[RyuJIT] Implement Interlocked.And and Interlocked.Or for arm64-v8.1

src/coreclr/jit/codegenarm64.cpp

@@ -2786,10 +2786,10 @@ void CodeGen::genJumpTable(GenTree* treeNode)
 }
 
 //------------------------------------------------------------------------
-// genLockedInstructions: Generate code for a GT_XADD or GT_XCHG node.
+// genLockedInstructions: Generate code for a GT_XADD, GT_XAND, GT_XXOR or GT_XCHG node.
 //
 // Arguments:
-//    treeNode - the GT_XADD/XCHG node
+//    treeNode - the GT_XADD/XAND/XXOR/XCHG node
 //
 void CodeGen::genLockedInstructions(GenTreeOp* treeNode)
 {
@@ -2810,6 +2810,19 @@ void CodeGen::genLockedInstructions(GenTreeOp* treeNode)
 
         switch (treeNode->gtOper)
         {
+            case GT_XXOR:
+                GetEmitter()->emitIns_R_R_R(INS_ldsetal, dataSize, dataReg, (targetReg == REG_NA) ? REG_ZR : targetReg,
+                                            addrReg);
+                break;
+            case GT_XAND:
+            {
+                // Grab a temp reg to perform `MVN` for dataReg first.
+                regNumber tempReg = treeNode->GetSingleTempReg();
+                GetEmitter()->emitIns_R_R(INS_mvn, dataSize, tempReg, dataReg);
+                GetEmitter()->emitIns_R_R_R(INS_ldclral, dataSize, tempReg, (targetReg == REG_NA) ? REG_ZR : targetReg,
+                                            addrReg);
+                break;
+            }
             case GT_XCHG:
                 GetEmitter()->emitIns_R_R_R(INS_swpal, dataSize, dataReg, targetReg, addrReg);
                 break;
@@ -2831,6 +2844,9 @@ void CodeGen::genLockedInstructions(GenTreeOp* treeNode)
     }
     else
     {
+        // These are imported normally if Atomics aren't supported.
+        assert(!treeNode->OperIs(GT_XXOR, GT_XAND));
+
         regNumber exResultReg  = treeNode->ExtractTempReg(RBM_ALLINT);
         regNumber storeDataReg = (treeNode->OperGet() == GT_XCHG) ? dataReg : treeNode->ExtractTempReg(RBM_ALLINT);
         regNumber loadReg      = (targetReg != REG_NA) ? targetReg : storeDataReg;
@@ -6221,6 +6237,10 @@ void CodeGen::genArm64EmitterUnitTests()
     theEmitter->emitIns_R_R_R(INS_ldadd, EA_8BYTE, REG_R8, REG_R9, REG_R10);
     theEmitter->emitIns_R_R_R(INS_ldadda, EA_8BYTE, REG_R8, REG_R9, REG_R10);
     theEmitter->emitIns_R_R_R(INS_ldaddal, EA_8BYTE, REG_R8, REG_R9, REG_R10);
+    theEmitter->emitIns_R_R_R(INS_ldclral, EA_4BYTE, REG_R8, REG_R9, REG_R10);
+    theEmitter->emitIns_R_R_R(INS_ldclral, EA_8BYTE, REG_R8, REG_R9, REG_R10);
+    theEmitter->emitIns_R_R_R(INS_ldsetal, EA_4BYTE, REG_R8, REG_R9, REG_R10);
+    theEmitter->emitIns_R_R_R(INS_ldsetal, EA_8BYTE, REG_R8, REG_R9, REG_R10);
     theEmitter->emitIns_R_R_R(INS_ldaddl, EA_8BYTE, REG_R8, REG_R9, REG_R10);
     theEmitter->emitIns_R_R_R(INS_swpb, EA_4BYTE, REG_R8, REG_R9, REG_R10);
     theEmitter->emitIns_R_R_R(INS_swpab, EA_4BYTE, REG_R8, REG_R9, REG_R10);

src/coreclr/jit/codegenarmarch.cpp

@@ -407,6 +407,8 @@ void CodeGen::genCodeForTreeNode(GenTree* treeNode)
 
