1 //! Propagates assignment destinations backwards in the CFG to eliminate redundant assignments.
5 //! MIR building can insert a lot of redundant copies, and Rust code in general often tends to move
6 //! values around a lot. The result is a lot of assignments of the form `dest = {move} src;` in MIR.
7 //! MIR building for constants in particular tends to create additional locals that are only used
8 //! inside a single block to shuffle a value around unnecessarily.
10 //! LLVM by itself is not good enough at eliminating these redundant copies (eg. see
11 //! <https://github.com/rust-lang/rust/issues/32966>), so this leaves some performance on the table
12 //! that we can regain by implementing an optimization for removing these assign statements in rustc
13 //! itself. When this optimization runs fast enough, it can also speed up the constant evaluation
14 //! and code generation phases of rustc due to the reduced number of statements and locals.
16 //! # The Optimization
18 //! Conceptually, this optimization is "destination propagation". It is similar to the Named Return
19 //! Value Optimization, or NRVO, known from the C++ world, except that it isn't limited to return
20 //! values or the return place `_0`. On a very high level, independent of the actual implementation
21 //! details, it does the following:
23 //! 1) Identify `dest = src;` statements with values for `dest` and `src` whose storage can soundly
25 //! 2) Replace all mentions of `src` with `dest` ("unifying" them and propagating the destination
27 //! 3) Delete the `dest = src;` statement (by making it a `nop`).
29 //! Step 1) is by far the hardest, so it is explained in more detail below.
33 //! We have a pair of places `p` and `q`, whose memory we would like to merge. In order for this to
34 //! be sound, we need to check a number of conditions:
36 //! * `p` and `q` must both be *constant* - it does not make much sense to talk about merging them
37 //! if they do not consistently refer to the same place in memory. This is satisfied if they do
38 //! not contain any indirection through a pointer or any indexing projections.
40 //! * We need to make sure that the goal of "merging the memory" is actually structurally possible
41 //! in MIR. For example, even if all the other conditions are satisfied, there is no way to
42 //! "merge" `_5.foo` and `_6.bar`. For now, we ensure this by requiring that both `p` and `q` are
43 //! locals with no further projections. Future iterations of this pass should improve on this.
45 //! * Finally, we want `p` and `q` to use the same memory - however, we still need to make sure that
46 //! each of them has enough "ownership" of that memory to continue "doing its job." More
47 //! precisely, what we will check is that whenever the program performs a write to `p`, then it
48 //! does not currently care about what the value in `q` is (and vice versa). We formalize the
49 //! notion of "does not care what the value in `q` is" by checking the *liveness* of `q`.
51 //! Because of the difficulty of computing liveness of places that have their address taken, we do
52 //! not even attempt to do it. Any places that are in a local that has its address taken is
53 //! excluded from the optimization.
55 //! The first two conditions are simple structural requirements on the `Assign` statements that can
56 //! be trivially checked. The third requirement however is more difficult and costly to check.
58 //! ## Future Improvements
60 //! There are a number of ways in which this pass could be improved in the future:
62 //! * Merging storage liveness ranges instead of removing storage statements completely. This may
63 //! improve stack usage.
65 //! * Allow merging locals into places with projections, eg `_5` into `_6.foo`.
67 //! * Liveness analysis with more precision than whole locals at a time. The smaller benefit of this
68 //! is that it would allow us to dest prop at "sub-local" levels in some cases. The bigger benefit
69 //! of this is that such liveness analysis can report more accurate results about whole locals at
70 //! a time. For example, consider:
72 //! ```ignore (syntax-highliting-only)
79 //! Because the current analysis only thinks in terms of locals, it does not have enough
80 //! information to report that `_1` is dead in the "unrelated code" section.
82 //! * Liveness analysis enabled by alias analysis. This would allow us to not just bail on locals
83 //! that ever have their address taken. Of course that requires actually having alias analysis
84 //! (and a model to build it on), so this might be a bit of a ways off.
