https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121606
Bug ID: 121606
Summary: -march=native on AVX10.1 capable host warns about
-mno-evex512
Product: gcc
Version: 15.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: marko.makela at mariadb dot com
Target Milestone: ---
The following test program throws a warning when built with -march=native on a
system that supports AVX10.1.
#include <immintrin.h>
__attribute__((target("avx10.1"))) __m128i xor3(__m128i a,__m128i b,__m128i c)
{ return _mm_ternarylogic_epi64(a, b, c, 0xf0 ^ 0xcc ^ 0xaa); }
s.c:3:1: warning: ‘-mno-evex512’ or ‘-mno-avx512XXX’ cannot disable AVX10
instructions when AVX10.1-512 is available in GCC 15, behavior will change to
it will disable that part of AVX512 instructions since GCC 16
3 | { return _mm_ternarylogic_epi64(a, b, c, 0xf0 ^ 0xcc ^ 0xaa); }
| ^
Curiously, this seems to contradict the output of the following:
gcc-15 -march=native -Q --help=target|grep evex512
-mevex512 [enabled]
Without -march=native there is no warning. In either case, the intrinsic
correctly translates into vpternlogq 0x96 (ternary exclusive or).
I am using the Debian package gcc-15, which reports its version as 15.2.0-1.
The above code was extracted from https://github.com/dr-m/crc32_simd (file
crc32_avx512.cc). That has been tested with GCC 11 through 14 when using the
equivalent target("evex512,avx512f,avx512dq,avx512bw,avx512vl"). I just tested
it with the Debian gcc-14 version 14.3.0-5.
In case it matters, the host CPU is an AMD Ryzen AI 9 HX 370 w/ Radeon 890M
(AMD Zen 5 microarchitecture).
