[PATCH v3 03/11] riscv: Implement cmpxchg8/16() using Zabha

[Alexandre Ghiti]

On 17/07/2024 17:29, Conor Dooley wrote:
> On Wed, Jul 17, 2024 at 10:26:34AM -0500, Andrew Jones wrote:
>> On Wed, Jul 17, 2024 at 08:19:49AM GMT, Alexandre Ghiti wrote:
>>> -#define __arch_cmpxchg_masked(sc_sfx, prepend, append, r, p, o, n)	\
>>> +#define __arch_cmpxchg_masked(sc_sfx, cas_sfx, prepend, append, r, p, o, n)	\
>>>   ({									\
>>> +	__label__ no_zabha_zacas, end;					\
>>> +									\
>>> +	if (IS_ENABLED(CONFIG_RISCV_ISA_ZABHA) &&			\
>>> +	    IS_ENABLED(CONFIG_RISCV_ISA_ZACAS)) {			\
>>> +		asm goto(ALTERNATIVE("j %[no_zabha_zacas]", "nop", 0,	\
>>> +				     RISCV_ISA_EXT_ZABHA, 1)		\
>>> +			 : : : : no_zabha_zacas);			\
>>> +		asm goto(ALTERNATIVE("j %[no_zabha_zacas]", "nop", 0,	\
>>> +				     RISCV_ISA_EXT_ZACAS, 1)		\
>>> +			 : : : : no_zabha_zacas);			\
>> I came late to the call, but I guess trying to get rid of these asm gotos
>> was the topic of the discussion. The proposal was to try and use static
>> branches, but keep in mind that we've had trouble with static branches
>> inside macros in the past when those macros are used in many places[1]
>>
>> [1] commit 0b1d60d6dd9e ("riscv: Fix build with CONFIG_CC_OPTIMIZE_FOR_SIZE=y")
> The other half of the suggestion was not using an asm goto, but instead
> trying to patch the whole thing in the alternative, for the problematic
> section with llvm < 17.


And I'm not a big fan of this solution since it would imply patching the 
5-7 instructions for LR/SC into nops which would probably slow (a bit) 
the amocas/amoswap sequence. I agree it should not be that big, but that 
it is just to fix an llvm issue, so not worth it to me!


>>> +									\
>>> +		__asm__ __volatile__ (					\
>>> +			prepend						\
>>> +			"	amocas" cas_sfx " %0, %z2, %1\n"	\
>>> +			append						\
>>> +			: "+&r" (r), "+A" (*(p))			\
>>> +			: "rJ" (n)					\
>>> +			: "memory");					\
>>> +		goto end;						\
>>> +	}								\
>>> +									\
>>> +no_zabha_zacas:;							\
>> unnecessary ;
>>
>>>   	u32 *__ptr32b = (u32 *)((ulong)(p) & ~0x3);			\
>>>   	ulong __s = ((ulong)(p) & (0x4 - sizeof(*p))) * BITS_PER_BYTE;	\
>>>   	ulong __mask = GENMASK(((sizeof(*p)) * BITS_PER_BYTE) - 1, 0)	\
>>> @@ -133,6 +155,8 @@
>>>   		: "memory");						\
>>>   									\
>>>   	r = (__typeof__(*(p)))((__retx & __mask) >> __s);		\
>>> +									\
>>> +end:;									\
>>>   })
>>>   
>>>   #define __arch_cmpxchg(lr_sfx, sc_cas_sfx, prepend, append, r, p, co, o, n)	\