The conversion of a move pattern where both operands are AX_REG
should be prevented.

2022-09-01  Uroš Bizjak  <ubiz...@gmail.com>

gcc/ChangeLog:

    PR target/106707
    * config/i386/i386.md (moves to/from AX_REG into xchg peephole2):
    Do not convert a move pattern where both operands are AX_REG.

gcc/testsuite/ChangeLog:

    PR target/106707
    * gcc.target/i386/pr106707.c: New test.

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

Pushed to master.

Uros.
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index a4a18cf89f5..1aef1af594d 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -3043,8 +3043,8 @@ (define_peephole2
   [(set (match_operand:SWI48 0 "general_reg_operand")
        (match_operand:SWI48 1 "general_reg_operand"))]
  "optimize_size > 1
-  && (REGNO (operands[0]) == AX_REG
-      || REGNO (operands[1]) == AX_REG)
+  && ((REGNO (operands[0]) == AX_REG)
+      != (REGNO (operands[1]) == AX_REG))
   && optimize_insn_for_size_p ()
   && peep2_reg_dead_p (1, operands[1])"
   [(parallel [(set (match_dup 0) (match_dup 1))
diff --git a/gcc/testsuite/gcc.target/i386/pr106707.c 
b/gcc/testsuite/gcc.target/i386/pr106707.c
new file mode 100644
index 00000000000..a127ccd4679
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr106707.c
@@ -0,0 +1,19 @@
+/* PR target/106707 */
+/* { dg-do compile { target int128 } } */
+/* { dg-options "-Oz -g -fno-cprop-registers -fno-dce" } */
+
+typedef unsigned __attribute__((__vector_size__ (8))) V;
+
+unsigned __int128 ii;
+unsigned x, y;
+
+V v;
+
+void
+foo (long a)
+{
+  long l = a != x;
+  int i = __builtin_add_overflow_p (y * ii, 0, 0);
+  V u = ii < x | v, w = x <= u < i & y <= x / ii;
+  v = __builtin_shufflevector (v, w, 1, 2) + (V) l;
+}

Reply via email to