Re: [PATCH] arm64: prefetch: Change assembly to be compatible with gcc and clang

From: Matthias Kaehlcke
Date: Thu Apr 27 2017 - 16:35:26 EST


Hi,

Thanks for your comments!

El Mon, Apr 24, 2017 at 02:34:47PM +0100 Will Deacon ha dit:

> On Thu, Apr 20, 2017 at 09:42:07AM +0100, Mark Rutland wrote:
> > On Wed, Apr 19, 2017 at 02:22:11PM -0700, Matthias Kaehlcke wrote:
> > > clang fails to build with the current code:
> > >
> > > arch/arm64/include/asm/processor.h:172:15: error: invalid operand in
> > > inline asm: 'prfm pldl1keep, ${0:a}'
> > >
> > > Apparently clang does not support the 'a' modifier. Change the
> > > constraint from 'p' ('An operand that is a valid memory address is
> > > allowed') to 'Q' ('A memory address which uses a single base register
> > > with no offset'), which works for both gcc and clang.
> >
> > It looks like the current %a0 template and p constraint were inherited
> > from arch/arm, as they've been there from day one on arm64.
> >
> > Looking at the arch/arm history, the "a" operand modifier and "p"
> > constraint were introduced in commit:
> >
> > 16f719de62809e22 ("[ARM] 5196/1: fix inline asm constraints for preload")
> >
> > ... so as to avoid GCC assuming prefetch of a pointer implied it was not
> > NULL. Until that point, we'd used no operand modifier and "o"
> > constraint.
> >
> > It's not clear to me whether "o", "p", and "Q" constraints differ in
> > this regard on AArch64, or if the issue regarding NULL is still
> > relevant. The GCC docs say the "p" constraint is used for "a valid
> > memory address", which does sound like it shouldn't be NULL.
> >
> > Otherwise, this does look consistent with what we do elsewhere.
>
> I really don't like using 'Q' here, for two reasons:
>
> 1. It means we likely allocate a register where we don't need to, because
> we're going to need to use [Xn] as the addressing mode, which means
> adding any immediate offsets.
>
> 2. As you mention, 16f719de62809e22 says that GCC will use this as an
> indication that the address is non-NULL.
>
> We also can't just remove the 'a', because that will result in assembly
> failures. I haven't dug into exactly why, but I suspect it's because "p"
> can generate a label, which then won't assemble if surrounded by '[]'.

I was suggested to use the intrinsic __pldx() from ACLE, however it
isn't supported by gcc and there seem to be no other uses of ACLE in
the kernel. For clang it should be possible to use its
__builtin_arm_prefetch(), however I suspect that the compiler
dependency wouldn't be well received.

I guess another option is to get clang to support the 'a' modifier,
though it seems to be another modifier that is undocumented (I only
saw it mentioned in some ancient gcc documentation).

Cheers

Matthias