[PATCH] D62604: [CodeGen] Generic Hardware Loop Support

[Markus Lavin via Phabricator via llvm-commits]

markus added a comment.

What happens if a target decides late (like `addPreEmitPass()` late) that a hardware loop is no longer possible due to reasons that are not detectable from the IR level hook. For PowerPC this does not seem to be an issue, or at least I cannot find it in the code, but for other targets there could be limitations that you need to analyse the actual machine instructions to find out about. Since the original IV chain has been replaced by intrinsic `llvm.loop.decrement.reg` (or whatever it iselected to) it seems we now would need to manually insert instructions to update the loop counter which could possibly be a bit hairy to do this late.

On the other hand (but not saying that it is better) if one treats the intrinsic as a pure hint and keep the original loop counter intact then the problem here becomes easier as it would just be a matter of deleting instructions if hardware loop insertion is successful and doing nothing if it fails.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D62604/new/

https://reviews.llvm.org/D62604