[PATCH] D13417: [MachineCombiner] make slack optional in critical path cost calculation (PR25016)

[Sanjay Patel via llvm-commits]

spatel added a comment.

> The crux is that the trace metrics assumes all calls can execute in parallel. That looks like an unrealistic assumption. This suggests there is a deeper modeling problem/question we need to understand first before jumping into solution space.


That was my "2nd thought" comment in this thread on Oct 6. And I agree that we could try to improve machine trace metrics, but my "3rd thought" comments from Oct 7 explain my current opinion about the problem: for a reassociation transform, we should be more conservative in our usage of MTM. It will save compile time and makes us immune to possible bugs in MTM or whatever CPU models it is relying on. I don't see any downside to doing this.


http://reviews.llvm.org/D13417