default to $ARCH-apple-macosx10.7.0 LLVM triple for darwin targets
Over in #60378, we made `rustc` switch LLVM target triples dynamically
based on the `MACOSX_DEPLOYMENT_TARGET` environment variable. This
change was made to align with `clang`'s behavior, and therefore make
cross-language LTO feasible on OS X. Otherwise, `rustc` would produce
LLVM bitcode files with a target triple of `x86_64-apple-darwin`,
`clang` would produce LLVM bitcode files with a target triple of
`x86_64-apple-macosx$VERSION`, and the linker would complain.
This change worked fine, except for one corner case: if you didn't have
`MACOSX_DEPLOYMENT_TARGET` set, and you wanted to do LTO on just Rust
code, you'd get warning messages similar to:
```
warning: Linking two modules of different target triples: ' is 'x86_64-apple-macosx10.7.0' whereas 'main.7rcbfp3g-cgu.4' is 'x86_64-apple-darwin'
```
This message occurs because libstd is compiled with
`MACOSX_DEPLOYMENT_TARGET` set to 10.7. The LLVM bitcode distributed in
libstd's rlibs, then, is tagged with the target triple of
`x86_64-apple-macosx10.7.0`, while the bitcode `rustc` produces for
"user" code is tagged with the target triple of `x86_64-apple-darwin`.
It's not good to have LTO on just Rust code (probably much more common
than cross-language LTO) warn by default. These warnings also break
Cargo's testsuite.
This change defaults to acting as though `MACOSX_DEPLOYMENT_TARGET` was
set to 10.7. "user" code will then be given a target triple that is
equivalent to the target triple libstd bitcode is already using. The
above warning will therefore go away.
`rustc` already assumes that compiling without
`MACOSX_DEPLOYMENT_TARGET` means that we're compiling for a target
compatible with OS X 10.7 (e.g. that things like TLS work properly). So
this change is really just making things conform more closely to the
status quo.
(It's also worth noting that before and after this patch, compiling with
`MACOSX_DEPLOYMENT_TARGET` set to, say, 10.9, works just fine: target
triples with an "apple" version ignore OS versions when checking
compatibility, so bitcode with a `x86_64-apple-macosx10.7.0` triple works just
fine with bitcode with a `x86_64-apple-macosx10.9.0` triple.)
Allow subdirectories to be tested by x.py test
Fixes https://github.com/rust-lang/rust/issues/60718.
As far as I can tell, multiple `--test-args` flags are ignored (only the first is respected), so if you specify a subdirectory, you won't also be able to filter using `--test-args`. If you don't specify a subdirectory, `--test-args` will continue working as usual, so this is strictly an improvement on the current state of affairs.
forego caching for all participants in cycles, apart from root node
This is a targeted fix for #60010, which uncovered a pretty bad failure of our caching strategy in the face of coinductive cycles. The problem is explained in the comment in the PR on the new field, `in_cycle`, but I'll reproduce it here:
> Starts out as false -- if, during evaluation, we encounter a
> cycle, then we will set this flag to true for all participants
> in the cycle (apart from the "head" node). These participants
> will then forego caching their results. This is not the most
> efficient solution, but it addresses #60010. The problem we
> are trying to prevent:
>
> - If you have `A: AutoTrait` requires `B: AutoTrait` and `C: NonAutoTrait`
> - `B: AutoTrait` requires `A: AutoTrait` (coinductive cycle, ok)
> - `C: NonAutoTrait` requires `A: AutoTrait` (non-coinductive cycle, not ok)
>
> you don't want to cache that `B: AutoTrait` or `A: AutoTrait`
> is `EvaluatedToOk`; this is because they were only considered
> ok on the premise that if `A: AutoTrait` held, but we indeed
> encountered a problem (later on) with `A: AutoTrait. So we
> currently set a flag on the stack node for `B: AutoTrait` (as
> well as the second instance of `A: AutoTrait`) to supress
> caching.
>
> This is a simple, targeted fix. The correct fix requires
> deeper changes, but would permit more caching: we could
> basically defer caching until we have fully evaluated the
> tree, and then cache the entire tree at once.
I'm not sure what the impact of this fix will be in terms of existing crates or performance: we were accepting incorrect code before, so there will perhaps be some regressions, and we are now caching less.
As the comment above notes, we could do a lot better than this fix, but that would involve more invasive rewrites. I thought it best to start with something simple.
r? @pnkfelix -- but let's do crater/perf run
cc @arielb1
as_ptr returns a read-only pointer
Add comments to `as_ptr` methods to warn that these are read-only pointers, and writing to them is UB.
