Re: [PATCH] bcm: address clang inline asm incompatibility

[Ard Biesheuvel <ard.biesheuvel@linaro.org>]

On 28 January 2015 at 17:08, Alex Elder <elder@linaro.org> wrote:
> On 01/28/2015 10:17 AM, Ard Biesheuvel wrote:
>> On 28 January 2015 at 14:11, Alex Elder <elder@linaro.org> wrote:
>>> On 01/28/2015 05:15 AM, Ard Biesheuvel wrote:
>>>> On 28 January 2015 at 05:18, Behan Webster <behanw@converseincode.com> wrote:
>>>>> From: Alex Elder <elder@linaro.org>
>>>>>
>>>>> My GCC-based build environment likes to call register r12 by the
>>>>> name "ip" in inline asm.  Behan Webster informed me that his Clang-
>>>>> based build environment likes "r12" instead.
>>>>>
>>>>> Try to make them both happy.
>>>>>
>>>>> Signed-off-by: Alex Elder <elder@linaro.org>
>>>>> Signed-off-by: Behan Webster <behanw@converseincode.com>
>>>>> ---
>>>>>  arch/arm/mach-bcm/bcm_kona_smc.c | 9 +++++++--
>>>>>  1 file changed, 7 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/arch/arm/mach-bcm/bcm_kona_smc.c b/arch/arm/mach-bcm/bcm_kona_smc.c
>>>>> index a55a7ec..3937bd5 100644
>>>>> --- a/arch/arm/mach-bcm/bcm_kona_smc.c
>>>>> +++ b/arch/arm/mach-bcm/bcm_kona_smc.c
>>>>> @@ -106,9 +106,14 @@ int __init bcm_kona_smc_init(void)
>>>>>   * request result appropriately.  This result value is found in r0
>>>>>   * when the "smc" request completes.
>>>>>   */
>>>>> +#ifdef __clang__
>>>>> +#define R12    "r12"
>>>>> +#else  /* !__clang__ */
>>>>> +#define R12    "ip"    /* gcc calls r12 "ip" */
>>>>> +#endif /* !__clang__ */
>>>>
>>>> Why not just use r12 for both?
>>>
>>> Yes, that would have been an obvious fix.  But the
>>> assembler (in the GCC environment) doesn't accept that.
>>>
>>
>> Mine has no problems with it at all
>>
>> $ echo 'mov r12, #0' | arm-linux-gnueabihf-gcc -c -x assembler-with-cpp -
>>
>> and grepping for r12 under arch/arm suggests the same
>
> The use of "r12" is fine.  But it's not just the assembler,
> I believe it also involves gcc.
>
> The problem is with the use of the __asmeq(x, y) macro.
>

Ah right. Apologies for assuming that you had missed something obvious here.
But __asmeq is not the toolchain, it is a local construct #define'd in
compiler.h

> If I assign the "ip" variable with "r12":
>         register u32 ip asm("r12");     /* Also called ip */
>
> Then that's fine.  However, this line then causes an error:
>                 __asmeq("%0", "r12")
>
> Apparently gcc uses register "ip" when it sees asm("r12").  So
> attempting to verify the desired register got used with __asmeq()
> causes a string mismatch--"ip" is not equal to "r12".
>
> So I could use:
>
>         register u32 ip asm("r12");     /* Also called ip */
>                 ...
>                 __asmeq("%0", "ip")
>
> And that will build.  But it's a little non-intuitive, and
> I suspect that clang might (rightfully) have a failure in
> this __asmeq() call.
>

In that case, I would strongly suggest fixing the __asmeq () macro
instead, and teach it that ("r12","ip") and ("ip","r12") are fine too.

The thing is, inline asm is a dodgy area to begin with in terms of
clang-to-gcc compatibility. On arm64, we have been seeing issues where
the width of the register -which is fixed on gcc- is selected based on
the size of that variable, i.e., an int32 gets a w# register and int64
gets a x# register. Imagine debugging that, e.g., a str %0, [xx] that
writes 8 bytes on GCC suddenly only writing 4 bytes when built with
clang.

If we also start using the preprocessor to conditionalise what is
emitted by inline asm, the waters get even murkier and it becomes even
harder to claim parity between the two.

-- 
Ard.