[Bug target/64338] [5 Regression] ICE in swap_condition, at jump.c:628

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64338

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 34393
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=34393&action=edit
gcc5-pr64338.patch

ix86_expand_int_movcc is called there with (ltgt (reg:CCFPU flags) (const_int
0)) condition, and I see several issues with that:
1) the code assumes that reverse_condition (whatever kind of that) always
succeeds, which is not always true; we shouldn't swap things if we can't
reverse condition
2) for CCFP/CCFPU modes, we should treat them like floating point modes, LTGT
is reversible
3) in the first hunk, there are dead assignments:
compare_code = reverse_condition_maybe_unordered (compare_code);
and
compare_code = reverse_condition (compare_code);
when only a few lines below that there is unconditional:
compare_code = UNKNOWN;
4) (not handled in the patch), diff uses HOST_WIDE_INT variable, signed, and
the
values can be arbitrary, so we can end up with undefined behavior in the
compiler if cf - ct overflows, or if diff is HWI minimum on diff = -diff etc.