Open psfinaki opened 1 week ago
Hi @EgorBo, so this is an alternative approach to enable PGO, without doing profile generation ourselves. However - this doesn't seem to work so far, as in, benchmarks don't show any improvements.
Any ideas what could be wrong?
@EgorBo any pointers on how to figure out why is profile we get from nuget is not applying, and if we generate it locally, it shows a difference?
@psfinaki @vzarytovskii I presume you test it on fsi
? have you added fsi
as a training scenario to dotnet-optimization repo? Otherwise it produces some MIBC trained to TechEmpower, console apps and for F# it's only "build Giraffe" scenario
@psfinaki @vzarytovskii I presume you test it on
fsi
? have you addedfsi
as a training scenario to dotnet-optimization repo? Otherwise it produces some MIBC trained to TechEmpower, console apps and for F# it's only "build Giraffe" scenario
We have tested on fsc
, since that's the only F# scenario in the optimization repo as of now.
@psfinaki did you try to build Giraffe with and without profile applied in this PR? Does it change anything?
So far I have only tried running published fsc
against an empty test.fsx
script. That was enough to show the differece in that other PR with us creating the profile.
Let me try Giraffe.
@vzarytovskii @EgorBo I'm afraid I don't see any difference in Giraffe compilation either :( it's the same execution time basically.
@vzarytovskii @EgorBo I'm afraid I don't see any difference in Giraffe compilation either :( it's the same execution time basically.
Do you a MIBC that works locally? can you send it to me? I'll check weigths in it
@EgorBo did you, by any chance, look at it? Curious what's the difference between those.
@EgorBo did you, by any chance, looked at it? Curious what's the difference between those.
The only @psfinaki collected looks extremely weird to me. I propose we merge this PR as is since it generally looks corret to me, the problem must be inside the dotnet-optimization repo, once we figure it out and fix it should propagate a correct MIBC and optimize things out, hopefully 🙂 I'll take a look on Friday
Description
This should eventually make the FSC and FSI start about 30% faster. The implementation is inspired by the runtime targets.
Fixes #12636 Fixes #13328