== This week ==
* Submitted a fix for the performance regression caused by my
arm_comparison_operator patch. Applied upstream after approval
from Ramana (thanks). Will backport to Linaro towards the end
of next week if there are no reported problems.
* Went back to looking at -fsched-pressure. To recap, a colleague
ran SPEC for s390 comparing:
(a) normal -O3 based flags
(b) (a) + -fsched-pressure without my patch
(c) (a) + -fsched-pressure with my patch
(c) got the best geomean result, but there were some individual
tests for which (b) was significantly worse than (a), and for
which (c) only partly closed the gap.
Found one problem. It looks like -fsched-pressure only really
operates on the issue rate and instruction latencies; it doesn't
seem to use the DFA. This seems to be unintentional, and fixing
it showed some nice results.
Also, the -fsched-pressure patch that I wrote at Connect set the
starting pressure based on the set of registers that are both live
on entry to the block _and_ used within the enclosing loop,
This still seems to be a bit too conservative, in that it makes
the scheduler go out of its way to preserve loop invariants,
even if there are too many of them. Experimented with changing
"used" to "defined". This too seemed to be a win.
* Got access to some PowerPC GNU/Linux machines that are suitable
for running SPEC. Set up my account there and got SPEC building.
The idea is to use this to get more cross-target evidence for the
-fsched-pressure submission(s).
* Discussion about the SMS register-scheduling patches after great
feedback from Ayal. While drafting a still-unsent reply justifying
the main part of the patch, I found I was also explaining why another
part of the patch (specifically the prologue/epilogue part) was wrong.
Thought about that a bit today.
* Submitted fix for LP 641126.
== Next week ==
* More SMS register scheduling.
* More -fsched-pressure.
* Hopefully remerge the arm_comparison_operator patch with this week's fix.
Richard
* Working on getting everything in place for cross-compiling Firefox for
ARM. Trying to understand how the configuration script and make file works.
* Working on a test that will run Sunspider and extract the results. The
challenging part is that results are only presented on the page, not e.g.
written to stdout or to file. My approach to create an html file, embed the
page with the test in an iframe, and read out the results when the test is
done.
* Running SPEC2K on the Snowball board. An updated kernel solved the issue
with great variations in the test results. Some tests results look a bit
strange, so I will look at what those tests do to see what part of the
system is stressed.
Best Regards
Åsa
Hi,
* widening shifts patch - submitted upstream
* change default vector size patch - submitted to linaro-gcc
* automatic choice of vector size for basic block vectorization - testing
* vectorizer bug fixes
Next week we have New Year holiday on Wednesday (half day) and Thursday.
Ira
The Linaro Toolchain Working Group is pleased to announce the 2011.09
release of both Linaro GCC 4.6 and Linaro GCC 4.5.
Linaro GCC 4.6 2011.09-1 is the seventh release in the 4.6 series. Based
off the latest GCC 4.6.1+svn178681, it contains a range of vectoriser
and core performance improvements as well as fixing a number of
bugs.
Interesting changes include:
* Updates to 4.6.1+svn178681
* Improves performance by making better use of conditional compares
* Improves performance by properly scheduling widening multiplies
* Improves size and speed by improving constant generation in Thumb-2
* Implements support for widening multiples in toe core
* Improves vectorised code by reducing the over-promotion of intermediates
* Improves performance by reducing redundant moves between VFP and ARM
* Finishes off supporting the Android team in integrating Linaro GCC
Fixes:
* LP: #823548 Can't use -flto with skia
* LP: #823711 libvirt version 0.9.2-4ubuntu8 failed to build on armel
* LP: #827990 internal compiler error: in decode_addr_const, at varasm.c:2632
* LP: #836401 ICE on a | (b << negative-constant)
* LP: #838994 ICE building perl w/ -marm
* LP: #843775 ICE optimizing widening multiply-and-accumulate
Linaro GCC 4.5 2011.09 is the fourteenth release in the 4.5
series. Based off the latest GCC 4.5.3+svn178560, this is a
maintenance focused release.
Interesting changes in 4.5 include:
* Updates to 4.5.3+svn178560
Fixes:
* LP: #823711 libvirt version 0.9.2-4ubuntu8 failed to build on armel
The source tarballs are available from:
https://launchpad.net/gcc-linaro/+milestone/4.6-2011.09https://launchpad.net/gcc-linaro/+milestone/4.5-2011.09
Downloads are available from the Linaro GCC page on Launchpad:
https://launchpad.net/gcc-linaro
More information on the features and issues are available from the
release page:
https://launchpad.net/gcc-linaro/4.6/4.6-2011.09https://launchpad.net/gcc-linaro/4.5/4.5-2011.09
Mailing list: http://lists.linaro.org/mailman/listinfo/linaro-toolchain
Bugs: https://bugs.launchpad.net/gcc-linaro/
Questions? https://ask.linaro.org/
Interested in commercial support? inquire at support(a)linaro.org
-- Michael
I tried to bootstrap current GCC trunk and our latest gcc-linaro-4.6
in profile guided, link time optimisation, and SMS modes. The results
are here:
https://wiki.linaro.org/MichaelHope/Sandbox/PGOLTOSMSStatus1
Short story: you can't bootstrap in LTO or PGO on ARM as they run out
of memory. i686 LTO is broken on trunk and gcc-linaro-4.6. SMS is
fine in general.
I'll run these once a week and keep an eye on them. A -fwhopr instead
of -flto may help on ARM. I don't know why the PGO build runs out of
memory.
-- Michael
Måns pointed me at the IDCT throughput test that's included with
libav. I've written up a page on how to build and run it at:
https://wiki.linaro.org/MichaelHope/Sandbox/LibAvDCT
Included are results with and without the vectoriser. In all cases
the vectoriser improves things, including increasing the SIMPLE-C
version by 11 % and the peak by 17 %.
The coefficient of variance is low so the results are consistent. I
haven't investigated the benchmark itself to see if its valid - we
could be vectorising the loop overhead instead of the IDCT itself.
-- Michael
Please coordinate with Jon Masters at RedHat/Fedora and Adam Conrad at
Ubuntu/Debian on this. (Cc'ing the cross-distro list, through which the
recent ARM summit at Linux Plumbers was organized.)
Cheers,
- Michael
On Sep 16, 2011 8:41 AM, "David Gilbert" <david.gilbert(a)linaro.org> wrote:
> OK, so we seem to have agreement here that what we want is autodetect
> for eglibc and
> forget about the triplet; well technically that probably makes my life
> easier, and I don't
> think it's too hard a sell.
>
> Dave
>
> _______________________________________________
> linaro-toolchain mailing list
> linaro-toolchain(a)lists.linaro.org
> http://lists.linaro.org/mailman/listinfo/linaro-toolchain
* Linaro GCC
Spun 4.5 and 4.6 2011.09 GCC release tarballs. Uploaded them to
Michael's server, and kicked off the tests.
Continued work on my new constant optimization experiments. I now have
it tracking all the constants and am looking at how to detect the
optimization opportunities. So far it only calculates how exprensive it
would be to generate a value by adding to an existing constant, which is
a start at least. I'm having difficulties detecting whether changing an
insn will make it's parent (dependency-wise) obsolete, or not (and
therefore whether to count its costs - there's no problem for
instructions that overwrite an entire register, but ones that write to
portions of registers (such as MOVT) make more complex dependency
chains, and the def-use chains don't seem to be sorted into the order of
use.
* Other
Half day vacation on Thursday.
