Hi John,
On 5/3/2024 12:12 PM, John Hubbard wrote:
> On 5/3/24 11:37 AM, Reinette Chatre wrote:
>> On 5/3/2024 9:52 AM, John Hubbard wrote:
>>> On 5/3/24 1:00 AM, Ilpo Järvinen wrote:
>>>> On Thu, 2 May 2024, John Hubbard wrote:
>>> ...
>>>>> diff --git a/tools/testing/selftests/resctrl/mbm_test.c b/tools/testing/selftests/resctrl/mbm_test.c
>>>>> index d67ffa3ec63a..c873793d016d 100644
>>>>> --- a/tools/testing/selftests/resctrl/mbm_test.c
>>>>> +++ b/tools/testing/selftests/resctrl/mbm_test.c
>>>>> @@ -33,7 +33,7 @@ show_bw_info(unsigned long *bw_imc, unsigned long *bw_resc, size_t span)
>>>>> avg_bw_imc = sum_bw_imc / 4;
>>>>> avg_bw_resc = sum_bw_resc / 4;
>>>>> - avg_diff = (float)labs(avg_bw_resc - avg_bw_imc) / avg_bw_imc;
>>>>> + avg_diff = (float)(avg_bw_resc - avg_bw_imc) / avg_bw_imc;
>>>>> avg_diff_per = (int)(avg_diff * 100);
>>>>> ret = avg_diff_per > MAX_DIFF_PERCENT;
>>>>
>>>> But how are these two cases same after your change when you ended up
>>>> removing taking the absolute value entirely?
>>>
>>> All of the arguments are unsigned integers, so all arithmetic results
>>> are interpreted as unsigned, so taking the absolute value of that is
>>> always a no-op.
>>
>> It does not seem as though clang can see when values have been casted.
>> I tried to do so explicitly with a:
>> avg_diff = labs((long)avg_bw_resc - avg_bw_imc) / (float)avg_bw_imc;
>
> The subtraction result will get promoted to an unsigned long, before being
> passed into labs(3).
>
>>
>> But that still triggers:
>> warning: taking the absolute value of unsigned type 'unsigned long' has no effect [-Wabsolute-value]
>
> As expected, yes.
>
>>
>> Looks like we may need to be more explicit types and not rely on casting so much
>> to make the compiler happy.
>>
>
> I assumed that this code did not expect to handle negative numbers,
> because it is using unsigned math throughout.
>
> If you do expect it to handle cases where, for example, this happens:
>
> avg_bw_imc > avg_bw_resc
The existing code seems to handle this ok. A sample program with this
scenario comparing existing computation with your first proposal is
below:
#include <stdio.h>
#include <stdlib.h>
void main(void) {
unsigned long avg_bw_resc = 20000;
unsigned long avg_bw_imc = 40000;
float avg_diff;
/* Existing code */
avg_diff = (float)labs(avg_bw_resc - avg_bw_imc) / avg_bw_imc;
printf("Existing code: avg_diff = %f\n", avg_diff);
/* Original proposed fix */
avg_diff = (float)(avg_bw_resc - avg_bw_imc) / avg_bw_imc;
printf("Original proposed fix: avg_diff = %f\n", avg_diff);
}
output:
Existing code: avg_diff = 0.500000
Original proposed fix: avg_diff = 461168590192640.000000
>
> ...then a proper solution is easy, and looks like this:
>
> diff --git a/tools/testing/selftests/resctrl/mbm_test.c b/tools/testing/selftests/resctrl/mbm_test.c
> index c873793d016d..b87f91a41494 100644
> --- a/tools/testing/selftests/resctrl/mbm_test.c
> +++ b/tools/testing/selftests/resctrl/mbm_test.c
> @@ -17,8 +17,8 @@
> static int
> show_bw_info(unsigned long *bw_imc, unsigned long *bw_resc, size_t span)
> {
> - unsigned long avg_bw_imc = 0, avg_bw_resc = 0;
> - unsigned long sum_bw_imc = 0, sum_bw_resc = 0;
> + long avg_bw_imc = 0, avg_bw_resc = 0;
> + long sum_bw_imc = 0, sum_bw_resc = 0;
> int runs, ret, avg_diff_per;
> float avg_diff = 0;
>
> Should I resend the patch with that approach?
ok. That indeed makes the computations easier to understand. I assume
you intend to fix the snippet in mba_test.c also?
Reinette
