Proposal: Remove or reduce math-finite.h

[Wilco Dijkstra]

Hi Joseph,

>> These are old x87 assembler implementations. Performance would
>> improve if they were removed.
>
> Do you have benchmark figures to justify that (like the ones HJ provided 
> when removing float versions of such functions)?

I ran exp using the expf trace on x86 and x64. There is no contest, the
generic math code shows around 6 times the throughput:

x86 exp using x87 assembler:

   "workload-spec2017.wrf": {
    "duration": 2.99568e+10,
    "iterations": 2.24096e+08,
    "reciprocal-throughput": 123.62,
    "latency": 143.737,
    "max-throughput": 8.08933e+06,
    "min-throughput": 6.95716e+06
   }

x64 exp using generic math:

   "workload-spec2017.wrf": {
    "duration": 3.00681e+10,
    "iterations": 9.34528e+08,
    "reciprocal-throughput": 20.9929,
    "latency": 43.3563,
    "max-throughput": 4.76351e+07,
    "min-throughput": 2.30647e+07
   }

Wilco