[RISC-V][LoongArch64] New passing info for floating-point structs #103945
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…carries also exact field sizes and offsets
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
dotnet-policy-service
bot
added
the
community-contribution
RISC-V testing failed on init-build |
tomeksowi
commented
Jun 25, 2024
RISC-V testing failed on init-build |
… FpStruct info calculation
RISC-V testing failed on init-build |
RISC-V test results for starfive-prio1-checked (
|
RISC-V test results for qemu-prio1-checked (
|
…omputeReturnFlags() This fixes JIT/HardwareIntrinsics/General/Vector* tests.
RISC-V testing failed on init-build |
RISC-V test results for qemu-prio1-checked (
|
… because they were identical. Move it to Common\Internal/Runtime because it's no longer exposed in JIT interface.
jkotas
reviewed
Jul 12, 2024
|
|
||
| EE_TO_JIT_TRANSITION_LEAF(); | ||
| // unittests | ||
| static_assert(CORINFO_TYPE_BYTE + 0 * 2 == CORINFO_TYPE_BYTE, ""); |
There was a problem hiding this comment.
IIRC, RISCV and Loongarch calling conventions require proper sign or zero extension. How does it work here given that this does not contain information about the actual sign?
There was a problem hiding this comment.
I thought so too but the integer field of a struct passed according to floating-point calling convention is not type-extended like int primitives.
jkotas
reviewed
Jul 12, 2024
| EE_TO_JIT_TRANSITION_LEAF(); | ||
| // unittests | ||
| static_assert(CORINFO_TYPE_BYTE + 0 * 2 == CORINFO_TYPE_BYTE, ""); | ||
| static_assert(CORINFO_TYPE_BYTE + 1 * 2 == CORINFO_TYPE_SHORT, ""); |
There was a problem hiding this comment.
It looks a bit odd that this is reporting 1-byte as unsigned and everything else as signed. If the sign is unnecessary (see my previous question), should this be always signed or always unsigned?
There was a problem hiding this comment.
I think CORINFO_TYPE_BYTE means signed, otherwise there wouldn't be an UBYTE.
runtime/src/coreclr/inc/corinfo.h
Lines 705 to 706 in 42b2b19
It can stay signed but it doesn't really matter for ABI as answered in the previous question.
jkotas
reviewed
Jul 12, 2024
jkotas
reviewed
Jul 12, 2024
| { | ||
| // Whether the struct should be passed by integer calling convention (cannot be passed by FP calling convention). | ||
| // If true, all other fields of CORINFO_FPSTRUCT_LOWERING are zeroed. | ||
| bool byIntegerCallConv; |
There was a problem hiding this comment.
For my edification, what's the calling convention of struct { object o; float a; }?
Is it passed using this special lowering or using integer call conv?
There was a problem hiding this comment.
According to floating-point calling convention (special lowering). Object is a pointer, which for the convention is simply an integer field.
There was a problem hiding this comment.
Should the Object field from my example be reported as CORINFO_TYPE_CLASS in this struct then? It does not look right that it is reported as int.
There was a problem hiding this comment.
I think we agreed that JIT doesn't need to know IntKind (it was used to tell ordinary integers from GC refs), because it can probe for GC refs from the layout it knows already.
There was a problem hiding this comment.
Currently JIT only needs to know from CORINFO_FPSTRUCT_LOWERING whether the type is floating or integer and the type size. GC info is taken from class layout.
I'd like to merge as is if it's not a problem, this PR blocks #104237 and subsequent changes stemming from #101796. If it turns out that the integer type needs to be more precisely specified for some reason, we can always amend it and FpStructInRegistersInfo.
There was a problem hiding this comment.
Ping. Can this be merged?
There was a problem hiding this comment.
Could you please add a comment that the types returned in loweredElements only represent the size and not the actual type?
(An alternative would be to change loweredElements to uint32_t loweredElementsSizes[] to make it obvious that the API just returns the size and not the actual type.)
There was a problem hiding this comment.
