== This Week ==
* TCWG-777 (3/10)
- Fixed ICE with the patch - toolchain builds with the patch.
- Sent patch to tcwg list for review
- Investigating why combine fails to combine
arm_andsi3_insn/arm_cmpsi_insn into zeroextractsi_compare0_scratch
when the pass uses ud-chains to find def but works
when def is found using ad-hoc way.
* TCWG-871 (4/10)
- Getting familiar with firefox build system
- LTO build lto/non-lto on x86 and arm (doc)
- Figuring out how to resolve "plugin needed to handle lto object" error.
tried the following:
binutils configured with: --enable-plugin, --enable-lto
gcc configured with: --enable-lto --with-ld-plugin=<just built ld>
Doesn't appear to work.
* TCWG-835 (1/10)
- Build failure with spec (PR67399)
* Misc (2/10)
- Meetings
- Looked at rtl dataflow (df.h and df*.c) and ree.c
== Next Week ==
Continue with TCWG-777, TCWG-835, TCWG-871
== This week ==
* TCWG-832 -Exploit vector multiply by scalar instructions when multiple
scalars are used as
coefficients in a loop (2/10)
- Initial investigation.
* TCWG-833 - Exploit Wide Add operations when appropriate (4/10)
- Ramana is reviewing Aarch32 patch
- Recoded Aarch64 support to use vect_select
- Debugging Aarch64 lto, tree-dump regression suite failures
* TCWG-834 - Use non-unit stride loads by preference when applicable (1/10)
- TCWG 834 is Bugzilla 67323 upstream; Richard Biener took ownership
as a vectorizer failure
- My plan was to write a test case that failed until fixed, but
Richard indicated this is not
standard practice
- Bugzilla 57195 (mode iterator bug) blocked compiling new pattern (1/10)
- Separated patch from Bugzilla 67321 patch and sent upstream
- Pinged upstream for comments
- Bugzilla 67320 - Incorrect standard names for wide addition (1/10)
- committed upstream in trunk
* Misc (1/10)
- Conference calls
== Next week ==
- Resolve Aarch64 TCWG-833 patch, validate and upstream
- Additional investigation into TCWG-832
# Progress #
* TCWG-857 [3/10]
With one kernel fix, HW breakpoint works for unaligned address
(2-byte aligned).
Debug linux kernel with KGDB. KGDB exposes an existing GDB bug, I
fixed it, but need some time thinking about how to submit it upstream.
* TCWG-567, arm watchpoint fixes [4/10]. All fails are fixed, but need
another round of test to confirm. ARM HW watchpoint doesn't work on
4.0.0 kernel, but I don't investigate on it.
* TCWG-757 [2/10], some patches review.
* Misc, meeting, [1/10]
# Plan #
* TCWG-857.
* Upstream the leftover of aarch64 multi-arch patches.
--
Yao
== Progress ==
o Linaro GCC validation (9/10)
* Analysed x86_64 fstack-protector and sse2 issue
* This is due to "ulimit -v" usage which brakes asan testing
and gcc testsuite caching mechanism.
* Validation is now stable on Hetzner/Austin hardware
* Documented summer validation issues
* Validate and committed on-going backports
o Upstream GCC (0/10)
* Armeb OOM fix committed into gcc-5
o Misc (1/10)
* Various meetings
== Plan ==
o 3 days off
o Continue backports/validation
== Progress ==
* Widening pass (TCWG-547) - 8/10
- Fixed all but one execution test failure.
- aarch64 and x86 are clean
- arm has one but this looks like a latent issue (in expand); looking
into it
- Latest trunk with aarch64 miscompiles stage2 fwprop (-fno-forwprop
works).
This happens with the commit 94f92c36a83d66a893c3bc6f00a038ba3dbe2a6f
[PR64164] Drop copyrename, use coalescible partition as base when
optimizing
- Tried forcing the same promote_mode for x86_64 and can reproduce it
with x86_64 also.
- Looking into it to see if this is an error with the commit
* Misc - 2/10
- gcc-patches, gcc-bugs list
== Plan ==
- Continue with widening pass
Upcoming Absences
* Away from this Wednesday until next Tuesday, inclusive
Benchmark infrastructure - TCWG-360 [1/10]
* Finished filling out cards/bugs for benchmarking work
* Which makes the rest of this report more fine-grained
Multinode wrapper - TCWG-888 [2/10]
* Completed oversubscription workaround
* Herded most tests through
* Uncovered one significant bug, seems fixed
Centralized source/results storage - TCWG-722 [1/10]
* Wrote draft of source/results handling rules
* Confirmed that old repos are subsets of new repos
Noise control experiments - TCWG-897 [2/10]
* Learned more about firmware, openembedded
* Juno running a minimal filesystem, needs some tidying up
Misc [4/10]
* Including a further 1/10 of ARM management
=Plan=
* Herd remaining multinode tests through
** Depends on LAVA lab coming back from power cut
** And queues being short enough
* Progress (finish?) Juno bring-up
* Share source/results rules draft
== This week ==
* TCWG-833 - Exploit Wide Add operations when appropriate (8/10)
- Reworked patch to use vect_select instead of unspec as requested
by Ramana
- Bugzilla 57195 (mode iterator bug) blocked compiling new pattern
- Created patch which successfully bootstrapped
- Created Bugzilla 67320 - Incorrect standard names for wide
addition (documentation bug)
- Successfully regression tested Aarch32 changes
- Send Aarch32 patch upstream for review
- Reworking patch for Aarch64 to use vect_select as well
- Created Bugzilla 67321 - [ARM] Exploit wide adds when appropriate
- Created Bugzilla 67322 - [Aarch64] Exploit wide adds when appropriate
* TCWG-834 - Use non-unit stride loads by preference when applicable (1/10)
- Created Bugzilla 67323 to track until fixed in GCC 6 by Richard Biener
- Began writing test case that fails
* Misc (1/10)
- Conference calls
== Next week ==
- Respond to Aarch32 TCWG-833 upstream requests and hopefully check-in patch
- Complete Aarch64 TCWG-833 patch, validate and upstream
- Finish testcase for TCWG-834 and submit upstream
== Progress ==
* Maintenance (CARD-1833 2/10)
- Fixing libc++abi build on AArch64
- Trying to remove a hack in ARMTargetInfo about default CPUs
- Bisecting PR24292
- Working with ARM to fix it, backport
* Buildbots (CARD-1823 8/10)
- Working with Adhemerval on VMA 42bits sanitizer
- Setting up a libc++ "full" build on the second stage, since
gcc can't build it yet. 171 tests fail.
- Ubuntu Vivid (39-bits VA / 4k pages) with Uboot is
the winner of stability, speed and easiness.
- Power outage meant I had to unbreak *a lot* of broken stuff
* Background (2/10)
- Code review, meetings, discussions, etc.
- Some Android/TSAN/AArch64 shenanigans
== Plan ==
Holidays all week
== Issues ==
Yes, it does add up to 12/10, since I had to work most of Saturday and
some of Sunday to fix the mess that the power cuts made on our lab.
If we had generators, that wouldn't have happened.
