This increases regionck performance greatly - type-checking on
librustc decreased from 9.1s to 8.1s. Because of Amdahl's law,
total performance is improved only by about 1.5% (LLVM wizards,
this is your opportunity to shine!).
before:
576.91user 4.26system 7:42.36elapsed 125%CPU (0avgtext+0avgdata 1142192maxresident)k
after:
566.50user 4.84system 7:36.84elapsed 125%CPU (0avgtext+0avgdata 1124304maxresident)k
I am somewhat worried really need to find out why we have this Red Queen's
Race going on here. Originally I suspected it may be a problem from RFC1214's
warnings, but it seems to be an effect from other changes.
However, the increase seems to be mostly in LLVM's time, so I guess
it's the LLVM wizards' problem.
TyClosure variant; thread this through wherever closure substitutions
are expected, which leads to a net simplification. Simplify trans
treatment of closures in particular.
region-bound is expected to change in Rust 1.3, but don't use it for
anything in this commit. Note that this is not a "significant" part of
the type (it's not part of the formal model) so we have to normalize
this away or trans starts to get confused because two equal types wind
up with distinct LLVM types.
The former stopped making sense when we started interning substs and made
TraitRef a 2-word copy type, and I'm moving the latter into an arena as
they live as long as the type context.
Function params which outlive everything in the body (incl
temporaries). Thus if we assign them their own `CodeExtent`, the
region inference can properly show that it is sound to have
temporaries with destructors that reference the parameters (because
such temporaries will be dropped before the parameters are).
This allows us to address issue 23338 in a clean way.
As a drive-by, fix a mistake in the tyencode for
`CodeExtent::BlockRemainder`.
We try to move the data when the length can be encoded in
the much smaller number of bytes. This interferes with indices and
type abbreviations however, so this commit introduces a public
interface to get and mark a "stable" (i.e. not affected by
relaxation) position of the current pointer.
The relaxation logic only moves a small data, currently at most
256 bytes, as moving the data can be costly. There might be
further opportunities to allow more relaxation by moving fields
around, which I didn't seriously try.
type-outlives works for closure types so that it ensures that all upvars
outlive the region in question. This gives the same guarantees but
without introducing artificial regions (and gives better error messages
to boot).
immediately surrounding a node that is a terminating_scope
(e.g. statements, looping forms) during which the destructors run (the
destructors for temporaries from the execution of that node, that is).
Introduced DestructionScopeData newtype wrapper around ast::NodeId, to
preserve invariant that FreeRegion and ScopeChain::BlockScope carry
destruction scopes (rather than arbitrary CodeExtents).
Insert DestructionScope and block Remainder into enclosing CodeExtents
hierarchy.
Add more doc for DestructionScope, complete with ASCII art.
Switch to constructing DestructionScope rather than Misc in a number
of places, mostly related to `ty::ReFree` creation, and use
destruction-scopes of node-ids at various calls to
liberate_late_bound_regions.
middle::resolve_lifetime: Map BlockScope to DestructionScope in `fn resolve_free_lifetime`.
Add the InnermostDeclaringBlock and InnermostEnclosingExpr enums that
are my attempt to clarify the region::Context structure, and that
later commmts build upon.
Improve the debug output for `CodeExtent` attached to `ty::Region::ReScope`.
Loosened an assertion in `rustc_trans::trans::cleanup` to account for
`DestructionScope`. (Perhaps this should just be switched entirely
over to `DestructionScope`, rather than allowing for either `Misc` or
`DestructionScope`.)
----
Even though the DestructionScope is new, this particular commit should
not actually change the semantics of any current code.
This new variant introduces finer-grain code extents, i.e. we now
track that a binding lives only for a suffix of a block, and
(importantly) will be dropped when it goes out of scope *before* the
bindings that occurred earlier in the block.
Both of these notions are neatly captured by marking the block (and
each suffix) as an enclosing scope of the next suffix beneath it.
This is work that is part of the foundation for issue #8861.
(It actually has been seen in earlier posted pull requests; I have
just factored it out into its own PR to ease my own rebasing.)
