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

[Behan Webster]

On 01/28/15 12:11, Ard Biesheuvel wrote:
> On 28 January 2015 at 19:38, Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
>> On 28 January 2015 at 19:27, Alex Elder <elder@xxxxxxxxxx> wrote:
>>> On 01/28/2015 01:17 PM, Ard Biesheuvel wrote:
>>>> On 28 January 2015 at 17:20, Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
>>>>> On 28 January 2015 at 17:08, Alex Elder <elder@xxxxxxxxxx> wrote:
>>>>>> On 01/28/2015 10:17 AM, Ard Biesheuvel wrote:
>>>>>>> On 28 January 2015 at 14:11, Alex Elder <elder@xxxxxxxxxx> wrote:
>>>>>>>> On 01/28/2015 05:15 AM, Ard Biesheuvel wrote:
>>>>>>>>> On 28 January 2015 at 05:18, Behan Webster <behanw@xxxxxxxxxxxxxxxxxx> wrote:
>>>>>>>>>> From: Alex Elder <elder@xxxxxxxxxx>
>>>>>>>>>>
>>>>>>>>>> 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@xxxxxxxxxx>
>>>>>>>>>> Signed-off-by: Behan Webster <behanw@xxxxxxxxxxxxxxxxxx>
>>>>>>>>>> ---
>>>>>>>>>>  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.
>>>>>
>>>> Something like this perhaps?
>>> So __asmeq() yields true if the register names (strings) are
>>> equal, or if one is "ip" and the other is "r12" (in either order).
>>>
>>> I can't comment on whether it's right in all build environments but
>>> this looks OK to me, to handle this special case.
>>>
>>> I would much rather you generate that patch.  Is that OK?
>>>
>> Sure, I can cook up a patch if you guys can confirm that it fixes your
>> use case. (I tested GCC myself but I don't have clang installed)
>>
> Actually, if clang is guaranteed to emit the correct register name
> inside the inline asm for register asm variables used in input or
> output constraints, I think it makes sense to #define __asmeq as a nop
> if __clang__ is defined. (Note that __asmeq only exists to work around
> a specific GCC bug)
As far as I'm aware neither clang nor gcc will guarantee this completely
in all places where asmeq has been used. Register assignments are
handled differently, and at different levels of the architecture between
the 2 compilers.

Certainly asmeq has caught these kinds of bad assumptions in a number of
places in the kernel while we've ported to clang (like when naked
functions are used, and the calling convention for -O2 is assumed).

I personally would prefer code which doesn't rely on variables being in
the correct registers. However I'm aware that in the case of the smc
instruction there really isn't a choice, as that's the way it works (it
assumes values in certain registers). :(

I think the code you've come up with is the best solution for now.

Behan

-- 
Behan Webster
behanw@xxxxxxxxxxxxxxxxxx


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/