RyuJIT: reconsider arm32 callfinally code pattern

[BruceForstall]

The JIT doesn't allow retless callfinally on ARM. Perhaps we should allow them, and generate "normal" calls instead of the mov lr, <continuation>; jmp <finally> form we generate now. Then we might be able to get rid of all the BBF_FINALLY_TARGET code, but we would need to set FEATURE_EH_CALLFINALLY_THUNKS=1. It might generate some worse code, which shouldn't normally matter because we now have finally cloning. We'd have to make sure the VM stack walker is ok with it. It would make ARM32 work just like ARM64, which might be advantageous. And if we could get rid of BBF_FINALLY_TARGET, that would be nice. If ARM32 chips implement a return address stack keyed off bl instructions (I don't know if they do), it might improve performance.

category:design
theme:eh

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

Issue Details

---

The JIT doesn't allow retless callfinally on ARM. Perhaps we should allow them, and generate "normal" calls instead of the mov lr, <continuation>; jmp <finally> form we generate now. Then we might be able to get rid of all the BBF_FINALLY_TARGET code, but we would need to set FEATURE_EH_CALLFINALLY_THUNKS=1. It might generate some worse code, which shouldn't normally matter because we now have finally cloning. We'd have to make sure the VM stack walker is ok with it. It would make ARM32 work just like ARM64, which might be advantageous. And if we could get rid of BBF_FINALLY_TARGET, that would be nice. If ARM32 chips implement a return address stack keyed off bl instructions (I don't know if they do), it might improve performance.

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