Re: [PATCH] arm/arm64: smccc: Use xN for arm64 register constraints with clang

[Greg Hackmann <ghackmann@google.com>]

On 03/22/2018 03:44 PM, Matthias Kaehlcke wrote:
> El Thu, Mar 22, 2018 at 10:26:18PM +0000 Nick Desaulniers ha dit:
> 
>> Note that a patch in this form has previously been implemented by:
>>
>> Andrey Konovalov <andreyknvl@google.com>:
>> https://gist.github.com/xairy/ee11682ea86044a45c0291c528cd936f
>>
>> and another by:
>>
>> Greg Hackmann <ghackmann@google.com>:
>> https://android-review.googlesource.com/c/kernel/common/+/645181
>>
>> If you used either as a reference, you may want to credit them with a
>> `Suggested-by:` in the commit message.
> 
> Not really, but I think I prefer Greg's version over mine and might
> use it in a respin if nobody raises objections.

NAK.  There's a reason I didn't send my change upstream.

As Marc pointed out (https://lkml.org/lkml/2018/3/16/987), the "r"
prefix tells gcc to pick the appropriate register width.  "x" makes it
unconditionally use the entire 64-bit register width.  Just swapping out
one for the other changes the macro's semantics.

Unfortunately since this was breaking builds in android-4.14 and we
didn't have an immediate-term fix, I bit the bullet and added the above
commit -- but *only* as a short-term workaround.  For the one caller we
currently have in 4.14.y, gcc was using the entire 64-bit width for all
its inputs anyway, so "r" vs. "x" didn't make a difference.  But that
might not be true if/when someone introduces other SMCCC 1.1 callers.

Unfortunately I don't see a better way to deal with this than waiting
for clang to support "r"-style constraints on ARM64.