Hi! Some time between today and tomorrow we plan to switch Fedora development primary compiler from GCC 4.0.2-RH to 4.1.0-RH prerelease. We hope GCC 4.1.0 will be officially released in time for the Fedora Core 5 release, but if we want to switch, we need to do it now so that the compiler and packages built with it are sufficiently tested. For (so far preliminary) list of changes see http://gcc.gnu.org/gcc-4.1/changes.html (though, safe builtins and -fstack-protector has been already provided in GCC 4.0.2-RH). GCC 4.1.0-RH rpms are based on snapshots from svn://gcc.gnu.org/svn/gcc/branches/redhat/gcc-4_1-branch This branch is a merge between the upstream svn://gcc.gnu.org/svn/gcc/branches/gcc-4_1-branch and svn://gcc.gnu.org/svn/gcc/branches/gomp-20050608-branch which adds (preliminary) OpenMP 2.5 support. From our experience with building rawhide packages with this compiler, we know about a few outstanding GCC bugs that affect them: http://gcc.gnu.org/PR25240 (_Pragma parsing bug, prevents glibc build) http://gcc.gnu.org/PR25221 (reload ICE when compiling GIMP on i?86) http://gcc.gnu.org/PR25023 (ICE with pushw in struct by value passing and -g on i586 kernel) (of course there are many more regressions: http://gcc.gnu.org/bugzilla/buglist.cgi?query_format=advanced&short_desc_type=allwordssubstr&short_desc=4.1&target_milestone=4.0.3&target_milestone=4.1.0&bug_status=UNCONFIRMED&bug_status=NEW&bug_status=ASSIGNED&bug_status=SUSPENDED&bug_status=WAITING&bug_status=REOPENED&priority=P1&priority=P2&priority=P3 but those weren't seen in rawhide packages (so far)). The C++ frontend is again more strict then in GCC 4.0.x and rejects more invalid code. Problems seen in multiple packages include: "extra qualification..." errors - http://gcc.gnu.org/PR16782 friend injection change - struct A { typedef int S; friend void foo (S); friend void foo (); }; struct B { typedef int T; T t; void foo () { ::foo (t); } }; used to be accepted but is not any longer as ill-formed code. To fix this, the function needs to be declared at namespace scope, not just in a friend declaration. methods defined in wrong namespaces - struct A { void foo (); }; namespace { void A::foo () {} }; This is ill-formed, the definition needs to be done in the namespace containing the class. extern int bar (const char *); template <class T> struct A { void bar (const T &) { } }; template <class T> struct B : public A <T> { B () : x (0) { } void baz () { bar (this->x); } const T x; }; void foo () { B <double> b; b.baz (); } This is ill-formed, as bar is not template dependent and so it is looked up and found in the global namespace. The fix is to make the call template dependent, e.g. by using this->bar (this->x); relying on some C++ STL headers to include <cassert> - some of the headers no longer include <cassert> (on purpose), if you see this, just add explicit #include <cassert>. Also packages that use -Werror might need some changes, because some new warnings were added (especially strict aliasing warnings in C++) and with -Werror that means the package doesn't build any longer. Please report any internal compiler errors as well as other problems where GCC is on fault (and that aren't in the above list) to http://bugzilla.redhat.com/bugzilla/ as usually, ideally the /tmp/cc*.out preprocessed source GCC driver creates for ICEs, or by manually preprocessing the source, verifying the bug is still reproduceable and also mentioning exact compiler version, architecture and all compiler options needed to reproduce it. If you want to test OpenMP, use -fopenmp option. Jakub
Attachment:
pgpcCVvmj1m44.pgp
Description: PGP signature
