Re: [PATCH 2/3] x86-64: Unify x86_*_percpu() functions.

[Jeremy Fitzhardinge <jeremy@goop.org>]

Ingo Molnar wrote:
> * Jeremy Fitzhardinge <jeremy@goop.org> wrote:
>
>   
>> Brian Gerst wrote:
>>     
>>> Merge the 32-bit and 64-bit versions of these functions.  Unlike 32-bit,
>>> the segment base is the current cpu's PDA instead of the offset from the
>>> original per-cpu area.  This is because GCC hardcodes the stackprotector
>>> canary at %gs:40.  Since the assembler is incapable of relocating against
>>> multiple symbols, the code ends up looking like:
>>>
>>> 	movq $per_cpu__var, reg
>>> 	subq $per_cpu__pda, reg
>>> 	movq %gs:(reg), reg
>>>
>>> This is still atomic since the offset is a constant (just calculated at
>>> runtime) and not dependant on the cpu number.
>>>   
>>>       
>> Yeah, it's a real pity we can't convince the linker to do this simple 
>> computation as a single %gs:ADDR addressing mode.  On the other hand, if 
>> the compiler can reuse the computation of %reg 2-3 times, then the 
>> generated code could well end up being denser.
>>     
>
> There's a nice project for linker hackers?
>
> I'd like to see some kernel image size measurements done on x86 defconfig 
> to see how much real impact this has on code density. Unless the impact is 
> horribly unacceptable, removing ~200 lines of weird x86-specific APIs is a 
> definitive plus.

Yep, I'm all for it.  I don't think there'll be much of a size impact at 
all.

    J