LLVM older that 3.6 has a bug that cause assertions when compiling certain
constructs. For 3.5 there's still a chance that the bug might get fixed
in 3.5.2, so let's keep allowing to compile with it for it for now.
Previously it would fail on a trivial case like
/// Summary line
/// <trailing space>
/// Regular content
Compliant markdown preprocessor would render that as two separate paragraphs, but our summary line
extractor interprets both lines as the same paragraph and includes both into the short summary resulting in
This adds search by type (for functions/methods) support to Rustdoc. Target issue is at https://github.com/rust-lang/rfcs/issues/658.
I've described my approach here: https://github.com/rust-lang/rfcs/issues/658#issuecomment-76484200. I'll copy the text in here as well:
---
Hi, it took me longer than I wished, but I have implemented this in a not-too-complex way that I think can be extended to support more complex features (like the ones mentioned [here](https://github.com/rust-lang/rust/issues/12866#issuecomment-66945317)).
The idea is to generate a JSON representation of the types of methods/functions in the existing index, and then make the JS understand when it should look by type (and not by name).
I tried to come up with a JSON representation that can be extended to support generics, bounds, ref/mut annotations and so on. Here are a few samples:
Function:
```rust
fn to_uppercase(c: char) -> char
```
```json
{
"inputs": [
{"name": "char"}
],
"output": {
"name": "char",
}
}
```
Method (implemented or defined in trait):
```rust
// in struct Vec
// self is considered an argument as well
fn capacity(&self) -> usize
```
```json
{
"inputs": [
{"name": "vec"}
],
"output": {
"name": "usize"
}
}
```
This simple format can be extended by adding more fields, like `generic: bool`, a `bounds` mapping and so on.
I have a working implementation in https://github.com/rust-lang/rust/compare/master...mihneadb:rustdoc-search-by-type. You can check out a live demo [here](http://data.mihneadb.net/doc/std/index.html?search=charext%20-%3E%20char).
The feature list is not that long:
- search by types (you *can* use generics as well, as long as you use the exact name - e.g. [`vec,t -> `](http://data.mihneadb.net/doc/std/index.html?search=vec%2C%20t%20-%3E))
- order of arguments does not matter
- `self` is took into account as well (e.g. search for `vec -> usize`)
- does not use "complex" annotations (e.g. you don't search for `&char -> char` but for `char -> char`)
My goal is to get a working, minimal "base" merged so that others can build upon it. How should I proceed? Do I open a PR (badly in need of code review since this is my first non "hello world"-ish rust code)?
---
LLVM older that 3.6 has a bug that cause assertions when compiling certain
constructs. For 3.5 there's still a chance that the bug might get fixed
in 3.5.2, so let's keep allowing to compile with it for it for now.
Previously it would fail on a trivial case like
/// Summary line
/// <trailing space>
/// Regular content
Compliant markdown preprocessor would render that as two separate paragraphs, but our summary line
extractor would interpret both lines as the same paragraph and include both into the short summary.
at least that's what the docs say: http://doc.rust-lang.org/std/slice/fn.from_raw_parts.html
A few situations got prettier. In some situations the mutability of the resulting and source pointers differed (and was cast away by transmute), the mutability matches now.
It wasn't clear to me that early_error was correct here, but it seems to
work. This code is reachable from `rustdoc`, which is problematic, because
early_error panics. rustc handles the panics gracefully (without ICEing or
crashing), but rustdoc does not. It's not the first such rustdoc problem,
though:
$ rustdoc hello.rs --extern std=bad-std
error: extern location for std does not exist: bad-std
hello.rs:1:1: 1:1 error: can't find crate for `std`
hello.rs:1
^
error: aborting due to 2 previous errors
thread '<unnamed>' panicked at 'Box<Any>', /home/rustbuild/src/rust-buildbot/slave/nightly-dist-rustc-linux/build/src/libsyntax/diagnostic.rs:151
thread '<unnamed>' panicked at 'called `Result::unwrap()` on an `Err` value: "rustc failed"', /home/rustbuild/src/rust-buildbot/slave/nightly-dist-rustc-linux/build/src/libcore/result.rs:744
thread '<main>' panicked at 'child thread None panicked', /home/rustbuild/src/rust-buildbot/slave/nightly-dist-rustc-linux/build/src/libstd/thread.rs:661
Reduce code size overhead from core::panicking::panic
core::panicking::panic currently creates an Arguments structure using
format_args!("{}", expr), which formats the expr str using the Display::fmt.
Display::fmt pulls in Formatter::pad, which then also pulls in string-related
code for truncation and padding.
If core::panicking::panic instead creates an Arguments structure with a string
piece, it is possible that the Display::fmt function for str can be optimized
out of the program.
In my testing with a 32-bit x86 bare metal program, the change tended to save
between ~100 bytes and ~5500 bytes, depending on what other panic* functions
the program invokes and whether the panic_fmt lang item uses the Arguments
value.
