| 1 |
/* PR target/97642 */ |
| 2 |
/* { dg-do run { target *-*-linux* } } */ |
| 3 |
/* { dg-options "-O2 -mavx512dq -mavx512vl -mavx512bw" } */ |
| 4 |
/* { dg-require-effective-target avx512vl } */ |
| 5 |
/* { dg-require-effective-target avx512dq } */ |
| 6 |
/* { dg-require-effective-target avx512bw } */ |
| 7 |
|
| 8 |
#include <assert.h> |
| 9 |
#include <immintrin.h> |
| 10 |
#include <stdint.h> |
| 11 |
#include <sys/mman.h> |
| 12 |
|
| 13 |
#define N 5 |
| 14 |
|
| 15 |
// Faults with GCC because usage of vpblendd |
| 16 |
__m256i __attribute__((noinline)) mask_load(uint32_t * arr) { |
| 17 |
__m256i tmp; |
| 18 |
return _mm256_mask_loadu_epi32(tmp, (1 << N) - 1, arr); |
| 19 |
} |
| 20 |
|
| 21 |
// Faults |
| 22 |
__m256i __attribute__((noinline)) blend_load_asm(uint32_t * arr) { |
| 23 |
__m256i tmp = _mm256_set1_epi64x(0); |
| 24 |
asm volatile("vpblendd %[m], (%[arr]), %[tmp], %[tmp]\n\t" |
| 25 |
: [ tmp ] "+x"(tmp) |
| 26 |
: [ arr ] "r"(arr), [ m ] "i"(((1 << N) - 1)) |
| 27 |
:); |
| 28 |
return tmp; |
| 29 |
} |
| 30 |
|
| 31 |
// Does not fault |
| 32 |
__m256i __attribute__((noinline)) mask_load_asm(uint32_t * arr) { |
| 33 |
__m256i tmp; |
| 34 |
asm volatile( |
| 35 |
"movb %[m], %%al\n\t" |
| 36 |
"kmovb %%eax, %%k1\n\t" |
| 37 |
"vmovdqu32 (%[arr]), %[tmp] %{%%k1} %{z%}\n\t" |
| 38 |
: [ tmp ] "+x"(tmp) |
| 39 |
: [ arr ] "r"(arr), [ m ] "i"(((1 << N) - 1)) |
| 40 |
: "eax", "k1"); |
| 41 |
return tmp; |
| 42 |
} |
| 43 |
|
| 44 |
|
| 45 |
void __attribute__((noinline)) mask_store(uint32_t * arr, __m256i v) { |
| 46 |
return _mm256_mask_storeu_epi32(arr, (1 << N) - 1, v); |
| 47 |
} |
| 48 |
|
| 49 |
|
| 50 |
#define NPAGES (2) |
| 51 |
#define END_OF_PAGE (1024 - N) |
| 52 |
|
| 53 |
#ifndef LOAD_METHOD |
| 54 |
#define LOAD_METHOD mask_load // mask_load_asm does not fault |
| 55 |
#endif |
| 56 |
|
| 57 |
|
| 58 |
int |
| 59 |
main() { |
| 60 |
if (!(__builtin_cpu_supports ("avx512dq") |
| 61 |
&& __builtin_cpu_supports ("avx512vl") |
| 62 |
&& __builtin_cpu_supports ("avx512bw"))) |
| 63 |
return 0; |
| 64 |
|
| 65 |
uint32_t * addr = |
| 66 |
(uint32_t *)mmap(NULL, NPAGES * 4096, PROT_READ | PROT_WRITE, |
| 67 |
MAP_ANONYMOUS | MAP_PRIVATE, -1, 0); |
| 68 |
|
| 69 |
for (uint32_t i = 0; i < NPAGES; i += 2) { |
| 70 |
|
| 71 |
uint32_t page_offset = 1024 * i + END_OF_PAGE; |
| 72 |
uint32_t next_page_offset = 1024 * (i + 1); |
| 73 |
|
| 74 |
assert(!mprotect(addr + next_page_offset, 4096, PROT_NONE)); |
| 75 |
mask_store(addr + page_offset, LOAD_METHOD(addr + page_offset)); |
| 76 |
} |
| 77 |
} |