Fix incorrect SIMD temp allocation for Vector256 with AVX2 disabled #58820

BruceForstall · 2021-09-08T17:46:35Z

The NI_Vector256_GetElement intrinsic, in some situations, requires
a stack temporary. With AVX2 disabled, this temporary was getting
allocated as a TYP_SIMD16 instead of a TYP_SIMD32, leading to overwriting
the local variable.

Add a type argument to the temp variable allocation, and allocate the
temp as the largest sized type required by any use.

Fixes #58295

The NI_Vector256_GetElement intrinsic, in some situations, requires a stack temporary. With AVX2 disabled, this temporary was getting allocated as a TYP_SIMD16 instead of a TYP_SIMD32, leading to overwriting the local variable. Add a type argument to the temp variable allocation, and allocate the temp as the largest sized type required by any use. Fixes dotnet#58295

ghost · 2021-09-08T17:46:41Z

Tagging subscribers to this area: @JulieLeeMSFT
See info in area-owners.md if you want to be subscribed.

Issue Details

The NI_Vector256_GetElement intrinsic, in some situations, requires
a stack temporary. With AVX2 disabled, this temporary was getting
allocated as a TYP_SIMD16 instead of a TYP_SIMD32, leading to overwriting
the local variable.

Add a type argument to the temp variable allocation, and allocate the
temp as the largest sized type required by any use.

Fixes #58295

Author:	BruceForstall
Assignees:	-
Labels:	`area-CodeGen-coreclr`
Milestone:	-

BruceForstall · 2021-09-08T17:56:01Z

Windows x64 diffs:


Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 222
Total bytes of diff: 190
Total bytes of delta: -32 (-14.41% of base)
Total relative delta: -0.29
    diff is an improvement.
    relative diff is an improvement.

Detail diffs



Top file improvements (bytes):
         -16 : 207372.dasm (-14.29% of base)
         -16 : 207371.dasm (-14.55% of base)

2 total files with Code Size differences (2 improved, 0 regressed), 0 unchanged.

Top method improvements (bytes):
         -16 (-14.29% of base) : 207372.dasm - GitHub_22815.Program:test128(byte):bool
         -16 (-14.55% of base) : 207371.dasm - GitHub_22815.Program:test128(ubyte):bool

Top method improvements (percentages):
         -16 (-14.55% of base) : 207371.dasm - GitHub_22815.Program:test128(ubyte):bool
         -16 (-14.29% of base) : 207372.dasm - GitHub_22815.Program:test128(byte):bool

2 total methods with Code Size differences (2 improved, 0 regressed), 0 unchanged.

Linux x64:


Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 978
Total bytes of diff: 931
Total bytes of delta: -47 (-4.81% of base)
Total relative delta: -0.33
    diff is an improvement.
    relative diff is an improvement.

Detail diffs



Top file improvements (bytes):
         -16 : 210097.dasm (-14.29% of base)
         -16 : 210096.dasm (-14.29% of base)
         -10 : 198748.dasm (-2.72% of base)
          -5 : 195198.dasm (-1.30% of base)

4 total files with Code Size differences (4 improved, 0 regressed), 0 unchanged.

Top method improvements (bytes):
         -16 (-14.29% of base) : 210097.dasm - GitHub_22815.Program:test128(byte):bool
         -16 (-14.29% of base) : 210096.dasm - GitHub_22815.Program:test128(ubyte):bool
         -10 (-2.72% of base) : 198748.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector128`1[Byte],int):this
          -5 (-1.30% of base) : 195198.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector128`1[Byte],int):this

Top method improvements (percentages):
         -16 (-14.29% of base) : 210097.dasm - GitHub_22815.Program:test128(byte):bool
         -16 (-14.29% of base) : 210096.dasm - GitHub_22815.Program:test128(ubyte):bool
         -10 (-2.72% of base) : 198748.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector128`1[Byte],int):this
          -5 (-1.30% of base) : 195198.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector128`1[Byte],int):this

4 total methods with Code Size differences (4 improved, 0 regressed), 0 unchanged.

These diffs are due to allocating a SIMD16 instead of a SIMD32, so the frame shrinks. We don't have any SPMI collections where AVX2 is disabled.

BruceForstall · 2021-09-08T17:58:02Z

@tannergooding @dotnet/jit-contrib PTAL

@tannergooding We could alternately go with your fix proposed in #58295 (comment), although this one has the minor benefit of shrinking the temp when the larger temp is not needed. Thoughts?

tannergooding · 2021-09-08T18:00:29Z

Thoughts?

I think this fix is fine. I thought we were using getSIMDInitTempVarNum in more places and this would be a more complex change to make work like you have.

src/coreclr/jit/lsraarm64.cpp

tannergooding

LGTM.

BruceForstall · 2021-09-08T19:19:42Z

@tannergooding With your suggested change (included here), there are a few SPMI arm64 asm diffs similar to the x64 ones noted above: TYP_SIMD16 temps shrunk to TYP_SIMD8, thus shrinking frame sizes.

coreclr_tests.pmi.windows.arm64.checked.mch:


Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 2184
Total bytes of diff: 2176
Total bytes of delta: -8 (-0.37% of base)
Total relative delta: -0.02
    diff is an improvement.
    relative diff is an improvement.

