Optimize RefCell refcount tracking
Address the performance concern raised in https://github.com/rust-lang/rust/pull/51466#issuecomment-398255276
cc @dtolnay @nnethercote @rust-lang/wg-compiler-performance
cc @RalfJung @jhjourdan for soundness concerns
Can somebody kick off a perf run on this? I'm not sure how that's done, but I understand it has to be started manually.
The idea of this change is to switch to representing mutable refcount as values below 0 to eliminate some branching that was required with the old algorithm.
Emscripten changed the default behavior recently:
bd050e64bb/ChangeLog.markdown (v1381-05172018)
It now defaults to WebAssembly and requires an explicit flag to generate asm.js.
WASM=0 is also valid for older emcc and thus doesn't break it.
Closes#51856
Make FileMap::{lines, multibyte_chars, non_narrow_chars} non-mutable.
This PR removes most of the interior mutability from `FileMap`, which should be beneficial, especially in a multithreaded setting. This is achieved by initializing the state in question when the filemap is constructed instead of during lexing. Hopefully this doesn't degrade performance.
cc @wesleywiser
The method relied on the FileMap still being under construction in
order for it to do what the name promises. It's now independent of
the current state.
Namely: labels, type parameters, bindings in patterns, parameter names in functions without body.
All of these do not need hygiene after lowering to HIR, only span locations.
Suggestion for 'static impl Trait return
When encountering a named or anonymous sup requirement (for example,
`&'a self`) and a `'static` impl Trait return type, suggest adding the
`'_` lifetime constraing to the return type.
Fix#43719, #51282.
```
error: cannot infer an appropriate lifetime
--> $DIR/static-return-lifetime-infered.rs:17:16
|
LL | fn iter_values_anon(&self) -> impl Iterator<Item=u32> {
| ----------------------- this return type evaluates to the `'static` lifetime...
LL | self.x.iter().map(|a| a.0)
| ------ ^^^^
| |
| ...but this borrow...
|
note: ...can't outlive the anonymous lifetime #1 defined on the method body at 16:5
--> $DIR/static-return-lifetime-infered.rs:16:5
|
LL | / fn iter_values_anon(&self) -> impl Iterator<Item=u32> {
LL | | self.x.iter().map(|a| a.0)
LL | | }
| |_____^
help: you can add a constraint to the return type to make it last less than `'static` and match the anonymous lifetime #1 defined on the method body at 16:5
|
LL | fn iter_values_anon(&self) -> impl Iterator<Item=u32> + '_ {
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
```
convert NLL ops to caches
This is a extension of <https://github.com/rust-lang/rust/pull/51460>. It uses a lot more caching than we used to do. This caching is not yet as efficient as it could be, but I'm curious to see the current perf results.
This is the high-level idea: in the MIR type checker, use [canonicalized queries](https://rust-lang-nursery.github.io/rustc-guide/traits/canonical-queries.html) for all the major operations. This is helpful because the MIR type check is operating in a context where all types are fully known (mostly, anyway) but regions are completely renumbered. This means we often wind up with duplicate queries like `Foo<'1, '2> :Bar` and `Foo<'3, '4>: Bar`. Canonicalized queries let us re-use the results. By the final commit in this PR, we can essentially just "read off" the resulting region relations and add them to the NLL type check.
Rollup of 7 pull requests
Successful merges:
- #49987 (Add str::split_ascii_whitespace.)
- #50342 (Document round-off error in `.mod_euc()`-method, see issue #50179)
- #51658 (Only do sanity check with debug assertions on)
- #51799 (Lower case some feature gate error messages)
- #51800 (Add a compiletest header for edition)
- #51824 (Fix the error reference for LocalKey::try_with)
- #51842 (Document that Layout::from_size_align does not allow align=0)
Failed merges:
r? @ghost
Fix the error reference for LocalKey::try_with
There's no such thing as `ThreadLocalError` and the method obviously returns `AccessError`, so adjusting (probably only outdated docs).
Add a compiletest header for edition
r? @nikomatsakis
Are the `-Zunstable-options` options needed in these tests? It looks like they aren't. If not, I can remove them.
Only do sanity check with debug assertions on
r? @nnethercote
I'm slighty confused. These changes address code that the `unused-warnings` benchmark doesn't go through, yet I see a 5% improvement to nightly on the `check` run, and no improvement on the other runs.
Maybe this change allows unrelated code in the same function to be better optimized?
Document round-off error in `.mod_euc()`-method, see issue #50179
Due to a round-off error the method `.mod_euc()` of both `f32` and `f64` can produce mathematical invalid outputs. If `self` in magnitude is much small than the modulus `rhs` and negative, `self + rhs` in the first branch cannot be represented in the given precision and results into `rhs`. In the mathematical strict sense, this breaks the definition. But given the limitation of floating point arithmetic it can be thought of the closest representable value to the true result, although it is not strictly in the domain `[0.0, rhs)` of the function. It is rather the left side asymptotical limit. It would be desirable that it produces the mathematical more sound approximation of `0.0`, the right side asymptotical limit. But this breaks the property, that `self == self.div_euc(rhs) * rhs + a.mod_euc(rhs)`.
The discussion in issue #50179 did not find an satisfying conclusion to which property is deemed more important. But at least we can document the behaviour. Which this pull request does.
Add str::split_ascii_whitespace.
As mentioned in #48656.
While `str::split_whitespace` now works in `libcore`, it still makes sense to offer this method, considering how it is still more performant in cases where only ASCII is necessary.
