount(tree); + + lsraDispNode(tree, mode, produce != 0 && mode != LSRA_DUMP_REFPOS); + + if (mode != LSRA_DUMP_REFPOS) { - regNumber tempReg = REG_NA; - bool useSwap = false; - if (emitter::isFloatReg(targetReg)) + if (consume > 0) { -#ifdef TARGET_ARM - if (sourceIntervals[fromReg]->registerType == TYP_DOUBLE) + printf("; "); + + bool first = true; + for (GenTree* operand : tree->Operands()) { - // ARM32 requires a double temp register for TYP_DOUBLE. - tempReg = tempRegDbl; + DumpOperandDefs(operand, first, mode, operandString, operandStringLength); } - else -#endif // TARGET_ARM - tempReg = tempRegFlt; - } -#ifdef TARGET_XARCH - else - { - useSwap = true; - } -#else // !TARGET_XARCH - - else - { - tempReg = tempRegInt; } - -#endif // !TARGET_XARCH - if (useSwap || tempReg == REG_NA) + } + else + { + // Print each RefPosition on a new line, but + // printing all the kills for each node on a single line + // and combining the fixed regs with their associated def or use + bool killPrinted = false; + RefPosition* lastFixedRegRefPos = nullptr; + for (; refPosIterator != refPositions.end() && + (currentRefPosition->refType == RefTypeUse || currentRefPosition->refType == RefTypeFixedReg || + currentRefPosition->refType == RefTypeKill || currentRefPosition->refType == RefTypeDef) && + (currentRefPosition->nodeLocation == tree->gtSeqNum || + currentRefPosition->nodeLocation == tree->gtSeqNum + 1); + ++refPosIterator, currentRefPosition = &refPosIterator) { - // First, we have to figure out the destination register for what's currently in fromReg, - // so that we can find its sourceInterval. - regNumber otherTargetReg = REG_NA; - - // By chance, is fromReg going where it belongs? - if (location[source[fromReg]] == targetReg) + Interval* interval = nullptr; + if (currentRefPosition->isIntervalRef()) { - otherTargetReg = fromReg; - // If we can swap, we will be done with otherTargetReg as well. - // Otherwise, we'll spill it to the stack and reload it later. - if (useSwap) - { - regMaskTP fromRegMask = genRegMask(fromReg); - targetRegsToDo &= ~fromRegMask; - } + interval = currentRefPosition->getInterval(); } - else + switch (currentRefPosition->refType) { - // Look at the remaining registers from targetRegsToDo (which we expect to be relatively - // small at this point) to find out what's currently in targetReg. - regMaskTP mask = targetRegsToDo; - while (mask != RBM_NONE && otherTargetReg == REG_NA) - { - regMaskTP nextRegMask = genFindLowestBit(mask); - regNumber nextReg = genRegNumFromMask(nextRegMask); - mask &= ~nextRegMask; - if (location[source[nextReg]] == targetReg) + case RefTypeUse: + if (currentRefPosition->IsPhysRegRef()) { - otherTargetReg = nextReg; + printf("\n Use:R%d(#%d)", + currentRefPosition->getReg()->regNum, currentRefPosition->rpNum); } - } - } - assert(otherTargetReg != REG_NA); - - if (useSwap) - { - // Generate a "swap" of fromReg and targetReg - insertSwap(block, insertionPoint, sourceIntervals[source[otherTargetReg]]->varNum, targetReg, - sourceIntervals[sourceReg]->varNum, fromReg); - location[sourceReg] = REG_NA; - location[source[otherTargetReg]] = (regNumberSmall)fromReg; - - INTRACK_STATS(updateLsraStat(STAT_RESOLUTION_MOV, block->bbNum)); - } - else - { - // Spill "targetReg" to the stack and add its eventual target (otherTargetReg) - // to "targetRegsFromStack", which will be handled below. - // NOTE: This condition is very rare. Setting COMPlus_JitStressRegs=0x203 - // has been known to trigger it in JIT SH. - - // First, spill "otherInterval" from targetReg to the stack. - Interval* otherInterval = sourceIntervals[source[otherTargetReg]]; - setIntervalAsSpilled(otherInterval); - addResolution(block, insertionPoint, otherInterval, REG_STK, targetReg); - JITDUMP(" (%s)\n", resolveTypeName[resolveType]); - location[source[otherTargetReg]] = REG_STK; - - regMaskTP otherTargetRegMask = genRegMask(otherTargetReg); - targetRegsFromStack |= otherTargetRegMask; - stackToRegIntervals[otherTargetReg] = otherInterval; - targetRegsToDo &= ~otherTargetRegMask; - - // Now, move the interval that is going to targetReg. - addResolution(block, insertionPoint, sourceIntervals[sourceReg], targetReg, fromReg); - JITDUMP(" (%s)\n", resolveTypeName[resolveType]); - location[sourceReg] = REG_NA; - - // Add its "fromReg" to "targetRegsReady", only if: - // - It was one of the target register we originally determined. - // - It is not the eventual target (otherTargetReg) because its - // value will be retrieved from STK. - if (source[fromReg] != REG_NA && fromReg != otherTargetReg) - { - regMaskTP fromRegMask = genRegMask(fromReg); - targetRegsReady |= fromRegMask; -#ifdef TARGET_ARM - if (genIsValidDoubleReg(fromReg)) + else { - // Ensure that either: - // - the Interval targeting fromReg is not double, or - // - the other half of the double is free. - Interval* otherInterval = sourceIntervals[source[fromReg]]; - regNumber upperHalfReg = REG_NEXT(fromReg); - if ((otherInterval->registerType == TYP_DOUBLE) && (location[upperHalfReg] != REG_NA)) + assert(interval != nullptr); + printf("\n Use:"); + interval->microDump(); + printf("(#%d)", currentRefPosition->rpNum); + if (currentRefPosition->isFixedRegRef && !interval->isInternal) { - targetRegsReady &= ~fromRegMask; + assert(genMaxOneBit(currentRefPosition->registerAssignment)); + assert(lastFixedRegRefPos != nullptr); + printf(" Fixed:%s(#%d)", getRegName(currentRefPosition->assignedReg(), + isFloatRegType(interval->registerType)), + lastFixedRegRefPos->rpNum); + lastFixedRegRefPos = nullptr; + } + if (currentRefPosition->isLocalDefUse) + { + printf(" LocalDefUse"); + } + if (currentRefPosition->lastUse) + { + printf(" *"); } } -#endif // TARGET_ARM + break; + case RefTypeDef: + { + // Print each def on a new line + assert(interval != nullptr); + printf("\n Def:"); + interval->microDump(); + printf("(#%d)", currentRefPosition->rpNum); + if (currentRefPosition->isFixedRegRef) + { + assert(genMaxOneBit(currentRefPosition->registerAssignment)); + printf(" %s", getRegName(currentRefPosition->assignedReg(), + isFloatRegType(interval->registerType))); + } + if (currentRefPosition->isLocalDefUse) + { + printf(" LocalDefUse"); + } + if (currentRefPosition->lastUse) + { + printf(" *"); + } + if (interval->relatedInterval != nullptr) + { + printf(" Pref:"); + interval->relatedInterval->microDump(); + } } + break; + case RefTypeKill: + if (!killPrinted) + { + printf("\n Kill: "); + killPrinted = true; + } + printf(getRegName(currentRefPosition->assignedReg(), + isFloatRegType(currentRefPosition->getReg()->registerType))); + printf(" "); + break; + case RefTypeFixedReg: + lastFixedRegRefPos = currentRefPosition; + break; + default: + printf("Unexpected RefPosition type at #%d\n", currentRefPosition->rpNum); + break; } - targetRegsToDo &= ~targetRegMask; - } - else - { - compiler->codeGen->regSet.rsSetRegsModified(genRegMask(tempReg) DEBUGARG(true)); -#ifdef TARGET_ARM - if (sourceIntervals[fromReg]->registerType == TYP_DOUBLE) - { - assert(genIsValidDoubleReg(targetReg)); - assert(genIsValidDoubleReg(tempReg)); - - addResolutionForDouble(block, insertionPoint, sourceIntervals, location, tempReg, targetReg, - resolveType); - } - else -#endif // TARGET_ARM - { - assert(sourceIntervals[targetReg] != nullptr); - - addResolution(block, insertionPoint, sourceIntervals[targetReg], tempReg, targetReg); - JITDUMP(" (%s)\n", resolveTypeName[resolveType]); - location[targetReg] = (regNumberSmall)tempReg; - } - targetRegsReady |= targetRegMask; } } - } - } - - // Finally, perform stack to reg moves - // All the target regs will be empty at this point - while (targetRegsFromStack != RBM_NONE) - { - regMaskTP targetRegMask = genFindLowestBit(targetRegsFromStack); - targetRegsFromStack &= ~targetRegMask; - regNumber targetReg = genRegNumFromMask(targetRegMask); - - Interval* interval = stackToRegIntervals[targetReg]; - assert(interval != nullptr); - - addResolution(block, insertionPoint, interval, targetReg, REG_STK); - JITDUMP(" (%s)\n", resolveTypeName[resolveType]); + printf("\n"); + } + if (enregisterLocalVars && mode == LSRA_DUMP_POST) + { + dumpOutVarToRegMap(block); + } + printf("\n"); } + printf("\n\n"); } -#if TRACK_LSRA_STATS - -#if TRACK_LSRA_STATS -const char* LinearScan::getStatName(unsigned stat) -{ - LsraStat lsraStat = (LsraStat)stat; - assert(lsraStat != LsraStat::COUNT); - - static const char* const lsraStatNames[] = { -#define LSRA_STAT_DEF(stat, name) name, -#include "lsra_stats.h" -#undef LSRA_STAT_DEF -#define REG_SEL_DEF(stat, value, shortname) #stat, -#include "lsra_score.h" -#undef REG_SEL_DEF - }; - - assert(stat < ArrLen(lsraStatNames)); - return lsraStatNames[lsraStat]; -} -#endif // TRACK_LSRA_STATS - -// ---------------------------------------------------------- -// updateLsraStat: Increment LSRA stat counter. -// -// Arguments: -// stat - LSRA stat enum -// bbNum - Basic block to which LSRA stat needs to be -// associated with. -// -void LinearScan::updateLsraStat(LsraStat stat, unsigned bbNum) +void LinearScan::dumpLsraAllocationEvent( + LsraDumpEvent event, Interval* interval, regNumber reg, BasicBlock* currentBlock, RegisterScore registerScore) { - if (bbNum > bbNumMaxBeforeResolution) + if (!(VERBOSE)) { - // This is a newly created basic block as part of resolution. - // These blocks contain resolution moves that are already accounted. return; } - - ++(blockInfo[bbNum].stats[(unsigned)stat]); -} - -// ----------------------------------------------------------- -// dumpLsraStats - dumps Lsra stats to given file. -// -// Arguments: -// file - file to which stats are to be written. -// -void LinearScan::dumpLsraStats(FILE* file) -{ - unsigned sumStats[LsraStat::COUNT] = {0}; - BasicBlock::weight_t wtdStats[LsraStat::COUNT] = {0}; - - fprintf(file, "----------\n"); - fprintf(file, "LSRA Stats"); -#ifdef DEBUG - if (!VERBOSE) + if ((interval != nullptr) && (reg != REG_NA) && (reg != REG_STK)) { - fprintf(file, " : %s\n", compiler->info.compFullName); + registersToDump |= getRegMask(reg, interval->registerType); + dumpRegRecordTitleIfNeeded(); } - else + + switch (event) { - // In verbose mode no need to print full name - // while printing lsra stats. - fprintf(file, "\n"); - } -#else - fprintf(file, " : %s\n", compiler->eeGetMethodFullName(compiler->info.compCompHnd)); -#endif + // Conflicting def/use + case LSRA_EVENT_DEFUSE_CONFLICT: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("DUconflict "); + dumpRegRecords(); + break; + case LSRA_EVENT_DEFUSE_CASE1: + printf(indentFormat, " Case #1 use defRegAssignment"); + dumpRegRecords(); + break; + case LSRA_EVENT_DEFUSE_CASE2: + printf(indentFormat, " Case #2 use useRegAssignment"); + dumpRegRecords(); + break; + case LSRA_EVENT_DEFUSE_CASE3: + printf(indentFormat, " Case #3 use useRegAssignment"); + dumpRegRecords(); + dumpRegRecords(); + break; + case LSRA_EVENT_DEFUSE_CASE4: + printf(indentFormat, " Case #4 use defRegAssignment"); + dumpRegRecords(); + break; + case LSRA_EVENT_DEFUSE_CASE5: + printf(indentFormat, " Case #5 set def to all regs"); + dumpRegRecords(); + break; + case LSRA_EVENT_DEFUSE_CASE6: + printf(indentFormat, " Case #6 need a copy"); + dumpRegRecords(); + if (interval == nullptr) + { + printf(indentFormat, " NULL interval"); + dumpRegRecords(); + } + else if (interval->firstRefPosition->multiRegIdx != 0) + { + printf(indentFormat, " (multiReg)"); + dumpRegRecords(); + } + break; - fprintf(file, "----------\n"); - fprintf(file, "Total Tracked Vars: %d\n", compiler->lvaTrackedCount); - fprintf(file, "Total Reg Cand Vars: %d\n", regCandidateVarCount); - fprintf(file, "Total number of Intervals: %d\n", - static_cast((intervals.size() == 0 ? 0 : (intervals.size() - 1)))); - fprintf(file, "Total number of RefPositions: %d\n", static_cast(refPositions.size() - 1)); + case LSRA_EVENT_SPILL: + dumpRefPositionShort(activeRefPosition, currentBlock); + assert(interval != nullptr && interval->assignedReg != nullptr); + printf("Spill %-4s ", getRegName(interval->assignedReg->regNum)); + dumpRegRecords(); + break; - // compute total number of spill temps created - unsigned numSpillTemps = 0; - for (int i = 0; i < TYP_COUNT; i++) - { - numSpillTemps += maxSpill[i]; - } - fprintf(file, "Total Number of spill temps created: %d\n", numSpillTemps); - fprintf(file, "..........\n"); - bool addedBlockHeader = false; - bool anyNonZeroStat = false; + // Restoring the previous register + case LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL: + case LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL_AFTER_SPILL: + assert(interval != nullptr); + if ((activeRefPosition == nullptr) || (activeRefPosition->refType == RefTypeBB)) + { + printf(emptyRefPositionFormat, ""); + } + else + { + dumpRefPositionShort(activeRefPosition, currentBlock); + } + printf((event == LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL) ? "Restr %-4s " : "SRstr %-4s ", + getRegName(reg)); + dumpRegRecords(); + break; - for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) - { - unsigned lsraStat = blockInfo[0].stats[statIndex]; + case LSRA_EVENT_DONE_KILL_GC_REFS: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("Done "); + break; - if (lsraStat != 0) - { - if (!addedBlockHeader) + case LSRA_EVENT_NO_GC_KILLS: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("None "); + break; + + // Block boundaries + case LSRA_EVENT_START_BB: + // The RefTypeBB comes after the RefTypeDummyDefs associated with that block, + // so we may have a RefTypeDummyDef at the time we dump this event. + // In that case we'll have another "EVENT" associated with it, so we need to + // print the full line now. + if (activeRefPosition->refType != RefTypeBB) { - addedBlockHeader = true; - fprintf(file, FMT_BB " [%8.2f]: ", 0, blockInfo[0].weight); - fprintf(file, "%s = %d", getStatName(statIndex), lsraStat); + dumpNewBlock(currentBlock, activeRefPosition->nodeLocation); + dumpRegRecords(); } else { - fprintf(file, ", %s = %d", getStatName(statIndex), lsraStat); + dumpRefPositionShort(activeRefPosition, currentBlock); } + break; - sumStats[statIndex] += lsraStat; - wtdStats[statIndex] += (lsraStat * blockInfo[0].weight); - anyNonZeroStat = true; - } - } - if (anyNonZeroStat) - { - fprintf(file, "\n"); - } + // Allocation decisions + case LSRA_EVENT_NEEDS_NEW_REG: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("Free %-4s ", getRegName(reg)); + dumpRegRecords(); + break; - for (BasicBlock* block = compiler->fgFirstBB; block != nullptr; block = block->bbNext) - { - if (block->bbNum > bbNumMaxBeforeResolution) - { - continue; - } + case LSRA_EVENT_ZERO_REF: + assert(interval != nullptr && interval->isLocalVar); + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("NoRef "); + dumpRegRecords(); + break; - addedBlockHeader = false; - anyNonZeroStat = false; - for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) - { - unsigned lsraStat = blockInfo[block->bbNum].stats[statIndex]; + case LSRA_EVENT_FIXED_REG: + case LSRA_EVENT_EXP_USE: + case LSRA_EVENT_KEPT_ALLOCATION: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("Keep %-4s ", getRegName(reg)); + break; - if (lsraStat != 0) + case LSRA_EVENT_COPY_REG: + assert(interval != nullptr && interval->recentRefPosition != nullptr); + dumpRefPositionShort(activeRefPosition, currentBlock); + if (allocationPassComplete || (registerScore == 0)) { - if (!addedBlockHeader) - { - addedBlockHeader = true; - fprintf(file, FMT_BB " [%8.2f]: ", block->bbNum, block->bbWeight); - fprintf(file, "%s = %d", getStatName(statIndex), lsraStat); - } - else - { - fprintf(file, ", %s = %d", getStatName(statIndex), lsraStat); - } + printf("Copy %-4s ", getRegName(reg)); + } + else + { + printf("%-5s(C) %-4s ", getScoreName(registerScore), getRegName(reg)); + } + break; - sumStats[statIndex] += lsraStat; - wtdStats[statIndex] += (lsraStat * block->bbWeight); - anyNonZeroStat = true; + case LSRA_EVENT_MOVE_REG: + assert(interval != nullptr && interval->recentRefPosition != nullptr); + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("Move %-4s ", getRegName(reg)); + dumpRegRecords(); + break; + + case LSRA_EVENT_ALLOC_REG: + dumpRefPositionShort(activeRefPosition, currentBlock); + if (allocationPassComplete || (registerScore == 0)) + { + printf("Alloc %-4s ", getRegName(reg)); + } + else + { + printf("%-5s(A) %-4s ", getScoreName(registerScore), getRegName(reg)); } - } - if (anyNonZeroStat) - { - fprintf(file, "\n"); - } - } + break; - fprintf(file, "..........\n"); - for (int regSelectI = 0; regSelectI < LsraStat::COUNT; regSelectI++) - { - if (regSelectI == firstRegSelStat) - { - fprintf(file, "..........\n"); - } - // TODO-review: I don't see a point of displaying Stats (SpillCount, etc.) if they are zero. Thoughts? - if ((regSelectI < firstRegSelStat) || (sumStats[regSelectI] != 0)) - { - // Print register selection stats - if (regSelectI >= firstRegSelStat) + case LSRA_EVENT_REUSE_REG: + dumpRefPositionShort(activeRefPosition, currentBlock); + if (allocationPassComplete || (registerScore == 0)) { - fprintf(file, "Total %s [#%2d] : %d Weighted: %f\n", getStatName(regSelectI), - (regSelectI - firstRegSelStat + 1), sumStats[regSelectI], wtdStats[regSelectI]); + printf("Reuse %-4s ", getRegName(reg)); } else { - fprintf(file, "Total %s : %d Weighted: %f\n", getStatName(regSelectI), sumStats[regSelectI], - wtdStats[regSelectI]); + printf("%-5s(A) %-4s ", getScoreName(registerScore), getRegName(reg)); } - } - } - printf("\n"); -} - -// ----------------------------------------------------------- -// dumpLsraStatsCsvFormat - dumps Lsra stats to given file in csv format. -// -// Arguments: -// file - file to which stats are to be written. -// -void LinearScan::dumpLsraStatsCsv(FILE* file) -{ - unsigned sumStats[LsraStat::COUNT] = {0}; + break; - // Write the header if the file is empty - if (ftell(file) == 0) - { - // header - fprintf(file, "\"Method Name\""); - for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) - { - fprintf(file, ",\"%s\"", LinearScan::getStatName(statIndex)); - } - fprintf(file, ",\"PerfScore\"\n"); - } + case LSRA_EVENT_NO_ENTRY_REG_ALLOCATED: + assert(interval != nullptr && interval->isLocalVar); + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("LoRef "); + break; - // bbNum == 0 - for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) - { - sumStats[statIndex] += blockInfo[0].stats[statIndex]; - } + case LSRA_EVENT_NO_REG_ALLOCATED: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("NoReg "); + break; - // blocks - for (BasicBlock* block = compiler->fgFirstBB; block != nullptr; block = block->bbNext) - { - if (block->bbNum > bbNumMaxBeforeResolution) - { - continue; - } + case LSRA_EVENT_RELOAD: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("ReLod %-4s ", getRegName(reg)); + dumpRegRecords(); + break; - for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) - { - sumStats[statIndex] += blockInfo[block->bbNum].stats[statIndex]; - } - } + case LSRA_EVENT_SPECIAL_PUTARG: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("PtArg %-4s ", getRegName(reg)); + break; - fprintf(file, "\"%s\"", compiler->info.compFullName); - for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++) - { - fprintf(file, ",%u", sumStats[statIndex]); - } - fprintf(file, ",%.2f\n", compiler->info.compPerfScore); -} -#endif // TRACK_LSRA_STATS + case LSRA_EVENT_UPPER_VECTOR_SAVE: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("UVSav %-4s ", getRegName(reg)); + break; -#ifdef DEBUG -void dumpRegMask(regMaskTP regs) -{ - if (regs == RBM_ALLINT) - { - printf("[allInt]"); - } - else if (regs == (RBM_ALLINT & ~RBM_FPBASE)) - { - printf("[allIntButFP]"); - } - else if (regs == RBM_ALLFLOAT) - { - printf("[allFloat]"); - } - else if (regs == RBM_ALLDOUBLE) - { - printf("[allDouble]"); - } - else - { - dspRegMask(regs); - } -} + case LSRA_EVENT_UPPER_VECTOR_RESTORE: + dumpRefPositionShort(activeRefPosition, currentBlock); + printf("UVRes %-4s ", getRegName(reg)); + break; -static const char* getRefTypeName(RefType refType) -{ - switch (refType) - { -#define DEF_REFTYPE(memberName, memberValue, shortName) \ - case memberName: \ - return #memberName; -#include "lsra_reftypes.h" -#undef DEF_REFTYPE - default: - return nullptr; - } -} + // We currently don't dump anything for these events. + case LSRA_EVENT_DEFUSE_FIXED_DELAY_USE: + case LSRA_EVENT_SPILL_EXTENDED_LIFETIME: + case LSRA_EVENT_END_BB: + case LSRA_EVENT_FREE_REGS: + case LSRA_EVENT_INCREMENT_RANGE_END: + case LSRA_EVENT_LAST_USE: + case LSRA_EVENT_LAST_USE_DELAYED: + break; -static const char* getRefTypeShortName(RefType refType) -{ - switch (refType) - { -#define DEF_REFTYPE(memberName, memberValue, shortName) \ - case memberName: \ - return shortName; -#include "lsra_reftypes.h" -#undef DEF_REFTYPE default: - return nullptr; + printf("????? %-4s ", getRegName(reg)); + dumpRegRecords(); + break; } } //------------------------------------------------------------------------ -// getScoreName: Returns the texual name of register score -const char* LinearScan::getScoreName(RegisterScore score) +// dumpRegRecordHeader: Dump the header for a column-based dump of the register state. +// +// Arguments: +// None. +// +// Return Value: +// None. +// +// Assumptions: +// Reg names fit in 4 characters (minimum width of the columns) +// +// Notes: +// In order to make the table as dense as possible (for ease of reading the dumps), +// we determine the minimum regColumnWidth width required to represent: +// regs, by name (e.g. eax or xmm0) - this is fixed at 4 characters. +// intervals, as Vnn for lclVar intervals, or as I for other intervals. +// The table is indented by the amount needed for dumpRefPositionShort, which is +// captured in shortRefPositionDumpWidth. +// +void LinearScan::dumpRegRecordHeader() { - switch (score) + printf("The following table has one or more rows for each RefPosition that is handled during allocation.