Hi all,
I'd like to run the binutils testsuite in a native configuration
regularly to find any regressions. Do we have any infrastructure for
doing this?
It looks like LAVA would be useful but it looks complex and I am not
sure if we use that already for running toolchain tests (I think there
was a Connect session about that?).
Thanks,
--
Will Newton
Toolchain Working Group, Linaro
== Progress ==
* Disable-peeling:
- benchmarking locally on snowball
- failed to build reference branch on cbuild, for an unknown reason
* Turnoff 64bits ops in Neon:
- committed upstream in 4.9 (now stage 1)
- benches of backport into Linaro-4.7 OK: no significant diff,
except +12.6% on spec2k's crafty)
- Filed merge request.
* Vectorizer cost model:
- benches of backport inot Linaro-4.7 OK (no significant diff, as expected).
- Filed merge request.
* Revert-coalesce:
- benches ran in cbuild, but no *-diff.txt file generated.
* Libsanitizer:
- testsuite results OK. All unsupported tests are normal.
* Neon intrinsics codegen:
- got internal feedback
* GDB reverse debug:
- filed 2 bug reports after internal requests.
* Misc:
- filed several bug reports against cbuild to track the problems I observed.
== Next ==
* Disable-peeling:
- analyze bench results
- understand/'fix what's wrong with cbuild
* Turnoff 64bits ops in Neon:
- backport to Linaro-4.8
* Revert-coalesce:
- manually compute bench differences if unable to have cbuild do it.
* Libsanitizer:
- propose patch upstream.
== Progress ==
Knocked out more GDB remote configuration failures that were failing due to
timeouts.
Had 1:1 with Matt for direction on GDB improvement plan.
Looked around for Ireland visa process and filled out visa application.
== Plan ==
Complete review of remote timeout based test case failures.
Start work on the GDB improvement blueprint in JIRA from this week.
Get remaining documentation for Ireland visa and complete application.
== Progress ==
- Looking at TLS relocs and coming up with test cases for garbage
collecting them.
- Patch for missedout testcases in gc section is up streamed. Thanks to Marcus.
== Plan ==
- Complete gc section test cases for TLS relocs.
- Understand and evaluate gprof support work for Aarch64
Misc
-----
Leave on Wednesday and Thursday (27th and 28th)
== Progress ==
- Worked on VRP based zero/sign extension elimination at tree level
- Some regression test cases are failing; investigating them
== Plan ==
- Analyse the reason for test case failure and fix them
- Extend the code to handle all the possible cases (currently
implemented for subset to get the framework working)
== Issues ==
* None
== Progress ==
* Libunwind AArch64 support:
- basic implementation for building lib and testsuite done
- but even a simple test case doesn't work for the moment
== Plan ==
* Libunwind AArch64 support:
- continue AArch64 specific implementation and informtation gathering
== Progress ==
* EuroLLVM paper review
* Disabled EH tests on ARM for now
* Fixed TSVC: NEON vs. VFP VMUL.f32 (and IEEE 754 correctness)
* Changed FeatureNEONForFP to false on EABI ARM targets unless unsafe-math
* Trying Arndale again, with u-boot patched, not good
* Broken Panda buildbots, trying to fix
== Issues ==
Got ill on Tuesday and only worked partially, if at all, the rest of the
week
Hopefully will be better next week...
== Plan ==
* Fix our panda buildbot, maybe add more
* Final try on Arndale GCC bootstrap before giving up
* Continue fixing test-suite failures (only 5 left)
Progress:
* VIRT-4 [Guest migration support for KVM]
** VIRT-51: vexpress migration:
** wrote and submitted patches for all the devices that need fixes
** tracked down a painful bug where the guest just hung on vmload
to a missing state-save for a timer device, submitted patch
** remaining work for this item is just responding to patch review
-- PMM
Hi folks,
I found an issue while fixing a test using the wrong VMUL.f32, and I'd like
to know what should be our choice on this topic that is slightly
controversial.
Basically, LLVM chooses to lower single-precision FMUL to NEON's VMUL.f32
instead of VFP's version because, on some cores (A8, A5 and Apple's Swift),
the VFP variant is really slow.
This is all cool and dandy, but NEON is not IEEE 754 compliant, so the
result is slightly different. So slightly that only one test, that was
really pushing the boundaries (ie. going below FLT_MIN) did catch it.
There are two ways we can go here:
1. Strict IEEE compatibility and *only* lower NEON's VMUL if unsafe-math is
on. This will make generic single-prec. code slower but you can always turn
unsafe-math on if you want more speed.
2. Continue using NEON for f32 by default and put a note somewhere that
people should turn this option (FeatureNEONForFP) off on A5/A8 if they
*really* care about maximum IEEE compliance.
Apple already said that for Darwin, 2 is still the option of choice. Do we
agree and ignore this issue? Or for GNU/EABI we want strict conformance by
default?
GCC uses fmuls...
cheers,
--renato