 #ifdef TARGET_ARM64
         case GT_XCHG:
+        case GT_XXOR:
+        case GT_XAND:
         case GT_XADD:
             genLockedInstructions(treeNode->AsOp());
             break;

src/coreclr/jit/codegenxarch.cpp

@@ -1706,6 +1706,11 @@ void CodeGen::genCodeForTreeNode(GenTree* treeNode)
             genLockedInstructions(treeNode->AsOp());
             break;
 
+        case GT_XXOR:
+        case GT_XAND:
+            NYI("Interlocked.Or and Interlocked.And aren't implemented for x86 yet.");
+            break;
+
         case GT_MEMORYBARRIER:
         {
             CodeGen::BarrierKind barrierKind =

src/coreclr/jit/decomposelongs.cpp

@@ -247,6 +247,8 @@ GenTree* DecomposeLongs::DecomposeNode(GenTree* tree)
 #endif // FEATURE_SIMD
 
         case GT_LOCKADD:
+        case GT_XXOR:
+        case GT_XAND:
         case GT_XADD:
         case GT_XCHG:
         case GT_CMPXCHG:

src/coreclr/jit/emitarm64.cpp

@@ -6091,6 +6091,8 @@ void emitter::emitIns_R_R_R(
         case INS_ldadda:
         case INS_ldaddal:
         case INS_ldaddl:
+        case INS_ldclral:
+        case INS_ldsetal:
         case INS_swpb:
         case INS_swpab:
         case INS_swpalb:

src/coreclr/jit/gentree.cpp

@@ -5474,7 +5474,7 @@ GenTree* GenTree::gtGetParent(GenTree*** parentChildPtrPtr) const
 
 bool GenTree::OperRequiresAsgFlag()
 {
-    if (OperIs(GT_ASG) || OperIs(GT_XADD, GT_XCHG, GT_LOCKADD, GT_CMPXCHG, GT_MEMORYBARRIER))
+    if (OperIs(GT_ASG) || OperIs(GT_XADD, GT_XXOR, GT_XAND, GT_XCHG, GT_LOCKADD, GT_CMPXCHG, GT_MEMORYBARRIER))
     {
         return true;
     }
@@ -5549,6 +5549,8 @@ bool GenTree::OperIsImplicitIndir() const
     switch (gtOper)
     {
         case GT_LOCKADD:
+        case GT_XXOR:
+        case GT_XAND:
         case GT_XADD:
         case GT_XCHG:
         case GT_CMPXCHG:

src/coreclr/jit/gentree.h

@@ -1519,7 +1519,18 @@ struct GenTree
 
     static bool OperIsAtomicOp(genTreeOps gtOper)
     {
-        return (gtOper == GT_XADD || gtOper == GT_XCHG || gtOper == GT_LOCKADD || gtOper == GT_CMPXCHG);
+        switch (gtOper)
+        {
+            case GT_XADD:
+            case GT_XXOR:
+            case GT_XAND:
+            case GT_XCHG:
+            case GT_LOCKADD:
+            case GT_CMPXCHG:
+                return true;
+            default:
+                return false;
+        }
     }
 
     bool OperIsAtomicOp() const

src/coreclr/jit/gtlist.h

@@ -53,6 +53,8 @@ GTNODE(ARR_LENGTH       , GenTreeArrLen      ,0,(GTK_UNOP|GTK_EXOP))      // arr
 GTNODE(INTRINSIC        , GenTreeIntrinsic   ,0,(GTK_BINOP|GTK_EXOP))     // intrinsics
 
