733 lines
31 KiB
Rust
Raw Normal View History

// Copyright 2012-2017 The Rust Project Developers. See the COPYRIGHT
// file at the top-level directory of this distribution and at
// http://rust-lang.org/COPYRIGHT.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.
use rustc::hir;
use rustc::hir::def_id::DefId;
use rustc::middle::region;
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
use rustc::mir::{self, Location, Place, Mir};
use rustc::mir::visit::{PlaceContext, Visitor};
use rustc::ty::{self, Region, TyCtxt};
use rustc::ty::RegionKind;
use rustc::ty::RegionKind::ReScope;
use rustc::util::nodemap::{FxHashMap, FxHashSet};
use rustc_data_structures::bitslice::{BitwiseOperator};
use rustc_data_structures::indexed_set::{IdxSet};
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
use rustc_data_structures::indexed_vec::{Idx, IndexVec};
use dataflow::{BitDenotation, BlockSets, InitialFlow};
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
pub use dataflow::indexes::{BorrowIndex, ReserveOrActivateIndex};
use borrow_check::nll::region_infer::RegionInferenceContext;
use borrow_check::nll::ToRegionVid;
use syntax_pos::Span;
use std::fmt;
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
use std::hash::Hash;
use std::rc::Rc;
2017-12-13 01:06:39 -06:00
/// `Borrows` stores the data used in the analyses that track the flow
/// of borrows.
///
/// It uniquely identifies every borrow (`Rvalue::Ref`) by a
/// `BorrowIndex`, and maps each such index to a `BorrowData`
/// describing the borrow. These indexes are used for representing the
/// borrows in compact bitvectors.
pub struct Borrows<'a, 'gcx: 'tcx, 'tcx: 'a> {
tcx: TyCtxt<'a, 'gcx, 'tcx>,
mir: &'a Mir<'tcx>,
scope_tree: Rc<region::ScopeTree>,
root_scope: Option<region::Scope>,
2017-12-13 01:06:39 -06:00
/// The fundamental map relating bitvector indexes to the borrows
/// in the MIR.
borrows: IndexVec<BorrowIndex, BorrowData<'tcx>>,
2017-12-13 01:06:39 -06:00
/// Each borrow is also uniquely identified in the MIR by the
/// `Location` of the assignment statement in which it appears on
/// the right hand side; we map each such location to the
/// corresponding `BorrowIndex`.
location_map: FxHashMap<Location, BorrowIndex>,
2017-12-13 01:06:39 -06:00
/// Every borrow in MIR is immediately stored into a place via an
/// assignment statement. This maps each such assigned place back
/// to its borrow-indexes.
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
assigned_map: FxHashMap<Place<'tcx>, FxHashSet<BorrowIndex>>,
2017-12-13 01:06:39 -06:00
/// Every borrow has a region; this maps each such regions back to
/// its borrow-indexes.
region_map: FxHashMap<Region<'tcx>, FxHashSet<BorrowIndex>>,
local_map: FxHashMap<mir::Local, FxHashSet<BorrowIndex>>,
region_span_map: FxHashMap<RegionKind, Span>,
nonlexical_regioncx: Option<Rc<RegionInferenceContext<'tcx>>>,
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
// Two-phase borrows actually requires two flow analyses; they need
// to be separate because the final results of the first are used to
// construct the gen+kill sets for the second. (The dataflow system
// is not designed to allow the gen/kill sets to change during the
// fixed-point iteration.)
