https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122074
Bug ID: 122074
Summary: Wrong code for avx512 intrinsic
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: rockeet at gmail dot com
Target Milestone: ---
#include <immintrin.h>
using byte_t = unsigned char;
size_t avx512_search_byte_max32(const byte_t* data, size_t len, byte_t key) {
return _tzcnt_u32(_mm256_mask_cmpge_epi8_mask(_bzhi_u32(-1, len),
*(__m256i_u*)data, _mm256_set1_epi8(key)));
}
see https://godbolt.org/z/Ef157Pj3n , gcc generate:
"avx512_search_byte_max32(unsigned char const*, unsigned long, unsigned char)":
vmovdqu ymm1, YMMWORD PTR [rdi]
vpbroadcastb ymm0, edx
mov eax, -1
bzhi eax, eax, esi
kmovd k1, eax
vpcmpb k0{k1}, ymm1, ymm0, 5
kmovd eax, k0
tzcnt eax, eax
mov eax, eax
vzeroupper
ret
intel icc generate:
avx512_search_byte_max32(unsigned char const*, unsigned long, unsigned char):
..B3.1: # Preds ..B3.0
mov eax, -1 #20.20
vpbroadcastb ymm0, edx #20.20
bzhi ecx, eax, esi #20.20
kmovd k1, ecx #20.20
vpcmpb k0{k1}, ymm0, YMMWORD PTR [rdi], 2 #20.20
kmovd r8d, k0 #20.20
tzcnt eax, r8d #20.9
vzeroupper #20.9
ret #20.9
The issue is: if [data, data+32) spans page boundary and bytes corresponding to
0 in mask is invalid address(such as segfault/bus error), code generated by gcc
will coredump.