draft 1

Author: vadimcn

text/0000-dllimport.md

@@ -0,0 +1,152 @@
+- Feature Name: dllimport
+- Start Date: 2016-08-13
+- RFC PR: (leave this empty)
+- Rust Issue: (leave this empty)
+
+# Summary
+[summary]: #summary
+
+Make compiler aware of the association between library names adorning `extern` blocks
+and symbols defined within the block.  Add attributes and command line switches that leverage
+this association.
+
+# Motivation
+[motivation]: #motivation
+
+Most of the time a linkage directive is only needed to inform the linker about
+what native libraries need to be linked into a program. On some platforms,
+however, the compiler needs more detailed knowledge about what's being linked
+from where in order to ensure that symbols are wired up correctly.
+
+On Windows, when a symbol is imported from a dynamic library, the code that accesses 
+this symbol must be generated differently than for symbols imported from a static library.
+
+Currently the compiler is not aware of associations between the libraries and symbols 
+imported from them, so it cannot alter code generation based on library kind.
+
+# Detailed design
+[design]: #detailed-design
+
+### Library <-> symbol association
+
+The compiler shall assume that symbols defined within extern block
+are imported from the library mentioned in the `#[link]` attribute adorning the block.
+
+### Changes to code generation
+
+On platforms other than Windows the above association will have no effect.
+On Windows, however, `#[link(..., kind="dylib")` shall be presumed to mean linking to a dll, 
+whereas `#[link(..., kind="static")` shall mean static linking.  In the former case, all symbols
+associated with that library will be marked with LLVM [dllimport][1] storage class.
+
+[1]: http://llvm.org/docs/LangRef.html#dll-storage-classes
+
+### Library name and kind variance
+
+Many native libraries are linked via the command line via `-l` which is passed
+in through Cargo build scripts instead of being written in the source code
+itself. As a recap, a native library may change names across platforms or
+distributions or it may be linked dynamically in some situations and
+statically in others which is why build scripts are leveraged to make these
+dynamic decisions. In order to support this kind of dynamism, the following 
+modifications are proposed:
+
+- A new library kind, "abstract".  An "abstract" library by itself does not 
+  cause any libraries to be linked. Its purpose is to establish an identifier, 
+  that may be later referred to from the command line flags.
+- Extend syntax of the `-l` flag to `-l [KIND=]lib[:NEWNAME]`.  The `NEWNAME` 
+  part may be used to override name of a library specified in the source.
+- Add new meaning to the `KIND` part: if "lib" is already specified in the source, 
+  this will override its kind with KIND.  Note that this override is possible only 
+  for libraries defined in the current crate.
+
+Example:
+
+```rust
+// mylib.rs
+#[link(name = "foo", kind="dylib")]
+extern {
+    // dllimport applied 
+}
+
+#[link(name = "bar", kind="static")]
+extern {
+    // dllimport not applied
+}
+
+#[link(name = "baz", kind="abstract")]
+extern {
+    // dllimport not applied, "baz" not linked
+}
+```
+
+```
+rustc mylib.rs -l static=foo # change foo's kind to "static", dllimport will not be applied  
+rustc mylib.rs -l foo:newfoo # link newfoo instead of foo   
+rustc mylib.rs -l dylib=baz:quoox # specify baz's kind as "dylib", change link name to quoox.
+```
+
+### Unbundled static libs (optional)
+
+It had been pointed out that sometimes one may wish to link to a static system library 
+(i.e. one that is always available to the linker) without bundling it into .lib's and .rlib's.   
+For this use case we'll introduce another library "kind", "static-nobundle".
+Such libraries would be treated in the same way as "static", minus the bundling.
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+For libraries to work robustly on MSVC, the correct `#[link]` annotation will
+be required. Most cases will "just work" on MSVC due to the compiler strongly
+favoring static linkage, but any symbols imported from a dynamic library or
+exported as a Rust dynamic library will need to be tagged appropriately to
+ensure that they work in all situations. Worse still, the `#[link]` annotations
+on an `extern` block are not required on any other platform to work correctly,
+meaning that it will be common that these attributes are left off by accident.
+
+
+# Alternatives
+[alternatives]: #alternatives
+
+- Instead of enhancing `#[link]`, a `#[linked_from = "foo"]` annotation could be added.
+  This has the drawback of not being able to handle native libraries whose
+  name is unpredictable across platforms in an easy fashion, however.
+  Additionally, it adds an extra attribute to the comipler that wasn't known
+  previously.
+
+- Support a `#[dllimport]` on extern blocks (or individual symbols, or both).
+  This has the following drawbacks, however:
+  - This attribute would duplicate the information already provided by 
+    `#[link(kind="...")]`.
+  - It is not always known whether `#[dllimport]` is needed. Native
+    libraires are not always known whether they're linked dynamically or
+    statically (e.g. that's what a build script decides), so `dllimport`
+    will need to be guarded by `cfg_attr`.
+
+- When linking native libraries, the compiler could attempt to locate each
+  library on the filesystem and probe the contents for what symbol names are
+  exported from the native library. This list could then be cross-referenced
+  with all symbols declared in the program locally to understand which symbols
+  are coming from a dylib and which are being linked statically. Some downsides
+  of this approach may include:
+
+    - It's unclear whether this will be a performant operation and not cause
+      undue runtime overhead during compiles.
+
+    - On Windows linking to a DLL involves linking to its "import library", so
+      it may be difficult to know whether a symbol truly comes from a DLL or
+      not.
+
+    - Locating libraries on the system may be difficult as the system linker
+      often has search paths baked in that the compiler does not know about.
+
+- As was already mentioned, "kind" override can affect codegen of the current crate only.
+  This overloading the `-l` flag for this purpose may be confusinfg to developers.
+  A new codegen flag might be a better fit for this, for example `-C libkind=KIND=LIB`.
+
+# Unresolved questions
+[unresolved]: #unresolved-questions
+
+- Should un-overridden "abstract" kind cause an error, a warning, or be silently ignored?
+- Do we even need "abstract"?  Since kind can be overridden, there's no harm in providing a default in the source.
+- Should we allow dropping a library specified in the source from linking via `-l lib:` (i.e. "rename to empty")?