ARM64-SVE: gathervector extends #103370

a74nh · 2024-06-12T19:32:54Z

Does not include GatherVectorWithByteOffsets.

There were handful of reviewed APIs that did not make sense, as they were all zero/sign extending from 32bits to 32bits.
eg:
Vector<int> GatherVectorUInt32ZeroExtend(Vector<int> mask, uint* address, Vector<int> indices);
For these, there was no matching C++ API and the SVE instruction it pointed to was invalid (as you can't extend 32bits to 32bits). Not quite sure how they got into the API. I did not include any of these in this PR.

As per #103159, all 32bit address APIs have been commented out.

dotnet-issue-labeler · 2024-06-12T19:33:00Z

Note regarding the new-api-needs-documentation label:

This serves as a reminder for when your PR is modifying a ref *.cs file and adding/modifying public APIs, please make sure the API implementation in the src *.cs file is documented with triple slash comments, so the PR reviewers can sign off that change.

dotnet-issue-labeler · 2024-06-12T19:33:01Z

Note regarding the new-api-needs-documentation label:

This serves as a reminder for when your PR is modifying a ref *.cs file and adding/modifying public APIs, please make sure the API implementation in the src *.cs file is documented with triple slash comments, so the PR reviewers can sign off that change.

dotnet-policy-service · 2024-06-12T19:33:24Z

Tagging subscribers to this area: @dotnet/area-system-runtime-intrinsics
See info in area-owners.md if you want to be subscribed.

a74nh · 2024-06-12T19:36:42Z

@dotnet/arm64-contrib @kunalspathak : This one is ready.

a74nh · 2024-06-13T10:20:57Z

stress test results. Failures look like predicate issues and the same as those for gathervector()

kunalspathak

LGTM. Thanks for writing reusable test templates :) One comment about the memory load flag.

kunalspathak · 2024-06-13T13:32:51Z

src/coreclr/jit/hwintrinsiclistarm64sve.h

@@ -76,6 +76,16 @@ HARDWARE_INTRINSIC(Sve,           FusedMultiplySubtract,
 HARDWARE_INTRINSIC(Sve,           FusedMultiplySubtractBySelectedScalar,                            -1,      4,      true,  {INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_sve_fmls,       INS_sve_fmls},    HW_Category_SIMDByIndexedElement,  HW_Flag_Scalable|HW_Flag_HasImmediateOperand|HW_Flag_HasRMWSemantics|HW_Flag_FmaIntrinsic|HW_Flag_LowVectorOperation)
 HARDWARE_INTRINSIC(Sve,           FusedMultiplySubtractNegated,                                     -1,     -1,      false, {INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_sve_fnmls,      INS_sve_fnmls},   HW_Category_SIMD,                  HW_Flag_Scalable|HW_Flag_EmbeddedMaskedOperation|HW_Flag_HasRMWSemantics|HW_Flag_LowMaskedOperation|HW_Flag_FmaIntrinsic|HW_Flag_SpecialCodeGen)
 HARDWARE_INTRINSIC(Sve,           GatherVector,                                                     -1,     -1,      false, {INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_sve_ld1w,       INS_sve_ld1w,       INS_sve_ld1d,       INS_sve_ld1d,       INS_sve_ld1w,       INS_sve_ld1d},    HW_Category_SIMD,                  HW_Flag_Scalable|HW_Flag_SpecialCodeGen|HW_Flag_ExplicitMaskedOperation|HW_Flag_LowMaskedOperation)
+HARDWARE_INTRINSIC(Sve,           GatherVectorByteZeroExtend,                                       -1,     -1,      false, {INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_sve_ld1b,       INS_sve_ld1b,       INS_sve_ld1b,       INS_sve_ld1b,       INS_invalid,        INS_invalid},     HW_Category_SIMD,                  HW_Flag_Scalable|HW_Flag_SpecialCodeGen|HW_Flag_ExplicitMaskedOperation|HW_Flag_LowMaskedOperation)


I didn't realize earlier but all the load/gatherload need to be marked as HW_Category_MemoryLoad

I didn't realize earlier but all the load/gatherload need to be marked as HW_Category_MemoryLoad

That's going to require some debugging, as they don't fit a normal style of a load API (I get an assert failure).

hwintrinsic.cpp has a if ((category == HW_Category_MemoryLoad) && op1->OperIs(GT_CAST)) check. This is only checked for memory loads with a single argument. All other types of memory loads are unchecked.

I moved this code so that it is checked for all HW_Category_MemoryLoad nodes. I reused OperIsMemoryLoad() but that means it needs doing after the ret node is created.

Inside OperIsMemoryLoad() I'm not sure what to do for the cases where the load is a vector of addresses. My thought was that we should not set paddr in these cases.

hwintrinsic.cpp has a if ((category == HW_Category_MemoryLoad) && op1->OperIs(GT_CAST)) check. This is only checked for memory loads with a single argument. All other types of memory loads are unchecked.

can we leave it as is then and not touch it?

can we leave it as is then and not touch it?

My assumption was this was a bug. The check needs to be added for all loads where the address is not in arg1.

What happens if the address is a GT_CAST and it is not removed?

