== Week of June 30th ==
- Continued working on toolchain testing improvements (CARD-1378, 6/10)
-- Troubleshooted Jenkins jobs
-- Various cbuild2 improvements
- Migrated master VM of TCWG dev environment to AWS (2/10)
-- It is now at tcwg-dev-env-master.linaro.org
- Various meetings and discussions (2/10)
-- Attempted to test QEMU's aarch32 crypto support
-- Organized travel for GCC GSoC student to GNU Tools Cauldron. Thanks Linaro for providing sponsorship for this!
--
Maxim Kuvyrkov
www.linaro.org
=Progress=
memcpy regression on A9 - TCWG-390 [6/10]
* Not down to the branch predictor after all
* Can be fixed for A9 with explicit pld, but this isn't free
cbuild2 benchmarking - TCWG-360 [1/10]
* Dug out my branch, merged in trunk, refreshed memory
Meetings/mail/etc [3/10]
* Including a little patch review for Roland's lowlevellock.h (TCWG-435)
=Plan=
Close TCWG-390 as wontfix
Lots more cbuild2 benchmark automation
Keep up with lowlevellock.h
== Progress ==
* Recovered from flu on leave 7th and 8th July (4/10)
=LTO bootstrap -tcwg180 (2/10)=
Git bisect experiments shows bootstrap compare errors apprearing and
disappearing at trunk at different revisons.
Alteast 3 revisons LTO bootstrap passes
r210740 2488876068f541a341472c3aeedadccc255f0696 PASS
r210854 271fe9cf111f6b581451bbfb4d346757877896cd PASS
r210973 487fea840105dbb0ff516b797e7d86bea82ef74a PASS
It is hard to do backport of revisons which fixes this bug. Need to
discuss with LTO people in GCC mailing list
=Reload spill fix -tcwg180 (2/10)=
Emailed my status on benchmarking
https://gcc.gnu.org/ml/gcc-patches/2014-07/msg00450.html
This bug cannot be fixed without performance degradation. So if it
cannot be up streamed, the reporter can use the patch to build his
package atleast. Planning to close this bug if I did not receive any
comments from maintainers.
=Misc (2/10)=
* GCC internal team meeting
* Meeting with Ryan
== Plan ==
* Benchmark coremark -O3 vs -O3 -flto.
* Email to marcus on ABI for large memory model in Aarch64.
* Cauldron preparation.
== Planed leaves ==
* Jul 18- 25: Travel to Cambridge.
== Progress ==
- 8th and 9th on holiday (4/10)
- Zero/sign extension elimination (TCWG-291) 3/10
* Patch one is accepted
* Posted the modified patch 2 after some discussions
* Started analysing the code generated for coremark and spec2k
* Looks there are more places that can be improved. I will post
additional patches after completing it
- Launchpad bugs (3/10)
* https://bugs.launchpad.net/gcc-linaro/+bug/1320965
* https://bugs.launchpad.net/gcc-linaro/+bug/1331112
== Plan ==
- sha1 regressions
- 15th and 16th on holiday
== Progress ==
* One day off.
* Test and send out the local NLS patch for community review (1/10).
* Test ccmp patches. Can not reproduce a FAIL in previous regression
test with the latest trunk(TCWG-488, 1/10).
* Test codes to skip arm_split_constant (TCWG-486, 3/10). Performance
results seam OK for spec2000, coremark dhystone and eembc.
* R/M toolchain related work (3/10).
== Plans ==
* Ping pending patches.
* Update ccmp patches.
== Planed leaves ==
* Jul. 15 - 25: Travel to Cambridge.
* Aug. 4 -8: Annual leaves.
== This week ==
Provide ldp/stp peephole optimization for Aarch64 [TCWG-446] [0/10]
- Halted development due to duplicate ARM development project by
Bin Cheng
Launchpad 1318831 - Invalid unpoisoning of stack redzones on ARM [3/10]
- Verifying test results
Launchpad 1267761 - miscompilation of unsigned comparison on aarch64 [2/10]
- Testing under way on Jenkins
Neon intrinsic tests [4/10]
- background work to gain familiarity with effort
Misc Meetings [1/10]
== Next week ==
- Verify testing on Launchpad 1267761 and 1318831
- Begin investigating compiler bugs discovered by neon intrinsic tests
== Future ==
== Progress ==
* GCC trunk cross-validation
- reported a few regressions
* Automation Framework (2/10)
- looked at Jenkins configuration & logs of some failures
* AArch64 libsanitizer (1/10)
- tried to use cbuild2 in the lab to build & test on HW, but all my
build attempts failed for various reasons
* Neon intrinsics tests (3/10)
- feedback from upstream requests some renaming changes + GNU coding
style fixes
- since reformating the existing testsuite is a manual and
error-prone task, it will bring additional delay in improving GCC
testsuite.
- it's probably more efficient to run the existing tests regularly
and start reporting & fixing the bugs identified, not waiting for the
full conversion
- updated the existing tests for fp16 and poly* types support with
GCC (reference built with armcc)
- interestingly, compiling the original testcases with armcc for
aarch64 caused ICE on 28 source files.
* Misc (meetings, conf-calls, ....) (4/10)
== Next ==
Holidays for 2 weeks
== Progress ==
* Monday holiday
* Automation Framework (CARD-1378 6/8)
- TCWG D01 moved in, Chromebooks out
- Movng LLVM buildbots to D01s
- D01s have no NEON support... :S
- We might have to move some Chromes back in
- Moving hackboxes to Chromebook 2s
* Background (2/8)
- Code review, meetings, discussions, etc.
== Plan ==
* Deal with NEON issue in the rack
* Finish Cauldron presentation
* Try to get the D01s as buildbots, even without NEON
== Progress ==
* AArch64 GDB reverse watchpoints issue. [TCWG-503] [4/10]
-- Found a new fix in generic process record implementation.
Testing stalled by too many 2500 failures seen on arm-linux-gdb in
general and some 850 record/replay implementation.
* AArch64 GDB handling of functions with empty prologue. [TCWG-504] [4/10]
-- Analysis of prologue generated by aarch64-gcc
-- Make an understanding of code implementation for arm-linux-gdb.
-- Try to reproduce reported bugs.
* Testing and Analysis of testsuite failures in arm-linux-gdb [TCWG-509] [1/10]
* Browsing gdb related upstream discussions. [1/10]
== Plan ==
* AArch64 GDB reverse watchpoints issue. [TCWG-503]
-- Test and submit fix for target record for the case where we cant
step breakpoints.
* Investigate issues in arm-linux-gdb testsuite results. [TCWG-509]
* Travel preparations to UK for Tools cauldron and TCWG sprint.
-- Obtain health certificate for international travelers.
* On Holiday from 14th to 17th July 2014.
=Progress=
memcpy regression on A9 - TCWG-390 [7/10]
* Tweaked bench.py some more
* Cobbled together a bare metal version of cortex-strings benchmark
* Was surprised to find that the problem does not relate to Linux
* Have some evidence that points at the branch predictor
Meetings/mail/etc [3/10]
=Plan=
See if A9 performance is recoverable without sacrificing other targets
Push some of my improvements into the cortex-strings repo
Possibly write up my cortex-strings-benchmarking-in-lava method
Back to cbuild benchmark automation