/// The `Reservations` analysis is the first of the two flow analyses
/// tracking (phased) borrows. It computes where a borrow is reserved;
/// i.e. where it can reach in the control flow starting from its
/// initial `assigned = &'rgn borrowed` statement, and ending
/// whereever `'rgn` itself ends.
pub(crate) struct Reservations<'a, 'gcx: 'tcx, 'tcx: 'a>(pub(crate) Borrows<'a, 'gcx, 'tcx>);
/// The `ActiveBorrows` analysis is the second of the two flow
/// analyses tracking (phased) borrows. It computes where any given
/// borrow `&assigned = &'rgn borrowed` is *active*, which starts at
/// the first use of `assigned` after the reservation has started, and
/// ends whereever `'rgn` itself ends.
pub(crate) struct ActiveBorrows<'a, 'gcx: 'tcx, 'tcx: 'a>(pub(crate) Borrows<'a, 'gcx, 'tcx>);
impl<'a, 'gcx, 'tcx> Reservations<'a, 'gcx, 'tcx> {
pub(crate) fn new(b: Borrows<'a, 'gcx, 'tcx>) -> Self { Reservations(b) }
pub(crate) fn location(&self, idx: ReserveOrActivateIndex) -> &Location {
self.0.location(idx.borrow_index())
}
}
impl<'a, 'gcx, 'tcx> ActiveBorrows<'a, 'gcx, 'tcx> {
pub(crate) fn new(r: Reservations<'a, 'gcx, 'tcx>) -> Self { ActiveBorrows(r.0) }
pub(crate) fn location(&self, idx: ReserveOrActivateIndex) -> &Location {
self.0.location(idx.borrow_index())
}
}
// temporarily allow some dead fields: `kind` and `region` will be
// needed by borrowck; `borrowed_place` will probably be a MovePathIndex when
// that is extended to include borrowed data paths.
#[allow(dead_code)]
#[derive(Debug)]
pub struct BorrowData<'tcx> {
pub(crate) location: Location,
pub(crate) kind: mir::BorrowKind,
pub(crate) region: Region<'tcx>,
pub(crate) borrowed_place: mir::Place<'tcx>,
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
pub(crate) assigned_place: mir::Place<'tcx>,
}
impl<'tcx> fmt::Display for BorrowData<'tcx> {
fn fmt(&self, w: &mut fmt::Formatter) -> fmt::Result {
let kind = match self.kind {
mir::BorrowKind::Shared => "",
mir::BorrowKind::Unique => "uniq ",
mir::BorrowKind::Mut => "mut ",
};
let region = format!("{}", self.region);
let region = if region.len() > 0 { format!("{} ", region) } else { region };
write!(w, "&{}{}{:?}", region, kind, self.borrowed_place)
}
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
impl ReserveOrActivateIndex {
fn reserved(i: BorrowIndex) -> Self { ReserveOrActivateIndex::new((i.index() * 2)) }
fn active(i: BorrowIndex) -> Self { ReserveOrActivateIndex::new((i.index() * 2) + 1) }
pub(crate) fn is_reservation(self) -> bool { self.index() % 2 == 0 }
pub(crate) fn is_activation(self) -> bool { self.index() % 2 == 1}
pub(crate) fn kind(self) -> &'static str {
if self.is_reservation() { "reserved" } else { "active" }
}
pub(crate) fn borrow_index(self) -> BorrowIndex {
BorrowIndex::new(self.index() / 2)
}
}
impl<'a, 'gcx, 'tcx> Borrows<'a, 'gcx, 'tcx> {
2017-10-30 08:28:07 -04:00
pub fn new(tcx: TyCtxt<'a, 'gcx, 'tcx>,
mir: &'a Mir<'tcx>,
nonlexical_regioncx: Option<Rc<RegionInferenceContext<'tcx>>>,
def_id: DefId,
body_id: Option<hir::BodyId>)
2017-10-30 08:28:07 -04:00
-> Self {
let scope_tree = tcx.region_scope_tree(def_id);
let root_scope = body_id.map(|body_id| {
region::Scope::CallSite(tcx.hir.body(body_id).value.hir_id.local_id)
});
let mut visitor = GatherBorrows {
tcx,
mir,
idx_vec: IndexVec::new(),
location_map: FxHashMap(),
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
assigned_map: FxHashMap(),
region_map: FxHashMap(),
local_map: FxHashMap(),
region_span_map: FxHashMap()
};
visitor.visit_mir(mir);
return Borrows { tcx: tcx,
mir: mir,
borrows: visitor.idx_vec,
scope_tree,
root_scope,
location_map: visitor.location_map,