\n" + "The first column provides the basic information about the RefPosition, with its type (e.g. Def,\n" + "Use, Fixd) followed by a '*' if it is a last use, and a 'D' if it is delayRegFree, and then the\n" + "action taken during allocation (e.g. Alloc a new register, or Keep an existing one).\n" + "The subsequent columns show the Interval occupying each register, if any, followed by 'a' if it is\n" + "active, a 'p' if it is a large vector that has been partially spilled, and 'i'if it is inactive.\n" + "Columns are only printed up to the last modifed register, which may increase during allocation,\n" + "in which case additional columns will appear. \n" + "Registers which are not marked modified have ---- in their column.\n\n"); + + // First, determine the width of each register column (which holds a reg name in the + // header, and an interval name in each subsequent row). + int intervalNumberWidth = (int)log10((double)intervals.size()) + 1; + // The regColumnWidth includes the identifying character (I or V) and an 'i', 'p' or 'a' (inactive, + // partially-spilled or active) + regColumnWidth = intervalNumberWidth + 2; + if (regColumnWidth < 4) { -#define REG_SEL_DEF(stat, value, shortname) \ - case stat: \ - return shortname; -#include "lsra_score.h" -#undef REG_SEL_DEF - default: - return " - "; + regColumnWidth = 4; } -} - -void RefPosition::dump() -{ - printf("IsPhysRegRef()) - { - this->getReg()->tinyDump(); - } - else if (getInterval()) - { - this->getInterval()->tinyDump(); - } - if (this->treeNode) - { - printf("%s", treeNode->OpName(treeNode->OperGet())); - if (this->treeNode->IsMultiRegNode()) - { - printf("[%d]", this->multiRegIdx); - } - } - printf(" " FMT_BB " ", this->bbNum); + maxNodeLocation = (maxNodeLocation == 0) + ? 1 + : maxNodeLocation; // corner case of a method with an infinite loop without any gentree nodes + assert(maxNodeLocation >= 1); + assert(refPositions.size() >= 1); + int nodeLocationWidth = (int)log10((double)maxNodeLocation) + 1; + int refPositionWidth = (int)log10((double)refPositions.size()) + 1; + int refTypeInfoWidth = 4 /*TYPE*/ + 2 /* last-use and delayed */ + 1 /* space */; + int locationAndRPNumWidth = nodeLocationWidth + 2 /* .# */ + refPositionWidth + 1 /* space */; + int shortRefPositionDumpWidth = locationAndRPNumWidth + regColumnWidth + 1 /* space */ + refTypeInfoWidth; + sprintf_s(shortRefPositionFormat, MAX_FORMAT_CHARS, "%%%dd.#%%-%dd ", nodeLocationWidth, refPositionWidth); + sprintf_s(emptyRefPositionFormat, MAX_FORMAT_CHARS, "%%-%ds", shortRefPositionDumpWidth); - printf("regmask="); - dumpRegMask(registerAssignment); + // The width of the "allocation info" + // - a 8-character allocation decision + // - a space + // - a 4-character register + // - a space + int allocationInfoWidth = 8 + 1 + 4 + 1; - printf(" minReg=%d", minRegCandidateCount); + // Next, determine the width of the legend for each row. This includes: + // - a short RefPosition dump (shortRefPositionDumpWidth), which includes a space + // - the allocation info (allocationInfoWidth), which also includes a space - if (this->lastUse) - { - printf(" last"); - } - if (this->reload) - { - printf(" reload"); - } - if (this->spillAfter) - { - printf(" spillAfter"); - } - if (this->writeThru) - { - printf(" writeThru"); - } - if (this->moveReg) - { - printf(" move"); - } - if (this->copyReg) - { - printf(" copy"); - } - if (this->isFixedRegRef) + regTableIndent = shortRefPositionDumpWidth + allocationInfoWidth; + + // BBnn printed left-justified in the NAME Typeld and allocationInfo space. + int bbNumWidth = (int)log10((double)compiler->fgBBNumMax) + 1; + // In the unlikely event that BB numbers overflow the space, we'll simply omit the predBB + int predBBNumDumpSpace = regTableIndent - locationAndRPNumWidth - bbNumWidth - 9; // 'BB' + ' PredBB' + if (predBBNumDumpSpace < bbNumWidth) { - printf(" fixed"); + sprintf_s(bbRefPosFormat, MAX_LEGEND_FORMAT_CHARS, "BB%%-%dd", shortRefPositionDumpWidth - 2); } - if (this->isLocalDefUse) + else { - printf(" local"); + sprintf_s(bbRefPosFormat, MAX_LEGEND_FORMAT_CHARS, "BB%%-%dd PredBB%%-%dd", bbNumWidth, predBBNumDumpSpace); } - if (this->delayRegFree) + + if (compiler->shouldDumpASCIITrees()) { - printf(" delay"); + columnSeparator = "|"; + line = "-"; + leftBox = "+"; + middleBox = "+"; + rightBox = "+"; } - if (this->outOfOrder) + else { - printf(" outOfOrder"); + columnSeparator = "\xe2\x94\x82"; + line = "\xe2\x94\x80"; + leftBox = "\xe2\x94\x9c"; + middleBox = "\xe2\x94\xbc"; + rightBox = "\xe2\x94\xa4"; } + sprintf_s(indentFormat, MAX_FORMAT_CHARS, "%%-%ds", regTableIndent); - if (this->RegOptional()) - { - printf(" regOptional"); - } - printf(">\n"); -} + // Now, set up the legend format for the RefPosition info + sprintf_s(legendFormat, MAX_LEGEND_FORMAT_CHARS, "%%-%d.%ds%%-%d.%ds%%-%ds%%s", nodeLocationWidth + 1, + nodeLocationWidth + 1, refPositionWidth + 2, refPositionWidth + 2, regColumnWidth + 1); -void RegRecord::dump() -{ - tinyDump(); + // Print a "title row" including the legend and the reg names. + lastDumpedRegisters = RBM_NONE; + dumpRegRecordTitleIfNeeded(); } -void Interval::dump() +void LinearScan::dumpRegRecordTitleIfNeeded() { - printf("Interval %2u:", intervalIndex); - - if (isLocalVar) - { - printf(" (V%02u)", varNum); - } - else if (IsUpperVector()) - { - assert(relatedInterval != nullptr); - printf(" (U%02u)", relatedInterval->varNum); - } - printf(" %s", varTypeName(registerType)); - if (isInternal) - { - printf(" (INTERNAL)"); - } - if (isSpilled) - { - printf(" (SPILLED)"); - } - if (isSplit) - { - printf(" (SPLIT)"); - } - if (isStructField) - { - printf(" (field)"); - } - if (isPromotedStruct) - { - printf(" (promoted struct)"); - } - if (hasConflictingDefUse) - { - printf(" (def-use conflict)"); - } - if (hasInterferingUses) - { - printf(" (interfering uses)"); - } - if (isSpecialPutArg) - { - printf(" (specialPutArg)"); - } - if (isConstant) + if ((lastDumpedRegisters != registersToDump) || (rowCountSinceLastTitle > MAX_ROWS_BETWEEN_TITLES)) { - printf(" (constant)"); + lastUsedRegNumIndex = 0; + int lastRegNumIndex = compiler->compFloatingPointUsed ? REG_FP_LAST : REG_INT_LAST; + for (int regNumIndex = 0; regNumIndex <= lastRegNumIndex; regNumIndex++) + { + if ((registersToDump & genRegMask((regNumber)regNumIndex)) != 0) + { + lastUsedRegNumIndex = regNumIndex; + } + } + dumpRegRecordTitle(); + lastDumpedRegisters = registersToDump; } - if (isWriteThru) +} + +void LinearScan::dumpRegRecordTitleLines() +{ + for (int i = 0; i < regTableIndent; i++) { - printf(" (writeThru)"); + printf("%s", line); } - - printf(" RefPositions {"); - for (RefPosition* refPosition = this->firstRefPosition; refPosition != nullptr; - refPosition = refPosition->nextRefPosition) + for (int regNumIndex = 0; regNumIndex <= lastUsedRegNumIndex; regNumIndex++) { - printf("#%u@%u", refPosition->rpNum, refPosition->nodeLocation); - if (refPosition->nextRefPosition) + regNumber regNum = (regNumber)regNumIndex; + if (shouldDumpReg(regNum)) { - printf(" "); + printf("%s", middleBox); + for (int i = 0; i < regColumnWidth; i++) + { + printf("%s", line); + } } } - printf("}"); - - // this is not used (yet?) - // printf(" SpillOffset %d", this->spillOffset); - - printf(" physReg:%s", getRegName(physReg)); + printf("%s\n", rightBox); +} +void LinearScan::dumpRegRecordTitle() +{ + dumpRegRecordTitleLines(); - printf(" Preferences="); - dumpRegMask(this->registerPreferences); + // Print out the legend for the RefPosition info + printf(legendFormat, "Loc ", "RP# ", "Name ", "Type Action Reg "); - if (relatedInterval) + // Print out the register name column headers + char columnFormatArray[MAX_FORMAT_CHARS]; + sprintf_s(columnFormatArray, MAX_FORMAT_CHARS, "%s%%-%d.%ds", columnSeparator, regColumnWidth, regColumnWidth); + for (int regNumIndex = 0; regNumIndex <= lastUsedRegNumIndex; regNumIndex++) { - printf(" RelatedInterval "); - relatedInterval->microDump(); + regNumber regNum = (regNumber)regNumIndex; + if (shouldDumpReg(regNum)) + { + const char* regName = getRegName(regNum); + printf(columnFormatArray, regName); + } } + printf("%s\n", columnSeparator); - printf("\n"); + rowCountSinceLastTitle = 0; + + dumpRegRecordTitleLines(); } -// print out very concise representation -void Interval::tinyDump() +void LinearScan::dumpRegRecords() { - printf("varNum); - } - else if (isInternal) + static char columnFormatArray[18]; + + for (regNumber regNum = REG_FIRST; regNum <= (regNumber)lastUsedRegNumIndex; regNum = REG_NEXT(regNum)) { - printf(" internal"); + if (shouldDumpReg(regNum)) + { + printf("%s", columnSeparator); + RegRecord& regRecord = physRegs[regNum]; + Interval* interval = regRecord.assignedInterval; + if (interval != nullptr) + { + dumpIntervalName(interval); + char activeChar = interval->isActive ? 'a' : 'i'; +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + if (interval->isPartiallySpilled) + { + activeChar = 'p'; + } +#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE + printf("%c", activeChar); + } + else if ((genRegMask(regNum) & regsBusyUntilKill) != RBM_NONE) + { + printf(columnFormatArray, "Busy"); + } + else + { + sprintf_s(columnFormatArray, MAX_FORMAT_CHARS, "%%-%ds", regColumnWidth); + printf(columnFormatArray, ""); + } + } } - printf("> "); + printf("%s\n", columnSeparator); + rowCountSinceLastTitle++; } -// print out extremely concise representation -void Interval::microDump() +void LinearScan::dumpIntervalName(Interval* interval) { - if (isLocalVar) + if (interval->isLocalVar) { - printf("", varNum, intervalIndex); - return; + printf(intervalNameFormat, 'V', interval->varNum); } - else if (IsUpperVector()) + else if (interval->IsUpperVector()) { - assert(relatedInterval != nullptr); - printf(" (U%02u)", relatedInterval->varNum); + printf(intervalNameFormat, 'U', interval->relatedInterval->varNum); } - char intervalTypeChar = 'I'; - if (isInternal) + else if (interval->isConstant) { - intervalTypeChar = 'T'; + printf(intervalNameFormat, 'C', interval->intervalIndex); + } + else + { + printf(intervalNameFormat, 'I', interval->intervalIndex); } - printf("<%c%u>", intervalTypeChar, intervalIndex); } -void RegRecord::tinyDump() +void LinearScan::dumpEmptyRefPosition() { - printf(" ", getRegName(regNum)); + printf(emptyRefPositionFormat, ""); } -void LinearScan::dumpDefList() +//------------------------------------------------------------------------ +// dumpNewBlock: Dump a line for a new block in a column-based dump of the register state. +// +// Arguments: +// currentBlock - the new block to be dumped +// +void LinearScan::dumpNewBlock(BasicBlock* currentBlock, LsraLocation location) { if (!VERBOSE) { return; } - JITDUMP("DefList: { "); - bool first = true; - for (RefInfoListNode *listNode = defList.Begin(), *end = defList.End(); listNode != end; - listNode = listNode->Next()) + + // Always print a title row before a RefTypeBB (except for the first, because we + // will already have printed it before the parameters) + if ((currentBlock != compiler->fgFirstBB) && (currentBlock != nullptr)) { - GenTree* node = listNode->treeNode; - JITDUMP("%sN%03u.t%d. %s", first ? "" : "; ", node->gtSeqNum, node->gtTreeID, GenTree::OpName(node->OperGet())); - first = false; + dumpRegRecordTitle(); } - JITDUMP(" }\n"); -} - -void LinearScan::lsraDumpIntervals(const char* msg) -{ - printf("\nLinear scan intervals %s:\n", msg); - for (Interval& interval : intervals) + // If the activeRefPosition is a DummyDef, then don't print anything further (printing the + // title line makes it clearer that we're "about to" start the next block). + if (activeRefPosition->refType == RefTypeDummyDef) { - // only dump something if it has references - // if (interval->firstRefPosition) - interval.dump(); + dumpEmptyRefPosition(); + printf("DDefs "); + printf(regNameFormat, ""); + return; + } + printf(shortRefPositionFormat, location, activeRefPosition->rpNum); + if (currentBlock == nullptr) + { + printf(regNameFormat, "END"); + printf(" "); + printf(regNameFormat, ""); + } + else + { + printf(bbRefPosFormat, currentBlock->bbNum, + currentBlock == compiler->fgFirstBB ? 0 : blockInfo[currentBlock->bbNum].predBBNum); } - - printf("\n"); } -// Dumps a tree node as a destination or source operand, with the style -// of dump dependent on the mode -void LinearScan::lsraGetOperandString(GenTree* tree, - LsraTupleDumpMode mode, - char* operandString, - unsigned operandStringLength) +// Note that the size of this dump is computed in dumpRegRecordHeader(). +// +void LinearScan::dumpRefPositionShort(RefPosition* refPosition, BasicBlock* currentBlock) { - const char* lastUseChar = ""; - if (tree->OperIsScalarLocal() && ((tree->gtFlags & GTF_VAR_DEATH) != 0)) + static RefPosition* lastPrintedRefPosition = nullptr; + if (refPosition == lastPrintedRefPosition) { - lastUseChar = "*"; + dumpEmptyRefPosition(); + return; } - switch (mode) + lastPrintedRefPosition = refPosition; + if (refPosition->refType == RefTypeBB) { - case LinearScan::LSRA_DUMP_PRE: - case LinearScan::LSRA_DUMP_REFPOS: - _snprintf_s(operandString, operandStringLength, operandStringLength, "t%d%s", tree->gtTreeID, lastUseChar); - break; - case LinearScan::LSRA_DUMP_POST: - { - Compiler* compiler = JitTls::GetCompiler(); - - if (!tree->gtHasReg()) - { - _snprintf_s(operandString, operandStringLength, operandStringLength, "STK%s", lastUseChar); - } - else - { - regNumber reg = tree->GetRegNum(); - int charCount = _snprintf_s(operandString, operandStringLength, operandStringLength, "%s%s", - getRegName(reg, genIsValidFloatReg(reg)), lastUseChar); - operandString += charCount; - operandStringLength -= charCount; - - if (tree->IsMultiRegNode()) - { - unsigned regCount = tree->IsMultiRegLclVar() - ? compiler->lvaGetDesc(tree->AsLclVar()->GetLclNum())->lvFieldCnt - : tree->GetMultiRegCount(); - for (unsigned regIndex = 1; regIndex < regCount; regIndex++) - { - regNumber reg = tree->GetRegByIndex(regIndex); - charCount = _snprintf_s(operandString, operandStringLength, operandStringLength, ",%s%s", - getRegName(reg, genIsValidFloatReg(reg)), lastUseChar); - operandString += charCount; - operandStringLength -= charCount; - } - } - } - } - break; - default: - printf("ERROR: INVALID TUPLE DUMP MODE\n"); - break; + dumpNewBlock(currentBlock, refPosition->nodeLocation); + return; } -} -void LinearScan::lsraDispNode(GenTree* tree, LsraTupleDumpMode mode, bool hasDest) -{ - Compiler* compiler = JitTls::GetCompiler(); - const unsigned operandStringLength = 6 * MAX_MULTIREG_COUNT + 1; - char operandString[operandStringLength]; - const char* emptyDestOperand = " "; - char spillChar = ' '; - - if (mode == LinearScan::LSRA_DUMP_POST) + printf(shortRefPositionFormat, refPosition->nodeLocation, refPosition->rpNum); + if (refPosition->isIntervalRef()) { - if ((tree->gtFlags & GTF_SPILL) != 0) - { - spillChar = 'S'; - } - if (!hasDest && tree->gtHasReg()) + Interval* interval = refPosition->getInterval(); + dumpIntervalName(interval); + char lastUseChar = ' '; + char delayChar = ' '; + if (refPosition->lastUse) { - // A node can define a register, but not produce a value for a parent to consume, - // i.e. in the "localDefUse" case. - // There used to be an assert here that we wouldn't spill such a node. - // However, we can have unused lclVars that wind up being the node at which - // it is spilled. This probably indicates a bug, but we don't realy want to - // assert during a dump. - if (spillChar == 'S') - { - spillChar = '$'; - } - else + lastUseChar = '*'; + if (refPosition->delayRegFree) { - spillChar = '*'; + delayChar = 'D'; } - hasDest = true; - } - } - printf("%c N%03u. ", spillChar, tree->gtSeqNum); - - LclVarDsc* varDsc = nullptr; - unsigned varNum = UINT_MAX; - if (tree->IsLocal()) - { - varNum = tree->AsLclVarCommon()->GetLclNum(); - varDsc = &(compiler->lvaTable[varNum]); - if (varDsc->lvLRACandidate) - { - hasDest = false; } + printf(" %s%c%c ", getRefTypeShortName(refPosition->refType), lastUseChar, delayChar); } - if (hasDest) + else if (refPosition->IsPhysRegRef()) { - if (mode == LinearScan::LSRA_DUMP_POST && tree->gtFlags & GTF_SPILLED) - { - assert(tree->gtHasReg()); - } - lsraGetOperandString(tree, mode, operandString, operandStringLength); - printf("%-15s =", operandString); + RegRecord* regRecord = refPosition->getReg(); + printf(regNameFormat, getRegName(regRecord->regNum)); + printf(" %s ", getRefTypeShortName(refPosition->refType)); } else { - printf("%-15s ", emptyDestOperand); + assert(refPosition->refType == RefTypeKillGCRefs); + // There's no interval or reg name associated with this. + printf(regNameFormat, " "); + printf(" %s ", getRefTypeShortName(refPosition->refType)); } - if (varDsc != nullptr) +} + +//------------------------------------------------------------------------ +// LinearScan::IsResolutionMove: +// Returns true if the given node is a move inserted by LSRA +// resolution. +// +// Arguments: +// node - the node to check. +// +bool LinearScan::IsResolutionMove(GenTree* node) +{ + if (!IsLsraAdded(node)) { - if (varDsc->lvLRACandidate) - { - if (mode == LSRA_DUMP_REFPOS) - { - printf(" V%02u(L%d)", varNum, getIntervalForLocalVar(varDsc->lvVarIndex)->intervalIndex); - } - else - { - lsraGetOperandString(tree, mode, operandString, operandStringLength); - printf(" V%02u(%s)", varNum, operandString); - if (mode == LinearScan::LSRA_DUMP_POST && tree->gtFlags & GTF_SPILLED) - { - printf("R"); - } - } - } - else - { - printf(" V%02u MEM", varNum); - } + return false; } - else if (tree->OperIs(GT_ASG)) + + switch (node->OperGet()) { - assert(!tree->gtHasReg()); - printf(" asg%s ", GenTree::OpName(tree->OperGet())); + case GT_LCL_VAR: + case GT_COPY: + return node->IsUnusedValue(); + + case GT_SWAP: + return true; + + default: + return false; } - else +} + +//------------------------------------------------------------------------ +// LinearScan::IsResolutionNode: +// Returns true if the given node is either a move inserted by LSRA +// resolution or an operand to such a move. +// +// Arguments: +// containingRange - the range that contains the node to check. +// node - the node to check. +// +bool LinearScan::IsResolutionNode(LIR::Range& containingRange, GenTree* node) +{ + for (;;) { - compiler->gtDispNodeName(tree); - if (tree->OperKind() & GTK_LEAF) + if (IsResolutionMove(node)) { - compiler->gtDispLeaf(tree, nullptr); + return true; + } + + if (!IsLsraAdded(node) || (node->OperGet() != GT_LCL_VAR)) + { + return false; } + + LIR::Use use; + bool foundUse = containingRange.TryGetUse(node, &use); + assert(foundUse); + + node = use.User(); } } //------------------------------------------------------------------------ -// DumpOperandDefs: dumps the registers defined by a node. +// verifyFinalAllocation: Traverse the RefPositions and verify various invariants. // // Arguments: -// operand - The operand for which to compute a register count. +// None. // -// Returns: -// The number of registers defined by `operand`. +// Return Value: +// None. // -void LinearScan::DumpOperandDefs( - GenTree* operand, bool& first, LsraTupleDumpMode mode, char* operandString, const unsigned operandStringLength) +// Notes: +// If verbose is set, this will also dump a table of the final allocations. +void LinearScan::verifyFinalAllocation() { - assert(operand != nullptr); - assert(operandString != nullptr); - if (!operand->IsLIR()) + if (VERBOSE) { - return; + printf("\nFinal allocation\n"); } - int dstCount = ComputeOperandDstCount(operand); - - if (dstCount != 0) + // Clear register assignments. + for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) { - // This operand directly produces registers; print it. - if (!first) - { - printf(","); - } - lsraGetOperandString(operand, mode, operandString, operandStringLength); - printf("%s", operandString); - first = false; + RegRecord* physRegRecord = getRegisterRecord(reg); + physRegRecord->assignedInterval = nullptr; } - else if (operand->isContained()) + + for (Interval& interval : intervals) { - // This is a contained node. Dump the defs produced by its operands. - for (GenTree* op : operand->Operands()) - { - DumpOperandDefs(op, first, mode, operandString, operandStringLength); - } + interval.assignedReg = nullptr; + interval.physReg = REG_NA; } -} - -void LinearScan::TupleStyleDump(LsraTupleDumpMode mode) -{ - BasicBlock* block; - LsraLocation currentLoc = 1; // 0 is the entry - const unsigned operandStringLength = 6 * MAX_MULTIREG_COUNT + 1; - char operandString[operandStringLength]; - // currentRefPosition is not used for LSRA_DUMP_PRE - // We keep separate iterators for defs, so that we can print them - // on the lhs of the dump - RefPositionIterator refPosIterator = refPositions.begin(); - RefPosition* currentRefPosition = &refPosIterator; + DBEXEC(VERBOSE, dumpRegRecordTitle()); - switch (mode) + BasicBlock* currentBlock = nullptr; + GenTree* firstBlockEndResolutionNode = nullptr; + LsraLocation currentLocation = MinLocation; + for (RefPosition& refPosition : refPositions) { - case LSRA_DUMP_PRE: - printf("TUPLE STYLE DUMP BEFORE LSRA\n"); - break; - case LSRA_DUMP_REFPOS: - printf("TUPLE STYLE DUMP WITH REF POSITIONS\n"); - break; - case LSRA_DUMP_POST: - printf("TUPLE STYLE DUMP WITH REGISTER ASSIGNMENTS\n"); - break; - default: - printf("ERROR: INVALID TUPLE DUMP MODE\n"); - return; - } + RefPosition* currentRefPosition = &refPosition; + Interval* interval = nullptr; + RegRecord* regRecord = nullptr; + regNumber regNum = REG_NA; + activeRefPosition = currentRefPosition; - if (mode != LSRA_DUMP_PRE) - { - printf("Incoming Parameters: "); - for (; refPosIterator != refPositions.end() && currentRefPosition->refType != RefTypeBB; - ++refPosIterator, currentRefPosition = &refPosIterator) + if (currentRefPosition->refType != RefTypeBB) { - Interval* interval = currentRefPosition->getInterval(); - assert(interval != nullptr && interval->isLocalVar); - printf(" V%02d", interval->varNum); - if (mode == LSRA_DUMP_POST) + if (currentRefPosition->IsPhysRegRef()) { - regNumber reg; - if (currentRefPosition->registerAssignment == RBM_NONE) + regRecord = currentRefPosition->getReg(); + regRecord->recentRefPosition = currentRefPosition; + regNum = regRecord->regNum; + } + else if (currentRefPosition->isIntervalRef()) + { + interval = currentRefPosition->getInterval(); + interval->recentRefPosition = currentRefPosition; + if (currentRefPosition->registerAssignment != RBM_NONE) { - reg = REG_STK; + if (!genMaxOneBit(currentRefPosition->registerAssignment)) + { + assert(currentRefPosition->refType == RefTypeExpUse || + currentRefPosition->refType == RefTypeDummyDef); + } + else + { + regNum = currentRefPosition->assignedReg(); + regRecord = getRegisterRecord(regNum); + } + } + } + } + + LsraLocation newLocation = currentRefPosition->nodeLocation; + currentLocation = newLocation; + + switch (currentRefPosition->refType) + { + case RefTypeBB: + { + if (currentBlock == nullptr) + { + currentBlock = startBlockSequence(); } else { - reg = currentRefPosition->assignedReg(); + // Verify the resolution moves at the end of the previous block. + for (GenTree* node = firstBlockEndResolutionNode; node != nullptr; node = node->gtNext) + { + assert(enregisterLocalVars); + // Only verify nodes that are actually moves; don't bother with the nodes that are + // operands to moves. + if (IsResolutionMove(node)) + { + verifyResolutionMove(node, currentLocation); + } + } + + // Validate the locations at the end of the previous block. + if (enregisterLocalVars) + { + VarToRegMap outVarToRegMap = outVarToRegMaps[currentBlock->bbNum]; + VarSetOps::Iter iter(compiler, currentBlock->bbLiveOut); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex)) + { + if (localVarIntervals[varIndex] == nullptr) + { + assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); + continue; + } + regNumber regNum = getVarReg(outVarToRegMap, varIndex); + interval = getIntervalForLocalVar(varIndex); + if (interval->physReg != regNum) + { + assert(regNum == REG_STK); + assert((interval->physReg == REG_NA) || interval->isWriteThru); + } + interval->physReg = REG_NA; + interval->assignedReg = nullptr; + interval->isActive = false; + } + } + + // Clear register assignments. + for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) + { + RegRecord* physRegRecord = getRegisterRecord(reg); + physRegRecord->assignedInterval = nullptr; + } + + // Now, record the locations at the beginning of this block. + currentBlock = moveToNextBlock(); } - LclVarDsc* varDsc = &(compiler->lvaTable[interval->varNum]); - printf("("); - regNumber assignedReg = varDsc->GetRegNum(); - regNumber argReg = (varDsc->lvIsRegArg) ? varDsc->GetArgReg() : REG_STK; - assert(reg == assignedReg || varDsc->lvRegister == false); - if (reg != argReg) + if (currentBlock != nullptr) { - printf(getRegName(argReg, isFloatRegType(interval->registerType))); - printf("=>"); + if (enregisterLocalVars) + { + VarToRegMap inVarToRegMap = inVarToRegMaps[currentBlock->bbNum]; + VarSetOps::Iter iter(compiler, currentBlock->bbLiveIn); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex)) + { + if (localVarIntervals[varIndex] == nullptr) + { + assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); + continue; + } + regNumber regNum = getVarReg(inVarToRegMap, varIndex); + interval = getIntervalForLocalVar(varIndex); + interval->physReg = regNum; + interval->assignedReg = &(physRegs[regNum]); + interval->isActive = true; + physRegs[regNum].assignedInterval = interval; + } + } + + if (VERBOSE) + { + dumpRefPositionShort(currentRefPosition, currentBlock); + dumpRegRecords(); + } + + // Finally, handle the resolution moves, if any, at the beginning of the next block. + firstBlockEndResolutionNode = nullptr; + bool foundNonResolutionNode = false; + + LIR::Range& currentBlockRange = LIR::AsRange(currentBlock); + for (GenTree* node : currentBlockRange.NonPhiNodes()) + { + if (IsResolutionNode(currentBlockRange, node)) + { + assert(enregisterLocalVars); + if (foundNonResolutionNode) + { + firstBlockEndResolutionNode = node; + break; + } + else if (IsResolutionMove(node)) + { + // Only verify nodes that are actually moves; don't bother with the nodes that are + // operands to moves. + verifyResolutionMove(node, currentLocation); + } + } + else + { + foundNonResolutionNode = true; + } + } } - printf("%s)", getRegName(reg, isFloatRegType(interval->registerType))); } - } - printf("\n"); - } - for (block = startBlockSequence(); block != nullptr; block = moveToNextBlock()) - { - currentLoc += 2; + break; - if (mode == LSRA_DUMP_REFPOS) - { - bool printedBlockHeader = false; - // We should find the boundary RefPositions in the order of exposed uses, dummy defs, and the blocks - for (; refPosIterator != refPositions.end() && - (currentRefPosition->refType == RefTypeExpUse || currentRefPosition->refType == RefTypeDummyDef || - (currentRefPosition->refType == RefTypeBB && !printedBlockHeader)); - ++refPosIterator, currentRefPosition = &refPosIterator) - { - Interval* interval = nullptr; - if (currentRefPosition->isIntervalRef()) + case RefTypeKill: + assert(regRecord != nullptr); + assert(regRecord->assignedInterval == nullptr); + dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum, currentBlock); + break; + case RefTypeFixedReg: + assert(regRecord != nullptr); + dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum, currentBlock); + break; + + case RefTypeUpperVectorSave: + dumpLsraAllocationEvent(LSRA_EVENT_UPPER_VECTOR_SAVE, nullptr, REG_NA, currentBlock); + break; + + case RefTypeUpperVectorRestore: + dumpLsraAllocationEvent(LSRA_EVENT_UPPER_VECTOR_RESTORE, nullptr, REG_NA, currentBlock); + break; + + case RefTypeDef: + case RefTypeUse: + case RefTypeParamDef: + case RefTypeZeroInit: + assert(interval != nullptr); + + if (interval->isSpecialPutArg) { - interval = currentRefPosition->getInterval(); + dumpLsraAllocationEvent(LSRA_EVENT_SPECIAL_PUTARG, interval, regNum); + break; } - switch (currentRefPosition->refType) + if (currentRefPosition->reload) { - case RefTypeExpUse: - assert(interval != nullptr); - assert(interval->isLocalVar); - printf(" Exposed use of V%02u at #%d\n", interval->varNum, currentRefPosition->rpNum); - break; - case RefTypeDummyDef: - assert(interval != nullptr); - assert(interval->isLocalVar); - printf(" Dummy def of V%02u at #%d\n", interval->varNum, currentRefPosition->rpNum); - break; - case RefTypeBB: - block->dspBlockHeader(compiler); - printedBlockHeader = true; - printf("=====\n"); - break; - default: - printf("Unexpected RefPosition type at #%d\n", currentRefPosition->rpNum); - break; + interval->isActive = true; + assert(regNum != REG_NA); + interval->physReg = regNum; + interval->assignedReg = regRecord; + regRecord->assignedInterval = interval; + dumpLsraAllocationEvent(LSRA_EVENT_RELOAD, nullptr, regRecord->regNum, currentBlock); } - } - } - else - { - block->dspBlockHeader(compiler); - printf("=====\n"); - } - if (enregisterLocalVars && mode == LSRA_DUMP_POST && block != compiler->fgFirstBB && - block->bbNum <= bbNumMaxBeforeResolution) - { - printf("Predecessor for variable locations: " FMT_BB "\n", blockInfo[block->bbNum].predBBNum); - dumpInVarToRegMap(block); - } - if (block->bbNum > bbNumMaxBeforeResolution) - { - SplitEdgeInfo splitEdgeInfo; - splitBBNumToTargetBBNumMap->Lookup(block->bbNum, &splitEdgeInfo); - assert(splitEdgeInfo.toBBNum <= bbNumMaxBeforeResolution); - assert(splitEdgeInfo.fromBBNum <= bbNumMaxBeforeResolution); - printf("New block introduced for resolution from " FMT_BB " to " FMT_BB "\n", splitEdgeInfo.fromBBNum, - splitEdgeInfo.toBBNum); - } - - for (GenTree* node : LIR::AsRange(block).NonPhiNodes()) - { - GenTree* tree = node; - - int produce = tree->IsValue() ? ComputeOperandDstCount(tree) : 0; - int consume = ComputeAvailableSrcCount(tree); - - lsraDispNode(tree, mode, produce != 0 && mode != LSRA_DUMP_REFPOS); + if (regNum == REG_NA) + { + // If this interval is still assigned to a register + if (interval->physReg != REG_NA) + { + // then unassign it if no new register was assigned to the RefTypeDef + if (RefTypeIsDef(currentRefPosition->refType)) + { + assert(interval->assignedReg != nullptr); + if (interval->assignedReg->assignedInterval == interval) + { + interval->assignedReg->assignedInterval = nullptr; + } + interval->physReg = REG_NA; + interval->assignedReg = nullptr; + } + } - if (mode != LSRA_DUMP_REFPOS) - { - if (consume > 0) + dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, interval); + } + else if (RefTypeIsDef(currentRefPosition->refType)) { - printf("; "); + interval->isActive = true; - bool first = true; - for (GenTree* operand : tree->Operands()) + if (VERBOSE) { - DumpOperandDefs(operand, first, mode, operandString, operandStringLength); + if (interval->isConstant && (currentRefPosition->treeNode != nullptr) && + currentRefPosition->treeNode->IsReuseRegVal()) + { + dumpLsraAllocationEvent(LSRA_EVENT_REUSE_REG, nullptr, regRecord->regNum, currentBlock); + } + else + { + dumpLsraAllocationEvent(LSRA_EVENT_ALLOC_REG, nullptr, regRecord->regNum, currentBlock); + } } } - } - else - { - // Print each RefPosition on a new line, but - // printing all the kills for each node on a single line - // and combining the fixed regs with their associated def or use - bool killPrinted = false; - RefPosition* lastFixedRegRefPos = nullptr; - for (; refPosIterator != refPositions.end() && - (currentRefPosition->refType == RefTypeUse || currentRefPosition->refType == RefTypeFixedReg || - currentRefPosition->refType == RefTypeKill || currentRefPosition->refType == RefTypeDef) && - (currentRefPosition->nodeLocation == tree->gtSeqNum || - currentRefPosition->nodeLocation == tree->gtSeqNum + 1); - ++refPosIterator, currentRefPosition = &refPosIterator) + else if (currentRefPosition->copyReg) { - Interval* interval = nullptr; - if (currentRefPosition->isIntervalRef()) + dumpLsraAllocationEvent(LSRA_EVENT_COPY_REG, interval, regRecord->regNum, currentBlock); + } + else if (currentRefPosition->moveReg) + { + assert(interval->assignedReg != nullptr); + interval->assignedReg->assignedInterval = nullptr; + interval->physReg = regNum; + interval->assignedReg = regRecord; + regRecord->assignedInterval = interval; + if (VERBOSE) { - interval = currentRefPosition->getInterval(); + dumpEmptyRefPosition(); + printf("Move %-4s ", getRegName(regRecord->regNum)); } - switch (currentRefPosition->refType) + } + else + { + dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum, currentBlock); + } + if (currentRefPosition->lastUse || (currentRefPosition->spillAfter && !currentRefPosition->writeThru)) + { + interval->isActive = false; + } + if (regNum != REG_NA) + { + if (currentRefPosition->spillAfter) { - case RefTypeUse: - if (currentRefPosition->IsPhysRegRef()) - { - printf("\n Use:R%d(#%d)", - currentRefPosition->getReg()->regNum, currentRefPosition->rpNum); - } - else - { - assert(interval != nullptr); - printf("\n Use:"); - interval->microDump(); - printf("(#%d)", currentRefPosition->rpNum); - if (currentRefPosition->isFixedRegRef && !interval->isInternal) - { - assert(genMaxOneBit(currentRefPosition->registerAssignment)); - assert(lastFixedRegRefPos != nullptr); - printf(" Fixed:%s(#%d)", getRegName(currentRefPosition->assignedReg(), - isFloatRegType(interval->registerType)), - lastFixedRegRefPos->rpNum); - lastFixedRegRefPos = nullptr; - } - if (currentRefPosition->isLocalDefUse) - { - printf(" LocalDefUse"); - } - if (currentRefPosition->lastUse) - { - printf(" *"); - } - } - break; - case RefTypeDef: + if (VERBOSE) { - // Print each def on a new line - assert(interval != nullptr); - printf("\n Def:"); - interval->microDump(); - printf("(#%d)", currentRefPosition->rpNum); - if (currentRefPosition->isFixedRegRef) - { - assert(genMaxOneBit(currentRefPosition->registerAssignment)); - printf(" %s", getRegName(currentRefPosition->assignedReg(), - isFloatRegType(interval->registerType))); - } - if (currentRefPosition->isLocalDefUse) + // If refPos is marked as copyReg, then the reg that is spilled + // is the homeReg of the interval not the reg currently assigned + // to refPos. + regNumber spillReg = regNum; + if (currentRefPosition->copyReg) { - printf(" LocalDefUse"); + assert(interval != nullptr); + spillReg = interval->physReg; } - if (currentRefPosition->lastUse) + dumpRegRecords(); + dumpEmptyRefPosition(); + if (currentRefPosition->writeThru) { - printf(" *"); + printf("WThru %-4s ", getRegName(spillReg)); } - if (interval->relatedInterval != nullptr) + else { - printf(" Pref:"); - interval->relatedInterval->microDump(); + printf("Spill %-4s ", getRegName(spillReg)); } } - break; - case RefTypeKill: - if (!killPrinted) + } + else if (currentRefPosition->copyReg) + { + regRecord->assignedInterval = interval; + } + else + { + if (RefTypeIsDef(currentRefPosition->refType)) + { + // Interval was assigned to a different register. + // Clear the assigned interval of current register. + if (interval->physReg != REG_NA && interval->physReg != regNum) { - printf("\n Kill: "); - killPrinted = true; + interval->assignedReg->assignedInterval = nullptr; } - printf(getRegName(currentRefPosition->assignedReg(), - isFloatRegType(currentRefPosition->getReg()->registerType))); - printf(" "); - break; - case RefTypeFixedReg: - lastFixedRegRefPos = currentRefPosition; - break; - default: - printf("Unexpected RefPosition type at #%d\n", currentRefPosition->rpNum); - break; + } + interval->physReg = regNum; + interval->assignedReg = regRecord; + regRecord->assignedInterval = interval; } } - } - printf("\n"); - } - if (enregisterLocalVars && mode == LSRA_DUMP_POST) - { - dumpOutVarToRegMap(block); - } - printf("\n"); - } - printf("\n\n"); -} - -void LinearScan::dumpLsraAllocationEvent( - LsraDumpEvent event, Interval* interval, regNumber reg, BasicBlock* currentBlock, RegisterScore registerScore) -{ - if (!(VERBOSE)) - { - return; - } - if ((interval != nullptr) && (reg != REG_NA) && (reg != REG_STK)) - { - registersToDump |= getRegMask(reg, interval->registerType); - dumpRegRecordTitleIfNeeded(); - } - - switch (event) - { - // Conflicting def/use - case LSRA_EVENT_DEFUSE_CONFLICT: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("DUconflict "); - dumpRegRecords(); - break; - case LSRA_EVENT_DEFUSE_CASE1: - printf(indentFormat, " Case #1 use defRegAssignment"); - dumpRegRecords(); - break; - case LSRA_EVENT_DEFUSE_CASE2: - printf(indentFormat, " Case #2 use useRegAssignment"); - dumpRegRecords(); - break; - case LSRA_EVENT_DEFUSE_CASE3: - printf(indentFormat, " Case #3 use useRegAssignment"); - dumpRegRecords(); - dumpRegRecords(); - break; - case LSRA_EVENT_DEFUSE_CASE4: - printf(indentFormat, " Case #4 use defRegAssignment"); - dumpRegRecords(); - break; - case LSRA_EVENT_DEFUSE_CASE5: - printf(indentFormat, " Case #5 set def to all regs"); - dumpRegRecords(); - break; - case LSRA_EVENT_DEFUSE_CASE6: - printf(indentFormat, " Case #6 need a copy"); - dumpRegRecords(); - if (interval == nullptr) - { - printf(indentFormat, " NULL interval"); - dumpRegRecords(); - } - else if (interval->firstRefPosition->multiRegIdx != 0) - { - printf(indentFormat, " (multiReg)"); - dumpRegRecords(); - } - break; - - case LSRA_EVENT_SPILL: - dumpRefPositionShort(activeRefPosition, currentBlock); - assert(interval != nullptr && interval->assignedReg != nullptr); - printf("Spill %-4s ", getRegName(interval->assignedReg->regNum)); - dumpRegRecords(); - break; - - // Restoring the previous register - case LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL: - case LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL_AFTER_SPILL: - assert(interval != nullptr); - if ((activeRefPosition == nullptr) || (activeRefPosition->refType == RefTypeBB)) - { - printf(emptyRefPositionFormat, ""); - } - else - { - dumpRefPositionShort(activeRefPosition, currentBlock); - } - printf((event == LSRA_EVENT_RESTORE_PREVIOUS_INTERVAL) ? "Restr %-4s " : "SRstr %-4s ", - getRegName(reg)); - dumpRegRecords(); - break; + break; + case RefTypeKillGCRefs: + // No action to take. + // However, we will assert that, at resolution time, no registers contain GC refs. + { + DBEXEC(VERBOSE, printf(" ")); + regMaskTP candidateRegs = currentRefPosition->registerAssignment; + while (candidateRegs != RBM_NONE) + { + regMaskTP nextRegBit = genFindLowestBit(candidateRegs); + candidateRegs &= ~nextRegBit; + regNumber nextReg = genRegNumFromMask(nextRegBit); + RegRecord* regRecord = getRegisterRecord(nextReg); + Interval* assignedInterval = regRecord->assignedInterval; + assert(assignedInterval == nullptr || !varTypeIsGC(assignedInterval->registerType)); + } + } + break; - case LSRA_EVENT_DONE_KILL_GC_REFS: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("Done "); - break; + case RefTypeExpUse: + case RefTypeDummyDef: + // Do nothing; these will be handled by the RefTypeBB. + DBEXEC(VERBOSE, dumpRefPositionShort(currentRefPosition, currentBlock)); + DBEXEC(VERBOSE, printf(" ")); + break; - case LSRA_EVENT_NO_GC_KILLS: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("None "); - break; + case RefTypeInvalid: + // for these 'currentRefPosition->refType' values, No action to take + break; + } - // Block boundaries - case LSRA_EVENT_START_BB: - // The RefTypeBB comes after the RefTypeDummyDefs associated with that block, - // so we may have a RefTypeDummyDef at the time we dump this event. - // In that case we'll have another "EVENT" associated with it, so we need to - // print the full line now. - if (activeRefPosition->refType != RefTypeBB) - { - dumpNewBlock(currentBlock, activeRefPosition->nodeLocation); - dumpRegRecords(); - } - else + if (currentRefPosition->refType != RefTypeBB) + { + DBEXEC(VERBOSE, dumpRegRecords()); + if (interval != nullptr) { - dumpRefPositionShort(activeRefPosition, currentBlock); - } - break; - - // Allocation decisions - case LSRA_EVENT_NEEDS_NEW_REG: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("Free %-4s ", getRegName(reg)); - dumpRegRecords(); - break; + if (currentRefPosition->copyReg) + { + assert(interval->physReg != regNum); + regRecord->assignedInterval = nullptr; + assert(interval->assignedReg != nullptr); + regRecord = interval->assignedReg; + } + if (currentRefPosition->spillAfter || currentRefPosition->lastUse) + { + assert(!currentRefPosition->spillAfter || currentRefPosition->IsActualRef()); - case LSRA_EVENT_ZERO_REF: - assert(interval != nullptr && interval->isLocalVar); - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("NoRef "); - dumpRegRecords(); - break; + if (RefTypeIsDef(currentRefPosition->refType)) + { + // If an interval got assigned to a different register (while the different + // register got spilled), then clear the assigned interval of current register. + if (interval->physReg != REG_NA && interval->physReg != regNum) + { + interval->assignedReg->assignedInterval = nullptr; + } + } - case LSRA_EVENT_FIXED_REG: - case LSRA_EVENT_EXP_USE: - case LSRA_EVENT_KEPT_ALLOCATION: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("Keep %-4s ", getRegName(reg)); - break; + interval->physReg = REG_NA; + interval->assignedReg = nullptr; - case LSRA_EVENT_COPY_REG: - assert(interval != nullptr && interval->recentRefPosition != nullptr); - dumpRefPositionShort(activeRefPosition, currentBlock); - if (allocationPassComplete || (registerScore == 0)) - { - printf("Copy %-4s ", getRegName(reg)); - } - else - { - printf("%-5s(C) %-4s ", getScoreName(registerScore), getRegName(reg)); + // regRegcord could be null if the RefPosition does not require a register. + if (regRecord != nullptr) + { + regRecord->assignedInterval = nullptr; + } +#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE + else if (interval->isUpperVector && !currentRefPosition->RegOptional()) + { + // These only require a register if they are not RegOptional, and their lclVar + // interval is living in a register and not already partially spilled. + if ((currentRefPosition->refType == RefTypeUpperVectorSave) || + (currentRefPosition->refType == RefTypeUpperVectorRestore)) + { + Interval* lclVarInterval = interval->relatedInterval; + assert((lclVarInterval->physReg == REG_NA) || lclVarInterval->isPartiallySpilled); + } + } +#endif + else + { + assert(currentRefPosition->RegOptional()); + } + } } - break; + } + } - case LSRA_EVENT_MOVE_REG: - assert(interval != nullptr && interval->recentRefPosition != nullptr); - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("Move %-4s ", getRegName(reg)); - dumpRegRecords(); - break; + // Now, verify the resolution blocks. + // Currently these are nearly always at the end of the method, but that may not always be the case. + // So, we'll go through all the BBs looking for blocks whose bbNum is greater than bbNumMaxBeforeResolution. + for (BasicBlock* currentBlock = compiler->fgFirstBB; currentBlock != nullptr; currentBlock = currentBlock->bbNext) + { + if (currentBlock->bbNum > bbNumMaxBeforeResolution) + { + // If we haven't enregistered an lclVars, we have no resolution blocks. + assert(enregisterLocalVars); - case LSRA_EVENT_ALLOC_REG: - dumpRefPositionShort(activeRefPosition, currentBlock); - if (allocationPassComplete || (registerScore == 0)) + if (VERBOSE) { - printf("Alloc %-4s ", getRegName(reg)); + dumpRegRecordTitle(); + printf(shortRefPositionFormat, 0, 0); + assert(currentBlock->bbPreds != nullptr && currentBlock->bbPreds->getBlock() != nullptr); + printf(bbRefPosFormat, currentBlock->bbNum, currentBlock->bbPreds->getBlock()->bbNum); + dumpRegRecords(); } - else + + // Clear register assignments. + for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) { - printf("%-5s(A) %-4s ", getScoreName(registerScore), getRegName(reg)); + RegRecord* physRegRecord = getRegisterRecord(reg); + physRegRecord->assignedInterval = nullptr; } - break; - - case LSRA_EVENT_REUSE_REG: - dumpRefPositionShort(activeRefPosition, currentBlock); - if (allocationPassComplete || (registerScore == 