[Bug tree-optimization/67755] New: Bad edge probability/block freqency for tree jump threading

[pthaugen at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67755

            Bug ID: 67755
           Summary: Bad edge probability/block freqency for tree jump
                    threading
           Product: gcc
           Version: 5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pthaugen at gcc dot gnu.org
                CC: bergner at gcc dot gnu.org, dje at gcc dot gnu.org,
                    tejohnson at google dot com
  Target Milestone: ---
              Host: powerpc64-unknown-linux-gnu
            Target: powerpc64-unknown-linux-gnu
             Build: powerpc64-unknown-linux-gnu

Noticed a problem where static branch hint was being applied to a branch
resulting in performance degradation, bisected change to r215739. Reduced to
following testcase, compiled with "-S -m64 -O1".

void (*zend_block_interruptions) (void);

int * _zend_mm_alloc_int (int * heap, long int size)
{
  int *best_fit;
  long int true_size = (size < 15 ? 32 : size);

  if (zend_block_interruptions)
    zend_block_interruptions ();

  if (__builtin_expect ((true_size < 543), 1))
    best_fit = heap + 2;
  else
    best_fit = heap;

  return best_fit;
}


For the snippet of code:
  if (zend_block_interruptions)
    zend_block_interruptions ();

094t.copyrename3 dump shows probability of 69.8%/30.2% (true/false) for the
function pointer <> null test. After jump threading, 095t.dom1 shows
probability of 100%/0% for the original block and 39.7%/60.3% for the
duplicated block. The original block is then statically predicted since it has
100% probability. Also note that the probabilities for the duplicated block
have approx been reversed, which also seems wrong.