[LLVM-203] Code-Size investigation with PGO
- Finished up the remain tasks and resolved the ticket.
- Tidied up the patches and added tests so that I can potentially
share them with other interested parties.
-- Writing tests took a lot longer than I thought as it involves
faking a profile and there were quite a few caveats to doing that.
Some prototyping in LLVM and LLD for some internal ABI discussions.
* 1 day off
== Progress ==
* FDPIC
- Cortex-M fix for binutils committed. uClibc-ng patch on-going
- GCC: handling feedback on v3 patches.
Trying to build/check xtensa uclinux toolchain to check my patches
do not affect it. No satisfactory reference build yet.
* GCC upstream validation:
- reported a few regressions
- one of last week's regressions is causing timeouts on some
validations, thus delaying the full queue. I haven't isolated the
cause yet
- dealing with some random results, again
* GCC:
- bug report on aarch64 about misaligned accesses. Waiting for more
details to reproduce the problem.
* misc (conf-calls, meetings, emails, ....)
- Benchmarking harness hopefully fixed
- reviewing infra script patches
- bugzilla cleanup
== Next ==
FDPIC:
- GCC: followup v3 patches
- uclibc-ng: look at how to test fdpic mode with openadk
- use qemu-system mode to run more tests
Benchmarking:
- fix harness until they support gcc-8
=== Work done during this past week ===
* GNU-296 / GCC PR85434 / CVE-2018-12886: upstream review
+ avoid useless cmp after eors
+ fix another missing earlyclobber
+ full Thumb-1, Thumb-2 and Arm bootstrap with testing with and
without -fPIC, with and without -fstack-protector-all
* Continue patch to further update cpus and architectures in bfd:
external review
+ also teach GAS about Armv5TEJ and newer wrt. bfd architectures tags
+ fix ld action in run_dump_test
+ add testing for all this and submit for external review
* Resurrect code size optimization for fmul/dmul from Tony Wang:
+ rebase on trunk
* Line management
=== Plan for week 43 ===
* LLVM-432 (Support arithmetic on FileCheck regex variable):
+ extend testcase coverage (add tests for latest syntax change and
add more negative testing)
+ finish cleaning up the code
* Try to reproduce perf issue mentioned in week #30's weekly report on
latest perf
* Line management
Progress:
* Attended OSS Europe/KVM Forum
- as usual, a good week (helpful conversations, interesting sessions)
- full conference writeup later in the week
thanks
-- PMM
[VIRT-198 # QEMU: SVE Emulation Support ]
Last of the patches merged to mainline. Epic is now closed.
[VIRT-282 # QEMU: Accelerate TCG with KVM ]
A hallway talk with Paolo lead to a write-up, and Alex encouraged me to create
the epic. The epic description contains a link to the write-up, if anyone is
interested. I'd like to at least create a vm, measure some round-trip costs,
and properly gauge the level of difficulty. Beyond that... we'll see.
[Upstream]
Patch review:
- risc-v decodetree patches v2.
Produced some patches against decodetree itself in response.
I'm hopeful to see a much cleaner v3.
- per-cpu locks
[KVM Forum]
- Unsurprisingly, lots of people worked on speculation mitigation this year.
- Lots of focus on "ram", and the allocation and management thereof.
- Four talks on improving nested virtualization.
- The rest to follow in the trip report.
r~
== Progress ==
* FDPIC
- Posted binutils and uClibc-ng patches to fix cortex-M support, under
discussion
- GCC: handling feedback on v3 patches.
Experimented thumb-1 builds, failed.
* GCC upstream validation:
- reported a few regressions
- dealing with some random results, again
* GCC:
- bug report on aarch64 about misaligned accesses. Waiting for more
details to reproduce the problem.
* misc (conf-calls, meetings, emails, ....)
- Working on fixes of our benchmarking harness to support new gcc-8 releases.
== Next ==
FDPIC:
- GCC: followup v3 patches
- uclibc-ng: look at how to test fdpic mode with openadk
- use qemu-system mode to run more tests
Benchmarking:
- fix harness until they support gcc-8
== This Week ==
* GNU-405: Implement division using vrecpe / vrecps (4/10)
- Patch validated and posted upstream.
* SVE ACLE intrinsics (4/10)
- Going thru documentation.
* GNU-235: Provide value-range info for erf family of functions (1/10)
- Working on patch.
* Misc (1/10)
- Meetings
== Next Week ==
- Continue GNU-235, SVE ACLE intrinsics
Upstream Work ([VIRT-109])
==========================
- started looking at {Qemu-arm} {RFC v4 00/71} per-CPU locks
Message-Id: <20181025151103.GA19931@flamenco>
- this is a precursor to Emilio's {RFC 00/48} Plugin support
Message-Id: <20181025172057.20414-1-cota(a)braap.org>
[VIRT-109] https://projects.linaro.org/browse/VIRT-109
Other Tasks
===========
- Attended KVM Forum 2018
- I am now Spectre'd out ;-)
- Some interesting discussions on upstream CI
- will save the rest of my notes for the conference report
- Delivered QEMU Keynote/Status report @ KVM Forum 2018
- Here are [the slides]
[the slides]
http://people.linaro.org/~alex.bennee/org/presentations/kvm18-qemu-keynote.…
Current Review Queue
====================
* {Qemu-arm} {RFC v4 00/71} per-CPU locks
Message-Id: <20181025151103.GA19931@flamenco>
* {RFC 00/48} Plugin support
Message-Id: <20181025172057.20414-1-cota(a)braap.org>
* {PATCH v2 0/3} Modern shell scripting (use $() instead of ``)
Message-Id: <20181018031723.23459-1-maozhongyi(a)cmss.chinamobile.com>
* {PATCH 0/7} Acceptance Tests: basic architecture support
Message-Id: <20181004151429.7232-1-crosa(a)redhat.com>
* {PATCH v7 00/19} Fixing record/replay and adding reverse debugging
Message-Id: <20181010133333.24538.53169.stgit@pasha-VirtualBox>
* {PATCH v2 0/3} Bootstrap Python venv and acceptance/functional tests
Message-Id: <20181009041826.19462-1-crosa(a)redhat.com>
--
Alex Bennée
[LLVM-203] Investigation into profiling and code-size optimizations
- Collected the remaining data I needed over the weekend.
- Wrote up report
- Rebased patches on tip of trunk
- Attached results and report to Jira issue.
- A one line summary of the results is that if you are lucky you can
get close to peak performance at close to Os code size if your program
happens to spend most of its time in a few small places. If you are
unlucky then increased inlining and unrolling can still result in an
overall code size increase over -O3 but the effect will be limited.
[LLVM-158] Monitor and maintain buildbots
- Relatively quiet week, a couple of patches pinged for fixes/reverts.
== This Week ==
* TCWG-1234: Coremark regression (7/10)
- Fixed golang regressions with the patch.
- Posted patch upstream to change hoisting order and apply cost model.
* Public holiday (2/10)
* Misc (1/10)
- Meetings
== Next Week ==
- TCWG-319, SVE
