JIT: Late expansion for casts

[EgorBo]

Move cast expansion from importer to a late phase. For now, only for profiled isinst.

Closes#96837
Contributes to #84025

[ghost]

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

Issue Details

---

Move cast expansion from importer to a late phase. For now, only for profiled isinst.

Closes #96837
Contributes to #84025

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[EgorBo]

/azp run runtime-coreclr pgo, runtime-coreclr libraries-pgo, runtime-coreclr pgostress

[azure-pipelines]

Azure Pipelines successfully started running 3 pipeline(s).

[EgorBo]

@jakobbotsch @dotnet/jit-contrib PTAL. Diffs. There are few things in the diffs:

1. Improvements where we were able to hoist/CSE casts.

2. Improvements where we realized block is cold and it doesn't need cast expansion, E.g.
   
   (it was previously expanded)

3. Regressions due to extra nullcheck this phase emits and no other phase can clean it up. But I already have a follow up patch to fix it via assertionprop

4. Possible regressions because of CSE'd class handle argument - I wasn't able to find them but it may happen. This phase works as is if I move it after assertionprop so I can rely on VN to get it. Probably will do that separately.

Will move more casts to this phase from importer (all of them hopefully). For now only recently introduced profiled casts are handled.

[jakobbotsch]

LGTM, but looks like CI is a bit unhappy.

[EgorBo]

/azp run runtime-coreclr pgo, runtime-coreclr libraries-pgo

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[EgorBo]

LGTM, but looks like CI is a bit unhappy.

Ah, I broke the control flow a bit - nullcheckBb used to jump straight to the use block without assigning tmp's value (but in most cases it was zeroed anyway so CI mostly passed.

Instead, I now do:

Case 1:

tmp=obj;
if (tmp!=null)
{
if (tmp.pMT!=likelyClass)
tmp=helperCallelseno-op; // this block will be collected
}
use(tmp);

Case 2 (likelyClass is known to never pass the type check):

tmp=obj;
if (tmp!=null)
{
if (tmp.pMT!=likelyClass)
tmp=helperCallelsetmp=null;
}
use(tmp);

I can avoid the initial tmp = obj; by introducing an extra BB but I think this shape is simpler.

[EgorBo]

Failure is #97103