Re: [Xenomai] [Xenomai-git] Jan Kiszka : switchtest: Add SSE and AVX check

[Jan Kiszka <jan.kiszka@web.de>]

On 2013-04-16 00:02, Jan Kiszka wrote:
> On 2013-04-13 18:11, Gilles Chanteperdrix wrote:
>> On 02/04/2013 07:57 PM, GIT version control wrote:
>>
>>> Module: xenomai-2.6
>>> Branch: master
>>> Commit: 192597326a0becd1980cb6c5cc9395af18a19c60
>>> URL:    http://git.xenomai.org/?p=xenomai-2.6.git;a=commit;h=192597326a0becd1980cb6c5cc9395af18a19c60
>>>
>>> Author: Jan Kiszka <jan.kiszka@siemens.com>
>>> Date:   Tue Jan 29 18:46:13 2013 +0100
>>>
>>> switchtest: Add SSE and AVX check
>>>
>>> Add a test for switching the lower SSE registers xmm0..7 or AVX
>>> registers ymm0..7, provided the CPU supports the corresponding
>>> feature. As xmm and ymm share their storage, we only need to check
>>> one of the features.
>>>
>>> Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
>>>
>>> ---
>>
>>
>>>  static inline unsigned fp_regs_check(unsigned val)
>>>  {
>>>  	unsigned i, result = val;
>>> +	uint64_t vec[8][4];
>>>  	unsigned e[8];
>>>  
>>>  	for (i = 0; i < 8; i++)
>>>  		__asm__ __volatile__("fistpl %0":"=m"(e[7 - i]));
>>> +	if (fp_features & FP_FEATURE_AVX) {
>>> +		__asm__ __volatile__(
>>> +			"vmovupd %%ymm0,%0;"
>>> +			"vmovupd %%ymm1,%1;"
>>> +			"vmovupd %%ymm2,%2;"
>>> +			"vmovupd %%ymm3,%3;"
>>> +			"vmovupd %%ymm4,%4;"
>>> +			"vmovupd %%ymm5,%5;"
>>> +			"vmovupd %%ymm6,%6;"
>>> +			"vmovupd %%ymm7,%7;"
>>> +			:
>>> +			: "m" (vec[0][0]), "m" (vec[1][0]),
>>> +			  "m" (vec[2][0]), "m" (vec[3][0]),
>>> +			  "m" (vec[4][0]), "m" (vec[5][0]),
>>> +			  "m" (vec[6][0]), "m" (vec[7][0]));
>>> +	} else if (fp_features & FP_FEATURE_SSE) {
>>> +		__asm__ __volatile__(
>>> +			"movupd %%xmm0,%0;"
>>> +			"movupd %%xmm1,%1;"
>>> +			"movupd %%xmm2,%2;"
>>> +			"movupd %%xmm3,%3;"
>>> +			"movupd %%xmm4,%4;"
>>> +			"movupd %%xmm5,%5;"
>>> +			"movupd %%xmm6,%6;"
>>> +			"movupd %%xmm7,%7;"
>>> +			:
>>> +			: "m" (vec[0][0]), "m" (vec[1][0]),
>>> +			  "m" (vec[2][0]), "m" (vec[3][0]),
>>> +			  "m" (vec[4][0]), "m" (vec[5][0]),
>>> +			  "m" (vec[6][0]), "m" (vec[7][0]));
>>> +	}
>>>  
>>>  	for (i = 0; i < 8; i++)
>>>  		if (e[i] != val) {
>>> @@ -65,8 +148,33 @@ static inline unsigned fp_regs_check(unsigned val)
>>>  			result = e[i];
>>>  		}
>>>  
>>> +	if (fp_features & FP_FEATURE_AVX) {
>>> +		for (i = 0; i < 8; i++) {
>>> +			int error = 0;
>>> +			if (vec[i][0] != val) {
>>> +				result = vec[i][0];
>>> +				error = 1;
>>> +			}
>>> +			if (vec[i][2] != val) {
>>> +				result = vec[i][2];
>>> +				error = 1;
>>> +			}
>>> +			if (error)
>>> +				printk("ymm%d: %llu/%llu != %u/%u\n",
>>> +				       i, (unsigned long long)vec[i][0],
>>> +				       (unsigned long long)vec[i][2],
>>> +				       val, val);
>>> +		}
>>> +	} else if (fp_features & FP_FEATURE_SSE) {
>>> +		for (i = 0; i < 8; i++)
>>> +			if (vec[i][0] != val) {
>>> +				printk("xmm%d: %llu != %u\n",
>>> +				       i, (unsigned long long)vec[i][0], val);
>>> +				result = vec[i][0];
>>> +			}
>>> +	}
>>> +
>>>  	return result;
>>>  }
>>
>>
>> This routine causes a warning from gcc and looks indeed wrong: if the
>> "vec" variable is used as an output variable of the inline assembly, it
>> should be in the output section of the inline assembly, not the input
>> section.
>>
> 
> Yes, seems wrong. Will try to look into it the next days.

Done, you can find the obvious fix in my for-upstream queue.

Jan


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 263 bytes
Desc: OpenPGP digital signature
URL: <http://www.xenomai.org/pipermail/xenomai/attachments/20130416/7ad3152c/attachment.pgp>