Re: [Valgrind-developers] Absolutely mystifying problem with fxrstor

[Philippe E. <ph...@wa...>]

Julian Seward wrote:
> Hi
> 
> I wonder if anyone can cast any light on the following.  I am
> completely mystified, not to mention stuck.
> 
> I'm doing stuff to add SSE/SSE2 support.  This means changing various
> fsave/frstor instructions, which move the FPU state back and forth
> between the simulated and real cpus, into their SSE equivalents,
> fxsave and fxrstor.  The SSE state is larger than the FPU/MMX 
> state so various structures have had their size increased.
> There is also a 16-byte alignment constraint on addresses in
> fxsave/fxrstor, so I've ensured that too.
> 
> Now, I think I've done everything right.  Nevertheless, one
> of my fxrstor's, in vg_syscalls.S, is segfaulting.  I have no
> idea why.  It's this
> 
> 	fxrstor	VG_(m_state_static)+64
> 
> (previously of course)
> 
> 	frstor	VG_(m_state_static)+64

Are you sure the state is not corrupted between this pair
of call ?


> and the address is duly 16-byte aligned, and the memory from
> that address for 512 bytes (the SSE state size) appears
> suitably accessible.  A few lines earlier there is
> 
> 	fxrstor	VG_(real_sse_state_saved_over_syscall)
> 
> and that works fine.
> 
> Help!  I'm stuck.  Is there some other magic constraints on
> fxrstor I need to know about?  I read the fine print in the P4
> documentation carefully, but I cannot see anything other than the
> 16-byte-alignment constraint.

there is other constraint to write mxcsr register see
11.6.6 in intel vol 1 doc, usefull only if you manipulate
state directly (actual code doesn't do that). Note than
fxrstor doc says than bit 6 is reserved but the documentation
is wrong, bit 6 can be set if you take enough care.

Also an errata on position of floating point tag word
(offset 4 not 5) but ditto as above if you don't touch
m_state_static+64 + ... it doesn't matter

Since the current code doesn't touch (m_state_static+64)[]
the most probable is a corruption of the state.
If you didn't already fix the problem what about posting
cvs diff -u ?

regards,
Philippe Elie