JIT: Switch tailcall implicit byref optimization to use last use information #81033
Conversation
…rmation The last use information is more general in most cases (for example, the old logic cannot allow the optimization when there are previous field writes to the implicit byref). In addition the old logic can interact with the new last-use copy elision optimization in an incorrect way. Fix dotnet#81019
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch, @kunalspathak Issue DetailsThe last use information is more general in most cases (for example, the Fix #81019
|
|
SuperPMI jobs seemingly did not trigger..? |
The zeroing that the tailcall-to-loop optimization does was zeroing the promoted copies implicit byrefs even when promotion of them was undone. This was introducing unexpected references to the promoted fields. Fix dotnet#81081
|
|
|
cc @dotnet/jit-contrib PTAL @AndyAyersMS Diffs. The new misses are tailcall candidates (that hit a new The regressions are in similar situations as the case inside #81019 -- i.e. the old logic prefers omitting the copy at some non-last use because it believes a subsequent use won't alias. That allows tailcalling, so it is usually better, but the case is quite rare as can be seen in the diffs. |
|
Ping @AndyAyersMS |
Yes, I felt both good and bad removing that code... :-) |
ghost
locked as resolved and limited conversation to collaborators
The last use information is more general in most cases (for example, the
old logic cannot allow the optimization when there are previous field
writes to the implicit byref). In addition the old logic can interact
with the new last-use copy elision optimization in an incorrect way.
Fix #81019