JIT: Stop using ref counts in forward sub #85909
Conversation
Instead of going to great lengths to keep the ref counts correct in physical promotion just stop using them in forward sub. The last use information generalizes what forward sub needs except for when the tracking limit is reached, which shows in the improvements we actually get from consulting the ref counts -- the improvements are very modest and in tests only. This makes early liveness the only consumer of the "regular" ref counts computed by local morph. Weighted ref counts are still used for undoing regular promotion of implicit byref parameters.
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsInstead of going to great lengths to keep the ref counts correct in physical promotion just stop using them in forward sub. The last use information generalizes what forward sub needs except for when the tracking limit is reached, which shows in the improvements we actually get from consulting the ref counts -- the improvements are very modest and in tests only. This makes early liveness the only consumer of the "regular" ref counts computed by local morph. Weighted ref counts are still used for undoing regular promotion of implicit byref parameters.
|
|
cc @dotnet/jit-contrib PTAL @AndyAyersMS Let me know if you have any objections to this... |
AndyAyersMS
left a comment
AndyAyersMS
left a comment
There was a problem hiding this comment.
I'm fine with removing reliance on early ref counts.
src/coreclr/jit/forwardsub.cpp
Outdated
| // | ||
| bool livenessBased = false; | ||
| if (varDsc->lvRefCnt(RCS_EARLY) != 2) | ||
| if (!fgDidEarlyLiveness) |
There was a problem hiding this comment.
Bail up in fgForwardSub, and assert `fgDidEarlyLiveness' is true here?
2 participants
Instead of going to great lengths to keep the ref counts correct in physical promotion just stop using them in forward sub. The last use information generalizes what forward sub needs except for when the tracking limit is reached, which shows in the improvements we actually get from consulting the ref counts -- the improvements are very modest and in tests only.
This makes early liveness the only consumer of the "regular" ref counts computed by local morph. Weighted ref counts are still used for undoing regular promotion of implicit byref parameters.