[PATCH] D25588: Statistic/Timer: Include timers in PrintStatisticsJSON().

[Matthias Braun via llvm-commits]

MatzeB added a comment.

In https://reviews.llvm.org/D25588#571012, @bruno wrote:

> This is very nice. A couple of questions/comments:
>
> - Looks like if you pass `-stats -stats-json -time-passes` then all timers and statistics are printed together, what happens if you just pass `-stats-json -time-passes`, would that only output the timers in json format? Also, when printed together, are they part of the same "group"? it's not clear from the tests how the output layout works.


The way this currently works is that the statistics printing code grabs the timers when printing (as a side effect this clears the timers so they won't get printed again by the timer printing code). This means that with "-stats-json -time-passes" without "-stats" you will get no stats printed and the timers will end up being printed in the traditional format. I don't think we need such a mode either as I expect the data to be processed by scripts which could just as well do the filtering afterwards.

The current output format is simply a list of key/values. However I introduced some hierarchy in the key names, so you will see key names like "mem2reg.NumSingleStore", "time.regalloc.seed.user" or "time.pass.Scalar Evolution Analysis.wall" so you can filter somewhat by prefixes/suffixes of the keys.

> - Do you plan to somehow track different runs of the same pass? Right now that could be inferred  by looking into the output order, but it would be nice (not necessarily in this patch) to easily access that information; I remember it to be really useful when comparing compile time between two different compilers, I could then notice when one compiler run the same pass more times than the other and it would also help identify when specific runs were slower due to the influence of other opts.

This patch does not change the way timers are run, so all timers with the same name and therefore all runs of a pass with the same name are merged together. It would indeed be nice to introduce a model here where rather dump some form statistic events rather than the summed stats; But that is a more fundamental change that needs to be done in a different commit. I'd be happy to help discuss, design and review if someone wants to attack this.


Repository:
  rL LLVM

https://reviews.llvm.org/D25588