4 day week, on holiday monday

Submitted a Linaro Connect presentation.

[TCWG-1424] Investigate profile feedback for code-size
Frustrating week trying to plug profile feedback into builds and
benchmarks and get sensible results
- First attempt was to use clang 2 stage pgo build
-- New pass manager has a bug that prevents clang from linking when I
enable clang instrumented profiling (local ref into discarded comdat
group containing counters)
-- Sample based profiling requires building create_llvm_prof from
Google's autofdo repository. Does not build out of the box after it
switched to using git submodules for its dependencies, managed to fix
up makefile to get it built.
-- Tried a sample profiling run but with perf's sampling rate and
clang's running time mean that the number of samples is too small to
give meaningful results.
-- The version of perf on my machine doesn't support incremental
profiling so I can't accumulate data from multiple runs.
-- Inlining before my pass to mark functions for size optimisation
accounts for majority of code size increase.

- Second attempt was integrating clang/llvm profiling into spec
-- Spent way longer than I'd expected working out how to write a
config file that handles sample and instrumentation based profile
runs. Both forms require the profile generated from perf or an
instrumented build to be post-processed before being used as an input
for the next stage. The sample profiling also needs the path to the
training executable which isn't obvious how to do.
-- Got SpecInt running with all the various combinations that I want
to run, haven't had time to do full runs to get results yet
_______________________________________________
linaro-toolchain mailing list
linaro-toolchain@lists.linaro.org
https://lists.linaro.org/mailman/listinfo/linaro-toolchain

Reply via email to