tree: 695ddc285388c60c69078d88970a4658e8bc6818 [path history] [tgz]
afdo_tools/bisection/ is the main script and entrypoint for this AFDO profile analysis tool. This tool attempts to determine which part of a “bad” profile is bad. It does this using several analysis techniques which iterate over provided good and bad profiles to isolate the problematic portion of the bad profile. Goodness and badness are determined by the user, by passing a user-provided bash script. If the program runs successfully to completion, results will be output to the path specified by analysis_output_file as a JSON with the following keys:

  • seed: Float, the seed to randomness for this analysis
  • bisect_results: a sub-JSON with the following keys:
    • ranges: 2d list, where each element is a list of functions that are problematic in conjunction with one another.
    • individuals: individual functions with a bad profile
  • good_only_functions: Boolean: is the bad profile just missing some function profiles (that only the good profile has?)
  • bad_only_functions: Boolean: does the bad profile have extra function profiles (i.e. the good profile doesn't have these functions) causing bad-ness?

Resuming offers the ability to resume profile analysis in case it was interrupted and the user does not want to restart analysis from the beginning. On every iteration of the analysis, it saves state to disk (as specified by the state_file flag). By default the tool will resume from this state file, and this behavior can be disabled by providing the no_resume flag when running the script.


Example Invocation

python --good_prof good.txt --bad_prof bad.txt --external_decider --analysis_output_file afdo_results.json

Required flags:

  • good_prof: A “good” text-based AFDO profile as outputted by bin/llvm-profdata (within an LLVM build).
  • bad_prof: A “bad” text-based AFDO profile as outputted by bin/llvm-profdata (within an LLVM build).
  • external_decider: A user-provided bash script that, given a text-based AFDO profile as above, has one of the following exit codes:
    • 0: The given profile is GOOD.
    • 1: The given profile is BAD.
    • 125: The goodness of the given profile cannot be accurately determined by the benchmarking script.
    • 127: Something went wrong while running the benchmarking script, no information about the profile (and this result will cause analysis to abort).
  • analysis_output_file: The path of a file to which to write the output. analysis results.

Optional flags:

Note that these are all related to the state-saving feature which is described above in “Resuming”, so feel free to return to this later.

  • state_file: An explicit path for saving/restoring intermediate state. Defaults to $(pwd)/afdo_analysis_state.json.
  • no_resume: If enabled, the analysis will not attempt to resume from previous state; instead, it will start from the beginning. Defaults to False, i.e. by default will always try to resume from previous state if possible.
  • remove_state_on_completion: If enabled, the state file will be removed upon the completion of profile analysis. If disabled, the state file will be renamed to <state_file_name>.completed.<date> to prevent reusing this as intermediate state. Defaults to False.
  • seed: A float specifying the seed for randomness. Defaults to seconds since epoch. Note that this can only be passed when --no_resume is True, since otherwise there is ambiguity in which seed to use.