Implement Math.expm1 using port from fdlibm.

Implement Math.expm1 using port from fdlibm.


R=rtoy@chromium.org
BUG=v8:3479
LOG=N

Committed: https://code.google.com/p/v8/source/detail?r=23238

[Yang, 2014-08-13 13:37:53]

On 2014/08/13 13:33:19, Yang wrote:

Title was wrong. Fixed.

[Raymond Toy, 2014-08-13 16:25:17]

LGTM

Just a few nits.

https://codereview.chromium.org/465353002/diff/1/test/mjsunit/es6/math-expm1.js
File test/mjsunit/es6/math-expm1.js (right):

https://codereview.chromium.org/465353002/diff/1/test/mjsunit/es6/math-expm1....
test/mjsunit/es6/math-expm1.js:24: }
For an additional sanity check, why not use the mathematical property that
expm1(n*log(2)) = exp(n*log(2)) - 1 = 2^n - 1? For simplicity, just use integer
n. For bonus points allow n = m/2 for integer m.  This only requires computing
sqrt(2). Or perhaps Math.pow is accurate to < 1 ulp?

The above sanity check would fail if exp(x) is wrong, and we know that exp(x) is
(now) less accurate than expm1.

https://codereview.chromium.org/465353002/diff/1/third_party/fdlibm/fdlibm.js
File third_party/fdlibm/fdlibm.js (right):

https://codereview.chromium.org/465353002/diff/1/third_party/fdlibm/fdlibm.js...
third_party/fdlibm/fdlibm.js:429: macro KLOG1P(x)
What changed here? And how is this related to expm1?

https://codereview.chromium.org/465353002/diff/1/third_party/fdlibm/fdlibm.js...
third_party/fdlibm/fdlibm.js:513: (KLOG1P(4) + z * (KLOG1P(5) + z *
KLOG1P(6)))))));
This seems unrelated to expm1. And I can't see what actually changed here..

[Yang, 2014-08-20 14:18:24]

https://codereview.chromium.org/465353002/diff/1/test/mjsunit/es6/math-expm1.js
File test/mjsunit/es6/math-expm1.js (right):

https://codereview.chromium.org/465353002/diff/1/test/mjsunit/es6/math-expm1....
test/mjsunit/es6/math-expm1.js:24: }
On 2014/08/13 16:25:17, Raymond Toy wrote:
> For an additional sanity check, why not use the mathematical property that
> expm1(n*log(2)) = exp(n*log(2)) - 1 = 2^n - 1? For simplicity, just use
integer
> n. For bonus points allow n = m/2 for integer m.  This only requires computing
> sqrt(2). Or perhaps Math.pow is accurate to < 1 ulp?
> 
> The above sanity check would fail if exp(x) is wrong, and we know that exp(x)
is
> (now) less accurate than expm1.

I like the suggestion. However, n*Math.LN2 is less precise than it should be, so
I would have to use an assertEqualsDelta with expected*1E-14 as delta, which
makes this rather pointless.

https://codereview.chromium.org/465353002/diff/1/third_party/fdlibm/fdlibm.js
File third_party/fdlibm/fdlibm.js (right):

https://codereview.chromium.org/465353002/diff/1/third_party/fdlibm/fdlibm.js...
third_party/fdlibm/fdlibm.js:513: (KLOG1P(4) + z * (KLOG1P(5) + z *
KLOG1P(6)))))));
On 2014/08/13 16:25:17, Raymond Toy wrote:
> This seems unrelated to expm1. And I can't see what actually changed here..

it used to be KLOGP1. I realized that it should be KLOG1P instead. I thought I
could piggyback this small change here.

[Yang, 2014-08-20 14:24:18]

Committed patchset #1 manually as 23238 (presubmit successful).

[Raymond Toy, 2014-08-20 16:18:20]

https://codereview.chromium.org/465353002/diff/1/test/mjsunit/es6/math-expm1.js
File test/mjsunit/es6/math-expm1.js (right):

https://codereview.chromium.org/465353002/diff/1/test/mjsunit/es6/math-expm1....
test/mjsunit/es6/math-expm1.js:24: }
On 2014/08/20 14:18:24, Yang wrote:
> On 2014/08/13 16:25:17, Raymond Toy wrote:
> > For an additional sanity check, why not use the mathematical property that
> > expm1(n*log(2)) = exp(n*log(2)) - 1 = 2^n - 1? For simplicity, just use
> integer
> > n. For bonus points allow n = m/2 for integer m.  This only requires
computing
> > sqrt(2). Or perhaps Math.pow is accurate to < 1 ulp?
> > 
> > The above sanity check would fail if exp(x) is wrong, and we know that
exp(x)
> is
> > (now) less accurate than expm1.
> 
> I like the suggestion. However, n*Math.LN2 is less precise than it should be,
so
> I would have to use an assertEqualsDelta with expected*1E-14 as delta, which
> makes this rather pointless.

It has the great advantage that the test is independent of the implementation of
Math.exp and has analytically correct results.  Yes, n*Math.LN2 is not quite the
same as n*log(2), but that's life with floating-point.

https://codereview.chromium.org/465353002/diff/1/third_party/fdlibm/fdlibm.js
File third_party/fdlibm/fdlibm.js (right):

https://codereview.chromium.org/465353002/diff/1/third_party/fdlibm/fdlibm.js...
third_party/fdlibm/fdlibm.js:513: (KLOG1P(4) + z * (KLOG1P(5) + z *
KLOG1P(6)))))));
On 2014/08/20 14:18:24, Yang wrote:
> On 2014/08/13 16:25:17, Raymond Toy wrote:
> > This seems unrelated to expm1. And I can't see what actually changed here..
> 
> it used to be KLOGP1. I realized that it should be KLOG1P instead. I thought I
> could piggyback this small change here.

Sounds good. Sorry I missed that in the review of the log1p function.