Re: Updating in glibc and gnulib
* Bruno Haible: > Florian Weimer wrote: >> Does gnulib still override unconditionally? > > Gnulib does not override , and never did. Thanks for looking into this. gnulib's libc-config.h does this: | #ifndef __attribute_nonnull__ | /* either does not exist, or is too old for Gnulib. |Prepare to include , which is Gnulib's version of a |more-recent glibc . */ | … | /* Include our copy of glibc . */ | # include And as gnulib's uses the same _SYS_CDEFS_H header guard as glibc's, that effectively replaces with gnulib's . > That is, when a package that uses Gnulib does > #include > it will get the of the system (from glibc, *BSD, Cygwin, > etc.). Apparently not if it includes libc-config.h first. I think what happened is that the order of backporting commit 0b5ca7c3e551e5502f3be3b06453324fe8604e82 Author: Paul Eggert Date: Tue Sep 21 07:47:45 2021 -0700 regex: copy back from Gnulib (which brought in __attribute_nonnull__) and commit a643f60c53876be0d57b4b7373770e6cb356fd13 Author: Siddhesh Poyarekar Date: Wed Oct 20 18:12:41 2021 +0530 Make sure that the fortified function conditionals are constant was reversed on the glibc 2.34 branch, so the version check based on __attribute_nonnull__ would signal that system is too old. But with the second commit for fortified functions, glibc 2.34 headers started requiring other macros not present in gnulib's copy, so some projects using copied gnulib sources would start to fail. I backported the regex sync to glibc 2.34 in November, so this should now be solved because we now define __attribute_nonnull__ even on the 2.34 branch. I think only the 2.34 branch had this problem. I think we should have backported the __attribute_nonnull__-defining commit to glibc 2.34 earlier, when we noticed problems. Updating the gnulib-bundled copy only (which is what happened at first) wasn't the best resolution. Thanks, Florian
Re: Updating in glibc and gnulib
Florian Weimer wrote: > Does gnulib still override unconditionally? Gnulib does not override , and never did. That is, when a package that uses Gnulib does #include it will get the of the system (from glibc, *BSD, Cygwin, etc.). Only when a package uses the compiler option "-I /usr/include/sys" and #include there would be a conflict between what Gnulib ships and the installed file in /usr/include/sys. But nobody does that, because there would already be a conflict between and . What Gnulib does is to ship a copy of glibc's as , not . And it is used for a few compilation units only (essentially regex, fnmatch, glob, and a few others). > Why does gnulib bundle ? Gnulib takes the source code of regex, fnmatch, glob, and a few other functions from glibc and makes them portable to other platforms. Since this source code contains references to __glibc_unlikely, __attribute_warn_unused_result__, etc., Gnulib uses the copy of cdefs.h to define these macros in the way the source code shared with glibc expects it. Some of these symbols are also defined on other platforms, sometimes differently. Therefore Gnulib has to be careful to not override essential symbols of these other platforms. It's not trivial, but there appears to be enough room to navigate through the two constraints. [1] > In the past, some gnulib-using programs supplied their own copy > of instead, even when building against glibc. This caused > build failures in the glibc headers because they (quite reasonably) > assumed that defines the macros for that glibc version. You need to take this up with the respective packages. As I said above, Gnulib overrides , not . There were some problems around the 'glob' code, but they were resolved more than 1.5 years ago. This area is still a bit fragile, though. > We could move glibc's internal definitions to a new header, reducing > in scope, but presumably that means gnulib would just > starting bundling that other header, and we would have the same issue > once more. Yes, such a move would be pointless. Bruno [1] https://lists.gnu.org/archive/html/bug-gnulib/2023-01/msg00238.html
Updating in glibc and gnulib
Why does gnulib bundle ? We edit this file regularly in glibc. In the past, some gnulib-using programs supplied their own copy of instead, even when building against glibc. This caused build failures in the glibc headers because they (quite reasonably) assumed that defines the macros for that glibc version. Does gnulib still override unconditionally? A version check will be difficult because sometimes, we have to backport header fixes to older versions, and that may require adding additional macros in . We could move glibc's internal definitions to a new header, reducing in scope, but presumably that means gnulib would just starting bundling that other header, and we would have the same issue once more. Thanks, Florian