Monday 21st Febuary 2011

This month's meetings

<<MonthCalendar: execution failed [Argument "anniversary" must be a boolean value, not "workinggroups/toolchain/meetingtemplate"] (see also the log)>>

Attendees

Agenda

  • Review action items from last meeting
  • FSF assignment is in place for GCC, GDB, Binutils, and GLIBC
  • libunwind approval - going well, deadline is the 24th
  • A short summary on latrace
  • 604753 Process shared mutex bug next steps

  • Strings in Ubuntu
  • Versatile Express QEMU
  • Performance work
    • Benchmarks
    • Headroom
    • ARMv7 vs ARMv5
    • NEON vs non-NEON
    • CoreMark

Minutes

FSF assignment:

  • Now have it in place

libunwind:

  • Works under the Linaro contribution rules
  • Seven day window ends on the 24th
  • No reason to wait until then, just keep things in a local branch

latrace:

  • Goal is investigate developer tools, looked into what people use
  • Looked into alternatives to ltrace as it has issues on many platforms
  • Uses the auditing API built into glibc
  • Has a bug or two, but generic
  • Can trace other libraries, as part of the loader
  • Michael: what's the age, health, community like?
  • Ken to look into it, does have a mailing list, has activity
    • Has git repository
    • Has mailing list since September 2010

604753:

  • Shared mutex seems to be fixed with the new eglibc
  • One test case (test timeouts) in APR may still exist
  • Matthias: is there any request to backport to Maverick or Lucid?
    • Michael: hasn't seen any
    • What's the policy?
    • These multi-core issues are preexisting. No regression, so no strict need to backport

Strings in Ubuntu:

  • Matthias: yes, can supply a patch that overwrites the existing files
  • Dave: what license? Matthias: keep the BSD and he'll update the license
  • Later step is to relicense and push everywhere
  • Michael: can we develop workloads using something like latrace?
  • Dave: will do, will take a bit longer
  • Can we show that it is better for typical workloads?
    • How about micro benchmarks?
      • memchr can be worse when character is within the first few bytes
      • strchr has no real worse-case
      • memset - all versions on large sets end up at memory bandwidth limits
  • Is staying clear of memcpy()/memmove() as they're already at memory bandwidth
  • A15 will be interesting in the future

libffi:

  • License has been approved, free to work on

Versatile Express QEMU:

  • Looked at patches from B Labs
  • Tweaks needed to get Linaro kernel booting
  • Removing stuff that was specific to their product
  • How does the original 2 week estimate look?
    • Is booting the kernel image, so no 'bear traps'
    • Will continue on clean up, estimate seems lower
    • Does have display, should have keyboard and mouse
    • Should have networking, mass storage (MMC)
    • ...as these are the same as versatile

Benchmarks:

  • Michael: went over report on DENBench vs compilers vs flags
  • In the front pack on most all
  • Significant regressions in some areas
  • What's the next step?
    • Pick regressed benchmarks
    • Analyse a bit further, especially hotspots and hopefully to RTL level
    • Then file tickets
  • NEON vs non-NEON is generally neutral
    • Many cases where NEON regresses
    • Ira: is working on a cost model
  • ARMv7 vs ARMv5
    • ARMv7 is generally an improvement
    • Some areas show a large improvement
    • A couple show ~10 % regressions
    • Chung-Lin: has a lead on CoreMark

      • The zero extend instruction introduced in ARMv6
      • Replaces an AND, xx, #255
      • Not as thoroughly supported as the AND, breaks if-conversion
      • Investigating further

Upcoming Linaro GCC release:

  • Matthias: mid-release updates
  • Michael: plan is to update two weeks into each monthly cycle. Had missed in the last two releases at least
  • Andrew: will pick up the SVN revision that comes out of the build system

WorkingGroups/ToolChain/Meetings/Archive/2011-02-21 (last modified 2013-08-30 11:48:03)