AArch64: sdot/udot not generated with variable or high trip counts

[rscottmanley]

The following test is not generating sdot/udot (depending on types) when the tripcount is variable, or if the loop is not unwound (the cutoff is at 60 trips, though may differ per target).

#include <stdint.h>
int32_t f(int8_t * restrict x, int8_t * restrict y, int n)
{
  int32_t r = 0;
  for (int j = 0; j < n; ++j) {
      r += x[j] * y[j];
  }
  return r;
}

clang (does not generate sdot): https://godbolt.org/z/KznKr1Kh8

gcc (generates sdot): https://godbolt.org/z/h1xqM1xMc

If you replace 'n' with some value < 60, you will see the instructions. Same problem with unsigned. This appears to be a problem with neoverse-n1, neoverse-v1 and neoverse-v2

[rscottmanley]

@paulwalker-arm @sjoerdmeijer

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Scott Manley (rscottmanley)

The following test is not generating sdot/udot (depending on types) when the tripcount is variable, or if the loop is not unwound (the cutoff is at 60 trips, though may differ per target).

#include <stdint.h>
int32_t f(int8_t * restrict x, int8_t * restrict y, int n)
{
  int32_t r = 0;
  for (int j = 0; j < n; ++j) {
      r += x[j] * y[j];
  }
  return r;
}

clang (does not generate sdot): https://godbolt.org/z/KznKr1Kh8

gcc (generates sdot): https://godbolt.org/z/h1xqM1xMc

If you replace 'n' with some value < 60, you will see the instructions. Same problem with unsigned. This appears to be a problem with neoverse-n1, neoverse-v1 and neoverse-v2

[paulwalker-arm]

You might be interested in #69587 and #69583. They're SVE focused but the plan is to expand the coverage to fixed length vectors as well.

[davemgreen]

This is done in the loop vectorizer now.