code review 3377041: math: fix Exp2 for integer exponents

math: fix Exp2 for integer exponents

For non-negative integer exponents, the results for
floating-point and integer arithmetic should be the same.

[peterGo, 2010-12-01 21:10:56]

Hello rsc (cc: golang-dev@googlegroups.com),

I'd like you to review this change.

[ejsherry, 2010-12-04 02:47:56]

I don't think special casing 0 to 63 is such a good idea. Exp2 will still be
inaccurate for most arguments. I think problem is x*Ln2 isn't precise enough
when x gets larger than ±1. I have a version which does argument reduction
before calling Exp and it seems to behave well. I'll try and get a CL up soon.

The tests probably don't need all those conversions; testing that
Exp2(float64(n))==Ldexp(1,n) should suffice. And n should range from -1022-52 to
1023.

I think it's worth looking into using SSE in the ASM version. The SSE version of
Exp is faster than the x87 version. Converting it to Exp2 may be the way to go.

[rsc1, 2010-12-06 19:38:45]

Closing in favor of ejsherry's CL 3481041