[PATCH] D26855: New unsafe-fp-math implementation for X86 target

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In https://reviews.llvm.org/D26855#608851, @avt77 wrote:

> BTW, I have a question:  this is the first patch in the possible series of patches. This patch introduces the the suggested approach on the example of FDiv - Recip implementation. Should I complete this patch with FDIV only or should I extend it with rsqrt? The patch is rather big that's why I suggest to make it really working with fdiv only and commit. After that I'll be able to make other similar patches faster and safely. Is it OK?


I think this is the right direction because we want to make use of the actual CPU scheduler model instead of estimates, but implementing this will almost certainly lead to unintended consequences that need to be worked around. For example, one thing I noticed when adding the reassociation patterns to machine combiner: we could significantly increase compile time because calculating critical path/resource height with MachineTraceMetrics can be expensive.

You should send a note to llvm-dev about this plan, the motivation, the benefits, etc. and reference this patch as the starting point. The changes you are proposing could be adapted by all targets, so there should be feedback from various backend people to avoid duplicated effort.


https://reviews.llvm.org/D26855