reassoc has certain tricks which in the end depend on the ability
to undo them. For DFP creating a -1. constant is easy but
re-identifying is appearantly not - real_minus_onep rejects those
outright for DFP. So we have to disable (at least) this one trick.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
Richard.
2022-04-13 Richard Biener <rguent...@suse.de>
PR tree-optimization/105263
* tree-ssa-reassoc.cc (try_special_add_to_ops): Do not consume
negates in multiplication chains with DFP.
* gcc.dg/pr105263.c: New testcase.
---
gcc/testsuite/gcc.dg/pr105263.c | 15 +++++++++++++++
gcc/tree-ssa-reassoc.cc | 4 +++-
2 files changed, 18 insertions(+), 1 deletion(-)
create mode 100644 gcc/testsuite/gcc.dg/pr105263.c
diff --git a/gcc/testsuite/gcc.dg/pr105263.c b/gcc/testsuite/gcc.dg/pr105263.c
new file mode 100644
index 00000000000..5cb7fcd09a2
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr105263.c
@@ -0,0 +1,15 @@
+/* { dg-do compile { target dfp } } */
+/* { dg-options "-O -ffast-math -w -Wno-psabi" } */
+
+typedef _Decimal64 __attribute__((__vector_size__ (8))) U;
+typedef _Decimal64 __attribute__((__vector_size__ (16))) V;
+
+V v;
+
+U
+foo (U u)
+{
+ u *= u;
+ u *= -(U){ v[1] };
+ return u;
+}
diff --git a/gcc/tree-ssa-reassoc.cc b/gcc/tree-ssa-reassoc.cc
index 4ab3c3361f9..f41b3c9a004 100644
--- a/gcc/tree-ssa-reassoc.cc
+++ b/gcc/tree-ssa-reassoc.cc
@@ -5857,7 +5857,9 @@ try_special_add_to_ops (vec<operand_entry *> *ops,
&& gimple_assign_rhs_code (def_stmt) == NEGATE_EXPR
&& !HONOR_SNANS (TREE_TYPE (op))
&& (!HONOR_SIGNED_ZEROS (TREE_TYPE (op))
- || !COMPLEX_FLOAT_TYPE_P (TREE_TYPE (op))))
+ || !COMPLEX_FLOAT_TYPE_P (TREE_TYPE (op)))
+ && (!FLOAT_TYPE_P (TREE_TYPE (op))
+ || !DECIMAL_FLOAT_MODE_P (element_mode (op))))
{
tree rhs1 = gimple_assign_rhs1 (def_stmt);
tree cst = build_minus_one_cst (TREE_TYPE (op));
--
2.34.1
