simple inference graph profiler #34613
Conversation
JeffBezanson
added
compiler:inference
base/compiler/typeinfer.jl
Outdated
| @@ -453,8 +453,15 @@ function resolve_call_cycle!(linfo::MethodInstance, parent::InferenceState) | |||
| return false | |||
| end | |||
|
|
|||
| const inference_graph = IdDict{Any,IdSet{Any}}() | |||
| const record_edges = Ref(false) | |||
There was a problem hiding this comment.
| const record_edges = Ref(false) | |
| const record_edges = Bool[false] |
maybe (for bootstrap)
There was a problem hiding this comment.
Is Ref not defined early enough? Maybe a comment to that effect since that was the first thing I wondered when seeing the current state of the PR.
There was a problem hiding this comment.
Ref is available, but no constructors. Use RefValue
JeffBezanson
force-pushed
the
jb/inferenceedges
branch
from
02ff9d9
to
8366e8b
Compare
|
I'm not sure if this is worth merging. On one hand, it was useful information for you. On the other hand, it's inactive code and might be unstable as an API over time (it's a very specific subset of information, and we will probably want more general result reports some day). |
|
I did not really intend to merge this; it's just here in case anybody wants to use this method again. |
|
I think we can effectively declare this present now |
This is the method I used in #34596.
Here is a gist with the analysis code and an example run:
https://gist.github.com/JeffBezanson/e381ea9cbe790d6d8f9bcf473a664f8c
Basically this is way to find which calls are leading to the most inference work, roughly speaking. It's not time-based and might not be helpful in every case, but in this case it gave a useful profile.