gcc-4.8.1 + graphite flags

[Rion]

I just reemerged world with gcc-4.8.1 an graphite c[xx]flags.
the system seems to be mostly stable, but I had to mask grahite flags for some packages.

here they are

Code: Select all

dev-qt/qtcore no-graphite.conf
dev-qt/qtscript no-graphite.conf
dev-qt/qtwebkit no-graphite.conf
dev-qt/qtgui no-graphite.conf # segfaults
sys-devel/bison no-graphite.conf # just broken.
app-editors/vim no-graphite.conf # very strange unusable behaviour
app-editors/gvim no-graphite.conf # very strange unusable behaviour
app-shells/bash no-graphite.conf # segfaults
app-arch/bzip2 no-graphite.conf # can't decompress anything
media-libs/x264 no-graphite.conf # won't build
media-video/ffmpeg no-graphite.conf # won't build
sys-apps/dbus no-graphite.conf # doesn't start
sys-devel/gdb no-graphite.conf # broken memory handling
dev-libs/openssl no-graphite.conf # tls doesn't work
media-video/mjpegtools no-graphite.conf # gcc OOM
media-libs/gegl no-graphite.conf # won't build
x11-libs/pango no-graphite.conf # cursor handling broken
app-office/libreoffice no-graphite.conf # does not pass tests
dev-vcs/subversion no-graphite.conf # gcc OOM
mate-extra/mate-power-manager no-graphite.conf # gcc OOM
media-video/vlc no-graphite.conf # gcc OOM
dev-qt/qt-creator no-graphite.conf # crashes on project open (sigabrt)
dev-util/valgrind no-graphite.conf # Valgrind: FATAL: aspacem assertion failed. nsegments[nsegments_used-1].end == Addr_MAX
sys-devel/gcc no-graphite.conf # looks like causes some apps to crash (qtcreator for example). have to make more tests..
net-libs/rb_libtorrent no-graphite.conf # crashes
dev-util/geany no-graphite.conf # crashes
sys-apps/gawk no-graphite.conf # somehow breaks configure step of php. leads to Fatal Error on call of any php func
x11-base/xorg-server no-graphite.conf # crash on exit?
dev-python/wxpython no-graphite.conf # only 2.8? causes some apps to crash. phatch for example
media-gfx/inkscape no-graphite.conf # 0.48.4-r1. unusable interface
dev-db/mariadb no-graphite.conf # breaks libreoffice compile
dev-lang/v8 no-graphite.conf # won't build
media-libs/libtheora no-graphite.conf # crashes
net-libs/gnutls no-graphite.conf  # 3.2.3. breaks certificate checking procedure. broken wget for example
app-text/poppler no-graphite.conf # 0.22.5. crashes evince on search
dev-lang/erlang no-graphite.conf # 16.1. segfault during compiation
kde-base/konsole no-graphite.conf # broken copy/paste by middle click
dev-lang/perl no-graphite.conf # 5.18.1 segfault during build
sys-block/partitionmanager no-graphite.conf # 1.0.3_p20130623. can't change part params. defaults are wrong
dev-libs/glib no-graphite.conf # 2.36.4-r1. random crashes. for example file-roller on opening some archives
app-emulation/virtualbox no-graphite.conf # 4.3.0. ice
media-libs/sdl-mixer no-graphite.conf # crashes. oolite for example

my flags:

Code: Select all

CFLAGS="-march=corei7-avx -O2 -pipe -ftree-vectorize -fgraphite-identity -maes -floop-parallelize-all -floop-interchange -floop-block -fira-loop-pressure"

and w/o graphite:

Code: Select all

CFLAGS="-march=corei7-avx -O2 -pipe -maes"

just fyi, maybe useful for somebody.

I'll update this list if found more unstable packages.

[kernelOfTruth]

thanks for sharing

[mbar]

And thanks from me

[PaulBredbury]

I've just started playing with graphite. With e.g. mplayer and SDL, it will try to use all available memory, bringing the PC to its knees with disk thrashing. To prevent that, use ulimit, e.g. with 4gb RAM:

Code: Select all

ulimit -v 1200000

I've put that in my root's ~/.bashrc, to make the compilation fail instead of pinching all RAM.

Edit: Gentoo's isl is out-of-date - latest isl is 0.12, which presumably needs gcc compiled with --disable-isl-version-check, since gcc 4.8-20130711's configure only looks for isl 0.10 & 0.11

[ZeuZ_NG]

Is there any update on wheter some of the listed applications are now able to start?
I'm going to try both LTO and Graphite but since it's in a low powered device I would like to save some time.
PS: Thanks for the information!

[Jaglover]

I've used

Code: Select all

-floop-interchange -floop-strip-mine -floop-block

for long time, everything builds. I tried LTO for some time and wasn't convinced it's worth the trouble. No LTO for now. Using 4.7.3.

[ZeuZ_NG]

Too much problems with LTO for no discernible benefit?
I'm using 4.8.1 so I'm on the edge of breaking guess I shouldn't push it

[mv]

I've used

Code: Select all

-floop-interchange -floop-strip-mine -floop-block

for long time, everything builds.

The problem with these are not built failures but runtime failures. Since 4.8.0, I disabled graphite since it has broken too many things (accidental crashes, strange behavior etc. pp); in earlier versions of gcc there were only occasional build failures with graphite but since 4.8.0 it has become practically unusable (at least world-wide).
In contrast with lto: I do not remember any package which built with LTO but produced runtime failures. Of course, many packages won't build with LTO.

