breaks on XFree86 build at -O2

[Daniel Stone]

With -O2, gcc sometimes produces files that have truncated symbols; given the
buildds are OK, this may be a local problem, but it is 100% reproducible on
certain files within the XFree86 tree:
daniels@nanasawa:~/canonical/xfree86/xfree86-4.3.0.dfsg.1/build-tree/xc/programs/xkbcomp%
make
gcc -m32 -g -O2 -fno-strict-aliasing -ansi -pedantic -Wall -Wpointer-arith
-Wstrict-prototypes -Wmissing-prototypes
-Wmissing-declarations -Wredundant-decls
-Wnested-externs -Wundef -I../../include/extensions -I../..
-I../../exports/include -Dlinux -D__i386__ -D_POSIX_C_SOURCE=199309L
                   -D_POSIX_SOURCE -D_XOPEN_SOURCE -D_BSD_SOURCE
-D_SVID_SOURCE -D_GNU_SOURCE
     -DFUNCPROTO=15 -DNARROWPROTO -c -o keytypes.o keytypes.c
rm -f xkbcomp
gcc -m32 -o xkbcomp -g -O2 -fno-strict-aliasing -ansi -pedantic -Wall
-Wpointer-arith -Wstrict-prototypes -Wmissing-prototypes
-Wmissing-declarations -Wredundant-decls
-Wnested-externs -Wundef -L../../exports/lib xkbcomp.o xkbscan.o expr.o
vmod.o indicators.o misc.o alias.o keymap.o keycodes.o keytypes.o
compat.o action.o symbols.o geometry.o xkbpath.o listing.o
       xkbparse.o parseutils.o utils.o -lxkbfile -lXext -lX11
-Wl,-rpath-link,../../exports/lib
keytypes.o: could not read symbols: File truncated
collect2: ld returned 1 exit status
make: *** [xkbcomp] Error 1
daniels@nanasawa:~/canonical/xfree86/xfree86-4.3.0.dfsg.1/build-tree/xc/programs/xkbcomp%
rm keytypes.o
daniels@nanasawa:~/canonical/xfree86/xfree86-4.3.0.dfsg.1/build-tree/xc/programs/xkbcomp%
gcc -m32 -g -fno-strict-aliasing -ansi -pedantic -Wall -Wpointer-arith
-Wstrict-prototypes -Wmissing-prototypes
-Wmissing-declarations -Wredundant-decls
-Wnested-externs -Wundef -I../../include/extensions -I../..
-I../../exports/include -Dlinux -D__i386__ -D_POSIX_C_SOURCE=199309L
                   -D_POSIX_SOURCE -D_XOPEN_SOURCE -D_BSD_SOURCE
-D_SVID_SOURCE -D_GNU_SOURCE
     -DFUNCPROTO=15 -DNARROWPROTO -c -o keytypes.o keytypes.c
daniels@nanasawa:~/canonical/xfree86/xfree86-4.3.0.dfsg.1/build-tree/xc/programs/xkbcomp%
make xkbcomp
rm -f xkbcomp
gcc -m32 -o xkbcomp -g -O2 -fno-strict-aliasing -ansi -pedantic -Wall
-Wpointer-arith -Wstrict-prototypes -Wmissing-prototypes
-Wmissing-declarations -Wredundant-decls
-Wnested-externs -Wundef -L../../exports/lib xkbcomp.o xkbscan.o expr.o
vmod.o indicators.o misc.o alias.o keymap.o keycodes.o keytypes.o
compat.o action.o symbols.o geometry.o xkbpath.o listing.o
       xkbparse.o parseutils.o utils.o -lxkbfile -lXext -lX11
-Wl,-rpath-link,../../exports/lib
daniels@nanasawa:~/canonical/xfree86/xfree86-4.3.0.dfsg.1/build-tree/xc/programs/xkbcomp%

It's been a while since I've filed a gcc bug, so please let me know what sort of
debug output I need to give you.

[Daniel Stone (daniels)]

Also occurs (100% reproducibly) on xc/programs/xtrap/xtrapstats.c.

I'm running on an AthlonXP 2400+ (standard clockspeed, well-cooled, all that),
good Kingston RAM that checked out fine with memtest86. The only real
derivation is that I'm running a self-compiled 2.6.9-rc1-mm1, as I need proper
XFS support, and 2.6.7 hung with my media reader plugged into the USB port (will
check if this still occurs with 2.6.8.1); however, I wouldn't believe this would
affect anything, as it only affects the very same specific files (static across
many unpacks from different .orig tarballs -- initially I thought it was my bad,
deleted the orig, and downloaded it again).

[Daniel Stone (daniels)]

Yep, local problem; 2.6.8.1-1 works fine. Something I encountered later on in
the build[0] made me think it was the kernel, and after putting 2.6.8.1-1 in,
the XFS braindamage in there is a lot less bad than in .9-rc1-mm1, in which it
encounters random, reproducible(!!) short reads. Go figure.

[0]: commonregs.h would come up as being 'shorter than expected' from gcc(?!?),
and sure enough, even after repeated wget'ing from the full file, it would
always end up reading shorter than expected.