This stops the compiler ICEing on the use of SIMD types in FFI signatures. It emits correct code for LLVM intrinsics, but I am quite unsure about the ABI handling in general so I've added a new feature gate `simd_ffi` to try to ensure people don't use it without realising there's a non-trivial risk of codegen brokenness.

Closes #20043.

Loading methods from external crates was erroneously using the type's privacy
for each method instead of each method's privacy. This commit fixes that.

Closes #21202

This commit also moves privacy to its own crate because I thought that was where the bug was. Turns out it wasn't, but it helped me iterate at least!

**The implementation is a direct adaptation of libcxx's condition_variable implementation.**

I also added a wait_timeout_with method, which matches the second overload in C++'s condition_variable. The implementation right now is kind of dumb but it works. There is an outstanding issue with it: as is it doesn't support the use case where a user doesn't care about poisoning and wants to continue through poison.

r? @alexcrichton @aturon 

I don't know if this handling of SIMD types is correct for the C ABI on
all platforms, so lets add an even finer feature gate than just the
`simd` one.

The `simd` one can be used with (relatively) little risk of complete
nonsense, the reason for it is that it is likely that things will
change. Using the types in FFI with an incorrect ABI will at best give
absolute nonsense results, but possibly cause serious breakage too, so
this is a step up in badness, hence a new feature gate.

This just compiles a test using SIMD in FFI (mostly importing LLVM
intrinsics) for almost all rustc's supported platforms, but not linking
it or running it, so there's absolutely no guarantee that this is correct.

For a call like `foo.bar()` where the method `bar` can't be resolved,
the compiler will search for traits that have methods with name `bar` to
give a more informative error, providing a list of possibilities.

Closes #7643.

With the addition of separate search paths to the compiler, it was intended that
applications such as Cargo could require a `--extern` flag per `extern crate`
directive in the source. The system can currently be subverted, however, due to
the `existing_match()` logic in the crate loader.

When loading crates we first attempt to match an `extern crate` directive
against all previously loaded crates to avoid reading metadata twice. This "hit
the cache if possible" step was erroneously leaking crates across the search
path boundaries, however. For example:

    extern crate b;
    extern crate a;

If `b` depends on `a`, then it will load crate `a` when the `extern crate b`
directive is being processed. When the compiler reaches `extern crate a` it will
use the previously loaded version no matter what. If the compiler was not
invoked with `-L crate=path/to/a`, it will still succeed.

This behavior is allowing `extern crate` declarations in Cargo without a
corresponding declaration in the manifest of a dependency, which is considered
a bug.

This commit fixes this problem by keeping track of the origin search path for a
crate. Crates loaded from the dependency search path are not candidates for
crates which are loaded from the crate search path.

**The implementation is a direct adaptation of libcxx's
condition_variable implementation.**

pthread_cond_timedwait uses the non-monotonic system clock. It's
possible to change the clock to a monotonic via pthread_cond_attr, but
this is incompatible with static initialization. To deal with this, we
calculate the timeout using the system clock, and maintain a separate
record of the start and end times with a monotonic clock to be used for
calculation of the return value.

With the addition of separate search paths to the compiler, it was intended that
applications such as Cargo could require a `--extern` flag per `extern crate`
directive in the source. The system can currently be subverted, however, due to
the `existing_match()` logic in the crate loader.

When loading crates we first attempt to match an `extern crate` directive
against all previously loaded crates to avoid reading metadata twice. This "hit
the cache if possible" step was erroneously leaking crates across the search
path boundaries, however. For example:

    extern crate b;
    extern crate a;

If `b` depends on `a`, then it will load crate `a` when the `extern crate b`
directive is being processed. When the compiler reaches `extern crate a` it will
use the previously loaded version no matter what. If the compiler was not
invoked with `-L crate=path/to/a`, it will still succeed.

This behavior is allowing `extern crate` declarations in Cargo without a
corresponding declaration in the manifest of a dependency, which is considered
a bug.

This commit fixes this problem by keeping track of the origin search path for a
crate. Crates loaded from the dependency search path are not candidates for
crates which are loaded from the crate search path.

As a result of this fix, this is a likely a breaking change for a number of
Cargo packages. If the compiler starts informing that a crate can no longer be
found, it likely means that the dependency was forgotten in your Cargo.toml.

[breaking-change]

Loading methods from external crates was erroneously using the type's privacy
for each method instead of each method's privacy. This commit fixes that.

Closes #21202

This fixes #16530 by hashing nullary structs [the same way as the empty tuple] (https://github.com/rust-lang/rust/blob/master/src/libcore/hash/mod.rs#L185). Other approaches are possible, but this was the simplest.

This PR adds rules for negative implementations. It follows pretty much what the [RFC](https://github.com/rust-lang/rfcs/blob/master/text/0019-opt-in-builtin-traits.md) says with 1 main difference:

Instead of positive implementations override negative implementations, this have been implemented in a way that a negative implementation of `Trait` for `T` will overlap with a positive implementation, causing a coherence error.

@nikomatsakis r?

cc #13231

[breaking-change]

If `a.method();` can't be resolved, we first look for implemented traits
globally and suggest those. If there are no such traits found, we only
then fall back to suggesting from the unfiltered list of traits.

This seems to match what clang does on arm, but I cannot do any
experimentation with mips, but it matches how the LLVM intrinsics are
defined in any case...

Unlike the intrinics in C, this types the SSE values base on integer
size. This matches the LLVM intrinsics which have concrete vector types
(`<4 x i32>` etc.), and is no loss of expressivity: if one is using a C
function that really takes an untyped integral SSE value, just give it
whatever Rust type makes most sense.

This seems to work on x86-64, but I am not able to test on other
platforms.

cc #20043

Fix commented graphs in src/doc/trpl/ownership.md

Reviewed-by: huonw

Fix std::sync::condvar::Condvar::notify_one docs

Reviewed-by: alexcrichton

Two minor fixes for save-analysis

Reviewed-by: huonw