Stream: t-compiler/wg-self-profile

Topic: improving the perf.rlo details view


mw (Oct 01 2019 at 08:31, on Zulip):

While working on the event review in rustc I noticed that quite a lot of our events are no queries and that they are nested. That makes it hard to read some of the LLVM stuff, for example.

mw (Oct 01 2019 at 08:32, on Zulip):

Especially for things that are not queries it is useful to have both the "self-time" and the time including all the sub-tasks.

mw (Oct 01 2019 at 08:33, on Zulip):

e.g. I want to know the entire amount of time spent in ThinLTO but I always want the time spent in each of the sub-tasks of ThinLTO

mw (Oct 01 2019 at 08:33, on Zulip):

the question is: how to visualize this properly?

mw (Oct 01 2019 at 08:34, on Zulip):

one option are flame-graphs, they contain pretty much exactly that information (it's unclear how to handle queries for those though)

mw (Oct 01 2019 at 08:35, on Zulip):

a simple alternative is to just add another column for the "inclusive time", as kCacheGrind has it.

mw (Oct 01 2019 at 08:36, on Zulip):

pasted image

mw (Oct 01 2019 at 08:37, on Zulip):

Since you can sort for either incl. or self time, that's very useful already.

mw (Oct 01 2019 at 08:37, on Zulip):

Do we know how big a flamegraph is for our benchmarks?

Last update: Nov 17 2019 at 07:35UTC