[avr-libc-dev] pow(x,y) broken for x<0

[Edwin Olson]

Hi folks,

I believe the implementation of pow is inconsistent with the standard 
POSIX definition. The one provided by avr libc assumes x>0 so that it 
can approximate x^y with exp( log(x) * y).

The POSIX definition allows x<0, provided that y is integral valued.

In my case, I was computing a euclidean distance ((x1-x0)^2 + 
(y1-y0)^2), using pow for no good reason, but it crashed the atmel with 
an apparant stack overflow. (Presumably log(x) is ill-behaved for x<0).

x<0 could be relatively easily handled with a little wrapper which 
computes pow (abs(x), y) * (-1 if y is odd).

Thanks,

-Ed