Updating Vector<T> to support nint and nuint #50832

tannergooding · 2021-04-07T06:54:25Z

This resolves #36160

In order for this to work, GenTreeJitIntrinsic was updated to track the baseType as a CorInfoType rather than as a var_types (as discussed in #36160 (comment)). There is additionally a new JitType2PreciseVarType method which ensures that a CorInfoType is converted to a var_types in a way that preserves the sign (JITtype2varType will convert CORINFO_TYPE_UINT to CORINFO_TYPE_INT otherwise).

I ran jit-diff --pmi for --frameworks, --benchmarks, and --tests after the second commit to ensure there were no diffs due to the above changes. I then added the three latter commits supporting Vector<nint> and Vector<nuint>.

Edit: The vast majority of the PR is hardware intrinsic tests being added via the NotSupported template (so they are auto-generated, like most of the hwintrinsic tests).
Review of the PR can be simplified by just reviewing one of the generated tests since they all follow the same pattern.

Edit: This work was completed:
~~I am marking this as a draft until I can get some more tests up, including negative validating for nint/nuint on Vector64/128/256<T>.~~
~~I will also take a look at updating the couple places where we still have using nuint_t = System.UInt64, such as in Utf16Utility.Validation.cs~~

ghost · 2021-04-07T06:54:30Z

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, to 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.

tannergooding · 2021-04-08T15:20:26Z

@fanyang-mono, could you assist with the mono failures here?

It looks like there are two kinds of failures:

Mono isn't correctly treating Vector128<nint>.Zero as unsupported
Mono is faulting in src/mono/mono/mini/simd-intrinsics.c due to an assertion that shouldn't be reached (looks like its not handling the MONO_TYPE_ for nint and nuint).

Related, it looks like Mono has some hard-coded logic around ThrowForUnsupportedVectorBaseType (in src/mono/mono/mini/intrinsics.c) and that is being renamed/split into two methods for this PR.

The latter two issues look fairly straightforward to resolve, I'm just not familiar with what the correct MONO_TYPE_ is for nint or nuint.
The first issue I'm not sure where to start looking.

fanyang-mono · 2021-04-08T17:55:40Z

@fanyang-mono, could you assist with the mono failures here?

It looks like there are two kinds of failures:

Mono isn't correctly treating Vector128<nint>.Zero as unsupported

Mono is faulting in src/mono/mono/mini/simd-intrinsics.c due to an assertion that shouldn't be reached (looks like its not handling the MONO_TYPE_ for nint and nuint).

Related, it looks like Mono has some hard-coded logic around ThrowForUnsupportedVectorBaseType (in src/mono/mono/mini/intrinsics.c) and that is being renamed/split into two methods for this PR.

The latter two issues look fairly straightforward to resolve, I'm just not familiar with what the correct MONO_TYPE_ is for nint or nuint.
The first issue I'm not sure where to start looking.

nint and nuint is currently not support by mono. @vargaz Zoltan, could you please provide a little more details, if there is any?

tannergooding · 2021-04-08T18:00:23Z

nint and nuint is currently not support by mono

I'm not sure I understand this, they are just aliases for System.IntPtr and System.UIntPtr as far as the JIT type system is concerned?

vargaz · 2021-04-08T18:06:14Z

IntPtr/UIntPtr is MONO_TYPE_I/MONO_TYPE_U.

fanyang-mono · 2021-04-08T18:07:14Z

nint and nuint is currently not support by mono

I'm not sure I understand this, they are just aliases for System.IntPtr and System.UIntPtr as far as the JIT type system is concerned?

If that's the case, they should be mapped to MONO_TYPE_I and MONO_TYPE_U. Sorry that I misunderstood them with struct types.

tannergooding · 2021-04-08T18:46:39Z

I believe I've fixed the first issue (70432e5):

Mono isn't correctly treating Vector128.Zero as unsupported

and the third issue (c3107e0):

it looks like Mono has some hard-coded logic around ThrowForUnsupportedVectorBaseType (in src/mono/mono/mini/intrinsics.c) and that is being renamed/split into two methods for this PR.

I've also fixed part of the second issue (also c3107e0):

Mono is faulting in src/mono/mono/mini/simd-intrinsics.c due to an assertion that shouldn't be reached (looks like its not handling the MONO_TYPE_ for nint and nuint).

