Monday 16th January 2012

This month's meetings

WorkingGroups/ToolChain/Meetings
<< <  2012 / 1 >  >>
Mon Tue Wed Thu Fri Sat Sun
            1
2 3 4 5 6 7 8
9 10 11 12 13 14 15
16 17 18 19 20 21 22
23 24 25 26 27 28 29
30 31          

Attendees

  • Åsa Sandahl
  • Ken Werner
  • Michael Hope
  • Peter Maydell
  • Ramana Radhakrishnan
  • Richard Earnshaw
  • Thiago Bauermann
  • Ulrich Weigand
  • Zhenqiang Chen

Agenda

  • Review action items from last meeting
  • SunSpider as a benchmark

    • VM?
    • Host?
  • Sections to strip in the binary toolchain
  • Blueprints for this month
    • Strided load/store fwprop?
    • Partial-partial PRE?
    • vld1 for quad loads?
    • Initial A15 system model?
  • -O3 writeup progress
    • Benchmarks?
    • Correctness?
  • ucontext routines

Action Items from this Meeting

  • TBA

Action Items from Previous Meeting

  • DONE: Michael to have a peek at EGLIBC test failures (tests basically OK on Lucid)
  • DONE: Michael to ask gcc-help about minimum versions (2.95 or above if no bootstrap)
  • DONE: Michael to change Richard's name to Andrew's for reviews
  • ACTION: Peter and Rusty to talk re: QEMU release process and fallbacks
  • ACTION: Michael to add benchmarking to the jobs list

Minutes

Actions:

  • See above

SunSpider:

  • STE experiments were on Android with the built-in WebKet/V8 based browser
  • 5 % regression overall, 23 tests, few were significant regressions
  • Ubuntu or Android?
  • V8 engine standalone? or as part of a browser? Chromium?
  • Cross or native?
  • Never did get Firefox cross compiled...
  • Does SunSpider need a browser? Or is just the VM OK?

    • ACTION: Asa to ask internally see if anyone knows
  • First step: see if it can run on V8 and how big the build-deps are
  • Can you automate the Android browser? Used STE internal tool
  • ACTION: Asa to look at native on x86 and if it can do the job
  • ACTION: Asa to create a blueprint for first step in SunSpider

Sections to strip in the binary toolchain:

  • Want to strip but not over strip
  • .debug_frame
    • May not be able to do a backtrace without
    • But why be different to native?
    • Native can add these afterwards but we can't - perhaps we can't?
  • Why strip?
    • Does strip the core so GDB won't step into
    • ACTION: Michael to reproduce the gdbserver libgcc step into
  • Covers all known sections
  • Easy to verify by dumping sections and looking for debug

Thiago joined:

  • Joining to work on the GDB side of things
  • Working on this side since 2007
  • Still finishing off a previous project, start full time next week

Blueprints for this month:

  • Strided load/store fwprop - not sure can be resolved in time
    • Patch is straightforward, much more aggressive in merging subreg
    • Means many more subreg accesses
    • ...which exposes passes that don't handle subregs very well
    • ...including passes from 1998!
  • Partial-partial PRE
    • Doesn't understand the patch well enough to discuss upstream
  • vld1 for quad loads?
    • Will hand off to Ulrich
    • General state? Got it to generating vld1 instead of vldm
    • Didn't like the change in supported addressing modes as it limits single word loads
    • Not sure if disable, pessimise, fix in reload?
  • Initial A15 system model?
    • Tidied things up, sent patchset upstream on Friday
    • Done to match the acceptance criteria
    • Peter will set milestone when committed
  • Any other work?
    • One of the PGO patches
      • Enable hot/cold partitioning
      • Held up because of testing
      • Haven't run SPEC, will run and finish off
      • ACTION: Michael to make a blueprint on enabling hot/cold partitioning in PGO

O3 writeup progress:

  • Benchmarks
    • Started a run over the weekend
    • Need to see how those went
    • Against the 2012.01 release
    • Has set
    • First information is geomean and top five regressions
  • Correctness
    • Building at -O2 on ARMv7 and Thumb
    • Building the Qt image
    • Next step is -O3
    • Some packages are setting their own options
    • Hopefully tacked onto the end so self options may take over

New A15 instruction coverage:

  • Typed in to risu
  • Has a bias towards generating interesting floating point numbers
  • Fills FP registers, runs ~100 instructions, then refills
  • GCC doesn't use FMAC at all
  • ACTION: Michael to send Peter a couple of paragraphs

WorkingGroups/ToolChain/Meetings/Archive/2012-01-16 (last modified 2013-08-30 11:49:59)