Why is mangling different for separate compilation?

[Atila Neves via Digitalmars-d]

On Saturday, 27 February 2016 at 06:34:19 UTC, Joakim wrote:
> On Friday, 26 February 2016 at 12:45:35 UTC, Atila Neves wrote:
>> foo.d:
>> -----
>> void main() {
>>     import bar;
>>     foreach(ut; __traits(getUnitTests, bar)) ut();
>> }
>> -----
>>
>> bar.d:
>> -----
>> unittest { assert(1 == 2); }
>> -----
>>
>>
>> # compile all at once
>> dmd -unittest foo.d bar.d # fine
>>
>>
>> # compile separately
>> dmd -c -unittest foo.d
>> dmd -c -unittest bar.d
>> dmd foo.o bar.o
>> foo.o:foo.d:function _Dmain: error: undefined reference to 
>> '_D3bar16__unittestL2_531FZv'
>> collect2: error: ld returned 1 exit status
>> --- errorlevel 1
>>
>>
>> objdump shows me that bar.o has a '_D3bar14__unittestL2_1FZv' 
>> symbol, which is nearly, but not quite, what foo.o is trying 
>> to call.
>>
>> Huh?
>
> As Walter notes, it's because the mangling changes based on 
> whether you separately compile the unit tests.  It's why you 
> needed to generate the main D source file for unit-threaded 
> too, becuase __traits(getUnitTests) gets confused by this same 
> issue if the modules containing tests are separately compiled:
>
> https://github.com/atilaneves/unit-threaded/blob/master/source/unit_threaded/runtime.d#L175

Well, guess how I found out about this issue? ;) It turns out all 
my unit test files were always in the same package due to 
laziness, and as soon as I tried compiling separately, things 
broke. Sigh.

I've had similar problems in the past with template mixins. It 
seems D's compile-time features don't mix with any kind of 
separate compilation, which is a shame.

Atila