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
assigned_map: visitor.assigned_map,
region_map: visitor.region_map,
local_map: visitor.local_map,
2017-10-30 08:28:07 -04:00
region_span_map: visitor.region_span_map,
nonlexical_regioncx };
struct GatherBorrows<'a, 'gcx: 'tcx, 'tcx: 'a> {
tcx: TyCtxt<'a, 'gcx, 'tcx>,
mir: &'a Mir<'tcx>,
idx_vec: IndexVec<BorrowIndex, BorrowData<'tcx>>,
location_map: FxHashMap<Location, BorrowIndex>,
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
assigned_map: FxHashMap<Place<'tcx>, FxHashSet<BorrowIndex>>,
region_map: FxHashMap<Region<'tcx>, FxHashSet<BorrowIndex>>,
local_map: FxHashMap<mir::Local, FxHashSet<BorrowIndex>>,
region_span_map: FxHashMap<RegionKind, Span>,
}
impl<'a, 'gcx, 'tcx> Visitor<'tcx> for GatherBorrows<'a, 'gcx, 'tcx> {
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
fn visit_assign(&mut self,
block: mir::BasicBlock,
assigned_place: &mir::Place<'tcx>,
rvalue: &mir::Rvalue<'tcx>,
location: mir::Location) {
fn root_local(mut p: &mir::Place<'_>) -> Option<mir::Local> {
loop { match p {
mir::Place::Projection(pi) => p = &pi.base,
mir::Place::Static(_) => return None,
mir::Place::Local(l) => return Some(*l)
}}
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
if let mir::Rvalue::Ref(region, kind, ref borrowed_place) = *rvalue {
if is_unsafe_place(self.tcx, self.mir, borrowed_place) { return; }
let borrow = BorrowData {
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
location, kind, region,
borrowed_place: borrowed_place.clone(),
assigned_place: assigned_place.clone(),
};
let idx = self.idx_vec.push(borrow);
self.location_map.insert(location, idx);
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
insert(&mut self.assigned_map, assigned_place, idx);
insert(&mut self.region_map, &region, idx);
if let Some(local) = root_local(borrowed_place) {
insert(&mut self.local_map, &local, idx);
}
}
return self.super_assign(block, assigned_place, rvalue, location);
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
fn insert<'a, K, V>(map: &'a mut FxHashMap<K, FxHashSet<V>>,
k: &K,
v: V)
where K: Clone+Eq+Hash, V: Eq+Hash
{
map.entry(k.clone())
.or_insert(FxHashSet())
.insert(v);
}
}
fn visit_rvalue(&mut self,
rvalue: &mir::Rvalue<'tcx>,
location: mir::Location) {
if let mir::Rvalue::Ref(region, kind, ref place) = *rvalue {
// double-check that we already registered a BorrowData for this
let mut found_it = false;
for idx in &self.region_map[region] {
let bd = &self.idx_vec[*idx];
if bd.location == location &&
bd.kind == kind &&
bd.region == region &&
bd.borrowed_place == *place
{
found_it = true;
break;
}
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
assert!(found_it, "Ref {:?} at {:?} missing BorrowData", rvalue, location);
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
return self.super_rvalue(rvalue, location);
}
fn visit_statement(&mut self,
block: mir::BasicBlock,
statement: &mir::Statement<'tcx>,
location: Location) {
if let mir::StatementKind::EndRegion(region_scope) = statement.kind {
self.region_span_map.insert(ReScope(region_scope), statement.source_info.span);
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
return self.super_statement(block, statement, location);
}
}
}
pub fn borrows(&self) -> &IndexVec<BorrowIndex, BorrowData<'tcx>> { &self.borrows }
2017-12-10 17:00:20 +00:00
pub fn scope_tree(&self) -> &Rc<region::ScopeTree> { &self.scope_tree }
pub fn location(&self, idx: BorrowIndex) -> &Location {
&self.borrows[idx].location
}
2017-10-30 08:28:07 -04:00
/// Add all borrows to the kill set, if those borrows are out of scope at `location`.
