[PATCH] D23334: [Sparc][Leon] Speed optimisation to reduce double floating-point instructions to singles.

[James Y Knight via llvm-commits]

jyknight added a comment.

This seems like the wrong level at which to do such a transformation.

It would be better as a generic IR pass IMO -- but I'm also wondering how this makes sense to even do at all. Why wouldn't code that wants single-float math just do single-float math?


Repository:
  rL LLVM

https://reviews.llvm.org/D23334