JIT: Use edge weights in LinearScan::findPredBlockForLiveIn

[amanasifkhalid]

Unrelated to the recent LSRA block order changes, but I noticed that for blocks with multiple predecessors, LinearScan::findPredBlockForLiveIn picks the predecessor with the largest block weight to get the current block's live-in variable register locations from. It would be more precise to factor in the predecessor edges' likelihoods, too -- for example, if one predecessor block is marginally heavier than the other, but is far less likely to flow into their shared successor, the latter seems like a better choice.

Diffs were a mixed bag locally, though now that we always have edge likelihoods available, this seems like something we should be doing on principle.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[amanasifkhalid]

cc @dotnet/jit-contrib, @kunalspathak PTAL. Diffs differ quite a bit by platform, though this looks like a PerfScore win most of the time. The last superpmi-collect run seems to have dropped quite a few collections, so we're missing some platforms -- if you'd like, I can post diffs for this change against an older GUID on win-x64, or kick off a new collect run?

[kunalspathak]

This sounds reasonable to me. Do we know why we do not see asmdiffs for win/x64, etc.?

[amanasifkhalid]

Do we know why we do not see asmdiffs for win/x64, etc.?

I'm guessing I kicked this run off too early, before all of the collections had propagated. Newer runs seem back to normal, so I'll kick another one off now.

[AndyAyersMS]

FWIW I merged a PR that updated the GUID yesterday (#108153) so maybe some of the new collection runs failed? I should have a new asp.net collection soonish.

[amanasifkhalid]

FWIW I merged a PR that updated the GUID yesterday (#108153) so maybe some of the new collection runs failed?

I was thinking that might be the case, but the SPMI diffs run from one of my other PRs with the updated GUID had collections for every platform, and the latest run on this PR is still missing collections...

[amanasifkhalid]

@kunalspathak @AndyAyersMS here are the diffs on win-x64 and win-arm64: gist. The arm64 diffs are smaller overall, as I had to skip some collections due to them stalling during asmdiffs on my machine (they didn't stall when replaying them, so this doesn't seem to be a JIT issue...).

[AndyAyersMS]

stalling during asmdiffs

Seems odd though...

[AndyAyersMS]

Looks like you slotted this this "after" the critical edge avoidance heuristic.

I wonder if it's better to use the edge weight heuristics there as well... if the "critical" edge is low weight then seems like we'd prefer to split it and put resolution moves there (that is, it seems like we could build a cost model of sorts and decide which we prefer, the unique pred or the "other pred").