Implement always-fallible TryFrom for usize/isize conversions that are infallible on some platforms
This reverts commit 837d6c7023 "Remove TryFrom impls that might become conditionally-infallible with a portability lint".
This fixes#49415 by adding (restoring) missing `TryFrom` impls for integer conversions to or from `usize` or `isize`, by making them always fallible at the type system level (that is, with `Error=TryFromIntError`) even though they happen to be infallible on some platforms (for some values of `size_of::<usize>()`).
They had been removed to allow the possibility to conditionally having some of them be infallible `From` impls instead, depending on the platforms, and have the [portability lint](https://github.com/rust-lang/rfcs/pull/1868) warn when they are used in code that is not already opting into non-portability. For example `#[allow(some_lint)] usize::from(x: u64)` would be valid on code that only targets 64-bit platforms.
This PR gives up on this possiblity for two reasons:
* Based on discussion with @aturon, it seems that the portability lint is not happening any time soon. It’s better to have the conversions be available *at all* than keep blocking them for so long. Portability-lint-gated platform-specific APIs can always be added separately later.
* For code that is fine with fallibility, the alternative would force it to opt into "non-portability" even though there would be no real portability issue.
Add some performance guidance to std::fs and std::io docs
Adds more documentation about performance to various "read" functions in `fs` and `io`, and to `BufReader`/`BufWriter`, with the goal of helping developers choose the best option for a given task.
Remove "empty buffer" doc in read_until
This appears copied from fill_buf, but the above paragraph already indicates that a lack of delimiter at the end is EOF.
On MIPS, error number 98 is not EADDRINUSE (it is EPROTOTYPE). To fix the
resulting test failure this causes, use a more portable error number in
the example documentation. EINVAL shold be more reliable because it was
defined in the original Unix as 22 so hopefully most derivatives have
defined it the same way.
Better Debug impl for io::Error.
This PR includes the below changes:
1. The former impl wrapped the entire thing in `Error { repr: ... }` which was unhelpful; this has been removed.
2. The `Os` variant of `io::Error` included the code and message, but not the kind; this has been fixed.
3. The `Custom` variant of `io::Error` included a `Custom(Custom { ... })`, which is now just `Custom { ... }`.
Example of previous impl:
```rust
Error {
repr: Custom(
Custom {
kind: InvalidData,
error: Error {
repr: Os {
code: 2,
message: "no such file or directory"
}
}
}
)
}
```
Example of new impl:
```rust
Custom {
kind: InvalidData,
error: Os {
code: 2,
kind: NotFound,
message: "no such file or directory"
}
}
```
docs: do not call integer overflows as underflows
In the API docs, integer overflow is sometimes called underflow. Underflow is really when the magnitude of a floating-point number is too small so the number underflows to subnormal or zero. With integers it is always overflow, even if the expected result is less than the minimum number that can be represented.
The read_exact implementation for &[u8] is optimized and usually allows LLVM to reduce a read_exact call for small numbers of bytes to a bounds check and a register load instead of a generic memcpy. On a workload I have that decompresses, deserializes (via bincode), and processes some data, this leads to a 40% speedup by essentially eliminating the deserialization overhead entirely.
Optimize `read_to_end`.
This patch makes `read_to_end` use Vec's memory-growth pattern rather
than using a custom pattern.
This has some interesting effects:
- If memory is reserved up front, `read_to_end` can be faster, as it
starts reading at the buffer size, rather than always starting at 32
bytes. This speeds up file reading by 2x in one of my use cases.
- It can reduce the number of syscalls when reading large files.
Previously, `read_to_end` would settle into a sequence of 8192-byte
reads. With this patch, the read size follows Vec's allocation
pattern. For example, on a 16MiB file, it can do 21 read syscalls
instead of 2057. In simple benchmarks of large files though, overall
speed is still dominated by the actual I/O.
- A downside is that Read implementations that don't implement
`initializer()` may see increased memory zeroing overhead.
I benchmarked this on a variety of data sizes, with and without
preallocated buffers. Most benchmarks see no difference, but reading
a small/medium file with a pre-allocated buffer is faster.
