| Message ID | 1380576057.18603.465.camel@ted |
|---|---|
| State | New |
| Headers | show |
On Mon, 2013-09-30 at 22:20 +0100, Richard Purdie wrote: > This is a combination of bug report and patch. I'm not sure if you'll want to fix it > like this but it illustrates the problem and should be easy to fix based on this. > > When we restore the mxcsr register with FXRSTOR, we need to update the various SSE > status flags in CPUX86State by calling update_sse_status(). If we don't, we end up > using the status bits from some other context with interesting results. > > I used a function prototype since it makes the fix clear, some code rearrangement > might be needed ultimately. > > Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org> Ping? This seems to be quite nasty SSE register corruption and it would be nice to get it fixed. Happy to rework the patch if needed, am just not sure the best way to handle moving the function if that is what is needed. Cheers, Richard > Index: qemu-1.5.0/target-i386/fpu_helper.c > =================================================================== > --- qemu-1.5.0.orig/target-i386/fpu_helper.c 2013-09-30 18:46:39.283377648 +0000 > +++ qemu-1.5.0/target-i386/fpu_helper.c 2013-09-30 18:46:56.895377232 +0000 > @@ -1149,6 +1149,8 @@ > } > } > > +static void update_sse_status(CPUX86State *env); > + > void helper_fxrstor(CPUX86State *env, target_ulong ptr, int data64) > { > int i, fpus, fptag, nb_xmm_regs; > @@ -1180,6 +1182,7 @@ > if (env->cr[4] & CR4_OSFXSR_MASK) { > /* XXX: finish it */ > env->mxcsr = cpu_ldl_data(env, ptr + 0x18); > + update_sse_status(env); > /* cpu_ldl_data(env, ptr + 0x1c); */ > if (env->hflags & HF_CS64_MASK) { > nb_xmm_regs = 16; > >
Index: qemu-1.5.0/target-i386/fpu_helper.c =================================================================== --- qemu-1.5.0.orig/target-i386/fpu_helper.c 2013-09-30 18:46:39.283377648 +0000 +++ qemu-1.5.0/target-i386/fpu_helper.c 2013-09-30 18:46:56.895377232 +0000 @@ -1149,6 +1149,8 @@ } } +static void update_sse_status(CPUX86State *env); + void helper_fxrstor(CPUX86State *env, target_ulong ptr, int data64) { int i, fpus, fptag, nb_xmm_regs; @@ -1180,6 +1182,7 @@ if (env->cr[4] & CR4_OSFXSR_MASK) { /* XXX: finish it */ env->mxcsr = cpu_ldl_data(env, ptr + 0x18); + update_sse_status(env); /* cpu_ldl_data(env, ptr + 0x1c); */ if (env->hflags & HF_CS64_MASK) { nb_xmm_regs = 16;
This is a combination of bug report and patch. I'm not sure if you'll want to fix it like this but it illustrates the problem and should be easy to fix based on this. When we restore the mxcsr register with FXRSTOR, we need to update the various SSE status flags in CPUX86State by calling update_sse_status(). If we don't, we end up using the status bits from some other context with interesting results. I used a function prototype since it makes the fix clear, some code rearrangement might be needed ultimately. Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>