86 //! * Various perf improvents. There are a bunch of comments in here marked `PERF` with ideas for
87 //! how to do things more efficiently. However, the complexity of the pass as a whole should be
92 //! A [previous attempt][attempt 1] at implementing an optimization like this turned out to be a
93 //! significant regression in compiler performance. Fixing the regressions introduced a lot of
94 //! undesirable complexity to the implementation.
96 //! A [subsequent approach][attempt 2] tried to avoid the costly computation by limiting itself to
97 //! acyclic CFGs, but still turned out to be far too costly to run due to suboptimal performance
98 //! within individual basic blocks, requiring a walk across the entire block for every assignment
99 //! found within the block. For the `tuple-stress` benchmark, which has 458745 statements in a
100 //! single block, this proved to be far too costly.
102 //! [Another approach after that][attempt 3] was much closer to correct, but had some soundness
103 //! issues - it was failing to consider stores outside live ranges, and failed to uphold some of the
104 //! requirements that MIR has for non-overlapping places within statements. However, it also had
105 //! performance issues caused by `O(l² * s)` runtime, where `l` is the number of locals and `s` is
106 //! the number of statements and terminators.
108 //! Since the first attempt at this, the compiler has improved dramatically, and new analysis
109 //! frameworks have been added that should make this approach viable without requiring a limited
110 //! approach that only works for some classes of CFGs:
111 //! - rustc now has a powerful dataflow analysis framework that can handle forwards and backwards
112 //! analyses efficiently.
113 //! - Layout optimizations for generators have been added to improve code generation for
114 //! async/await, which are very similar in spirit to what this optimization does.
116 //! Also, rustc now has a simple NRVO pass (see `nrvo.rs`), which handles a subset of the cases that
117 //! this destination propagation pass handles, proving that similar optimizations can be performed
120 //! ## Pre/Post Optimization
122 //! It is recommended to run `SimplifyCfg` and then `SimplifyLocals` some time after this pass, as
123 //! it replaces the eliminated assign statements with `nop`s and leaves unused locals behind.
125 //! [liveness]: https://en.wikipedia.org/wiki/Live_variable_analysis
126 //! [attempt 1]: https://github.com/rust-lang/rust/pull/47954
127 //! [attempt 2]: https://github.com/rust-lang/rust/pull/71003
128 //! [attempt 3]: https://github.com/rust-lang/rust/pull/72632
130 use std::collections::hash_map::{Entry, OccupiedEntry};
133 use rustc_data_structures::fx::FxHashMap;
134 use rustc_index::bit_set::BitSet;
135 use rustc_middle::mir::visit::{MutVisitor, PlaceContext, Visitor};
136 use rustc_middle::mir::{dump_mir, PassWhere};
137 use rustc_middle::mir::{
138 traversal, BasicBlock, Body, InlineAsmOperand, Local, LocalKind, Location, Operand, Place,
139 Rvalue, Statement, StatementKind, TerminatorKind,
141 use rustc_middle::ty::TyCtxt;
142 use rustc_mir_dataflow::impls::MaybeLiveLocals;
143 use rustc_mir_dataflow::{Analysis, ResultsCursor};
145 pub struct DestinationPropagation;
147 impl<'tcx> MirPass<'tcx> for DestinationPropagation {
148 fn is_enabled(&self, sess: &rustc_session::Session) -> bool {
149 // For now, only run at MIR opt level 3. Two things need to be changed before this can be
150 // turned on by default:
151 // 1. Because of the overeager removal of storage statements, this can cause stack space
152 // regressions. This opt is not the place to fix this though, it's a more general
154 // 2. Despite being an overall perf improvement, this still causes a 30% regression in
155 // keccak. We can temporarily fix this by bounding function size, but in the long term
156 // we should fix this by being smarter about invalidating analysis results.
