LLVM crash in performance_nested_vectorization_gemm (arm-64, llvm-main)

[stevesuzuki-arm]

Description

arm-64, llvm-main CI failure
performance_nested_vectorization_gemm

performance_nested_vectorization_gemm: /__w/halide-llvm/halide-llvm/src_cache/3b4499c9eba52ca462a0a6b3764f6bbdfa399da6/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:329: unsigned int getNumElements(llvm::Type*): Assertion `!isa<ScalableVectorType>(Ty) && "ScalableVectorType is not supported."' failed.

At least, #9128 is related.

Filed LLVM issue: llvm/llvm-project#197132

Reproducing case

Run ctest -R performance_nested_vectorization_gemm on SVE2 enabled host compiled with the latest halide_llvm 23.0.0.dev92488+g3b4499c9

Halide commit (if known)

Found in CI of #9131

Target

arm-64-linux-sve2-vector_bits_128

Operating system

arm-64-linux

Additional context

No response

[alexreinking]

Seems like a CI race between upgrading LLVM (the checks passed on that PR) and merging one of the SVE-related changes. It's possible we need to skip that test until LLVM fixes their backend.

[stevesuzuki-arm]

LLVM Issue has been fixed and closed. I've confirmed the test failure in Halide is not reproduced with LLVM 86f7e93921fd.
However, a new failure happens in simd_op_check_arm, test_op_vneg_f64_235_vecwidth2, arm-32 target

[alexreinking]

@stevesuzuki-arm -- is there a tracking issue with LLVM for that test?

[stevesuzuki-arm]

@stevesuzuki-arm -- is there a tracking issue with LLVM for that test?

No, AFAIK. Right now, I have no idea whether it is LLVM issue or Halide test assumption issue.

[alexreinking]

Let me manually launch a Halide LLVM update PR

[stevesuzuki-arm]

However, a new failure happens in simd_op_check_arm, test_op_vneg_f64_235_vecwidth2, arm-32 target

#9140 fixes it