[PATCH] D79785: [ARM] Register pressure with -mthumb forces register reload before each call

[Prathamesh via Phabricator via llvm-commits]

prathamesh added a comment.

In D79785#2095601 <https://reviews.llvm.org/D79785#2095601>, @ostannard wrote:

> I'd expect to see some extra checks to decide if reverting the indirect-call optimisation is profitable or not. If the loaded address is used more than three times, without being spilled, then I think it's better to leave the calls as indirect ones. If you don't think this is ever likely to be worthwhile for thumb 1, then it would be better to modify ARMTargetLowering::LowerCall to not do the transformation in the first place.


Hi Oliver,
Instead of using a heuristic in LowerCall, we decided to convert blx -> bl using foldMemoryOperand hook only if the register holding function's address got spilled,
and the patch implements the hook for Thumb1.
For the above test-case, it seems the load of function's address was rematerialized 3 times, which was eliminated by converting the three blx to bl's.
This is till not optimal, but better than the original behavior, without affecting other test-cases and heuristics.

Thanks,
Prathamesh


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D79785/new/

https://reviews.llvm.org/D79785