[llvm-commits] [PATCH][FastMath, InstCombine] Fadd/Fsub optimizations

[Shuxin Yang]

On 12/11/12 4:14 PM, Eli Friedman wrote:
> Then what cost are you trying to avoid by saving a pointer to the
> FAddCombine?  Constructing an FAddCombine should be basically free.
The cost of FAddCombine::FAddCombine is nothing.

But I don't like to add this stmt to InstCombiner as I have to expose to 
the FAdd* classes
to the InstCombine.h:

FAddCombine TheGuy;

I just feel more comfortable with "FAddCombine *TheGuy";


> My high-level concern with this patch is basically that this is a lot
> of code, the algorithm is complex, and the implementation is hard to
> reason about; I'm trying to find places where can simplify if
> possible.  And I'm immediately skeptical of any performance claim
> without numbers.
Some 700 lines of code is nothing.
Among them 650 lines are about utility, the core has about 50 lines.