[PATCH] Add support for a directory argument to llvm-link

[Rafael Espíndola]

>> When do you do that as part of regular development?
>
>
> For example, when creating large bitcode files from entire projects (e.g.
> for analysis of llvm optimizations or performance testing (of llvm libs), or
> just for stress testing). Until LLD rules the world and has turnkey LTO and
> an emit-bitcode option (a while yet), there will be a need for flexibly
> linking a ton of bitcode modules.

Have you ever had to do that? I have debugged quiet a few lto
failures, from bootstrapping clang to building firefox. Both OS X
(ld64) and linux (gold) and in every case the reduction was done
directly with the native linker.

>> No, it doesn't. Quiet the opposite. Having a feature that makes it
>> easier to use llvm-mc in production is undesirable.
>
>
> There's a big difference between a prototype for a product and going to
> production. Also, for prototypes being able to quickly pull things together
> and have them work is important and convenience of use is a big part of
> that.
>
> These prototypes are not alien to upstream development. Prototypes feed back
> experience that informs upstream development and allocation of development
> effort.

I appreciate the difference. But the experience of what the llvm-ld
had grown to makes me really not what to add features to llvm-link.
The easier we make for people to confuse it with a production linker,
the more likely someone is to fall into that trap.

Cheers,
Rafael