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
///
/// `is_activations` tracks whether we are in the Reservations or
/// the ActiveBorrows flow analysis, and does not set the
/// activation kill bits in the former case. (Technically, we
/// could set those kill bits without such a guard, since they are
/// never gen'ed by Reservations in the first place. But it makes
/// the instrumentation and graph renderings nicer to leave
/// activations out when of the Reservations kill sets.)
2017-10-30 08:28:07 -04:00
fn kill_loans_out_of_scope_at_location(&self,
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
sets: &mut BlockSets<ReserveOrActivateIndex>,
location: Location,
is_activations: bool) {
if let Some(ref regioncx) = self.nonlexical_regioncx {
// NOTE: The state associated with a given `location`
// reflects the dataflow on entry to the statement. If it
// does not contain `borrow_region`, then then that means
// that the statement at `location` kills the borrow.
//
// We are careful always to call this function *before* we
// set up the gen-bits for the statement or
// termanator. That way, if the effect of the statement or
// terminator *does* introduce a new loan of the same
// region, then setting that gen-bit will override any
// potential kill introduced here.
2017-10-30 08:28:07 -04:00
for (borrow_index, borrow_data) in self.borrows.iter_enumerated() {
let borrow_region = borrow_data.region.to_region_vid();
if !regioncx.region_contains_point(borrow_region, location) {
sets.kill(&ReserveOrActivateIndex::reserved(borrow_index));
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
if is_activations {
sets.kill(&ReserveOrActivateIndex::active(borrow_index));
}
2017-10-30 08:28:07 -04:00
}
}
}
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
/// Models statement effect in Reservations and ActiveBorrows flow
/// analyses; `is activations` tells us if we are in the latter
/// case.
fn statement_effect_on_borrows(&self,
sets: &mut BlockSets<ReserveOrActivateIndex>,
location: Location,
is_activations: bool) {
let block = &self.mir.basic_blocks().get(location.block).unwrap_or_else(|| {
panic!("could not find block at location {:?}", location);
});
let stmt = block.statements.get(location.statement_index).unwrap_or_else(|| {
panic!("could not find statement at location {:?}");
});
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
// Do kills introduced by NLL before setting up any potential
// gens. (See NOTE in kill_loans_out_of_scope_at_location.)
self.kill_loans_out_of_scope_at_location(sets, location, is_activations);
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
if is_activations {
// INVARIANT: `sets.on_entry` accurately captures
// reservations on entry to statement (b/c
// accumulates_intrablock_state is overridden for
// ActiveBorrows).
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
//
// Now compute the activations generated by uses within
// the statement based on that reservation state.
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
let mut find = FindPlaceUses { sets, assigned_map: &self.assigned_map };
find.visit_statement(location.block, stmt, location);
}
match stmt.kind {
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
// EndRegion kills any borrows (reservations and active borrows both)
mir::StatementKind::EndRegion(region_scope) => {
if let Some(borrow_indexes) = self.region_map.get(&ReScope(region_scope)) {
2017-10-30 08:28:07 -04:00
assert!(self.nonlexical_regioncx.is_none());
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
for idx in borrow_indexes {
sets.kill(&ReserveOrActivateIndex::reserved(*idx));
if is_activations {
sets.kill(&ReserveOrActivateIndex::active(*idx));
}
}
} else {
// (if there is no entry, then there are no borrows to be tracked)
}
}
mir::StatementKind::Assign(_, ref rhs) => {
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
// NOTE: if/when the Assign case is revised to inspect
// the assigned_place here, make sure to also
// re-consider the current implementations of the
// propagate_call_return method.
if let mir::Rvalue::Ref(region, _, ref place) = *rhs {
if is_unsafe_place(self.tcx, self.mir, place) { return; }
if let RegionKind::ReEmpty = region {
// If the borrowed value is dead, the region for it
// can be empty. Don't track the borrow in that case.
