AVX load instructions can be made to ignore segfaults on unused data

mask-test.cc

#include <immintrin.h>
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <sys/mman.h>

// Compile with clang++ -g -mavx512bw -mavx512f -o mask-test mask-test.cc

int main() {
  char* addr = (char*)mmap(0, 8192, PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, -1, 0);
  int result = mprotect(addr + 4096, 4096, PROT_NONE);
  if (result != 0) abort();
  __mmask32 mask = 0xFFFFFFFFULL; 
  for (int i = 0; i < 64; i += 2) {
    // This address appears to go over the end of the page into the
    // protected page when i != 0 and we load a 512 bit (64 byte) value.
    char* bad = addr + 4096 - 64 + i;
    // Load 512 bits into register, but use the mask to zero some of the
    // 16 bit words.
    __m512i input = _mm512_maskz_loadu_epi16(mask, bad);
    // Make sure the optimizer doesn't remove the load instruction.
    uint16_t values[32];
    _mm512_store_si512((__m512i*)values, input);
    printf("Values[31] = 0x%04x\n", values[31]);
    // For the next iteration we add a bit of masking which will both zero the
    // value in the register, but also suppresses faults!
    mask >>= 1;
  }
}