157 sess.mir_opt_level() >= 3
160 fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
161 let def_id = body.source.def_id();
162 let mut allocations = Allocations::default();
163 trace!(func = ?tcx.def_path_str(def_id));
165 let borrowed = rustc_mir_dataflow::impls::borrowed_locals(body);
167 // In order to avoid having to collect data for every single pair of locals in the body, we
168 // do not allow doing more than one merge for places that are derived from the same local at
169 // once. To avoid missed opportunities, we instead iterate to a fixed point - we'll refer to
170 // each of these iterations as a "round."
172 // Reaching a fixed point could in theory take up to `min(l, s)` rounds - however, we do not
173 // expect to see MIR like that. To verify this, a test was run against `[rust-lang/regex]` -
174 // the average MIR body saw 1.32 full iterations of this loop. The most that was hit were 30
175 // for a single function. Only 80/2801 (2.9%) of functions saw at least 5.
177 // [rust-lang/regex]:
178 // https://github.com/rust-lang/regex/tree/b5372864e2df6a2f5e543a556a62197f50ca3650
179 let mut round_count = 0;
181 // PERF: Can we do something smarter than recalculating the candidates and liveness
183 let mut candidates = find_candidates(
186 &mut allocations.candidates,
187 &mut allocations.candidates_reverse,
190 let mut live = MaybeLiveLocals
191 .into_engine(tcx, body)
192 .iterate_to_fixpoint()
193 .into_results_cursor(body);
194 dest_prop_mir_dump(tcx, body, &mut live, round_count);
196 FilterInformation::filter_liveness(
199 &mut allocations.write_info,
203 // Because we do not update liveness information, it is unsound to use a local for more
204 // than one merge operation within a single round of optimizations. We store here which
205 // ones we have already used.
206 let mut merged_locals: BitSet<Local> = BitSet::new_empty(body.local_decls.len());
208 // This is the set of merges we will apply this round. It is a subset of the candidates.
209 let mut merges = FxHashMap::default();
211 for (src, candidates) in candidates.c.iter() {
212 if merged_locals.contains(*src) {
216 candidates.iter().find(|dest| !merged_locals.contains(**dest)) else {
219 if !tcx.consider_optimizing(|| {
220 format!("{} round {}", tcx.def_path_str(def_id), round_count)
224 merges.insert(*src, *dest);
225 merged_locals.insert(*src);
226 merged_locals.insert(*dest);
228 trace!(merging = ?merges);
230 if merges.is_empty() {
235 apply_merges(body, tcx, &merges, &merged_locals);
242 /// Container for the various allocations that we need.
244 /// We store these here and hand out `&mut` access to them, instead of dropping and recreating them
245 /// frequently. Everything with a `&'alloc` lifetime points into here.
248 candidates: FxHashMap<Local, Vec<Local>>,
249 candidates_reverse: FxHashMap<Local, Vec<Local>>,
250 write_info: WriteInfo,
251 // PERF: Do this for `MaybeLiveLocals` allocations too.
255 struct Candidates<'alloc> {
256 /// The set of candidates we are considering in this optimization.
258 /// We will always merge the key into at most one of its values.
260 /// Whether a place ends up in the key or the value does not correspond to whether it appears as
261 /// the lhs or rhs of any assignment. As a matter of fact, the places in here might never appear
262 /// in an assignment at all. This happens because if we see an assignment like this:
264 /// ```ignore (syntax-highlighting-only)
268 /// We will still report that we would like to merge `_1` and `_2` in an attempt to allow us to
269 /// remove that assignment.
270 c: &'alloc mut FxHashMap<Local, Vec<Local>>,
271 /// A reverse index of the `c` set; if the `c` set contains `a => Place { local: b, proj }`,
272 /// then this contains `b => a`.
273 // PERF: Possibly these should be `SmallVec`s?
