I've been checking over the benchmarks as a lead up to the 2010.09
release. We're in a good way compared to both 4.4.4 and 4.5.1 for
most non-trivial tests.
* pybench is 10.9 % faster than 4.4.4 and 7.7 % faster than 4.5.1.
* linpack is 46.4 % faster than 4.4.4 and the same as 4.5.1.
* ffmpeg h.264 video decode (with hand written assembler versions
turned off) is 15.4 % faster than 4.4.4 and 1.2 % faster than 4.5.1.
All results are statistically invalid and against poor workloads, but
I'll work on that.
See http://ex.seabright.co.nz/helpers/benchcompare for more.
-- Michael
Loïc Minier wrote:
> I see you moved the wiki page to the public space, thanks
>
> Couple of notes:
> * make sure you use the rename action on the page, I think this will
> preserver history (I didn't check whether you did or not, but I think
> not)
No, I didn't. I use copy and paste. I'll use rename action.
> * add a page at the old location with "#redirect NewPage" or
> "#refresh 0 http://newurl/"
OK, got it. Thanks for your help on wiki.
...are available here:
https://wiki.linaro.org/WorkingGroups/ToolChain/Meetings/2010-09-06
A copy and activity reports are included below.
-- Michael
Attendees
• Name Email IRC Nick
Andrew Stubbs ams(a)codesourcery.com ams_cs
Chung-Lin Tang cltang(a)codesourcery.com cltang
Julian Brown julian(a)codesourcery.com jbrown
Loïc Minier lool(a)linaro.org lool
Marcin Juszkiewicz marcin.juszkiewicz(a)linaro.org hrw
Matthias Klose doko(a)canonical.com doko
Michael Hope michael.hope(a)linaro.org michaelh
Peter Maydell peter.maydell(a)linaro.org pm215
Richard Earnshaw richard.earnshaw(a)arm.com rearnshaw
Ulrich Weigand ulrich.weigand(a)linaro.org uweigand
Yao Qi yao(a)codesourcery.com yao
Agenda
• Licensing of string routines
• State of valgrind
• State of GDB
• Open tickets
□ 600298, 616141, 604753: SMP/sync related
□ 605059 4.4.5
□ 629671 ICE in reload_cse_simplify_operands in thumb-1 mode
□ 590696 Wrong use of objdump during cross build
• Upcoming release
• Creating blueprints
Blueprint Assignee
Initial delivery of Linaro GCC 4.4 ams
Cross Compiler Packages hrw
Action Items from this Meeting
• ACTION: Richard to check with the legal department on string licensing
issues
• ACTION: Peter to talk with valgrind upstream re: Linaro releasing a
ARM-focused version
• ACTION: Michael to organise an 'experimental' PPA that toolchain output can
go into
• ACTION: Michael to talk with Cody Somerville re: building on ARM
• ACTION: Michael to set up a GDB 7.2 based off the release tarball
• ACTION: Andrew to pull sync changes back into 4.4 for this release
• ACTION: Michael to assign appropriate sync ticket to Andrew to track the
backport
• ACTION: Andrew to merge the current post 4.4.4 release branch into our 4.4
for this release
• ACTION: Julian to do a basic investigation into 629671
• ACTION: Andrew to merge the cross-compile objdump ticket into this release
and re-kick upstream process
Action Items from Previous Meeting
• ACTION: Michael to re-check with TSC that we can assign copyright but keep
ability to relicense
• DONE: Yao to continue on GDB for a week then switch to investigation
• ACTION: Peter to check into the state and progress of valgrind for the
meeting on the 30th.
• ACTION: Chung-Lin to shift the CSL backport list out onto the Linaro wiki
• ACTION: Michael to see about doing an archive rebuild with 4.5
• DONE: Michael to send IBM's list to Yao
Minutes
String routines:
• Michael asked Richard about getting the current str* routines by ARM
transferred to Linaro
• Linaro will then get these into other C libraries
• FSF prefers LGPL and copyright for glibc
• Linaro prefers MIT/X11 everywhere so that fixes and improvements can be
shared
• Richard is concerned about the copyright assignment and any patent grant
• ACTION: Richard to check with the legal department on string licensing
issues
• Extreme fallback is to re-write the routines to all be under Linaro
copyright. memcpy() and similar may need this
Valgrind:
• Peter has been looking at how it works on the ARM platform
• Upstream is very responsive to issues
• Now works on Firefox and OO.org
• Upstrem doesn't have any particular release cycle
• ARM changes are pretty extensive and can't be extracted
• Peter suggested making valgrind available in a PPA to start with
• NEON detection at startup is remaining issue
• What next?
