Hi!
As the testcase shows, while reassociation of vector multiplication
has been enabled, the eliminate_using_constants optimization
isn't enabled for integral vectors (only for floating point vectors),
so e.g. if we have more than one { ~0, ~0, ... } vectors in there
coming from NEGATE_EXPRs, we don't actually fold them into a single
{ ~0, ~0, ... } or { 1, 1, ... } vector (the former then into
NEGATE_EXPR on the result, the latter removed altogether).
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2016-10-31 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/77860
* tree-ssa-reassoc.c (eliminate_using_constants): Handle
also integral complex and vector constants.
* gcc.dg/pr77860.c: New test.
--- gcc/tree-ssa-reassoc.c.jj 2016-10-31 13:28:12.000000000 +0100
+++ gcc/tree-ssa-reassoc.c 2016-10-31 13:47:00.364216248 +0100
@@ -924,7 +924,7 @@ eliminate_using_constants (enum tree_cod
tree type = TREE_TYPE (oelast->op);
if (oelast->rank == 0
- && (INTEGRAL_TYPE_P (type) || FLOAT_TYPE_P (type)))
+ && (ANY_INTEGRAL_TYPE_P (type) || FLOAT_TYPE_P (type)))
{
switch (opcode)
{
--- gcc/testsuite/gcc.dg/pr77860.c.jj 2016-10-31 13:50:34.019649814 +0100
+++ gcc/testsuite/gcc.dg/pr77860.c 2016-10-31 13:49:35.000000000 +0100
@@ -0,0 +1,13 @@
+/* PR tree-optimization/77860 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fno-tree-vrp -fno-tree-forwprop -Wno-psabi" } */
+
+typedef unsigned short V __attribute__((vector_size (16)));
+
+V
+foo (V x, V y)
+{
+ V a = -x;
+ V b = -y;
+ return a * b;
+}
Jakub
