[PATCH] [x86] set default reciprocal (division and square root) codegen to match GCC
spatel at rotateright.com
Fri Jun 19 18:58:41 PDT 2015
In http://reviews.llvm.org/D10396#191356, @ahatanak wrote:
> I just want to make sure I understand where this is headed. The recip flag and refinement step are eventually going to be added to the IR fast-math flags and propagated to the DAG nodes, is that correct?
That's correct, assuming there are no showstoppers for FMF in the DAG. The alternative would be to add a function attribute for this. Without one of those solutions, it's my understanding that we can't specify recip codegen when using LTO.
More information about the llvm-commits