[dmd-internals] How to avoid 4-word padding for symbols?

[Michel Fortin]

Turns out it doesn't work because those symbols refer to each other, so by coalescing them in one symbol I'm creating a self-referenced symbol and the backend doesn't like that.

I resorted to adding a 'noalign' flag to outdata() in the backend. It's a small change and it works, but it feels like a hack. Ideally there should be a way to specify the alignment per segment (there might already be one, but I can't figure it out). Since the Objective-C stuff belongs to separate segments it'd make things really easy.

I know there is the 'align' parameter in mach_getsegment(), but it doesn't seem to have any effect other than changing alignment value for the section in the object file.


Le 2010-11-07 à 21:05, Walter Bright a écrit :

> Symbols are aligned when written out to the data segment.
> 
> To defeat the alignment, symbols that should be adjacent should be merged into one symbol.
> 
> Michel Fortin wrote:
>> It's probably a small thing that I'm missing, but I have a small problem where the code below causes undesired padding to be added between two symbols in the object file. This code writes 5 words to the object file:
>> 
>>    symbol = symbol_name(sname, SCstatic, type_fake(TYnptr));
>>    symbol->Sdt = dt;
>>    symbol->Sseg = objc_getsegment(SEGprotocol);
>>    outdata(symbol);
>> 
>> If you call it 2 times, it generates 5*2 + 3 padding bytes between the two (added dashes to show padding):
>> 
>> [25] 002d2 0034 000da6    0 0f94   2 10000000 00 00 __protocol __OBJC
>> 0000:   0  0  0  0  6  3  0  0  0  0  0  0  0  0  0  0   ................
>> 0010:   0  0  0  0 -0--0--0--0--0--0--0--0--0--0--0--0   ................
>> 0020:   0  0  0  0 36  3  0  0  0  0  0  0  0  0  0  0   ....6...........
>> 0030:   0  0  0  0                     ....            
>> This padding is causing a crash when starting the Objective-C runtime because it expects the two to be contiguous.
>> 
>> What needs to be changed to remove this padding?
>> 
>>  
> _______________________________________________
> dmd-internals mailing list
> dmd-internals at puremagic.com
> http://lists.puremagic.com/mailman/listinfo/dmd-internals

-- 
Michel Fortin
michel.fortin at michelf.com
http://michelf.com/