[PATCH] Fix neg overflow expansion (PR middle-end/77594)

[Jakub Jelinek]

Hi!

As mentioned in the PR, during expand_arith_overflow expansion
expand_neg_overflow falls through into expand_addsub_overflow, so expands
roughly the same code twice and keeps perhaps the less efficient version
around.

Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?

2016-09-16  Jakub Jelinek  <ja...@redhat.com>
            Eric Botcazou  <ebotca...@adacore.com>

        PR middle-end/77594
        * internal-fn.c (expand_arith_overflow) <case MINUS_EXPR>: Don't fall
        through into expand_addsub_overflow after expand_neg_overflow.

        * gcc.target/i386/pr77594.c: New test.

--- gcc/internal-fn.c.jj        2016-08-29 12:17:14.000000000 +0200
+++ gcc/internal-fn.c   2016-09-15 09:42:09.693980531 +0200
@@ -1833,7 +1833,10 @@ expand_arith_overflow (enum tree_code co
            {
            case MINUS_EXPR:
              if (integer_zerop (arg0) && !unsr_p)
-               expand_neg_overflow (loc, lhs, arg1, false);
+               {
+                 expand_neg_overflow (loc, lhs, arg1, false);
+                 return;
+               }
              /* FALLTHRU */
            case PLUS_EXPR:
              expand_addsub_overflow (loc, code, lhs, arg0, arg1,
--- gcc/testsuite/gcc.target/i386/pr77594.c.jj  2016-09-15 09:44:51.879910346 
+0200
+++ gcc/testsuite/gcc.target/i386/pr77594.c     2016-09-15 09:44:43.000000000 
+0200
@@ -0,0 +1,11 @@
+/* PR middle-end/77594 */
+/* { dg-do compile } */
+/* { dg-options "-O0" } */
+
+int
+foo (int a, int *b)
+{
+  return __builtin_sub_overflow (0, a, b);
+}
+
+/* { dg-final { scan-assembler-times "\tjn?o\t" 1 } } */

        Jakub