Added comment. The alternative would need another bool array for floating vs. integer and would stray from CORINFO_SWIFT_LOWERING if the need to handle ABI lowering in shared code ever comes up.
jakobbotsch
approved these changes
Jul 15, 2024
JulieLeeMSFT
added
the
needs-author-action
|
@tomeksowi, please take a look at test failure. |
dotnet-policy-service
bot
removed
the
needs-author-action
This was referenced Jul 25, 2024
Closed
The CI failures look unrelated to this change. |
directhex
pushed a commit
that referenced
this pull request
Jul 26, 2024
…03945) * Replace StructFloatFieldInfoFlags with FpStructInRegistersInfo which carries also exact field sizes and offsets * Replace StructFloatFieldInfoFlags with FpStruct::Flags in profiler * Remove FpStructInRegistersInfo::FromOldFlags() * Fix duplicating types in HandleInlineArray * Remove signedness from FpStruct::IntKind because most probably we won't need it * Remove old StructFloatFieldInfoFlags calculating routine * Typo in TARGET_LOONGARCH64 * Remove m_returnedFpFieldOffsets from ArgIterator * Add missing ENREGISTERED_PARAMTYPE_MAXSIZE condition to C# version of FpStruct info calculation * Rename RISCV64PassStructInRegister to match settled casing for RiscV in class names * Update hardcoded flags for float and double in ArgIteratorTemplate::ComputeReturnFlags() This fixes JIT/HardwareIntrinsics/General/Vector* tests. * Fix build on other platforms * Update LoongArch to use FpStructInRegistersInfo * Remove unused old flag masks * LoongArch64 typo Co-authored-by: Qiao Pengcheng <[email protected]> * Missing FpStruct namespace Co-authored-by: Qiao Pengcheng <[email protected]> * Missing FpStruct namespace Co-authored-by: Qiao Pengcheng <[email protected]> * Missing FpStruct namespace Co-authored-by: Qiao Pengcheng <[email protected]> * Use FpStruct namespace everywhere in JIT * JIT review * Update StructFloatFieldInfoFlags description * Revert to hitherto instruction set order as it's not the point of this PR * Unify get{LoongArch,RiscV}64PassFpStructInRegistersInfo JIT interfaces * Use JIT_TO_EE_TRANSITION instead of _LEAF because MethodTable::GetFpStructInRegistersInfo may throw * Remove FpStruct::IntKind, we should have similar info in ClassLayout in JIT * Change JIT interface to return a struct similar to CORINFO_SWIFT_LOWERING to facilitate code unification in the future * Change JIT to use new Swift-like getFpStructLowering * Cache CORINFO_FPSTRUCT_LOWERING * Update LoongArch classifier to use CORINFO_FPSTRUCT_LOWERING * Update StructFloatInfoFlags doc comment on C# * Move StructFloatFieldInfoFlags and FpStructInRegistersInfo out of the JIT interface * Merge LoongArch and RISC-V AOT calculation of FpStructInRegistersInfo because they were identical. Move it to Common\Internal/Runtime because it's no longer exposed in JIT interface. * Don't zero-initialize CORINFO_FPSTRUCT_LOWERING * Add note for CORINFO_FPSTRUCT_LOWERING::loweredElements type --------- Co-authored-by: Qiao Pengcheng <[email protected]>
ilonatommy
added a commit
that referenced
this pull request
Aug 1, 2024
…05110) * Remove support for older LLVM versions, and re-order linker flags We have generally tried to support linking multiple versions of LLVM within our git tree. Every new LLVM version moves symbols around between libraries, and as a result, every new version of LLVM requires different linker flags to build. The command line tool `llvm-config` should tell you the exact flags you need, but it is a problem for us when cross-compiling to rely on this, and as a result, we transcribe the result of various llvm-config outputs directly into Mono's CMakeLists.txt. In an effort to support multiple versions of LLVM, flags common between all supported versions were kept in one place, then the version-specific flags appended afterwards. And this has worked fine for years. However: 1. Whilst we only link with `lld`, it is common for contributors and source-build to link with `gold`, `bfd`, or some other GNU-flavoured linker, where library order is essential 2. The list of common libraries to link has remained unchanged for years, but the symbol intra-dependencies may have changed a long time ago, so common symbol order cannot be assumed to remain valid between LLVM versions This has resulted in a long-standing problem for people using e.g. Debian or Ubuntu or GitHub CodeSpaces, rather than always building with one of our dockerfiles representing our "real" build environment, when targeting platforms which use Mono and link LLVM. * Bumping clang and llvm - make