[Bug tree-optimization/108692] [11/12/13 Regression] Miscompilation of orc_test.c since r11-5160

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108692

--- Comment #4 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:6ad1c1027628f094260037536f6b6fcdb63b5add

commit r13-5742-g6ad1c1027628f094260037536f6b6fcdb63b5add
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Wed Feb 8 18:41:21 2023 +0100

    vect-patterns: Fix up vect_widened_op_tree [PR108692]

    The following testcase is miscompiled on aarch64-linux since r11-5160.
    Given
      <bb 3> [local count: 955630225]:
      # i_22 = PHI <i_20(6), 0(5)>
      # r_23 = PHI <r_19(6), 0(5)>
    ...
      a.0_5 = (unsigned char) a_15;
      _6 = (int) a.0_5;
      b.1_7 = (unsigned char) b_17;
      _8 = (int) b.1_7;
      c_18 = _6 - _8;
      _9 = ABS_EXPR <c_18>;
      r_19 = _9 + r_23;
    ...
    where SSA_NAMEs 15/17 have signed char, 5/7 unsigned char and rest is int
    we first pattern recognize c_18 as
    patt_34 = (a.0_5) w- (b.1_7);
    which is still correct, 5/7 are unsigned char subtracted in wider type,
    but then vect_recog_sad_pattern turns it into
    SAD_EXPR <a_15, b_17, r_23>
    which is incorrect, because 15/17 are signed char and so it is
    sum of absolute signed differences rather than unsigned sum of
    absolute unsigned differences.
    The reason why this happens is that vect_recog_sad_pattern calls
    vect_widened_op_tree with MINUS_EXPR, WIDEN_MINUS_EXPR on the
    patt_34 = (a.0_5) w- (b.1_7); statement's vinfo and vect_widened_op_tree
    calls vect_look_through_possible_promotion on the operands of the
    WIDEN_MINUS_EXPR, which looks through the further casts.
    vect_look_through_possible_promotion has careful code to stop when there
    would be nested casts that need to be preserved, but the problem here
    is that the WIDEN_*_EXPR operation itself has an implicit cast on the
    operands already - in this case of WIDEN_MINUS_EXPR the unsigned char
    5/7 SSA_NAMEs are widened to unsigned short before the subtraction,
    and vect_look_through_possible_promotion obviously isn't told about that.

    Now, I think when we see those WIDEN_{MULT,MINUS,PLUS}_EXPR codes, we had
    to look through possible promotions already when creating those and so
    vect_look_through_possible_promotion again isn't really needed, all we need
    to do is arrange what that function will do if the operand isn't result
    of any cast.  Other option would be let
vect_look_through_possible_promotion
    know about the implicit promotion from the WIDEN_*_EXPR, but I'm afraid
    that would be much harder.

    2023-02-08  Jakub Jelinek  <jakub@redhat.com>

            PR tree-optimization/108692
            * tree-vect-patterns.cc (vect_widened_op_tree): If rhs_code is
            widened_code which is different from code, don't call
            vect_look_through_possible_promotion but instead just check op is
            SSA_NAME with integral type for which vect_is_simple_use is true
            and call set_op on this_unprom.

            * gcc.dg/pr108692.c: New test.