fdlibm: reduce accuracy slightly in favor of performance.

fdlibm: reduce accuracy slightly in favor of performance.

[Raymond Toy, 2014-06-02 16:55:53]

I don't see that saving 2ms (from 17) here is worth the loss of accuracy.  

Accurate reduction doesn't come for free and this is much better than the
current version for medium to large arguments.

https://codereview.chromium.org/305403002/diff/1/src/math.js
File src/math.js (right):

https://codereview.chromium.org/305403002/diff/1/src/math.js#newcode199
src/math.js:199: y = x - pio2_1 - pio2_2 - pio2_2t;
I don't think this is computing the same values as before.

What happened to the if statement? The value of y0 is different on the two
branches.  Can you prove that y is the same value as y0 on both branches.

And I think you need to be careful about the order of operations in computing y.
 It's not clear to me that the compiler won't notice that pio2_1, pio2_2, and
pio2_2t are constants that the compiler won't just coalesce into one constant,
thereby destroying all of the extra bits you actually want to use.