[PATCH][lld][PECOFF] Add Support for entry point symbol name

[Nick Kledzik]

On Aug 23, 2013, at 10:42 AM, Jesús Serrano <dagonoweda at gmail.com> wrote:
> It's perfectly ok to me that you submit the patch just as it is, as you said, it is useful.
> 
> In the other hand, I've been thinking how to implement the final link behavior. The search of the entry point symbol must be done obviously after the input reading. I think that also after the resolve phase, since the " _main" symbol could be defined in a shared library. One possible solution would be to search inside a new Pass before the writing. What do you think?
A pass is too heavy weight, and passes cannot lookup atoms by name…

The way to do this is like how the mach-o and ELF writers find _main.  The PECOFF Writer should implement addFiles() to add a file to the link which has a zero size atom which has references to UndefinedAtoms for each of those potential entry points (_main, _wWinMain, etc).  These UndefinedAtoms have a canBeNull() of canBeNullAtBuildTime, so that the resolver will not error if they are not found.  Then when the PECOFF Writer is told to write the file, it looks back at the magic atom it added and look to see which of the possible entry points were found and uses the highest priority one.

-Nick


> 
> 
> On 23/08/2013 1:19, Rui Ueyama wrote:
>> Thank you for the investigation. This is more complex than I thought and very interesting. This behavior needs to be implemented for link.exe compatibility. It doesn't make sense to implement this in the driver as you wrote and should be in the linking context.
>> 
>> I want submit your patch for now, as it's useful. You'll be able to remove the code from the driver and add new code to the linking context. Does this sound good?
>> 
>> On Tue, Aug 20, 2013 at 1:07 PM, Jesús Serrano <dagonoweda at gmail.com> wrote:
>> 
>> I've attached a new patch with this changes. Now the windows driver sets the entry symbol name according to the corresponding options. However, this won't work in all cases. Depending onwhether the crt is compiled as multibyte or unicode, the entry point symbol name changes from "mainCRTStartup" to "wmainCRTStartup", and the same for the windows subsystem. The problem with this is that the use of unicode is unknown at link time (or at least I don't know how to retrieve this option from object files). One possible solution would be to have two entry point names in the linking context, and find a symbol mathing any of them, but I don't like it so much. Any ideas on this?
>> 
>> I've looking into this, and seems that LINK.EXE behavior for executable images is to "select" the CRT entry point depending on the symbols defined in the object files.
>>  If neither the subsystem nor the entry point are defined in command line arguments, the linker searches for the symbols "_main", "_wmain", "_WinMain" and "_wWinMain". Depending on the found symbols, the linker chooses the subsystem and the CRT entry point, having "_main" the highest priority and "_wWinMain" the lesser. This is, if "_main" is defined, the subsystem will be CONSOLE and the CRT entry point "_mainCRTStartup", and so on. If several of these entry point symbols for the same subsystem are defined, the linker chooses the symbol of higher priority and prints a warning.
>> If the subsystem is defined in the command line, then the linker searches only for the corresponding entry point symbols to choose between the unicode and the multibyte CRT entry point.
>> If the entry point symbol is defined in the command line, but not the subsystem, the linker forces to choose one and throws an error.
>> 
>> If we want to follow this behavior, it has no sense to let the driver to choose the entry point symbol name. This could be handled by the linking context after the file parsing or the by writer, what do you think about that?
>> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130823/5adf77ef/attachment.html>