Cross compilation + #[phase(syntax)] = sadness

[alexcrichton]

I believe that when you specify a crate with #[phase(syntax)], that should resolve a crate for the host architecture, not the target architecture. This means that if I'm doing a cross-compile, I'm going to attempt to dlopen libraries for the target architecture, which will probably fail if I'm cross compiling.

cc @sfackler

[alexcrichton]

Another problem I just uncovered:

Let's say that I have my own syntax extension inside of directory foo. To use this syntax, extension I will pass -L foo to the compiler. The compiler will then open up foo/libmymacro-ffffff-1.0.dylib in an attempt to read the metadata and find the macro registrar. By default, crate loading will load dependencies as well. This means that when libmymacro is built for the host architecture, it will load the target architecture's libstd metadata (!!!).

The new SVH infrastructure has detected this bug because the SVH is different between target/host crates. This could be fixed with a combination of #10814 and perhaps some more clever infrastructure for loading syntax extensions with a different target architecture.