This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Revert part of my commit from yesterday (fix PR c++/20023, tree-optimization/20009)
- From: Jakub Jelinek <jakub at redhat dot com>
- To: gcc-patches at gcc dot gnu dot org
- Date: Thu, 17 Feb 2005 09:53:41 -0500
- Subject: [PATCH] Revert part of my commit from yesterday (fix PR c++/20023, tree-optimization/20009)
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
Hi!
The convert.c part of the patch to PR middle-end/19857 fix was not
strictly necessary and although it did not cause any make check regressions on
any of the 7 arches I used to bootstrap/regtest, it apparently caused some
failures.
E.g. in c++/20023 case, the difference is that the switch condition
used to look like:
<nop_expr (size_t) <convert_expr (int) <param_decl (pointer) >>>
which folds into itself and therefore expand_case sees an integer
SWITCH_COND. But with my convert.c change convert_to_integer
returns <nop_expr (size_t) <convert_expr (unsigned int) <param_decl (pointer) >>>
and as fold strips sign nops, it folds this into
<nop_expr (size_t) <param_decl (pointer) >>.
I believe it is not wrong even at this point, converting from pointer
to size_t does not at least on i386-*-linux* need any runtime code and
therefore I think a NOP_EXPR is appropriate. But e.g. expand_case
is not prepared to see POINTER_TYPE's after stripping nops.
Given how close we are to 4.0 branching, I think best is to revert
this part of my patch and perhaps look into this in the future.
Ok?
2005-02-17 Jakub Jelinek <jakub@redhat.com>
PR c++/20023
PR tree-optimization/20009
* convert.c (convert_to_integer): Revert 2005-02-16 change.
* gcc.c-torture/compile/20050217-1.c: New test.
* g++.dg/opt/switch3.C: New test.
--- gcc/convert.c.jj 2005-02-16 14:52:36.000000000 +0100
+++ gcc/convert.c 2005-02-17 15:38:03.307819471 +0100
@@ -387,8 +387,7 @@ convert_to_integer (tree type, tree expr
expr = integer_zero_node;
else
expr = fold (build1 (CONVERT_EXPR,
- lang_hooks.types.type_for_size
- (POINTER_SIZE, TYPE_UNSIGNED (type)),
+ lang_hooks.types.type_for_size (POINTER_SIZE, 0),
expr));
return convert_to_integer (type, expr);
--- gcc/testsuite/gcc.c-torture/compile/20050217-1.c 2005-01-27 14:27:08.338732320 +0100
+++ gcc/testsuite/gcc.c-torture/compile/20050217-1.c 2005-02-17 15:10:08.702023097 +0100
@@ -0,0 +1,14 @@
+/* PR c++/20023 */
+
+void f (void);
+typedef __SIZE_TYPE__ size_t;
+void g (void *a)
+{
+ size_t b = (size_t) a;
+ switch (b)
+ {
+ case 1:
+ f ();
+ break;
+ }
+}
--- gcc/testsuite/g++.dg/opt/switch3.C 2005-01-27 14:27:08.338732320 +0100
+++ gcc/testsuite/g++.dg/opt/switch3.C 2005-02-17 15:10:08.702023097 +0100
@@ -0,0 +1,16 @@
+// PR c++/20023
+// { dg-do compile }
+// { dg-options "-O2" }
+
+void f (void);
+typedef __SIZE_TYPE__ size_t;
+void g (void *a)
+{
+ size_t b = (size_t) a;
+ switch (b)
+ {
+ case 1:
+ f ();
+ break;
+ }
+}
Jakub