Optimize a ^ ((a ^ b) & mask) to (~mask & a) | (b & mask).

From the perspective of the pipeline, `andn + and + ior` version take
2 cycles(AND and ANDN doesn't have dependence), but xor + and + xor
will take 3 cycles.

-       xorl    %edi, %esi
        andl    %edx, %esi
-       movl    %esi, %eax
-       xorl    %edi, %eax
+       andn    %edi, %edx, %eax
+       orl     %esi, %eax

gcc/ChangeLog:

	PR target/94790
	* config/i386/i386.md (*xor2andn): New define_insn_and_split.

gcc/testsuite/ChangeLog:

	PR target/94790
	* gcc.target/i386/pr94790-1.c: New test.
	* gcc.target/i386/pr94790-2.c: Ditto.

3 files changed, 61 insertions, 0 deletions

diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index 376df1d..9937643 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -10453,6 +10453,44 @@
    (set_attr "znver1_decode" "double")
    (set_attr "mode" "DI")])
 
+;; PR target/94790: Optimize a ^ ((a ^ b) & mask) to (~mask & a) | (b & mask)
+(define_insn_and_split "*xor2andn"
+  [(set (match_operand:SWI248 0 "nonimmediate_operand")
+	(xor:SWI248
+	  (and:SWI248
+	    (xor:SWI248
+	      (match_operand:SWI248 1 "nonimmediate_operand")
+	      (match_operand:SWI248 2 "nonimmediate_operand"))
+	    (match_operand:SWI248 3 "nonimmediate_operand"))
+	  (match_dup 1)))
+    (clobber (reg:CC FLAGS_REG))]
+  "(TARGET_BMI || TARGET_AVX512BW)
+   && ix86_pre_reload_split ()"
+  "#"
+  "&& 1"
+  [(parallel [(set (match_dup 4)
+		(and:SWI248
+		  (not:SWI248
+		    (match_dup 3))
+		  (match_dup 1)))
+	      (clobber (reg:CC FLAGS_REG))])
+   (parallel [(set (match_dup 5)
+		(and:SWI248
+		  (match_dup 2)
+		  (match_dup 3)))
+	      (clobber (reg:CC FLAGS_REG))])
+   (parallel [(set (match_dup 0)
+		(ior:SWI248
+		  (match_dup 4)
+		  (match_dup 5)))
+	      (clobber (reg:CC FLAGS_REG))])]
+{
+  operands[1] = force_reg (<MODE>mode, operands[1]);
+  operands[3] = force_reg (<MODE>mode, operands[3]);
+  operands[4] = gen_reg_rtx (<MODE>mode);
+  operands[5] = gen_reg_rtx (<MODE>mode);
+})
+
 ;; See comment for addsi_1_zext why we do use nonimmediate_operand
 (define_insn "*<code>si_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r")
diff --git a/gcc/testsuite/gcc.target/i386/pr94790-1.c b/gcc/testsuite/gcc.target/i386/pr94790-1.c
new file mode 100755
index 0000000..1ff5fa2
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr94790-1.c
@@ -0,0 +1,14 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -mbmi" } */
+/* { dg-final { scan-assembler-times "andn\[ \\t\]" 2 } } */
+/* { dg-final { scan-assembler-not "xorl\[ \\t\]" } } */
+
+unsigned r1(unsigned a, unsigned b, unsigned mask)
+{
+  return a ^ ((a ^ b) & mask);
+}
+
+unsigned r2(unsigned a, unsigned b, unsigned mask)
+{
+  return (~mask & a) | (b & mask);
+}
diff --git a/gcc/testsuite/gcc.target/i386/pr94790-2.c b/gcc/testsuite/gcc.target/i386/pr94790-2.c
new file mode 100755
index 0000000..ad11d94
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr94790-2.c
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -mbmi" } */
+/* { dg-final { scan-assembler-not "andn\[ \\t\]" } } */
+/* { dg-final { scan-assembler-times "xorl\[ \\t\]" 2 } } */
+
+unsigned r1(unsigned a, unsigned b, unsigned mask)
+{
+  return a ^ ((a ^ b) & mask) + (a ^ b);
+}