On June 12, 2015 1:17:14 PM GMT+02:00, Jakub Jelinek <ja...@redhat.com> wrote:
>Hi!
>
>The following testcase ICEs in 4.8/4.9.  The problem is that we don't
>fold
>NON_LVALUE_EXPR of INTEGER_CST into the INTEGER_CST, but it seems huge
>amount of fold-const.c code has the assumption that if
>argN = opN;
>...
>STRIP_NOPS (argN);
>...
>if (TREE_CODE (argN) == INTEGER_CST)
>
>then argN == opN, so passes e.g. argN instead of opN, assuming that
>argN must have type already.  Eric has fixed this for 5+ already
>by dropping the NON_LVALUE_EXPR if the argument is isn't a possible
>lvalue.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux on both 4.9 and
>4.8
>branches with --enable-checking=yes,rtl, ok for those branches?
>Is the testcase alone ok for 5.2/trunk?

OK and yes.

Thanks,
Richard.

>A safer alternative would be just to fold NON_LVALUE_EXPR of
>INTEGER_CST
>into the INTEGER_CST.
>
>2015-06-12  Jakub Jelinek  <ja...@redhat.com>
>
>       PR middle-end/63608
>       * gcc.c-torture/compile/pr63608.c: New test.
>
>       Backported from mainline
>       2014-05-16  Eric Botcazou  <ebotca...@adacore.com>
>
>       * fold-const (fold_unary_loc) <NON_LVALUE_EXPR>: New case.
>       <CASE_CONVERT>: Pass arg0 instead of op0 to fold_convert_const.
>
>--- gcc/fold-const.c   (revision 210517)
>+++ gcc/fold-const.c   (revision 210518)
>@@ -7850,6 +7850,11 @@ fold_unary_loc (location_t loc, enum tre
>       return fold_convert_loc (loc, type, op0);
>       return NULL_TREE;
> 
>+    case NON_LVALUE_EXPR:
>+      if (!maybe_lvalue_p (op0))
>+      return fold_convert_loc (loc, type, op0);
>+      return NULL_TREE;
>+
>     CASE_CONVERT:
>     case FLOAT_EXPR:
>     case FIX_TRUNC_EXPR:
>@@ -8113,7 +8118,7 @@ fold_unary_loc (location_t loc, enum tre
>           }
>       }
> 
>-      tem = fold_convert_const (code, type, op0);
>+      tem = fold_convert_const (code, type, arg0);
>       return tem ? tem : NULL_TREE;
> 
>     case ADDR_SPACE_CONVERT_EXPR:
>--- gcc/testsuite/gcc.c-torture/compile/pr63608.c.jj   2015-06-12
>10:55:16.484746890 +0200
>+++ gcc/testsuite/gcc.c-torture/compile/pr63608.c      2015-06-12
>10:53:58.000000000 +0200
>@@ -0,0 +1,14 @@
>+/* PR middle-end/63608 */
>+
>+typedef long T;
>+typedef unsigned long U;
>+unsigned long a;
>+
>+unsigned long
>+foo (int b)
>+{
>+  T c = 0;
>+  const U d = 2248593032UL;
>+  a = (c = +d) | (~4L & ~b);
>+  return c;
>+}
>
>       Jakub


Reply via email to