Hi!
The following testcase ICEs since r15-1671, because the match.pd pattern
now allows a cast and the function checks whether the ifn is supported
on a wrong type. .SAT_{ADD,SUB,MUL} are binary ifns, so they care about
the type of their first operand:
#define binary_direct { 0, 0, true }
where
/* optabs can be parameterized by one or two modes. These fields describe
how to select those modes from the types of the return value and
arguments. A value of -1 says that the mode is determined by the
return type while a value N >= 0 says that the mode is determined by
the type of argument N. A value of -2 says that this internal
function isn't directly mapped to an optab. */
but in this function (unlike the function right below it for the
same ifns) checks the type of the lhs which since that change can be
actually a different type (expansion performs the operation on the
argument types and then casts the result to the lhs type).
So, e.g. on x86_64 -m32, it checks wether ussubsi3 insn can be used
(which it can), but then actually uses it on DImode arguments and
ussubdi3 is TARGET_64BIT only. Similarly for -m64 it checks ussubsi3 too
instead of ussubti3 (which doesn't exist).
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk
and after a while for 15.3?
2026-01-03 Jakub Jelinek <[email protected]>
PR tree-optimization/123372
* tree-ssa-math-opts.cc
(build_saturation_binary_arith_call_and_replace): Pass type of op_0
rather than type of lhs as second argument to
direct_internal_fn_supported_p.
* gcc.dg/tree-ssa/pr123372.c: New test.
--- gcc/tree-ssa-math-opts.cc.jj 2026-01-02 09:56:10.390332748 +0100
+++ gcc/tree-ssa-math-opts.cc 2026-01-02 18:45:57.087546705 +0100
@@ -4083,7 +4083,7 @@ build_saturation_binary_arith_call_and_r
internal_fn fn, tree lhs,
tree op_0, tree op_1)
{
- if (direct_internal_fn_supported_p (fn, TREE_TYPE (lhs), OPTIMIZE_FOR_BOTH))
+ if (direct_internal_fn_supported_p (fn, TREE_TYPE (op_0), OPTIMIZE_FOR_BOTH))
{
gcall *call = gimple_build_call_internal (fn, 2, op_0, op_1);
gimple_call_set_lhs (call, lhs);
--- gcc/testsuite/gcc.dg/tree-ssa/pr123372.c.jj 2026-01-02 18:49:45.419717435
+0100
+++ gcc/testsuite/gcc.dg/tree-ssa/pr123372.c 2026-01-02 18:51:22.256093427
+0100
@@ -0,0 +1,25 @@
+/* PR tree-optimization/123372 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fgimple" } */
+
+#ifdef __SIZEOF_INT128__
+#define T unsigned __int128
+#else
+#define T unsigned long long
+#endif
+
+unsigned int __GIMPLE (ssa,startwith("phiopt4"))
+foo (T a, T b)
+{
+ T _4;
+ unsigned int _1;
+ bool _5;
+ unsigned int _6;
+
+ __BB(2):
+ _5 = a_2(D) >= b_3(D);
+ _4 = a_2(D) - b_3(D);
+ _6 = (unsigned int) _4;
+ _1 = _5 ? _6 : 0U;
+ return _1;
+}
Jakub
