[PATCH] D53453: [ARM] Make InstrEmitter mark CPSR defs dead for Thumb1.

[Dave Green via Phabricator via llvm-commits]

dmgreen added subscribers: philipginsbach, qcolombet.
dmgreen added a comment.

Interesting. I didn't realise it worked like that. I had presumed that lot of passes would have to be taught about the optional defs, as opposed to them not being marked as dead correctly.

In my testing, the codesize is kind of up and down, looking a little bit up overall. It's roughly a wash, with one big increase in bzip for some reason.

But the performance looks like a big improvement. It fixes the regression we got from https://reviews.llvm.org/rL339472 and shows lots of other I'd previously attributed to https://reviews.llvm.org/D33935. To the extent that https://reviews.llvm.org/D33935 on top of this actually looks it hurts more than it helps. I'm not sure why yet, I would expect them to help each other out more than they hurt each other.


Repository:
  rL LLVM

https://reviews.llvm.org/D53453