Re: [PATCH RFC V2 3/5] jump_label: if a key has already been initialized, don't nop it out

[Richard Henderson <rth@redhat.com>]

On 10/06/2011 11:26 AM, Steven Rostedt wrote:
> On Thu, 2011-10-06 at 14:10 -0400, Jason Baron wrote:
> 
>>> Looks like jmp2 is about 5% faster than jmp5 on Sandybridge with this
>>> benchmark.
>>>
>>> But insignificant difference on Nehalem.
>>>
>>>     J
>>
>> It would be cool if we could make the total width 2-bytes, when
>> possible.  It might be possible by making the initial 'JUMP_LABEL_INITIAL_NOP'
>> as a 'jmp' to the 'l_yes' label. And then patching that with a no-op at boot
>> time or link time - letting the compiler pick the width. In that way we could
>> get the optimal width...
> 
> Why not just do it?
> 
> jump_label is encapsulated in arch_static_branch() which on x86 looks
> like:
> 
> static __always_inline bool arch_static_branch(struct jump_label_key *key)
> {
> 	asm goto("1:"
> 		JUMP_LABEL_INITIAL_NOP
> 		".pushsection __jump_table,  \"aw\" \n\t"
> 		_ASM_ALIGN "\n\t"
> 		_ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> 		".popsection \n\t"
> 		: :  "i" (key) : : l_yes);
> 	return false;
> l_yes:
> 	return true;
> }
> 
> 
> That jmp to l_yes should easily be a two byte jump.

Until the compiler decides to re-order the code.  That's the problem --
in the general case you do not know how far away the destination is really
going to be.

There are a couple of possibilities for improvement:

(1) Do as Jason suggests above and let the assembler figure out the size
of the branch that is needed.  Without adding more data to __jump_table,
you'll want to be extremely careful about checking the two pointers to 
see what size branch has been installed.

(2) Always reserve 5 bytes of space, but if the distance is small enough
patch in a 2-byte jump.  That doesn't help with the icache footprint.

(3) There is no 3.  I was going to say something clever about gas .ifne
conditionals, but a quick test revealed they don't work for forward
declarations.



r~