Simplify precompiles #669
Simplify precompiles #669
Conversation
This switches to "do a little work" for precompilation, rather than explicitly spelling out the specific signatures. This should be more maintainable and potentially more exhaustive (or could be made to be so) since it precompiles across runtime dispatch boundaries. This also uses `Base.inferencebarrier` in several places to prevent an argument from being specialized.
timholy
force-pushed
the
teh/pc_external
branch
from
530e4c2 to
d49f829
Compare
|
I've investigated with help from @aviatesk, and determined that the stubborn MethodInstance was having its compiled code discarded because of a |
| @@ -122,7 +122,7 @@ end | |||
| function methods_by_execution(mod::Module, ex::Expr; kwargs...) | |||
| methodinfo = MethodInfo() | |||
| docexprs = DocExprs() | |||
| value, frame = methods_by_execution!(JuliaInterpreter.Compiled(), methodinfo, docexprs, mod, ex; kwargs...) | |||
| value, frame = methods_by_execution!(Base.inferencebarrier(JuliaInterpreter.Compiled()), methodinfo, docexprs, mod, ex; kwargs...) | |||
There was a problem hiding this comment.
This seems to be an actual usecase example of JuliaLang/julia#41931. As far as I understand, does it turn out to be effective to turn off inference of recurse(...) within a callstack of methods_by_execution!? If recurse(...) is less-frequently called actually, then that sounds a reasonable hack to improve latency.
There was a problem hiding this comment.
Yes, this is exactly a @noinfer. All the execution time specific to a particular recurse is in JuliaInterpreter anyway.
|
Wow, the GC corruption on 1.7 is a bit scary---we might need to debug that and patch 1.7 before merging this. But the test failure on nightly is JuliaLang/julia#42822, which appears to be deliberate. (CC @IanButterworth.) |
This switches to "do a little work" for precompilation,
rather than explicitly spelling out the specific signatures.
This should be more maintainable and potentially more
exhaustive (or could be made to be so) since it
precompiles across runtime dispatch boundaries.
This also uses
Base.inferencebarrierin several places toprevent an argument from being specialized.
I was using this as a test case for JuliaLang/julia#43990 to see if I can eliminate all inference from Revise. It turns out there's one CodeInstance whose inferred code is either not cached or is deleted after compilation. Needs investigation. But the total inference time is <50ms, so in practice it's not a big deal.