Hi!

The following testcase is miscompiled, because it optimizes
(x & -128) ? -128 : 0 to (x & -128) when it shouldn't.  The problem is if
the type of the COND_EXPR has smaller precision than the BIT_AND_EXPR in the
test, we verify that integer_pow2p (arg1), which is true in this case (-128
in signed char is a power of two), and then operand_equal_p between that
value and the constant in BIT_AND_EXPR's second argument (operand_equal_p
compares only the value, and -128 == -128).  The following patch verifies
also that the BIT_AND_EXPR's second operand is a power of two.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk and
release branches?

2018-09-12  Jakub Jelinek  <ja...@redhat.com>

        PR middle-end/87248
        * fold-const.c (fold_ternary_loc) <case COND_EXPR>: Verify also that
        BIT_AND_EXPR's second operand is a power of two.  Formatting fix.

        * c-c++-common/torture/pr87248.c: New test.

--- gcc/fold-const.c.jj 2018-09-06 09:41:59.000000000 +0200
+++ gcc/fold-const.c    2018-09-08 00:12:28.332418784 +0200
@@ -11607,10 +11607,16 @@ fold_ternary_loc (location_t loc, enum t
          && integer_pow2p (arg1)
          && TREE_CODE (TREE_OPERAND (arg0, 0)) == BIT_AND_EXPR
          && operand_equal_p (TREE_OPERAND (TREE_OPERAND (arg0, 0), 1),
-                             arg1, OEP_ONLY_CONST))
+                             arg1, OEP_ONLY_CONST)
+         /* operand_equal_p compares just value, not precision, so e.g.
+            arg1 could be 8-bit -128 and be power of two, but BIT_AND_EXPR
+            second operand 32-bit -128, which is not a power of two (or vice
+            versa.  */
+         && integer_pow2p (TREE_OPERAND (TREE_OPERAND (arg0, 0), 1)))
        return pedantic_non_lvalue_loc (loc,
-                                   fold_convert_loc (loc, type,
-                                                     TREE_OPERAND (arg0, 0)));
+                                       fold_convert_loc (loc, type,
+                                                         TREE_OPERAND (arg0,
+                                                                       0)));
 
       /* Disable the transformations below for vectors, since
         fold_binary_op_with_conditional_arg may undo them immediately,
--- gcc/testsuite/c-c++-common/torture/pr87248.c.jj     2018-09-08 
01:13:43.431334239 +0200
+++ gcc/testsuite/c-c++-common/torture/pr87248.c        2018-09-08 
01:14:55.446593520 +0200
@@ -0,0 +1,36 @@
+/* PR middle-end/87248 */
+/* { dg-do run } */
+
+void
+foo (signed char *p, int q)
+{
+  *p = q & (-__SCHAR_MAX__ - 1) ? (-__SCHAR_MAX__ - 1) : 0;
+}
+
+int
+bar (long long x)
+{
+  return x & (-__INT_MAX__ - 1) ? (-__INT_MAX__ - 1) : 0;
+}
+
+int
+main ()
+{
+#if __INT_MAX__ > 4 * __SCHAR_MAX__
+  signed char a[4];
+  foo (a, __SCHAR_MAX__ + 1U);
+  foo (a + 1, 2 * (__SCHAR_MAX__ + 1U));
+  foo (a + 2, -__INT_MAX__ - 1);
+  foo (a + 3, (__SCHAR_MAX__ + 1U) / 2);
+  if (a[0] != (-__SCHAR_MAX__ - 1) || a[1] != a[0] || a[2] != a[0] || a[3] != 
0)
+    __builtin_abort ();
+#endif
+#if __LONG_LONG_MAX__ > 4 * __INT_MAX__
+  if (bar (__INT_MAX__ + 1LL) != (-__INT_MAX__ - 1)
+      || bar (2 * (__INT_MAX__ + 1LL)) != (-__INT_MAX__ - 1)
+      || bar (-__LONG_LONG_MAX__ - 1) != (-__INT_MAX__ - 1)
+      || bar ((__INT_MAX__ + 1LL) / 2) != 0)
+    __builtin_abort ();
+#endif
+  return 0;
+}

        Jakub

Reply via email to