JIT: Stop using ref counts in forward sub (#85909)
authorJakob Botsch Nielsen <Jakob.botsch.nielsen@gmail.com>
Tue, 9 May 2023 07:24:34 +0000 (09:24 +0200)
committerGitHub <noreply@github.com>
Tue, 9 May 2023 07:24:34 +0000 (09:24 +0200)
commit547c506abe05e510bd43330fc8f6d4c5961e9223
tree1bfca8c3e49c66dc2371727c822fca9489cfc065
parent7d54c0524da14337c916c754061a2a300fdfabb6
JIT: Stop using ref counts in forward sub (#85909)

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.
src/coreclr/jit/forwardsub.cpp
src/coreclr/jit/promotiondecomposition.cpp