On Thu, 2015-11-05 at 15:33 -0500, David Miller wrote:

From: David Miller <davem@davemloft.net>
Date: Wed, 28 Oct 2015 18:44:15 -0700 (PDT)

quoted

From: David Woodhouse <dwmw2@infradead.org>
Date: Wed, 28 Oct 2015 23:53:52 +0900

quoted

I know MIPS will need some additional libgcc functionality brought
into the kernel, and Ralf has already been looking at that; not sure
about other platforms. Let's find out :)

It should properly build on Sparc, but I still need to do my due
diligence to see if gcc will actually emit the little-endian loads
and stores or not...

It does not, instead it emits libgcc calls.

So I don't I can use this on sparc.

....

quoted

I wonder if we could provide a generic weak version in the lib/
directory which just uses the existing __arch_swab32 et al? Does GCC
use a standard function call ABI for __bswap[sd]i2 or can it be
'special'?

That's undesirable compared to inlining the asm.

So... GCC knows *exactly* what this builtin function does.

It is better to inline it than to call an out-of-line function.

Yet GCC emits an out-of-line call instead of just inlining it.

Is that not something we should be filing a GCC PR for?

Basically, we are working around a GCC misbehaviour by using opaque
inline asm instead of the GCC builtins that *currently* happen to
result in suboptimal code.

We've mostly learned our lesson *not* to do that kind of thing :)

-- 
dwmw2