return
}
let index = self.location_map.get(&location).unwrap_or_else(|| {
panic!("could not find BorrowIndex for location {:?}", location);
});
assert!(self.region_map.get(region).unwrap_or_else(|| {
panic!("could not find BorrowIndexs for region {:?}", region);
}).contains(&index));
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
sets.gen(&ReserveOrActivateIndex::reserved(*index));
}
}
mir::StatementKind::StorageDead(local) => {
// Make sure there are no remaining borrows for locals that
// are gone out of scope.
//
// FIXME: expand this to variables that are assigned over.
if let Some(borrow_indexes) = self.local_map.get(&local) {
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
sets.kill_all(borrow_indexes.iter()
.map(|b| ReserveOrActivateIndex::reserved(*b)));
if is_activations {
sets.kill_all(borrow_indexes.iter()
.map(|b| ReserveOrActivateIndex::active(*b)));
}
}
}
mir::StatementKind::InlineAsm { .. } |
mir::StatementKind::SetDiscriminant { .. } |
mir::StatementKind::StorageLive(..) |
mir::StatementKind::Validate(..) |
mir::StatementKind::Nop => {}
}
}
2017-10-30 08:28:07 -04:00
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
/// Models terminator effect in Reservations and ActiveBorrows
/// flow analyses; `is activations` tells us if we are in the
/// latter case.
fn terminator_effect_on_borrows(&self,
sets: &mut BlockSets<ReserveOrActivateIndex>,
location: Location,
is_activations: bool) {
let block = &self.mir.basic_blocks().get(location.block).unwrap_or_else(|| {
panic!("could not find block at location {:?}", location);
});
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
// Do kills introduced by NLL before setting up any potential
// gens. (See NOTE in kill_loans_out_of_scope_at_location.)
self.kill_loans_out_of_scope_at_location(sets, location, is_activations);
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
let term = block.terminator();
if is_activations {
// INVARIANT: `sets.on_entry` accurately captures
// reservations on entry to terminator (b/c
// accumulates_intrablock_state is overridden for
// ActiveBorrows).
//
// Now compute effect of the terminator on the activations
// themselves in the ActiveBorrows state.
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
let mut find = FindPlaceUses { sets, assigned_map: &self.assigned_map };
find.visit_terminator(location.block, term, location);
}
match term.kind {
mir::TerminatorKind::Resume |
mir::TerminatorKind::Return |
mir::TerminatorKind::GeneratorDrop => {
// When we return from the function, then all `ReScope`-style regions
// are guaranteed to have ended.
// Normally, there would be `EndRegion` statements that come before,
// and hence most of these loans will already be dead -- but, in some cases
// like unwind paths, we do not always emit `EndRegion` statements, so we
// add some kills here as a "backup" and to avoid spurious error messages.
for (borrow_index, borrow_data) in self.borrows.iter_enumerated() {
if let ReScope(scope) = borrow_data.region {
// Check that the scope is not actually a scope from a function that is
// a parent of our closure. Note that the CallSite scope itself is
// *outside* of the closure, for some weird reason.
if let Some(root_scope) = self.root_scope {
if *scope != root_scope &&
self.scope_tree.is_subscope_of(*scope, root_scope)
{
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
sets.kill(&ReserveOrActivateIndex::reserved(borrow_index));
if is_activations {
sets.kill(&ReserveOrActivateIndex::active(borrow_index));
}
}
}
}
}
}
mir::TerminatorKind::SwitchInt {..} |
mir::TerminatorKind::Drop {..} |
mir::TerminatorKind::DropAndReplace {..} |
mir::TerminatorKind::Call {..} |
mir::TerminatorKind::Assert {..} |
mir::TerminatorKind::Yield {..} |
mir::TerminatorKind::Goto {..} |
mir::TerminatorKind::FalseEdges {..} |
mir::TerminatorKind::Unreachable => {}
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
}
}
impl<'a, 'gcx, 'tcx> ActiveBorrows<'a, 'gcx, 'tcx> {
pub(crate) fn borrows(&self) -> &IndexVec<BorrowIndex, BorrowData<'tcx>> {
self.0.borrows()
}
/// Returns the span for the "end point" given region. This will
/// return `None` if NLL is enabled, since that concept has no
/// meaning there. Otherwise, return region span if it exists and
/// span for end of the function if it doesn't exist.
