On linux with dmd 0.148, I'm seeing some odd linking issues that I haven't tracked down just yet.

$ cat test_box.d
import std.boxer;

int main()
{
    Box foo = box(1);
    int bar = unbox!(int)(foo);

    return 0;
}

$ dmd -g -v
test_box.d
parse     test_box
semantic  test_box
semantic2 test_box
semantic3 test_box
code      test_box
generating code for function 'main'
generating code for function 'unbox'
generating code for function 'unboxCastInteger'
gcc test_box.o -o test_box -g -lphobos -lpthread -lm -Xlinker -L/workplace/bradrob/dmd-0.148/dmd/lib
test_box.o(.gnu.linkonce.t_D3std5boxer24__T16unboxCastIntegerTiZ16unboxCastIntegerFS3std5boxer3BoxZi+0x18): In function `_D3std5boxer24__T16unboxCastIntegerTiZ16unboxCastIntegerFS3std5boxer3BoxZi':
/workplace/bradrob/d-playground/test_box.d:5: undefined reference to `_assert_3std5boxer'
collect2: ld returned 1 exit status
--- errorlevel 1


Commenting out the assert in std/boxer.d inside unboxCastInteger(T) at line 429 allows the link to succeed.

Similarly, if I use something other than an int and hit one of the other unboxCast[pick a type] templates I get the same problem.

Thoughts?

Thanks,
Brad

On Thu, 2 Mar 2006 17:38:00 -0800, Brad Roberts wrote:

> On linux with dmd 0.148, I'm seeing some odd linking issues that I haven't tracked down just yet.

On Windows, you have to compile anything that imports std.boxer with the "-release" switch to get the linker happy.

-- 
Derek
(skype: derek.j.parnell)
Melbourne, Australia
"Down with mediocracy!"
3/03/2006 3:18:27 PM

On Fri, 3 Mar 2006, Derek Parnell wrote:

> On Thu, 2 Mar 2006 17:38:00 -0800, Brad Roberts wrote:
> 
> > On linux with dmd 0.148, I'm seeing some odd linking issues that I haven't tracked down just yet.
> 
> On Windows, you have to compile anything that imports std.boxer with the "-release" switch to get the linker happy.
> 
> -- 
> Derek
> (skype: derek.j.parnell)
> Melbourne, Australia
> "Down with mediocracy!"
> 3/03/2006 3:18:27 PM

Not a bad temporary work around, but only temporary since that disables all asserts.

Thanks,
Brad

Derek Parnell wrote:
> On Thu, 2 Mar 2006 17:38:00 -0800, Brad Roberts wrote:
> 
>> On linux with dmd 0.148, I'm seeing some odd linking issues that I haven't tracked down just yet.
> 
> On Windows, you have to compile anything that imports std.boxer with the
> "-release" switch to get the linker happy.

Do we know the reason for that? Has it been reduced to a minimal test case?

Don Clugston schrieb am 2006-03-03:
> Derek Parnell wrote:
>> On Thu, 2 Mar 2006 17:38:00 -0800, Brad Roberts wrote:
>> 
>>> On linux with dmd 0.148, I'm seeing some odd linking issues that I haven't tracked down just yet.
>> 
>> On Windows, you have to compile anything that imports std.boxer with the "-release" switch to get the linker happy.
>
> Do we know the reason for that?

DMD generates different symbols depending on the state of the "-release" switch.

> Has it been reduced to a minimal test case?

Not as far as I know.

Thomas

On Sun, 5 Mar 2006, Thomas Kuehne wrote:

> Don Clugston schrieb am 2006-03-03:
> > Derek Parnell wrote:
> >> On Thu, 2 Mar 2006 17:38:00 -0800, Brad Roberts wrote:
> >> 
> >>> On linux with dmd 0.148, I'm seeing some odd linking issues that I haven't tracked down just yet.
> >> 
> >> On Windows, you have to compile anything that imports std.boxer with the "-release" switch to get the linker happy.
> >
> > Do we know the reason for that?
> 
> DMD generates different symbols depending on the state of the "-release" switch.
> 
> > Has it been reduced to a minimal test case?
> 
> Not as far as I know.
> 
> Thomas

I tried a little bit to reduce the problem.  If you're looking for a single file contained test case then I'm not sure that's doable.  I suspect this might relate to the other threads going on right now about symbols changing visibility based on compiler options.

Random thought:  Now that dmd can produce .di files, THOSE are what should be used now rather than the full .d's during subsequent compilation.  I haven't checked their output, but I assume that the unused paths around version, debug, release, etc are omitted such that the .di file _exactly_ matches the .o/.obj files produced.

Later,
Brad