[PATCH] D11724: COFF: Add test for ld/section created import library

[Rui Ueyama via llvm-commits]

On Thu, Aug 6, 2015 at 9:32 AM, Martell Malone <martellmalone at gmail.com>
wrote:

> Yeah I thought the same thing would b true.
> Okay I'll start working towards this then for the gendef tool.
> Did you know that llvm-objdump can't dump implib sections btw ?
> I started working on a patch for that.
> I'll add you as a subscriber when I send in the fix
>
>
> Would you be able to send me the patch you have for dll style libs ?
>

I don't have a patch for that. My idea was identifying dlltool-style import
library by the existence of ".idata$7" section (because only the GNU
extension uses that section) and create a ImportFile for that file, but as
that was too hacky, I didn't actually write code for that.


>
> Thanks again for the help Rui
>
> On Thu, Aug 6, 2015 at 5:28 PM, Rui Ueyama <ruiu at google.com> wrote:
>
>> On Thu, Aug 6, 2015 at 9:24 AM, Martell Malone <martellmalone at gmail.com>
>> wrote:
>>
>>> On second thought, we don't have to solve that only with the import
>>>> library. You can create a (regular) COFF object containing a thunk for an
>>>> aliased function, and include that object file to a .lib, no?
>>>
>>> Yes that seems like a good solution to this problem I might be able to
>>> add something to the genlib tool to support this.
>>> I might have to mockup one for yaml2obj.
>>> Can we have a regular coff object in the same lib as an implib ?
>>>
>>
>> I believe so -- the import library is just a regular ar file (.lib file)
>> after all and there's no special bit or something that distinguishes import
>> libraries and other .lib files, but you may want to try if in doubt. You
>> can add a regular COFF file to an existing import library using lib command.
>>
>>
>>>
>>> On Thu, Aug 6, 2015 at 5:17 PM, Rui Ueyama <ruiu at google.com> wrote:
>>>
>>>> On second thought, we don't have to solve that only with the import
>>>> library. You can create a (regular) COFF object containing a thunk for an
>>>> aliased function, and include that object file to a .lib, no?
>>>>
>>>> On Thu, Aug 6, 2015 at 8:22 AM, Rui Ueyama <ruiu at google.com> wrote:
>>>>
>>>>> I don't think the short import library supports that. With that you
>>>>> can undecorate names, but AFAIK you cannot define arbitrary aliases for
>>>>> dllexported symbols. Does mingw-w64 heavily relies on that feature?
>>>>>
>>>>> On Thu, Aug 6, 2015 at 8:14 AM, Martell Malone <
>>>>> martellmalone at gmail.com> wrote:
>>>>>
>>>>>> Before I let this slip by aswell there is one big issue with
>>>>>> switching over to the implib format for mingw-w64
>>>>>>
>>>>>> LIBRARY "user32.dll"
>>>>>> EXPORTS
>>>>>> MessageBoxA == MessageBoxW
>>>>>>
>>>>>> dlltool objct format provides us with the option of having an alias.
>>>>>> Can we do this in implib ?
>>>>>>
>>>>>> What the above .def means that where MessagBoxW is called it is
>>>>>> joined to MessageBoxA.
>>>>>>
>>>>>>
>>>>>> On Thu, Aug 6, 2015 at 3:58 PM, Martell Malone <
>>>>>> martellmalone at gmail.com> wrote:
>>>>>>
>>>>>>> Yeah, I found that too and it (especially the way how it creates a
>>>>>>>> gap between .idata$3 and .idata$4) looks really hacky. I also found that
>>>>>>>> GNU ld has a special logic to order .idata$<n> sections.
>>>>>>>> At first I thought that I could mimic GNU ld and MSVC linker to
>>>>>>>> generate the .idata section, but seems like it would really mess up the DLL
>>>>>>>> import table generation code. We probably should keep the existing logic
>>>>>>>
>>>>>>> Yeah I don't think copying gnu ld is the way to go, it is a very
>>>>>>> hacky project to say the least :)
>>>>>>>
>>>>>>>
>>>>>>> It feels to me that it makes more sense to add a new option to
>>>>>>>> dlltool to generate short import libraries. It shouldn't be that hard.
>>>>>>>> Martell, what do you think?
>>>>>>>
>>>>>>> I already wrote a replacement tool for this called genlib that will
>>>>>>> go into the mingw-w64 project.
>>>>>>> I want to remove mingw-w64's dependancy on binutils so that we can
>>>>>>> have a clang based toolchain without binutils at all.
>>>>>>> The notion of having dlltool as part of binutils made no sense in
>>>>>>> the first place, It should have been part of mingw to begin with.
>>>>>>> The name was chosen to correlate to on of mingw-w64's other tools
>>>>>>> called gendef which creates the def files from parsing dll's.
>>>>>>> I'm still finalizing the code in this and doing some tests.
>>>>>>>
>>>>>>> I then tried to detect dlltool-style import files and read them as
>>>>>>>> if they were short import libraries, so that I can keep the existing code.
>>>>>>>> That didn't work well because it's not easy to detect dlltool-style import
>>>>>>>> files in a reliable manner without sacrificing performance.
