Re: [PATCH 0/7] preempt_count rework -v2

["H. Peter Anvin" <hpa@zytor.com>]

On 09/11/2013 06:13 AM, Peter Zijlstra wrote:
> On Tue, Sep 10, 2013 at 02:43:06PM -0700, Linus Torvalds wrote:
>> That said, looking at your patch, I get the *very* strong feeling that
>> we could make a macro that does all the repetitions for us, and then
>> have a
>>
>>   GENERATE_RMW(atomic_sub_and_test, LOCK_PREFIX "subl", "e", "")
> 
> The below seems to compile..
> 
> +
> +#define GENERATE_ADDcc(var, val, lock, cc)				\
> +do {									\
> +	const int add_ID__ = (__builtin_constant_p(val) &&		\
> +			((val) == 1 || (val) == -1)) ? (val) : 0;	\
> +									\
> +	switch (sizeof(var)) {						\
> +	case 4:								\
> +		if (add_ID__ == 1) {					\
> +			asm volatile goto(lock "incl %0;"		\
> +					  "j" cc " %l[cc_label]"	\
> +					  : : "m" (var)			\
> +					  : "memory" : cc_label);	\
> +		} else if (add_ID__ == -1) {				\
> +			asm volatile goto(lock "decl %0;"		\
> +					  "j" cc " %l[cc_label]"	\
> +					  : : "m" (var)			\
> +					  : "memory" : cc_label);	\
> +		} else {						\
> +			asm volatile goto(lock "addl %1, %0;"		\
> +					  "j" cc " %l[cc_label]"	\
> +					  : : "m" (var), "er" (val)	\
> +					  : "memory" : cc_label);	\
> +		}							\
> +		break;							\
> +									\
> +	case 8:								\
> +		if (add_ID__ == 1) {					\
> +			asm volatile goto(lock "incq %0;"		\
> +					  "j" cc " %l[cc_label]"	\
> +					  : : "m" (var)			\
> +					  : "memory" : cc_label);	\
> +		} else if (add_ID__ == -1) {				\
> +			asm volatile goto(lock "decq %0;"		\
> +					  "j" cc " %l[cc_label]"	\
> +					  : : "m" (var)			\
> +					  : "memory" : cc_label);	\
> +		} else {						\
> +			asm volatile goto(lock "addq %1, %0;"		\
> +					  "j" cc " %l[cc_label]"	\
> +					  : : "m" (var), "er" (val)	\
> +					  : "memory" : cc_label);	\
> +		}							\
> +		break;							\
> +									\

At least in the "asm goto" case you can use:

	lock "add%z0 %1,%0;"

... and skip the switch statement.

There was a bug in some old (gcc 3.x?) early x86-64 versions which would
treat %z0 as if it was %Z0 which means it would emit "ll" instead of "q"
but that doesn't apply to any gcc which has "asm goto"...

	-hpa