[It was pointed out](https://internals.rust-lang.org/t/as-ptr-vs-as-mut-ptr/9940) that `CStr` does not even have an `as_mut_ptr`. I originally was going to add one, but there is no method at all that would mutate a `CStr`. Was that a deliberate choice or should I add an `as_mut_ptr` (similar to [what I did for `str`](https://github.com/rust-lang/rust/pull/58200))?
Add implementations of last in terms of next_back on a bunch of DoubleEndedIterators
Provided a `DoubleEndedIterator` has finite length, `Iterator::last` is equivalent to `DoubleEndedIterator::next_back`. But searching forwards through the iterator when it's unnecessary is obviously not good for performance. I ran into this on one of the collection iterators.
I tried adding appropriate overloads for a bunch of the iterator adapters like filter, map, etc, but I ran into a lot of type inference failures after doing so.
The other interesting case is what to do with `Repeat`. Do we consider it part of the contract that `Iterator::last` will loop forever on it? The docs do say that the iterator will be evaluated until it returns None. This is also relevant for the adapters, it's trivially easy to observe whether a `Map` adapter invoked its closure a zillion times or just once for the last element.
Destabilize the `Error::type_id` function
This commit destabilizes the `Error::type_id` function in the standard library.
This does so by effectively reverting #58048, restoring the `#[unstable]`
attribute. The security mailing list has recently been notified of a
vulnerability relating to the stabilization of this function. First stabilized
in Rust 1.34.0, a stable function here allows users to implement a custom
return value for this function:
struct MyType;
impl Error for MyType {
fn type_id(&self) -> TypeId {
// Enable safe casting to `String` by accident.
TypeId::of::<String>()
}
}
This, when combined with the `Error::downcast` family of functions, allows
safely casting a type to any other type, clearly a memory safety issue! A
formal announcement has been made to the [security mailing list](https://groups.google.com/forum/#!topic/rustlang-security-announcements/aZabeCMUv70) as well as [the blog](https://blog.rust-lang.org/2019/05/13/Security-advisory.html)
This commit simply destabilizes the `Error::type_id` which, although breaking
for users since Rust 1.34.0, is hoped to have little impact and has been deemed
sufficient to mitigate this issue for the stable channel. The long-term fate of
the `Error::type_id` API will be discussed at #60784.
Over in #60378, we made `rustc` switch LLVM target triples dynamically
based on the `MACOSX_DEPLOYMENT_TARGET` environment variable. This
change was made to align with `clang`'s behavior, and therefore make
cross-language LTO feasible on OS X. Otherwise, `rustc` would produce
LLVM bitcode files with a target triple of `x86_64-apple-darwin`,
`clang` would produce LLVM bitcode files with a target triple of
`x86_64-apple-macosx$VERSION`, and the linker would complain.
This change worked fine, except for one corner case: if you didn't have
`MACOSX_DEPLOYMENT_TARGET` set, and you wanted to do LTO on just Rust
code, you'd get warning messages similar to:
```
warning: Linking two modules of different target triples: ' is 'x86_64-apple-macosx10.7.0' whereas 'main.7rcbfp3g-cgu.4' is 'x86_64-apple-darwin'
```
This message occurs because libstd is compiled with
`MACOSX_DEPLOYMENT_TARGET` set to 10.7. The LLVM bitcode distributed in
libstd's rlibs, then, is tagged with the target triple of
`x86_64-apple-macosx10.7.0`, while the bitcode `rustc` produces for
"user" code is tagged with the target triple of `x86_64-apple-darwin`.
It's not good to have LTO on just Rust code (probably much more common
than cross-language LTO) warn by default. These warnings also break
Cargo's testsuite.
This change defaults to acting as though `MACOSX_DEPLOYMENT_TARGET` was
set to 10.7. "user" code will then be given a target triple that is
equivalent to the target triple libstd bitcode is already using. The
above warning will therefore go away.
`rustc` already assumes that compiling without
`MACOSX_DEPLOYMENT_TARGET` means that we're compiling for a target
compatible with OS X 10.7 (e.g. that things like TLS work properly). So
this change is really just making things conform more closely to the
status quo.
(It's also worth noting that before and after this patch, compiling with
`MACOSX_DEPLOYMENT_TARGET` set to, say, 10.9, works just fine: target
triples with an "apple" version ignore OS versions when checking
compatibility, so bitcode with a `x86_64-apple-macosx10.7.0` triple works just
fine with bitcode with a `x86_64-apple-macosx10.9.0` triple.)
