ICC / IFC 8.0 is out!

[Cheesefoam]

EDIT: This message has been updated to contain the most recent information and successful package merges with ICC 8.0.055 and IFC 8.0.034

You can find the ebuilds here:
Bug #35602: ICC 8.0.055 Ebuild
Bug #35609: IFC 8.0.034 Ebuild

Packages known to work:
Works with libtool 1.5 installed (perhaps even without):

• xine-ui-0.9.22
• blas-19980702 Need to change FC="g77" to FC="ifc" in ebuild
• lapack-3.0 Need to change FC="g77" to FC="ifc" in ebuild
• flac-1.1.0
• bzip2-1.0.2-r2
• unzip-5.50-r2
• gzip-1.3.3-r2
• povray-3.5c
• sed-4.0.7
• mpg123-0.59-r3
• mpeg-lib-1.3.1-r1
• libcdio
• cdrtools
• gtk+
• tar
• bison
• xmame See bug 35607 for 2 small patches
• bash See bug 36000 for patch to the latest bash ebuild
• pyorbit
• pkgconfig
• glib
• atk
• pango
• libpng
• jpeg
• tiff
• sys-apps/debianutils-1.16.7-r4
• sys-apps/coreutils-5.0.91-r2
• sys-apps/psmisc-21.2-r4
• sys-apps/grep-2.5.1-r1
• sys-apps/less-381
• sys-apps/which-2.16
• sys-apps/sed-4.0.8
• app-arch/cpio-2.5
• sys-apps/diffutils-2.8.4-r4
• net-misc/rsync-2.5.7
• sys-apps/file-4.06
• sys-apps/net-tools-1.60-r7
• sys-apps/wget-1.9.1-r2
• sys-apps/util-linux-2.13-r3
• sys-apps/pam-login-3.14
• sys-apps/shadow-4.0.3-r9
• app-shells/sash-3.6
• app-editors/nano-1.3.0
• dev-python/python-fchksum-1.7.1
• sys-apps/slocate-2.7-r3
• sys-apps/findutils-4.1.20-r1
• sys-devel/flex-2.5.4a-r5
• net-misc/dhcpcd-1.3.22_p4-r3
• sys-fs/e2fsprogs-1.34
• sys-devel/make-3.80
• app-arch/sharutils-4.2.1-r6
• sys-libs/readline-4.3-r5
• sys-libs/ncurses-5.3-r4
• sys-devel/libtool-1.5
• sys-devel/m4-1.4-r1
• sys-devel/bison-1.875
• sys-apps/texinfo-4.6
• sys-apps/groff-1.18.1-r3
• sys-apps/man-1.5m
• sys-devel/patch-2.5.9
• app-crypt/hashalot-0.1.0
• sys-libs/cracklib-2.7-r8
• dev-libs/popt-1.7-r1
• sys-apps/kbd-1.08-r4
• sys-apps/ed-0.2-r3
• sys-apps/fbset-2.1
• sys-devel/gettext-0.12.1
• sys-devel/bc-1.06-r5
• sys-devel/bin86-0.16.13
• sys-libs/pwdb-0.62
• sys-libs/db-4.1.25_p1-r3

Packages reported to work (still being investigated)

• gaim
• qt
• qt-3.3.0_beta1
• arts

Packages known *NOT* to work:

• MozillaFirebird-0.7
• mpqc-2.2.0
• kdelibs 3.2 beta 2
• sys-apps/procps-3.1.14-r1 - Fails to compile when compiled with ICC on file proc/readproc.c.
• net-misc/openssh-3.7.1_p2-r1 - Fails to compile with a bunch of undefined refernces to _intel_fast_mem{set,cpy}
• app-arch/ncompress-4.2.4 - Fails to compile with a missing closing quote error on line 1870 of compress42.c
• sys-apps/gawk-3.1.3 - Causes some breakage in the rc-scripts: /lib/rcscripts/filefuncs.so has undefined symbols.
• sys-libs/pam-0.77 - Pam fails to recognize that CC=ICC has been set and uses gcc anyways.
• net-misc/iptuils-021109 - Same as pam, ignores CC.
• dev-libs/expat-1.95.7 - Causes KDE Applications to segfault
• dev-libs/openssl-0.9.7c-r1 - Gets stuck in a loop on creating certs
• dev-libs/glib-2.2.3 - Causes artsd to segfault
• sys-apps/module-init-tools-0.9.15_pre4 - fails on insmod.c (code 2)
• sys-fs/devfsd-1.3.25-r5 - Causes boot to fail with "Symbol: "mksymlink" not found in GLOBAL." error.

So far, the general consensus is that the following flags do not work well for emerges, and should be avoided:

Code: Select all

-ip   -ipo   -cxxlib-gcc   -fast   -static

Suggested "safe" optimizations are:

Code: Select all

-O2 -x<X>
(Where <X> is the appropriate code for your processor, e.g. -xN)

[haugboel]

So I guess you can't just copy the 7.* ebuild. Have you tried to extract files with rpm2targz and do the installation manually
Downloading ifc as I write but over 128K mobile phone... I'll post details later if I get it working.

