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

[Shuxin Yang]

I recall another reason I ditch pattern-match approach.  It is extremely 
clunky to estimate how many instructions we are going to save.
Each time it come cross a muti-use instruction, it simply give up.

However, my implementation look ahead regardless single-use and 
multi-use, and evaluate the # of instruction the optimization
save at the later stage, which enable us to find more opportunities.

>>
> I start with pattern match, and gave up later on. It seems to be too 
> messy.