Re: lib/iomap.c mmio_{in,out}s* vs. __raw_* accessors

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

> The _only_ reason to use "ioread32be()" would be because the machine is 
> actually natively BE, and you want to avoid swab. That's kind of the point 
> of using "be32_to_cpu(__raw_readl(addr)))" like we do now - it will do the 
> byte swap only if it's necessary.
> 
> In contrast, your "swab(readl())" does _two_ byteswaps - once to turn it 
> into LE, then to turn it back into BE.

I'm not doing a swab(readl()), I'm doing a swab(insl()) and have the
arch provide a native BE accessor for readl_be(). The idea is that I
don't want to add _be accessors for PIO and PIO is slow anyway. But I'm
providing one for MMIO, along with the repeat versions. Have a second
look. 

> So if we can't just rip it out, then we sure as hell shouldn't replace it 
> with something that is obviously worse either.
> 
> In other words - I don't see the reasoning here again. You seem to want to 
> make the code just worse. 

Wait, let's make thing clear, there are 2 things here:

 - MMIO : For that, I'm providing readw_be etc... which my patch defines
based on __raw_* just as your suggest, I'm just adding a way for the
arch to provide its own.

 - PIO : This is broken -now-. The current code doesn't swap at all in
the PIO case, thus you get LE result when using ioread32be() on PIO. I
propose to fix that with swab() because PIO sucks already, there is no
"__raw" for PIO and it doesn't deserve new accessors nor speed.

Cheers,
Ben.