[PATCH] D38164: [MachineScheduler] Favor instructions that do not increase pressure.

[Matthias Braun via Phabricator via llvm-commits]

MatzeB added a subscriber: atrick.
MatzeB added a comment.

I think the cryptocode I was looking at back then was john the ripper (openbenchmarking.org / pts/john-the-ripper).

Just tried with/without your changes and I see the descrypt part improving and the LM part regressing. So whether this is an improvement or regression may depend on the situation and we should find/look at examples.

I assume Andy designed the current behavior to give more freedom to the latency minimizing code? So an out of order Cortex-A57 may not be the best test case for it; then again if it improves the spilling situation and leads to a win on out of order cores we could switch the behavior based on out-of-order/in-order.


https://reviews.llvm.org/D38164