Thursday 10th 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        



  • Stand up call



  • Running benchmarks on the ursa boards
  • Done some EEMBC and SPEC runs and looked at the processing scripts
  • Made a wiki page to make the next persons job easier
  • Comparing ref and train datasets in SPEC
    • see if the changes appear in both
    • Doing a -O2 vs a -O3 run to show the differences
  • Talked with zyga re: private reports


  • Finished implementation of register pressure estimate
  • Ran on EEMBC, DENBench, and libav
  • Still see some regressions due to other SMS areas
  • where SMS fails but doloop is still applied
  • Checking a posted patch that helps this


  • Working on SLP of conditions, and with any number of operands
  • Needs more as expression type != condition type for particular case
  • Needs to extend pattern recognition to recognise non-constants
  • Gives a 2.6 x speed up in pixel weighting


  • Internal CSL work
  • Ran the release
  • Benchmarking for generic tuning, taking a while
  • Looked at how to do benchmarks via LAVA
  • ACTION: Michael to make it accessable


  • Benchmarked the A15 optimised memcpy against the Linaro A9 versions
  • On the A9, the A9 version is significantly faster
  • Still fighting some race conditions in membase
  • Some other non-Linaro activities


  • Remote core file generation
  • Has an initial set of patches, posted, getting review
  • Patches interact with other features so not yet accepted
  • Skipped this months release due to no changes


  • Recovering from travel...
  • Running the qemu-linaro release
  • Beagle model has bit rotted a bit and display has stopped working


  • Short week due to travel
  • Releases
  • Catch up after Connect, firing off emails

From Ken:

  • found that libunwind segfaults in case unwinding via DWARF debug is enabled
    • usually the debug info gets stripped from the binary on android and backtracing gets done via the ARM specific unwind tables (so this isn't a major issue for Android)
    • however, the segfault is not Android specific and can also occur on Linux when building libunwind for remote unwinding only
  • workaround in place, but when discussing this on the ml it turned out that:
    • libunwind-ptrace should not be compiled for REMOTE_ONLY case at all (it was intended for host!=target)
    • this means that our current Android approach is rendered to be a dead end
    • we need another way to get libunwind to build on Android
  • I started to work on a patch that adds a new configure switch for building libunwind-ptrace functionality only (superset of REMOTE_ONLY)

WorkingGroups/ToolChain/Meetings/Archive/2011-11-10 (last modified 2013-08-30 11:47:46)