On 11 November 2015 at 00:45, Savolainen, Petri (Nokia - FI/Espoo) <
petri.savolainen(a)nokia.com> wrote:
>
>
> > -----Original Message-----
> > From: lng-odp [mailto:lng-odp-bounces@lists.linaro.org] On Behalf Of
> > EXT Nicolas Morey-Chaisemartin
> > Sent: Tuesday, November 10, 2015 5:13 PM
> > To: Zoltan Kiss; linaro-toolchain(a)lists.linaro.org
> > Cc: lng-odp
> > Subject: Re: [lng-odp] Runtime inlining
> >
> > As I said in the call last week, the problem is wider than that.
> >
> > ODP specifies a lot of types but not their sizes, a lot of
> > enums/defines (things like ODP_PKTIO_INVALID) but not their value
> > either.
> > For our port a lot of those values were changed for
> > performance/implementation reason. So I'm not even compatible between
> > one version of our ODP port and another one.
> >
> > The only way I can see to solve this is for ODP to fix the size of all
> > these types.
> > Default/Invalid values are not that easy, as a pointer would have a
> > completely different behaviour from structs/bitfields
> >
> > Nicolas
> >
>
> Type sizes do not need to be fixed in general, but only when an
> application is build for binary compatibility (the use case we are talking
> here). Binary compatibility and thus the fixed type sizes are defined per
> ISA.
>
> We can e.g. define a configure target (for our reference implementation ==
> linux-generic) "--binary-compatible=armv8.x" or
> "--binary-compatible=x86_64". When you build your application with that
> option, "platform dependent" types and constants would be fixed to
> pre-defined values specified in (new) ODP API arch files.
>
> So instead of building against
> odp/platform/linux-generic/include/odp/plat/queue_types.h ...
>
> typedef ODP_HANDLE_T(odp_queue_t);
> #define ODP_QUEUE_INVALID _odp_cast_scalar(odp_queue_t, 0)
> #define ODP_QUEUE_NAME_LEN 32
>
>
> ... you'd build against odp/arch/armv8.x/include/odp/queue_types.h ...
>
With the introduction of odp/arch at the top level I think we should also
move platform/linux-generic/arch to the same location
> typedef uintptr_t odp_queue_t;
> #define ODP_QUEUE_INVALID ((uintptr_t)0)
> #define ODP_QUEUE_NAME_LEN 64
>
>
> ... or odp/arch/x86_64/include/odp/queue_types.h
>
> typedef uint64_t odp_queue_t;
> #define ODP_QUEUE_INVALID ((uint64_t)0xffffffffffffffff)
> #define ODP_QUEUE_NAME_LEN 32
>
>
> For highest performance on a fixed target platform, you'd still build
> against the platform directly
>
> odp/platform/<soc_vendor_xyz>/include/odp/plat/queue_types.h
>
> typedef xyz_queue_desc_t * odp_queue_t;
> #define ODP_QUEUE_INVALID ((xyz_queue_desc_t *)0xdeadbeef)
> #define ODP_QUEUE_NAME_LEN 20
>
>
> -Petri
>
>
>
>
> _______________________________________________
> lng-odp mailing list
> lng-odp(a)lists.linaro.org
> https://lists.linaro.org/mailman/listinfo/lng-odp
>
--
Mike Holmes
Technical Manager - Linaro Networking Group
Linaro.org <http://www.linaro.org/> *│ *Open source software for ARM SoCs
Holiday [2/10]
Juno crash analysis [2/10]
* Spent some time fiddling with kexec on AArch64
* Worked in one very specific case
* Another patch series is (apparently) coming, will look out for it
and try again
SPEC-on-Android [2/10]
* Supporting Qian on getting this working
* Wrote a readme for the repository, fixed a Makefile bug that Qian's
cross-compiler happened to tickle
Jenkins benchmarking job - TCWG-348 [1/10]
* Tested, tidied up pbl hacks to generate JSON
* Tested my pbl with Jenkins prototype jobs
* A few minor bug fixes/enhancements for pbl
LAVA jobs for uinstance - TCWG-432 [1/10]
* Reworked jobs to support uinstance, maintaining backward
compatibility as far as possible
* Started adding support to submit results to bundle stream
Misc [2/10]
* Debian FS ready to submit
* Usual meetings/mail/etc background
=Plan=
Look at doing pbl hacks properly in Fathi's in-development refactored p-b-l
Pull together Jenkins/LAVA/pbl, ready to test when uinstance is available
Write up noise control report
(If time, if patches land) have another go at crashdump
== Progress ==
o Linaro GCC (4/10)
* Delivered GCC 4.9 2015.10 snapshot
* More backports forGCC 5 2015.11
* Many instabilities on Hetzner this week
o Upstream work (2/10)
* Sanitizing gfortran testsuite
o Release tools (2/10)
* Added RCs and binaries support to our snapshot.linaro.org
publishing job
o Misc (2/10)
* Various meetings
* Some support
== Plan ==
o Track missing backports dependencies
o Continue ongoing tasks.
== This week ==
* TCWG-369 - Exploit wide add operations when appropriate for Aarch64 (4/10)
- Determined that vectorizer is failing for all targets that have
widening adds with
V8HI to V4SI support (aarch64, ia64, powerPC).