I've removed the GT_CAST changes for now. I'll raise a bug so we can double check it later. PR is all still working.

kunalspathak · 2024-06-14T20:12:37Z

/ba-g the failures are known and test time out.

kunalspathak · 2024-06-17T22:46:23Z

src/coreclr/jit/gentree.cpp

+                if (varTypeIsSIMD(addr->gtType))
+                {
+                    // The address is a vector of addresses.
+                    // Return true, but do not set pAddr.


@a74nh - do you remember why we do not set pAddr here? We get AV in value numbering because of that.

Sve_GatherVector_Bases_double_ulong was failing for me with below call stack:

> clrjit_universal_arm64_x64.dll!Compiler::fgValueNumberAddExceptionSetForIndirection(GenTree * tree, GenTree * baseAddr) Line 13618 C++ clrjit_universal_arm64_x64.dll!Compiler::fgValueNumberHWIntrinsic(GenTreeHWIntrinsic * tree) Line 12461 C++ clrjit_universal_arm64_x64.dll!Compiler::fgValueNumberTree(GenTree * tree) Line 12056 C++ clrjit_universal_arm64_x64.dll!Compiler::fgValueNumberBlock(BasicBlock * blk) Line 10554 C++ clrjit_universal_arm64_x64.dll!Compiler::fgValueNumber() Line 10340 C++ clrjit_universal_arm64_x64.dll!CompilerPhaseWithStatus::DoPhase() Line 132 C++

Opened #103606

dotnet-issue-labeler bot added area-System.Runtime.Intrinsics new-api-needs-documentation labels Jun 12, 2024

dotnet-policy-service bot added the community-contribution Indicates that the PR has been added by a community member label Jun 12, 2024

ARM64-SVE: gathervector extends

6b722db

a74nh force-pushed the gatherload_extend_github branch from 7dd8a48 to 6b722db Compare June 12, 2024 19:35

a74nh marked this pull request as ready for review June 12, 2024 19:36

kunalspathak added the arm-sve Work related to arm64 SVE/SVE2 support label Jun 12, 2024

This was referenced Jun 12, 2024

System.IO.Net5Compat.Tests and System.IO.Tests suddenly exiting with error 137 #100558

Open

SIGKILL (OOM?) while running LibraryImportGenerator.Tests w/o actionable log messages or artifacts dotnet/dnceng#2496

Open

Comment out 32bit address APIs

763c9b4

a74nh added 2 commits June 13, 2024 11:29

Replace triple quotes with doubles

f9c0a06

restore commented API

3478b1c

kunalspathak mentioned this pull request Jun 13, 2024

Arm64: Implement SVE APIs #99957

Closed

kunalspathak approved these changes Jun 13, 2024

View reviewed changes

a74nh added 2 commits June 14, 2024 10:03

Set HW_Category_MemoryLoad for all gatherloads

3de7737

Fix cast checking for all load types

cf06dc0

This was referenced Jun 14, 2024

CI timeout in System.Data.Tests.DataSetTest2.Bug537229_BinFormatSerializer_Test #103350

Open

SslStreamTlsResumeTests.ClientDisableTlsResume_Succeeds failed in CI #103449

Open

Remove GT_CAST changes

64783ea

This was referenced Jun 14, 2024

GC\Regressions\v2.0-beta2\452950\452950\452950.cmd failing on Mono minijit Windows x64 #99729

Open

at GC_Regressions._v2_0_beta2_452950_452950_452950_._v2_0_beta2_452950_452950_452950_cmd() #100174

Closed

Dead lettering tests #101524

Closed

build-analysis bot mentioned this pull request Jun 14, 2024

x64 test failure: DeserializeStoredObjects strings differ. #103367

Closed

kunalspathak merged commit 0d60428 into dotnet:main Jun 14, 2024
160 of 167 checks passed

a74nh deleted the gatherload_extend_github branch June 17, 2024 09:57

kunalspathak reviewed Jun 17, 2024

View reviewed changes

github-actions bot locked and limited conversation to collaborators Jul 18, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

ARM64-SVE: gathervector extends #103370

ARM64-SVE: gathervector extends #103370

a74nh commented Jun 12, 2024

dotnet-issue-labeler bot commented Jun 12, 2024

dotnet-issue-labeler bot commented Jun 12, 2024

dotnet-policy-service bot commented Jun 12, 2024

a74nh commented Jun 12, 2024

a74nh commented Jun 13, 2024

kunalspathak left a comment

kunalspathak Jun 13, 2024

a74nh Jun 13, 2024

a74nh Jun 14, 2024

kunalspathak Jun 14, 2024

a74nh Jun 14, 2024

a74nh Jun 14, 2024

kunalspathak commented Jun 14, 2024

kunalspathak Jun 17, 2024

kunalspathak Jun 17, 2024

ARM64-SVE: gathervector extends #103370

ARM64-SVE: gathervector extends #103370

Conversation

a74nh commented Jun 12, 2024

dotnet-issue-labeler bot commented Jun 12, 2024

dotnet-issue-labeler bot commented Jun 12, 2024

dotnet-policy-service bot commented Jun 12, 2024

a74nh commented Jun 12, 2024

a74nh commented Jun 13, 2024

kunalspathak left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kunalspathak commented Jun 14, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment