On Mon, Dec 02, 2019 at 08:28:27AM +0100, Uros Bizjak wrote:
> > I'll have a look tomorrow.
>
> general_reg_operand should be used in the pattern.
That indeed works, ok for trunk if it passes bootstrap/regtest?
2019-12-02 Uroš Bizjak <ubiz...@gmail.com>
Jakub Jelinek <ja...@redhat.com>
PR target/92744
* config/i386/i386.md (peephole2 for *swap<mode>): Use
general_reg_operand predicates instead of register_operand.
* g++.dg/dfp/pr92744.C: New test.
--- gcc/config/i386/i386.md.jj 2019-11-19 10:28:18.187668272 +0100
+++ gcc/config/i386/i386.md 2019-12-02 09:55:04.989644441 +0100
@@ -2788,10 +2788,10 @@ (define_insn "*swap<mode>"
(set_attr "bdver1_decode" "double")])
(define_peephole2
- [(set (match_operand:SWI 0 "register_operand")
- (match_operand:SWI 1 "register_operand"))
+ [(set (match_operand:SWI 0 "general_reg_operand")
+ (match_operand:SWI 1 "general_reg_operand"))
(set (match_dup 1)
- (match_operand:SWI 2 "register_operand"))
+ (match_operand:SWI 2 "general_reg_operand"))
(set (match_dup 2) (match_dup 0))]
"peep2_reg_dead_p (3, operands[0])
&& optimize_insn_for_size_p ()"
--- gcc/testsuite/g++.dg/dfp/pr92744.C.jj 2019-12-02 10:01:04.133064486
+0100
+++ gcc/testsuite/g++.dg/dfp/pr92744.C 2019-12-02 10:00:31.577571093 +0100
@@ -0,0 +1,20 @@
+// PR target/92744
+// { dg-do compile }
+// { dg-options "-Os -fno-tree-ccp" }
+
+typedef float T __attribute__((mode(SD)));
+struct A { T a; };
+void foo ();
+
+bool
+operator!= (A x, A y)
+{
+ return x.a != y.a;
+}
+
+void
+bar (A x, A y)
+{
+ if (x != y)
+ foo ();
+}
Jakub
