Re: mozilla-* porter help needed for ARM
On Tue, Oct 04, 2005 at 08:40:59AM -0700, Peter Naulls wrote:
> On the contrary, autoconf makes many things consistent and predictable
> when cross compiling, and improvements to autoconf and other tools have
> made the possibility of cross compiling much more likely for a given
> package. You're more likely to have a problem with a package that
> doesn't use it.
> However, the setup for Debian packages isn't always ideal for cross
> building even though Mozilla certainly does.
Yeah having full support for cross compiling would be nice, but then
again the current buildd system at least ensures it really does build on
the architecture itself.
Cross compiling also can give slightly different results than a native
For example if you have something like this:
#define SPEED 1000
#define REDUCEDSPEED SPEED*0.1
On a native compile, it precalculates REDUCEDSPEED as 100 in any place
that needs it and inserts it at compile time.
Doing a cross compile, it seems gcc doesn't assume it knows how the
architecture does floating point calculations and rounding and such, so
it inserts the actual floating point call instead and the calculation is
done at runtime.
Had the code said SPEED/10 instead, it would have done the interget
calculation at compile time either way since it knows how integers work
on all systems.
This means the resulting binary may run slightly slower if cross
compiled than if native compiled, while in the case of kernel code where
floating point is a no no, you end up with code that can't be cross
compiled and used sincce you can't load the module into the kernel due
to the floating point calls (and lack of floating point hardware on some
architectures that emulate it when user space needs it).