Re: More annoying code generation by clang

[H. Peter Anvin]

On April 8, 2024 1:49:34 AM PDT, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>On Thu, Apr 04, 2024 at 03:53:49PM -0700, Linus Torvalds wrote:
>
>> diff --git a/arch/x86/include/asm/barrier.h b/arch/x86/include/asm/barrier.h
>> index 66e57c010392..6159d2cbbfde 100644
>> --- a/arch/x86/include/asm/barrier.h
>> +++ b/arch/x86/include/asm/barrier.h
>> @@ -33,20 +33,15 @@
>>   * Returns:
>>   *     0 - (index < size)
>>   */
>> +#define array_index_mask_nospec(idx,sz) ({	\
>> +	typeof((idx)+(sz)) __idx = (idx);	\
>> +	typeof(__idx) __sz = (sz);		\
>> +	typeof(__idx) __mask;			\
>> +	asm volatile ("cmp %1,%2; sbb %0,%0"	\
>> +			:"=r" (__mask)		\
>> +			:"ir"(__sz),"r" (__idx)	\
>> +			:"cc");			\
>> +	__mask; })
>
>Should this not carry a comment about the "ir" constraint wanting to be
>"g" except for clang being daft?
>
>(I really wish clang would go fix this, it keeps coming up time and
>again).
>
>>  
>>  /* Prevent speculative execution past this barrier. */
>>  #define barrier_nospec() asm volatile("lfence":::"memory")
>

If the only reason for "ir" as opposed to "g" (= "irm") is clang then it really needs to be called out. Or better yet, don't do anything and let the clang people actually fix their code generation.