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

[Andrew V. Tischenko via Phabricator via llvm-commits]

avt77 added a comment.

Chandler, in fact this patch should not show any improvement in generating code. If you look in changes made in tests you'll see that the newly generated code is almost identical to the previous one (only some names, order of instructions, etc.). The idea of the patch is moving of such kind of optimization from the rather high level (DAGCombiner) to the really low level (MachineCombiner), Here we see real target machine instructions and as result we can use real cost model to estimate the real cost of possible transformation (in the given case the transformation is the replacement of one instruction (div) with some sequence of instructions). The transformation itself already exists inside Clang but the patch suggests to implement it in another place and that's it. If we agree with this new place of implementation then it will be the base for future possible similar optimizations like rsqrt, etc. And in addition this (and follow up) patch(es) will allow us to remove 'fake' subtarget features like FeatureFastScalarFSQRT / FeatureFastVectorFSQRT, etc. The question from Gerolf was not about the quality of the generated code (it's the same like we have now) but about the compilation time only.

Of course I'll try to collect the required perf numbers but they should be the same like we have now. Do we really need it?


https://reviews.llvm.org/D26855