The remainder of the second issue looks to be that a couple code paths will still assert (type_to_expand_op, type_to_insert_op, and type_to_extract_var_op) since MONO_TYPE_I8 and MONO_TYPE_U8 aren't handled.
It isn't clear if this is better handled by picking an existing op (e.g. OP_EXPAND_I4 when targeting 32-bit and OP_EXPAND_I8 when targeting 64-bit) or if it is better to add a new op (e.g. OP_EXPAND_I) and plumb the handling through.

fanyang-mono · 2021-04-08T18:49:09Z

I believe I've fixed the first issue (70432e5):

Mono isn't correctly treating Vector128.Zero as unsupported

and the third issue (c3107e0):

it looks like Mono has some hard-coded logic around ThrowForUnsupportedVectorBaseType (in src/mono/mono/mini/intrinsics.c) and that is being renamed/split into two methods for this PR.

I've also fixed part of the second issue (also c3107e0):

Mono is faulting in src/mono/mono/mini/simd-intrinsics.c due to an assertion that shouldn't be reached (looks like its not handling the MONO_TYPE_ for nint and nuint).

The remainder of the second issue looks to be that a couple code paths will still assert (type_to_expand_op, type_to_insert_op, and type_to_extract_var_op) since MONO_TYPE_I8 and MONO_TYPE_U8 aren't handled.
It isn't clear if this is better handled by picking an existing op (e.g. OP_EXPAND_I4 when targeting 32-bit and OP_EXPAND_I8 when targeting 64-bit) or if it is better to add a new op (e.g. OP_EXPAND_I) and plumb the handling through.

Maybe something like this

case MONO_TYPE_I:
#if TARGET_SIZEOF_VOID_P == 8
<treat it the same as I8>
#else
<treat it as I4>
#endif

tannergooding · 2021-04-08T19:12:07Z

I believe the relevant mono paths are now updated to correctly handle Vector<nint> and Vector<nuint> and to block the same for Vector64/128/256<T> (at least until we take those to API review and hopefully unblock them as well). Hopefully CI passes now, thanks for the help! 😄

tannergooding · 2021-04-08T21:44:51Z

CC. @echesakovMSFT, @dotnet/jit-contrib for review

lambdageek · 2021-04-08T23:06:55Z

@vargaz could you review the mono commit 8eb6fe9

vargaz · 2021-04-09T11:10:18Z

The mono changes look ok to me.

…n for nint/nuint

… functions for Numerics vs Intrinsics

…vially pulled from the operand instead

…rowForUnsupportedNumericsVectorBaseType and ThrowForUnsupportedIntrinsicsVectorBaseType

…_U` on relevant code paths

…handles

tannergooding · 2021-04-14T16:27:25Z

Rebased to resolve a conflict, believe I've also resolved the remaining ARM64 issue (it passed on my Windows box).

…MD operations

tannergooding · 2021-04-15T02:40:51Z

The mono changes look ok to me.

@vargaz, I had to push another commit for Mono: 0f4764d

A bit of the handling existed in mini-amd64.c and so was missed in my first pass, but caught by the tests.
I think I've gotten it all this time, but we'll see what CI says 😄

tannergooding · 2021-04-15T19:33:39Z

Everything is passing now except for two LLVM AOT scenarios. It isn't clear to me what's wrong here given that the regular Mono leg passed.

@vargaz or @fanyang-mono, could you provide some guidance on how to root cause the failure here?

tannergooding · 2021-04-15T19:40:35Z

Actually, I may have found it, there is a single line calling out an assert about halfway through the log:

Assertion: should not be reached at /__w/1/s/src/mono/mono/mini/mini-llvm.c:580

tannergooding · 2021-04-16T01:39:09Z

Only test failures are unrelated and due to #51346

This should be ready for final review/merge.

GrabYourPitchforks · 2021-04-17T19:27:12Z

My small contribution to the review: commit 57fc73d LGTM!

echesakov

LGTM
Thank you for your contribution, @tannergooding!

cc @dotnet/jit-contrib

ghost added area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI new-api-needs-documentation labels Apr 7, 2021

tannergooding force-pushed the intrinsic-type branch 3 times, most recently from b686f43 to f3be3d6 Compare April 7, 2021 22:12

tannergooding removed the new-api-needs-documentation label Apr 7, 2021

tannergooding force-pushed the intrinsic-type branch 2 times, most recently from 3053f08 to 755f6e0 Compare April 8, 2021 02:30

tannergooding marked this pull request as ready for review April 8, 2021 21:44

tannergooding requested review from CoffeeFlux, lambdageek, SamMonoRT and vargaz as code owners April 8, 2021 21:44

tannergooding requested a review from echesakov April 8, 2021 21:44

JulieLeeMSFT assigned tannergooding Apr 8, 2021

JulieLeeMSFT added this to the 6.0.0 milestone Apr 8, 2021

This was referenced Apr 9, 2021

Helix DNS failures causing work item failures in CI #50750

Closed

[tests] System.Text.Json.Tests segfault, for Libraries Test Run release coreclr OSX x64 Release #47805

