== This week ==
* TCWG-146 - Detect smin/umin idion (1/10)
- Patch sent upstream for approval
* TCWG-140 - Transform end of loop conditions to min_expr (4/10)
- Patch and investigating validation regressions
* TCWG-833 - Exploit Wide Add operations when appropriate (4/10)
- Investigation into why vectorizer does not exploit wide adds
* Misc (1/10)
- Conference calls
- Conference call with Kugan and Prathamesh to discuss GCC Git workflow
- Conference call with Charles and Prathamesh to discuss
autovectorization
== Next week ==
- Vacation
== Progress ==
(TCWG-831) post-indexed addressing [3/10]
. vectorization project kick-off call
. code browsing/reading to understand mailing list feedback about previous patch
(TCWG-775) NEON error messages [6/10]
. completed conversion of some ARM intrinsics to give same error
messages as AArch64 work
. reworked tests so they can be shared between AArch64, ARM.
. re-submitted previous patch with updated tests
Misc [1/10]
email, irc, gerrit reviews, connect travel booking, AArch64 qemu
big-endian experiment
== Plans ==
submit patch for work done so far on ARM NEON error messages
cortex-a53 workarounds
Benchmark automation - TCWG-360 [3/10]
* Created a partial Jenkins prototype
* Considered some security issues
Benchmarking presentation [5/10]
* Drafted some slides, did some reading
Misc [2/10]
=Plan=
More of the above
== Progress ==
* TCWG-849 (1/10)
- Committed improvement for VRP
https://gcc.gnu.org/viewcvs/gcc?view=revision&revision=225108
* Add REG_EQUAL for arm_emit_movpair (4/10)
- Posted patches for review
* TACT -TCWG-851 (3/10)
- Started with the small examples.
- Ran into an error while tuning; looking into it
* Git work flow for upstream patches -TCWG-848 (1/10)
- Had a chat with Michael and Prathamesh
- Tried the work-flow and now started documenting them
* Misc (1/10)
- gcc-patches, gcc-bugs list
- Meetings
== Plan ==
- GCC Bugs
- TACT driven optimization exploration for gcc
* TCWG-830 (4/10)
- Observing tree dumps
- Peeling for alignment happens at -O3 but not at -O2 -ftree-vectorize
Reason: in vect_enhance_data_refs_alignment() for:
a) -O2 -ftree-vectorize: max_allowed_peel == 0
b) -O3: max_allowed_peel == (unsigned) -1;
which equals UINT_MAX and therefore peeling gets allowed.
- Workaround: Pass -param vect-max-peeling-for-alignment=0
- Peeling for alignment with O2 can be enabled by passing
-fvect-cost-model (we don't want this!)
Reason:
opts.c:
/* Tune vectorization related parametees according to cost model. */
if (opts->x_flag_vect_cost_model == VECT_COST_MODEL_CHEAP)
{
maybe_set_param_value (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS,
6, opts->x_param_values, opts_set->x_param_values);
maybe_set_param_value (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS,
0, opts->x_param_values, opts_set->x_param_values);
maybe_set_param_value (PARAM_VECT_MAX_PEELING_FOR_ALIGNMENT,
0, opts->x_param_values, opts_set->x_param_values);
}
The above if condition becomes false when -fvect-cost-model is passed.
- Proposed patch (untested): http://pastebin.com/ftp0mrwH
Patch follows the workaround and passes --param vect-max-peeling-for-alignment=0
if unaligned access is supported.
* TCWG-777 (4/10)
- Observing tree and rtl dumps
- Workaround: for -O1 pass -fno-tree-fre -fno-tree-dominator-opts
Test-case: http://pastebin.com/cjBcSpiT
Generated assembly at -O1 without workaround: http://pastebin.com/jmQGZhN9
Generated assembly at -O1 with workaround: http://pastebin.com/JGj05z66
Is that the expected output for no unnecessary temps in assembly with
workaround ?
Is it profitable over the assembly generated without workaround ?
- Approach currently taken:
a) New pass "remove-temps" (for lack of better name), after nrv (added
as last gimple pass).
b) Transforms:
if (ssa_var != 0)
to
new_ssa_var = SSA_NAME_DEF_STMT (ssa_var)
if (new_ssa_var != 0)
This "unfolds" cse on expressions within if, which was done by fre
(and if fre was disabled then by dom pass).
c) However this approach results in dead stores.
eg:
_8 = flags_7(D) & 1;
if (_8 != 0)
...
is transformed to:
_8 = flags_7(D) & 1;
_32 = flags_7(D) & 1;
if (_32 != 0)
...
so store to _8 is dead store.
I tried to run dse after remove-temps but that didn't work.
RTL 194r.jump eliminates the above dead store as "trivially dead insn".
However I don't think it's a good idea to have dead stores like these
in gimple and rely
on RTL to eliminate them. I could try to make the pass bit smarter to
not generate redundant stores like _32 != 0 in above case.
d) Patch (no intent to commit as-is): http://pastebin.com/AGXnSkrZ
Generated assembly at -O1 with the patch: http://pastebin.com/VmHCVpGC
Patch eliminates temporaries at -O1 but not at -O2.
