<sup>**context:** moving back to a layered approach to type checking.</sup>
It looks like they'd not ended up tightly coupled in the time one was owned by the other. Every instance outside of `FnCtxt.inh` was from an `InferCtxt` created and dropped in the same function body.
This conflicts slightly with #30652, but there too it looks like the `FulfillmentContext` is from an `InferCtxt` that is created and dropped within the same function body (across one call to a module-private function).
That said, I heard that the PR that originally moved `FulfillmentContext` into `InferCtxt` was big, which leaves me concerned that I'm missing something.
r? @nikomatsakis
Why do this: The RegionGraph representation previously conflated all
of the non-variable regions (i.e. the concrete regions such as
lifetime parameters to the current function) into a single dummy node.
A single dummy node leads DFS on a graph `'a -> '_#1 -> '_#0 -> 'b` to
claim that `'_#1` is reachable from `'_#0` (due to `'a` and `'b` being
conflated in the graph representation), which is incorrect (and can
lead to soundness bugs later on in compilation, see #30438).
Splitting the dummy node ensures that DFS will never introduce new
ancestor relationships between nodes for variable regions in the
graph.
Issue #31109 uncovered two semi-related problems:
* A panic in `str::parse::<f64>`
* A panic in `rustc::middle::const_eval::lit_to_const` where the result of float parsing was unwrapped.
This series of commits fixes both issues and also drive-by-fixes some things I noticed while tracking down the parsing panic.
The scope of these refactorings is a little bit bigger than the title implies. See each commit for details.
I’m submitting this for nitpicking now (the first 4 commits), because I feel the basic idea/implementation is sound and should work. I will eventually expand this PR to cover the translator changes necessary for all this to work (+ tests), ~~and perhaps implement a dynamic dropping scheme while I’m at it as well.~~
r? @nikomatsakis
This change also modifies the dep graph infrastructure to keep track of the number of active tasks, so that even if we are not building the full dep-graph, we still get assertions when there is no active task and one does something that would add a read/write edge. This is particularly helpful since, if the assertions are *not* active, you wind up with the error happening in the message processing thread, which is too late to know the correct backtrace.
~~Before landing, I need to do some performance measurements. Those are underway.~~
See measurements below. No real effect on time.
r? @michaelwoerister
Have the `ObligationForest` keep some per-tree state (or type `T`) and have it give a mutable reference for use when processing obligations. In this case, it will be a hashmap. This obviously affects the work that @soltanmm has been doing on snapshotting. I partly want to toss this out there for discussion.
Fixes#31157. (The test in question goes to approx. 30s instead of 5 minutes for me.)
cc #30977.
cc @aturon @arielb1 @soltanmm
r? @aturon who reviewed original `ObligationForest`
The purpose of the translation item collector is to find all monomorphic instances of functions, methods and statics that need to be translated into LLVM IR in order to compile the current crate.
So far these instances have been discovered lazily during the trans path. For incremental compilation we want to know the set of these instances in advance, and that is what the trans::collect module provides.
In the future, incremental and regular translation will be driven by the collector implemented here.
r? @nikomatsakis
cc @rust-lang/compiler
Translation Item Collection
===========================
This module is responsible for discovering all items that will contribute to
to code generation of the crate. The important part here is that it not only
needs to find syntax-level items (functions, structs, etc) but also all
their monomorphized instantiations. Every non-generic, non-const function
maps to one LLVM artifact. Every generic function can produce
from zero to N artifacts, depending on the sets of type arguments it
is instantiated with.
This also applies to generic items from other crates: A generic definition
in crate X might produce monomorphizations that are compiled into crate Y.
We also have to collect these here.
The following kinds of "translation items" are handled here:
- Functions
- Methods
- Closures
- Statics
- Drop glue
The following things also result in LLVM artifacts, but are not collected
here, since we instantiate them locally on demand when needed in a given
codegen unit:
- Constants
- Vtables
- Object Shims
General Algorithm
-----------------
Let's define some terms first:
- A "translation item" is something that results in a function or global in
the LLVM IR of a codegen unit. Translation items do not stand on their
own, they can reference other translation items. For example, if function
`foo()` calls function `bar()` then the translation item for `foo()`
references the translation item for function `bar()`. In general, the
definition for translation item A referencing a translation item B is that
the LLVM artifact produced for A references the LLVM artifact produced
for B.
- Translation items and the references between them for a directed graph,
where the translation items are the nodes and references form the edges.
Let's call this graph the "translation item graph".
- The translation item graph for a program contains all translation items
that are needed in order to produce the complete LLVM IR of the program.
The purpose of the algorithm implemented in this module is to build the
translation item graph for the current crate. It runs in two phases:
1. Discover the roots of the graph by traversing the HIR of the crate.
2. Starting from the roots, find neighboring nodes by inspecting the MIR
representation of the item corresponding to a given node, until no more
new nodes are found.
The roots of the translation item graph correspond to the non-generic
syntactic items in the source code. We find them by walking the HIR of the
crate, and whenever we hit upon a function, method, or static item, we
create a translation item consisting of the items DefId and, since we only
consider non-generic items, an empty type-substitution set.
Given a translation item node, we can discover neighbors by inspecting its
MIR. We walk the MIR and any time we hit upon something that signifies a
reference to another translation item, we have found a neighbor. Since the
translation item we are currently at is always monomorphic, we also know the
concrete type arguments of its neighbors, and so all neighbors again will be
monomorphic. The specific forms a reference to a neighboring node can take
in MIR are quite diverse. Here is an overview:
The most obvious form of one translation item referencing another is a
function or method call (represented by a CALL terminator in MIR). But
calls are not the only thing that might introduce a reference between two
function translation items, and as we will see below, they are just a
specialized of the form described next, and consequently will don't get any
special treatment in the algorithm.
