This reverts the reversion of r10-5979 amending the CTOR case
with a comment as to why the conversion is not necessary there.
It also adds a testcase (but not for the CTOR case).
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
2022-03-07 Richard Biener <rguent...@suse.de>
PR tree-optimization/104782
* tree-vect-slp.cc (vectorize_slp_instance_root_stmt):
Re-instantiate r10-5979 fix, add comment.
* gcc.dg/vect/pr104782.c: New testcase.
---
gcc/testsuite/gcc.dg/vect/pr104782.c | 18 ++++++++++++++++++
gcc/tree-vect-slp.cc | 7 +++++++
2 files changed, 25 insertions(+)
create mode 100644 gcc/testsuite/gcc.dg/vect/pr104782.c
diff --git a/gcc/testsuite/gcc.dg/vect/pr104782.c
b/gcc/testsuite/gcc.dg/vect/pr104782.c
new file mode 100644
index 00000000000..7b8ca6ca25b
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr104782.c
@@ -0,0 +1,18 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-O3" } */
+/* { dg-additional-options "-march=armv8.2-a+sve -msve-vector-bits=128" {
target aarch64-*-* } } */
+
+int a, b, c;
+static int d;
+short *q;
+void f() {
+ int *p = &d;
+ b = 9;
+ for (b = 9; b; b--) {
+ a = 2;
+ for (c = 2; c <= 9; c++) {
+ for (int i = 0; i < 3; i++)
+ *p |= (*q)++;
+ }
+ }
+}
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index f9bbc872a99..4ac2b70303c 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -7382,6 +7382,10 @@ vectorize_slp_instance_root_stmt (slp_tree node,
slp_instance instance)
gimple *child_stmt = SLP_TREE_VEC_STMTS (node)[0];
tree vect_lhs = gimple_get_lhs (child_stmt);
tree root_lhs = gimple_get_lhs (instance->root_stmts[0]->stmt);
+ if (!useless_type_conversion_p (TREE_TYPE (root_lhs),
+ TREE_TYPE (vect_lhs)))
+ vect_lhs = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (root_lhs),
+ vect_lhs);
rstmt = gimple_build_assign (root_lhs, vect_lhs);
}
else if (SLP_TREE_NUMBER_OF_VEC_STMTS (node) > 1)
@@ -7392,6 +7396,9 @@ vectorize_slp_instance_root_stmt (slp_tree node,
slp_instance instance)
vec<constructor_elt, va_gc> *v;
vec_alloc (v, nelts);
+ /* A CTOR can handle V16HI composition from VNx8HI so we
+ do not need to convert vector elements if the types
+ do not match. */
FOR_EACH_VEC_ELT (SLP_TREE_VEC_STMTS (node), j, child_stmt)
CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
gimple_get_lhs (child_stmt));
--
2.34.1
