`__BIGGEST_ALIGNMENT__` not the same as gcc for AVX and later

[llvmbot]

Bugzilla Link	42554
Version	8.0
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@topperc,@DougGregor,@RKSimon,@zygoloid

Extended Description

When compiled -march=native (on znver1) with gcc and clang I get the following:

• gcc 9.1.1 20190503 (Red Hat 9.1.1-1)

    alignof(max_align_t)=16  __BIGGEST_ALIGNMENT__=32  malloc() align=16
    __SSE__  __AVX__  __AVX2__

• clang 8.0.0 (Fedora 8.0.0-1.fc30)
  (4.2.1 Compatible Clang 8.0.0 (Fedora 8.0.0-1.fc30))

   alignof(max_align_t)=16  __BIGGEST_ALIGNMENT__=16  malloc() align=16
   __SSE__  __AVX__  __AVX2__

When compiled without -march=native, both give:

   alignof(max_align_t)=16  __BIGGEST_ALIGNMENT__=16  malloc() align=16
   __SSE__

My guess is that clang's __BIGGEST_ALIGNMENT__ for -march=native is a bug... except that gcc v4.2.1 predates AVX ???

FWIW: in all cases glibc malloc() aligned to 16. My reading of the glibc code tells me that its alignment is 2*sizeof(size_t), which is in fact the same as alignof(max_align_t), but is in no way related.

[AaronBallman]

https://godbolt.org/z/TjMsWqeo3

Clang: #define __BIGGEST_ALIGNMENT__ 16
GCC: #define __BIGGEST_ALIGNMENT__ 32

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (llvmbot)

| | | | --- | --- | | Bugzilla Link | [42554](https://llvm.org/bz42554) | | Version | 8.0 | | OS | Linux | | Reporter | LLVM Bugzilla Contributor | | CC | @topperc,@DougGregor,@RKSimon,@zygoloid |

Extended Description

When compiled -march=native (on znver1) with gcc and clang I get the following:

• gcc 9.1.1 20190503 (Red Hat 9.1.1-1)

    alignof(max_align_t)=16  __BIGGEST_ALIGNMENT__=32  malloc() align=16
    __SSE__  __AVX__  __AVX2__

• clang 8.0.0 (Fedora 8.0.0-1.fc30)
  (4.2.1 Compatible Clang 8.0.0 (Fedora 8.0.0-1.fc30))

   alignof(max_align_t)=16  __BIGGEST_ALIGNMENT__=16  malloc() align=16
   __SSE__  __AVX__  __AVX2__

When compiled without -march=native, both give:

   alignof(max_align_t)=16  __BIGGEST_ALIGNMENT__=16  malloc() align=16
   __SSE__

My guess is that clang's __BIGGEST_ALIGNMENT__ for -march=native is a bug... except that gcc v4.2.1 predates AVX ???

FWIW: in all cases glibc malloc() aligned to 16. My reading of the glibc code tells me that its alignment is 2*sizeof(size_t), which is in fact the same as alignof(max_align_t), but is in no way related.