□ Packaging is straight forward
□ Don't want to steal upstream's thunder or release something
inappropriate
□ ACTION: Peter to talk with valgrind upstream re: Linaro releasing a
ARM-focused version
• Could bring into the Linaro overlay PPA
• ACTION: Michael to organise an 'experimental' PPA that toolchain output can
go into
• ACTION: Michael to talk with Cody Somerville re: building on ARM
GDB:
• 7.2 is now available
• Time to start up a gdb-linaro based on that
• Matthias mentioned that we will have GDB 7.2 on Maverick
• How should we manage the source
□ QEMU is over git
□ Could use bzr or git
□ bzr with Launchpad can't handle multiple branches when pulling from git
□ GDB is unique in how it's mixed in with the rest of the projects hosted
on sourceare
□ Branches as such are trucky
□ Could just base off tarballs
□ ACTION: Michael to set up a GDB 7.2 based off the release tarball
Tickets:
• ACTION: Andrew to pull sync changes back into 4.4 for this release
• ACTION: Michael to assign appropriate sync ticket to Andrew to track the
backport
• ACTION: Andrew to merge the current post 4.4.4 release branch into our 4.4
for this release
• ACTION: Julian to do a basic investigation into 629671
• ACTION: Andrew to merge the cross-compile objdump ticket into this release
and re-kick upstream process
Patch tracker:
• Andrew noted that it is now fully populated with the GCC data
• Has assigned various patches that still need to go upstream to Yao and
Julian
Next meeting is on 2010-09-08 on the public code.
--- Chung-Lin Tang
== Linaro Toolchain ==
* Google ARM patch sets: committed a second set to SG++ 4.5 trunk on
Tues. AndrewS pushed both sets to Linaro. Worked on a third set, those
related to PR42235, but this time regression test results were not so
clean. Will look into, but considering whether to stop the backports
here.
* LP:628526, submitted a patch to gcc-patches for explicitly turning
off stack protection in libgcc build flags, awaiting response.
* LP:601030, eglibc 2.11/12 problem with ___longjmp_chk on x86-64.
Problem seems to be clear, fix quite simple, but so far cannot seem to
reproduce and verify. Also unclear if I should send the fix to eglibc
or glibc, the idea of the latter making me a bit nervous... :P
== libffi ==
* Got an acknowledgement from the libffi maintainer that he'll review
the VFP hard-float support patch soon.
== This week ==
* Look into remaining Google approved patches, mainly those related to
PR42235 and PR42575.
* Try to reproduce LP:601030 and send patch soon.
* Linaro GCC investigations.
--- Andrew Stubbs
== Linaro GCC ==
* Michael has get the new patch tracker into a usable state. I've
transferred all the data from the old wiki tracker, and looked up the
remaining data as far as I can. The new tracker should now be fully
populated with data. It's here, for the moment:
http://ex.seabright.co.nz/helpers/patchtrack
* Start Yao and Julian on the optimization investigation tasks.
* Continue trawl through the CS bugs looking for candidates to push
to the Linaro tracker.
== Other ==
* Public holiday on Monday.
* Attended the monthly CS/Linaro sync meeting.
--- Yao Qi
== Linaro GDB wrap up ==
* LP:615993 gdb.base/sigstep.exp failures
Patch was committed to gdb mainline and 7.2 branch.
* LP:615995 gdb.base/watch-vfork.exp failures
Discussed with Pedro, create a patch, which fixed failures on ARM,
but can't fix failures on x86(they are caused by different problems).
Leave the x86 failures there, and patch is being reviewed in
gdb-patches.
== Linaro GCC ==
* CS306:Investigate on thumb2 improvement
Read/understand previous effort related on code size
improvement from CSL wiki pages.
Experiment with CSL scripts for size benchmarking. With Dan's
help, run benchmarking in a correct/reasonable way.
'Reproduce' some inefficient code mentioned by Julian. Some of them
are still there.
== Misc ==
* LP:605042
Revert one patch, and rebuild it. No seg fault is found.