LLVM_LIBDIR_<triple> is only defined for host triples, not target triples.
FWIW, the same is true for LLVM_STDCPP_RUSTFLAGS_<triple>, where we
explicitly define it as empty when --enable-llvm-static-stdcpp is not
specified, but it's still undefined for cross-compiled triples.
The new `std::io` module has had some time to bake now, and this commit
stabilizes its functionality. There are still portions of the module which
remain unstable, and below contains a summart of the actions taken.
This commit also deprecates the entire contents of the `old_io` module in a
blanket fashion. All APIs should now have a reasonable replacement in the
new I/O modules.
Stable APIs:
* `std::io` (the name)
* `std::io::prelude` (the name)
* `Read`
* `Read::read`
* `Read::{read_to_end, read_to_string}` after being modified to return a `usize`
for the number of bytes read.
* `ReadExt`
* `Write`
* `Write::write`
* `Write::{write_all, write_fmt}`
* `WriteExt`
* `BufRead`
* `BufRead::{fill_buf, consume}`
* `BufRead::{read_line, read_until}` after being modified to return a `usize`
for the number of bytes read.
* `BufReadExt`
* `BufReader`
* `BufReader::{new, with_capacity}`
* `BufReader::{get_ref, get_mut, into_inner}`
* `{Read,BufRead} for BufReader`
* `BufWriter`
* `BufWriter::{new, with_capacity}`
* `BufWriter::{get_ref, get_mut, into_inner}`
* `Write for BufWriter`
* `IntoInnerError`
* `IntoInnerError::{error, into_inner}`
* `{Error,Display} for IntoInnerError`
* `LineWriter`
* `LineWriter::{new, with_capacity}` - `with_capacity` was added
* `LineWriter::{get_ref, get_mut, into_inner}` - `get_mut` was added)
* `Write for LineWriter`
* `BufStream`
* `BufStream::{new, with_capacities}`
* `BufStream::{get_ref, get_mut, into_inner}`
* `{BufRead,Read,Write} for BufStream`
* `stdin`
* `Stdin`
* `Stdin::lock`
* `Stdin::read_line` - added method
* `StdinLock`
* `Read for Stdin`
* `{Read,BufRead} for StdinLock`
* `stdout`
* `Stdout`
* `Stdout::lock`
* `StdoutLock`
* `Write for Stdout`
* `Write for StdoutLock`
* `stderr`
* `Stderr`
* `Stderr::lock`
* `StderrLock`
* `Write for Stderr`
* `Write for StderrLock`
* `io::Result`
* `io::Error`
* `io::Error::last_os_error`
* `{Display, Error} for Error`
Unstable APIs:
(reasons can be found in the commit itself)
* `Write::flush`
* `Seek`
* `ErrorKind`
* `Error::new`
* `Error::from_os_error`
* `Error::kind`
Deprecated APIs
* `Error::description` - available via the `Error` trait
* `Error::detail` - available via the `Display` implementation
* `thread::Builder::{stdout, stderr}`
Changes in functionality:
* `old_io::stdio::set_stderr` is now a noop as the infrastructure for printing
backtraces has migrated to `std::io`.
[breaking-change]
This implements a wish suggested in #17410, detecting enum variants that are never constructed, even in the presence of `#[derive(Clone)]`. The implementation is general and not specific to `#[derive(Clone)]`.
r? @jakub-
When this attribute is applied to a function, its return value gets the
noalias attribute, which is how you tell LLVM that the function returns
a \"new\" pointer that doesn't alias anything accessible to the caller,
i.e. it acts like a memory allocator.
Plain malloc doesn't need this attribute because LLVM already knows
about malloc and adds the attribute itself.
Show how to get a pointer without destroying the box.
Use `boxed::into_raw` instead of `mem::transmute`.
I removed the `let my_num: *const i32 = mem::transmute(my_num);` case altogether because we own the box, a `*mut` pointer is good anywhere a `*const` is needed, `from_raw` takes a mutable pointer and casting from a `*const` caused an ICE.
The new `std::io` module has had some time to bake now, and this commit
stabilizes its functionality. There are still portions of the module which
remain unstable, and below contains a summart of the actions taken.
This commit also deprecates the entire contents of the `old_io` module in a
blanket fashion. All APIs should now have a reasonable replacement in the
new I/O modules.
Stable APIs:
* `std::io` (the name)
* `std::io::prelude` (the name)
* `Read`
* `Read::read`
* `Read::{read_to_end, read_to_string}` after being modified to return a `usize`
for the number of bytes read.
* `Write`
* `Write::write`
* `Write::{write_all, write_fmt}`
* `BufRead`
* `BufRead::{fill_buf, consume}`
* `BufRead::{read_line, read_until}` after being modified to return a `usize`
for the number of bytes read.
