[Bug middle-end/46488] [4.5 regression] server/core_filters.c from apache httpd 2.2.17 miscompiled at -O3

[pcpa at mandriva dot com.br]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46488

--- Comment #27 from Paulo César Pereira de Andrade <pcpa at mandriva dot com.br> 2010-11-29 21:26:32 UTC ---
Created attachment 22572
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22572
test case

I learned most of apache/apr/apr-util internals while debugging this problem in
mandriva cooker,
so, maybe the test case can be made simpler... The crash in mandriva is only
confirmed to happen
in i586. Sample run of the test case:

[pcpa@localhost ~]$ gcc -O0 -fstrict-aliasing -D_LARGEFILE64_SOURCE
-I/usr/include/apr-1 -g3 apr-test.c; LC_ALL=C ./a.out
It works!
[pcpa@localhost ~]$ gcc -O3 -fstrict-aliasing -D_LARGEFILE64_SOURCE
-I/usr/include/apr-1 -g3 apr-test.c; LC_ALL=C ./a.out
a.out: apr-test.c:59: check: Assertion `e->link.next->link.prev == e' failed.
Abortado
[pcpa@localhost ~]$ gcc -O3 -fno-strict-aliasing -D_LARGEFILE64_SOURCE
-I/usr/include/apr-1 -g3 apr-test.c; LC_ALL=C ./a.out
It works!

[pardon choosing the "It works!" string :-)]

It also appears to correct the issue with -fstrict-aliasing by adding
temporaries, e.g:
-#define APR_RING_SPLICE_HEAD(hp, ep1, epN, elem, link)                 \
-       APR_RING_SPLICE_AFTER(APR_RING_SENTINEL((hp), elem, link),      \
-                            (ep1), (epN), link)
+#define APR_RING_SPLICE_HEAD(hp, ep1, epN, elem, link) do {            \
+       struct elem *lep = APR_RING_SENTINEL((hp), elem, link);         \
+       APR_RING_SPLICE_AFTER(lep, (ep1), (epN), link);                 \
+    } while (0)

The attribute noinline in the test case was to "somehow" make gcc not
understand where
the pointers come from, otherwise, if removing any of them, it will also work.