[PATCH] Add a fallback mechanism for undefined atom.

[Rui Ueyama]

On Thu, Aug 29, 2013 at 3:15 PM, kledzik at apple.com <kledzik at apple.com>wrote:

>
>   Interesting feature.
>
>   What happens if a.o has an undefine foo which says its fallback is bar,
> but in b.o an undefine for foo says its fallback is baz?  Is that an error,
> or does each fallback apply to its original usage.  The problem is that the
> Resolver is currently coalescing UndefinedAtoms based on their name. I does
> not know that fallbacks need to match too.
>

Good question. Such corner case behavior is not really defined in the MS COFF
spec, so I tried to see how MS link.exe behaves. link.exe seems to follow
the latter; it's allowed that the symbols with the same name in different
object files to have different fallback symbols, and each fallback apply to
its original usage. Because undefined atoms are as you said coalesced by
name, this patch is not strictly compatible with link.exe. I'd think is
okay in practice, as the feature is rarely used, though. (The support of
this feature is needed because the C++ standard library uses it, but
regular programs usually don't.)

  Another variant is to have Undefined return a alternate name as a string,
> rather than returning a new UndefinedAtom object.
>
>   Another approach is to use weak aliases.  That is if a.o has an undefine
> for foo with a fallback of bar, that when parsing that .o file into atoms
> produces an UndefinedAtom for foo, but also a DefinedAtom with name=foo,
> weak (mergeAsWeak), hidden (scopeLinkageUnit), size=0, isAlias=true which
> aliases to bar.  So, it any definition of foo does show up, it will
> override the weak alias.  If not the weak alias will be used and change the
> reference to bar.
>

If any (real) definition of bar will show up, link will have to fail with an
error message that bar was not found. That's different from weak symbol; if
no real definition is found, the weak symbol will just be used without error.
The implementation of isAlias support seems pretty immature. It really
looked like a stub, and nothing has really been wired up to it..

Also, a weak symbol with internal linkage sounds odd and hard to
imagine a correct
semantics. Should it be coalesced away with the symbol in other file
because it's weak, or shouldn't because it's static? Clang does not allow
the combination of __attribute((weak)) and static, probably to avoid the
issue.


>
>   I don't think the isAlias support is all wired up in lld.  But weak
> aliases are something common in ELF. So would should  get it working, and
> doing so may provide a way to implement this COFF feature.
>
>   One issue is how this interacts with static libraries.  The weak alias
> for foo make look like there is a definition, so any static archives will
> not be searched for foo.  That seems different than what you want.  I'm not
> sure what the gnu linker does about this with weak aliases.  We may need a
> new resolver option as to how to handle this.
>
>
> ================
> Comment at: lib/Core/Resolver.cpp:214
> @@ +213,3 @@
> +      // for COFF "weak external" symbol.
> +      const UndefinedAtom *fallbackUndefAtom = undefAtom->fallback();
> +      if (fallbackUndefAtom) {
> ----------------
> You need to re-test:
>   if (!_symbolTable.isDefined(undefName)
> because searchLibraries() may have already found a definition.
>
>
> http://llvm-reviews.chandlerc.com/D1550
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130829/1620e74b/attachment.html>