[Bug tree-optimization/37242] New: missed load PRE-like opportunity

["bonzini at gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

In the attached code, the if-conversion opportunity in PR37240 does not benefit
a Pentium 4.  However, there is another optimization possible on both systems,
namely changing

    while ((maxIdx += maxIdx) < last) {
        if (numbers[maxIdx] < numbers[maxIdx + 1]) maxIdx++;
        if (tmp >= numbers[maxIdx]) break;
        numbers[top] = numbers[maxIdx];
        top = maxIdx;
    }

to

    while ((maxIdx += maxIdx) < last) {
        int a = numbers[maxIdx], b = numbers[maxIdx + 1];
        if (a < b) maxIdx++, a = b;
        if (tmp >= a) break;
        numbers[top] = a;
        top = maxIdx;
    }

It seems to me that numbers[maxIdx] is partially redundant (it is available if
maxIdx++ is not executed).  If an additional load of numbers[maxIdx + 1] is
inserted in the "then" branch, it can also be found to be fully redundant so
that copy propagation generates the optimized code.

This gives a ~3% performance increase on i686-pc-linux-gnu for this benchmark.


-- 
           Summary: missed load PRE-like opportunity
           Product: gcc
           Version: 4.3.2
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: bonzini at gnu dot org
GCC target triplet: i686-pc-linux-gnu
 BugsThisDependsOn: 37239


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37242