| Message ID | 20110121194204.GL2724@tyan-ft48-01.lab.bos.redhat.com |
|---|---|
| State | New |
| Headers | show |
On 01/21/2011 02:42 PM, Jakub Jelinek wrote: > The following patch fixes it by not calling begin_for_scope > in that case again. OK. > BTW, I was kind of surprised pt.c calls begin_for_stmt > for both FOR_STMT and RANGE_FOR_STMT, I'd kind of expect > it should be begin_range_for_stmt in the latter case, but > haven't investigated in detail... Calling begin_for_stmt is correct; RANGE_FOR_STMT only exists in templates, at tsubst time we convert them into the equivalent C-style for. Jason
On Jan 21, 2011, at 11:42 AM, Jakub Jelinek wrote:
> With -fno-for-scope begin_for_scope in templates sets
Hum... I wonder, can we just rm -rf -fno-for-scope now... Useful a decade ago, I'd hope that everybody has already fixed their code.
On Fri, Jan 21, 2011 at 3:29 PM, Mike Stump <mikestump@comcast.net> wrote: > On Jan 21, 2011, at 11:42 AM, Jakub Jelinek wrote: >> With -fno-for-scope begin_for_scope in templates sets > > Hum... I wonder, can we just rm -rf -fno-for-scope now... Useful a decade ago, I'd hope that everybody has already fixed their code. > I'm in favor of removing -fno-for-scope and -fno-elide-copy-constructor.
--- gcc/cp/semantics.c.jj 2011-01-18 12:20:08.000000000 +0100 +++ gcc/cp/semantics.c 2011-01-21 17:56:47.000000000 +0100 @@ -860,8 +860,9 @@ begin_for_stmt (tree scope, tree init) if (scope == NULL_TREE) { - gcc_assert (!init); - scope = begin_for_scope (&init); + gcc_assert (!init || !(flag_new_for_scope > 0)); + if (!init) + scope = begin_for_scope (&init); } FOR_INIT_STMT (r) = init; TREE_CHAIN (r) = scope; @@ -962,8 +963,9 @@ begin_range_for_stmt (tree scope, tree i if (scope == NULL_TREE) { - gcc_assert (!init); - scope = begin_for_scope (&init); + gcc_assert (!init || !(flag_new_for_scope > 0)); + if (!init) + scope = begin_for_scope (&init); } /* RANGE_FOR_STMTs do not use nor save the init tree, so we --- gcc/testsuite/g++.dg/cpp0x/range-for10.C.jj 2011-01-21 18:10:06.000000000 +0100 +++ gcc/testsuite/g++.dg/cpp0x/range-for10.C 2011-01-21 18:10:56.000000000 +0100 @@ -0,0 +1,18 @@ +// PR c++/47388 +// { dg-do compile } +// { dg-options "-fno-for-scope -std=c++0x" } + +template <int> +void +foo () +{ + int a[] = { 1, 2, 3, 4 }; + for (int i : a) + ; +} + +void +bar () +{ + foo <0> (); +} --- gcc/testsuite/g++.dg/template/for1.C.jj 2011-01-21 18:12:16.000000000 +0100 +++ gcc/testsuite/g++.dg/template/for1.C 2011-01-21 18:12:10.000000000 +0100 @@ -0,0 +1,23 @@ +// PR c++/47388 +// { dg-do compile } +// { dg-options "-fno-for-scope" } + +template <int> +void +foo () +{ + int i; + for (i = 0; i < 16; i++) + ; + for (int j = 0; j < 16; j++) + ; + if (j != 16) + for (;;) + ; +} + +void +bar () +{ + foo <0> (); +}
Hi! With -fno-for-scope begin_for_scope in templates sets init to push_stmt_list () and returns NULL for scope. cp_parser_for first calls this (thus scope is NULL, but init is non-NULL, and then cp_parser_{c,range}_for call begin_*for_stmt which asserts that if scope is NULL then init is NULL too (which is true for invocations from pt.c, and for -ffor-scope, but not for -fno-for-scope if processing_template_decl). The following patch fixes it by not calling begin_for_scope in that case again. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? BTW, I was kind of surprised pt.c calls begin_for_stmt for both FOR_STMT and RANGE_FOR_STMT, I'd kind of expect it should be begin_range_for_stmt in the latter case, but haven't investigated in detail... 2011-01-27 Jakub Jelinek <jakub@redhat.com> PR c++/47388 * semantics.c (begin_for_stmt): If -fno-for-scope, don't assume init must be NULL if scope is NULL. (begin_range_for_stmt): Likewise. * g++.dg/cpp0x/range-for10.C: New test. * g++.dg/template/for1.C: New test. Jakub