This fixes minor fallout from the IOR->ADD change for rotates that I
installed a little while ago.
Basically the SH backend has a special pattern for setting the T
register that has elements similar to a rotate. With the IOR->ADD
change that pattern no longer matches and we get scan-asm failures.
Fixing isn't a trivial case of just replacing IOR with ADD as the
IOR->ADD change changes some of the simplifications/canonicalizations
along the way.
The net is we need a pattern with a slightly different structure. I've
regression tested this on sh3[eb]-linux-gnu and bootstrapped
sh4-linux-gnu (without a regression test).
Pushing to the trunk,
Jeff
gcc/
* config/sh/sh.md (neg_zero_extract_4b): New pattern.
diff --git a/gcc/config/sh/sh.md b/gcc/config/sh/sh.md
index 19481c07db6..92a1efeb811 100644
--- a/gcc/config/sh/sh.md
+++ b/gcc/config/sh/sh.md
@@ -10093,6 +10093,25 @@ (define_insn_and_split "*neg_zero_extract_4"
[(set (reg:SI T_REG) (zero_extract:SI (xor:SI (match_dup 0) (match_dup 1))
(const_int 1) (match_dup 2)))])
+;; Same thing, but when we use a PLUS rather than IOR/XOR for the rotation
+;; which causes things to simplify somewhat differently.
+(define_insn_and_split "*neg_zero_extract_4b"
+ [(set (reg:SI T_REG)
+ (and:SI (not:SI (plus:SI
+ (lshiftrt:SI (match_operand:SI 0 "arith_reg_operand")
+ (match_operand 1 "const_int_operand"))
+ (ashift:SI (match_operand:SI 2 "arith_reg_operand")
+ (match_operand 3 "const_int_operand"))))
+ (const_int 1)))]
+ "TARGET_SH1 && can_create_pseudo_p ()
+ && INTVAL (operands[3]) > 0
+ && INTVAL (operands[1]) + INTVAL (operands[3]) == 32"
+ "#"
+ "&& 1"
+ [(set (reg:SI T_REG) (zero_extract:SI (xor:SI (match_dup 0) (match_dup 4))
+ (const_int 1) (match_dup 1)))]
+ { operands[4] = GEN_INT (1 << INTVAL (operands[1])); })
+
(define_insn_and_split "*neg_zero_extract_5"
[(set (reg:SI T_REG)
(and:SI (not:SI (subreg:SI
