On 10/04/19 16:57, Sean Christopherson wrote:
> On Wed, Apr 10, 2019 at 12:55:53PM +0000, David Laight wrote:
>> From: Paolo Bonzini
>>> Sent: 10 April 2019 10:55
>>>
>>> This check will soon be done on every nested vmentry and vmexit,
>>> "parallelize" it using bitwise operations.
>>>
>>> Signed-off-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
>>> ---
>> ...
>>> diff --git a/arch/x86/kvm/x86.h b/arch/x86/kvm/x86.h
>>> index 28406aa1136d..7bc7ac9d2a44 100644
>>> --- a/arch/x86/kvm/x86.h
>>> +++ b/arch/x86/kvm/x86.h
>>> @@ -347,4 +347,12 @@ static inline void kvm_after_interrupt(struct kvm_vcpu *vcpu)
>>> __this_cpu_write(current_vcpu, NULL);
>>> }
>>>
>>> +static inline bool kvm_pat_valid(u64 data)
>>> +{
>>> + if (data & 0xF8F8F8F8F8F8F8F8)
>>> + return false;
>>> + /* 0, 1, 4, 5, 6, 7 are valid values. */
>>> + return (data | ((data & 0x0202020202020202) << 1)) == data;
>>> +}
>>> +
>>
>> How about:
>> /*
>> * Each byte must be 0, 1, 4, 5, 6 or 7.
>> * Convert 001x to 011x then 100x so 2 and 3 fail the test.
>> */
>> data |= (data ^ 0x0404040404040404ULL)) + 0x0202020202020202ULL;
>> if (data & 0xF8F8F8F8F8F8F8F8ULL)
>> return false;
>
> Woah. My vote is for Paolo's version as the separate checks allow the
> reader to walk through step-by-step. The generated assembly isn't much
> different from a performance perspective since the TEST+JNE will be not
> taken in the fast path.
>
> Fancy:
> 0x000000000004844f <+255>: movabs $0xf8f8f8f8f8f8f8f8,%rcx
> 0x0000000000048459 <+265>: xor %eax,%eax
> 0x000000000004845b <+267>: test %rcx,%rdx
> 0x000000000004845e <+270>: jne 0x4848b <kvm_mtrr_valid+315>
> 0x0000000000048460 <+272>: movabs $0x202020202020202,%rax
> 0x000000000004846a <+282>: and %rdx,%rax
> 0x000000000004846d <+285>: add %rax,%rax
> 0x0000000000048470 <+288>: or %rdx,%rax
> 0x0000000000048473 <+291>: cmp %rdx,%rax
> 0x0000000000048476 <+294>: sete %al
> 0x0000000000048479 <+297>: retq
>
> Really fancy:
> 0x0000000000048447 <+247>: movabs $0x404040404040404,%rcx
> 0x0000000000048451 <+257>: movabs $0x202020202020202,%rax
> 0x000000000004845b <+267>: xor %rdx,%rcx
> 0x000000000004845e <+270>: add %rax,%rcx
> 0x0000000000048461 <+273>: movabs $0xf8f8f8f8f8f8f8f8,%rax
> 0x000000000004846b <+283>: or %rcx,%rdx
> 0x000000000004846e <+286>: test %rax,%rdx
> 0x0000000000048471 <+289>: sete %al
> 0x0000000000048474 <+292>: retq
Yeah, the three constants are expensive. Too bad the really fancy
version sums twos and xors fours; if it were the opposite, it could have
used lea and then I would have chosen that one just for the coolness factor.
(Quoting Avi, "mmu.c is designed around the fact that x86 has an
instruction to do "x = 12 + 9*y").
Paolo
