Implement RFC 1210: impl specialization
This PR implements [impl specialization](https://github.com/rust-lang/rfcs/pull/1210),
carefully following the proposal laid out in the RFC.
The implementation covers the bulk of the RFC. The remaining gaps I know of are:
- no checking for lifetime-dependent specialization (a soundness hole);
- no `default impl` yet;
- no support for `default` with associated consts;
I plan to cover these gaps in follow-up PRs, as per @nikomatsakis's preference.
The basic strategy is to build up a *specialization graph* during
coherence checking. Insertion into the graph locates the right place
to put an impl in the specialization hierarchy; if there is no right
place (due to partial overlap but no containment), you get an overlap
error. Specialization is consulted when selecting an impl (of course),
and the graph is consulted when propagating defaults down the
specialization hierarchy.
You might expect that the specialization graph would be used during
selection -- i.e., when actually performing specialization. This is
not done for two reasons:
- It's merely an optimization: given a set of candidates that apply,
we can determine the most specialized one by comparing them directly
for specialization, rather than consulting the graph. Given that we
also cache the results of selection, the benefit of this
optimization is questionable.
- To build the specialization graph in the first place, we need to use
selection (because we need to determine whether one impl specializes
another). Dealing with this reentrancy would require some additional
mode switch for selection. Given that there seems to be no strong
reason to use the graph anyway, we stick with a simpler approach in
selection, and use the graph only for propagating default
implementations.
Trait impl selection can succeed even when multiple impls can apply,
as long as they are part of the same specialization family. In that
case, it returns a *single* impl on success -- this is the most
specialized impl *known* to apply. However, if there are any inference
variables in play, the returned impl may not be the actual impl we
will use at trans time. Thus, we take special care to avoid projecting
associated types unless either (1) the associated type does not use
`default` and thus cannot be overridden or (2) all input types are
known concretely.
r? @nikomatsakis
This commit fixes our support for cross compiling a compiler to run on FreeBSD.
Over the weekend I managed to get a cross compiler from Linux to FreeBSD [1]
which I hope to soon use to start producing FreeBSD nightly compilers. With the
`make dist` support added in #32237 we should be able to produce standard
rustc/rust-std packages for FreeBSD through a new slave with this cross compiler.
Currently, however, we don't "Just Work" when cross compiling FreeBSD and a
number of changes were required (part of this PR). They include:
* A few build fixes were needed in LLVM. Our own branch has been rebased on the
actual 3.8 release and I applied one extra commit [2] which contains two fixes:
1. The LLVM CMake build system passes the `-Wl,-z,defs` flag on many
platforms, but *not* when `CMAKE_SYSTEM_NAME` is "FreeBSD". Unfortunately
this doesn't take into account when we're cross compiling, and as predicted
the build will fail if `-Wl,-z,defs` is passed (see [3] for more info). To
fix this we test `TARGET_TRIPLE` instead of the `CMAKE_SYSTEM_NAME` which
is what we're compiling for which fixes the problem.
2. The `PATH_MAX` constant is apparently defined in a different location than
many other Unix systems, so a file which required this just needed some
help to keep compiling.
* Support for compiling compiler-rt with CMake has been added to rustbuild. It
looks like it just emulates Linux in what it compiles as it didn't seem to
naturally produce anything else... At least the architecture is right, so
seems good for now at least!
[1]: https://github.com/alexcrichton/port-of-rust/blob/master/prebuilt/freebsd/Dockerfile
[2]: https://github.com/rust-lang/llvm/commit/be89e4b5
[3]: https://bugs.webkit.org/show_bug.cgi?id=138420
danger of inference variables floating around without their inference
context.
The main insight here is that, when we are translating substitutions
between two impls, *we already know that the more specific impl holds*,
so we do not need to add its obligations to the parameter
environment. Instead, we can just thread through the inference context
we used to show select the more specific impl in the first place.
projection sensitive to "mode" (most importantly, trans vs middle).
This commit introduces several pieces of iteration infrastructure in the
specialization graph data structure, as well as various helpers for
finding the definition of a given item, given its kind and name.
