malloc/memset -> calloc optimization should happen even if there are BBs between the ptr check and the memset

[pinskia]

take:

void *combine_alt (int s, bool c, bool *a)
{
  void *r = __builtin_malloc (s);
  if (!r)  return 0;
  if (a) *a = 1;
  __builtin_memset (r, 0, s);
  return r;
}

This should be optimized to:

void *combine_alt_tgt (int s, bool c, bool *a)
{
  void *r = __builtin_calloc (s, 1);
  if (!r)  return 0;
  if (a) *a = 1;
  return r;
}

Right now the check in shouldCreateCalloc is too restrictive. Instead of checking MemsetBB != FalseBB, it should be checking that MemsetBB post dominates the FalseBB.

Note if you remove the check for !r, shouldCreateCalloc is too restrictive there too. that is because MallocBB == MemsetBB check should be similarly a post dominate check.

Note I found this while fixing up GCC finally (https://gcc.gnu.org/PR83022). And I thought I would report the missed optimization back to LLVM.