274 reverse: &'alloc mut FxHashMap<Local, Vec<Local>>,
277 //////////////////////////////////////////////////////////
280 // Applies the actual optimization
282 fn apply_merges<'tcx>(
283 body: &mut Body<'tcx>,
285 merges: &FxHashMap<Local, Local>,
286 merged_locals: &BitSet<Local>,
288 let mut merger = Merger { tcx, merges, merged_locals };
289 merger.visit_body_preserves_cfg(body);
292 struct Merger<'a, 'tcx> {
294 merges: &'a FxHashMap<Local, Local>,
295 merged_locals: &'a BitSet<Local>,
298 impl<'a, 'tcx> MutVisitor<'tcx> for Merger<'a, 'tcx> {
299 fn tcx(&self) -> TyCtxt<'tcx> {
303 fn visit_local(&mut self, local: &mut Local, _: PlaceContext, _location: Location) {
304 if let Some(dest) = self.merges.get(local) {
309 fn visit_statement(&mut self, statement: &mut Statement<'tcx>, location: Location) {
310 match &statement.kind {
311 // FIXME: Don't delete storage statements, but "merge" the storage ranges instead.
312 StatementKind::StorageDead(local) | StatementKind::StorageLive(local)
313 if self.merged_locals.contains(*local) =>
315 statement.make_nop();
320 self.super_statement(statement, location);
321 match &statement.kind {
322 StatementKind::Assign(box (dest, rvalue)) => {
324 Rvalue::Use(Operand::Copy(place) | Operand::Move(place)) => {
325 // These might've been turned into self-assignments by the replacement
326 // (this includes the original statement we wanted to eliminate).
328 debug!("{:?} turned into self-assignment, deleting", location);
329 statement.make_nop();
341 //////////////////////////////////////////////////////////
342 // Liveness filtering
344 // This section enforces bullet point 2
346 struct FilterInformation<'a, 'body, 'alloc, 'tcx> {
347 body: &'body Body<'tcx>,
348 live: &'a mut ResultsCursor<'body, 'tcx, MaybeLiveLocals>,
349 candidates: &'a mut Candidates<'alloc>,
350 write_info: &'alloc mut WriteInfo,
354 // We first implement some utility functions which we will expose removing candidates according to
355 // different needs. Throughout the livenss filtering, the `candidates` are only ever accessed
356 // through these methods, and not directly.
357 impl<'alloc> Candidates<'alloc> {
358 /// Just `Vec::retain`, but the condition is inverted and we add debugging output
359 fn vec_filter_candidates(
362 mut f: impl FnMut(Local) -> CandidateFilter,
366 let remove = f(*dest);
367 if remove == CandidateFilter::Remove {
368 trace!("eliminating {:?} => {:?} due to conflict at {:?}", src, dest, at);
370 remove == CandidateFilter::Keep
374 /// `vec_filter_candidates` but for an `Entry`
375 fn entry_filter_candidates(
376 mut entry: OccupiedEntry<'_, Local, Vec<Local>>,
378 f: impl FnMut(Local) -> CandidateFilter,
381 let candidates = entry.get_mut();
382 Self::vec_filter_candidates(p, candidates, f, at);
383 if candidates.len() == 0 {
388 /// For all candidates `(p, q)` or `(q, p)` removes the candidate if `f(q)` says to do so
389 fn filter_candidates_by(
392 mut f: impl FnMut(Local) -> CandidateFilter,
395 // Cover the cases where `p` appears as a `src`
396 if let Entry::Occupied(entry) = self.c.entry(p) {
397 Self::entry_filter_candidates(entry, p, &mut f, at);
399 // And the cases where `p` appears as a `dest`
400 let Some(srcs) = self.reverse.get_mut(&p) else {
403 // We use `retain` here to remove the elements from the reverse set if we've removed the
404 // matching candidate in the forward set.
