JIT: Use loop-aware RPO for initial block layout #108903
Conversation
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
|
@AndyAyersMS PTAL. Diffs (surprisingly?) show net PerfScore regressions. I suspect this is from cases where we previously were breaking up the loop body to create quick exit paths (i.e. when returning early in a loop). In such cases, we're taking a PerfScore penalty by breaking up fallthrough for conditional "exit early or continue" jumps in loops, but we're improving the locality of the loop body, which isn't modeled by PerfScore. Perhaps we should leave cases where we're justified in breaking up loops to 3-opt? |
Yes, I'd say deciding to break up loop bodies is a "global" decision that is better left to 3-opt. |
2 participants
Part of #107749. Now that we have a utility for computing loop-aware RPO traversals, use it for getting the initial block layout. This isn't enough on its own to get rid of
fgMoveHotJumpscompletely (churn is quite big if I try to remove that method), but having a better initial layout should help decrease the search space a bit once we replacefgMoveHotJumpswith 3-opt.cc @dotnet/jit-contrib