[PATCH] mtd: cfi: Deiline large functions

[Denys Vlasenko]

On 05/21/2015 10:36 AM, Brian Norris wrote:
> On Thu, May 21, 2015 at 09:50:38AM +0200, Denys Vlasenko wrote:
>>>> cfi_udelay(): 74 bytes, 26 callsites
>>>
>>> ^^ This is pretty dead-simple. If it's generating bad code, we might
>>> look at fixing it up instead. Almost all of its call sites are with
>>> constant input, so it *should* just become:
>>>
>>> 	udelay(1);
>>> 	cond_resched();
>>>
>>> in most cases. For the non-constant cases, we might still do an
>>> out-of-line implementation. Or maybe we just say it's all not worth it,
>>> and we just stick with what you have. But I'd like to consider
>>> alternatives to out-lining this one.
>>
>> You want to consider not-deinlining (IOW: speed-optimizing)
> 
> Inlining isn't always about speed.
> 
>> a *fixed time delay function*?
>>
>> Think about what delay functions do...
> 
> I wasn't really looking at speed. Just memory usage.

I don't follow.

A single, not-inlined cfi_udelay(1) call is
a minimal possible code size. Even

udelay(1);
cond_resched();

ought to be bigger.

> And I was only pointing this out because udelay() has a different
> implementation for the __builtin_constant_p() case. You can't take
> advantage of that for non-inlined versions of cfi_udelay().
> 
> But that may be irrelevant anyway, now that I think again. At best,
> you're trading one function call (arm_delay_ops.const_udelay() on ARM)
> for another (cfi_udelay()), since you can never completely optimize out
> the latter.

*delay() and *sleep() functions are special: they do NOT
want to be executed as fast as possible. They are *pausing*
execution. They are *intended* to be "slow".

You should not strive to optimize out function call overhead
when you call one of these. Otherwise, it would mean that you
essentially do this for e.g. udelay(NUM):

"I want to pause for NUM us, (which is about NUM*3000 CPU cycles),
let's optimize out call+ret so that we speed up execution
by 5 cycles".

Do you see why it does not make sense?