docs less ambiguous. (#105401) * Bump main to RC1 (#105338) * Update SDK to preview 6 (#104696) * Update SDK to preview 6 * Update Shared.csproj Fix `error NU1903: Package 'System.Text.Json' 8.0.0 has a known high severity vulnerability` * Fix with existing version. --------- Co-authored-by: Ilona Tomkowicz <[email protected]> Co-authored-by: Viktor Hofer <[email protected]> * Change DefaultMaximumErrorResponseLength to KB from Byte (#105396) * Change DefaultMaximumErrorResponseLength to KB from Byte * Handle overflow * Review feedback * Fix warning for MakeGenericType annotation mismatch (#104921) Fixes warning code when a generic type whose type parameters have DAM annotations is used with MakeGenericType, over a type that doesn't have matching annotations. The code IL2070 used to mention the 'this' argument. Instead it should have been IL2071 which mentions the generic argument as the cause of the mismatch. Similar for MakeGenericMethod with IL2090 and IL2091. * Set GCStressIncompatible on GenericContext tests (#104686) Co-authored-by: Vladimir Sadov <[email protected]> * Add runtime config parameter to force ijwhost to load assemblies in an isolated context (#105337) * Add support for isolated load context in LoadInMemoryAssemblyInContext by passing -1 as loadContext * Have ijwhost check a runtime config parameter to determine if it should run in an isolated load context * Added test for ijwhost isolated load context runtime config option * [RISC-V] Fix passing float and uint arguments in VM (#105021) * Add tests * Fix passing float and uint arguments in VM * Change test lib name so it doesn't clash with managed DLL on Windows * Fix platform analyzer attribute order for MacCatalyst (#105409) We need to make sure the attribute for MacCatalyst comes _after_ the iOS one due to how MacCatalyst is a superset of iOS: https://learn.microsoft.com/en-us/dotnet/standard/analyzers/platform-compat-analyzer#platform-inclusion This caused an error in aspnetcore in the latest dependency flow because the analyzer thought AesGcm is _only_ supported on MacCatalyst: > error CA1416: (NETCORE_ENGINEERING_TELEMETRY=Build) This call site is reachable on all platforms. 'AesGcm.Decrypt(ReadOnlySpan<byte>, ReadOnlySpan<byte>, ReadOnlySpan<byte>, Span<byte>, ReadOnlySpan<byte>)' is only supported on: 'maccatalyst' 13.0 and later. * Use correct `ExceptionArgument` value in `System.IO.Pipelines` (#105418) * Remove zlib from requirements script and instruction files (#105419) * Remove zlib from requirements instructions * Remove zlib from native requirements installation script * Revert "Remove zlib from requirements script and instruction files (#105419)" (#105449) This reverts commit 3ec6286. * Ensure that WaitForPendingFinalizers has seen the expected Full GC count (#105289) * Ensure that WaitForPendingFinalizers has seen the expected Full GC * NativeAOT and some renames * a testcase * make the test not unsafe and make OuterLoop * Use unsigned math when comparing collection ticks * cast the diff to int when comparing gc ticks * Migrate to zlib-ng, part 3: Remove zlib and zlib-intel source code and license mentions (second attempt) (#105371) * Remove zlib/ * Remove zlib-intel/ * Remove third party notice * Remove patches * Remove version txts * Remove cgmanifest.json entries * Remove installer third party notice * Update docs --------- Co-authored-by: Jan Kotas <[email protected]> * [browser] Trigger relink on `EmccMaximumHeapSize` change (#105027) * Edit test + trigger relink. * Remove logging to speed up the test + decrease loop runs to prevent "Browser has been disconnected" error. * Feedback - properties are not bool-only anymore. * Fix: workload needed when heap size set. --------- Co-authored-by: Larry Ewing <[email protected]> * Delete erroneous Socket test (#105448) * Pull the python dependency from the EmsdkVersion where possible (#105437) * Use `BinaryPrimitives` more in the ILCompiler (#105404) * JIT: use VNVisitReachingVNs in IsVNNeverNegative (#105197) * Set xunit env var to not print output for passing tests (#105392) * Set xunit env var to not print output for passing tests Fixes #103445 * Update xunit.console.targets * Update xunit.console.targets * Some more automated C# modernization in corelib (#105151) * Fix IDE0056 on corelib (indexing can be simplified) * Fix IDE (null check can be simplified) * Fix IDE0078 (use pattern matching) * Fix IDE0019 (use pattern matching) * Fix IDE0066 (use switch expression) * Fix IDE0250 (struct can be made readonly) * Fix nullability warning and