Monday 21st November 2011

This month's meetings

<< <  2011 / 11 >  >>
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        



  • Review action items from last meeting
  • 892880 MIPS: link error with optspace

  • 889984 libgcc FP routines missing DWARF CFI

  • 831094 peeling interaction with set loop

  • Next steps in 64 bit sync primitives
  • Backports in general
    • When to propose for 4.6, what effort
  • Backport requests:
    • 890657 mm/page-writeback.c:820: internal compiler error: Segmentation fault

    • 714921 memory usage due to -fvar-tracking (from Fedora)

    • 675347 Qt and volatile bitfields

  • -O3 performance
    • Partial PRE?
    • MOVW/MOVT?
    • 64 bit operations?
  • Getting multiarch patches into gcc-linaro
  • Generic tuning benchmarks and possible triggers
  • OpenOCD card

Action Items from Previous Meeting

  • DONE: Michael to get perf everywhere
  • DONE: Michael send links to the binaries of the the benchmarks


libgcc and optspace:

  • Richard hasn't seen it before
  • Is happy to have a look at it
  • Keep as a Launchpad bug for it

libgcc and CFI:

  • Strip debug information?
  • or add CFI?
  • Not having CFI is a bug that should be fixed
  • Have the macros - use those or the libgcc bits and bytes macro versions?
  • Which should we use? Should it be possible to build with another toolchain or assembler?
  • How old are the CFI directives? > a couple of years?

  • Already tied to gas through the macros?
  • ACTION: See how many versions of gas have support
  • If its two or three or more then it's OK
  • Does need to support coff but not with any random assembler
  • Ubuntu package build:
    • Strips all debug information and puts them into other packages
    • Which is fine for shared, but if you link in from a static then that is stripped
    • Does it strip the frame backtrace? (it shouldn't)
    • On all other platforms eh_frame is kept, but ARM has something else that is similar
    • Some ARM Inc tools use the debug data to generate exception handling
    • Change binutils to not strip debug frame if target==ARM?
    • Interpretation of the ABI is you should always generate debug frame
    • Approximate with -fexceptions? No, as only generates ARM specific sections
    • Why?
      • Can debug into statically linked libraries
      • Can use to generate unwind data via the linker if the ARM exception tables aren't there
      • Also isn't loaded
      • Can unwind frames
      • Is more precise unwinding - in theory can have at every instruction address
      • Is more generic than ARM exception tables but more a theoretical difference
      • ARM exceptions can only throw from call sites, not other instructions, or async like signal handlers
  • Is it in our queue? How would it help an end developer or performance?

Peeling and a set loop:

  • A set of eight words
  • So few that peeling is a bad idea in general
  • Checked with a cost model enabled and skipped due to too few iteration
  • Ira will have a look into it
  • How low can the alignment be?
    • The A8 likes things to be quad word aligned (especially due to loads via the L2 bypass)
    • Depends on the micro architecture
    • Michael needs to re-run his micro benchmark
    • Can specify the alignment
      • Will fault if you do
      • Is an advantage on the A8
      • Might not be on the A9. Might remove one check though

Next steps in 64 bit sync primitives:

  • Any other libraries/programs that use them?


  • What is our policy?
  • Strictest is that backports should be for regressions
    • Andrew couldn't find the word 'regression'
  • Backend maintainers do have some discression, low risk will be considered
  • Michael's thoughts:
    • Encourage author to backport
    • Don't backport on speculation, only on direct request
    • Backport ours if meets criteria
  • How far back can you go to count it as a regression?
    • Historically 2.95!
  • Any bug fixes in ours that should be in FSF 4.6?
  • 714921 Michael to ping Julian, else propose a patch

WorkingGroups/ToolChain/Meetings/Archive/2011-11-21 (last modified 2013-08-30 11:49:07)