https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97028
Bug ID: 97028
Summary: Compilation errors for AVX512 intrinsic with
-masm=intel
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: nikita.shulga at gmail dot com
Target Milestone: ---
Following code snippet, containing no asm blocks can be compiled just fine with
'-Os -mavx512f' options, by compilation fails using gcc-9.0 or newer if
'-masm=intel' is added：
#include <immintrin.h>
float foo(float* con) {
__mmask16 msk = 0x00ff;
__m512 a = _mm512_maskz_loadu_ps(msk, con);
__m512 b = _mm512_set1_ps(con[1]);
__m512 c = _mm512_mul_ps(a,b);
return ((float *)&c)[0];
}
That is:
$ gcc -c -Os -mavx512f mul.c ; echo $?
0
$ gcc -c -Os -masm=intel -mavx512f mul.c ; echo $?
/tmp/ccmeN1OP.s: Assembler messages:
/tmp/ccmeN1OP.s:12: Error: invalid use of register
1
And the offending instruction looks as follows:
$ gcc -S -o - -Os -masm=intel -mavx512f mul.c|head -n 12|tail -n1
vmulps zmm0, zmm0, DWORD PTR [rdi+4]<{1to16}>
But compilation works just fine if gcc-8 is used instead.
