llvm dev - Mar 2014 - [LLVMdev] [RFC] Alias should not point to declarations

On 30 October 2013 16:16, Tim Northover <t.p.northover at gmail.com>
wrote:
> Hi Rafael,
>
>> With all that in mind, the attached patch changes the verifier to
>> reject aliases to declarations and updates that language reference.
>
> MachO has an R_INDR (== "indirect") symbol flag/type that (from
my
> understanding) exactly reflects this. The linker is supposed to record
> the alias and define both symbols when the referee is defined. In
> fact, I've been working on exposing this to C code very recently so
> I'd obviously quite like it to stick around.
>
> Currently I've not submitted a patch because, even though it's
> documented, ld64 doesn't have support just yet so I can't actually
> test things properly.
>
> I think the only benefit is going to be to compile-time and
> source-structure, but it would still be a shame to lose it when LLVM
> is so close to having the ideal model right now.
We chatted a bit or IRC and it does look like N_INDR  has the
behaviour of llvm aliases. An alias to a declaration is a definition
for example. It is documented here:
https://developer.apple.com/library/mac/documentation/DeveloperTools/Conceptual/MachORuntime/Reference/reference.html

I will try to make the ELF/COFF restrictions explicit with errors in
codegen instead.

Cheers,
Rafael

> We chatted a bit or IRC and it does look like N_INDR  has the
> behaviour of llvm aliases. An alias to a declaration is a definition
> for example. It is documented here:
>
https://developer.apple.com/library/mac/documentation/DeveloperTools/Conceptual/MachORuntime/Reference/reference.html
We discussed this a bit more on IRC today. Fixing pr10367 is more
important now that there is work under way to support comdats, which
would use aliases to implement RTTI for the msvc abi.

It is much easier to design a feature based on what existing tools
(linkers in this case) support and N_INDR is not currently
implemented. On the other hand, adding support for it once it is
implemented should be relatively easy.

The attached patch then effectively just replaces a report_fatal_error
with a proper verifier check that represent what can currently be
implemented.

Tim, can you confirm that it is OK?

Thanks,
Rafael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: t.patch
Type: text/x-patch
Size: 15096 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140312/682645ea/attachment.bin>

> Tim, can you confirm that it is OK?
Yep. I'll try to make sure nothing makes N_INDR support impossible,
but there's no particular reason to pretend it's already there as far
as I can see.

Cheers.

Tim.