Need help with FPU/MMX/SSE state save/restore

[Denis Vlasenko]

I need to use MMX and/or SSE registers in my code.

Have looked at arch/i386/lib/mmx.c,
	kernel_fpu_begin();
	muck_with_MMX_regs();
	kernel_fpu_end();
seems to be the way to go.

My code is in .S, so I did the equivalent
(see <------), but routine crashed and burned
as soon as it tripped over kernel_fpu_begin.
Commenting KERNEL_FPU_BEGIN made code work
(I was able to boot with root=/dev/fd0)
but it won't be usable (will trash userspace
FP/MMX/SSE registers).

Does it have anything to do with the fact I'm
using this code from csum_partial[_copy_generic] ?
--
vda

...
#define	KERNEL_FPU_BEGIN \
	call	kernel_fpu_begin
#define	KERNEL_FPU_END(r) \
	movl	%cr0, r	;\
	orl	$8, r	;\
	movl	r, %cr0
...
# "big chunks" loop
	PREFETCH((%esi))	# Prefetch _each_ cacheline
	PREFETCH(32(%esi))
	PREFETCH(64(%esi))
	PREFETCH(64+32(%esi))
	PREFETCH(128(%esi))
	PREFETCH(128+32(%esi))
	PREFETCH(192(%esi))
	PREFETCH(192+32(%esi))
	KERNEL_FPU_BEGIN	<--------------
	clc
#define ROUND(x,r) \
SRC(	movq	x(%esi), r	);	\
	adcl	x(%esi), %eax	;	\
	adcl	x+4(%esi), %eax	;	\
DST(	movntq	r, x(%edi)	);
10:
	PREFETCH(256(%esi))
	ROUND(%mm0)
	ROUND(8,%mm0)
	ROUND(16,%mm0)
	ROUND(24,%mm0)
	lea	ITER_SZ(%esi), %esi
	lea	ITER_SZ(%edi), %edi
	loop	10b

	adcl	$0, %eax
	sfence
	KERNEL_FPU_END(%ebx)	<------------