CODEWITHSUNDEEP
cgccauto-vectorization
freddofrog
freddofrog

Reputation: 101

gcc won't vectorize simple loop

I'm trying to vectorise simplified version of Example 4 from the gcc auto-vectorize documentation. For the life of me, I can't work out how to do it;

typedef int aint __attribute__ ((__aligned__(16)));
void foo1 (int n, aint * restrict px, aint *restrict qx) {

  /* feature: support for (aligned) pointer accesses.  */
  int *__restrict p = __builtin_assume_aligned (px, 16);
  int *__restrict q = __builtin_assume_aligned (qx, 16);

  while (n--){
    //*p++ += *q++; <- this is vectorized                                                                                                                                                                   
    p[n] += q[n]; // This isn't!                                                                                                                                                                            
  }
}

I'm running gcc 4.7.2 with gcc -o apps/craft_dbsplit.o -c -Wall -g -ggdb -O3 -msse2 -funsafe-math-optimizations -ffast-math -ftree-vectorize -ftree-vectorizer-verbose=5 -funsafe-loop-optimizations -std=c99

And it replies with:

Analyzing loop at apps/craft_dbsplit.c:388

388: dependence distance  = 0.
388: dependence distance == 0 between *D.9363_14 and *D.9363_14
388: dependence distance  = 0.
388: accesses have the same alignment.
388: dependence distance modulo vf == 0 between *D.9363_14 and *D.9363_14
388: vect_model_load_cost: unaligned supported by hardware.
388: vect_get_data_access_cost: inside_cost = 2, outside_cost = 0.
388: vect_model_store_cost: unaligned supported by hardware.
388: vect_get_data_access_cost: inside_cost = 2, outside_cost = 0.
388: Alignment of access forced using peeling.
388: Vectorizing an unaligned access.
388: vect_model_load_cost: aligned.
388: vect_model_load_cost: inside_cost = 1, outside_cost = 0 .
388: vect_model_load_cost: unaligned supported by hardware.
388: vect_model_load_cost: inside_cost = 2, outside_cost = 0 .
388: vect_model_simple_cost: inside_cost = 1, outside_cost = 0 .
388: not vectorized: relevant stmt not supported: *D.9363_14 = D.9367_20;

apps/craft_dbsplit.c:382: note: vectorized 0 loops in function.

Upvotes: 3

Views: 705

Answers (1)

user3185968
user3185968

Reputation:

The loop runs from high addresses to low addresses. Your gcc treats vector-operations as running from low addresses to high addresses, and thus doesn't realize it can vectorize. Your "optimization", making the loop a while (n--), is actually preventing the more relevant optimization. Try

#include <stddef.h>

void foo1 (size_t n, int *restrict px, int const *restrict qx)
{
  int *restrict p = __builtin_assume_aligned(px, 16);
  int const *restrict q = __builtin_assume_aligned(qx, 16);
  size_t i = 0;
  while (i < n)
    {
      p[i] += q[i];
      i++;
    }
}

Upvotes: 1

Related Questions