 GTNODE(LOCKADD          , GenTreeOp          ,0,(GTK_BINOP|GTK_NOVALUE))
+GTNODE(XAND             , GenTreeOp          ,0,GTK_BINOP)
+GTNODE(XXOR             , GenTreeOp          ,0,GTK_BINOP)
 GTNODE(XADD             , GenTreeOp          ,0,GTK_BINOP)
 GTNODE(XCHG             , GenTreeOp          ,0,GTK_BINOP)
 GTNODE(CMPXCHG          , GenTreeCmpXchg     ,0,GTK_SPECIAL)

src/coreclr/jit/importer.cpp

@@ -4337,6 +4337,25 @@ GenTree* Compiler::impIntrinsic(GenTree*                newobjThis,
                 break;
             }
 
+#ifdef TARGET_ARM64
+            // Intrinsify Interlocked.Or and Interlocked.And only for arm64-v8.1 (and newer)
+            // TODO-CQ: Implement for XArch (https://github.com/dotnet/runtime/issues/32239).
+            case NI_System_Threading_Interlocked_Or:
+            case NI_System_Threading_Interlocked_And:
+            {
+                if (opts.OptimizationEnabled() && compOpportunisticallyDependsOn(InstructionSet_Atomics))
+                {
+                    assert(sig->numArgs == 2);
+                    GenTree*   op2 = impPopStack().val;
+                    GenTree*   op1 = impPopStack().val;
+                    genTreeOps op  = (ni == NI_System_Threading_Interlocked_Or) ? GT_XXOR : GT_XAND;
+                    retNode        = gtNewOperNode(op, genActualType(callType), op1, op2);
+                    retNode->gtFlags |= GTF_GLOB_REF | GTF_ASG;
+                }
+                break;
+            }
+#endif // TARGET_ARM64
+
 #ifdef FEATURE_HW_INTRINSICS
             case NI_System_Math_FusedMultiplyAdd:
             {
@@ -4845,6 +4864,17 @@ NamedIntrinsic Compiler::lookupNamedIntrinsic(CORINFO_METHOD_HANDLE method)
                 result = NI_System_Threading_Thread_get_ManagedThreadId;
             }
         }
+        else if (strcmp(className, "Interlocked") == 0)
+        {
+            if (strcmp(methodName, "And") == 0)
+            {
+                result = NI_System_Threading_Interlocked_And;
+            }
+            else if (strcmp(methodName, "Or") == 0)
+            {
+                result = NI_System_Threading_Interlocked_Or;
+            }
+        }
     }
 #if defined(TARGET_XARCH) || defined(TARGET_ARM64)
     else if (strcmp(namespaceName, "System.Buffers.Binary") == 0)

src/coreclr/jit/instrsarm64.h

@@ -1194,6 +1194,12 @@ INST1(ldadda,      "ldadda",       LD|ST,  IF_LS_3E,  0xB8A00000)
 INST1(ldaddal,     "ldaddal",      LD|ST,  IF_LS_3E,  0xB8E00000)
                                    //  ldaddal  Rm, Rt, [Xn]        LS_3E  1X111000111mmmmm 000000nnnnnttttt   B8E0 0000   Rm Rt Rn ARMv8.1 LSE Atomics
 
+INST1(ldclral,     "ldclral",      LD|ST,  IF_LS_3E,  0xB8E01000)
+                                   //  ldclral  Rm, Rt, [Xn]        LS_3E  1X111000111mmmmm 000100nnnnnttttt   B8E0 1000   Rm Rt Rn ARMv8.1 LSE Atomics
+
+INST1(ldsetal,     "ldsetal",      LD|ST,  IF_LS_3E,  0xB8E03000)
+                                   //  ldsetal  Rm, Rt, [Xn]        LS_3E  1X111000111mmmmm 001100nnnnnttttt   B8E0 3000   Rm Rt Rn ARMv8.1 LSE Atomics
+
 INST1(ldaddl,      "ldaddl",       LD|ST,  IF_LS_3E,  0xB8600000)
                                    //  ldaddl   Rm, Rt, [Xn]        LS_3E  1X111000011mmmmm 000000nnnnnttttt   B860 0000   Rm Rt Rn ARMv8.1 LSE Atomics
 

src/coreclr/jit/liveness.cpp

@@ -287,6 +287,8 @@ void Compiler::fgPerNodeLocalVarLiveness(GenTree* tree)
 