406 if f(*src) == CandidateFilter::Keep {
409 let Entry::Occupied(entry) = self.c.entry(*src) else {
412 Self::entry_filter_candidates(
416 if dest == p { CandidateFilter::Remove } else { CandidateFilter::Keep }
425 #[derive(Copy, Clone, PartialEq, Eq)]
426 enum CandidateFilter {
431 impl<'a, 'body, 'alloc, 'tcx> FilterInformation<'a, 'body, 'alloc, 'tcx> {
432 /// Filters the set of candidates to remove those that conflict.
434 /// The steps we take are exactly those that are outlined at the top of the file. For each
435 /// statement/terminator, we collect the set of locals that are written to in that
436 /// statement/terminator, and then we remove all pairs of candidates that contain one such local
437 /// and another one that is live.
439 /// We need to be careful about the ordering of operations within each statement/terminator
440 /// here. Many statements might write and read from more than one place, and we need to consider
441 /// them all. The strategy for doing this is as follows: We first gather all the places that are
442 /// written to within the statement/terminator via `WriteInfo`. Then, we use the liveness
443 /// analysis from *before* the statement/terminator (in the control flow sense) to eliminate
444 /// candidates - this is because we want to conservatively treat a pair of locals that is both
445 /// read and written in the statement/terminator to be conflicting, and the liveness analysis
446 /// before the statement/terminator will correctly report locals that are read in the
447 /// statement/terminator to be live. We are additionally conservative by treating all written to
448 /// locals as also being read from.
449 fn filter_liveness<'b>(
450 candidates: &mut Candidates<'alloc>,
451 live: &mut ResultsCursor<'b, 'tcx, MaybeLiveLocals>,
452 write_info_alloc: &'alloc mut WriteInfo,
453 body: &'b Body<'tcx>,
455 let mut this = FilterInformation {
459 // We don't actually store anything at this scope, we just keep things here to be able
460 // to reuse the allocation.
461 write_info: write_info_alloc,
462 // Doesn't matter what we put here, will be overwritten before being used
463 at: Location { block: BasicBlock::from_u32(0), statement_index: 0 },
465 this.internal_filter_liveness();
468 fn internal_filter_liveness(&mut self) {
469 for (block, data) in traversal::preorder(self.body) {
470 self.at = Location { block, statement_index: data.statements.len() };
471 self.live.seek_after_primary_effect(self.at);
472 self.write_info.for_terminator(&data.terminator().kind);
473 self.apply_conflicts();
475 for (i, statement) in data.statements.iter().enumerate().rev() {
476 self.at = Location { block, statement_index: i };
477 self.live.seek_after_primary_effect(self.at);
478 self.write_info.for_statement(&statement.kind, self.body);
479 self.apply_conflicts();
484 fn apply_conflicts(&mut self) {
485 let writes = &self.write_info.writes;
487 let other_skip = self.write_info.skip_pair.and_then(|(a, b)| {
496 self.candidates.filter_candidates_by(
499 if Some(q) == other_skip {
500 return CandidateFilter::Keep;
502 // It is possible that a local may be live for less than the
503 // duration of a statement This happens in the case of function
504 // calls or inline asm. Because of this, we also mark locals as
505 // conflicting when both of them are written to in the same
507 if self.live.contains(q) || writes.contains(&q) {
508 CandidateFilter::Remove
510 CandidateFilter::Keep
519 /// Describes where a statement/terminator writes to
520 #[derive(Default, Debug)]
523 /// If this pair of locals is a candidate pair, completely skip processing it during this
524 /// statement. All other candidates are unaffected.