== This Week ==
* Continue my work on CS306.
--- Peter Maydell
RAG:
Red:
Amber: virtio-system writeup not going as fast as expected
Green: ARM legal OK now received
Milestones:
| Planned | Estimate | Actual |
finish virtio-system | 2010-08-27 | ? | |
I need to replan this (no forward progress this week
because more important stuff intervened)
Progress:
virtio-system:
- actually trying a SATA disk revealed that the PB926 PCI
interrupt mapping was wrong; now fixed after consulting
the schematics and a round or two of patch testing with Arnd
- I have a PB1176 board but it doesn't seem to talk to
the serial port on poweron. Will try a firmware reflash
but it might just be broken...
- no progress on writeup because other things intervened.
valgrind:
- went through the motions of getting a valgrind svn snapshot
into the ubuntu packaging
- tested on pegatron (A8, maverick, thumb2), found four bugs:
+ BX PC not implemented (fixed upstream)
https://bugs.kde.org/show_bug.cgi?id=249775
+ RBIT not implemented (fixed upstream)
https://bugs.kde.org/show_bug.cgi?id=249924
+ pwrite64 syscall not implemented (fixed upstream)
https://bugs.kde.org/show_bug.cgi?id=249996
+ test for presence of neon wrong
https://bugs.kde.org/show_bug.cgi?id=249775
With a bodge for the last and the fixes for the first 3,
valgrind now successfully runs openoffice and firefox.
other:
- Investigated https://bugs.launchpad.net/bugs/628471 : qemu-maemo
doesn't work with new linaro beagleboard kernels. It looks
like we now try to probe for NAND (which failed earlier
for other reasons which I suspect are a now-fixed bug),
and qemu-maemo's NAND implementation doesn't map anything at
the address the nand code is trying to poll for a status bit.
- first post to qemu-devel :-) (review of somebody's
patch to not confuse SMC with BKPT in the arm decoder)
Plans:
virtio-system:
- hoping to get the qemu patches into the ubuntu qemu-maemo
package, which will avoid the need to talk about patching qemu
- finish the writeup and put it on the wiki
- test PCI patches on PB1176
valgrind:
- respin a valgrind with proper fixes for everything and
put it in a PPA somewhere
other:
- come up with some fix or workaround for #628471
- put the rebased ubuntu qemu-maemo work up onto gitorious
so other people can see it
Absences:
Friday 5 November and 20 other days in this calendar year
Looks good. I've created a real project, added a README/LICENSE, and
merged your changes. See:
https://launchpad.net/tcwg-web
There was a funny render difference between Firefox and Chromium -
revisions with no bugs lead to a rowspan of zero which Firefox doesn't
like. I also pulled some common code out into a function and used the
built-in variable 'loop'. 'loop' is quite nice as it provides values
like .index, .first, .odd, and so on based on your position in the
loop.
-- Michael
On Fri, Sep 3, 2010 at 11:02 PM, Andrew Stubbs <ams(a)codesourcery.com> wrote:
> Hi Michael,
>
> I've been playing with you patch tracker, and come up with this:
>
> https://code.launchpad.net/~ams-codesourcery/+junk/tcwg-web
>
> I don't seem to be able to propose an official merge request to your branch,
> but it's just a quick implementation anyway, and could probably be cleaned
> up.
>
> The patch renders each ticket in it's own row (without changing the way the
> first two columns are rendered). This means they can have their own colour
> and we can maybe see better what status goes with what bug.
>
> To see an example of what it does, see revision 4.4:93544
>
> Andrew
>
I want to share status of my cross compiler packages work with all of you.
Some time ago I did a split of them into two:
- armel-cross-toolchain-base (1.36 now)
- armel-cross-toolchain (1.29 now)
Where first one provides binutils, linux headers, libc6 and libgcc
packages. Second provides final gcc.
Today I got a-c-t-base to a moment when it builds fine on PPA [1]. 1.36 got
sent for rebuild to fix missing gcc-4.5-arm-linux-gnueabi-base package. When
it will build then a-c-toolchain package will get uploaded for build.
Result will deprecate my current repository at people.canonical.com [2]
because PPA gives signed repository.
On Monday I will probably have to update both components because there was
gcc-4.4 upload so probably gcc-4.5 will follow (so I will be able to drop one
patch).