         // We'll assume these are use-then-defs of memory.
         case GT_LOCKADD:
+        case GT_XXOR:
+        case GT_XAND:
         case GT_XADD:
         case GT_XCHG:
         case GT_CMPXCHG:
@@ -2052,6 +2054,8 @@ void Compiler::fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALAR
                 break;
 
             case GT_LOCKADD:
+            case GT_XXOR:
+            case GT_XAND:
             case GT_XADD:
             case GT_XCHG:
             case GT_CMPXCHG:

src/coreclr/jit/lower.cpp

@@ -318,10 +318,14 @@ GenTree* Lowering::LowerNode(GenTree* node)
             CheckImmedAndMakeContained(node, node->AsCmpXchg()->gtOpComparand);
             break;
 
+        case GT_XXOR:
+        case GT_XAND:
         case GT_XADD:
             CheckImmedAndMakeContained(node, node->AsOp()->gtOp2);
             break;
 #elif defined(TARGET_XARCH)
+        case GT_XXOR:
+        case GT_XAND:
         case GT_XADD:
             if (node->IsUnusedValue())
             {

src/coreclr/jit/lowerarmarch.cpp

@@ -76,6 +76,8 @@ bool Lowering::IsContainableImmed(GenTree* parentNode, GenTree* childNode) const
 #ifdef TARGET_ARM64
             case GT_CMPXCHG:
             case GT_LOCKADD:
+            case GT_XXOR:
+            case GT_XAND:
             case GT_XADD:
                 return comp->compOpportunisticallyDependsOn(InstructionSet_Atomics)
                            ? false

src/coreclr/jit/lsraarm64.cpp

@@ -425,6 +425,8 @@ int LinearScan::BuildNode(GenTree* tree)
         break;
 
         case GT_LOCKADD:
+        case GT_XXOR:
+        case GT_XAND:
         case GT_XADD:
         case GT_XCHG:
         {
@@ -440,6 +442,11 @@ int LinearScan::BuildNode(GenTree* tree)
                     buildInternalIntRegisterDefForNode(tree);
                 }
             }
+            else if (tree->OperIs(GT_XAND))
+            {
+                // for ldclral we need an internal register.
+                buildInternalIntRegisterDefForNode(tree);
+            }
 
             assert(!tree->gtGetOp1()->isContained());
             RefPosition* op1Use = BuildUse(tree->gtGetOp1());

src/coreclr/jit/lsraxarch.cpp

@@ -449,6 +449,8 @@ int LinearScan::BuildNode(GenTree* tree)
         }
         break;
 
+        case GT_XXOR:
+        case GT_XAND:
         case GT_XADD:
         case GT_XCHG:
         {

src/coreclr/jit/namedintrinsiclist.h

@@ -53,6 +53,9 @@ enum NamedIntrinsic : unsigned short
     NI_IsSupported_Dynamic,
     NI_Throw_PlatformNotSupportedException,
 
