[cfe-dev] [libcxx/test] Puzzling floating point behaviour

[Edward Meewis]

On 11-12-11 12:18, Edward Meewis wrote:
> On 10-12-11 19:20, Howard Hinnant wrote:
>> On Dec 10, 2011, at 2:04 AM, Edward Meewis wrote:
>>
>>> On 10-12-11 5:34, Eli Friedman wrote:
>>>> On Fri, Dec 9, 2011 at 8:23 PM, Edward Meewis
>>>> <ed at extraordinarymachine.nl>    wrote:
>>>>> On 10-12-11 1:34, Eli Friedman wrote:
>>>>>> On Fri, Dec 9, 2011 at 7:58 AM, Edward Meewis
>>>>>> <ed at extraordinarymachine.nl>      wrote:
>>>>>>> <snip>
>>>>>>>
>>>>>>>
>>>>>>> I'm guessing long doubles get demoted somewhere. Does anyone have an
>>>>>>> idea what the problem may be?
>>>>>> I'm not getting quite the same results as you, but your testcase isn't
>>>>>> really testing what you want to test: std::exp is not the same as
>>>>>> ::exp.
>>>>>>
>>>>>> -Eli
>>>>> Thanks, Eli, that's probably a big clue, but I don't get it. This is
>>>>> probably a stupid question, but can you explain? I would expect all exp's to
>>>>> be in std::. What's the difference?
>>>> std::exp is the properly overloaded exp defined by the C++ standard.
>>>> ::exp isn't overloaded, so it loses precision.
>>>>
>>>> -Eli
>>> I am certainly not convinced that things are working properly. The attached program and test script tests 8 cases, all compiled with fresh clang++ and fresh libc++:
>>>
>>> library    WITH_NS_STD    WITH_EXP_OVERLOADED  Result
>>> -------------------------------------------------------
>>> libc++     no             no                   d incorrect, dx1 correct
>>> libc++     no             yes                  d, dx1 correct
>>> libc++     yes            no                   d, dx1 incorrect
>>> libc++     yes            yes                  d, dx1 incorrect
>>> libstdc++  no             no                   d, dx1 correct
>>> libstdc++  no             yes                  doesn't compile
>>> libstdc++  yes            no                   d, dx1 correct
>>> libstdc++  yes            yes                  d, dx1 correct
>>>
>>> I'm going to give it a rest for a while, any ideas are highly appriciated!
>>>
>>> Regards, Ed.
>>> <pow_complex_test.cpp><test_pow.sh>_______________________________________________
>>> cfe-dev mailing list
>>> cfe-dev at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-dev
>> Eli alluded to the difference.  In 7 places of pow_complex_test.cpp you're picking up the wrong math function when WITH_NS_STD is not defined.
>>
>> In<math.h>   you have:
>>
>> double exp(double);
>> double sin(double);
>> double cos(double);
>>
>> In<cmath>   you *additionally* have:
>>
>> namespace std
>> {
>> float exp(float);
>> double exp(double);
>> long double exp(long double);
>>
>> float sin(float);
>> double sin(double);
>> long double sin(long double);
>>
>> float cos(float);
>> double cos(double);
>> long double cos(long double);
>> }
>>
>> In your WITH_EXP_OVERLOAD you silently convert long double to double and call ::exp, ::cos and ::sin unless WITH_NS_STD is defined.
>>
>> Ditto in main() when computing dx1.  You should qualify all of these calls with std:: to prevent this, e.g.g:
>>
>>       long double __e = std::exp(__x.real());
>>       return std::complex<long double>(__e * std::cos(__i), __e * std::sin(__i));
>>
>> Howard
>>
> Ah, I think the penny dropped.
>
> In which case: shouldn't the template for std::pow be specialized into
> double and long double functions, with one calling exp and the latter
> calling expl?
>
> Anyway, thanks to all who replied,
>
> -- Ed.
>
Strike that remark. I'm dealing with a linker issue on my system. Here's 
why:

* Linked with '-nodefaultlibs -lm' gives incorrect results (err~2e-10)
* Linked without those linker options gets me err = 0.

Something gets pulled from libgcc* that makes things work. Don't know 
why yet.

Leason: stray of the golden path and you learn a lot. For those who like 
a linker puzzle (and walk into the dire strait of my private 
investigations), the compile/link logs are attached.

-- Ed.
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: libm.log
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20111211/78f9b2f8/attachment.ksh>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: nolibm.log
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20111211/78f9b2f8/attachment-0001.ksh>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: pow_complex_test.cpp
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20111211/78f9b2f8/attachment-0002.ksh>