]> git.ipfire.org Git - thirdparty/gcc.git/commitdiff
match.pd: Add missing type check to reduc(ctor) pattern [PR121772]
authorAlex Coplan <alex.coplan@arm.com>
Tue, 9 Sep 2025 11:57:14 +0000 (12:57 +0100)
committerAlex Coplan <alex.coplan@arm.com>
Thu, 25 Sep 2025 13:36:02 +0000 (14:36 +0100)
In this PR we have a reduction of a vector constructor, where the
type of the constructor is int16x8_t and the elements are int16x4_t;
i.e. it is representing a concatenation of two vectors.

This triggers a match.pd pattern which looks like it was written to
handle reductions of vector constructors where the elements of the ctor
are scalars, not vectors.  There is no type check to enforce this
property, which leads to the pattern replacing a reduction to scalar
with an int16x4_t vector in this case, which of course is a type error,
leading to an invalid GIMPLE ICE.

This patch adds a type check to the pattern, only going ahead with the
transformation if the element type of the ctor matches that of the
reduction.

gcc/ChangeLog:

PR tree-optimization/121772
* match.pd: Add type check to reduc(ctor) pattern.

gcc/testsuite/ChangeLog:

PR tree-optimization/121772
* gcc.target/aarch64/torture/pr121772.c: New test.

(cherry picked from commit a7a9b7badc0ba95b510c7e61da6439fca78e31d3)

gcc/match.pd
gcc/testsuite/gcc.target/aarch64/torture/pr121772.c [new file with mode: 0644]

index ba036e528370a19ebc76aaa41f28eb082bdb2515..3d1752055b347e15d3d25ed9f1c1c376cf006905 100644 (file)
@@ -10826,6 +10826,7 @@ and,
                         ? gimple_assign_rhs1 (SSA_NAME_DEF_STMT (@0)) : @0);
            tree elt = ctor_single_nonzero_element (ctor); }
       (if (elt
+          && types_match (type, TREE_TYPE (elt))
           && !HONOR_SNANS (type)
           && !HONOR_SIGNED_ZEROS (type))
        { elt; }))))
diff --git a/gcc/testsuite/gcc.target/aarch64/torture/pr121772.c b/gcc/testsuite/gcc.target/aarch64/torture/pr121772.c
new file mode 100644 (file)
index 0000000..3b4cf4d
--- /dev/null
@@ -0,0 +1,5 @@
+/* { dg-do compile } */
+#include <arm_neon.h>
+int16_t f(int16x4_t b) {
+  return vaddvq_s16(vcombine_s16(b, vdup_n_s16 (0)));
+}