Type error on SIMD native types not caught by compiler

[jfalcou]

@SadiinsoSnowfall and I were cleaning up some more of our SIMD library code and found that code like:

auto f(__m128d a, __m128 b) { return a + b; }

Compiles when it should fail.

Reproduction :

https://godbolt.org/z/eE4b4Kdn8

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Joel Falcou (jfalcou)

@SadiinsoSnowfall and I were cleaning up some more of our SIMD library code and found that code like:

auto f(__m128d a, __m128 b) { return a + b; }

Compiles when it should fail.

Reproduction :

https://godbolt.org/z/eE4b4Kdn8

[RKSimon]

This is caught with -flax-vector-conversions=none

[jfalcou]

It should just not compile no ?

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Joel Falcou (jfalcou)

@SadiinsoSnowfall and I were cleaning up some more of our SIMD library code and found that code like:

auto f(__m128d a, __m128 b) { return a + b; }

Compiles when it should fail.

Reproduction :

https://godbolt.org/z/eE4b4Kdn8

[RKSimon]

#17538

[jfalcou]

OK, but as for Agner Fog, I can't really rely on people to use this. Is there any plan to fix it from within the intrinsic file ?

[RKSimon]

OK, but as for Agner Fog, I can't really rely on people to use this. Is there any plan to fix it from within the intrinsic file ?

The intrinsics headers are tested with -flax-vector-conversions=none already so they should be clean, there's no intention to force apply compile flags based off including the intrinsics headers.

[jfalcou]

OK, but as for Agner Fog, I can't really rely on people to use this. Is there any plan to fix it from within the intrinsic file ?

The intrinsics headers are tested with -flax-vector-conversions=none already so they should be clean, there's no intention to force apply compile flags based off including the intrinsics headers.

clearly they are not then ?

[phoebewang]

OK, but as for Agner Fog, I can't really rely on people to use this. Is there any plan to fix it from within the intrinsic file ?

The intrinsics headers are tested with -flax-vector-conversions=none already so they should be clean, there's no intention to force apply compile flags based off including the intrinsics headers.

clearly they are not then ?

$ grep -r flax-vector-conversions clang/test/CodeGen/X86/ | wc -l
199