Patch from Lennert Buytenhek

Hi,

I need the patch below (against 2.6.8-rc1-ds1) to make nwfpe properly
emulate arithmetic with doubles on a big endian ARM platform.

From reading the mailing list archives and from helpful comments I've
received from people on this list, I gather that this has come up in
the past, but it appears that Russell King was never really convinced
as to why this patch is needed.  I think I understand what's going on,
and will try to explain.

On little endian ARM, the double value 1.0 looks like this when stored
in memory in FPA word ordering:
bytes: 0x00 0x00 0xf0 0x3f 0x00 0x00 0x00 0x00
u32s:  0x3ff00000 0x00000000
u64:   0x000000003ff00000

On big endian, it looks like this:
bytes: 0x3f 0xf0 0x00 0x00 0x00 0x00 0x00 0x00
u32s:  0x3ff00000 0x00000000
u64:   0x3ff0000000000000

It appears to be this way because once upon a time, somebody decided
that the sub-words of a double will use native endian word ordering
within themselves, but the two separate words will always be stored
with the most significant one first.  God knows why they did it this
way, but they did.

Anyway.  The key observation is that nwfpe internally stores double
values in the type 'float64', which is basically just a typedef for
unsigned long long.  It never accesses 'float64's on the byte level
by casting pointers around or anything like that, it just uses direct
u64 arithmetic primitives (add, shift, or, and) for float64
manipulations and that's it.

So.  For little endian platforms, 1.0 looks like:
0x00 0x00 0xf0 0x3f 0x00 0x00 0x00 0x00
But since nwfpe treats it as a u64, it wants it to look like:
0x00 0x00 0x00 0x00 0x00 0x00 0xf0 0x3f
So, that's why the current code swaps the words around when getting
doubles from userspace and putting them back (see fpa11_cpdt.c,
loadDouble and storeDouble.)

On big endian, 1.0 looks like:
0x3f 0xf0 0x00 0x00 0x00 0x00 0x00 0x00
Since nwfpe treats it as a u64, it wants it to look like:
0x3f 0xf0 0x00 0x00 0x00 0x00 0x00 0x00
Hey!  That's exactly the same.  So in this case, it shouldn't be
swapping the halves around.  However, it currently does that swapping
unconditionally, and that's why floating point emulation messes up.

This is how I understand things -- hope it makes sense to other people
too.


cheers,
Lennert