Additionally I made 'gcc-defaults-armel-cross' package (available in [1])
which makes installing of cross compilers a bit easier (no need to worry which
version to install - just "apt-get install gcc-arm-linux-gnueabi" is enough).
Selection of cross gcc version is done in other way then native one. Native is
using "gcc" package which contains /usr/bin/gcc as symlink to /usr/bin/gcc-4.4
file. Cross gcc uses "update-alternatives" to setup /usr/bin/arm-linux-
gnueabi-gcc file. I want to fix it in 11.04 so cross gcc will use same method
as native one.
1. https://edge.launchpad.net/~hrw/+archive/arm-cross-compiler
2. http://people.canonical.com/~hrw/ubuntu-maverick-armel-cross-compilers/
Regards,
--
JID: hrw(a)jabber.org
Website: http://marcin.juszkiewicz.com.pl/
LinkedIn: http://www.linkedin.com/in/marcinjuszkiewicz
Michael,
a quick update to our discussion today: actually, GDB 7.2 has already been
released earlier today:
http://sourceware.org/ml/gdb-announce/2010/msg00003.html
Mit freundlichen Gruessen / Best Regards
Ulrich Weigand
--
Dr. Ulrich Weigand | Phone: +49-7031/16-3727
STSM, GNU compiler and toolchain for Linux on System z and Cell/B.E.
IBM Deutschland Research & Development GmbH
Vorsitzender des Aufsichtsrats: Martin Jetter | Geschäftsführung: Dirk
Wittkopp
Sitz der Gesellschaft: Böblingen | Registergericht: Amtsgericht
Stuttgart, HRB 243294
Hi Alexander. I've looked into the problem and the linker error is
caused by a mix of stack protector options between libgcc and the C
library.
GCC includes a feature called the stack smashing protector which
detects writing past the end of a stack based object. It's quite nice
as it gives decent protection against buffer overruns which are the
most common type of security vulnerability.
The implementation is straight forward: when compiled with
-fstack-protector, any function with a stack-based character array
will have extra stack checking code inserted into the prologue and
epilogue. The prologue allocates a canary value at the top of stack
and fills it in with the value of '__stack_chk_guard' provided by
libssp. The epilogue checks this value and calls `__stack_chk_fail`
if it has been changed. The stack protector can interfere with some
code and isn't applicable in others.
The problem here is caused by a stack up of things:
* glibc knows about -fstack-protector and turns it on and off for
different functions and libraries
* gcc knows about -fstack-protector and includes libssp if required
* glibc knows about libgcc and statically links against it to ensure
availability
* Meego seems to turn on -fstack-protector by default (as does Ubuntu)
This results in the libgcc function '_gcc_Unwind_Backtrace' being
built with the stack protector and the glibc library 'libanl' without.
At static link time GCC sees that the stack protector is off and
skips linking against libssp, causing the missing symbol error.
The solution is to add -fno-stack-protector to the libgcc build
options and rebuild the compiler. I've heard (but can't track down
the link) that the ARM libgcc unwind functions must be built this way
in any case.
See
http://svn.debian.org/wsvn/gcccvs/branches/sid/gcc-4.5/debian/patches/gcc-d…
for how Debian does this.
Hope that helps,
-- Michael
On Fri, Aug 27, 2010 at 9:06 PM, <Alexander.Kanevskiy(a)nokia.com> wrote:
> Hi Michael.
>
> I've created for you account in MeeGo OBS (build system that we use in MeeGo
> is OpenSuSE build system)
>
> login: michaelh
> password: wog-feg-da
> Web client url: https://build.meego.com
> API url: https://api.meego.com
>
> The build log that had problem with glibc 2.12 + gcc 4.5 you can find here:
>
> https://build.meego.com/package/live_build_log?arch=armv7el&package=glibc&pr
> oject=home%3Akad%3Abranches%3ATrunk%3ATesting&repository=standard
>
> Might be you have some idea what went wrong, as our toolchain people were
> not able to find why combination of latest gcc plus glibc 2.11.x works, but
> not gcc 4.5 + glibc 2.12.0 :(
>
> This log is from my home project inside OBS, where stuff is already a bit
> outdated. I'll ask Jan-Simon from Linux Fundation to point to right place
> where latest builds are present, so you can experiment with them.
>
> --
> Best regards, Alexander Kanevskiy.
>
>
>
