[llvm-commits] [llvm] r75619 - in /llvm/trunk/lib: CodeGen/AsmPrinter/AsmPrinter.cpp Target/DarwinTargetAsmInfo.cpp

[Chris Lattner]

On Jul 17, 2009, at 3:35 PM, Dale Johannesen wrote:
> On Jul 17, 2009, at 3:05 PMPDT, Chris Lattner wrote:
> On Jul 14, 2009, at 9:56 AM, Dale Johannesen wrote:
>>>>>> getValueName(GV).substr(0,strlen(getLessPrivateGlobalPrefix()))  
>>>>>> ==
>>>>> +        Mang->getMangledName(GV).substr(0,
>>>>> +
>>>>> strlen(getLessPrivateGlobalPrefix())) ==
>>>>>       getLessPrivateGlobalPrefix())))
>>>>
>>>> Dale, I think you added "less private global". Do you remember?
>>>
>>> Only what's in the comment:
>>>   /// LessPrivateGlobalPrefix - This prefix is used for symbols
>>> that should
>>>   /// be passed through the assembler but be removed by the
>>> linker.  This
>>>   /// is "l" on Darwin, currently used for some ObjC metadata.
>>
>> Dale, I see a series of commits where this code was added, refined,  
>> etc, culimating in this specific patch:
>> http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080908/066962.html
>
> There's some later messages which talk about this.  The first one  
> has an example.  There was originally a radar but I think the first  
> message is more helpful.
> http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080908/066985.html
> http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080908/067012.html

This is a really unfortunate way to do this.  I filed rdar://7071300  
to track this work, but it has several dependencies.  In particular,  
we need to introduce a new llvm.used.assembler directive, which is  
just like llvm.used, but doesn't cause .no_dead_strip to get emitted.   
We need to add a new linkage type corresponding to "l" labels in  
darwin, which itself would be a huge cleanup, then we need to switch  
clang/llvm-gcc to work with this.

I don't want to lose this, but I don't have time to tackle it right now.

>> Unfortunately there are no testcases and I have no idea why this  
>> stuff is being done.  Can you explain it?
>>
>> I think this is really the wrong way to handle this.  It add a  
>> bunch of ad-hoc hooks and the code in  
>> DarwinTargetAsmInfo::emitUsedDirectiveFor (besides being a very  
>> misleading name for the hook) is gross string manipulation that  
>> should not happen in the code generator like that.
>
> Patches welcome!  I'm certainly not wedded to the current approach  
> or the name (I don't like the name either, just couldn't think of  
> anything better).

I will rename this.  The name implies it actually emits something.

-Chris