On Wed, 05.01.2011 at 20:36:53 +0100, Jilles Tjoelker wrote: > On Wed, Jan 05, 2011 at 05:55:45PM +0100, Ulrich Spörlein wrote: > > On Wed, 05.01.2011 at 09:34:49 -0500, John Baldwin wrote: > > > These are all marked as __dead2, so the compiler should "know" that these > > > do > > > not return. > > > And clang did the right thing here in the past. Beware that it does no > > inter-procedural analysis yet, so it will usually miss that usage() > > calls exit unconditionally. > > > *But*, it should grok that for err(3) and exit(3). Now there are some > > possible remedies: > > > - get IPA to work with clang, or at least file a bug > > > - mark functions as __dead2 (please don't do that) > > Why not?
Cause IMHO it adds clutter, is noisy and needs to be maintained manually, when we have these "computer" things that should deduct this by themselves. > I have done this in some cases because it leads to better code with gcc > (the system version in 9-current). See SVN commit r212508 to > bin/sh/parser.c. Although synexpect() and synerror() are static, adding > __dead2 to both makes the executable 576 bytes smaller on i386 (these > functions are called many times). Adding __dead2 to synexpect() only > causes a warning "noreturn function does return" (it calls synerror()). > Adding __dead2 to synerror() only also makes the executable smaller but > not as much as adding it to both. > > Reordering the functions in the file does not help to make gcc see that > the functions do not return. This is too bad and really makes me sad. It shouldn't be necessary to hand-hold the compilers like that. Could you try some tests with gcc 4.5 to confirm this is still required? Uli _______________________________________________ freebsd-current@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-current To unsubscribe, send any mail to "freebsd-current-unsubscr...@freebsd.org"
