https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105197

--- Comment #5 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Tamar Christina <tnfch...@gcc.gnu.org>:

https://gcc.gnu.org/g:78c718490bc2843d4dadcef8a0ae14aed1d15a32

commit r12-8080-g78c718490bc2843d4dadcef8a0ae14aed1d15a32
Author: Tamar Christina <tamar.christ...@arm.com>
Date:   Mon Apr 11 15:09:05 2022 +0100

    middle-end: Prevent the use of the cond inversion detection code when both
conditions are external. [PR105197]

    Previously ifcvt used to enforce that a mask A and the inverse of said mask
be
    represented as ~A. So for the masks

      _25 = _6 != 0;
      _44 = _4 != 0;

    ifcvt would produce for an operation requiring the inverse of said mask

      _26 = ~_25;
      _43 = ~_44;

    but now that VN is applied to the entire function body we get a
simplification
    on the mask and produce:

      _26 = _6 == 0;
      _43 = _4 == 0;

    This in itself is not a problem semantically speaking (though it does
create
    more masks that need to be tracked) but when vectorizing the masked
conditional
    we would still detect _26 and _43 to be inverses of _25 and _44 and mark
them
    as requiring their operands be swapped.

    When vectorizing we swap the operands but don't find the BIT_NOT_EXPR to
remove
    and so we leave the condition as is which produces invalid code:

    ------>vectorizing statement: _ifc__41 = _43 ? 0 : _ifc__40;
    created new init_stmt: vect_cst__136 = { 0, ... }
    add new stmt: _137 = mask__43.26_135 & loop_mask_111
    note:  add new stmt: vect__ifc__41.27_138 = VEC_COND_EXPR <_137,
vect__ifc__40.25_133, vect_cst__136>;

    This fixes disabling the inversion detection code when the loop isn't
masked
    since both conditional would be external.  We'd then not use the new
cond_code
    and would incorrectly still swap the operands.

    The resulting code is also better than GCC-11 with most operations now
    predicated on the loop mask rather than a ptrue.

    gcc/ChangeLog:

            PR target/105197
            * tree-vect-stmts.cc (vectorizable_condition): Prevent cond swap
when
            not masked.

    gcc/testsuite/ChangeLog:

            PR target/105197
            * gcc.target/aarch64/sve/pr105197-1.c: New test.
            * gcc.target/aarch64/sve/pr105197-2.c: New test.

Reply via email to