----
These finer grained scopes do mean that some code is newly rejected by
`rustc`; for example:
```rust
let mut map : HashMap<u8, &u8> = HashMap::new();
let tmp = Box::new(2);
map.insert(43, &*tmp);
```
This will now fail to compile with a message that `*tmp` does not live
long enough, because the scope of `tmp` is now strictly smaller than
that of `map`, and the use of `&u8` in map's type requires that the
borrowed references are all to data that live at least as long as the
map.
The usual fix for a case like this is to move the binding for `tmp`
up above that of `map`; note that you can still leave the initialization
in the original spot, like so:
```rust
let tmp;
let mut map : HashMap<u8, &u8> = HashMap::new();
tmp = box 2;
map.insert(43, &*tmp);
```
Similarly, one can encounter an analogous situation with `Vec`: one
would need to rewrite:
```rust
let mut vec = Vec::new();
let tmp = 'c';
vec.push(&tmp);
```
as:
```
let tmp;
let mut vec = Vec::new();
tmp = 'c';
vec.push(&tmp);
```
----
In some corner cases, it does not suffice to reorder the bindings; in
particular, when the types for both bindings need to reflect exactly
the *same* code extent, and a parent/child relationship between them
does not work.
In pnkfelix's experience this has arisen most often when mixing uses
of cyclic data structures while also allowing a lifetime parameter
`'a` to flow into a type parameter context where the type is
*invariant* with respect to the type parameter. An important instance
of this is `arena::TypedArena<T>`, which is invariant with respect
to `T`.
(The reason that variance is relevant is this: *if* `TypedArena` were
covariant with respect to its type parameter, then we could assign it
the longer lifetime when it is initialized, and then convert it to a
subtype (via covariance) with a shorter lifetime when necessary. But
`TypedArena` is invariant with respect to its type parameter, and thus
if `S` is a subtype of `T` (in particular, if `S` has a lifetime
parameter that is shorter than that of `T`), then a `TypedArena<S>` is
unrelated to `TypedArena<T>`.)
Concretely, consider code like this:
```rust
struct Node<'a> { sibling: Option<&'a Node<'a>> }
struct Context<'a> {
// because of this field, `Context<'a>` is invariant with respect to `'a`.
arena: &'a TypedArena<Node<'a>>,
...
}
fn new_ctxt<'a>(arena: &'a TypedArena<Node<'a>>) -> Context<'a> { ... }
fn use_ctxt<'a>(fcx: &'a Context<'a>) { ... }
let arena = TypedArena::new();
let ctxt = new_ctxt(&arena);
use_ctxt(&ctxt);
```
In these situations, if you try to introduce two bindings via two
distinct `let` statements, each is (with this commit) assigned a
distinct extent, and the region inference system cannot find a single
region to assign to the lifetime `'a` that works for both of the
bindings. So you get an error that `ctxt` does not live long enough;
but moving its binding up above that of `arena` just shifts the error
so now the compiler complains that `arena` does not live long enough.
SO: What to do? The easiest fix in this case is to ensure that the two
bindings *do* get assigned the same static extent, by stuffing both
bindings into the same let statement, like so:
```rust
let (arena, ctxt): (TypedArena, Context);
arena = TypedArena::new();
ctxt = new_ctxt(&arena);
use_ctxt(&ctxt);
```
Due to the new code rejections outlined above, this is a ...
[breaking-change]
In preparation for upcoming changes to the `Writer` trait (soon to be called
`Write`) this commit renames the current `write` method to `write_all` to match
the semantics of the upcoming `write_all` method. The `write` method will be
repurposed to return a `usize` indicating how much data was written which
differs from the current `write` semantics. In order to head off as much
unintended breakage as possible, the method is being deprecated now in favor of
a new name.
[breaking-change]
1. Produced more unique types than is necessary. This increases memory consumption.
2. Linking the type parameter to its definition *seems* like a good idea, but it
encourages reliance on the bounds listing.
3. It made pretty-printing harder and in particular was causing bad error messages
when errors occurred before the `TypeParameterDef` entries were fully stored.