Add #[doc(hidden)] attribute on compiler generated module.
Resolves unavoidable `missing_docs` warning/error on proc-macro crates.
Resolves#42008.
Changes not yet tested locally, however I wanted to submit first since `rustc` takes forever to compile.
coretest: Downgrade deny to warn
The `deny` causes a build failure in https://github.com/RalfJung/miri-test-libstd. Since we use `-D warnings` for rustc builds, `warn` should be enough to lead to compile errors here, without impeding external builds.
This commit destabilizes the `Error::type_id` function in the standard library.
This does so by effectively reverting #58048, restoring the `#[unstable]`
attribute. The security mailing list has recently been notified of a
vulnerability relating to the stabilization of this function. First stabilized
in Rust 1.34.0, a stable function here allows users to implement a custom
return value for this function:
struct MyType;
impl Error for MyType {
fn type_id(&self) -> TypeId {
// Enable safe casting to `String` by accident.
TypeId::of::<String>()
}
}
This, when combined with the `Error::downcast` family of functions, allows
safely casting a type to any other type, clearly a memory safety issue! A
security announcement will be shortly posted to the security mailing list as
well as the Rust Blog, and when those links are available they'll be filled in
for this PR as well.
This commit simply destabilizes the `Error::type_id` which, although breaking
for users since Rust 1.34.0, is hoped to have little impact and has been deemed
sufficient to mitigate this issue for the stable channel. The long-term fate of
the `Error::type_id` API will be discussed at #60784.
Use `Symbol` more
A `Symbol` can be equated with a string (e.g. `&str`). This involves a
TLS lookup to get the chars (and a Mutex lock in a parallel compiler)
and then a char-by-char comparison. This functionality is convenient but
avoids one of the main benefits of `Symbol`s, which is fast equality
comparisons.
This PR removes the `Symbol`/string equality operations, forcing a lot
of existing string occurrences to become `Symbol`s. Fortunately, these
are almost all static strings (many are attribute names) and we can add
static `Symbol`s as necessary, and very little extra interning occurs.
The benefits are (a) a slight speedup (possibly greater in a parallel
compiler), and (b) the code is a lot more principled about `Symbol` use.
The main downside is verbosity, particularly with more `use
syntax::symbol::symbols` items.
r? @Zoxc
These will be used in the subsequent commits. Many of them are
attributes.
The commit also adds the ability to handle symbols that aren't
identifiers (e.g. "proc-macro").
This shouldn't be possible any more, but if it does happen, emit an
error rather than maybe panicking later when NLL finds a the ReScope.
Impl trait in bindings is sufficiently broken that I don't think this
breaks anything that works for it.
Keep original literal tokens in AST
The original literal tokens (`token::Lit`) are kept in AST until lowering to HIR.
The tokens are kept together with their lowered "semantic" representation (`ast::LitKind`), so the size of `ast::Lit` is increased (this also increases the size of meta-item structs used for processing built-in attributes).
However, the size of `ast::Expr` stays the same.
The intent is to remove the "semantic" representation from AST eventually and keep literals as tokens until lowering to HIR (at least), and I'm going to work on that, but it would be good to land this sooner to unblock progress on the [lexer refactoring](https://github.com/rust-lang/rust/pull/59706).
Fixes a part of https://github.com/rust-lang/rust/issues/43081 (literal tokens that are passed to proc macros are always precise, including hexadecimal numbers, strings with their original escaping, etc)
Fixes a part of https://github.com/rust-lang/rust/issues/60495 (everything except for proc macro API doesn't need escaping anymore)
This also allows to eliminate a certain hack from the lexer (https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/pretty-printing.20comments/near/165005357).
cc @matklad
Make tests compatible with musl host
As an alternative to passing explicit linker all over the place I could try linking `cc` to musl gcc since this bootstraps itself.
Assigning for discussion:
r? @alexcrichton
Document the order of {Vec,VecDeque,String}::retain
It's natural for `retain` to work in order from beginning to end, but
this wasn't actually documented to be the case. If we actually promise
this, then the caller can do useful things like track the index of each
element being tested, as [discussed in the forum][1]. This is now
documented for `Vec`, `VecDeque`, and `String`.
[1]: https://users.rust-lang.org/t/vec-retain-by-index/27697
`HashMap` and `HashSet` also have `retain`, and the `hashbrown`
implementation does happen to use a plain `iter()` order too, but it's
not certain that this should always be the case for these types.
r? @scottmcm
