[Bug rtl-optimization/65651] Redundant cmp with zero instruction in loop for x86 target.

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65651

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |uros at gcc dot gnu.org

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Well, there is a significant difference between the two testcases, one uses the
result of the comparison just in == 0 test, thus CCZmode is appropriate, the
other uses it in two comparisons, one == 0 test and one < 0 test.
For combine to match *sub<mode>_2 insn, it has to match
ix86_match_ccmode (insn, CCGOCmode)
where CCGOCmode stands for:
   Add CCGOC to indicate comparisons against zero that allows
   unspecified garbage in the Carry and Overflow flag. This
   mode is used to simulate comparisons of (a-b) and (a+b)
   against zero using sub/cmp/add operations.
But the jle instruction tests ZF || SF <> OF and thus it isn't appropriate.
So the question is if the CCGOC test isn't too restrictive, say if CCGCmode
would be sufficient (but then we'd still need to arrange for the CCGCmode to be
used, rather than CCmode), or if the optimization you are looking for is simply
not possible.