In addition, associated type projection is now *mode-sensitive*, with
three possible modes:
- **Topmost**. This means that projection is only possible if there is a
non-`default` definition of the associated type directly on the
selected impl. This mode is a bit of a hack: it's used during early
coherence checking before we have built the specialization
graph (and therefore before we can walk up the specialization
parents to find other definitions). Eventually, this should be
replaced with a less "staged" construction of the specialization
graph.
- **AnyFinal**. Projection succeeds for any non-`default` associated
type definition, even if it is defined by a parent impl. Used
throughout typechecking.
- **Any**. Projection always succeeds. Used by trans.
The lasting distinction here is between `AnyFinal` and `Any` -- we wish
to treat `default` associated types opaquely for typechecking purposes.
In addition to the above, the commit includes a few other minor review fixes.
This commit leverages the specialization graph infrastructure to allow
specializing trait implementations to leave off associated types for
which their parents have provided defaults.
It also modifies the type projection code to avoid projecting associated
types unless either (1) all input types are fully known or (2) the
available associated type is "final", i.e. not marked `default`.
This restriction is required for soundness, due to examples like:
```rust
trait Foo {
type Assoc;
}
impl<T> Foo for T {
default type Assoc = ();
}
impl Foo for u8 {
type Assoc = String;
}
fn generic<T>() -> <T as Foo>::Assoc {
() //~ ERROR
}
fn main() {
let s: String = generic::<u8>();
println!("{}", s); // bad news
}
```
This commit leverages the specialization graph infrastructure to allow
specializing trait implementations to leave off methods for which their
parents have provided defaults.
It does not yet check that the `default` keyword is appropriately used
in such cases.
- Rewrites the overlap checker to instead build up a specialization
graph, checking for overlap errors in the process.
- Use the specialization order during impl selection.
This commit does not yet handle associated types correctly, and assumes
that all items are `default` and are overridden.
The module contains a few important components:
- The `specialize` function, which determines whether one impl is a
specialization of another.
- The `SpecializationGraph`, a per-trait graph recording the
specialization tree. The main purpose of the graph is to allow
traversals upwards (to less specialized impls) for discovering
un-overridden defaults, and for ensuring that overridden items are
allowed to be overridden.
The facet of a stage is rarely relevant when running a tool or building
something, it's all a question of what stage the *compiler* is built in. We've
already got a nice handy `Compiler` structure to carry this information, so
let's use it!
This refactors the signature of the `Build::cargo` function two ways:
1. The `stage` argument is removed, this was just duplicated with the `compiler`
argument's stage field.
2. The `target` argument is now required. This was a bug where if the `--target`
flag isn't passed then the snapshot stage0 compiler is always used, so we
won't pick up any changes.
Much of the other changes in this commit are just propagating these decisions
outwards. For example many of the `Step` variants no longer have a stage
argument as they're baked into the compiler.
Unfortunately on i686-pc-windows-gnu LLVM's answer to `--host-target` is
`x86_64-pc-windows-gnu` even though we're building in a 32-bit shell as well as
compiling 32-bit libraries. For now use Cargo's `HOST` environment variable to
determine whether we're doing a cross compilation or not.
When deriving Hash, RustcEncodable and RustcDecodable, the syntax extension
needs a type parameter to use in the inner method. They used to use __H, __S
and __D respectively. If this conflicts with a type parameter already declared
for the item, bad times result (see the test). There is no hygiene for type
parameters, but this commit introduces a better heuristic by concatenating the
names of all extant type parameters (and prepending __H).
This changes local variable names in all derives to remove leading
double-underscores. As far as I can tell, this doesn't break anything
because there is no user code in these generated functions except for
struct, field and type parameter names, and this doesn't cause shadowing
of those. But I am still a bit nervous.
This replaces some `if`s with `match`es. This was originally not possible
because using a global path in a match statement caused a "non-constant
path in constant expr" ICE. The issue is long since closed, though you still
hit it (as an error now, not an ICE) if you try to generate match patterns
using pat_lit(expr_path). But it works when constructing the patterns more
carefully.
