The most recent changes for pr108477 tripped regressions on the sh port:
Tests that now fail, but worked before (10 tests):
gcc: gcc.target/sh/pr54089-1.c scan-assembler-not and\t#1
gcc: gcc.target/sh/pr54089-1.c scan-assembler-times rotcr 25
gcc: gcc.target/sh/pr54089-1.c scan-assembler-times sett 1
gcc: gcc.target/sh/pr54089-1.c scan-assembler-times shll\t 1
gcc: gcc.target/sh/pr54089-7.c scan-assembler-not and
gcc: gcc.target/sh/pr54089-7.c scan-assembler-not movt
gcc: gcc.target/sh/pr54089-7.c scan-assembler-not rotr
gcc: gcc.target/sh/pr54089-7.c scan-assembler-times rotcr 4
gcc: gcc.target/sh/pr54089-8.c scan-assembler-times rotcl 28
gcc: gcc.target/sh/pr54089-9.c scan-assembler-times rotcl 4
I spot checked pr54089-1 and every case I looked at was better before
the change compared to after the change. So I chose the path of forcing
the AOP expander to use IOR. That fixes the regressions seen on
sh3/sh3eb. Those regressions would almost certainly pop on sh4/sh4eb,
but those haven't triggered in their weekly cycle yet.
Pushing to the trunk after verifying all the sh3/sh3eb regressions are
fixed.
Jeff
gcc/
* config/sh/sh.md (aopsi3): Define and route towards IOR.
diff --git a/gcc/config/sh/sh.md b/gcc/config/sh/sh.md
index 85980f31953b..5acca403608d 100644
--- a/gcc/config/sh/sh.md
+++ b/gcc/config/sh/sh.md
@@ -2776,6 +2776,19 @@ (define_insn "xorsi3"
"xor %2,%0"
[(set_attr "type" "arith")])
+;; When we can use PLUS, IOR, or XOR interchangably, use IOR.
+(define_expand "aopsi3"
+ [(set (match_operand:SI 0 "arith_reg_dest")
+ (ior:SI (match_operand:SI 1 "arith_reg_operand")
+ (match_operand:SI 2 "logical_operand")))]
+ ""
+ {
+ emit_insn (gen_iorsi3 (operands[0], operands[1], operands[2]));
+ DONE;
+ }
+)
+
+
;; The *logical_op_t pattern helps combine eliminating sign/zero extensions
;; of results where one of the inputs is a T bit store. Notice that this
;; pattern must not match during reload. If reload picks this pattern it