Troels

[stonent]

I guess you could hack the ebuild file to grab icc 8

[haugboel]

You need to know the exact file names to hack the ebuilds. They are for
-ifc: l_fc_p_8.0.034.tar.gz
-icc: l_cc_p_8.0.055.tar.gz
happy hacking. You still need to register to get the license file though.

On a side note this is a truely new version of icc/ifc, and contains among other things a merge with the Compaq Visual Fortran compiler. Where I work we normally get a tad better performance from CVF under windows than ifc under linux, and that is for an old version of CVF.

Oh boy - X-Mas comes early this year...

[Cheesefoam]

I tried hacking the old 7.x ICC/IFC files, but the results weren't pretty. Considering this is a first time attempt for me with ebuilds, it'll probably take me a while to come up with something that works.

On the positive side, accodring to Intel, licenses from 6 and 7 will work with 8 as long as they have not expired.

I'll work on it some more as I have time.

[Cheesefoam]

I'm working on it, but it isn't pretty so far. Currently, I keep getting sandbox access violations when the ebuild tried to make the destination directories for copying the files over:

Code: Select all

Obelix icc # emerge icc
Calculating dependencies ...done!
>>> emerge (1 of 1) dev-lang/icc-8.0.055 to /
*** Adjusting cvs-src permissions for portage user...
>>> md5 src_uri ;-) l_cc_p_8.0.055.tar.gz
>>> Unpacking source...
>>> Unpacking l_cc_p_8.0.055.tar.gz to /var/tmp/portage/icc-8.0.055/work
>>> Source unpacked.
 * Extracting: intel-icc8-8.0-44.i386.rpm
72994 blocks

>>> Install icc-8.0.055 into /var/tmp/portage/icc-8.0.055/image/ category dev-lang
mkdir: cannot create directory `/opt/intel/compiler80': File exists
ACCESS DENIED  mkdir:     /opt/intel/compiler80/ia32/intel_cc_80
cp: cannot create directory `/opt/intel/compiler80/ia32/intel_cc_80': Permission denied
man:
prepallstrip:
strip:
strip:
>>> Completed installing into /var/tmp/portage/icc-8.0.055/image/

--------------------------- ACCESS VIOLATION SUMMARY ---------------------------
LOG FILE = "/tmp/sandbox-icc-8.0.055-11868.log"

mkdir:     /opt/intel/compiler80/ia32/intel_cc_80
--------------------------------------------------------------------------------

What am I doing wrong here?

[Cheesefoam]

Nevermind - I fixed it. I'm uploading the files to bugs.gentoo.org.

Bug #35602: ICC 8.0.055 Ebuild
Bug #35609: IFC 8.0.034 Ebuild

I should point out that current these ebuilds install ICC to /opt/intel_cc_80 and IFC to /opt/intel_fc_80. I tried messing with the ebuilds to get them to play nice and both install into /opt/intel/compiler80, but I couldn't get it to work.

Also, it should be pointed out that the binary driver for IFC has changed names to "ifortbin". The scripts reflect this, but if you try to run ifcbin directly and it complains, this is why. You can run IFC with the old "ifc" script, or with a new one "ifort". Why Intel did this is beyond me.

[ebrostig]

I'm currently working on testing using icc 8 as a replacement for gcc, but I don't expect it to be an easy task.

Mozilla-Firebird failed during compile using icc.
Trying xfree 4.3.99.901 (The developer snapshot) at the moment.

Erik

[Cheesefoam]

Brave man. I started with xmms. It worked mostly, except for one library. You might want to add -cxxlib-gcc to the CFLAGS, for it forces ICC to use the GCC libraries & headers instead of its own. Might make things more portable.

Also, if you didn't see it, there's a new optimization flag, -ax<X>, where <X> is the processor type (i.e. -axN for P4). It does specific optimizations for the P4, but seems to be more aggressive in autoscheduling and optimizing for the processor.

[stonent]

The real test is compiling a kernel, complete with vector optimizations.

I'm drooling! Maybe even build GCC under ICC! Super fast emerges!

[discomfitor]

see my post http://forums.gentoo.org/viewtopic.php? ... highlight= for more info on using icc in ebuilds

[stonent]

Well I know the old ICC couldn't build kernels, but the new one is supposed to be able to compile code designed for GCC

So I'm gathering that the ICC license allows us to build our own packages, but not distribute them to someone else, correct?

[Cheesefoam]

2.6 series kernels are *supposed* to be able to be compiled with ICC.

I haven't actually tried, though.

[ebrostig]

Hmmm...
I knew Xfree was going to be a challenge

Funny that imake compiled with icc fails when trying to parse structures like:

Code: Select all

#if BLAH
#define HUH
#ifndef BLAH
#undef HUH
#endif
#endif

Complains about missing #endif's.

