On 2015-06-09 22:07, Sebastian Andrzej Siewior wrote:

On 2015-06-03 17:05:16 [+0200], Stefan Agner wrote:

quoted

As Boris guessed correctly, the reason I used the raw variant was due to
performance improvements due to the barrier. However, I will use

yeah, do you have any numbers by chance?

Not current ones for Linux. I added this in the U-Boot variant of the
driver.
 

quoted

{read|write}l_relaxed instead, which should offer endian abstraction
while not having the performance penalty due to extensive barriers...

well, even those
|$ git grep readl_relaxed arch/powerpc/
|arch/powerpc/include/asm/io.h:#define readl_relaxed(addr) readl(addr)

have the endian swap. So an abstraction like you provided earlier
would be nice :)

What do you mean by that?

Btw, I sent v5 of the patchset which use the relaxed variants. As
expected, performance did not suffer by that change on ARM.

--
Stefan