>>>>>>>
>>>>>>> While you might not want to merge because of this into the official
>>>>>>> project because of performance issues it might be something for the
>>>>>>> mingw-w64 users to avail of until ld supports import style libraries.
>>>>>>>
>>>>>>> If you have the patch for this I'd like have a look at it if
>>>>>>> possible.
>>>>>>> I'd like to try and apply this over the PECOFF for the clang 3.7
>>>>>>> package in our msys2 distro.
>>>>>>>
>>>>>>> The issue I have is I would have to get approval to switch to using
>>>>>>> genlib instead of dlltool as mingw-w64's default.
>>>>>>> This would not be approved until ld supports implibs and the next
>>>>>>> version of binutils released.
>>>>>>> As you probably well know how things work that could take months to
>>>>>>> get changed over.
>>>>>>> I'd like to use your patch as a base for a temporary stop over until
>>>>>>> this happens
>>>>>>>
>>>>>>> I'm sure the distro users of msys2 won't mind a performance hit
>>>>>>> until 3.8 rather then having no support at all.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Wed, Aug 5, 2015 at 10:51 PM, Rui Ueyama <ruiu at google.com> wrote:
>>>>>>>
>>>>>>>> On Mon, Aug 3, 2015 at 10:36 PM, Martell Malone <
>>>>>>>> martellmalone at gmail.com> wrote:
>>>>>>>>
>>>>>>>>> From my reading on how gnuld handles PE/COFF
>>>>>>>>> It uses a linker script that describes how it lays out its idata
>>>>>>>>> section.
>>>>>>>>> From i386pe.x
>>>>>>>>>
>>>>>>>>>   .idata BLOCK(__section_alignment__) :
>>>>>>>>>   {
>>>>>>>>>     /* This cannot currently be handled with grouped sections.
>>>>>>>>> 	See pe.em:sort_sections.  */
>>>>>>>>>     SORT(*)(.idata$2)
>>>>>>>>>     SORT(*)(.idata$3)
>>>>>>>>>     /* These zeroes mark the end of the import list.  */
>>>>>>>>>     LONG (0); LONG (0); LONG (0); LONG (0); LONG (0);
>>>>>>>>>     SORT(*)(.idata$4)
>>>>>>>>>     SORT(*)(.idata$5)
>>>>>>>>>     SORT(*)(.idata$6)
>>>>>>>>>     SORT(*)(.idata$7)
>>>>>>>>>   }
>>>>>>>>>
>>>>>>>>>
>>>>>>>> Yeah, I found that too and it (especially the way how it creates a
>>>>>>>> gap between .idata$3 and .idata$4) looks really hacky. I also found that
>>>>>>>> GNU ld has a special logic to order .idata$<n> sections.
>>>>>>>>
>>>>>>>> At first I thought that I could mimic GNU ld and MSVC linker to
>>>>>>>> generate the .idata section, but seems like it would really mess up the DLL
>>>>>>>> import table generation code. We probably should keep the existing logic.
>>>>>>>>
>>>>>>>> I then tried to detect dlltool-style import files and read them as
>>>>>>>> if they were short import libraries, so that I can keep the existing code.
>>>>>>>> That didn't work well because it's not easy to detect dlltool-style import
>>>>>>>> files in a reliable manner without sacrificing performance.
>>>>>>>>
>>>>>>>> It feels to me that it makes more sense to add a new option to
>>>>>>>> dlltool to generate short import libraries. It shouldn't be that hard.
>>>>>>>> Martell, what do you think?
>>>>>>>>
>>>>>>>>> Under ld/emultempl/pep.em in binutils it describes how it converts the MS import library to its format.
>>>>>>>>> see here http://github.com/bminor/binutils-gdb/blob/master/ld/emultempl/pep.em#L1625
>>>>>>>>> From the code in this function and the rest of pep.em we can see how it handles it.
>>>>>>>>> I'm sure this will make more sense to you however as you know what the MS format is.
>>>>>>>>>
>>>>>>>>> I assume the conversion will have to go the other way for us
>>>>>>>>> I think the most notable thing is the use of idata7 instead of idata6 for the dll name
>>>>>>>>> We could use that as a check?
>>>>>>>>> Then hijack it pulling out the function names once we see this being used and insert it into the sections like a MS generated one
>>>>>>>>>
>>>>>>>>> You might have a much cleaner solution however. :)
>>>>>>>>>
>>>>>>>>> If this isn't enough insight into what you need I can do more digging.
>>>>>>>>> Just ping me and let me know
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On Tue, Aug 4, 2015 at 3:15 AM, Rui Ueyama <ruiu at google.com>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>> ruiu added a comment.
>>>>>>>>>>
>>>>>>>>>> Do you know anything about how GNU ld handles these import
>>>>>>>>>> libraries? My
>>>>>>>>>> linker is able to read it and construct .idata section, but the
>>>>>>>>>> resulting
>>>>>>>>>> .idata section is not going to be in correct format. If GNU
>>>>>>>>>> linker is able
>>>>>>>>>> to generate a correct .idata section from this type of import
>>>>>>>>>> libraries,
>>>>>>>>>> there must be something I'm missing in my linker.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> http://reviews.llvm.org/D11724
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150806/fce7d74d/attachment.html>