0)) + // Set the incoming register assignments + VarToRegMap inVarToRegMap = getInVarToRegMap(currentBlock->bbNum); + VarSetOps::Iter iter(compiler, currentBlock->bbLiveIn); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex)) { - printf("Reuse %-4s ", getRegName(reg)); + if (localVarIntervals[varIndex] == nullptr) + { + assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); + continue; + } + regNumber regNum = getVarReg(inVarToRegMap, varIndex); + Interval* interval = getIntervalForLocalVar(varIndex); + interval->physReg = regNum; + interval->assignedReg = &(physRegs[regNum]); + interval->isActive = true; + physRegs[regNum].assignedInterval = interval; } - else + + // Verify the moves in this block + LIR::Range& currentBlockRange = LIR::AsRange(currentBlock); + for (GenTree* node : currentBlockRange.NonPhiNodes()) { - printf("%-5s(A) %-4s ", getScoreName(registerScore), getRegName(reg)); + assert(IsResolutionNode(currentBlockRange, node)); + if (IsResolutionMove(node)) + { + // Only verify nodes that are actually moves; don't bother with the nodes that are + // operands to moves. + verifyResolutionMove(node, currentLocation); + } } - break; - - case LSRA_EVENT_NO_ENTRY_REG_ALLOCATED: - assert(interval != nullptr && interval->isLocalVar); - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("LoRef "); - break; - - case LSRA_EVENT_NO_REG_ALLOCATED: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("NoReg "); - break; - - case LSRA_EVENT_RELOAD: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("ReLod %-4s ", getRegName(reg)); - dumpRegRecords(); - break; - - case LSRA_EVENT_SPECIAL_PUTARG: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("PtArg %-4s ", getRegName(reg)); - break; - - case LSRA_EVENT_UPPER_VECTOR_SAVE: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("UVSav %-4s ", getRegName(reg)); - break; - - case LSRA_EVENT_UPPER_VECTOR_RESTORE: - dumpRefPositionShort(activeRefPosition, currentBlock); - printf("UVRes %-4s ", getRegName(reg)); - break; - - // We currently don't dump anything for these events. - case LSRA_EVENT_DEFUSE_FIXED_DELAY_USE: - case LSRA_EVENT_SPILL_EXTENDED_LIFETIME: - case LSRA_EVENT_END_BB: - case LSRA_EVENT_FREE_REGS: - case LSRA_EVENT_INCREMENT_RANGE_END: - case LSRA_EVENT_LAST_USE: - case LSRA_EVENT_LAST_USE_DELAYED: - break; - - default: - printf("????? %-4s ", getRegName(reg)); - dumpRegRecords(); - break; - } -} - -//------------------------------------------------------------------------ -// dumpRegRecordHeader: Dump the header for a column-based dump of the register state. -// -// Arguments: -// None. -// -// Return Value: -// None. -// -// Assumptions: -// Reg names fit in 4 characters (minimum width of the columns) -// -// Notes: -// In order to make the table as dense as possible (for ease of reading the dumps), -// we determine the minimum regColumnWidth width required to represent: -// regs, by name (e.g. eax or xmm0) - this is fixed at 4 characters. -// intervals, as Vnn for lclVar intervals, or as I for other intervals. -// The table is indented by the amount needed for dumpRefPositionShort, which is -// captured in shortRefPositionDumpWidth. -// -void LinearScan::dumpRegRecordHeader() -{ - printf("The following table has one or more rows for each RefPosition that is handled during allocation.\n" - "The first column provides the basic information about the RefPosition, with its type (e.g. Def,\n" - "Use, Fixd) followed by a '*' if it is a last use, and a 'D' if it is delayRegFree, and then the\n" - "action taken during allocation (e.g. Alloc a new register, or Keep an existing one).\n" - "The subsequent columns show the Interval occupying each register, if any, followed by 'a' if it is\n" - "active, a 'p' if it is a large vector that has been partially spilled, and 'i'if it is inactive.\n" - "Columns are only printed up to the last modifed register, which may increase during allocation,\n" - "in which case additional columns will appear. \n" - "Registers which are not marked modified have ---- in their column.\n\n"); - - // First, determine the width of each register column (which holds a reg name in the - // header, and an interval name in each subsequent row). - int intervalNumberWidth = (int)log10((double)intervals.size()) + 1; - // The regColumnWidth includes the identifying character (I or V) and an 'i', 'p' or 'a' (inactive, - // partially-spilled or active) - regColumnWidth = intervalNumberWidth + 2; - if (regColumnWidth < 4) - { - regColumnWidth = 4; - } - sprintf_s(intervalNameFormat, MAX_FORMAT_CHARS, "%%c%%-%dd", regColumnWidth - 2); - sprintf_s(regNameFormat, MAX_FORMAT_CHARS, "%%-%ds", regColumnWidth); - - // Next, determine the width of the short RefPosition (see dumpRefPositionShort()). - // This is in the form: - // nnn.#mmm NAME TYPEld - // Where: - // nnn is the Location, right-justified to the width needed for the highest location. - // mmm is the RefPosition rpNum, left-justified to the width needed for the highest rpNum. - // NAME is dumped by dumpReferentName(), and is "regColumnWidth". - // TYPE is RefTypeNameShort, and is 4 characters - // l is either '*' (if a last use) or ' ' (otherwise) - // d is either 'D' (if a delayed use) or ' ' (otherwise) - - maxNodeLocation = (maxNodeLocation == 0) - ? 1 - : maxNodeLocation; // corner case of a method with an infinite loop without any gentree nodes - assert(maxNodeLocation >= 1); - assert(refPositions.size() >= 1); - int nodeLocationWidth = (int)log10((double)maxNodeLocation) + 1; - int refPositionWidth = (int)log10((double)refPositions.size()) + 1; - int refTypeInfoWidth = 4 /*TYPE*/ + 2 /* last-use and delayed */ + 1 /* space */; - int locationAndRPNumWidth = nodeLocationWidth + 2 /* .# */ + refPositionWidth + 1 /* space */; - int shortRefPositionDumpWidth = locationAndRPNumWidth + regColumnWidth + 1 /* space */ + refTypeInfoWidth; - sprintf_s(shortRefPositionFormat, MAX_FORMAT_CHARS, "%%%dd.#%%-%dd ", nodeLocationWidth, refPositionWidth); - sprintf_s(emptyRefPositionFormat, MAX_FORMAT_CHARS, "%%-%ds", shortRefPositionDumpWidth); - - // The width of the "allocation info" - // - a 8-character allocation decision - // - a space - // - a 4-character register - // - a space - int allocationInfoWidth = 8 + 1 + 4 + 1; - - // Next, determine the width of the legend for each row. This includes: - // - a short RefPosition dump (shortRefPositionDumpWidth), which includes a space - // - the allocation info (allocationInfoWidth), which also includes a space - - regTableIndent = shortRefPositionDumpWidth + allocationInfoWidth; - // BBnn printed left-justified in the NAME Typeld and allocationInfo space. - int bbNumWidth = (int)log10((double)compiler->fgBBNumMax) + 1; - // In the unlikely event that BB numbers overflow the space, we'll simply omit the predBB - int predBBNumDumpSpace = regTableIndent - locationAndRPNumWidth - bbNumWidth - 9; // 'BB' + ' PredBB' - if (predBBNumDumpSpace < bbNumWidth) - { - sprintf_s(bbRefPosFormat, MAX_LEGEND_FORMAT_CHARS, "BB%%-%dd", shortRefPositionDumpWidth - 2); - } - else - { - sprintf_s(bbRefPosFormat, MAX_LEGEND_FORMAT_CHARS, "BB%%-%dd PredBB%%-%dd", bbNumWidth, predBBNumDumpSpace); - } - - if (compiler->shouldDumpASCIITrees()) - { - columnSeparator = "|"; - line = "-"; - leftBox = "+"; - middleBox = "+"; - rightBox = "+"; - } - else - { - columnSeparator = "\xe2\x94\x82"; - line = "\xe2\x94\x80"; - leftBox = "\xe2\x94\x9c"; - middleBox = "\xe2\x94\xbc"; - rightBox = "\xe2\x94\xa4"; + // Verify the outgoing register assignments + { + VarToRegMap outVarToRegMap = getOutVarToRegMap(currentBlock->bbNum); + VarSetOps::Iter iter(compiler, currentBlock->bbLiveOut); + unsigned varIndex = 0; + while (iter.NextElem(&varIndex)) + { + if (localVarIntervals[varIndex] == nullptr) + { + assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); + continue; + } + regNumber regNum = getVarReg(outVarToRegMap, varIndex); + Interval* interval = getIntervalForLocalVar(varIndex); + // Either the register assignments match, or the outgoing assignment is on the stack + // and this is a write-thru interval. + assert(interval->physReg == regNum || (interval->physReg == REG_NA && regNum == REG_STK) || + (interval->isWriteThru && regNum == REG_STK)); + interval->physReg = REG_NA; + interval->assignedReg = nullptr; + interval->isActive = false; + } + } + } } - sprintf_s(indentFormat, MAX_FORMAT_CHARS, "%%-%ds", regTableIndent); - - // Now, set up the legend format for the RefPosition info - sprintf_s(legendFormat, MAX_LEGEND_FORMAT_CHARS, "%%-%d.%ds%%-%d.%ds%%-%ds%%s", nodeLocationWidth + 1, - nodeLocationWidth + 1, refPositionWidth + 2, refPositionWidth + 2, regColumnWidth + 1); - // Print a "title row" including the legend and the reg names. - lastDumpedRegisters = RBM_NONE; - dumpRegRecordTitleIfNeeded(); + DBEXEC(VERBOSE, printf("\n")); } -void LinearScan::dumpRegRecordTitleIfNeeded() +//------------------------------------------------------------------------ +// verifyResolutionMove: Verify a resolution statement. Called by verifyFinalAllocation() +// +// Arguments: +// resolutionMove - A GenTree* that must be a resolution move. +// currentLocation - The LsraLocation of the most recent RefPosition that has been verified. +// +// Return Value: +// None. +// +// Notes: +// If verbose is set, this will also dump the moves into the table of final allocations. +void LinearScan::verifyResolutionMove(GenTree* resolutionMove, LsraLocation currentLocation) { - if ((lastDumpedRegisters != registersToDump) || (rowCountSinceLastTitle > MAX_ROWS_BETWEEN_TITLES)) + GenTree* dst = resolutionMove; + assert(IsResolutionMove(dst)); + + if (dst->OperGet() == GT_SWAP) { - lastUsedRegNumIndex = 0; - int lastRegNumIndex = compiler->compFloatingPointUsed ? REG_FP_LAST : REG_INT_LAST; - for (int regNumIndex = 0; regNumIndex <= lastRegNumIndex; regNumIndex++) + GenTreeLclVarCommon* left = dst->gtGetOp1()->AsLclVarCommon(); + GenTreeLclVarCommon* right = dst->gtGetOp2()->AsLclVarCommon(); + regNumber leftRegNum = left->GetRegNum(); + regNumber rightRegNum = right->GetRegNum(); + LclVarDsc* leftVarDsc = compiler->lvaTable + left->GetLclNum(); + LclVarDsc* rightVarDsc = compiler->lvaTable + right->GetLclNum(); + Interval* leftInterval = getIntervalForLocalVar(leftVarDsc->lvVarIndex); + Interval* rightInterval = getIntervalForLocalVar(rightVarDsc->lvVarIndex); + assert(leftInterval->physReg == leftRegNum && rightInterval->physReg == rightRegNum); + leftInterval->physReg = rightRegNum; + rightInterval->physReg = leftRegNum; + leftInterval->assignedReg = &physRegs[rightRegNum]; + rightInterval->assignedReg = &physRegs[leftRegNum]; + physRegs[rightRegNum].assignedInterval = leftInterval; + physRegs[leftRegNum].assignedInterval = rightInterval; + if (VERBOSE) { - if ((registersToDump & genRegMask((regNumber)regNumIndex)) != 0) - { - lastUsedRegNumIndex = regNumIndex; - } + printf(shortRefPositionFormat, currentLocation, 0); + dumpIntervalName(leftInterval); + printf(" Swap "); + printf(" %-4s ", getRegName(rightRegNum)); + dumpRegRecords(); + printf(shortRefPositionFormat, currentLocation, 0); + dumpIntervalName(rightInterval); + printf(" \" "); + printf(" %-4s ", getRegName(leftRegNum)); + dumpRegRecords(); } - dumpRegRecordTitle(); - lastDumpedRegisters = registersToDump; + return; } -} - -void LinearScan::dumpRegRecordTitleLines() -{ - for (int i = 0; i < regTableIndent; i++) + regNumber dstRegNum = dst->GetRegNum(); + regNumber srcRegNum; + GenTreeLclVarCommon* lcl; + if (dst->OperGet() == GT_COPY) { - printf("%s", line); + lcl = dst->gtGetOp1()->AsLclVarCommon(); + srcRegNum = lcl->GetRegNum(); } - for (int regNumIndex = 0; regNumIndex <= lastUsedRegNumIndex; regNumIndex++) + else { - regNumber regNum = (regNumber)regNumIndex; - if (shouldDumpReg(regNum)) + lcl = dst->AsLclVarCommon(); + if ((lcl->gtFlags & GTF_SPILLED) != 0) { - printf("%s", middleBox); - for (int i = 0; i < regColumnWidth; i++) - { - printf("%s", line); - } + srcRegNum = REG_STK; } - } - printf("%s\n", rightBox); -} -void LinearScan::dumpRegRecordTitle() -{ - dumpRegRecordTitleLines(); - - // Print out the legend for the RefPosition info - printf(legendFormat, "Loc ", "RP# ", "Name ", "Type Action Reg "); - - // Print out the register name column headers - char columnFormatArray[MAX_FORMAT_CHARS]; - sprintf_s(columnFormatArray, MAX_FORMAT_CHARS, "%s%%-%d.%ds", columnSeparator, regColumnWidth, regColumnWidth); - for (int regNumIndex = 0; regNumIndex <= lastUsedRegNumIndex; regNumIndex++) - { - regNumber regNum = (regNumber)regNumIndex; - if (shouldDumpReg(regNum)) + else { - const char* regName = getRegName(regNum); - printf(columnFormatArray, regName); + assert((lcl->gtFlags & GTF_SPILL) != 0); + srcRegNum = dstRegNum; + dstRegNum = REG_STK; } } - printf("%s\n", columnSeparator); - - rowCountSinceLastTitle = 0; - - dumpRegRecordTitleLines(); -} - -void LinearScan::dumpRegRecords() -{ - static char columnFormatArray[18]; - for (regNumber regNum = REG_FIRST; regNum <= (regNumber)lastUsedRegNumIndex; regNum = REG_NEXT(regNum)) + Interval* interval = getIntervalForLocalVarNode(lcl); + assert(interval->physReg == srcRegNum || (srcRegNum == REG_STK && interval->physReg == REG_NA)); + if (srcRegNum != REG_STK) { - if (shouldDumpReg(regNum)) - { - printf("%s", columnSeparator); - RegRecord& regRecord = physRegs[regNum]; - Interval* interval = regRecord.assignedInterval; - if (interval != nullptr) - { - dumpIntervalName(interval); - char activeChar = interval->isActive ? 'a' : 'i'; -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - if (interval->isPartiallySpilled) - { - activeChar = 'p'; - } -#endif // FEATURE_PARTIAL_SIMD_CALLEE_SAVE - printf("%c", activeChar); - } - else if ((genRegMask(regNum) & regsBusyUntilKill) != RBM_NONE) - { - printf(columnFormatArray, "Busy"); - } - else - { - sprintf_s(columnFormatArray, MAX_FORMAT_CHARS, "%%-%ds", regColumnWidth); - printf(columnFormatArray, ""); - } - } + physRegs[srcRegNum].assignedInterval = nullptr; } - printf("%s\n", columnSeparator); - rowCountSinceLastTitle++; -} - -void LinearScan::dumpIntervalName(Interval* interval) -{ - if (interval->isLocalVar) + if (dstRegNum != REG_STK) { - printf(intervalNameFormat, 'V', interval->varNum); + interval->physReg = dstRegNum; + interval->assignedReg = &(physRegs[dstRegNum]); + physRegs[dstRegNum].assignedInterval = interval; + interval->isActive = true; } - else if (interval->IsUpperVector()) + else { - printf(intervalNameFormat, 'U', interval->relatedInterval->varNum); + interval->physReg = REG_NA; + interval->assignedReg = nullptr; + interval->isActive = false; } - else if (interval->isConstant) + if (VERBOSE) { - printf(intervalNameFormat, 'C', interval->intervalIndex); + printf(shortRefPositionFormat, currentLocation, 0); + dumpIntervalName(interval); + printf(" Move "); + printf(" %-4s ", getRegName(dstRegNum)); + dumpRegRecords(); + } +} +#endif // DEBUG + + +LinearScan::RegisterSelection::RegisterSelection(LinearScan* linearScan) +{ + this->linearScan = linearScan; + +#ifdef DEBUG + mappingTable = new ScoreMappingTable(linearScan->compiler->getAllocator(CMK_LSRA)); + +#define REG_SEL_DEF(stat, value, shortname) mappingTable->Set(stat, &LinearScan::RegisterSelection::try_##stat); +#include "lsra_score.h" +#undef REG_SEL_DEF + + LPCWSTR ordering = JitConfig.JitLsraOrdering(); + + if (ordering == nullptr) + { + memcpy(RegSelectionOrder, DefaultOrder, sizeof(RegisterScore) * REGSELECT_HEURISTIC_COUNT); } else { - printf(intervalNameFormat, 'I', interval->intervalIndex); + for (int orderId = 0; orderId < REGSELECT_HEURISTIC_COUNT; orderId++) + { + // Make sure we do not set repeated entries + assert(RegSelectionOrder[orderId] == NONE); + + RegSelectionOrder[orderId] = DefaultOrder[ordering[orderId] - 'A']; + } } +#endif // DEBUG } -void LinearScan::dumpEmptyRefPosition() +void LinearScan::RegisterSelection::reset(Interval* interval, RefPosition* refPos) { - printf(emptyRefPositionFormat, ""); + currentInterval = interval; + refPosition = refPos; + score = 0; // TODO:Kpathak - why is this used? + + regType = linearScan->getRegisterType(currentInterval, refPosition); + currentLocation = refPosition->nodeLocation; + nextRefPos = refPosition->nextRefPosition; + candidates = refPosition->registerAssignment; + preferences = currentInterval->registerPreferences; + + // This is not actually a preference, it's merely to track the lclVar that this + // "specialPutArg" is using. + relatedInterval = currentInterval->isSpecialPutArg ? nullptr : currentInterval->relatedInterval; + relatedPreferences = (relatedInterval == nullptr) ? RBM_NONE : relatedInterval->getCurrentPreferences(); + + rangeEndLocation = refPosition->getRangeEndLocation(); + relatedLastLocation = rangeEndLocation; // TODO:kpathak - need to see why this is not used after refactor? + preferCalleeSave = currentInterval->preferCalleeSave; + rangeEndRefPosition = nullptr; + lastRefPosition = currentInterval->lastRefPosition; + lastLocation = MinLocation; + prevRegRec = currentInterval->assignedReg; + + // These are used in the post-selection updates, and must be set for any selection. + freeCandidates = RBM_NONE; + matchingConstants = RBM_NONE; + unassignedSet = RBM_NONE; + + coversSet = RBM_NONE; + preferenceSet = RBM_NONE; + coversRelatedSet = RBM_NONE; + coversFullSet = RBM_NONE; + + foundRegBit = REG_NA; + found = false; + skipAllocation = false; + coversSetsCalculated = false; } -//------------------------------------------------------------------------ -// dumpNewBlock: Dump a line for a new block in a column-based dump of the register state. -// -// Arguments: -// currentBlock - the new block to be dumped -// -void LinearScan::dumpNewBlock(BasicBlock* currentBlock, LsraLocation location) +bool LinearScan::RegisterSelection::applySelection(int selectionScore, regMaskTP selectionCandidates) { - if (!VERBOSE) + regMaskTP newCandidates = candidates & selectionCandidates; + if (newCandidates != RBM_NONE) { - return; + score += selectionScore; + candidates = newCandidates; + return LinearScan::isSingleRegister(candidates); } + return false; +} - // Always print a title row before a RefTypeBB (except for the first, because we - // will already have printed it before the parameters) - if ((currentBlock != compiler->fgFirstBB) && (currentBlock != nullptr)) - { - dumpRegRecordTitle(); - } - // If the activeRefPosition is a DummyDef, then don't print anything further (printing the - // title line makes it clearer that we're "about to" start the next block). - if (activeRefPosition->refType == RefTypeDummyDef) - { - dumpEmptyRefPosition(); - printf("DDefs "); - printf(regNameFormat, ""); - return; - } - printf(shortRefPositionFormat, location, activeRefPosition->rpNum); - if (currentBlock == nullptr) - { - printf(regNameFormat, "END"); - printf(" "); - printf(regNameFormat, ""); - } - else +// Select a single register, if it is in the candidate set. +// Return true if so. +bool LinearScan::RegisterSelection::applySingleRegSelection(int selectionScore, regMaskTP selectionCandidate) +{ + assert(LinearScan::isSingleRegister(selectionCandidate)); + regMaskTP newCandidates = candidates & selectionCandidate; + if (newCandidates != RBM_NONE) { - printf(bbRefPosFormat, currentBlock->bbNum, - currentBlock == compiler->fgFirstBB ? 