[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH 9/9] x86/bitops: Use the POPCNT instruction when available


  • To: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
  • From: Jan Beulich <jbeulich@xxxxxxxx>
  • Date: Tue, 27 Aug 2024 17:04:54 +0200
  • Autocrypt: addr=jbeulich@xxxxxxxx; keydata= xsDiBFk3nEQRBADAEaSw6zC/EJkiwGPXbWtPxl2xCdSoeepS07jW8UgcHNurfHvUzogEq5xk hu507c3BarVjyWCJOylMNR98Yd8VqD9UfmX0Hb8/BrA+Hl6/DB/eqGptrf4BSRwcZQM32aZK 7Pj2XbGWIUrZrd70x1eAP9QE3P79Y2oLrsCgbZJfEwCgvz9JjGmQqQkRiTVzlZVCJYcyGGsD /0tbFCzD2h20ahe8rC1gbb3K3qk+LpBtvjBu1RY9drYk0NymiGbJWZgab6t1jM7sk2vuf0Py O9Hf9XBmK0uE9IgMaiCpc32XV9oASz6UJebwkX+zF2jG5I1BfnO9g7KlotcA/v5ClMjgo6Gl MDY4HxoSRu3i1cqqSDtVlt+AOVBJBACrZcnHAUSuCXBPy0jOlBhxPqRWv6ND4c9PH1xjQ3NP nxJuMBS8rnNg22uyfAgmBKNLpLgAGVRMZGaGoJObGf72s6TeIqKJo/LtggAS9qAUiuKVnygo 3wjfkS9A3DRO+SpU7JqWdsveeIQyeyEJ/8PTowmSQLakF+3fote9ybzd880fSmFuIEJldWxp Y2ggPGpiZXVsaWNoQHN1c2UuY29tPsJgBBMRAgAgBQJZN5xEAhsDBgsJCAcDAgQVAggDBBYC AwECHgECF4AACgkQoDSui/t3IH4J+wCfQ5jHdEjCRHj23O/5ttg9r9OIruwAn3103WUITZee e7Sbg12UgcQ5lv7SzsFNBFk3nEQQCACCuTjCjFOUdi5Nm244F+78kLghRcin/awv+IrTcIWF hUpSs1Y91iQQ7KItirz5uwCPlwejSJDQJLIS+QtJHaXDXeV6NI0Uef1hP20+y8qydDiVkv6l IreXjTb7DvksRgJNvCkWtYnlS3mYvQ9NzS9PhyALWbXnH6sIJd2O9lKS1Mrfq+y0IXCP10eS FFGg+Av3IQeFatkJAyju0PPthyTqxSI4lZYuJVPknzgaeuJv/2NccrPvmeDg6Coe7ZIeQ8Yj t0ARxu2xytAkkLCel1Lz1WLmwLstV30g80nkgZf/wr+/BXJW/oIvRlonUkxv+IbBM3dX2OV8 AmRv1ySWPTP7AAMFB/9PQK/VtlNUJvg8GXj9ootzrteGfVZVVT4XBJkfwBcpC/XcPzldjv+3 HYudvpdNK3lLujXeA5fLOH+Z/G9WBc5pFVSMocI71I8bT8lIAzreg0WvkWg5V2WZsUMlnDL9 mpwIGFhlbM3gfDMs7MPMu8YQRFVdUvtSpaAs8OFfGQ0ia3LGZcjA6Ik2+xcqscEJzNH+qh8V m5jjp28yZgaqTaRbg3M/+MTbMpicpZuqF4rnB0AQD12/3BNWDR6bmh+EkYSMcEIpQmBM51qM EKYTQGybRCjpnKHGOxG0rfFY1085mBDZCH5Kx0cl0HVJuQKC+dV2ZY5AqjcKwAxpE75MLFkr wkkEGBECAAkFAlk3nEQCGwwACgkQoDSui/t3IH7nnwCfcJWUDUFKdCsBH/E5d+0ZnMQi+G0A nAuWpQkjM1ASeQwSHEeAWPgskBQL
  • Cc: Roger Pau Monné <roger.pau@xxxxxxxxxx>, Xen-devel <xen-devel@xxxxxxxxxxxxxxxxxxxx>
  • Delivery-date: Tue, 27 Aug 2024 15:05:02 +0000
  • List-id: Xen developer discussion <xen-devel.lists.xenproject.org>

On 27.08.2024 16:59, Andrew Cooper wrote:
> On 27/08/2024 1:47 pm, Jan Beulich wrote:
>> On 27.08.2024 13:17, Andrew Cooper wrote:
>>> On 26/08/2024 2:07 pm, Jan Beulich wrote:
>>>> On 23.08.2024 01:06, Andrew Cooper wrote:
>>>>> @@ -475,4 +476,24 @@ static always_inline unsigned int arch_flsl(unsigned 
>>>>> long x)
>>>>>  }
>>>>>  #define arch_flsl arch_flsl
>>>>>  
>>>>> +static always_inline unsigned int arch_hweightl(unsigned long x)
>>>>> +{
>>>>> +    unsigned int r;
>>>>> +
>>>>> +    /*
>>>>> +     * arch_generic_hweightl() is written in ASM in order to preserve all
>>>>> +     * registers, as the compiler can't see the call.
>>>>> +     *
>>>>> +     * This limits the POPCNT instruction to using the same ABI as a 
>>>>> function
>>>>> +     * call (input in %rdi, output in %eax) but that's fine.
>>>>> +     */
>>>>> +    alternative_io("call arch_generic_hweightl",
>>>>> +                   "popcnt %[val], %q[res]", X86_FEATURE_POPCNT,
>>>>> +                   ASM_OUTPUT2([res] "=a" (r) ASM_CALL_CONSTRAINT),
>>>>> +                   [val] "D" (x));
>>>> If you made [val] an output ("+D") you could avoid preserving the register
>>>> in the function. And I'd expect the register wouldn't be re-used often
>>>> afterwards, so its clobbering likely won't harm code quality very much.
>>> "+D" means it's modified by the asm, which forces preservation of the
>>> register, if it's still needed afterwards.
>>>
>>> Plain "D" means not modified by the asm, which means it can be reused if
>>> necessary.
>> And we likely would prefer the former: If the register's value isn't
>> use afterwards (and that's the case as far as the function on its own
>> goes), the compiler will know it doesn't need to preserve anything.
>> That way, rather than always preserving (in the called function)
>> preservation will be limited to just the (likely few) cases where the
>> value actually is still needed afterwards.
> 
> Constraints are there to describe how the asm() behaves to the compiler.
> 
> You appear to be asking me to put in explicitly-incorrect constraints
> because you think it will game the optimiser.
> 
> Except the reasoning is backwards.  The only thing forcing "+D" will do
> is cause the compiler to preserve the value elsewhere if it's actually
> needed later, despite the contents of %rdi still being good for the purpose.
> 
> In other words, using "+D" for asm which is really only "D" (as this one
> is) will result in strictly worse code generation in the corner case you
> seem to be worried about.

Well, then leave it as is. (An extra benefit would have been that
arch_generic_hweightl() then would ended up with a odd-number-of-slots
stack frame. Not that this matters much, but having ABI-compliant
functions where possible seems always preferable, when possible.)

Jan



 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.