https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82735
Bug ID: 82735
Summary: _mm256_zeroupper does not invalidate previously
computed registers
Product: gcc
Version: 7.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: marcin.slusarz at intel dot com
Target Milestone: ---
$ cat main.c
#include <stdio.h>
#include <string.h>
void test(char *dest);
int main()
{
char buf[32];
memset(buf, 0x2, 32);
test(buf);
for (int i = 0; i < 32; ++i)
printf("%d ", buf[i]);
printf("\n");
}
$ cat zeroupper.c
#include <immintrin.h>
void test(char *dest)
{
__m256i ymm1 = _mm256_set1_epi8((char)0x1);
_mm256_storeu_si256((__m256i *)dest + 32, ymm1);
_mm256_zeroupper();
__m256i ymm2 = _mm256_set1_epi8((char)0x1);
_mm256_storeu_si256((__m256i *)dest, ymm2);
}
$ gcc -c -std=c99 -O2 main.c
$ gcc -c -std=c99 -O2 zeroupper.c -mavx
$ gcc zeroupper.o main.o -o zeroupper
$ ./zeroupper
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Expected output is:
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
I can reproduce it on gcc 4.9, 5.4.1, 6.3.0 and 7.2.1.
gcc 4.6.4 is not affected.
clang 3.5, 3.8, 3.9 and 4.0 produce the expected output.
