JIT: removed unneeded throw helper blocks

[AndyAyersMS]

Various optimizations can happen between the time the throw helper blocks are first requested (morph) and finally used (codegen). Prune away ones that aren't needed during the stack level setter.

Fixes#93948.

[ghost]

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

Issue Details

---

Various optimizations can happen between the time the throw helper blocks are first requested (morph) and finally used (codegen). Prune away ones that aren't needed during the stack level setter.

Fixes #93948.

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

[AndyAyersMS]

Diffs

For code size this is pure goodness, but the TP impact is a bit high. Some possible mitigations:

• there's not much benefit on min opts so we can do some of this more selectively (only if required)
• the caching done by fgFindExcptnTarget is likely not that effective, seems like we should just keep a full hash table. Want to verify this is the big extra cost first.

[AndyAyersMS]

Also what I'm doing likely won't work for SCK_FAIL_FAST but we don't seem to be using it? It is easy enough to just always keep these if they are pre-generated.

[jakobbotsch]

Since GTF_CALL and GTF_EXCEPT are up-to-date and overapproximations here you can optimize TP with this trick:

runtime/src/coreclr/jit/promotion.cpp

Lines 2340 to 2349 in 3805c17

	if (((*use)->gtFlags & (GTF_EXCEPT | GTF_CALL)) == 0)
	{
	assert(!(*use)->OperMayThrow(m_compiler));
	return use;
	}
	if (!(*use)->OperMayThrow(m_compiler))
	{
	return use;
	}

[AndyAyersMS]

TP looking okayish now (Diffs), realized the stack level setter was disabled for mac arm64, so enabled it there.

As part of this I will also close out #42673.

Final Diffs. If we could merge all this new analysis into lower we might be able to shave off some TP, but that seems more complicated. Since this phase previously did not run for osx-arm64 we can see the overall TP hit is around 0.5% in minopts and 0.2% when optimizing.

[AndyAyersMS]

Should be ready for final review.

FYI @dotnet/jit-contrib

[BruceForstall]

Minor notes

[BruceForstall]

Minor, but maybe you should follow the pattern set by GetBlockToEHPreds and many similar functions, to have a:

AddCodeDscMap* GetAddCodeDscMap()
{
if (fgAddCodeDscMap == nullptr)
{
fgAddCodeDscMap = new (getAllocator(CMK_Unknown)) AddCodeDscMap(getAllocator(CMK_Unknown));
}
return fgAddCodeDscMap;
}

Instead of having the allocation within fgAddCodeRef

[BruceForstall]

Remove or enable

[AndyAyersMS]

SPMI failures are known bad contexts for OSX. Ignoring.