Currently recompiling 'krusader' with icc and -g option since the first build seg faulted when I tried to start krusader.

Erik

[ebrostig]

2.6 series kernels are *supposed* to be able to be compiled with ICC.

I haven't actually tried, though.

I've tried and the compile failed utterly

Erik

[haugboel]

Here is a link to a japanese pdf file describing what have to be done in order to compile a kernel with icc 8.0.

http://www.suri.co.jp/~ohkubo-k/linux/i ... 030801.pdf

Unfortunately my japanese skills are rather limited but if anybody feels like translating it to the rest of the world it would be rocking.

[/url]

[Cheesefoam]

Unfortunately, it seems that the use of interprocedural optimization (-ipo) seems to b0rk a whole lot of compiles, due to the referencing of ld. Even in you do an

Code: Select all

export LD=xild

it often still does not take it. The kernel compile is a great example of this. Sadly, this is one of the coolest features of the ICC compiler.

EDIT: By the way, here's the script I usually use for doing an ICC emerge:

Code: Select all

export AR=xiar
export LD=xild
export CC=icc
export CXX=icc
export FC=ifc
export FF=ifc
export FORTRAN=ifc
export F77=ifc
emerge $*
export CC=gcc
export AR=ar
export LD=ld
export CXX=gcc
export FC=g77
export FF=g77
export FORTRAN=g77
export F77=g77

It should cover all of the bases (with the exception of the occasional ld problem.

I'm gonna try and keep a running list of things that I know compile and work with ICC 8.0, using these CFLAGS (*I noticed that perhaps -axN is not such a good choice, unless you are going for portability. Use the -xN option instead, as it skips adding instructions for backward compatility with other processors):

Code: Select all

-O2 -xN -cxxlibs-gcc -ipo -ipo_obj

Works with libtool 1.5 installed (perhaps even without):

Code: Select all

xine-ui
blas (Need to change FC="g77" to FC="ifc" in ebuild)
lapack (Need to change FC="g77" to FC="ifc" in ebuild)
flac
bzip2
unzip
gzip

[discomfitor]

under 7.0 I was able to get the vast majority of packages to compile and work successfully. I can't imagine that 8.0 is any different, so that means that all ebuilds that use standard configure scripts and don't have wacky makefiles should work fairly easily. perhaps I'll get back on icc and see how things have progressed...

[Cheesefoam]

What packages did you get to work, and how? The general experience has been that the vast majority of packages did *not* compile or work properly under 7.0.

Cases in point: XMMS, xine-lib, mozilla, openoffice, blas, lapack, mpqc, evolution, kde, gnome.....

7.0 had serious problems in that it could not create shared libraries, nor would binaries produced by an emerge propely link to gcc-compiled libraries.

EDIT: I managed to get libtool 1.5 *properly* working. I guess my previous attempts were b0rked. Now mpqc is compiling just fine. Good call on libtool 1.5! Why don't you submit it to bugs.gentoo.org as an ebuild to get incorporated into the Gentoo tree?

[stonent]

The cool thing I like about ICC is when it pops up the message that it is vectorizing something. I don't know how much that helps, but it sounds cool.

I wonder about making a stage 1 with icc set as default for everything. We could start with emerge system and start working from there.

We'd have to link the license file to a ramdisk I guess.

Then also notice if a compile fails, dump all the info into a report and then continue on to the next item. Then see how much gets compiled. Also for things that support it (like Python) do a make test on each item and log the output.

[Cheesefoam]

Well, we should keep a list of things that do work. I'm going to take a crack at MozillaFirebird tonight with libtool 1.5 and see what happens. I'm compiling mpqc right now to see if it works.

[discomfitor]

as I said, things like oo were not doable in 7.0 because I was too lazy to get into deep hack mode and actually change the makefiles and such. if you look at the other thread which I linked to earlier, there are directions for mangling the ebuilds of most packages into a semblance of workingness. these include getting shared libs to WORK, and the programs do indeed work. I never checked gcc-linkability because I never used gcc...my idea was to try compiling the entire distro with icc. I got pretty far, but then I ran into issues with X and stuff and it was too much of a hassle.

anyway, if you want me to explain some of the stuff I did, just ask...I check fairly often.

[CheshireCat]

Installed icc8 by hand for now, I'm not up to doing an ebuild for this mess at the moment. Looks like it works decently, I've built flac with it (almost 2x as fast, even though I'm running "P3" optimized code on an AthlonXP), and I am trying to do xine-lib right now. Libvorbis had some weirdness that seems to be related to libm, maybe -cxxlib-gcc will help there... looks good so far though.

[Cheesefoam]

Well, MozillaFirebird bailed. Too bad. The current mpqc build in Portage doesn't work either, though I am testing the most recent version of mpqc, 2.2.0 with ICC to see if it compiles. If it does, I'll submit a new ebuild for it.

I'll add flac to the list of working apps.

[stonent]

Geeze, what good is ICC if it can't compile anything?