using mpn_preinv_divrem_1

[Ashod Nakashian]

Torbjorn Granlund wrote:
> Ashod Nakashian <saghmos at xter.net> writes:
> 
>   I understand that. What I was asking was why mpn_divrem_1 isn't 
>   precomputing the inverse on-the-fly?
> 
> It sure does, look at the code, in mpn/generic/divrem_1.c.
> DIVREM_1_NORM_THRESHOLD and DIVREM_1_UNNORM_THRESHOLD
> controls this.
> 
>   Precomputing the inverse should pay off for a few limb's 
>   long numbers and more, assuming a (mul + overhead) is faster on the 
>   target CPU than a plain div.
>   
> That's the experience of the GMP hackers.
> 
>   Now, I'm willing to experiment with this on my AMD-64 3200+ to see how 
>   far I can get. But my next question is how do you compute the inverse of 
>   the divisor? As I stated in my previous mail, the traditional method, as 
>   I understand it, doesn't work. There is some assumption that I'm missing.
>   
> All details are in the paper "Division by Invariant Integers using
> Multiplication", <http://swox.com/~tege/divcnst-pldi94.pdf>.  I
> think the source code for invert_limb might also give you a hint.
> 
> --
> Torbjörn
> 
> 

Thanks!
-Ash