fix #42078, improve the idempotency of callsite inlining #42082
Conversation
aviatesk
added
the
compiler:optimizer
aviatesk
force-pushed
the
avi/fix42078
branch
from
f63ae02
to
caf272b
Compare
| frame.src = finish!(interp, result) | ||
| end | ||
| run_optimizer && (frame.cached = true) | ||
| typeinf(interp, frame) |
There was a problem hiding this comment.
The previous form here was intended to be an example of running an external optimization pipeline. I think cached=false also has some follow-on effects of deciding how to resolve (and optimize) cycles?
There was a problem hiding this comment.
Yeah, that's why I made this change. Currently the optimization behavior really changes depending on the cache configuration, and it causes typeinf_code to behave slightly different from actual execution, which is very confusing.
For example,
Line 176 in 5629856
@test_broken for long, but I found it has been functional actually(the problem was that the previous
cached = false hinders type_annotate! to do additional clean up work).
And this change is necessary for the test case added in this PR, because callsite inlining depends on typeinf_edge, which actually changes its behavior depending on caller's cache configuration.
External consumer can still plugin their own optimization pipeline by overloading OptimizationState, so I think this change is okay.
After #41328, inference can observe statement flags and try to re-infer a discarded source if it's going to be inlined. The re-inferred source will only be cached into the inference-local cache, and won't be cached globally.
aviatesk
force-pushed
the
avi/fix42078
branch
from
caf272b
to
5629856
Compare
|
Test failures on linux 32 seem unrelated to this PR. |
vtjnash
added
the
status:merge me
DilumAluthge
removed
the
status:merge me
…liaLang#42082) After JuliaLang#41328, inference can observe statement flags and try to re-infer a discarded source if it's going to be inlined. The re-inferred source will only be cached into the inference-local cache, and won't be cached globally.
…liaLang#42082) After JuliaLang#41328, inference can observe statement flags and try to re-infer a discarded source if it's going to be inlined. The re-inferred source will only be cached into the inference-local cache, and won't be cached globally.
After #41328, inference can observe statement flags and try to re-infer
a discarded source if it's going to be inlined.
The re-inferred source will only be cached into the inference-local
cache, and won't be cached globally.