I have not yet figured out the reason for that.
For if (flags & 1),
In dfinish pass for -O1, the generated RTL is from
zeroextractsi_compare0_scratch
while for -O2, the generated RTL is from andsi3_compare0
e) Is this a problem also on x86 ?
x86 generated assembly with -O1: http://pastebin.com/XMeTXXwK
* Misc (2/10)
- Getting familiar with vectorizer and NEON gcc intrinsics
- Reviewed git tutorials and starting preparation of git doc
- Conference calls
== Next Week ==
- Continue working on TCWG-830 and TCWG-777
- Header file flattening
- Travel to Mumbai on 2nd July (Thursday) for US Visa OFC appointment.
== Progress ==
* Maintenance (CARD-1833 4/10)
- Found the trail on the ADD/SUB with negative immediate
- Submitting RFC for discussion (http://llvm.org/PR20978)
- Bugzilla farming
- More LNT investigations (http://llvm.org/perf/ unstable)
* Releases (CARD-1431 1/10)
- Building, testing and uploading 3.6.2 RC1
* Background (5/10)
- Code review, meetings, discussions, etc.
- More stride vectorizer code review (lnN/stN implementation)
- More lab discussions (routers, lab split, new link)
- Changing my dev env to emacs (huge mind set flip)
== Plan ==
* Continue with ADD/SUB change
* Continue with Emacs setup
* Move benchmark bot to CMake
* Some other bugs
* One day off on Thu [2/10]
# Progress #
* Linaro GDB [4/10]
** TCWG-805, aarch64 native debugging multi-arch support.
Prepare for the patches submission.
It is a big patch series, and think about how to upstream them.
Write commit log including the rationale of the changes.
* FSF GDB [2/10]
** FSF GDB 7.10 release. Audit some GDB regressions caused by intel
mpx stuff.
** PR 18605. Write a patch and it is in testing.
** Other patches review.
* Misc [2/10]
** File expense report for Grenoble travel.
** Some discussions on aarch64 tracepoint.
# Plan #
* TCWG-805, upstream some patches on multi-arch debugging.
--
Yao
* One day off (Wed) (2/10)
== Progress ==
* linaro-5.1-2015.06 snapshot (1/10)
- dealt with tags, release notes
- shared it with B&B
* 4.8-2015.06 branch merge (1/10)
- investigated regression: incorrect automatic merge
- fixed, validation on-going
* 4.9 branch (2/10)
- updated our git linaro-4.9-branch to match the svn one
- ready for branch merge, will be done right after fsf release
* Misc (4/10)
- meetings, conf-calls, emails, reviews (GCC backports, ABE, backflip)
== Next ==
* more reviews for new backports
* backports, release, validation: update doc
* hopefully upstream work
Recently I came across two excellent post about accelerating clang/llvm
build with different compiler/optimization [1] [2].
I tried some of author advices getting very good results. Basically I
moved to optimized clang build, changed to gold linker and used another
memory allocator than system glibc one. Results in build time for all
the clang/llvm toolchain is summarized below (my machine is a i7-4510U,
2C/4T, 8GB, 256GB SSD):
GCC 4.8.4 + gold (Ubuntu 14.04)
real 85m17.640s
user 257m1.976s
sys 11m35.284s
LLVM 3.6 + gold (Ubuntu 14.04)
real 34m4.909s
user 128m43.382s
sys 3m51.643s
LLVM 3.7 + gold + tcmalloc
real 32m56.707s
user 121m40.562s
sys 3m52.358s
Gold linker also shows a *much* less RSS usage, I am able to fully use make -j4
while linking in 8GB without issue any swapping.
Two things I would add/check for the posts:
1. Change from libc to tcmalloc showed me a 3-4% improvement. I tried jemalloc,
but tcmalloc is faster. I am using currently system version 2.2, but I have
pushed an aggressive decommit patch to enable as default for 2.4 that might
show lower RSS and latency (I will check it later).
2. First I try to accelerate my build by offloading compilation using distcc.
Results were good, although the other machine utilization (i7, 4C/8T, 8GB)
showed mixes cpu utilization. The problem was linking memory utilization
using ld.bfd, which generates a lot of swapping with higher job count. I
will try using distcc with clang.
[1] http://blogs.s-osg.org/an-introduction-to-accelerating-your-build-with-clan…
[2] http://blogs.s-osg.org/a-conclusion-to-accelerating-your-build-with-clang/
Benchmark automation - TCWG-360 [7/10]
* Arndales stopped booting
** Package servers for elderly filesystem had gone
** Investigated some approaches to creating more stable filesystems
** Realized I could just updated image to point at old-releases, so
did that for now
* _More_ time thinking about interactions with Jenkins & LAVA. Fathi
gave me some Jenkins jobs to prototype in.
* Brain-dumped some of the present state of things into Collaborate
Misc - [3/10]
=Plan=
Jenkins prototyping
