Re: init fails w/ new build of gcc & glibc

[Mark Hatle <fray@mvista.com>]

> Mark,
>
> You mentioned these two exclusive ABIs a couple days ago. If you want to compile the
> toolchain and the kernel correctly to insure run-time compatibility (in my case for 8xx),
> would you do this
>
> for kernel fp emulation
> compile kernel with CONFIG_MATH_EMULATION=y, (do you use -msoft-float while compiling
> kernel?)
> then compile the toolchain without -msoft-float

Use the kernel parameters as they stand, I wouldn't mess with them....
(floating point is not supposed to happen in the kernel anyway, and when
it does it's almost always assembly that is trapped anyway.)

> or
>
> for soft-floating point - no kernel emulation
> compile kernel CONFIG_MATH_EMULATION is not set (do you use -msoft-float while compiling kernel?)
> then compile the toolchain with -msoft-float

Basically.. One thing to keep in mind on the 8xx (due to cache line
size) it'll never truely be compatable with the rest of the PPC world.
But on a processor like the 405 it "can be" compatable using the
CONFIG_MATH_EMULATION.  The catch is that you have a lot more overhead
doing emulation.... (The emulation itself is pretty fast, certainly no
slower then the software code invoked in -msoft-float..)  However, you
have much more context switching over head and possible cache
invalidation.  Also, your software could be using floating point
registers w/o you expliciting using float's or double's due to gcc
optimizations to argument passing.  All in all you will have a penalty
for using hard floating point on a soft-floating point architecture.

This is the reason that MontaVista has chosen to go the route of
soft-float...

--Mark

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/