gcc assembly for stack vs heap access

Question

Here are two links to godblot(replaced by the actual code), in one case I sum the elements of an array declared on the stack and the second sum elements of array allocated on the heap (please ignore the incorect size for malloc).

Heap (-std=c++17 -O3 -mavx2):

#include <stdio.h>
#include <stdlib.h>
int access(){
    int sum;
    //int vec[10000];
    int *vec = (int*)malloc(10000*sizeof(int));

    for(int i=0;i<10000;i++)
        sum+=vec[i];

    return sum;
}

Assembly for heap:

access():
  push rbp
  mov edi, 10000
  mov rbp, rsp
  and rsp, -32
  call malloc
  vpxor xmm1, xmm1, xmm1
  lea rdx, [rax+40000]
.L2:
  vmovdqu xmm2, XMMWORD PTR [rax]
  vinserti128 ymm0, ymm2, XMMWORD PTR [rax+16], 0x1
  add rax, 32
  vpaddd ymm1, ymm1, ymm0
  cmp rdx, rax
  jne .L2
  vmovdqa xmm0, xmm1
  vextracti128 xmm1, ymm1, 0x1
  vpaddd xmm1, xmm0, xmm1
  vpsrldq xmm0, xmm1, 8
  vpaddd xmm0, xmm1, xmm0
  vpsrldq xmm1, xmm0, 4
  vpaddd xmm0, xmm0, xmm1
  vmovd eax, xmm0
  vzeroupper
  leave
  ret

For stack C code just use the vec declared on the stack that is commented in the previous C code.

Stack() assembly code:

access():
  push rbp
  vpxor xmm0, xmm0, xmm0
  mov rbp, rsp
  and rsp, -32
  sub rsp, 39880
  lea rax, [rsp-120]
  lea rdx, [rsp+39880]
.L2:
  vpaddd ymm0, ymm0, YMMWORD PTR [rax]
  add rax, 32
  cmp rdx, rax
  jne .L2
  vmovdqa xmm1, xmm0
  vextracti128 xmm0, ymm0, 0x1
  vpaddd xmm0, xmm1, xmm0
  vpsrldq xmm1, xmm0, 8
  vpaddd xmm0, xmm0, xmm1
  vpsrldq xmm1, xmm0, 4
  vpaddd xmm0, xmm0, xmm1
  vmovd eax, xmm0
  vzeroupper
  leave
  ret

Any idea why in the heap case it needs to use more instructions. It also seems that it skips 16 bytes of data which it stores in xmm2 register?

If we change the compiler argument from -mavx2 to -mavx512f or remove completly remove it, the code looks similar. Is there an issue with the gcc trunk version?

Clang seems to unroll the loop and generates code that makes sense.

Not accessible. Please upload those examples together with your question to enhance this and let other learn from it. — dboy, May 08 '20 at 11:49
`malloc` does not guarantee 32 byte alignment. Use `aligned_alloc(32, 40000);` instead. — Jester, May 08 '20 at 11:55
It's different with `-mavx512vl` or `-mavx512f` because the `-mavx256-split-unaligned-load` default tuning option (helps Sandybridge and some AMD) only applies to 256-bit loads, and for some reason GCC defaults to 512-bit vectors when you just use `-mavx512f`. I guess the usual `-mprefer-vector-width=256` is only part of `-march=skylake-avx512`, not on by default, and you didn't use any tuning options. — Peter Cordes, May 08 '20 at 12:04

gcc assembly for stack vs heap access

0 Answers0