address PR feedback * Address PR feedback and revert a downlevel change * Wrap any `?? throw new`s that go beyond 120 characters * Fix ECMA 355 Partition download links (#105454) On https://github.com/dotnet/runtime/blob/main/docs/project/dotnet-standards.md the Partition with Notes download links using HTTP protocol fail to download in Chrome: >Mixed Content: The site at 'https://github.com/' was loaded over a secure connection, but the file at 'https://download.microsoft.com/download/7/3/3/733AD403-90B2-4064-A81E-01035A7FE13C/MS%20Partition%20I.pdf' was redirected through an insecure connection. This file should be served over HTTPS. See https://blog.chromium.org/2020/02/protecting-users-from-insecure.html for more details. (It might be reasonble for somebody else to followup fixing all domains on all pages with regex `\(http://([.a-z0-9-]+)` replacing with `(https://$1`, but I didn't test each blog site supports HTTPS.) * Add swiftcall signature check for `mono_class_try_get_swift_error_class` (#105408) * Add signature check for swiftcall * Handle null values for swift_error_ptr * Enable NativeAOT runtime tests on MacCatalyst (#102882) This PR updates the CLRTest.Execute.Bash.targets file to set the apple run command for MacCatalyst. The command apple just-run used on Apple mobile is not permitted, and apple test requires the a test runner. Additionally, it is necessary to locate Info.plist in the Contents/ directory and the binary in Contents/MacOS/ within the bundle. --------- Co-authored-by: Ivan Povazan <[email protected]> * Resolving an antigen failure (#105260) * Resolving an antigen failure * Fix method accessibility so xunit doesn't complain * Support field access on GetType() of T constrained to be Enum (#105351) Adds trimming support for instance.GetType().GetFields(), where instance is a variable of type `T` that is constrained to System.Enum. This includes a change to have ILLink's TypeProxy track a TypeReference instead of TypeDefinition, which was necessary to allow TypeProxy to represent a generic parameter. Note that this only supports the specific case where `GetType()` is called on a variable of type `T` that is constrained to `Enum`. A variable of type `Enum` is not supported, so the following will still warn: ```csharp static void M(Enum v) { v.GetType().GetFields(); } ``` * Update mono to support shuffle for constant inputs (#105299) * Support mono creating xconst in a few more places * Update mono to support shuffle for constant inputs * Ensure that arm64 also accelerates shuffle for non-constant inputs * Ensure OP_XZERO and OP_XONES are recognized as being constant * Ensure shuffle creates a correct instruction when the fsig doesn't match the necessary parameter count * Ensure that getting the index for floating-point shuffle is possible * Ensure the right class handle is passed down to LLVM so overload resolution can function * Make sure we update the original xconst if we mutate it * Return a new constant and instead of mutating the existing one * Insert relevant xcast nodes * Add some asserts around the ecount * Ensure we get the right element type * Ensure we don't create nodes unnecessarily for create_elementwise * Ensure that create_elementwise still works for other vector sizes * Ensure indentation of switch cases is correct for Mono * Make TooDeepJsonDocument test more consistent across platforms (#105445) * Make TooDeepJsonDocument test more consistent across platforms Run the test on a thread with as consistent a stack size as possible so that we don't inadvertently succeed due to having a really large stack. * Disable test on mono * Update the TypeLib embedding and add comments on API use (#105416) There is an undocumented semantic of Win32 Resource APIs. The missing semantic is that all resource type/name strings are transparently converted to uppercase when calling any of the Win32 Resource APIs. We don't want to apply this undocumented semantic to the reader/writer API so we document it instead. We are avoiding applying the behavior since ReadyToRun scenarios are designed to be a byte for byte copy of the resource, including name as it was written by other tooling. * Update docs for ByRefLike with generics for work in .NET 10 (#103318) Co-authored-by: Jan Kotas <[email protected]> * Fix double printing in StressLog and simplify stresslog macros (#105420) * Add support for nested types in the `corelib.h` parser for rooting descriptors (#105432) * Fixed for assertion failure due to not checking if we are processing Eph samples (#105164) * [PERF] Use