On June 4, 2016 3:36:13 PM GMT+02:00, Jakub Jelinek <ja...@redhat.com> wrote:
>Hi!
>
>On this testcase execute_update_addresses_taken creates
>VIEW_CONVERT_EXPR of a {CLOBBER}, which is of course invalid.
>Instead, this patch just creates {CLOBBER} of the right type and let's
>the following ssa update transform that to replacing uses with default
>def
>or whatever else is appropriate.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
Thanks,
Richard.
>2016-06-04 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/71405
> * tree-ssa.c (execute_update_addresses_taken): For clobber with
> incompatible type, build a new clobber with the right type instead
> of building a VIEW_CONVERT_EXPR around it.
>
> * g++.dg/torture/pr71405.C: New test.
>
>--- gcc/tree-ssa.c.jj 2016-05-24 10:56:01.000000000 +0200
>+++ gcc/tree-ssa.c 2016-06-04 11:59:36.561302230 +0200
>@@ -1622,9 +1622,16 @@ execute_update_addresses_taken (void)
> if (gimple_assign_lhs (stmt) != lhs
> && !useless_type_conversion_p (TREE_TYPE (lhs),
> TREE_TYPE (rhs)))
>- rhs = fold_build1 (VIEW_CONVERT_EXPR,
>- TREE_TYPE (lhs), rhs);
>-
>+ {
>+ if (gimple_clobber_p (stmt))
>+ {
>+ rhs = build_constructor (TREE_TYPE (lhs), NULL);
>+ TREE_THIS_VOLATILE (rhs) = 1;
>+ }
>+ else
>+ rhs = fold_build1 (VIEW_CONVERT_EXPR,
>+ TREE_TYPE (lhs), rhs);
>+ }
> if (gimple_assign_lhs (stmt) != lhs)
> gimple_assign_set_lhs (stmt, lhs);
>
>--- gcc/testsuite/g++.dg/torture/pr71405.C.jj 2016-06-04
>12:10:33.694768153 +0200
>+++ gcc/testsuite/g++.dg/torture/pr71405.C 2016-06-04
>12:09:28.000000000 +0200
>@@ -0,0 +1,22 @@
>+// PR tree-optimization/71405
>+// { dg-do compile }
>+
>+struct C
>+{
>+ C () {}
>+ int i;
>+};
>+
>+void *
>+operator new (__SIZE_TYPE__ x, void *y)
>+{
>+ return y;
>+}
>+
>+int
>+main ()
>+{
>+ int a;
>+ new (&a) C;
>+ return a;
>+}
>
> Jakub
