[phobos] Can somebody with OSX please reduce the Phobos unittest failure?

[Brad Roberts]

On 10/25/2010 11:38 PM, Walter Bright wrote:
> 
> 
> Brad Roberts wrote:
>> On 10/25/2010 8:58 PM, Walter Bright wrote:
>>  
>>> Don Clugston wrote:
>>>    
>>>> We have an OSX-only failure, which I think is a compiler bug.
>>>> It's failing in std.bigint, which only imports
>>>> std.internal.math.biguintCore and std.internal.math.bignumX86.
>>>> There's no dependency on anything else in Phobos, so it ought to be
>>>> pretty easy to reduce.
>>>> I can't reduce it, because I don't have OSX.
>>>>
>>>> The failing test is this one:
>>>>     assert(BigInt(0x1234_5678_9ABC_5A5AL).toLong() == 0x1234_5678_9ABC_5A5AL);
>>>>
>>>> This is blocking a release.
>>>>
>>>>         
>>> Found the problem - you're right, it's a compiler bug. A bad one. Oops!
>>>     
>>
>> An osx specific problem?
>>   
> 
> Doing a signed long compare calls a function in llmath.d. That function is the
> only one that returns its result in the flags. In OSX, the stack must always be
> aligned on 16 bytes when a function is called. The alignment is done by
> adding/subtracting from ESP, which resets the flags. Oops! The solution was to
> just inline the comparison code.

Nice and subtle.  My favorite kind of ick.