correct python executable on windows in venv (#105451) * Fix up Fuzzlyn CI scripts for new hardware intrinsics support (#105470) 1) Strip out the extensions in the seed name when using it for file/directory names, since the list of extensions is quite long 2) Limit the number of unreduced/uncategorized example seeds we show * Try to re-enable DeepEquals_TooDeepJsonDocument_ThrowsInsufficientExecutionStackException test on mono (#105509) * zlib-ng: avoid suppressing WD4242 and WD4244 (#105433) WD4242 and WD4244 are compiler warnings that should not be suppressed because the warn about possible loss of data. WD4242 shows up in zlib-ng/arch/*/slide_hash*.c files and comes from the arguments passed to the slide_hash_chain method. WD4244 happens in Windows when building in Debug configuration, in various zlib-ng/deflate*.c files, and comes from the arguments passed to the check_match method. Fixed by: - Adding asserts to verify the values are below the maximum allowed for their type. - Casting them the proper type before passing them as arguments to their methods. - Removing the WD suppressions, which unfortunately also propagated to other unrelated cmake files. - Fixed a similar loss of data error in an unrelated mono file where the warning suppression was propagated due to this. * Delete outdated comments (#105519) * Arm64/Sve: Add FFR register liveness tracking (#105348) * Add tracking of FFR register somewhat workable code cleanup Remove FFR Add all the GetFfr* wip Work with MskCns() model Use physReg approach Remove commented prototypes working Remove bunch of unnecessary code Remove SpecialImport from GetFFR/SetFFR/LoadFirstFaulting some more code cleanup some fixup * Change condition for PhysReg * jit format * review feedback * unspill for LoadVectorFirstFaulting as well * Use the right opReg * skip spilling tracking * review feedback * Use non-existent REG_FFR * Do not reload from FFR for GetFfr() * review feedback * Make just GrabTemp * fix build and formatting * missed another build failure for arm * Fix throwing exception when calling RunClassConstructor on a generic type with a static constructor (#105513) * Fix throwing exception when calling RunClassConstructor on a generic type with a static constructor #99183 seems to have done away with assuming that a generic type's static constructor was always "initialized". As a result, if you call RunClassConstructor on it, the runtime would throw an exception. Fixes #103891 * Apply suggestions from code review --------- Co-authored-by: Jan Kotas <[email protected]> * [RISC-V][LoongArch64] New passing info for floating-point structs (#103945) * Replace StructFloatFieldInfoFlags with FpStructInRegistersInfo which carries also exact field sizes and offsets * Replace StructFloatFieldInfoFlags with FpStruct::Flags in profiler * Remove FpStructInRegistersInfo::FromOldFlags() * Fix duplicating types in HandleInlineArray * Remove signedness from FpStruct::IntKind because most probably we won't need it * Remove old StructFloatFieldInfoFlags calculating routine * Typo in TARGET_LOONGARCH64 * Remove m_returnedFpFieldOffsets from ArgIterator * Add missing ENREGISTERED_PARAMTYPE_MAXSIZE condition to C# version of FpStruct info calculation * Rename RISCV64PassStructInRegister to match settled casing for RiscV in class names * Update hardcoded flags for float and double in ArgIteratorTemplate::ComputeReturnFlags() This fixes JIT/HardwareIntrinsics/General/Vector* tests. * Fix build on other platforms * Update LoongArch to use FpStructInRegistersInfo * Remove unused old flag masks * LoongArch64 typo Co-authored-by: Qiao Pengcheng <[email protected]> * Missing FpStruct namespace Co-authored-by: Qiao Pengcheng <[email protected]> * Missing FpStruct namespace Co-authored-by: Qiao Pengcheng <[email protected]> * Missing FpStruct namespace Co-authored-by: Qiao Pengcheng <[email protected]> * Use FpStruct namespace everywhere in JIT * JIT review * Update StructFloatFieldInfoFlags description * Revert to hitherto instruction set order as it's not the point of this PR * Unify get{LoongArch,RiscV}64PassFpStructInRegistersInfo JIT interfaces * Use JIT_TO_EE_TRANSITION instead of _LEAF because MethodTable::GetFpStructInRegistersInfo may throw * Remove FpStruct::IntKind, we should have similar info in ClassLayout in JIT * Change JIT interface to return a struct similar to CORINFO_SWIFT_LOWERING to facilitate code unification in the future * Change JIT to use new Swift-like getFpStructLowering * Cache CORINFO_FPSTRUCT_LOWERING * Update LoongArch