JIT: Spill newarr into temp

[hez2010]

Extracted from #104906

Spilling newarr into temp and introduce a new block flag for it. This is the prerequisites of escape analysis for arrays.

Question: should we always spill newarr instead of only do this when optimizations are enabled?

/cc: @AndyAyersMS

[dotnet-policy-service]

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

[AndyAyersMS]

There are quite a few diffs. Can you investigate?

Question: should we always spill newarr instead of only do this when optimizations are enabled?

I would follow the pattern we have for newobj, which always spills to a temp.

[hez2010]

I would follow the pattern we have for newobj, which always spills to a temp.

This will lead to huge diffs for tier-0 so I will do this later after I investigated the diff.

[AndyAyersMS]

Latest diffs look pretty good. Seems like if we're going to allocate and store an array into a static it is generally better to do the array creation before fetching the static's address.

Can you look at what happens if we also do this when not optimizing?

[hez2010]

Diffs

[AndyAyersMS]

Diffs

Ah, ok. Let's undo that last commit and just do this when optimizing.

Suggests we should probably look into doing the same thing for newobj, or never split up front, and instead search and split in the object allocation code.

[MichalPetryka]

Suggests we should probably look into doing the same thing for newobj, or never split up front, and instead search and split in the object allocation code.

Is a temp not always needed for newobj in order to call the ctor?

[AndyAyersMS]

Suggests we should probably look into doing the same thing for newobj, or never split up front, and instead search and split in the object allocation code.

Is a temp not always needed for newobj in order to call the ctor?

It is today, but you could imagine arranging things so that the ctor returns the object, perhaps.

[hez2010]

Ah, ok. Let's undo that last commit and just do this when optimizing.

Done.

[AndyAyersMS]

Thanks!

[AndyAyersMS]

I think the x86 failure was fixed, going to merge and see if we can get a green CI.