Improve EH dead code elimination

[BruceForstall]

Update fgRemoveDeadBlocks() to only treat EH handlers as reachable if their corresponding try block entry is reachable.

This causes more dead code to be eliminated in handlers that are not reachable.

There is a little more work to do to completely clean these up:

1. If a try is not reachable, we should get rid of its entire EH table entry. Currently, its block is generated as int3 (on xarch).
2. Even though the handler is not reachable, we still generate the prolog followed by int3 for the body.

Contributes to #82335

[ghost]

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

Issue Details

---

Update fgRemoveDeadBlocks() to only treat EH handlers as reachable if their corresponding try block entry is reachable.

This causes more dead code to be eliminated in handlers that are not reachable.

There is a little more work to do to completely clean these up:

1. If a try is not reachable, we should get rid of its entire EH table entry. Currently, its block is generated as int3 (on xarch).
2. Even though the handler is not reachable, we still generate the prolog followed by int3 for the body.

Contributes to #82335

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

[BruceForstall]

@AndyAyersMS PTAL
cc @dotnet/jit-contrib

[kunalspathak]

LGTM

[AndyAyersMS]

1. If a try is not reachable, we should get rid of its entire EH table entry.

We encode EH region numbers into the AddCodeDsc entries (starting in morph) so if we remove things from the EH table we'll need to fix these up too. Or perhaps rework all this so these entries hang off the EH table somehow and migrate organically along with their related regions.

[BruceForstall]

We encode EH region numbers into the AddCodeDsc entries

In particular, using bbThrowIndex(), to ensure fgAddCodeRef only shares throw blocks in the same EH region.

[AndyAyersMS]

Interesting that we don't see much of a TP win here.

I wonder if we should try doing this earlier?

[BruceForstall]

Interesting that we don't see much of a TP win here.

Would you expect a TP win because of less IR? It seems like unreachable handlers should be rare, especially after the early EH opts. This still doesn't remove an unreachable try or the unreachable handler (it just turns them into int3), or the "dead" EH region.

I wonder if we should try doing this earlier?

Seems like it needs to be late enough for all the normal opts creating dead flow to run.

[BruceForstall]

/azp run runtime-coreclr jitstress

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[BruceForstall]

Test failure is #93506