Detail diffs



Top file regressions (bytes):
           4 : 196147.dasm (0.54% of base)

Top file improvements (bytes):
          -4 : 196012.dasm (-0.56% of base)
          -4 : 196094.dasm (-1.12% of base)
          -4 : 196182.dasm (-1.05% of base)

4 total files with Code Size differences (3 improved, 1 regressed), 0 unchanged.

Top method regressions (bytes):
           4 ( 0.54% of base) : 196147.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector64`1[Byte],int):this

Top method improvements (bytes):
          -4 (-0.56% of base) : 196012.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector64`1[Byte],int):this
          -4 (-1.12% of base) : 196094.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector64`1[Byte],int):this
          -4 (-1.05% of base) : 196182.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector64`1[Byte],int):this

Top method regressions (percentages):
           4 ( 0.54% of base) : 196147.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector64`1[Byte],int):this

Top method improvements (percentages):
          -4 (-1.12% of base) : 196094.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector64`1[Byte],int):this
          -4 (-1.05% of base) : 196182.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector64`1[Byte],int):this
          -4 (-0.56% of base) : 196012.dasm - HVATests`1[Byte][System.Byte]:checkValues(System.String,System.Runtime.Intrinsics.Vector64`1[Byte],int):this

4 total methods with Code Size differences (3 improved, 1 regressed), 0 unchanged.

libraries.crossgen2.windows.arm64.checked.mch:


Summary of Code Size diffs:
(Lower is better)

Total bytes of base: 1124
Total bytes of diff: 1124
Total bytes of delta: 0 (0.00% of base)

Detail diffs



0 total files with Code Size differences (0 improved, 0 regressed), 3 unchanged.

0 total methods with Code Size differences (0 improved, 0 regressed), 3 unchanged.

tannergooding · 2021-09-08T19:35:09Z

Should this also add an outerloop leg that will catch this kind of failure in the future?

BruceForstall · 2021-09-08T19:37:37Z

Should this also add an outerloop leg that will catch this kind of failure in the future?

I'm testing a PR (#58822) to add JitStress=1 and JitStress=2 to the "runtime-coreclr jitstress-isas-x86" pipeline. We can discuss it over there.

BruceForstall · 2021-09-09T04:36:53Z

/backport to release/6.0

github-actions · 2021-09-09T04:37:08Z

Started backporting to release/6.0: https://github.com/dotnet/runtime/actions/runs/1215944929

ghost added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Sep 8, 2021

BruceForstall requested a review from tannergooding September 8, 2021 17:56

tannergooding reviewed Sep 8, 2021

View reviewed changes

src/coreclr/jit/lsraarm64.cpp Outdated Show resolved Hide resolved

tannergooding approved these changes Sep 8, 2021

View reviewed changes

Code review change: improve arm64 SIMD temp creation type

2dbef46

BruceForstall merged commit 73b29b5 into dotnet:main Sep 9, 2021

github-actions bot mentioned this pull request Sep 9, 2021

[release/6.0] Fix incorrect SIMD temp allocation for Vector256 with AVX2 disabled #58850

Merged

BruceForstall mentioned this pull request Sep 9, 2021

Add JitStress=1/2 runs to jitstress-isas-x86 pipeline #58822

Merged

ghost locked as resolved and limited conversation to collaborators Oct 9, 2021

BruceForstall deleted the Fix58295 branch December 28, 2022 01:13

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Fix incorrect SIMD temp allocation for Vector256 with AVX2 disabled #58820

Fix incorrect SIMD temp allocation for Vector256 with AVX2 disabled #58820

Uh oh!

BruceForstall commented Sep 8, 2021

Uh oh!

ghost commented Sep 8, 2021

Uh oh!

BruceForstall commented Sep 8, 2021

Uh oh!

BruceForstall commented Sep 8, 2021

Uh oh!

tannergooding commented Sep 8, 2021

Uh oh!

Uh oh!

tannergooding left a comment

Uh oh!

BruceForstall commented Sep 8, 2021

Uh oh!

tannergooding commented Sep 8, 2021

Uh oh!

BruceForstall commented Sep 8, 2021

Uh oh!

BruceForstall commented Sep 9, 2021

Uh oh!

github-actions bot commented Sep 9, 2021

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Fix incorrect SIMD temp allocation for Vector256 with AVX2 disabled #58820

Fix incorrect SIMD temp allocation for Vector256 with AVX2 disabled #58820

Uh oh!

Conversation

BruceForstall commented Sep 8, 2021

Uh oh!

ghost commented Sep 8, 2021

Uh oh!

BruceForstall commented Sep 8, 2021

Uh oh!

BruceForstall commented Sep 8, 2021

Uh oh!

tannergooding commented Sep 8, 2021

Uh oh!

Uh oh!

tannergooding left a comment

Choose a reason for hiding this comment

Uh oh!

BruceForstall commented Sep 8, 2021

coreclr_tests.pmi.windows.arm64.checked.mch:

libraries.crossgen2.windows.arm64.checked.mch:

Uh oh!

tannergooding commented Sep 8, 2021

Uh oh!

BruceForstall commented Sep 8, 2021

Uh oh!

BruceForstall commented Sep 9, 2021

Uh oh!

github-actions bot commented Sep 9, 2021

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants