Eliminate redundant compare between set{z,nz} and j{z,nz}:
setz %al; test %al,%al; jz <...> -> setz %al; jnz <...> and
setnz %al, test %al,%al; jz <...> -> setnz %al; jz <...>.

We can use the original Zero-flag value instead of setting the
temporary register and testing it for zero.

gcc/ChangeLog:

    * config/i386/i386.md (redundant compare peephole2):
    New peephole2 pattern.

Bootstrapped and regression tested on x86_64-linux-gnu {,-m32}.

Also tested by building and booting the linux kernel (where the
peephole2 triggers several times).

I was not able to craft the testcase, the original sequence is very
elusive and requires reuse of the value, set from the flags elsewhere
in the function in a very specific way.

Uros.
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index f83064ec335..e862368113c 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -18118,6 +18118,35 @@ (define_split
     FAIL;
 })
 
+;; Eliminate redundant compare between set{z,nz} and j{z,nz}:
+;; setz %al; test %al,%al; jz <...> -> setz %al; jnz <...> and
+;; setnz %al, test %al,%al; jz <...> -> setnz %al; jz <...>.
+(define_peephole2
+  [(set (match_operand:QI 0 "nonimmediate_operand")
+       (match_operator:QI 1 "bt_comparison_operator"
+         [(reg:CCZ FLAGS_REG) (const_int 0)]))
+   (set (reg:CCZ FLAGS_REG)
+       (compare:CCZ (match_dup 0) (const_int 0)))
+   (set (pc)
+       (if_then_else (match_operator 2 "bt_comparison_operator"
+                       [(reg:CCZ FLAGS_REG) (const_int 0)])
+                     (match_operand 3)
+                     (pc)))]
+  "peep2_regno_dead_p (3, FLAGS_REG)"
+  [(set (match_dup 0)
+       (match_op_dup 1 [(reg:CCZ FLAGS_REG) (const_int 0)]))
+   (set (pc)
+       (if_then_else (match_dup 2)
+                     (match_dup 3)
+                     (pc)))]
+{
+  if (GET_CODE (operands[1]) == EQ)
+    {
+      operands[2] = shallow_copy_rtx (operands[2]);
+      PUT_CODE (operands[2], reverse_condition (GET_CODE (operands[2])));
+    }
+})
+
 ;; The SSE store flag instructions saves 0 or 0xffffffff to the result.
 ;; subsequent logical operations are used to imitate conditional moves.
 ;; 0xffffffff is NaN, but not in normalized form, so we can't represent

Reply via email to