- Modified test cases to indicate expected failure with wide add
V8HI to V4SI support
- Patch sent upstream for approval
* Bugzilla 68223 - arm_[su]min_cmp pattern fails
- Resolved by reverting patch for tcwg-146 as pattern fail in some
corner cases. (3/10)
- Reverted patch checked in upstream
* Misc (1/10)
- Conference calls
* Illness, November 2nd (2/10)
== Next week ==
- TCWG-317 - Resolve lto big endian failures
== Progress ==
- Leave (2/10)
- Widening pass (TCWG-547) - 5/10
* Made the latest changes requested in the review
* Fixed bootstrap and bootstrap mis-compare for ppc64-linux-gnu
* Making uninitialized variable as anonymous ssa (as asked in review)
results in few ICEs.
* Posted updated patch for feedback
- Misc (3/10)
* started looking into LTO status
* Looked at LuaJIT for arm
* gcc/bug list
== Plan ==
* continue with widening pass based on feedback
* Look at implementing LuaJIT for aarch64
* LTO
== This week ==
* TCWG-72 (6/10)
- 5 iterations since the original patch. Changes include:
a) Integration into widening_mul patch
b) Rewriting the divmod transform so DIVMOD() is placed before the topmost
div/mod stmt
c) Removed check for widening mode and optab handler check in expand_DIVMOD
d) Fixed ICE when constant is one of the operands to div/mod stmt.
e) Fixed mis-compilation with a test-case when operands matched but in
opposite order.
f) Formatting nits and fixed test-cases.
- Richard suggested no need to check for post-domination conditions.
- Not sure on what condition to gate the transform.
Checking for availability of divmod/div/mod is not sufficient because arm
defines optab handler for mod which only matches r0 % n where n is
constant and power of 2
for other cases it's expanded via divmod libcall thru expand_divmod.
We would rather need
to check if the template for mod/div gets matched than just to check
if optab handler exists.
AFAIK this cannot be done during tree-ssa passes.
I can think of two approaches:
a) Do the transform to DIVMOD representation unconditionally in
widening_mul pass.
And then in expand_DIVMOD check if the template for mod can be matched.
If it does match then undo the transform from DIVMOD to original
representation and expand.
I am not sure how feasible it is to undo the transform at expansion
time, and start expanding the modified cfg.
b) Define a new target hook combine_divmod.
Default implementation could check for optab handler for div/mod/divmod.
and I could override it for arm-backend to additionally check if the
second operand is a constant and power of 2 and fail for this case
(since we want this to be expanded from modsi3 pattern).
Not sure if this is a good idea, I am replicating the information from
the modsi3 pattern.
If the pattern changes, the hook would also need to be changed.
* Convert ASM_FORMAT_PRIVATE_NAME to hook (2/10)
* TCWG-319 (1/10)
- Bencharmking for patch in progress
* Misc (1/10)
- Meetings
- Sync with Kugan
== Next Week ==
- Continue with TCWG-72
- Complete the patch with build, test and config-builds for
ASM_FROMAT_PRIVATE_NAME and submit upstream
- Continue benchmarking TCWG-319, TCWG-310
== Progress ==
* Buildbots (5/10)
- Some broken bots, bisecting, etc
- Helping a MIPS patch pass on ARM bot
* Maintenance (2/10)
- SciMark2 seems not to be unstable or slow any more in ARM64
- Some more investigations on Loop Load Elimination
- Profiling bigfib on APM and HiKey
* Background (3/10)
- Code review, meetings, discussions, general support, etc.
- Some FOSDEM fiddling
- Some power issues
== Progress ==
* Validation
- moved list of unstable tests to a separate repo, to make
maintenance easier (TCWG-425)
- Jenkins jobs maintenance & cleanup
- a few ABE reporting patches
- comparison of results between old & new lab
* GCC
- trunk monitoring, reported a few new failures.
- Send patch to fix vqtb[lx][34] intrinsics on aarch64_be
* Binutils
- Added a Jenkins job to build+check binutils on
a variety of configurations:
https://ci.linaro.org/view/tcwg-ci/job/tcwg-binutils/
- sent a small patch to fix a bug in the recent STM32L4XX erratum patch
== Next ==
* Validation:
- work on the switch to the new lab, once dev-01 is back online
- more tuning to avoid deadlocks
- re-measure build time on dev-01, to better tune other build jobs
* Two half day off. [2/10]
# Progress #
* TCWG-332, fails in gdb.threads/multiple-step-overs.exp. [1/10]
Testing the simpler approach suggested during the review.
* TCWG-387, done. [1/10] GDB patches are pushed in.
* TCWG-422, GNU vector extension support in ARM GDB. [2/10]
Patches are done, and being tested.
* TCWG-423, GNU vector extension support in AArch64 GDB. [2/10]
Writing patches. Find more issues for AArch64 that GDB doesn't
fully understand the AArch64 calling convention. Need more work here.
* Review ARM GDBserver software single step patch. [1/10]
* Misc, meeting, email, [1/10]
# Plan #
* Off on Wed afternoon.
* TCWG-422, post patches
* TCWG-423, continue.
--
Yao
