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.
This commit is contained in:
Jakub Jelinek
2023-01-04 10:54:38 +01:00
parent 8692b15ae7
commit 345dffd0d4
2 changed files with 17 additions and 2 deletions

View File

@@ -40,6 +40,7 @@ along with GCC; see the file COPYING3. If not see
#include "tm.h"
#include "tree-eh.h"
#include "langhooks.h"
#include "tree-pass.h"
/* Routine to determine if the types T1 and T2 are effectively
the same for GENERIC. If T1 or T2 is not a type, the test
@@ -71,7 +72,7 @@ single_use (tree t ATTRIBUTE_UNUSED)
static inline bool
canonicalize_math_p ()
{
return true;
return !cfun || (cfun->curr_properties & PROP_gimple_opt_math) == 0;
}
/* Return true if math operations that are beneficial only after
@@ -90,7 +91,7 @@ canonicalize_math_after_vectorization_p ()
static inline bool
optimize_vectors_before_lowering_p ()
{
return true;
return !cfun || (cfun->curr_properties & PROP_gimple_lvec) == 0;
}
/* Return true if successive divisions can be optimized.

View File

@@ -0,0 +1,14 @@
/* PR middle-end/108237 */
typedef unsigned char __attribute__((__vector_size__ (1))) U;
typedef unsigned long long __attribute__((__vector_size__ (16))) V;
U u;
V v;
V
foo (void)
{
V w = v != ((unsigned char) ((unsigned char) u == u) & v);
return w;
}