[Bug tree-optimization/123966] New: Eliminate atomic relaxed memory order load with unused result

[law at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123966

            Bug ID: 123966
           Summary: Eliminate atomic relaxed memory order load with unused
                    result
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: law at gcc dot gnu.org
  Target Milestone: ---

So Misha and I were looking at 520.omnet a while back and realized that there
was a load where we never used the result.  This was buried inside jemalloc's
free() routine.  So it's not crazy hot, but it is important.  And being a load
amplifies the value in removing the instrution over simple ALU ops.  The
sequence we saw for rv64 was:

        ld      a3,0(a3)
        andi    a3,a4,1

Note how the load writes a3, then the andi immediately overwrites the value in
a3 without reading the value in a3.

This was ultimately tracked down to an atomic load where we never used the
result.  It's kind of inherent in the jemalloc code paths.  It's a relaxed
memory order so I believe it is safe to elide the load if its result is not
used.


void idalloctm() {
  unsigned result;
  __atomic_load(&atomic_load_u_a, &result, 0);
}