generic-match-head: Don't assume GENERIC folding is done only early [PR108237]
authorJakub Jelinek <jakub@redhat.com>
Wed, 4 Jan 2023 09:54:38 +0000 (10:54 +0100)
committerJakub Jelinek <jakub@redhat.com>
Wed, 4 Jan 2023 09:54:38 +0000 (10:54 +0100)
commit345dffd0d4ebff7e705dfff1a8a72017a167120a
treed7e3dfdcb265446d2eb9d5180623ae79c60156e8
parent8692b15ae7c05e3224f285069e070c009d9f6efe
generic-match-head: Don't assume GENERIC folding is done only early [PR108237]

We ICE on the following testcase, because a valid V2DImode
!= comparison is folded into an unsupported V2DImode > comparison.
The match.pd pattern which does this looks like:
/* Transform comparisons of the form (X & Y) CMP 0 to X CMP2 Z
   where ~Y + 1 == pow2 and Z = ~Y.  */
(for cst (VECTOR_CST INTEGER_CST)
 (for cmp (eq ne)
      icmp (le gt)
  (simplify
   (cmp (bit_and:c@2 @0 cst@1) integer_zerop)
    (with { tree csts = bitmask_inv_cst_vector_p (@1); }
     (if (csts && (VECTOR_TYPE_P (TREE_TYPE (@1)) || single_use (@2)))
      (with { auto optab = VECTOR_TYPE_P (TREE_TYPE (@1))
                         ? optab_vector : optab_default;
              tree utype = unsigned_type_for (TREE_TYPE (@1)); }
       (if (target_supports_op_p (utype, icmp, optab)
            || (optimize_vectors_before_lowering_p ()
                && (!target_supports_op_p (type, cmp, optab)
                    || !target_supports_op_p (type, BIT_AND_EXPR, optab))))
        (if (TYPE_UNSIGNED (TREE_TYPE (@1)))
         (icmp @0 { csts; })
         (icmp (view_convert:utype @0) { csts; })))))))))
and that optimize_vectors_before_lowering_p () guarded stuff there
already deals with this problem, not trying to fold a supported comparison
into a non-supported one.  The reason it doesn't work in this case is that
it isn't GIMPLE folding which does this, but GENERIC folding done during
forwprop4 - forward_propagate_into_comparison -> forward_propagate_into_comparison_1
-> combine_cond_expr_cond -> fold_binary_loc -> generic_simplify
and we simply assumed that GENERIC folding happens only before
gimplification.

The following patch fixes that by checking cfun properties instead of
always returning true in those cases.

2023-01-04  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/108237
* generic-match-head.cc: Include tree-pass.h.
(canonicalize_math_p, optimize_vectors_before_lowering_p): Define
to false if cfun and cfun->curr_properties has PROP_gimple_opt_math
resp. PROP_gimple_lvec property set.

* gcc.c-torture/compile/pr108237.c: New test.
gcc/generic-match-head.cc
gcc/testsuite/gcc.c-torture/compile/pr108237.c [new file with mode: 0644]