[ZeuZ_NG]

I'm currently doing a world rebuild so will report back as soon as possible, although this is a newly built system and to be fairly honest I don't have much besides fluxbox and the Xorgserver plus the intel driver..
Perhaps a browser is a good measure? or too risky?

[saellaven]

in addition to your list, let me add two more

dev-db/mariadb no-graphite.conf # breaks libreoffice compile
dev-lang/v8 no-graphite.conf # won't build

[ZeuZ_NG]

in addition to your list, let me add two more

dev-db/mariadb no-graphite.conf # breaks libreoffice compile
dev-lang/v8 no-graphite.conf # won't build

How recently have you tried this? and with what version of gcc compiler suite?

[ZeuZ_NG]

my flags:

Code: Select all

CFLAGS="-march=corei7-avx -O2 -pipe -ftree-vectorize -fgraphite-identity -maes -floop-parallelize-all -floop-interchange -floop-block -fira-loop-pressure"

and w/o graphite:

Code: Select all

CFLAGS="-march=corei7-avx -O2 -pipe -maes"

Why not march=native?

[mv]

sys-devel/gcc no-graphite.conf # looks like causes some apps to crash (qtcreator for example). have to make more tests.

Ah, I see: I had compiled gcc-4.8.{0,1}* with graphite, of course. Maybe this is the reason for my many problems with graphite. On the other hand, it would be a strange accident that using gcc compiled with graphite just has problems for compiling with graphite. But who knows: Maybe the cause is that something in graphite is not reentrant or thread-safe and using graphite with a graphite-compiled gcc just forces such a behavior?

[ZeuZ_NG]

sys-devel/gcc no-graphite.conf # looks like causes some apps to crash (qtcreator for example). have to make more tests.

Ah, I see: I had compiled gcc-4.8.{0,1}* with graphite, of course. Maybe this is the reason for my many problems with graphite. On the other hand, it would be a strange accident that using gcc compiled with graphite just has problems for compiling with graphite. But who knows: Maybe the cause is that something in graphite is not reentrant or thread-safe and using graphite with a graphite-compiled gcc just forces such a behavior?

I've had no problem in a first shot building it with -floop-interchange -floop-strip-mine -floop-block although I've been reading:
https://wiki.gentoo.org/wiki/CFLAGS (the graphite/lto/openmp section) and http://software.intel.com/en-us/blogs/2 ... ance-hints

[mv]

I've had no problem in a first shot building it

As I said, the problem is not building it but too many runtime failures: The behaviour of vim (saving broken data after editing which obviously also other have experienced) is a typical example; since avoiding graphite solves these issues, it seems that this is not a bug of vim but actually of gcc w/graphite.

[saellaven]

in addition to your list, let me add two more

dev-db/mariadb no-graphite.conf # breaks libreoffice compile
dev-lang/v8 no-graphite.conf # won't build

How recently have you tried this? and with what version of gcc compiler suite?

gcc version 4.8.1 (Gentoo 4.8.1 p1.0, pie-0.5.6)

Sun Jun 30 02:02:03 2013 >>> dev-db/mariadb-5.5.31
Fri Aug 2 10:49:13 2013 >>> dev-lang/v8-3.18.5.14

there was a configure error in libreoffice-4.1.0.1 involving mis-parsing flags from mariadb that was only resolved by compiling mariadb without the graphite flags

v8 threw a compile error with the graphite flags

[ZeuZ_NG]

in addition to your list, let me add two more

dev-db/mariadb no-graphite.conf # breaks libreoffice compile
dev-lang/v8 no-graphite.conf # won't build

How recently have you tried this? and with what version of gcc compiler suite?

gcc version 4.8.1 (Gentoo 4.8.1 p1.0, pie-0.5.6)

Sun Jun 30 02:02:03 2013 >>> dev-db/mariadb-5.5.31
Fri Aug 2 10:49:13 2013 >>> dev-lang/v8-3.18.5.14

there was a configure error in libreoffice-4.1.0.1 involving mis-parsing flags from mariadb that was only resolved by compiling mariadb without the graphite flags

v8 threw a compile error with the graphite flags

Looks like I'll have to end up rebuilding a system from scratch then hehe..

[ZeuZ_NG]

Ok, so following my experience so far I can say that

Code: Select all

-floop-interchange -floop-strip-mine -floop-block

Seems to have negatively impacted performance..

[Rion]

Why not march=native?

distcc

[Jaglover]

Ok, so following my experience so far I can say that

Code: Select all

-floop-interchange -floop-strip-mine -floop-block

Seems to have negatively impacted performance..

Interesting. How did you measure that? Is this about build time or running performance?

[Rion]

Ok, so following my experience so far I can say that

Code: Select all

-floop-interchange -floop-strip-mine -floop-block

Seems to have negatively impacted performance..

Interesting. How did you measure that? Is this about build time or running performance?

btw I have noticed something similar with my flags too.

[ZeuZ_NG]

Ok, so following my experience so far I can say that

Code: Select all

-floop-interchange -floop-strip-mine -floop-block

Seems to have negatively impacted performance..

Interesting. How did you measure that? Is this about build time or running performance?

I've run some bc calculations with time before and after, it's not a good meassure but it overall feels slower..

[Rion]

probably I'll recompile the world with gcc-4.8.2 and w/o graphite.
just added one yet package to the list above..