0 : blockInfo[currentBlock->bbNum].predBBNum); + candidates = newCandidates; + return true; } -} - -// Note that the size of this dump is computed in dumpRegRecordHeader(). -// -void LinearScan::dumpRefPositionShort(RefPosition* refPosition, BasicBlock* currentBlock) + return false; +} + +void LinearScan::RegisterSelection::try_FREE() { - static RefPosition* lastPrintedRefPosition = nullptr; - if (refPosition == lastPrintedRefPosition) + assert(!found); + if (freeCandidates == RBM_NONE) { - dumpEmptyRefPosition(); return; } - lastPrintedRefPosition = refPosition; - if (refPosition->refType == RefTypeBB) + + found = applySelection(FREE, freeCandidates); +} + +// Apply the CONST_AVAILABLE (matching constant) heuristic. +// Note that we always need to define the 'matchingConstants' set. +void LinearScan::RegisterSelection::try_CONST_AVAILABLE() +{ + assert(!found); + if (freeCandidates == RBM_NONE) { - dumpNewBlock(currentBlock, refPosition->nodeLocation); return; } - printf(shortRefPositionFormat, refPosition->nodeLocation, refPosition->rpNum); - if (refPosition->isIntervalRef()) + + if (currentInterval->isConstant && RefTypeIsDef(refPosition->refType)) { - Interval* interval = refPosition->getInterval(); - dumpIntervalName(interval); - char lastUseChar = ' '; - char delayChar = ' '; - if (refPosition->lastUse) - { - lastUseChar = '*'; - if (refPosition->delayRegFree) - { - delayChar = 'D'; - } - } - printf(" %s%c%c ", getRefTypeShortName(refPosition->refType), lastUseChar, delayChar); + found = applySelection(CONST_AVAILABLE, matchingConstants); } - else if (refPosition->IsPhysRegRef()) +} + +// Apply the THIS_ASSIGNED heuristic. +void LinearScan::RegisterSelection::try_THIS_ASSIGNED() +{ + assert(!found); + if (freeCandidates == RBM_NONE) { - RegRecord* regRecord = refPosition->getReg(); - printf(regNameFormat, getRegName(regRecord->regNum)); - printf(" %s ", getRefTypeShortName(refPosition->refType)); + return; } - else + + if (prevRegRec != nullptr) { - assert(refPosition->refType == RefTypeKillGCRefs); - // There's no interval or reg name associated with this. - printf(regNameFormat, " "); - printf(" %s ", getRefTypeShortName(refPosition->refType)); + found = applySelection(THIS_ASSIGNED, freeCandidates & preferences & prevRegBit); } } -//------------------------------------------------------------------------ -// LinearScan::IsResolutionMove: -// Returns true if the given node is a move inserted by LSRA -// resolution. -// -// Arguments: -// node - the node to check. -// -bool LinearScan::IsResolutionMove(GenTree* node) +// Apply the COVERS heuristic. +void LinearScan::RegisterSelection::try_COVERS() { - if (!IsLsraAdded(node)) - { - return false; - } + assert(!found); - switch (node->OperGet()) - { - case GT_LCL_VAR: - case GT_COPY: - return node->IsUnusedValue(); + calculateSets(); - case GT_SWAP: - return true; + found = applySelection(COVERS, coversSet & preferenceSet); +} - default: - return false; - } +// Apply the OWN_PREFERENCE heuristic. +// Note that 'preferenceSet' already includes only freeCandidates. +void LinearScan::RegisterSelection::try_OWN_PREFERENCE() +{ + assert(!found); + +#ifdef DEBUG + calculateSets(); +#endif + + found = applySelection(OWN_PREFERENCE, (preferenceSet & freeCandidates)); } -//------------------------------------------------------------------------ -// LinearScan::IsResolutionNode: -// Returns true if the given node is either a move inserted by LSRA -// resolution or an operand to such a move. -// -// Arguments: -// containingRange - the range that contains the node to check. -// node - the node to check. -// -bool LinearScan::IsResolutionNode(LIR::Range& containingRange, GenTree* node) +// Apply the COVERS_RELATED heuristic. +void LinearScan::RegisterSelection::try_COVERS_RELATED() { - for (;;) - { - if (IsResolutionMove(node)) - { - return true; - } + assert(!found); - if (!IsLsraAdded(node) || (node->OperGet() != GT_LCL_VAR)) - { - return false; - } +#ifdef DEBUG + calculateSets(); +#endif - LIR::Use use; - bool foundUse = containingRange.TryGetUse(node, &use); - assert(foundUse); + found = applySelection(COVERS_RELATED, (coversRelatedSet & freeCandidates)); +} - node = use.User(); - } +// Apply the RELATED_PREFERENCE heuristic. +void LinearScan::RegisterSelection::try_RELATED_PREFERENCE() +{ + assert(!found); + + found = applySelection(RELATED_PREFERENCE, relatedPreferences & freeCandidates); } -//------------------------------------------------------------------------ -// verifyFinalAllocation: Traverse the RefPositions and verify various invariants. -// -// Arguments: -// None. -// -// Return Value: -// None. -// -// Notes: -// If verbose is set, this will also dump a table of the final allocations. -void LinearScan::verifyFinalAllocation() +// Apply the CALLER_CALLEE heuristic. +void LinearScan::RegisterSelection::try_CALLER_CALLEE() { - if (VERBOSE) - { - printf("\nFinal allocation\n"); - } + assert(!found); - // Clear register assignments. - for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) - { - RegRecord* physRegRecord = getRegisterRecord(reg); - physRegRecord->assignedInterval = nullptr; - } + found = applySelection(CALLER_CALLEE, callerCalleePrefs & freeCandidates); +} - for (Interval& interval : intervals) +// Apply the UNASSIGNED heuristic. +void LinearScan::RegisterSelection::try_UNASSIGNED() +{ + assert(!found); + +#ifdef DEBUG + calculateSets(); +#endif + + found = applySelection(UNASSIGNED, unassignedSet); +} + +// Apply the COVERS_FULL heuristic. +void LinearScan::RegisterSelection::try_COVERS_FULL() +{ + assert(!found); + +#ifdef DEBUG + calculateSets(); +#endif + + found = applySelection(COVERS_FULL, (coversFullSet & freeCandidates)); +} + +void LinearScan::RegisterSelection::try_BEST_FIT() +{ + assert(!found); + + if (freeCandidates == RBM_NONE) { - interval.assignedReg = nullptr; - interval.physReg = REG_NA; + return; } - DBEXEC(VERBOSE, dumpRegRecordTitle()); - - BasicBlock* currentBlock = nullptr; - GenTree* firstBlockEndResolutionNode = nullptr; - LsraLocation currentLocation = MinLocation; - for (RefPosition& refPosition : refPositions) + regMaskTP bestFitSet = RBM_NONE; + // If the best score includes COVERS_FULL, pick the one that's killed soonest. + // If none cover the full range, the BEST_FIT is the one that's killed later. + bool earliestIsBest = ((score & COVERS_FULL) != 0); + LsraLocation bestFitLocation = earliestIsBest ? MaxLocation : MinLocation; + for (regMaskTP bestFitCandidates = candidates; bestFitCandidates != RBM_NONE;) { - RefPosition* currentRefPosition = &refPosition; - Interval* interval = nullptr; - RegRecord* regRecord = nullptr; - regNumber regNum = REG_NA; - activeRefPosition = currentRefPosition; + regMaskTP bestFitCandidateBit = genFindLowestBit(bestFitCandidates); + bestFitCandidates &= ~bestFitCandidateBit; + regNumber bestFitCandidateRegNum = genRegNumFromMask(bestFitCandidateBit); - if (currentRefPosition->refType != RefTypeBB) + // Find the next RefPosition of the register. + LsraLocation nextIntervalLocation = linearScan->getNextIntervalRef(bestFitCandidateRegNum, regType); + LsraLocation nextPhysRefLocation = linearScan->getNextFixedRef(bestFitCandidateRegNum, regType); + nextPhysRefLocation = Min(nextPhysRefLocation, nextIntervalLocation); + // If the nextPhysRefLocation is a fixedRef for the rangeEndRefPosition, increment it so that + // we don't think it isn't covering the live range. + // This doesn't handle the case where earlier RefPositions for this Interval are also + // FixedRefs of this regNum, but at least those are only interesting in the case where those + // are "local last uses" of the Interval - otherwise the liveRange would interfere with the reg. + // TODO: This duplicates code in an earlier loop, and is basically here to duplicate previous + // behavior; see if we can avoid this. + if (nextPhysRefLocation == rangeEndLocation && rangeEndRefPosition->isFixedRefOfReg(bestFitCandidateRegNum)) { - if (currentRefPosition->IsPhysRegRef()) + INDEBUG(linearScan->dumpLsraAllocationEvent(LSRA_EVENT_INCREMENT_RANGE_END, currentInterval)); + nextPhysRefLocation++; + } + + if (nextPhysRefLocation == bestFitLocation) + { + bestFitSet |= bestFitCandidateBit; + } + else + { + bool isBetter = false; + if (nextPhysRefLocation > lastLocation) { - regRecord = currentRefPosition->getReg(); - regRecord->recentRefPosition = currentRefPosition; - regNum = regRecord->regNum; + // This covers the full range; favor it if the other doesn't, or if it's a closer match. + if ((bestFitLocation <= lastLocation) || (nextPhysRefLocation < bestFitLocation)) + { + isBetter = true; + } } - else if (currentRefPosition->isIntervalRef()) + else { - interval = currentRefPosition->getInterval(); - interval->recentRefPosition = currentRefPosition; - if (currentRefPosition->registerAssignment != RBM_NONE) + // This doesn't cover the full range; favor it if the other doesn't either, but this ends later. + if ((bestFitLocation <= lastLocation) && (nextPhysRefLocation > bestFitLocation)) { - if (!genMaxOneBit(currentRefPosition->registerAssignment)) - { - assert(currentRefPosition->refType == RefTypeExpUse || - currentRefPosition->refType == RefTypeDummyDef); - } - else - { - regNum = currentRefPosition->assignedReg(); - regRecord = getRegisterRecord(regNum); - } + isBetter = true; } } + if (isBetter) + { + bestFitSet = bestFitCandidateBit; + bestFitLocation = nextPhysRefLocation; + } } + } + assert(bestFitSet != RBM_NONE); + found = applySelection(BEST_FIT, bestFitSet); +} - LsraLocation newLocation = currentRefPosition->nodeLocation; - currentLocation = newLocation; +// Apply the IS_PREV_REG heuristic. TODO: Check if Only applies if we have freeCandidates. +// Oddly, the previous heuristics only considered this if it covered the range. +void LinearScan::RegisterSelection::try_IS_PREV_REG() +{ + // TODO: We do not check found here. + if ((prevRegRec != nullptr) && ((score & COVERS_FULL) != 0)) + { + found = applySingleRegSelection(IS_PREV_REG, prevRegBit); + } +} - switch (currentRefPosition->refType) +// Apply the REG_ORDER heuristic. Only applies if we have freeCandidates. +void LinearScan::RegisterSelection::try_REG_ORDER() +{ + assert(!found); + + if (freeCandidates == RBM_NONE) // TODO: Kpathak, is this needed? + { + return; + } + + // This will always result in a single candidate. That is, it is the tie-breaker + // for free candidates, and doesn't make sense as anything other than the last + // heuristic for free registers. + unsigned lowestRegOrder = UINT_MAX; + regMaskTP lowestRegOrderBit = RBM_NONE; + for (regMaskTP regOrderCandidates = candidates; regOrderCandidates != RBM_NONE;) + { + regMaskTP regOrderCandidateBit = genFindLowestBit(regOrderCandidates); + regOrderCandidates &= ~regOrderCandidateBit; + regNumber regOrderCandidateRegNum = genRegNumFromMask(regOrderCandidateBit); + unsigned thisRegOrder = linearScan->getRegisterRecord(regOrderCandidateRegNum)->regOrder; + if (thisRegOrder < lowestRegOrder) { - case RefTypeBB: - { - if (currentBlock == nullptr) - { - currentBlock = startBlockSequence(); - } - else - { - // Verify the resolution moves at the end of the previous block. - for (GenTree* node = firstBlockEndResolutionNode; node != nullptr; node = node->gtNext) - { - assert(enregisterLocalVars); - // Only verify nodes that are actually moves; don't bother with the nodes that are - // operands to moves. - if (IsResolutionMove(node)) - { - verifyResolutionMove(node, currentLocation); - } - } + lowestRegOrder = thisRegOrder; + lowestRegOrderBit = regOrderCandidateBit; + } + } + assert(lowestRegOrderBit != RBM_NONE); + found = applySingleRegSelection(REG_ORDER, lowestRegOrderBit); +} - // Validate the locations at the end of the previous block. - if (enregisterLocalVars) - { - VarToRegMap outVarToRegMap = outVarToRegMaps[currentBlock->bbNum]; - VarSetOps::Iter iter(compiler, currentBlock->bbLiveOut); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex)) - { - if (localVarIntervals[varIndex] == nullptr) - { - assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); - continue; - } - regNumber regNum = getVarReg(outVarToRegMap, varIndex); - interval = getIntervalForLocalVar(varIndex); - if (interval->physReg != regNum) - { - assert(regNum == REG_STK); - assert((interval->physReg == REG_NA) || interval->isWriteThru); - } - interval->physReg = REG_NA; - interval->assignedReg = nullptr; - interval->isActive = false; - } - } +void LinearScan::RegisterSelection::try_SPILL_COST() +{ + assert(!found); + + // The set of registers with the lowest spill weight. + regMaskTP lowestCostSpillSet = RBM_NONE; + // Apply the SPILL_COST heuristic and eliminate regs that can't be spilled. - // Clear register assignments. - for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) - { - RegRecord* physRegRecord = getRegisterRecord(reg); - physRegRecord->assignedInterval = nullptr; - } + // The spill weight for 'refPosition' (the one we're allocating now). + float thisSpillWeight = linearScan->getWeight(refPosition); + // The spill weight for the best candidate we've found so far. + float bestSpillWeight = FloatingPointUtils::infinite_float(); + // True if we found registers with lower spill weight than this refPosition. + bool foundLowerSpillWeight = false; - // Now, record the locations at the beginning of this block. - currentBlock = moveToNextBlock(); - } + for (regMaskTP spillCandidates = candidates; spillCandidates != RBM_NONE;) + { + regMaskTP spillCandidateBit = genFindLowestBit(spillCandidates); + spillCandidates &= ~spillCandidateBit; + regNumber spillCandidateRegNum = genRegNumFromMask(spillCandidateBit); + RegRecord* spillCandidateRegRecord = &linearScan->physRegs[spillCandidateRegNum]; - if (currentBlock != nullptr) - { - if (enregisterLocalVars) - { - VarToRegMap inVarToRegMap = inVarToRegMaps[currentBlock->bbNum]; - VarSetOps::Iter iter(compiler, currentBlock->bbLiveIn); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex)) - { - if (localVarIntervals[varIndex] == nullptr) - { - assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); - continue; - } - regNumber regNum = getVarReg(inVarToRegMap, varIndex); - interval = getIntervalForLocalVar(varIndex); - interval->physReg = regNum; - interval->assignedReg = &(physRegs[regNum]); - interval->isActive = true; - physRegs[regNum].assignedInterval = interval; - } - } + // Can and should the interval in this register be spilled for this one, + // if we don't find a better alternative? + if ((linearScan->getNextIntervalRef(spillCandidateRegNum, regType) == currentLocation) && + !spillCandidateRegRecord->assignedInterval->getNextRefPosition()->RegOptional()) + { + continue; + } + if (!linearScan->isSpillCandidate(currentInterval, refPosition, spillCandidateRegRecord)) + { + continue; + } - if (VERBOSE) - { - dumpRefPositionShort(currentRefPosition, currentBlock); - dumpRegRecords(); - } + float currentSpillWeight = linearScan->spillCost[spillCandidateRegNum]; +#ifdef TARGET_ARM + if (currentInterval->registerType == TYP_DOUBLE) + { + currentSpillWeight = max(currentSpillWeight, linearScan->spillCost[REG_NEXT(spillCandidateRegNum)]); + } +#endif + if (currentSpillWeight < bestSpillWeight) + { + bestSpillWeight = currentSpillWeight; + lowestCostSpillSet = spillCandidateBit; + } + else if (currentSpillWeight == bestSpillWeight) + { + lowestCostSpillSet |= spillCandidateBit; + } + } - // Finally, handle the resolution moves, if any, at the beginning of the next block. - firstBlockEndResolutionNode = nullptr; - bool foundNonResolutionNode = false; + if (lowestCostSpillSet == RBM_NONE) + { + return; + } - LIR::Range& currentBlockRange = LIR::AsRange(currentBlock); - for (GenTree* node : currentBlockRange.NonPhiNodes()) - { - if (IsResolutionNode(currentBlockRange, node)) - { - assert(enregisterLocalVars); - if (foundNonResolutionNode) - { - firstBlockEndResolutionNode = node; - break; - } - else if (IsResolutionMove(node)) - { - // Only verify nodes that are actually moves; don't bother with the nodes that are - // operands to moves. - verifyResolutionMove(node, currentLocation); - } - } - else - { - foundNonResolutionNode = true; - } - } - } - } + // We won't spill if this refPosition is RegOptional() and we have no candidates + // with a lower spill cost. + if ((bestSpillWeight >= thisSpillWeight) && refPosition->RegOptional()) + { + currentInterval->assignedReg = nullptr; + skipAllocation = true; + found = true; + } - break; + // We must have at least one with the lowest spill cost. + assert(lowestCostSpillSet != RBM_NONE); + found = applySelection(SPILL_COST, lowestCostSpillSet); +} - case RefTypeKill: - assert(regRecord != nullptr); - assert(regRecord->assignedInterval == nullptr); - dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum, currentBlock); - break; - case RefTypeFixedReg: - assert(regRecord != nullptr); - dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum, currentBlock); - break; +// Apply the FAR_NEXT_REF heuristic. +void LinearScan::RegisterSelection::try_FAR_NEXT_REF() +{ + assert(!found); - case RefTypeUpperVectorSave: - dumpLsraAllocationEvent(LSRA_EVENT_UPPER_VECTOR_SAVE, nullptr, REG_NA, currentBlock); - break; + LsraLocation farthestLocation = MinLocation; + regMaskTP farthestSet = RBM_NONE; + for (regMaskTP farthestCandidates = candidates; farthestCandidates != RBM_NONE;) + { + regMaskTP farthestCandidateBit = genFindLowestBit(farthestCandidates); + farthestCandidates &= ~farthestCandidateBit; + regNumber farthestCandidateRegNum = genRegNumFromMask(farthestCandidateBit); - case RefTypeUpperVectorRestore: - dumpLsraAllocationEvent(LSRA_EVENT_UPPER_VECTOR_RESTORE, nullptr, REG_NA, currentBlock); - break; + // Find the next RefPosition of the register. + LsraLocation nextIntervalLocation = + linearScan->getNextIntervalRef(farthestCandidateRegNum, currentInterval->registerType); + LsraLocation nextPhysRefLocation = Min(linearScan->nextFixedRef[farthestCandidateRegNum], nextIntervalLocation); + if (nextPhysRefLocation == farthestLocation) + { + farthestSet |= farthestCandidateBit; + } + else if (nextPhysRefLocation > farthestLocation) + { + farthestSet = farthestCandidateBit; + farthestLocation = nextPhysRefLocation; + } + } + // We must have at least one with the lowest spill cost. + assert(farthestSet != RBM_NONE); + found = applySelection(FAR_NEXT_REF, farthestSet); +} - case RefTypeDef: - case RefTypeUse: - case RefTypeParamDef: - case RefTypeZeroInit: - assert(interval != nullptr); +// Apply the PREV_REG_OPT heuristic. +void LinearScan::RegisterSelection::try_PREV_REG_OPT() +{ + assert(!found); - if (interval->isSpecialPutArg) - { - dumpLsraAllocationEvent(LSRA_EVENT_SPECIAL_PUTARG, interval, regNum); - break; - } - if (currentRefPosition->reload) - { - interval->isActive = true; - assert(regNum != REG_NA); - interval->physReg = regNum; - interval->assignedReg = regRecord; - regRecord->assignedInterval = interval; - dumpLsraAllocationEvent(LSRA_EVENT_RELOAD, nullptr, regRecord->regNum, currentBlock); - } - if (regNum == REG_NA) - { - // If this interval is still assigned to a register - if (interval->physReg != REG_NA) - { - // then unassign it if no new register was assigned to the RefTypeDef - if (RefTypeIsDef(currentRefPosition->refType)) - { - assert(interval->assignedReg != nullptr); - if (interval->assignedReg->assignedInterval == interval) - { - interval->assignedReg->assignedInterval = nullptr; - } - interval->physReg = REG_NA; - interval->assignedReg = nullptr; - } - } + regMaskTP prevRegOptSet = RBM_NONE; + for (regMaskTP prevRegOptCandidates = candidates; prevRegOptCandidates != RBM_NONE;) + { + regMaskTP prevRegOptCandidateBit = genFindLowestBit(prevRegOptCandidates); + prevRegOptCandidates &= ~prevRegOptCandidateBit; + regNumber prevRegOptCandidateRegNum = genRegNumFromMask(prevRegOptCandidateBit); + Interval* assignedInterval = linearScan->physRegs[prevRegOptCandidateRegNum].assignedInterval; + bool foundPrevRegOptReg = true; +#ifdef DEBUG + bool hasAssignedInterval = false; +#endif - dumpLsraAllocationEvent(LSRA_EVENT_NO_REG_ALLOCATED, interval); - } - else if (RefTypeIsDef(currentRefPosition->refType)) - { - interval->isActive = true; + if ((assignedInterval != nullptr) && (assignedInterval->recentRefPosition != nullptr)) + { + foundPrevRegOptReg &= + (assignedInterval->recentRefPosition->reload && assignedInterval->recentRefPosition->RegOptional()); +#ifdef DEBUG + hasAssignedInterval = true; +#endif + } +#ifndef TARGET_ARM + else + { + foundPrevRegOptReg = false; + } +#endif - if (VERBOSE) - { - if (interval->isConstant && (currentRefPosition->treeNode != nullptr) && - currentRefPosition->treeNode->IsReuseRegVal()) - { - dumpLsraAllocationEvent(LSRA_EVENT_REUSE_REG, nullptr, regRecord->regNum, currentBlock); - } - else - { - dumpLsraAllocationEvent(LSRA_EVENT_ALLOC_REG, nullptr, regRecord->regNum, currentBlock); - } - } - } - else if (currentRefPosition->copyReg) - { - dumpLsraAllocationEvent(LSRA_EVENT_COPY_REG, interval, regRecord->regNum, currentBlock); - } - else if (currentRefPosition->moveReg) - { - assert(interval->assignedReg != nullptr); - interval->assignedReg->assignedInterval = nullptr; - interval->physReg = regNum; - interval->assignedReg = regRecord; - regRecord->assignedInterval = interval; - if (VERBOSE) - { - dumpEmptyRefPosition(); - printf("Move %-4s ", getRegName(regRecord->regNum)); - } - } - else - { - dumpLsraAllocationEvent(LSRA_EVENT_KEPT_ALLOCATION, nullptr, regRecord->regNum, currentBlock); - } - if (currentRefPosition->lastUse || (currentRefPosition->spillAfter && !currentRefPosition->writeThru)) - { - interval->isActive = false; - } - if (regNum != REG_NA) - { - if (currentRefPosition->spillAfter) - { - if (VERBOSE) - { - // If refPos is marked as copyReg, then the reg that is spilled - // is the homeReg of the interval not the reg currently assigned - // to refPos. - regNumber spillReg = regNum; - if (currentRefPosition->copyReg) - { - assert(interval != nullptr); - spillReg = interval->physReg; - } - dumpRegRecords(); - dumpEmptyRefPosition(); - if (currentRefPosition->writeThru) - { - printf("WThru %-4s ", getRegName(spillReg)); - } - else - { - printf("Spill %-4s ", getRegName(spillReg)); - } - } - } - else if (currentRefPosition->copyReg) - { - regRecord->assignedInterval = interval; - } - else - { - if (RefTypeIsDef(currentRefPosition->refType)) - { - // Interval was assigned to a different register. - // Clear the assigned interval of current register. - if (interval->physReg != REG_NA && interval->physReg != regNum) - { - interval->assignedReg->assignedInterval = nullptr; - } - } - interval->physReg = regNum; - interval->assignedReg = regRecord; - regRecord->assignedInterval = interval; - } - } - break; - case RefTypeKillGCRefs: - // No action to take. - // However, we will assert that, at resolution time, no registers contain GC refs. +#ifdef TARGET_ARM + // If current interval is TYP_DOUBLE, verify if the other half register matches the heuristics. + // We have three cases: + // 1. One of the register of the pair have an assigned interval: Check if that register's refPosition + // matches the heuristics. If yes, add it to the set. + // 2. Both registers of the pair have an assigned interval: Conservatively "and" conditions for + // heuristics of their corresponding refPositions. If both register's heuristic matches, add them + // to the set. TODO-CQ-ARM: We may implement a better condition later. + // 3. None of the register have an assigned interval: Skip adding register and assert. + if (currentInterval->registerType == TYP_DOUBLE) + { + regNumber anotherHalfRegNum = linearScan->findAnotherHalfRegNum(prevRegOptCandidateRegNum); + assignedInterval = linearScan->physRegs[anotherHalfRegNum].assignedInterval; + if ((assignedInterval != nullptr) && (assignedInterval->recentRefPosition != nullptr)) + { + if (assignedInterval->recentRefPosition->reload && assignedInterval->recentRefPosition->RegOptional()) { - DBEXEC(VERBOSE, printf(" ")); - regMaskTP candidateRegs = currentRefPosition->registerAssignment; - while (candidateRegs != RBM_NONE) - { - regMaskTP nextRegBit = genFindLowestBit(candidateRegs); - candidateRegs &= ~nextRegBit; - regNumber nextReg = genRegNumFromMask(nextRegBit); - RegRecord* regRecord = getRegisterRecord(nextReg); - Interval* assignedInterval = regRecord->assignedInterval; - assert(assignedInterval == nullptr || !varTypeIsGC(assignedInterval->registerType)); - } + foundPrevRegOptReg &= (assignedInterval->recentRefPosition->reload && + assignedInterval->recentRefPosition->RegOptional()); } - break; +#ifdef