pub(crate) fn opt_region_end_span(&self, region: &Region) -> Option<Span> {
match self.0.nonlexical_regioncx {
Some(_) => None,
None => {
match self.0.region_span_map.get(region) {
Some(span) => Some(span.end_point()),
None => Some(self.0.mir.span.end_point())
}
}
}
}
}
/// `FindPlaceUses` is a MIR visitor that updates `self.sets` for all
/// of the borrows activated by a given statement or terminator.
///
/// ----
///
/// The `ActiveBorrows` flow analysis, when inspecting any given
/// statement or terminator, needs to "generate" (i.e. set to 1) all
/// of the bits for the borrows that are activated by that
/// statement/terminator.
///
/// This struct will seek out all places that are assignment-targets
/// for borrows (gathered in `self.assigned_map`; see also the
/// `assigned_map` in `struct Borrows`), and set the corresponding
/// gen-bits for activations of those borrows in `self.sets`
struct FindPlaceUses<'a, 'b: 'a, 'tcx: 'a> {
assigned_map: &'a FxHashMap<Place<'tcx>, FxHashSet<BorrowIndex>>,
sets: &'a mut BlockSets<'b, ReserveOrActivateIndex>,
}
impl<'a, 'b, 'tcx> FindPlaceUses<'a, 'b, 'tcx> {
fn has_been_reserved(&self, b: &BorrowIndex) -> bool {
self.sets.on_entry.contains(&ReserveOrActivateIndex::reserved(*b))
}
/// return whether `context` should be considered a "use" of a
/// place found in that context. "Uses" activate associated
/// borrows (at least when such uses occur while the borrow also
/// has a reservation at the time).
fn is_potential_use(context: PlaceContext) -> bool {
match context {
// storage effects on an place do not activate it
PlaceContext::StorageLive | PlaceContext::StorageDead => false,
// validation effects do not activate an place
//
// FIXME: Should they? Is it just another read? Or can we
// guarantee it won't dereference the stored address? How
// "deep" does validation go?
PlaceContext::Validate => false,
// FIXME: This is here to not change behaviour from before
// AsmOutput existed, but it's not necessarily a pure overwrite.
// so it's possible this should activate the place.
PlaceContext::AsmOutput |
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
// pure overwrites of an place do not activate it. (note
// PlaceContext::Call is solely about dest place)
PlaceContext::Store | PlaceContext::Call => false,
// reads of an place *do* activate it
PlaceContext::Move |
PlaceContext::Copy |
PlaceContext::Drop |
PlaceContext::Inspect |
PlaceContext::Borrow { .. } |
PlaceContext::Projection(..) => true,
}
}
}
impl<'a, 'b, 'tcx> Visitor<'tcx> for FindPlaceUses<'a, 'b, 'tcx> {
fn visit_place(&mut self,
place: &mir::Place<'tcx>,
context: PlaceContext<'tcx>,
location: Location) {
debug!("FindPlaceUses place: {:?} assigned from borrows: {:?} \
used in context: {:?} at location: {:?}",
place, self.assigned_map.get(place), context, location);
if Self::is_potential_use(context) {
if let Some(borrows) = self.assigned_map.get(place) {
for borrow_idx in borrows {
debug!("checking if index {:?} for {:?} is reserved ({}) \
and thus needs active gen-bit set in sets {:?}",
borrow_idx, place, self.has_been_reserved(&borrow_idx), self.sets);
if self.has_been_reserved(&borrow_idx) {
self.sets.gen(&ReserveOrActivateIndex::active(*borrow_idx));
} else {
// (This can certainly happen in valid code. I
// just want to know about it in the short
// term.)
