[Bug c/82735] New: _mm256_zeroupper does not invalidate previously computed registers

[marcin.slusarz at intel dot com]

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.