Re: [Mingw-users] expm1()

[Greg C. <chi...@co...>]

On 2005-01-26 02:48 AM, Danny Smith wrote:

> Greg Chicares wrote:
> 
>>>Maybe I could provide a narrative description of changes to
>>>  mingw/runtime/mingwex/math/expl.c
>>>that would be easy to translate. The intel instruction set actually
>>>treats expm1() as the general case, and you have to undo the 'minus
>>>one' part to implement plain exp(). It might be as simple as [...]
>>
>>...yes, it is that simple, and here's a clean-room 'not-a-patch'
>>that should make it easy to implement without copyright or
>>license issues:
>>
>>// Make a copy of mingwex/math/expl.c .
>>//
>>// Change the function name (I happened to use 'NEW_FUNCTION').
>>//
>>// The original function uses an asm instruction that already
>>// calculates expm1(), which is what we want; then it adds one,
>>// which is not what we want, so get rid of the code that adds one.
>>//
>>// The inline-asm function should be used only within its domain of
>>//   (-.5*M_LN2, +.5*M_LN2)
>>// A cover function ought to test whether its argument is outside
>>// that domain and, if it is, simply call exp() and return its
>>// result minus one.
> 
> If we restrict domain of an our asm __expm1 to (-.5*M_LN2, +.5*M_LN2), than much
> of the code
> in exp() is really unnecessary. We don't need to worry about the integral part
> of  the separaton
> 
> x log2(e) = i + f, |f| <= .5

You're right.

I was trying to change as little as possible in order to avoid
FSF issues, but this evening I remembered that mingwex isn't
part of the GNU project.

I'll rewrite patch 1109674 to reflect your suggestion, which
will save some cycles as well as making things clearer. I think
the domain can be made twice as broad, too.