Inefficient codegen for loop with known small trip count

[dzaima]

The code

#include<stdint.h>
void f(char* a, uint8_t l) {
  for (int i = 0; i < l; i++) a[i]++;
}

compiled with -O3 -march=haswell (or -O3 -mavx2 or similar; compiler explorer), generates unrolled xmm operations, even though ymm ones can trivially replace them.

Same thing happens with an assume of length:

void f(char* a, int l) {
  __builtin_assume(l < 256);
  for (int i = 0; i < l; i++) a[i]++;
}

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: dzaima (dzaima)

The code

#include<stdint.h>
void f(char* a, uint8_t l) {
  for (int i = 0; i < l; i++) a[i]++;
}

compiled with -O3 -march=haswell (or -O3 -mavx2 or similar; compiler explorer), generates unrolled xmm operations, even though ymm ones can trivially replace them.

Same thing happens with an assume of length:

void f(char* a, int l) {
  __builtin_assume(l < 256);
  for (int i = 0; i < l; i++) a[i]++;
}

[miguelraz]

@RKSimon I'd like to be assigned this issue and try my hand at it.

[RKSimon]

@RKSimon I'd like to be assigned this issue and try my hand at it.

Go for it! @fhahn might be able to give some tips regarding any loop vectorizer related issue.