* `BufReader`
* `BufReader::{new, with_capacity}`
* `BufReader::{get_ref, get_mut, into_inner}`
* `{Read,BufRead} for BufReader`
* `BufWriter`
* `BufWriter::{new, with_capacity}`
* `BufWriter::{get_ref, get_mut, into_inner}`
* `Write for BufWriter`
* `IntoInnerError`
* `IntoInnerError::{error, into_inner}`
* `{Error,Display} for IntoInnerError`
* `LineWriter`
* `LineWriter::{new, with_capacity}` - `with_capacity` was added
* `LineWriter::{get_ref, get_mut, into_inner}` - `get_mut` was added)
* `Write for LineWriter`
* `BufStream`
* `BufStream::{new, with_capacities}`
* `BufStream::{get_ref, get_mut, into_inner}`
* `{BufRead,Read,Write} for BufStream`
* `stdin`
* `Stdin`
* `Stdin::lock`
* `Stdin::read_line` - added method
* `StdinLock`
* `Read for Stdin`
* `{Read,BufRead} for StdinLock`
* `stdout`
* `Stdout`
* `Stdout::lock`
* `StdoutLock`
* `Write for Stdout`
* `Write for StdoutLock`
* `stderr`
* `Stderr`
* `Stderr::lock`
* `StderrLock`
* `Write for Stderr`
* `Write for StderrLock`
* `io::Result`
* `io::Error`
* `io::Error::last_os_error`
* `{Display, Error} for Error`
Unstable APIs:
(reasons can be found in the commit itself)
* `Write::flush`
* `Seek`
* `ErrorKind`
* `Error::new`
* `Error::from_os_error`
* `Error::kind`
Deprecated APIs
* `Error::description` - available via the `Error` trait
* `Error::detail` - available via the `Display` implementation
* `thread::Builder::{stdout, stderr}`
Changes in functionality:
* `old_io::stdio::set_stderr` is now a noop as the infrastructure for printing
backtraces has migrated to `std::io`.
* The `ReadExt`, `WriteExt`, and `BufReadExt` extension traits were all removed
by folding functionality into the corresponding trait.
[breaking-change]
When this attribute is applied to a function, its return value gets the
noalias attribute, which is how you tell LLVM that the function returns
a "new" pointer that doesn't alias anything accessible to the caller,
i.e. it acts like a memory allocator.
Plain malloc doesn't need this attribute because LLVM already knows
about malloc and adds the attribute itself.
This commit stabilizes essentially all of the new `std::path` API. The
API itself is changed in a couple of ways (which brings it in closer
alignment with the RFC):
* `.` components are now normalized away, unless they appear at the
start of a path. This in turn effects the semantics of e.g. asking for
the file name of `foo/` or `foo/.`, both of which yield `Some("foo")`
now. This semantics is what the original RFC specified, and is also
desirable given early experience rolling out the new API.
* The `parent` method is now `without_file` and succeeds if, and only
if, `file_name` is `Some(_)`. That means, in particular, that it fails
for a path like `foo/../`. This change affects `pop` as well.
In addition, the `old_path` module is now deprecated.
[breaking-change]
r? @alexcrichton
This commit stabilizes essentially all of the new `std::path` API. The
API itself is changed in a couple of ways (which brings it in closer
alignment with the RFC):
* `.` components are now normalized away, unless they appear at the
start of a path. This in turn effects the semantics of e.g. asking for
the file name of `foo/` or `foo/.`, both of which yield `Some("foo")`
now. This semantics is what the original RFC specified, and is also
desirable given early experience rolling out the new API.
* The `parent` function now succeeds if, and only if, the path has at
least one non-root/prefix component. This change affects `pop` as
well.
* The `Prefix` component now involves a separate `PrefixComponent`
struct, to better allow for keeping both parsed and unparsed prefix data.
In addition, the `old_path` module is now deprecated.
Closes#23264
[breaking-change]
The end result is that common fields (id, name, attributes, etc.) are stored in now-structures `ImplItem` and `TraitItem`.
The signature of a method is no longer duplicated between methods with a body (default/impl) and those without, they now share `MethodSig`.
This is also a [breaking-change] because of minor bugfixes and changes to syntax extensions:
* `pub fn` methods in a trait no longer parse - remove the `pub`, it has no meaning anymore
* `MacResult::make_methods` is now `make_impl_items` and the return type has changed accordingly
* `quote_method` is gone, because `P<ast::Method>` doesn't exist and it couldn't represent a full method anyways - could be replaced by `quote_impl_item`/`quote_trait_item` in the future, but I do hope we realize how silly that combinatorial macro expansion is and settle on a single `quote` macro + some type hints - or just no types at all (only token-trees)
r? @nikomatsakis This is necessary (hopefully also sufficient) for associated constants.
