> 在 2016年12月23日，00:18，Richard Sandiford <rdsandif...@googlemail.com> 写道：
>
> Yunqiang Su <yunqiang...@imgtec.com> writes:
>>> 在 2016年12月22日，23:48，Yunqiang Su <yunqiang...@imgtec.com> 写道：
>>>
>>>>
>>>> 在 2016年12月22日，23:31，Richard Sandiford
>>>> <rdsandif...@googlemail.com> 写道：
>>>>
>>>> Matthew Fortune <matthew.fort...@imgtec.com> writes:
>>>>> Sandra Loosemore <san...@codesourcery.com> writes:
>>>>>> On 12/21/2016 11:54 AM, Yunqiang Su wrote:
>>>>>>> By this patch, I add a build-time option ` --with-unfused-madd4=yes/no',
>>>>>>> and runtime option -m(no-)unfused-madd4,
>>>>>>> to disable generate madd.fmt instructions.
>>>>>>
>>>>>> Your patch also needs a documentation change so that the new
>>>>>> command-line option is listed in the GCC manual with other MIPS target
>>>>>> options.
>>>>>
>>>>> Any opinions on option names to control this? Is it best to target
>>>>> the specific
>>>>> feature that is non-compliant on loongson or apply a general
>>>>> -mfix-loongson
>>>>> type option?
>>>>>
>>>>> I'm not sure I have a strong opinion either way but there do seem to be
>>>>> multiple possible variants.
>>>>
>>>> Wasn't sure from this thread whether Loongson simply had a fused
>>>> implementation (without intermediate rounding) or whether the
>>>> instructions gave numerically incorrect results for some inputs.
>>>
>>> I test to define ISA_HAS_FUSED_MADD4 true and
>>> define ISA_HAS_UNFUSED_MADD4 false, and try to build a test case.
>>> With ISA_HAS_FUSED_MADD4, the result is about 1e-17,
>>> and with ISA_HAS_UNFUSED_MADD4, the result is about 1e-17,
>>> both of the are incorrect (the expect value is 0).
>>>
>>> The test case is
>>>
>>> #include <stdio.h>
>>>
>>> double a = 0.6;
>>> double b = 0.4;
>>> double c = 0.6;
>>> double d = 0.4;
>>>
>>> int main(void)
>>> {
>>> double x = a * b - c * d;
>>> printf("%le\n", x);
>>> return 0;
>>> }
>>>
>>>
>>>> It sounds from a later thread like it's generating incorrect results,
>>>> is that right? If so, then FWIW I agree an -mfix option would be more
>>>> consistent. E.g. one of the -mfix-vr4120 errata was an incorrect
>>>> integer division result and one of the -mfix-sb1 errata was an incorrect
>>>> single-precision float division result. The latter case could have been
>>>> handled by an option to disable DIV.S and DIV.PS, but the -mfix option
>>>> gave more control.
>>>>
>>>> If instead the problem is that the instructions are fused then that's
>>>> also what the original MIPS 4 parts did, so maybe an option to control
>>>> fusedness would make sense.
>>>
>>> The result to thread it fused or unfused, is different, while neither of
>>> them
>>> is correct.
>>
>> ohh, the result are same, and neither is correct.
>> both of them are 1.332268e-17.
>
> That's the expected result for an implementation in which the subtraction
> is fused with the first multiplication without intermediate rounding.
> The second 0.4 * 0.6 isn't exactly representable and rounds down, to a
> value slightly less than 0.24. Then the fused operation subtracts this
> value from the exact result of the first 0.4 * 0.6 (0.24), giving a
> value slightly greater than 0.
I see. But I think 1.332268e-17 is too big than expected.
1-e17 for double is totally unacceptable.
>
> I see Paul Hua's patch does add Loongson to the list of targets
> with a fused implementation (should have checked earlier, sorry).
> So I think after that patch we would do the right thing. (In particular,
> -ffp-contract=off would then disable the fusing.)
will -ffp-contract=off disable some other optimization?
If so, I don’t think that will be an ideal choice for distributions, like
Debian.
>
> Thanks,
> Richard