DEBUG + hasAssignedInterval = true; +#endif + } + } +#endif - case RefTypeExpUse: - case RefTypeDummyDef: - // Do nothing; these will be handled by the RefTypeBB. - DBEXEC(VERBOSE, dumpRefPositionShort(currentRefPosition, currentBlock)); - DBEXEC(VERBOSE, printf(" ")); - break; + if (foundPrevRegOptReg) + { + // TODO-Cleanup: Previously, we always used the highest regNum with a previous regOptional + // RefPosition, which is not really consistent with the way other selection criteria are + // applied. should probably be: prevRegOptSet |= prevRegOptCandidateBit; + prevRegOptSet = prevRegOptCandidateBit; + } - case RefTypeInvalid: - // for these 'currentRefPosition->refType' values, No action to take - break; +#ifdef DEBUG + // The assigned should be non-null, and should have a recentRefPosition, however since + // this is a heuristic, we don't want a fatal error, so we just assert (not noway_assert). + if (!hasAssignedInterval) + { + assert(!"Spill candidate has no assignedInterval recentRefPosition"); } +#endif + } + found = applySelection(PREV_REG_OPT, prevRegOptSet); +} - if (currentRefPosition->refType != RefTypeBB) +// Apply the REG_NUM heuristic. +void LinearScan::RegisterSelection::try_REG_NUM() +{ + assert(!found); + + found = applySingleRegSelection(REG_NUM, genFindLowestBit(candidates)); +} + +void LinearScan::RegisterSelection::calculateSets() +{ + if (freeCandidates == RBM_NONE || coversSetsCalculated) + { + return; + } + + preferenceSet = (candidates & preferences); + regMaskTP coversCandidates = (preferenceSet == RBM_NONE) ? candidates : preferenceSet; + for (; coversCandidates != RBM_NONE;) + { + regMaskTP coversCandidateBit = genFindLowestBit(coversCandidates); + coversCandidates &= ~coversCandidateBit; + regNumber coversCandidateRegNum = genRegNumFromMask(coversCandidateBit); + + // If we have a single candidate we don't need to compute the preference-related sets, but we + // do need to compute the unassignedSet. + if (!found) { - DBEXEC(VERBOSE, dumpRegRecords()); - if (interval != nullptr) + // Find the next RefPosition of the register. + LsraLocation nextIntervalLocation = linearScan->getNextIntervalRef(coversCandidateRegNum, regType); + LsraLocation nextPhysRefLocation = linearScan->getNextFixedRef(coversCandidateRegNum, regType); + LsraLocation coversCandidateLocation = Min(nextPhysRefLocation, nextIntervalLocation); + + // If the nextPhysRefLocation is a fixedRef for the rangeEndRefPosition, increment it so that + // we don't think it isn't covering the live range. + // This doesn't handle the case where earlier RefPositions for this Interval are also + // FixedRefs of this regNum, but at least those are only interesting in the case where those + // are "local last uses" of the Interval - otherwise the liveRange would interfere with the reg. + if (coversCandidateLocation == rangeEndLocation && + rangeEndRefPosition->isFixedRefOfReg(coversCandidateRegNum)) { - if (currentRefPosition->copyReg) + INDEBUG(linearScan->dumpLsraAllocationEvent(LSRA_EVENT_INCREMENT_RANGE_END, currentInterval)); + coversCandidateLocation++; + } + if (coversCandidateLocation > rangeEndLocation) + { + coversSet |= coversCandidateBit; + } + if ((coversCandidateBit & relatedPreferences) != RBM_NONE) + { + if (coversCandidateLocation > relatedInterval->lastRefPosition->nodeLocation) { - assert(interval->physReg != regNum); - regRecord->assignedInterval = nullptr; - assert(interval->assignedReg != nullptr); - regRecord = interval->assignedReg; + coversRelatedSet |= coversCandidateBit; } - if (currentRefPosition->spillAfter || currentRefPosition->lastUse) + } + else if (coversCandidateBit == refPosition->registerAssignment) + { + // If we had a fixed-reg def of a reg that will be killed before the use, prefer it to any other + // registers with the same score. (Note that we haven't changed the original registerAssignment + // on the RefPosition). + // Overload the RELATED_PREFERENCE value. + // TODO-CQ: Consider if this should be split out. + coversRelatedSet |= coversCandidateBit; + } + // Does this cover the full range of the interval? + if (coversCandidateLocation > lastLocation) + { + coversFullSet |= coversCandidateBit; + } + } + // The register is considered unassigned if it has no assignedInterval, OR + // if its next reference is beyond the range of this interval. + if (linearScan->nextIntervalRef[coversCandidateRegNum] > lastLocation) + { + unassignedSet |= coversCandidateBit; + } + } + + coversSetsCalculated = true; +} + +regMaskTP LinearScan::RegisterSelection::select(Interval* currentInterval, + RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)) +{ +#ifdef DEBUG + *registerScore = NONE; +#endif + + reset(currentInterval, refPosition); + + // process data-structures + if (RefTypeIsDef(refPosition->refType)) + { + if (currentInterval->hasConflictingDefUse) + { + linearScan->resolveConflictingDefAndUse(currentInterval, refPosition); + candidates = refPosition->registerAssignment; + } + // Otherwise, check for the case of a fixed-reg def of a reg that will be killed before the + // use, or interferes at the point of use (which shouldn't happen, but Lower doesn't mark + // the contained nodes as interfering). + // Note that we may have a ParamDef RefPosition that is marked isFixedRegRef, but which + // has had its registerAssignment changed to no longer be a single register. + else if (refPosition->isFixedRegRef && nextRefPos != nullptr && RefTypeIsUse(nextRefPos->refType) && + !nextRefPos->isFixedRegRef && genMaxOneBit(refPosition->registerAssignment)) + { + regNumber defReg = refPosition->assignedReg(); + RegRecord* defRegRecord = linearScan->getRegisterRecord(defReg); + + RefPosition* currFixedRegRefPosition = defRegRecord->recentRefPosition; + assert(currFixedRegRefPosition != nullptr && + currFixedRegRefPosition->nodeLocation == refPosition->nodeLocation); + + // If there is another fixed reference to this register before the use, change the candidates + // on this RefPosition to include that of nextRefPos. + RefPosition* nextFixedRegRefPosition = defRegRecord->getNextRefPosition(); + if (nextFixedRegRefPosition != nullptr && + nextFixedRegRefPosition->nodeLocation <= nextRefPos->getRefEndLocation()) + { + candidates |= nextRefPos->registerAssignment; + if (preferences == refPosition->registerAssignment) { - assert(!currentRefPosition->spillAfter || currentRefPosition->IsActualRef()); + preferences = candidates; + } + } + } + } - if (RefTypeIsDef(currentRefPosition->refType)) - { - // If an interval got assigned to a different register (while the different - // register got spilled), then clear the assigned interval of current register. - if (interval->physReg != REG_NA && interval->physReg != regNum) - { - interval->assignedReg->assignedInterval = nullptr; - } - } + preferences &= candidates; + if (preferences == RBM_NONE) + { + preferences = candidates; + } - interval->physReg = REG_NA; - interval->assignedReg = nullptr; +#ifdef DEBUG + candidates = linearScan->stressLimitRegs(refPosition, candidates); +#endif + assert(candidates != RBM_NONE); - // regRegcord could be null if the RefPosition does not require a register. - if (regRecord != nullptr) - { - regRecord->assignedInterval = nullptr; - } -#if FEATURE_PARTIAL_SIMD_CALLEE_SAVE - else if (interval->isUpperVector && !currentRefPosition->RegOptional()) - { - // These only require a register if they are not RegOptional, and their lclVar - // interval is living in a register and not already partially spilled. - if ((currentRefPosition->refType == RefTypeUpperVectorSave) || - (currentRefPosition->refType == RefTypeUpperVectorRestore)) - { - Interval* lclVarInterval = interval->relatedInterval; - assert((lclVarInterval->physReg == REG_NA) || lclVarInterval->isPartiallySpilled); - } - } + Interval* nextRelatedInterval = relatedInterval; + Interval* finalRelatedInterval = relatedInterval; + Interval* rangeEndInterval = relatedInterval; + + bool avoidByteRegs = false; +#ifdef TARGET_X86 + if ((relatedPreferences & ~RBM_BYTE_REGS) != RBM_NONE) + { + avoidByteRegs = true; + } #endif - else + + // Follow the chain of related intervals, as long as: + // - The next reference is a def. We don't want to use the relatedInterval for preferencing if its next reference + // is not a new definition (as it either is or will become live). + // - The next (def) reference is downstream. Otherwise we could iterate indefinitely because the preferences can be + // circular. + // - The intersection of preferenced registers is non-empty. + // + while (nextRelatedInterval != nullptr) + { + RefPosition* nextRelatedRefPosition = nextRelatedInterval->getNextRefPosition(); + + // Only use the relatedInterval for preferencing if the related interval's next reference + // is a new definition. + if ((nextRelatedRefPosition != nullptr) && RefTypeIsDef(nextRelatedRefPosition->refType)) + { + finalRelatedInterval = nextRelatedInterval; + nextRelatedInterval = nullptr; + + // First, get the preferences for this interval + regMaskTP thisRelatedPreferences = finalRelatedInterval->getCurrentPreferences(); + // Now, determine if they are compatible and update the relatedPreferences that we'll consider. + regMaskTP newRelatedPreferences = thisRelatedPreferences & relatedPreferences; + if (newRelatedPreferences != RBM_NONE && (!avoidByteRegs || thisRelatedPreferences != RBM_BYTE_REGS)) + { + // TODO-CQ: The following isFree() check doesn't account for the possibility that there's an + // assignedInterval whose recentRefPosition was delayFree. It also fails to account for + // the TYP_DOUBLE case on ARM. It would be better to replace the call to isFree with + // isRegAvailable(genRegNumFromMask(newRelatedPreferences), regType)), but this is retained + // to achieve zero diffs. + // + bool thisIsSingleReg = isSingleRegister(newRelatedPreferences); + if (!thisIsSingleReg || + (finalRelatedInterval->isLocalVar && + linearScan->isFree(linearScan->getRegisterRecord(genRegNumFromMask(newRelatedPreferences))))) + { + relatedPreferences = newRelatedPreferences; + // If this Interval has a downstream def without a single-register preference, continue to iterate. + if (nextRelatedRefPosition->nodeLocation > rangeEndLocation) { - assert(currentRefPosition->RegOptional()); + preferCalleeSave = (preferCalleeSave || finalRelatedInterval->preferCalleeSave); + rangeEndLocation = nextRelatedRefPosition->getRangeEndLocation(); + rangeEndInterval = finalRelatedInterval; + nextRelatedInterval = finalRelatedInterval->relatedInterval; } } } } + else + { + if (nextRelatedInterval == relatedInterval) + { + relatedInterval = nullptr; + relatedPreferences = RBM_NONE; + } + nextRelatedInterval = nullptr; + } + } + + // For floating point, we want to be less aggressive about using callee-save registers. + // So in that case, we just need to ensure that the current RefPosition is covered. + if (useFloatReg(currentInterval->registerType)) + { + rangeEndRefPosition = refPosition; + preferCalleeSave = currentInterval->preferCalleeSave; + } + else if (currentInterval->isWriteThru && refPosition->spillAfter) + { + // This is treated as a last use of the register, as there is an upcoming EH boundary. + rangeEndRefPosition = refPosition; + } + else + { + rangeEndRefPosition = refPosition->getRangeEndRef(); + // If we have a chain of related intervals, and a finalRelatedInterval that + // is not currently occupying a register, and whose lifetime begins after this one, + // we want to try to select a register that will cover its lifetime. + if ((rangeEndInterval != nullptr) && (rangeEndInterval->assignedReg == nullptr) && + !rangeEndInterval->isWriteThru && + (rangeEndInterval->getNextRefLocation() >= rangeEndRefPosition->nodeLocation)) + { + lastRefPosition = rangeEndInterval->lastRefPosition; + } + } + if ((relatedInterval != nullptr) && !relatedInterval->isWriteThru) + { + relatedLastLocation = relatedInterval->lastRefPosition->nodeLocation; + } + + if (preferCalleeSave) + { + regMaskTP calleeSaveCandidates = calleeSaveRegs(currentInterval->registerType); + if (currentInterval->isWriteThru) + { + // We'll only prefer a callee-save register if it's already been used. + regMaskTP unusedCalleeSaves = + calleeSaveCandidates & ~(linearScan->compiler->codeGen->regSet.rsGetModifiedRegsMask()); + callerCalleePrefs = calleeSaveCandidates & ~unusedCalleeSaves; + preferences &= ~unusedCalleeSaves; + } + else + { + callerCalleePrefs = calleeSaveCandidates; + } + } + else + { + callerCalleePrefs = callerSaveRegs(currentInterval->registerType); } - // Now, verify the resolution blocks. - // Currently these are nearly always at the end of the method, but that may not always be the case. - // So, we'll go through all the BBs looking for blocks whose bbNum is greater than bbNumMaxBeforeResolution. - for (BasicBlock* currentBlock = compiler->fgFirstBB; currentBlock != nullptr; currentBlock = currentBlock->bbNext) + // If this has a delayed use (due to being used in a rmw position of a + // non-commutative operator), its endLocation is delayed until the "def" + // position, which is one location past the use (getRefEndLocation() takes care of this). + rangeEndLocation = rangeEndRefPosition->getRefEndLocation(); + lastLocation = lastRefPosition->getRefEndLocation(); + + // We'll set this to short-circuit remaining heuristics when we have a single candidate. + found = false; + + // Is this a fixedReg? + regMaskTP fixedRegMask = RBM_NONE; + if (refPosition->isFixedRegRef) { - if (currentBlock->bbNum > bbNumMaxBeforeResolution) + assert(genMaxOneBit(refPosition->registerAssignment)); + fixedRegMask = refPosition->registerAssignment; + if (candidates == refPosition->registerAssignment) { - // If we haven't enregistered an lclVars, we have no resolution blocks. - assert(enregisterLocalVars); - - if (VERBOSE) + found = true; + if (linearScan->nextIntervalRef[genRegNumFromMask(candidates)] > lastLocation) { - dumpRegRecordTitle(); - printf(shortRefPositionFormat, 0, 0); - assert(currentBlock->bbPreds != nullptr && currentBlock->bbPreds->getBlock() != nullptr); - printf(bbRefPosFormat, currentBlock->bbNum, currentBlock->bbPreds->getBlock()->bbNum); - dumpRegRecords(); + unassignedSet = candidates; } + } + } - // Clear register assignments. - for (regNumber reg = REG_FIRST; reg < ACTUAL_REG_COUNT; reg = REG_NEXT(reg)) - { - RegRecord* physRegRecord = getRegisterRecord(reg); - physRegRecord->assignedInterval = nullptr; - } + // Eliminate candidates that are in-use or busy. + if (!found) + { + regMaskTP busyRegs = linearScan->regsBusyUntilKill | linearScan->regsInUseThisLocation; + candidates &= ~busyRegs; - // Set the incoming register assignments - VarToRegMap inVarToRegMap = getInVarToRegMap(currentBlock->bbNum); - VarSetOps::Iter iter(compiler, currentBlock->bbLiveIn); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex)) - { - if (localVarIntervals[varIndex] == nullptr) - { - assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); - continue; - } - regNumber regNum = getVarReg(inVarToRegMap, varIndex); - Interval* interval = getIntervalForLocalVar(varIndex); - interval->physReg = regNum; - interval->assignedReg = &(physRegs[regNum]); - interval->isActive = true; - physRegs[regNum].assignedInterval = interval; - } + // Also eliminate as busy any register with a conflicting fixed reference at this or + // the next location. + // Note that this will eliminate the fixedReg, if any, but we'll add it back below. + regMaskTP checkConflictMask = candidates & linearScan->fixedRegs; + while (checkConflictMask != RBM_NONE) + { + regMaskTP checkConflictBit = genFindLowestBit(checkConflictMask); + checkConflictMask &= ~checkConflictBit; + regNumber checkConflictReg = genRegNumFromMask(checkConflictBit); + LsraLocation checkConflictLocation = linearScan->nextFixedRef[checkConflictReg]; - // Verify the moves in this block - LIR::Range& currentBlockRange = LIR::AsRange(currentBlock); - for (GenTree* node : currentBlockRange.NonPhiNodes()) + if ((checkConflictLocation == currentLocation) || + (refPosition->delayRegFree && (checkConflictLocation == (currentLocation + 1)))) { - assert(IsResolutionNode(currentBlockRange, node)); - if (IsResolutionMove(node)) - { - // Only verify nodes that are actually moves; don't bother with the nodes that are - // operands to moves. - verifyResolutionMove(node, currentLocation); - } + candidates &= ~checkConflictBit; } + } + candidates |= fixedRegMask; + found = isSingleRegister(candidates); + } - // Verify the outgoing register assignments - { - VarToRegMap outVarToRegMap = getOutVarToRegMap(currentBlock->bbNum); - VarSetOps::Iter iter(compiler, currentBlock->bbLiveOut); - unsigned varIndex = 0; - while (iter.NextElem(&varIndex)) - { - if (localVarIntervals[varIndex] == nullptr) - { - assert(!compiler->lvaGetDescByTrackedIndex(varIndex)->lvLRACandidate); - continue; - } - regNumber regNum = getVarReg(outVarToRegMap, varIndex); - Interval* interval = getIntervalForLocalVar(varIndex); - // Either the register assignments match, or the outgoing assignment is on the stack - // and this is a write-thru interval. - assert(interval->physReg == regNum || (interval->physReg == REG_NA && regNum == REG_STK) || - (interval->isWriteThru && regNum == REG_STK)); - interval->physReg = REG_NA; - interval->assignedReg = nullptr; - interval->isActive = false; - } - } + // By chance, is prevRegRec already holding this interval, as a copyReg or having + // been restored as inactive after a kill? + // NOTE: this is not currently considered one of the selection criteria - it always wins + // if it is the assignedInterval of 'prevRegRec'. + if (!found && (prevRegRec != nullptr)) + { + prevRegBit = genRegMask(prevRegRec->regNum); + if ((prevRegRec->assignedInterval == currentInterval) && ((candidates & prevRegBit) != RBM_NONE)) + { + candidates = prevRegBit; + found = true; +#ifdef DEBUG + *registerScore = THIS_ASSIGNED; +#endif } } + else + { + prevRegBit = RBM_NONE; + } - DBEXEC(VERBOSE, printf("\n")); -} + if (!found && (candidates == RBM_NONE)) + { + assert(refPosition->RegOptional()); + currentInterval->assignedReg = nullptr; + return REG_NA; + } -//------------------------------------------------------------------------ -// verifyResolutionMove: Verify a resolution statement. Called by verifyFinalAllocation() -// -// Arguments: -// resolutionMove - A GenTree* that must be a resolution move. -// currentLocation - The LsraLocation of the most recent RefPosition that has been verified. -// -// Return Value: -// None. -// -// Notes: -// If verbose is set, this will also dump the moves into the table of final allocations. -void LinearScan::verifyResolutionMove(GenTree* resolutionMove, LsraLocation currentLocation) -{ - GenTree* dst = resolutionMove; - assert(IsResolutionMove(dst)); + // TODO-Cleanup: Previously, the "reverseSelect" stress mode reversed the order of the heuristics. + // It needs to be re-engineered with this refactoring. + // In non-debug builds, this will simply get optimized away + bool reverseSelect = false; +#ifdef DEBUG + reverseSelect = linearScan->doReverseSelect(); +#endif // DEBUG - if (dst->OperGet() == GT_SWAP) + freeCandidates = linearScan->getFreeCandidates(candidates, regType); + + // If no free candidates, then double check if refPosition is an actual ref. + if (freeCandidates == RBM_NONE) { - GenTreeLclVarCommon* left = dst->gtGetOp1()->AsLclVarCommon(); - GenTreeLclVarCommon* right = dst->gtGetOp2()->AsLclVarCommon(); - regNumber leftRegNum = left->GetRegNum(); - regNumber rightRegNum = right->GetRegNum(); - LclVarDsc* leftVarDsc = compiler->lvaTable + left->GetLclNum(); - LclVarDsc* rightVarDsc = compiler->lvaTable + right->GetLclNum(); - Interval* leftInterval = getIntervalForLocalVar(leftVarDsc->lvVarIndex); - Interval* rightInterval = getIntervalForLocalVar(rightVarDsc->lvVarIndex); - assert(leftInterval->physReg == leftRegNum && rightInterval->physReg == rightRegNum); - leftInterval->physReg = rightRegNum; - rightInterval->physReg = leftRegNum; - leftInterval->assignedReg = &physRegs[rightRegNum]; - rightInterval->assignedReg = &physRegs[leftRegNum]; - physRegs[rightRegNum].assignedInterval = leftInterval; - physRegs[leftRegNum].assignedInterval = rightInterval; - if (VERBOSE) + // We won't spill if this refPosition is not an actual ref. + if (!refPosition->IsActualRef()) { - printf(shortRefPositionFormat, currentLocation, 0); - dumpIntervalName(leftInterval); - printf(" Swap "); - printf(" %-4s ", getRegName(rightRegNum)); - dumpRegRecords(); - printf(shortRefPositionFormat, currentLocation, 0); - dumpIntervalName(rightInterval); - printf(" \" "); - printf(" %-4s ", getRegName(leftRegNum)); - dumpRegRecords(); + currentInterval->assignedReg = nullptr; + return REG_NA; } - return; } - regNumber dstRegNum = dst->GetRegNum(); - regNumber srcRegNum; - GenTreeLclVarCommon* lcl; - if (dst->OperGet() == GT_COPY) + else { - lcl = dst->gtGetOp1()->AsLclVarCommon(); - srcRegNum = lcl->GetRegNum(); + // Set the 'matchingConstants' set. + if (currentInterval->isConstant && RefTypeIsDef(refPosition->refType)) + { + matchingConstants = linearScan->getMatchingConstants(candidates, currentInterval, refPosition); + } } - else + +#define IF_FOUND_GOTO_DONE \ + if (found) \ + goto Selection_Done; + +#ifdef DEBUG + HeuristicFn fn; + for (int orderId = 0; orderId < REGSELECT_HEURISTIC_COUNT /*heuristicsCount*/; orderId++) { - lcl = dst->AsLclVarCommon(); - if ((lcl->gtFlags & GTF_SPILLED) != 0) + IF_FOUND_GOTO_DONE + + RegisterScore heuristicToApply = RegSelectionOrder[orderId]; + if (mappingTable->Lookup(heuristicToApply, &fn)) { - srcRegNum = REG_STK; + (this->*fn)(); + if (found) + { + *registerScore = heuristicToApply; + } + +#if TRACK_LSRA_STATS + INTRACK_STATS_IF(found, linearScan->updateLsraStat(linearScan->getLsraStatFromScore(heuristicToApply), + refPosition->bbNum)); +#endif // TRACK_LSRA_STATS } else { - assert((lcl->gtFlags & GTF_SPILL) != 0); - srcRegNum = dstRegNum; - dstRegNum = REG_STK; + assert(!"Unexpected heuristic value!"); } } +#else // RELEASE + // In release, just invoke the default order - Interval* interval = getIntervalForLocalVarNode(lcl); - assert(interval->physReg == srcRegNum || (srcRegNum == REG_STK && interval->physReg == REG_NA)); - if (srcRegNum != REG_STK) - { - physRegs[srcRegNum].assignedInterval = nullptr; - } - if (dstRegNum != REG_STK) - { - interval->physReg = dstRegNum; - interval->assignedReg = &(physRegs[dstRegNum]); - physRegs[dstRegNum].assignedInterval = interval; - interval->isActive = true; - } - else - { - interval->physReg = REG_NA; - interval->assignedReg = nullptr; - interval->isActive = false; - } - if (VERBOSE) +#define REG_SEL_DEF(stat, value, shortname) \ + try_##stat(); \ + IF_FOUND_GOTO_DONE +#include "lsra_score.h" +#undef REG_SEL_DEF +#endif // DEBUG +#undef IF_FOUND_GOTO_DONE + +Selection_Done: + if (skipAllocation) { - printf(shortRefPositionFormat, currentLocation, 0); - dumpIntervalName(interval); - printf(" Move "); - printf(" %-4s ", getRegName(dstRegNum)); - dumpRegRecords(); + return REG_NA; } + + calculateSets(); + + assert(found && isSingleRegister(candidates)); + foundRegBit = candidates; + return candidates; } -#endif // DEBUG diff --git a/src/coreclr/jit/lsra.h b/src/coreclr/jit/lsra.h index 1cf89d419b8a8..99039029fddc9 100644 --- a/src/coreclr/jit/lsra.h +++ b/src/coreclr/jit/lsra.h @@ -1169,7 +1169,7 @@ class LinearScan : public LinearScanInterface /***************************************************************************** * Register selection ****************************************************************************/ - regMaskTP getFreeCandidates(regMaskTP candidates, var_types regType) + regMaskTP getFreeCandidates(regMaskTP candidates, var_types regType) //TODO: Move this into RegisterSelection { regMaskTP result = candidates & m_AvailableRegs; #ifdef TARGET_ARM @@ -1183,54 +1183,141 @@ class LinearScan : public LinearScanInterface return result; } - struct registerSelector +#ifdef DEBUG + class RegisterSelection; + // For lsra ordering experimentation + + typedef void (LinearScan::RegisterSelection::*HeuristicFn)(); + typedef JitHashTable, HeuristicFn> ScoreMappingTable; +#define REGSELECT_HEURISTIC_COUNT 17 +#endif + + class RegisterSelection { - regMaskTP candidates; - int score; -#ifdef TARGET_ARM - var_types regType; -#endif // TARGET_ARM + public: + RegisterSelection(LinearScan* linearScan); - // Apply a simple mask-based selection heuristic, and return 'true' if we now have a single candidate. - bool applySelection(int selectionScore, regMaskTP selectionCandidates DEBUG_ARG(RegisterScore* registerScore)) - { - regMaskTP newCandidates = candidates & selectionCandidates; - if (newCandidates != RBM_NONE) + // Perform register selection and update currentInterval or refPosition TODO: + FORCEINLINE regMaskTP select(Interval* currentInterval, RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)); + + // If the register is from unassigned set such that it was not already + // assigned to the current interval + FORCEINLINE bool foundUnassignedReg() { - score += selectionScore; - candidates = newCandidates; - bool found = isSingleRegister(candidates); -#ifdef DEBUG - if (found) - { - *registerScore = (RegisterScore)selectionScore; - } -#endif - return found; + assert(found && isSingleRegister(foundRegBit)); + bool isUnassignedReg = ((foundRegBit & unassignedSet) != RBM_NONE); + return isUnassignedReg && !isAlreadyAssigned(); } - return false; - } - // Select a single register, if it is in the candidate set. - // Return true if so. - bool applySingleRegSelection(int selectionScore, - regMaskTP selectionCandidate DEBUG_ARG(RegisterScore* registerScore)) - { - assert(isSingleRegister(selectionCandidate)); - regMaskTP newCandidates = candidates & selectionCandidate; - if (newCandidates != RBM_NONE) + // Did register selector decide to spill this interval + FORCEINLINE bool isSpilling() + { + return (foundRegBit & freeCandidates) == RBM_NONE; + } + + // Is the value one of the constant that is already in a register + FORCEINLINE bool isMatchingConstant() + { + assert(found && isSingleRegister(foundRegBit)); + return (matchingConstants & foundRegBit) != RBM_NONE; + } + + // Did we apply CONST_AVAILABLE heuristics + FORCEINLINE bool isConstAvailable() { - score += selectionScore; - candidates = newCandidates; + return (score & CONST_AVAILABLE) != 0; + } + + private: #ifdef DEBUG - *registerScore = (RegisterScore)selectionScore; + RegisterScore DefaultOrder[REGSELECT_HEURISTIC_COUNT] = + { +#define REG_SEL_DEF(stat, value, shortname) stat, +#include "lsra_score.h" +#undef REG_SEL_DEF + }; + + RegisterScore RegSelectionOrder[REGSELECT_HEURISTIC_COUNT] = {NONE}; + ScoreMappingTable* mappingTable = nullptr; #endif - return true; + LinearScan* linearScan = nullptr; + int score = 0; + Interval* currentInterval = nullptr; + RefPosition* refPosition = nullptr; + + RegisterType regType = RegisterType::TYP_UNKNOWN; + LsraLocation currentLocation = MinLocation; + RefPosition* nextRefPos = nullptr; + + regMaskTP candidates; + regMaskTP preferences = RBM_NONE; + Interval* relatedInterval = nullptr; + + regMaskTP relatedPreferences = RBM_NONE; + LsraLocation rangeEndLocation; + LsraLocation relatedLastLocation; // TODO:kpathak - need to see why this is not used after refactor? + bool preferCalleeSave = false; + RefPosition* rangeEndRefPosition; + RefPosition* lastRefPosition; + regMaskTP callerCalleePrefs = RBM_NONE; + LsraLocation lastLocation; + RegRecord* prevRegRec = nullptr; + + regMaskTP prevRegBit = RBM_NONE; + + // These are used in the post-selection updates, and must be set for any selection. + regMaskTP freeCandidates; + regMaskTP matchingConstants; + regMaskTP unassignedSet; + regMaskTP foundRegBit; + + // Compute the sets for COVERS, OWN_PREFERENCE, COVERS_RELATED, COVERS_FULL and UNASSIGNED together, + // as they all require similar computation. + regMaskTP coversSet; + regMaskTP preferenceSet; + regMaskTP coversRelatedSet; + regMaskTP coversFullSet; + bool coversSetsCalculated = false; + bool found = false; + bool skipAllocation = false; + regNumber foundReg = REG_NA; + + // If the selected register is already assigned to the current internal + FORCEINLINE bool isAlreadyAssigned() + { + assert(found && isSingleRegister(candidates)); + return (prevRegBit & preferences) == foundRegBit; } - return false; - } + + bool applySelection(int selectionScore, regMaskTP selectionCandidates); + bool applySingleRegSelection(int selectionScore, regMaskTP selectionCandidate); + FORCEINLINE void calculateSets(); + FORCEINLINE void reset(Interval* interval, RefPosition* refPosition); + +#define REG_SEL_DEF(stat, value, shortname) FORCEINLINE void try_##stat(); +#include "lsra_score.h" +#undef REG_SEL_DEF + /*FORCEINLINE void tryFree(); + FORCEINLINE void tryConstAvailable(); + FORCEINLINE void tryThisAssigned(); + FORCEINLINE void tryCovers(); + FORCEINLINE void tryOwnPreference(); + FORCEINLINE void tryCoversRelated(); + FORCEINLINE void tryRelatedPreference(); + FORCEINLINE void tryCallerCallee(); + FORCEINLINE void tryUnassigned(); + FORCEINLINE void tryCoversFull(); + FORCEINLINE void tryBestFit(); + FORCEINLINE void tryIsPrevReg(); + FORCEINLINE void tryRegOrder(); + FORCEINLINE void trySpillCost(); + FORCEINLINE void tryFarNextRef(); + FORCEINLINE void tryPrevRegOpt(); + FORCEINLINE void tryRegNum();*/ }; + RegisterSelection* regSelector; + /***************************************************************************** * For Resolution phase ****************************************************************************/ @@ -1386,6 +1473,7 @@ class LinearScan : public LinearScanInterface unsigned regCandidateVarCount; void updateLsraStat(LsraStat stat, unsigned currentBBNum); void dumpLsraStats(FILE* file); + LsraStat getLsraStatFromScore(RegisterScore registerScore); LsraStat firstRegSelStat = STAT_FREE; public: From 0e1f3a619b356cc281f9b1c1b1093a4f7c679927 Mon Sep 17 00:00:00 2001 From: Kunal Pathak Date: Mon, 17 May 2021 10:16:26 -0700 Subject: [PATCH 2/7] Add missing return condition --- src/coreclr/jit/lsra.cpp | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/src/coreclr/jit/lsra.cpp b/src/coreclr/jit/lsra.cpp index d946d9a90e4ff..918a99d23f2b6 100644 --- a/src/coreclr/jit/lsra.cpp +++ b/src/coreclr/jit/lsra.cpp @@ -2758,6 +2758,11 @@ regNumber LinearScan::allocateReg(Interval* currentInterval, RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)) { regMaskTP foundRegBit = regSelector->select(currentInterval, refPosition DEBUG_ARG(registerScore)); + if (foundRegBit == REG_NA) + { + return REG_NA; + } + regNumber foundReg = genRegNumFromMask(foundRegBit); RegRecord* availablePhysRegRecord = getRegisterRecord(foundReg); Interval* assignedInterval = availablePhysRegRecord->assignedInterval; From 43a4ddf92557a79449eb3c3e83fcf614867f316f Mon Sep 17 00:00:00 2001 From: Kunal Pathak Date: Tue, 18 May 2021 11:54:30 -0700 Subject: [PATCH 3/7] Minor cleanup - summary docs - Removed DEFAULT_ORDER - Added order seq ID --- src/coreclr/jit/lsra.cpp | 170 ++++++++++++++++++++++++++--------- src/coreclr/jit/lsra.h | 36 ++------ src/coreclr/jit/lsra_score.h | 37 ++++---- 3 files changed, 152 insertions(+), 91 deletions(-) diff --git a/src/coreclr/jit/lsra.cpp b/src/coreclr/jit/lsra.cpp index 918a99d23f2b6..d5430ab527bd4 100644 --- a/src/coreclr/jit/lsra.cpp +++ b/src/coreclr/jit/lsra.cpp @@ -8559,7 +8559,7 @@ const char* LinearScan::getStatName(unsigned stat) #define LSRA_STAT_DEF(stat, name) name, #include "lsra_stats.h" #undef LSRA_STAT_DEF -#define REG_SEL_DEF(stat, value, shortname) #stat, +#define REG_SEL_DEF(stat, value, shortname, orderSeqId) #stat, #include "lsra_score.h" #undef REG_SEL_DEF }; @@ -8572,8 +8572,8 @@ LsraStat LinearScan::getLsraStatFromScore(RegisterScore registerScore) { switch (registerScore) { -#define REG_SEL_DEF(stat, value, shortname) \ - case RegisterScore::stat: \ +#define REG_SEL_DEF(stat, value, shortname, orderSeqId) \ + case RegisterScore::stat: \ return LsraStat::STAT_##stat; #include "lsra_score.h" #undef REG_SEL_DEF @@ -8852,7 +8852,7 @@ const char* LinearScan::getScoreName(RegisterScore score) { switch (score) { -#define REG_SEL_DEF(stat, value, shortname) \ +#define REG_SEL_DEF(stat, value, shortname, orderSeqId) \ case stat: \ return shortname; #include "lsra_score.h" @@ -10808,29 +10808,40 @@ LinearScan::RegisterSelection::RegisterSelection(LinearScan* linearScan) #ifdef DEBUG mappingTable = new ScoreMappingTable(linearScan->compiler->getAllocator(CMK_LSRA)); -#define REG_SEL_DEF(stat, value, shortname) mappingTable->Set(stat, &LinearScan::RegisterSelection::try_##stat); +#define REG_SEL_DEF(stat, value, shortname, orderSeqId) \ + mappingTable->Set(stat, &LinearScan::RegisterSelection::try_##stat); #include "lsra_score.h" #undef REG_SEL_DEF LPCWSTR ordering = JitConfig.JitLsraOrdering(); - if (ordering == nullptr) { - memcpy(RegSelectionOrder, DefaultOrder, sizeof(RegisterScore) * REGSELECT_HEURISTIC_COUNT); + ordering = L"ABCDEFGHIJKLMNOPQ"; } - else + + for (int orderId = 0; orderId < REGSELECT_HEURISTIC_COUNT; orderId++) { - for (int orderId = 0; orderId < REGSELECT_HEURISTIC_COUNT; orderId++) - { - // Make sure we do not set repeated entries - assert(RegSelectionOrder[orderId] == NONE); + // Make sure we do not set repeated entries + assert(RegSelectionOrder[orderId] == NONE); - RegSelectionOrder[orderId] = DefaultOrder[ordering[orderId] - 'A']; + switch (ordering[orderId]) + { +#define REG_SEL_DEF(enum_name, value, shortname, orderSeqId) \ + case orderSeqId: \ + RegSelectionOrder[orderId] = RegisterScore::##enum_name; \ + break; +#include "lsra_score.h" +#undef REG_SEL_DEF + default: + assert(!"Invalid lsraOrdering value."); } } #endif // DEBUG } +// ---------------------------------------------------------- +// reset: Resets the values of all the fields used for register selection. +// void LinearScan::RegisterSelection::reset(Interval* interval, RefPosition* refPos) { currentInterval = interval; @@ -10849,7 +10860,7 @@ void LinearScan::RegisterSelection::reset(Interval* interval, RefPosition* refPo relatedPreferences = (relatedInterval == nullptr) ? RBM_NONE : relatedInterval->getCurrentPreferences(); rangeEndLocation = refPosition->getRangeEndLocation(); - relatedLastLocation = rangeEndLocation; // TODO:kpathak - need to see why this is not used after refactor? + relatedLastLocation = rangeEndLocation; preferCalleeSave = currentInterval->preferCalleeSave; rangeEndRefPosition = nullptr; lastRefPosition = currentInterval->lastRefPosition; @@ -10872,6 +10883,16 @@ void LinearScan::RegisterSelection::reset(Interval* interval, RefPosition* refPo coversSetsCalculated = false; } +// ---------------------------------------------------------- +// applySelection: Apply the heuristic to the candidates. +// +// Arguments: +// selectionScore: The score corresponding to the heuristics we apply. +// selectionCandidates: The possible candidates for the heuristic to apply. +// +// Return Values: +// 'true' if there was a single register candidate available after the heuristic is applied. +// bool LinearScan::RegisterSelection::applySelection(int selectionScore, regMaskTP selectionCandidates) { regMaskTP newCandidates = candidates & selectionCandidates; @@ -10884,8 +10905,16 @@ bool LinearScan::RegisterSelection::applySelection(int selectionScore, regMaskTP return false; } -// Select a single register, if it is in the candidate set. -// Return true if so. +// ---------------------------------------------------------- +// applySingleRegSelection: Select a single register, if it is in the candidate set. +// +// Arguments: +// selectionScore: The score corresponding to the heuristics we apply. +// selectionCandidates: The possible candidates for the heuristic to apply. +// +// Return Values: +// 'true' if there was a single register candidate available after the heuristic is applied. +// bool LinearScan::RegisterSelection::applySingleRegSelection(int selectionScore, regMaskTP selectionCandidate) { assert(LinearScan::isSingleRegister(selectionCandidate)); @@ -10898,6 +10927,9 @@ bool LinearScan::RegisterSelection::applySingleRegSelection(int selectionScore, return false; } +// ---------------------------------------------------------- +// try_FREE: Apply the FREE heuristic. +// void LinearScan::RegisterSelection::try_FREE() { assert(!found); @@ -10909,8 +10941,11 @@ void LinearScan::RegisterSelection::try_FREE() found = applySelection(FREE, freeCandidates); } -// Apply the CONST_AVAILABLE (matching constant) heuristic. -// Note that we always need to define the 'matchingConstants' set. +// ---------------------------------------------------------- +// try_CONST_AVAILABLE: Apply the CONST_AVAILABLE (matching constant) heuristic. +// +// Note: we always need to define the 'matchingConstants' set. +// void LinearScan::RegisterSelection::try_CONST_AVAILABLE() { assert(!found); @@ -10925,7 +10960,9 @@ void LinearScan::RegisterSelection::try_CONST_AVAILABLE() } } -// Apply the THIS_ASSIGNED heuristic. +// ---------------------------------------------------------- +// try_THIS_ASSIGNED: Apply the THIS_ASSIGNED heuristic. +// void LinearScan::RegisterSelection::try_THIS_ASSIGNED() { assert(!found); @@ -10940,42 +10977,51 @@ void LinearScan::RegisterSelection::try_THIS_ASSIGNED() } } -// Apply the COVERS heuristic. +// ---------------------------------------------------------- +// try_COVERS: Apply the COVERS heuristic. +// void LinearScan::RegisterSelection::try_COVERS() { assert(!found); - calculateSets(); + calculateCoversSets(); found = applySelection(COVERS, coversSet & preferenceSet); } -// Apply the OWN_PREFERENCE heuristic. -// Note that 'preferenceSet' already includes only freeCandidates. +// ---------------------------------------------------------- +// try_OWN_PREFERENCE: Apply the OWN_PREFERENCE heuristic. +// +// Note: 'preferenceSet' already includes only freeCandidates. +// void LinearScan::RegisterSelection::try_OWN_PREFERENCE() { assert(!found); #ifdef DEBUG - calculateSets(); + calculateCoversSets(); #endif found = applySelection(OWN_PREFERENCE, (preferenceSet & freeCandidates)); } -// Apply the COVERS_RELATED heuristic. +// ---------------------------------------------------------- +// try_COVERS_RELATED: Apply the COVERS_RELATED heuristic. +// void LinearScan::RegisterSelection::try_COVERS_RELATED() { assert(!found); #ifdef DEBUG - calculateSets(); + calculateCoversSets(); #endif found = applySelection(COVERS_RELATED, (coversRelatedSet & freeCandidates)); } -// Apply the RELATED_PREFERENCE heuristic. +// ---------------------------------------------------------- +// try_RELATED_PREFERENCE: Apply the RELATED_PREFERENCE heuristic. +// void LinearScan::RegisterSelection::try_RELATED_PREFERENCE() { assert(!found); @@ -10983,7 +11029,9 @@ void LinearScan::RegisterSelection::try_RELATED_PREFERENCE() found = applySelection(RELATED_PREFERENCE, relatedPreferences & freeCandidates); } -// Apply the CALLER_CALLEE heuristic. +// ---------------------------------------------------------- +// try_CALLER_CALLEE: Apply the CALLER_CALLEE heuristic. +// void LinearScan::RegisterSelection::try_CALLER_CALLEE() { assert(!found); @@ -10991,30 +11039,37 @@ void LinearScan::RegisterSelection::try_CALLER_CALLEE() found = applySelection(CALLER_CALLEE, callerCalleePrefs & freeCandidates); } -// Apply the UNASSIGNED heuristic. +// ---------------------------------------------------------- +// try_UNASSIGNED: Apply the UNASSIGNED heuristic. +// void LinearScan::RegisterSelection::try_UNASSIGNED() { assert(!found); #ifdef DEBUG - calculateSets(); + calculateCoversSets(); #endif found = applySelection(UNASSIGNED, unassignedSet); } -// Apply the COVERS_FULL heuristic. +// ---------------------------------------------------------- +// try_COVERS_FULL: Apply the COVERS_FULL heuristic. +// void LinearScan::RegisterSelection::try_COVERS_FULL() { assert(!found); #ifdef DEBUG - calculateSets(); + calculateCoversSets(); #endif found = applySelection(COVERS_FULL, (coversFullSet & freeCandidates)); } +// ---------------------------------------------------------- +// try_BEST_FIT: Apply the BEST_FIT heuristic. +// void LinearScan::RegisterSelection::try_BEST_FIT() { assert(!found); @@ -11086,8 +11141,12 @@ void LinearScan::RegisterSelection::try_BEST_FIT() found = applySelection(BEST_FIT, bestFitSet); } -// Apply the IS_PREV_REG heuristic. TODO: Check if Only applies if we have freeCandidates. -// Oddly, the previous heuristics only considered this if it covered the range. +// ---------------------------------------------------------- +// try_IS_PREV_REG: Apply the IS_PREV_REG heuristic. +// +// Note: Oddly, the previous heuristics only considered this if it covered the range. +// TODO: Check if Only applies if we have freeCandidates. +// void LinearScan::RegisterSelection::try_IS_PREV_REG() { // TODO: We do not check found here. @@ -11097,7 +11156,9 @@ void LinearScan::RegisterSelection::try_IS_PREV_REG() } } -// Apply the REG_ORDER heuristic. Only applies if we have freeCandidates. +// ---------------------------------------------------------- +// try_REG_ORDER: Apply the REG_ORDER heuristic. Only applies if we have freeCandidates. +// void LinearScan::RegisterSelection::try_REG_ORDER() { assert(!found); @@ -11128,6 +11189,9 @@ void LinearScan::RegisterSelection::try_REG_ORDER() found = applySingleRegSelection(REG_ORDER, lowestRegOrderBit); } +// ---------------------------------------------------------- +// try_SPILL_COST: Apply the SPILL_COST heuristic. +// void LinearScan::RegisterSelection::try_SPILL_COST() { assert(!found); @@ -11199,7 +11263,9 @@ void LinearScan::RegisterSelection::try_SPILL_COST() found = applySelection(SPILL_COST, lowestCostSpillSet); } -// Apply the FAR_NEXT_REF heuristic. +// ---------------------------------------------------------- +// try_FAR_NEXT_REF: Apply the FAR_NEXT_REF heuristic. +// void LinearScan::RegisterSelection::try_FAR_NEXT_REF() { assert(!found); @@ -11231,7 +11297,9 @@ void LinearScan::RegisterSelection::try_FAR_NEXT_REF() found = applySelection(FAR_NEXT_REF, farthestSet); } -// Apply the PREV_REG_OPT heuristic. +// ---------------------------------------------------------- +// try_PREV_REG_OPT: Apply the PREV_REG_OPT heuristic. +// void LinearScan::RegisterSelection::try_PREV_REG_OPT() { assert(!found); @@ -11310,7 +11378,9 @@ void LinearScan::RegisterSelection::try_PREV_REG_OPT() found = applySelection(PREV_REG_OPT, prevRegOptSet); } -// Apply the REG_NUM heuristic. +// ---------------------------------------------------------- +// try_REG_NUM: Apply the REG_NUM heuristic. +// void LinearScan::RegisterSelection::try_REG_NUM() { assert(!found); @@ -11318,7 +11388,11 @@ void LinearScan::RegisterSelection::try_REG_NUM() found = applySingleRegSelection(REG_NUM, genFindLowestBit(candidates)); } -void LinearScan::RegisterSelection::calculateSets() +// ---------------------------------------------------------- +// calculateCoversSets: Calculate the necessary covers set registers to be used +// for heuristics lke COVERS, COVERS_RELATED, COVERS_FULL. +// +void LinearScan::RegisterSelection::calculateCoversSets() { if (freeCandidates == RBM_NONE || coversSetsCalculated) { @@ -11359,7 +11433,7 @@ void LinearScan::RegisterSelection::calculateSets() } if ((coversCandidateBit & relatedPreferences) != RBM_NONE) { - if (coversCandidateLocation > relatedInterval->lastRefPosition->nodeLocation) + if (coversCandidateLocation > relatedLastLocation) { coversRelatedSet |= coversCandidateBit; } @@ -11390,6 +11464,16 @@ void LinearScan::RegisterSelection::calculateSets() coversSetsCalculated = true; } +// ---------------------------------------------------------- +// select: For given `currentInterval` and `refPosition`, selects a register to be assigned. +// +// Arguments: +// currentInterval - Current interval for which register needs to be selected. +// refPosition - Refposition within the interval for which register needs to be selected. +// +// Return Values: +// Register bit selected (a single register) and REG_NA if no register was selected. +// regMaskTP LinearScan::RegisterSelection::select(Interval* currentInterval, RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)) { @@ -11684,7 +11768,7 @@ regMaskTP LinearScan::RegisterSelection::select(Interval* current #ifdef DEBUG HeuristicFn fn; - for (int orderId = 0; orderId < REGSELECT_HEURISTIC_COUNT /*heuristicsCount*/; orderId++) + for (int orderId = 0; orderId < REGSELECT_HEURISTIC_COUNT; orderId++) { IF_FOUND_GOTO_DONE @@ -11710,7 +11794,7 @@ regMaskTP LinearScan::RegisterSelection::select(Interval* current #else // RELEASE // In release, just invoke the default order -#define REG_SEL_DEF(stat, value, shortname) \ +#define REG_SEL_DEF(stat, value, shortname, orderSeqId) \ try_##stat(); \ IF_FOUND_GOTO_DONE #include "lsra_score.h" @@ -11724,7 +11808,7 @@ regMaskTP LinearScan::RegisterSelection::select(Interval* current return REG_NA; } - calculateSets(); + calculateCoversSets(); assert(found && isSingleRegister(candidates)); foundRegBit = candidates; diff --git a/src/coreclr/jit/lsra.h b/src/coreclr/jit/lsra.h index 99039029fddc9..20991122c6703 100644 --- a/src/coreclr/jit/lsra.h +++ b/src/coreclr/jit/lsra.h @@ -381,7 +381,7 @@ enum LsraStat #define LSRA_STAT_DEF(enum_name, enum_str) enum_name, #include "lsra_stats.h" #undef LSRA_STAT_DEF -#define REG_SEL_DEF(enum_name, value, short_str) STAT_##enum_name, +#define REG_SEL_DEF(enum_name, value, short_str, orderSeqId) STAT_##enum_name, #include "lsra_score.h" #undef REG_SEL_DEF COUNT @@ -408,7 +408,7 @@ struct LsraBlockInfo enum RegisterScore { -#define REG_SEL_DEF(enum_name, value, short_str) enum_name = value, +#define REG_SEL_DEF(enum_name, value, short_str, orderSeqId) enum_name = value, #include "lsra_score.h" #undef REG_SEL_DEF NONE = 0 @@ -1169,7 +1169,7 @@ class LinearScan : public LinearScanInterface /***************************************************************************** * Register selection ****************************************************************************/ - regMaskTP getFreeCandidates(regMaskTP candidates, var_types regType) //TODO: Move this into RegisterSelection + regMaskTP getFreeCandidates(regMaskTP candidates, var_types regType) { regMaskTP result = candidates & m_AvailableRegs; #ifdef TARGET_ARM @@ -1230,13 +1230,6 @@ class LinearScan : public LinearScanInterface private: #ifdef DEBUG - RegisterScore DefaultOrder[REGSELECT_HEURISTIC_COUNT] = - { -#define REG_SEL_DEF(stat, value, shortname) stat, -#include "lsra_score.h" -#undef REG_SEL_DEF - }; - RegisterScore RegSelectionOrder[REGSELECT_HEURISTIC_COUNT] = {NONE}; ScoreMappingTable* mappingTable = nullptr; #endif @@ -1255,7 +1248,7 @@ class LinearScan : public LinearScanInterface regMaskTP relatedPreferences = RBM_NONE; LsraLocation rangeEndLocation; - LsraLocation relatedLastLocation; // TODO:kpathak - need to see why this is not used after refactor? + LsraLocation relatedLastLocation; bool preferCalleeSave = false; RefPosition* rangeEndRefPosition; RefPosition* lastRefPosition; @@ -1291,29 +1284,12 @@ class LinearScan : public LinearScanInterface bool applySelection(int selectionScore, regMaskTP selectionCandidates); bool applySingleRegSelection(int selectionScore, regMaskTP selectionCandidate); - FORCEINLINE void calculateSets(); + FORCEINLINE void calculateCoversSets(); FORCEINLINE void reset(Interval* interval, RefPosition* refPosition); -#define REG_SEL_DEF(stat, value, shortname) FORCEINLINE void try_##stat(); +#define REG_SEL_DEF(stat, value, shortname, orderSeqId) FORCEINLINE void try_##stat(); #include "lsra_score.h" #undef REG_SEL_DEF - /*FORCEINLINE void tryFree(); - FORCEINLINE void tryConstAvailable(); - FORCEINLINE void tryThisAssigned(); - FORCEINLINE void tryCovers(); - FORCEINLINE void tryOwnPreference(); - FORCEINLINE void tryCoversRelated(); - FORCEINLINE void tryRelatedPreference(); - FORCEINLINE void tryCallerCallee(); - FORCEINLINE void tryUnassigned(); - FORCEINLINE void tryCoversFull(); - FORCEINLINE void tryBestFit(); - FORCEINLINE void tryIsPrevReg(); - FORCEINLINE void tryRegOrder(); - FORCEINLINE void trySpillCost(); - FORCEINLINE void tryFarNextRef(); - FORCEINLINE void tryPrevRegOpt(); - FORCEINLINE void tryRegNum();*/ }; RegisterSelection* regSelector; diff --git a/src/coreclr/jit/lsra_score.h b/src/coreclr/jit/lsra_score.h index 6cc66703815a7..9b3b182522ad6 100644 --- a/src/coreclr/jit/lsra_score.h +++ b/src/coreclr/jit/lsra_score.h @@ -17,29 +17,30 @@ // of the ordering on asmDiffs. However, there is much more room for tuning, // and perhaps additional criteria. // -// name score short_name -REG_SEL_DEF(FREE, 0x10000, "FREE ") // It is not currently assigned to an *active* interval +// name score short_name orderSeqId +REG_SEL_DEF(FREE, 0x10000, "FREE ", 'A') // It is not currently assigned to an *active* interval // These are the original criteria for comparing registers that are free. -REG_SEL_DEF(CONST_AVAILABLE, 0x08000, "CONST") // It is a constant value that is already in an acceptable register. -REG_SEL_DEF(THIS_ASSIGNED, 0x04000, "THISA") // It is in the interval's preference set and it is already assigned to this interval. -REG_SEL_DEF(COVERS, 0x02000, "COVRS") // It is in the interval's preference set and it covers the current range. -REG_SEL_DEF(OWN_PREFERENCE, 0x01000, "OWNPR") // It is in the preference set of this interval. -REG_SEL_DEF(COVERS_RELATED, 0x00800, "COREL") // It is in the preference set of the related interval and covers its entire lifetime. -REG_SEL_DEF(RELATED_PREFERENCE, 0x00400, "RELPR") // It is in the preference set of the related interval. -REG_SEL_DEF(CALLER_CALLEE, 0x00200, "CRCE ") // It is in the right "set" for the interval (caller or callee-save). -REG_SEL_DEF(UNASSIGNED, 0x00100, "UNASG") // It is not currently assigned to any (active or inactive) interval -REG_SEL_DEF(COVERS_FULL, 0x00080, "COFUL") // It covers the full range of the interval from current position to the end. -REG_SEL_DEF(BEST_FIT, 0x00040, "BSFIT") // The available range is the closest match to the full range of the interval. -REG_SEL_DEF(IS_PREV_REG, 0x00020, "PRVRG") // This register was previously assigned to the interval. -REG_SEL_DEF(REG_ORDER, 0x00010, "ORDER") // Tie-breaker + +REG_SEL_DEF(CONST_AVAILABLE, 0x08000, "CONST", 'B') // It is a constant value that is already in an acceptable register. +REG_SEL_DEF(THIS_ASSIGNED, 0x04000, "THISA", 'C') // It is in the interval's preference set and it is already assigned to this interval. +REG_SEL_DEF(COVERS, 0x02000, "COVRS", 'D') // It is in the interval's preference set and it covers the current range. +REG_SEL_DEF(OWN_PREFERENCE, 0x01000, "OWNPR", 'E') // It is in the preference set of this interval. +REG_SEL_DEF(COVERS_RELATED, 0x00800, "COREL", 'F') // It is in the preference set of the related interval and covers its entire lifetime. +REG_SEL_DEF(RELATED_PREFERENCE, 0x00400, "RELPR", 'G') // It is in the preference set of the related interval. +REG_SEL_DEF(CALLER_CALLEE, 0x00200, "CRCE ", 'H') // It is in the right "set" for the interval (caller or callee-save). +REG_SEL_DEF(UNASSIGNED, 0x00100, "UNASG", 'I') // It is not currently assigned to any (active or inactive) interval +REG_SEL_DEF(COVERS_FULL, 0x00080, "COFUL", 'J') // It covers the full range of the interval from current position to the end. +REG_SEL_DEF(BEST_FIT, 0x00040, "BSFIT", 'K') // The available range is the closest match to the full range of the interval. +REG_SEL_DEF(IS_PREV_REG, 0x00020, "PRVRG", 'L') // This register was previously assigned to the interval. +REG_SEL_DEF(REG_ORDER, 0x00010, "ORDER", 'M') // Tie-breaker // These are the original criteria for comparing registers that are in use. -REG_SEL_DEF(SPILL_COST, 0x00008, "SPILL") // It has the lowest cost of all the candidates. -REG_SEL_DEF(FAR_NEXT_REF, 0x00004, "FNREF") // It has a farther next reference than the best candidate thus far. -REG_SEL_DEF(PREV_REG_OPT, 0x00002, "PRGOP") // The previous RefPosition of its current assigned interval is RegOptional. +REG_SEL_DEF(SPILL_COST, 0x00008, "SPILL", 'N') // It has the lowest cost of all the candidates. +REG_SEL_DEF(FAR_NEXT_REF, 0x00004, "FNREF", 'O') // It has a farther next reference than the best candidate thus far. +REG_SEL_DEF(PREV_REG_OPT, 0x00002, "PRGOP", 'P') // The previous RefPosition of its current assigned interval is RegOptional. // TODO-CQ: Consider using REG_ORDER as a tie-breaker even for busy registers. -REG_SEL_DEF(REG_NUM, 0x00001, "RGNUM") // It has a lower register number. +REG_SEL_DEF(REG_NUM, 0x00001, "RGNUM", 'Q') // It has a lower register number. // clang-format on From 87a3e989ecdf59d9641b91f7730812b3a09de4b3 Mon Sep 17 00:00:00 2001 From: Kunal Pathak Date: Tue, 18 May 2021 11:55:54 -0700 Subject: [PATCH 4/7] jit format --- src/coreclr/jit/lsra.cpp | 15 ++-- src/coreclr/jit/lsra.h | 167 ++++++++++++++++++++------------------- 2 files changed, 91 insertions(+), 91 deletions(-) diff --git a/src/coreclr/jit/lsra.cpp b/src/coreclr/jit/lsra.cpp index d5430ab527bd4..bafaf2ebd9608 100644 --- a/src/coreclr/jit/lsra.cpp +++ b/src/coreclr/jit/lsra.cpp @@ -626,7 +626,7 @@ LinearScan::LinearScan(Compiler* theCompiler) , refPositions(theCompiler->getAllocator(CMK_LSRA_RefPosition)) , listNodePool(theCompiler) { - regSelector = new (theCompiler, CMK_LSRA) RegisterSelection(this); + regSelector = new (theCompiler, CMK_LSRA) RegisterSelection(this); firstColdLoc = MaxLocation; #ifdef DEBUG @@ -2757,7 +2757,7 @@ bool LinearScan::isMatchingConstant(RegRecord* physRegRecord, RefPosition* refPo regNumber LinearScan::allocateReg(Interval* currentInterval, RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)) { - regMaskTP foundRegBit = regSelector->select(currentInterval, refPosition DEBUG_ARG(registerScore)); + regMaskTP foundRegBit = regSelector->select(currentInterval, refPosition DEBUG_ARG(registerScore)); if (foundRegBit == REG_NA) { return REG_NA; @@ -8572,9 +8572,9 @@ LsraStat LinearScan::getLsraStatFromScore(RegisterScore registerScore) { switch (registerScore) { -#define REG_SEL_DEF(stat, value, shortname, orderSeqId) \ - case RegisterScore::stat: \ - return LsraStat::STAT_##stat; +#define REG_SEL_DEF(stat, value, shortname, orderSeqId) \ + case RegisterScore::stat: \ + return LsraStat::STAT_##stat; #include "lsra_score.h" #undef REG_SEL_DEF default: @@ -10800,7 +10800,6 @@ void LinearScan::verifyResolutionMove(GenTree* resolutionMove, LsraLocation curr } #endif // DEBUG - LinearScan::RegisterSelection::RegisterSelection(LinearScan* linearScan) { this->linearScan = linearScan; @@ -11474,7 +11473,7 @@ void LinearScan::RegisterSelection::calculateCoversSets() // Return Values: // Register bit selected (a single register) and REG_NA if no register was selected. // -regMaskTP LinearScan::RegisterSelection::select(Interval* currentInterval, +regMaskTP LinearScan::RegisterSelection::select(Interval* currentInterval, RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)) { #ifdef DEBUG @@ -11792,7 +11791,7 @@ regMaskTP LinearScan::RegisterSelection::select(Interval* current } } #else // RELEASE - // In release, just invoke the default order +// In release, just invoke the default order #define REG_SEL_DEF(stat, value, shortname, orderSeqId) \ try_##stat(); \ diff --git a/src/coreclr/jit/lsra.h b/src/coreclr/jit/lsra.h index 20991122c6703..628781c232a1e 100644 --- a/src/coreclr/jit/lsra.h +++ b/src/coreclr/jit/lsra.h @@ -1194,98 +1194,99 @@ class LinearScan : public LinearScanInterface class RegisterSelection { - public: - RegisterSelection(LinearScan* linearScan); + public: + RegisterSelection(LinearScan* linearScan); - // Perform register selection and update currentInterval or refPosition TODO: - FORCEINLINE regMaskTP select(Interval* currentInterval, RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)); + // Perform register selection and update currentInterval or refPosition TODO: + FORCEINLINE regMaskTP select(Interval* currentInterval, + RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)); - // If the register is from unassigned set such that it was not already - // assigned to the current interval - FORCEINLINE bool foundUnassignedReg() - { - assert(found && isSingleRegister(foundRegBit)); - bool isUnassignedReg = ((foundRegBit & unassignedSet) != RBM_NONE); - return isUnassignedReg && !isAlreadyAssigned(); - } + // If the register is from unassigned set such that it was not already + // assigned to the current interval + FORCEINLINE bool foundUnassignedReg() + { + assert(found && isSingleRegister(foundRegBit)); + bool isUnassignedReg = ((foundRegBit & unassignedSet) != RBM_NONE); + return isUnassignedReg && !isAlreadyAssigned(); + } - // Did register selector decide to spill this interval - FORCEINLINE bool isSpilling() - { - return (foundRegBit & freeCandidates) == RBM_NONE; - } + // Did register selector decide to spill this interval + FORCEINLINE bool isSpilling() + { + return (foundRegBit & freeCandidates) == RBM_NONE; + } - // Is the value one of the constant that is already in a register - FORCEINLINE bool isMatchingConstant() - { - assert(found && isSingleRegister(foundRegBit)); - return (matchingConstants & foundRegBit) != RBM_NONE; - } + // Is the value one of the constant that is already in a register + FORCEINLINE bool isMatchingConstant() + { + assert(found && isSingleRegister(foundRegBit)); + return (matchingConstants & foundRegBit) != RBM_NONE; + } - // Did we apply CONST_AVAILABLE heuristics - FORCEINLINE bool isConstAvailable() - { - return (score & CONST_AVAILABLE) != 0; - } + // Did we apply CONST_AVAILABLE heuristics + FORCEINLINE bool isConstAvailable() + { + return (score & CONST_AVAILABLE) != 0; + } - private: + private: #ifdef DEBUG - RegisterScore RegSelectionOrder[REGSELECT_HEURISTIC_COUNT] = {NONE}; - ScoreMappingTable* mappingTable = nullptr; + RegisterScore RegSelectionOrder[REGSELECT_HEURISTIC_COUNT] = {NONE}; + ScoreMappingTable* mappingTable = nullptr; #endif - LinearScan* linearScan = nullptr; - int score = 0; - Interval* currentInterval = nullptr; - RefPosition* refPosition = nullptr; - - RegisterType regType = RegisterType::TYP_UNKNOWN; - LsraLocation currentLocation = MinLocation; - RefPosition* nextRefPos = nullptr; - - regMaskTP candidates; - regMaskTP preferences = RBM_NONE; - Interval* relatedInterval = nullptr; - - regMaskTP relatedPreferences = RBM_NONE; - LsraLocation rangeEndLocation; - LsraLocation relatedLastLocation; - bool preferCalleeSave = false; - RefPosition* rangeEndRefPosition; - RefPosition* lastRefPosition; - regMaskTP callerCalleePrefs = RBM_NONE; - LsraLocation lastLocation; - RegRecord* prevRegRec = nullptr; - - regMaskTP prevRegBit = RBM_NONE; - - // These are used in the post-selection updates, and must be set for any selection. - regMaskTP freeCandidates; - regMaskTP matchingConstants; - regMaskTP unassignedSet; - regMaskTP foundRegBit; - - // Compute the sets for COVERS, OWN_PREFERENCE, COVERS_RELATED, COVERS_FULL and UNASSIGNED together, - // as they all require similar computation. - regMaskTP coversSet; - regMaskTP preferenceSet; - regMaskTP coversRelatedSet; - regMaskTP coversFullSet; - bool coversSetsCalculated = false; - bool found = false; - bool skipAllocation = false; - regNumber foundReg = REG_NA; - - // If the selected register is already assigned to the current internal - FORCEINLINE bool isAlreadyAssigned() - { - assert(found && isSingleRegister(candidates)); - return (prevRegBit & preferences) == foundRegBit; - } + LinearScan* linearScan = nullptr; + int score = 0; + Interval* currentInterval = nullptr; + RefPosition* refPosition = nullptr; + + RegisterType regType = RegisterType::TYP_UNKNOWN; + LsraLocation currentLocation = MinLocation; + RefPosition* nextRefPos = nullptr; + + regMaskTP candidates; + regMaskTP preferences = RBM_NONE; + Interval* relatedInterval = nullptr; + + regMaskTP relatedPreferences = RBM_NONE; + LsraLocation rangeEndLocation; + LsraLocation relatedLastLocation; + bool preferCalleeSave = false; + RefPosition* rangeEndRefPosition; + RefPosition* lastRefPosition; + regMaskTP callerCalleePrefs = RBM_NONE; + LsraLocation lastLocation; + RegRecord* prevRegRec = nullptr; + + regMaskTP prevRegBit = RBM_NONE; + + // These are used in the post-selection updates, and must be set for any selection. + regMaskTP freeCandidates; + regMaskTP matchingConstants; + regMaskTP unassignedSet; + regMaskTP foundRegBit; + + // Compute the sets for COVERS, OWN_PREFERENCE, COVERS_RELATED, COVERS_FULL and UNASSIGNED together, + // as they all require similar computation. + regMaskTP coversSet; + regMaskTP preferenceSet; + regMaskTP coversRelatedSet; + regMaskTP coversFullSet; + bool coversSetsCalculated = false; + bool found = false; + bool skipAllocation = false; + regNumber foundReg = REG_NA; + + // If the selected register is already assigned to the current internal + FORCEINLINE bool isAlreadyAssigned() + { + assert(found && isSingleRegister(candidates)); + return (prevRegBit & preferences) == foundRegBit; + } - bool applySelection(int selectionScore, regMaskTP selectionCandidates); - bool applySingleRegSelection(int selectionScore, regMaskTP selectionCandidate); - FORCEINLINE void calculateCoversSets(); - FORCEINLINE void reset(Interval* interval, RefPosition* refPosition); + bool applySelection(int selectionScore, regMaskTP selectionCandidates); + bool applySingleRegSelection(int selectionScore, regMaskTP selectionCandidate); + FORCEINLINE void calculateCoversSets(); + FORCEINLINE void reset(Interval* interval, RefPosition* refPosition); #define REG_SEL_DEF(stat, value, shortname, orderSeqId) FORCEINLINE void try_##stat(); #include "lsra_score.h" From a36289bf52b677ab9707afb423bf2c9b0eb22d94 Mon Sep 17 00:00:00 2001 From: Kunal Pathak Date: Tue, 18 May 2021 15:32:09 -0700 Subject: [PATCH 5/7] fix linux build --- src/coreclr/jit/lsra.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/coreclr/jit/lsra.cpp b/src/coreclr/jit/lsra.cpp index bafaf2ebd9608..fc853b4367a70 100644 --- a/src/coreclr/jit/lsra.cpp +++ b/src/coreclr/jit/lsra.cpp @@ -10815,7 +10815,7 @@ LinearScan::RegisterSelection::RegisterSelection(LinearScan* linearScan) LPCWSTR ordering = JitConfig.JitLsraOrdering(); if (ordering == nullptr) { - ordering = L"ABCDEFGHIJKLMNOPQ"; + ordering = W("ABCDEFGHIJKLMNOPQ"); } for (int orderId = 0; orderId < REGSELECT_HEURISTIC_COUNT; orderId++) @@ -10827,7 +10827,7 @@ LinearScan::RegisterSelection::RegisterSelection(LinearScan* linearScan) { #define REG_SEL_DEF(enum_name, value, shortname, orderSeqId) \ case orderSeqId: \ - RegSelectionOrder[orderId] = RegisterScore::##enum_name; \ + RegSelectionOrder[orderId] = enum_name; \ break; #include "lsra_score.h" #undef REG_SEL_DEF From 58bf6ffab30fe4740fb2f61a3458f8c92fab6a3a Mon Sep 17 00:00:00 2001 From: Kunal Pathak Date: Tue, 18 May 2021 15:44:10 -0700 Subject: [PATCH 6/7] review feedback --- src/coreclr/jit/lsra.cpp | 8 ++++---- src/coreclr/jit/lsra.h | 2 +- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/src/coreclr/jit/lsra.cpp b/src/coreclr/jit/lsra.cpp index fc853b4367a70..3f1746c58e17f 100644 --- a/src/coreclr/jit/lsra.cpp +++ b/src/coreclr/jit/lsra.cpp @@ -2758,7 +2758,7 @@ regNumber LinearScan::allocateReg(Interval* currentInterval, RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)) { regMaskTP foundRegBit = regSelector->select(currentInterval, refPosition DEBUG_ARG(registerScore)); - if (foundRegBit == REG_NA) + if (foundRegBit == RBM_NONE) { return REG_NA; } @@ -11729,7 +11729,7 @@ regMaskTP LinearScan::RegisterSelection::select(Interval* currentInterval, { assert(refPosition->RegOptional()); currentInterval->assignedReg = nullptr; - return REG_NA; + return RBM_NONE; } // TODO-Cleanup: Previously, the "reverseSelect" stress mode reversed the order of the heuristics. @@ -11749,7 +11749,7 @@ regMaskTP LinearScan::RegisterSelection::select(Interval* currentInterval, if (!refPosition->IsActualRef()) { currentInterval->assignedReg = nullptr; - return REG_NA; + return RBM_NONE; } } else @@ -11804,7 +11804,7 @@ regMaskTP LinearScan::RegisterSelection::select(Interval* currentInterval, Selection_Done: if (skipAllocation) { - return REG_NA; + return RBM_NONE; } calculateCoversSets(); diff --git a/src/coreclr/jit/lsra.h b/src/coreclr/jit/lsra.h index 628781c232a1e..d944c4530739d 100644 --- a/src/coreclr/jit/lsra.h +++ b/src/coreclr/jit/lsra.h @@ -1197,7 +1197,7 @@ class LinearScan : public LinearScanInterface public: RegisterSelection(LinearScan* linearScan); - // Perform register selection and update currentInterval or refPosition TODO: + // Perform register selection and update currentInterval or refPosition FORCEINLINE regMaskTP select(Interval* currentInterval, RefPosition* refPosition DEBUG_ARG(RegisterScore* registerScore)); From 94d5571c87bdc86fdf971c32c4ffe3a0067e5f85 Mon Sep 17 00:00:00 2001 From: Kunal Pathak Date: Tue, 18 May 2021 16:58:45 -0700 Subject: [PATCH 7/7] Remove TODO --- src/coreclr/jit/lsra.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/coreclr/jit/lsra.cpp b/src/coreclr/jit/lsra.cpp index 3f1746c58e17f..8417b1d63e9a7 100644 --- a/src/coreclr/jit/lsra.cpp +++ b/src/coreclr/jit/lsra.cpp @@ -10845,7 +10845,7 @@ void LinearScan::RegisterSelection::reset(Interval* interval, RefPosition* refPo { currentInterval = interval; refPosition = refPos; - score = 0; // TODO:Kpathak - why is this used? + score = 0; regType = linearScan->getRegisterType(currentInterval, refPosition); currentLocation = refPosition->nodeLocation; @@ -11162,7 +11162,7 @@ void LinearScan::RegisterSelection::try_REG_ORDER() { assert(!found); - if (freeCandidates == RBM_NONE) // TODO: Kpathak, is this needed? + if (freeCandidates == RBM_NONE) { return; }