[Bug tree-optimization/91986] New: missed loop unrolling optimization

[trass3r at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91986

            Bug ID: 91986
           Summary: missed loop unrolling optimization
           Product: gcc
           Version: 9.2.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: trass3r at gmail dot com
  Target Milestone: ---

#include <cstdint>
#include <utility>
const int N = 4;
void bitreverse(int* data)
{
    uint32_t j = 0;
    for (uint32_t i = 0; i < N; ++i)
    {
        if (j > i)
            std::swap(data[i], data[j]);

        uint32_t k = N/2;
        while (k <= j)
        {
            j -= k;
            k /= 2;
        }
        j += k;
    }
}

Even with -O3 gcc doesn't fully unroll the loop but still there are quite some
redundant instructions: https://godbolt.org/z/cA-S8J
It's similar with N=8: https://godbolt.org/z/niZPeS

Interestingly if you change the swap line slightly clang suddenly starts using
pshufd which seems quite clever: https://godbolt.org/z/5YJnJ1