debug!("encountered use of Place {:?} of borrow_idx {:?} \
at location {:?} outside of reservation",
place, borrow_idx, location);
}
}
}
}
self.super_place(place, context, location);
}
}
impl<'a, 'gcx, 'tcx> BitDenotation for Reservations<'a, 'gcx, 'tcx> {
type Idx = ReserveOrActivateIndex;
fn name() -> &'static str { "reservations" }
fn bits_per_block(&self) -> usize {
self.0.borrows.len() * 2
}
fn start_block_effect(&self, _entry_set: &mut IdxSet<ReserveOrActivateIndex>) {
// no borrows of code region_scopes have been taken prior to
// function execution, so this method has no effect on
// `_sets`.
}
fn statement_effect(&self,
sets: &mut BlockSets<ReserveOrActivateIndex>,
location: Location) {
debug!("Reservations::statement_effect sets: {:?} location: {:?}", sets, location);
self.0.statement_effect_on_borrows(sets, location, false);
}
fn terminator_effect(&self,
sets: &mut BlockSets<ReserveOrActivateIndex>,
location: Location) {
debug!("Reservations::terminator_effect sets: {:?} location: {:?}", sets, location);
self.0.terminator_effect_on_borrows(sets, location, false);
}
fn propagate_call_return(&self,
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
_in_out: &mut IdxSet<ReserveOrActivateIndex>,
_call_bb: mir::BasicBlock,
_dest_bb: mir::BasicBlock,
_dest_place: &mir::Place) {
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
// there are no effects on borrows from method call return...
//
// ... but if overwriting a place can affect flow state, then
// latter is not true; see NOTE on Assign case in
// statement_effect_on_borrows.
}
}
impl<'a, 'gcx, 'tcx> BitDenotation for ActiveBorrows<'a, 'gcx, 'tcx> {
type Idx = ReserveOrActivateIndex;
fn name() -> &'static str { "active_borrows" }
/// Overriding this method; `ActiveBorrows` uses the intrablock
/// state in `on_entry` to track the current reservations (which
/// then affect the construction of the gen/kill sets for
/// activations).
fn accumulates_intrablock_state() -> bool { true }
fn bits_per_block(&self) -> usize {
self.0.borrows.len() * 2
}
fn start_block_effect(&self, _entry_sets: &mut IdxSet<ReserveOrActivateIndex>) {
// no borrows of code region_scopes have been taken prior to
// function execution, so this method has no effect on
// `_sets`.
}
fn statement_effect(&self,
sets: &mut BlockSets<ReserveOrActivateIndex>,
location: Location) {
debug!("ActiveBorrows::statement_effect sets: {:?} location: {:?}", sets, location);
self.0.statement_effect_on_borrows(sets, location, true);
}
fn terminator_effect(&self,
sets: &mut BlockSets<ReserveOrActivateIndex>,
location: Location) {
debug!("ActiveBorrows::terminator_effect sets: {:?} location: {:?}", sets, location);
self.0.terminator_effect_on_borrows(sets, location, true);
}
fn propagate_call_return(&self,
_in_out: &mut IdxSet<ReserveOrActivateIndex>,
_call_bb: mir::BasicBlock,
_dest_bb: mir::BasicBlock,
_dest_place: &mir::Place) {
// there are no effects on borrows from method call return...
//
// ... but If overwriting a place can affect flow state, then
// latter is not true; see NOTE on Assign case in
// statement_effect_on_borrows.
