Messages in this thread | | | Date | Tue, 04 Sep 2012 15:09:42 +0300 | From | Avi Kivity <> | Subject | Re: [PATCH 2/8] KVM: x86 emulator: use aligned variants of SSE register ops |
| |
On 08/30/2012 02:30 AM, Mathias Krause wrote: > As the the compiler ensures that the memory operand is always aligned > to a 16 byte memory location,
I'm not sure it does. Is V4SI aligned? Do we use alignof() to propagate the alignment to the vcpu allocation code?
> use the aligned variant of MOVDQ for > read_sse_reg() and write_sse_reg(). > > diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c > index 1451cff..5a0fee1 100644 > --- a/arch/x86/kvm/emulate.c > +++ b/arch/x86/kvm/emulate.c > @@ -909,23 +909,23 @@ static void read_sse_reg(struct x86_emulate_ctxt *ctxt, sse128_t *data, int reg) > { > ctxt->ops->get_fpu(ctxt); > switch (reg) { > - case 0: asm("movdqu %%xmm0, %0" : "=m"(*data)); break; > - case 1: asm("movdqu %%xmm1, %0" : "=m"(*data)); break; > - case 2: asm("movdqu %%xmm2, %0" : "=m"(*data)); break; > - case 3: asm("movdqu %%xmm3, %0" : "=m"(*data)); break; > - case 4: asm("movdqu %%xmm4, %0" : "=m"(*data)); break; > - case 5: asm("movdqu %%xmm5, %0" : "=m"(*data)); break; > - case 6: asm("movdqu %%xmm6, %0" : "=m"(*data)); break; > - case 7: asm("movdqu %%xmm7, %0" : "=m"(*data)); break; > + case 0: asm("movdqa %%xmm0, %0" : "=m"(*data)); break; > + case 1: asm("movdqa %%xmm1, %0" : "=m"(*data)); break; > + case 2: asm("movdqa %%xmm2, %0" : "=m"(*data)); break; > + case 3: asm("movdqa %%xmm3, %0" : "=m"(*data)); break; > + case 4: asm("movdqa %%xmm4, %0" : "=m"(*data)); break; > + case 5: asm("movdqa %%xmm5, %0" : "=m"(*data)); break; > + case 6: asm("movdqa %%xmm6, %0" : "=m"(*data)); break; > + case 7: asm("movdqa %%xmm7, %0" : "=m"(*data)); break; > #ifdef CONFIG_X86_64 > - case 8: asm("movdqu %%xmm8, %0" : "=m"(*data)); break; > - case 9: asm("movdqu %%xmm9, %0" : "=m"(*data)); break; > - case 10: asm("movdqu %%xmm10, %0" : "=m"(*data)); break; > - case 11: asm("movdqu %%xmm11, %0" : "=m"(*data)); break; > - case 12: asm("movdqu %%xmm12, %0" : "=m"(*data)); break; > - case 13: asm("movdqu %%xmm13, %0" : "=m"(*data)); break; > - case 14: asm("movdqu %%xmm14, %0" : "=m"(*data)); break; > - case 15: asm("movdqu %%xmm15, %0" : "=m"(*data)); break; > + case 8: asm("movdqa %%xmm8, %0" : "=m"(*data)); break; > + case 9: asm("movdqa %%xmm9, %0" : "=m"(*data)); break; > + case 10: asm("movdqa %%xmm10, %0" : "=m"(*data)); break; > + case 11: asm("movdqa %%xmm11, %0" : "=m"(*data)); break; > + case 12: asm("movdqa %%xmm12, %0" : "=m"(*data)); break; > + case 13: asm("movdqa %%xmm13, %0" : "=m"(*data)); break; > + case 14: asm("movdqa %%xmm14, %0" : "=m"(*data)); break; > + case 15: asm("movdqa %%xmm15, %0" : "=m"(*data)); break; > #endif > default: BUG();
The vmexit costs dominates any win here by several orders of magnitude.
-- error compiling committee.c: too many arguments to function
| |