+    NI_System_Threading_Interlocked_And,
+    NI_System_Threading_Interlocked_Or,
+
 #ifdef FEATURE_HW_INTRINSICS
     NI_HW_INTRINSIC_START,
 #if defined(TARGET_XARCH)

src/coreclr/jit/optimizer.cpp

@@ -7101,7 +7101,7 @@ void Compiler::optHoistLoopBlocks(unsigned loopNum, ArrayStack<BasicBlock*>* blo
                         m_beforeSideEffect = false;
                     }
                 }
-                else if (tree->OperIs(GT_XADD, GT_XCHG, GT_LOCKADD, GT_CMPXCHG, GT_MEMORYBARRIER))
+                else if (tree->OperIs(GT_XADD, GT_XXOR, GT_XAND, GT_XCHG, GT_LOCKADD, GT_CMPXCHG, GT_MEMORYBARRIER))
                 {
                     // If this node is a MEMORYBARRIER or an Atomic operation
                     // then don't hoist and stop any further hoisting after this node
@@ -7915,6 +7915,8 @@ bool Compiler::optComputeLoopSideEffectsOfBlock(BasicBlock* blk)
                         break;
 
                     case GT_LOCKADD:
+                    case GT_XXOR:
+                    case GT_XAND:
                     case GT_XADD:
                     case GT_XCHG:
                     case GT_CMPXCHG:

src/coreclr/jit/valuenum.cpp

@@ -5251,7 +5251,7 @@ void ValueNumStore::vnDumpSimdType(Compiler* comp, VNFuncApp* simdType)
 static UINT8      vnfOpAttribs[VNF_COUNT];
 static genTreeOps genTreeOpsIllegalAsVNFunc[] = {GT_IND, // When we do heap memory.
                                                  GT_NULLCHECK, GT_QMARK, GT_COLON, GT_LOCKADD, GT_XADD, GT_XCHG,
-                                                 GT_CMPXCHG, GT_LCLHEAP, GT_BOX,
+                                                 GT_CMPXCHG, GT_LCLHEAP, GT_BOX, GT_XXOR, GT_XAND,
 
                                                  // These need special semantics:
                                                  GT_COMMA, // == second argument (but with exception(s) from first).
@@ -8232,6 +8232,8 @@ void Compiler::fgValueNumberTree(GenTree* tree)
                         noway_assert("LOCKADD should not appear before lowering");
                         break;
 
+                    case GT_XXOR: // Binop
+                    case GT_XAND: // Binop
                     case GT_XADD: // Binop
                     case GT_XCHG: // Binop
                     {

src/libraries/System.Private.CoreLib/src/System/Threading/Interlocked.cs

@@ -133,6 +133,7 @@ public static ulong Read(ref ulong location) =>
         /// <param name="value">The value to be combined with the integer at <paramref name="location1"/>.</param>
         /// <returns>The original value in <paramref name="location1"/>.</returns>
         /// <exception cref="NullReferenceException">The address of <paramref name="location1"/> is a null pointer.</exception>
+        [Intrinsic]
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static int And(ref int location1, int value)
         {
@@ -164,6 +165,7 @@ public static uint And(ref uint location1, uint value) =>
         /// <param name="value">The value to be combined with the integer at <paramref name="location1"/>.</param>
         /// <returns>The original value in <paramref name="location1"/>.</returns>
         /// <exception cref="NullReferenceException">The address of <paramref name="location1"/> is a null pointer.</exception>
+        [Intrinsic]
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static long And(ref long location1, long value)
         {
@@ -197,6 +199,7 @@ public static ulong And(ref ulong location1, ulong value) =>
         /// <param name="value">The value to be combined with the integer at <paramref name="location1"/>.</param>
         /// <returns>The original value in <paramref name="location1"/>.</returns>
         /// <exception cref="NullReferenceException">The address of <paramref name="location1"/> is a null pointer.</exception>
+        [Intrinsic]
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static int Or(ref int location1, int value)
         {
@@ -228,6 +231,7 @@ public static uint Or(ref uint location1, uint value) =>
         /// <param name="value">The value to be combined with the integer at <paramref name="location1"/>.</param>
         /// <returns>The original value in <paramref name="location1"/>.</returns>
         /// <exception cref="NullReferenceException">The address of <paramref name="location1"/> is a null pointer.</exception>
+        [Intrinsic]
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static long Or(ref long location1, long value)
         {