A function does not need to actually be called in order to be a neighbor of
another function. It suffices to just take a reference in order to introduce
an edge. Consider the following example:
```rust
fn print_val<T: Display>(x: T) {
println!("{}", x);
}
fn call_fn(f: &Fn(i32), x: i32) {
f(x);
}
fn main() {
let print_i32 = print_val::<i32>;
call_fn(&print_i32, 0);
}
```
The MIR of none of these functions will contain an explicit call to
`print_val::<i32>`. Nonetheless, in order to translate this program, we need
an instance of this function. Thus, whenever we encounter a function or
method in operand position, we treat it as a neighbor of the current
translation item. Calls are just a special case of that.
In a way, closures are a simple case. Since every closure object needs to be
constructed somewhere, we can reliably discover them by observing
`RValue::Aggregate` expressions with `AggregateKind::Closure`. This is also
true for closures inlined from other crates.
Drop glue translation items are introduced by MIR drop-statements. The
generated translation item will again have drop-glue item neighbors if the
type to be dropped contains nested values that also need to be dropped. It
might also have a function item neighbor for the explicit `Drop::drop`
implementation of its type.
A subtle way of introducing neighbor edges is by casting to a trait object.
Since the resulting fat-pointer contains a reference to a vtable, we need to
instantiate all object-save methods of the trait, as we need to store
pointers to these functions even if they never get called anywhere. This can
be seen as a special case of taking a function reference.
Since `Box` expression have special compiler support, no explicit calls to
`exchange_malloc()` and `exchange_free()` may show up in MIR, even if the
compiler will generate them. We have to observe `Rvalue::Box` expressions
and Box-typed drop-statements for that purpose.
Interaction with Cross-Crate Inlining
-------------------------------------
The binary of a crate will not only contain machine code for the items
defined in the source code of that crate. It will also contain monomorphic
instantiations of any extern generic functions and of functions marked with
The collection algorithm handles this more or less transparently. When
constructing a neighbor node for an item, the algorithm will always call
`inline::get_local_instance()` before proceeding. If no local instance can
be acquired (e.g. for a function that is just linked to) no node is created;
which is exactly what we want, since no machine code should be generated in
the current crate for such an item. On the other hand, if we can
successfully inline the function, we subsequently can just treat it like a
local item, walking it's MIR et cetera.
Eager and Lazy Collection Mode
------------------------------
Translation item collection can be performed in one of two modes:
- Lazy mode means that items will only be instantiated when actually
referenced. The goal is to produce the least amount of machine code
possible.
- Eager mode is meant to be used in conjunction with incremental compilation
where a stable set of translation items is more important than a minimal
one. Thus, eager mode will instantiate drop-glue for every drop-able type
in the crate, even of no drop call for that type exists (yet). It will
also instantiate default implementations of trait methods, something that
otherwise is only done on demand.
Open Issues
-----------
Some things are not yet fully implemented in the current version of this
module.
Since no MIR is constructed yet for initializer expressions of constants and
statics we cannot inspect these properly.
Ideally, no translation item should be generated for const fns unless there
is a call to them that cannot be evaluated at compile time. At the moment
this is not implemented however: a translation item will be produced
regardless of whether it is actually needed or not.
<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/rust-lang/rust/30900)
<!-- Reviewable:end -->
The purpose of the translation item collector is to find all monomorphic instances of functions, methods and statics that need to be translated into LLVM IR in order to compile the current crate.
So far these instances have been discovered lazily during the trans path. For incremental compilation we want to know the set of these instances in advance, and that is what the trans::collect module provides.
In the future, incremental and regular translation will be driven by the collector implemented here.
This PR adds some minor error correction to the parser - if there is a missing ident, we recover and carry on. It also makes compilation more robust so that non-fatal errors (which is still most of them, unfortunately) in parsing do not cause us to abort compilation. The effect is that a program with a missing or incorrect ident can get all the way to type checking.
In 95d904625b output was accidentally moved
from STDERR to STDOUT.
This commit also changes the order of debug output. Previously, it was:
```
/* id 22: … */ {
…
}
DEBUG:rustc::middle::dataflow:
```
Now, it is:
```
DEBUG:rustc::middle::dataflow: /* id 22: … */ {
…
}
```
This is a fix for #30741. It simplifies dep-graph tracking for trait matching. I was experimenting with having a greater resolution here, but decided to pare back to just have one dep node for "trait resolutions on trait `Foo`", which means that adding an impl to the trait `Foo` will invalidate all fns that had to do any trait matching at all on `Foo`. This seems like a reasonable starting place.
Independently, I realized I had neglected to record a dependency from trans on typeck -- this is obviously needed, since trans consumes a bunch of data structures that typeck produces (but which are not currently individually tracked) -- and because trans assumes that typeck has been done. Eventually those are going to go away and be replaced with MIR, which will be tracked, so this edge would presumably be derived automatically then, but it's an obvious enough thing to want for now.
r? @arielb1
cc @michaelwoerister -- this might indirectly fix the problem you observed with the trans cache, though it'd be nice to try and craft an independent test case for that.
was the major use-case, and to update the dep-graph. Other kinds of
predicates are now excluded from the cache because there is no easy way
to make a good dep-graph node for them, and because they are not
believed to be that useful. :)
Fixes#30741. (However, the test still gives wrong result for trans,
for an independent reason which is fixed in the next commit.)