classifier to use CORINFO_FPSTRUCT_LOWERING * Update StructFloatInfoFlags doc comment on C# * Move StructFloatFieldInfoFlags and FpStructInRegistersInfo out of the JIT interface * Merge LoongArch and RISC-V AOT calculation of FpStructInRegistersInfo because they were identical. Move it to Common\Internal/Runtime because it's no longer exposed in JIT interface. * Don't zero-initialize CORINFO_FPSTRUCT_LOWERING * Add note for CORINFO_FPSTRUCT_LOWERING::loweredElements type --------- Co-authored-by: Qiao Pengcheng <[email protected]> * Ensure we don't reuse temps when calling fgMorphArgs on LIR nodes (#105508) * Ensure constant evaluation of shifts on xarch broadcast the operand to the correct size (#105487) * Ensure constant evaluation of shifts on xarch broadcast the operand to the correct size * Ensure we don't try to execute AVX2 code on unsupported platforms * Use ConcurrentDictionary in runtimecounters test (#105520) * Use ConcurrentDictionary in runtimecounters test Fixes #105443 * Fix build break * Fix issue 98506 - Excessive exceptions generated in StackTraceSymbols (#105530) * Fix issue 98506 - Excessive exceptions generated in StackTraceSymbols * Code review feedback * Clean up some usages of LowLevelList<T> (#105407) * Fix ShuffleThunk cache heap (#105480) There was a problem with using heap from the related LoaderAllocator for shuffle thunk cache heap. I have tested it again and it seems that the issue is gone. So I am removing the workaround, making the cache use LoaderAllocator local heap. Close #55697 * [browser] Fix computing destination sub path and publish extension target path in Wasm SDK (#105458) * Bump flags to LLVM 19, not 16 --------- Co-authored-by: Larry Ewing <[email protected]> Co-authored-by: Alexander Köplinger <[email protected]> Co-authored-by: Ilona Tomkowicz <[email protected]> Co-authored-by: Carlos Sánchez López <[email protected]> Co-authored-by: Adeel Mujahid <[email protected]> Co-authored-by: Viktor Hofer <[email protected]> Co-authored-by: Ahmet Ibrahim Aksoy <[email protected]> Co-authored-by: Sven Boemer <[email protected]> Co-authored-by: Steve Pfister <[email protected]> Co-authored-by: Vladimir Sadov <[email protected]> Co-authored-by: Mike Oliphant <[email protected]> Co-authored-by: Tomasz Sowiński <[email protected]> Co-authored-by: xtqqczze <[email protected]> Co-authored-by: Jan Kotas <[email protected]> Co-authored-by: Stephen Toub <[email protected]> Co-authored-by: Paulus Pärssinen <[email protected]> Co-authored-by: Egor Bogatov <[email protected]> Co-authored-by: Carl Walsh <[email protected]> Co-authored-by: Milos Kotlar <[email protected]> Co-authored-by: Ivan Povazan <[email protected]> Co-authored-by: Tanner Gooding <[email protected]> Co-authored-by: Aaron Robinson <[email protected]> Co-authored-by: Jeremy Koritzinsky <[email protected]> Co-authored-by: Mukund Raghav Sharma (Moko) <[email protected]> Co-authored-by: Cameron Aavik <[email protected]> Co-authored-by: Jakob Botsch Nielsen <[email protected]> Co-authored-by: Kunal Pathak <[email protected]> Co-authored-by: Qiao Pengcheng <[email protected]> Co-authored-by: Mike McLaughlin <[email protected]> Co-authored-by: Huo Yaoyuan <[email protected]> Co-authored-by: Jan Vorlicek <[email protected]> Co-authored-by: Marek Fišera <[email protected]>
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Small structs containing one or two fields, at least one of them floating-point, can be passed (or returned) according to hardware floating-point calling convention: each field occupying one register. The existing implementation worked only for the narrow case when the fields were naturally placed. However, the ABIs on both platforms when enregistering fields disregard placement hints such as manual alignment, packing attributes, or padding with empty structs (when they are sized 1 byte like in C++ or .NET). This means additional information such as field offsets and sizes needs to be passed wherever registers<->memory copying of such structs happens.
This change introduces a new container (FpStructInRegistersInfo) for the flags and additional passing info, and some general clean-up around it, improved flag naming, adds logging, etc. The JIT interface uses CORINFO_FPSTRUCT_LOWERING -- a structure similar to CORINFO_SWIFT_LOWERING to facilitate code reuse in the future.
Functionally this PR should be a no-op, the additional info will be utilized in subsequent PRs so the functional changes don't get drowned out in the diffs.
Stems from #101796, part of #84834, cc @dotnet/samsung