Description:
------------
I ran into an issue with pow() that causes it to lose precision on 64 bit 
platforms if one or more of the arguments is a float.  

I've attached a patch that fixes this by checking if both arguments are exactly 
representable as an integer, and if so converting to an integer.

This attempts to retain as much precision as possible by letting the result be an 
exact representation where possible.

Additionally, the patch fixes a tested-for-bug on 64 bit platforms where:

pow(-9.2233720368548E+18, 9223372036854775807) == INF when it should be -INF.



Test script:
---------------
var_dump((int) (pow(2, (float) 55) - 1));

Expected result:
----------------
int(36028797018963967)

Actual result:
--------------
int(36028797018963968)