[llvm-dev] k registers in extended asm

[Scott Manley via llvm-dev]

Hi,

I am getting an abort in llc with 'main' and LLVM 13.0 when I run clang at
-O0 for the following test case:

#include <immintrin.h>
void small(float *b, __m512 c, __mmask16 mask)
{
  __m512 a = _mm512_setzero_ps();
  asm("vfmadd231ps (%1), %2, %0 %{%3%}": "+v"(a): "r"(b), "v"(c), "k"
(mask));
  _mm512_mask_storeu_ps(b, mask, a);
}

clang -S bug.c -march=skylake-avx512 -O0

bug.c:6:7: error: Register k0 can't be used as write mask
  asm("vfmadd231ps (%1), %2, %0 %{%3%}": "+v"(a): "r"(b), "v"(c), "k"
(mask));
      ^
<inline asm>:1:36: note: instantiated into assembly here
        vfmadd231ps (%rax), %zmm1, %zmm0 {%k0}


Works: clang -S bug.c -march=skylake-avx512 -O2 // chooses k1

The first observation I have is that there is no write of k0 on the fma
instruction. So even if I wrote my test case to explicitly use k0, there is
still the abort at -O0. This feels like a bug.

The second thing I have -- is it required that I specify the mask register
I want since k0 is "special"? Or is llc incorrectly choosing the k0
register in the first place?

Thanks!

Scott
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20211104/a217367a/attachment.html>