JIT: tail merge returns with multiple statements

[AndyAyersMS]

Remove the restriction that a mergeable return have one statement.

[dotnet-policy-service]

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

[AndyAyersMS]

@EgorBo PTAL
cc @dotnet/jit-contrib

Like #109661, I spotted this when I was experimenting with a very early flow opts pass. Seems to have good diffs. There are some cases now where we'll jump to a ret, but that sort of thing should be handled later on, once we have an idea how big the epilogs can be.

[EgorBo]

I remember that I wanted to relax it in a follow up PR after that change, but forgot 🙂

[EgorBo]

the diffs look nice, but I guess the CI failure is related? I don't recall it when I was disabling it.

Looks like tail call related

[AndyAyersMS]

the diffs look nice, but I guess the CI failure is related? I don't recall it when I was disabling it.

Looks like tail call related

Yes, they are quite likely related.

[AndyAyersMS]

the diffs look nice, but I guess the CI failure is related? I don't recall it when I was disabling it.
Looks like tail call related

Yes, they are quite likely related.

It's because X86 does not set FEATURE_TAILCALL_OPT and so does not set FEATURE_TAILCALL_OPT_SHARED_RETURN and so can't handle having a block with a (nonrecursive) tail call branch to a return block.

Not sure why x86 is special here, but assuming it needs to be this way, I can just use the old logic for x86, or check if the statement before the return is a tail call candidate, and defer merging if so.

FYI @jakobbotsch

[AndyAyersMS]

the diffs look nice, but I guess the CI failure is related? I don't recall it when I was disabling it.
Looks like tail call related

Yes, they are quite likely related.

It's because X86 does not set FEATURE_TAILCALL_OPT and so does not set FEATURE_TAILCALL_OPT_SHARED_RETURN and so can't handle having a block with a (nonrecursive) tail call branch to a return block.

Not sure why x86 is special here, but assuming it needs to be this way, I can just use the old logic for x86, or check if the statement before the return is a tail call candidate, and defer merging if so.

FYI @jakobbotsch

I ended up not splitting returns from any tail call as there is a fair bit of twisty logic in morph that expects certain shapes.