[PATCH] D14079: [FPEnv Core 14/14] Introduce F*_W_CHAIN instrs to prevent reordering

[Sergey Dmitrouk via llvm-commits]

sdmitrouk added a comment.

In http://reviews.llvm.org/D14079#276541, @jyknight wrote:

> Drive-by-comment...
>
> Is there an advantage to duplicating FADD, FSUB, FMUL, FDIV, and FREM into a second W_CHAIN set, instead of having only a single set of ops which have the chain operand? It would naively seem better to have less duplication? (That is, if the compilation flags say to not care about fp ordering, the input chain would still be there, but would always be set to the entry node.)


This was the first implementation, but the issue was that in that case you can't (or I don't know how) use same instruction twice to create one node with side effects and the other one without (and model changes in related registers probably too). Hal actually said that one could add new instructions and for me it also appeared like a safer choice to do not break current floating-point math.


Repository:
  rL LLVM

http://reviews.llvm.org/D14079