}
}
impl<'a, 'gcx, 'tcx> BitwiseOperator for Reservations<'a, 'gcx, 'tcx> {
#[inline]
fn join(&self, pred1: usize, pred2: usize) -> usize {
pred1 | pred2 // union effects of preds when computing reservations
}
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
impl<'a, 'gcx, 'tcx> BitwiseOperator for ActiveBorrows<'a, 'gcx, 'tcx> {
#[inline]
fn join(&self, pred1: usize, pred2: usize) -> usize {
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
pred1 | pred2 // union effects of preds when computing activations
}
}
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
impl<'a, 'gcx, 'tcx> InitialFlow for Reservations<'a, 'gcx, 'tcx> {
#[inline]
fn bottom_value() -> bool {
New `ActiveBorrows` dataflow for two-phase `&mut`; not yet borrowed-checked. High-level picture: The old `Borrows` analysis is now called `Reservations` (implemented as a newtype wrapper around `Borrows`); this continues to compute whether a `Rvalue::Ref` can reach a statement without an intervening `EndRegion`. In addition, we also track what `Place` each such `Rvalue::Ref` was immediately assigned to in a given borrow (yay for MIR-structural properties!). The new `ActiveBorrows` analysis then tracks the initial use of any of those assigned `Places` for a given borrow. I.e. a borrow becomes "active" immediately after it starts being "used" in some way. (This is conservative in the sense that we will treat a copy `x = y;` as a use of `y`; in principle one might further delay activation in such cases.) The new `ActiveBorrows` analysis needs to take the `Reservations` results as an initial input, because the reservation state influences the gen/kill sets for `ActiveBorrows`. In particular, a use of `a` activates a borrow `a = &b` if and only if there exists a path (in the control flow graph) from the borrow to that use. So we need to know if the borrow reaches a given use to know if it really gets a gen-bit or not. * Incorporating the output from one dataflow analysis into the input of another required more changes to the infrastructure than I had expected, and even after those changes, the resulting code is still a bit subtle. * In particular, Since we need to know the intrablock reservation state, we need to dynamically update a bitvector for the reservations as we are also trying to compute the gen/kills bitvector for the active borrows. * The way I ended up deciding to do this (after also toying with at least two other designs) is to put both the reservation state and the active borrow state into a single bitvector. That is why we now have separate (but related) `BorrowIndex` and `ReserveOrActivateIndex`: each borrow index maps to a pair of neighboring reservation and activation indexes. As noted above, these changes are solely adding the active borrows dataflow analysis (and updating the existing code to cope with the switch from `Borrows` to `Reservations`). The code to process the bitvector in the borrow checker currently just skips over all of the active borrow bits. But atop this commit, one *can* observe the analysis results by looking at the graphviz output, e.g. via ```rust #[rustc_mir(borrowck_graphviz_preflow="pre_two_phase.dot", borrowck_graphviz_postflow="post_two_phase.dot")] ``` Includes doc for `FindPlaceUses`, as well as `Reservations` and `ActiveBorrows` structs, which are wrappers are the `Borrows` struct that dictate which flow analysis should be performed.
2017-12-01 12:32:51 +01:00
false // bottom = no Rvalue::Refs are reserved by default
}
}
fn is_unsafe_place<'a, 'gcx: 'tcx, 'tcx: 'a>(
tcx: TyCtxt<'a, 'gcx, 'tcx>,
mir: &'a Mir<'tcx>,
place: &mir::Place<'tcx>
) -> bool {
2017-12-01 14:31:47 +02:00
use self::mir::Place::*;
use self::mir::ProjectionElem;
match *place {
Local(_) => false,
Static(ref static_) => tcx.is_static_mut(static_.def_id),
Projection(ref proj) => {
match proj.elem {
ProjectionElem::Field(..) |
ProjectionElem::Downcast(..) |
ProjectionElem::Subslice { .. } |
ProjectionElem::ConstantIndex { .. } |
ProjectionElem::Index(_) => {
is_unsafe_place(tcx, mir, &proj.base)
}
ProjectionElem::Deref => {
let ty = proj.base.ty(mir, tcx).to_ty(tcx);
match ty.sty {
ty::TyRawPtr(..) => true,
_ => is_unsafe_place(tcx, mir, &proj.base),
}
}
}
}
}
}