Closed

tannergooding added 8 commits April 13, 2021 15:02

Updating the Vector64/128/256<T> tests to have NotSupported validatio…

ed5db1f

…n for nint/nuint

Splitting ThrowHelper.ThrowForUnsupportedVectorBaseType into separate…

35764d7

… functions for Numerics vs Intrinsics

Updating Utf16Utility.Validation to directly use Vector<nuint>

57fc73d

Don't use the auxiliary type to hold a SIMD type, since it can be tri…

3d4fe63

…vially pulled from the operand instead

Split the mono handling for ThrowForUnsupportedVectorBaseType into Th…

fa2d443

…rowForUnsupportedNumericsVectorBaseType and ThrowForUnsupportedIntrinsicsVectorBaseType

Add basic handling for MONO_TYPE_I and MONO_TYPE_U to simd-intrinsics.c

18c2c1a

Ensure simd-intrinsics.c in Mono handles MONO_TYPE_I and `MONO_TYPE…

b311562

…_U` on relevant code paths

Ensure we don't assert when encountering synthesized Vector128<nint> …

c150826

…handles

tannergooding force-pushed the intrinsic-type branch from 8ada17b to c150826 Compare April 14, 2021 16:26

tannergooding added 3 commits April 14, 2021 10:25

Applying formatting patch

1ea8770

Fix the handling for Crc32 and Crc32C on ARM64

3b6f168

Updating Mono mini-amd64 to handle MONO_TYPE_I and MONO_TYPE_U for SI…

0f4764d

…MD operations

Handle OP_XCOMPARE.CMP_GE_UN for MONO_TYPE_U

a657763

Handle MONO_TYPE_I and MONO_TYPE_U for Vector types in mini-llvm

7f0dd03

tannergooding requested review from EgorBo and imhameed as code owners April 15, 2021 19:41

echesakov approved these changes Apr 20, 2021

View reviewed changes

tannergooding merged commit 17506ca into dotnet:main Apr 20, 2021

This was referenced Apr 21, 2021

Weird inlining of methods optimized by JIT #47434

Closed

Mark SIMD assignments as related to SIMD intrinsics #51731

Merged

JulieLeeMSFT mentioned this pull request Apr 27, 2021

What's new in .NET 6 Preview 4 dotnet/core#6098

Closed

AntonLapounov mentioned this pull request Apr 28, 2021

Fix hanging in JIT for ARM64 release compilation dotnet/runtimelab#1035

Closed

ghost locked as resolved and limited conversation to collaborators May 20, 2021

tannergooding deleted the intrinsic-type branch July 1, 2025 14:39

Updating Vector<T> to support nint and nuint #50832

Updating Vector<T> to support nint and nuint #50832

Uh oh!

Conversation

tannergooding commented Apr 7, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ghost commented Apr 7, 2021

Uh oh!

tannergooding commented Apr 8, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

fanyang-mono commented Apr 8, 2021

Uh oh!

tannergooding commented Apr 8, 2021

Uh oh!

vargaz commented Apr 8, 2021

Uh oh!

fanyang-mono commented Apr 8, 2021

Uh oh!

tannergooding commented Apr 8, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

fanyang-mono commented Apr 8, 2021

Uh oh!

tannergooding commented Apr 8, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

tannergooding commented Apr 8, 2021

Uh oh!

lambdageek commented Apr 8, 2021

Uh oh!

vargaz commented Apr 9, 2021

Uh oh!

tannergooding commented Apr 14, 2021

Uh oh!

tannergooding commented Apr 15, 2021

Uh oh!

tannergooding commented Apr 15, 2021

Uh oh!

tannergooding commented Apr 15, 2021

Uh oh!

tannergooding commented Apr 16, 2021

Uh oh!

GrabYourPitchforks commented Apr 17, 2021

Uh oh!

echesakov left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tannergooding commented Apr 7, 2021 •

edited

Loading

tannergooding commented Apr 8, 2021 •

edited

Loading

tannergooding commented Apr 8, 2021 •

edited

Loading

tannergooding commented Apr 8, 2021 •

edited

Loading