The following fixes PR81705, a simple omission in my last association 
patch.

Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.

Richard.

2017-08-04  Richard Biener  <rguent...@suse.de>

        PR middle-end/81705
        * fold-const.c (fold_binary_loc): Properly restrict
        minus_var0 && minus_var1 case when associating undefined overflow
        entities.

        * c-c++-common/ubsan/pr81705.c: New testcase.

Index: gcc/fold-const.c
===================================================================
--- gcc/fold-const.c    (revision 250865)
+++ gcc/fold-const.c    (working copy)
@@ -9592,12 +9592,13 @@ fold_binary_loc (location_t loc,
          if (POINTER_TYPE_P (atype)
              || (INTEGRAL_TYPE_P (atype) && !TYPE_OVERFLOW_WRAPS (atype)))
            {
-             if (var0 && var1)
+             if ((var0 && var1) || (minus_var0 && minus_var1))
                {
                  /* ???  If split_tree would handle NEGATE_EXPR we could
-                    simplify this down to the var0/minus_var1 cases.  */
-                 tree tmp0 = var0;
-                 tree tmp1 = var1;
+                    simply reject these cases and the allowed cases would
+                    be the var0/minus_var1 ones.  */
+                 tree tmp0 = var0 ? var0 : minus_var0;
+                 tree tmp1 = var1 ? var1 : minus_var1;
                  bool one_neg = false;
 
                  if (TREE_CODE (tmp0) == NEGATE_EXPR)
Index: gcc/testsuite/c-c++-common/ubsan/pr81705.c
===================================================================
--- gcc/testsuite/c-c++-common/ubsan/pr81705.c  (nonexistent)
+++ gcc/testsuite/c-c++-common/ubsan/pr81705.c  (working copy)
@@ -0,0 +1,12 @@
+/* { dg-do run } */
+/* { dg-options "-fsanitize=undefined -fsanitize-undefined-trap-on-error" } */
+
+int var_4 = -1716607962;
+int var_14 = 943738830;
+volatile int a;
+int main()
+{
+  //  (-(-1716607962) - 516151698) - -(9403738830)
+  a = (-var_4 - 516151698) - -var_14;
+  return 0;
+}

Reply via email to