525 skip_pair: Option<(Local, Local)>,
529 fn for_statement<'tcx>(&mut self, statement: &StatementKind<'tcx>, body: &Body<'tcx>) {
532 StatementKind::Assign(box (lhs, rhs)) => {
533 self.add_place(*lhs);
536 self.add_operand(op);
537 self.consider_skipping_for_assign_use(*lhs, op, body);
539 Rvalue::Repeat(op, _) => {
540 self.add_operand(op);
542 Rvalue::Cast(_, op, _)
543 | Rvalue::UnaryOp(_, op)
544 | Rvalue::ShallowInitBox(op, _) => {
545 self.add_operand(op);
547 Rvalue::BinaryOp(_, ops) | Rvalue::CheckedBinaryOp(_, ops) => {
548 for op in [&ops.0, &ops.1] {
549 self.add_operand(op);
552 Rvalue::Aggregate(_, ops) => {
554 self.add_operand(op);
557 Rvalue::ThreadLocalRef(_)
558 | Rvalue::NullaryOp(_, _)
559 | Rvalue::Ref(_, _, _)
560 | Rvalue::AddressOf(_, _)
562 | Rvalue::Discriminant(_)
563 | Rvalue::CopyForDeref(_) => (),
566 // Retags are technically also reads, but reporting them as a write suffices
567 StatementKind::SetDiscriminant { place, .. }
568 | StatementKind::Deinit(place)
569 | StatementKind::Retag(_, place) => {
570 self.add_place(**place);
572 StatementKind::Intrinsic(_)
574 | StatementKind::Coverage(_)
575 | StatementKind::StorageLive(_)
576 | StatementKind::StorageDead(_) => (),
577 StatementKind::FakeRead(_) | StatementKind::AscribeUserType(_, _) => {
578 bug!("{:?} not found in this MIR phase", statement)
583 fn consider_skipping_for_assign_use<'tcx>(
589 let Some(rhs) = rhs.place() else {
592 if let Some(pair) = places_to_candidate_pair(lhs, rhs, body) {
593 self.skip_pair = Some(pair);
597 fn for_terminator<'tcx>(&mut self, terminator: &TerminatorKind<'tcx>) {
600 TerminatorKind::SwitchInt { discr: op, .. }
601 | TerminatorKind::Assert { cond: op, .. } => {
602 self.add_operand(op);
604 TerminatorKind::Call { destination, func, args, .. } => {
605 self.add_place(*destination);
606 self.add_operand(func);
608 self.add_operand(arg);
611 TerminatorKind::InlineAsm { operands, .. } => {
612 for asm_operand in operands {
614 InlineAsmOperand::In { value, .. } => {
615 self.add_operand(value);
617 InlineAsmOperand::Out { place, .. } => {
618 if let Some(place) = place {
619 self.add_place(*place);
622 // Note that the `late` field in `InOut` is about whether the registers used
623 // for these things overlap, and is of absolutely no interest to us.
624 InlineAsmOperand::InOut { in_value, out_place, .. } => {
625 if let Some(place) = out_place {
626 self.add_place(*place);
628 self.add_operand(in_value);
630 InlineAsmOperand::Const { .. }
631 | InlineAsmOperand::SymFn { .. }
632 | InlineAsmOperand::SymStatic { .. } => (),
636 TerminatorKind::Goto { .. }
637 | TerminatorKind::Resume { .. }
638 | TerminatorKind::Abort { .. }
639 | TerminatorKind::Return
640 | TerminatorKind::Unreachable { .. } => (),
641 TerminatorKind::Drop { .. } => {
642 // `Drop`s create a `&mut` and so are not considered
644 TerminatorKind::DropAndReplace { .. }
645 | TerminatorKind::Yield { .. }
646 | TerminatorKind::GeneratorDrop
647 | TerminatorKind::FalseEdge { .. }
648 | TerminatorKind::FalseUnwind { .. } => {
649 bug!("{:?} not found in this MIR phase", terminator)
654 fn add_place<'tcx>(&mut self, place: Place<'tcx>) {
655 self.writes.push(place.local);
658 fn add_operand<'tcx>(&mut self, op: &Operand<'tcx>) {
660 // FIXME(JakobDegen): In a previous version, the `Move` case was incorrectly treated as
661 // being a read only. This was unsound, however we cannot add a regression test because
662 // it is not possible to set this off with current MIR. Once we have that ability, a
663 // regression test should be added.
664 Operand::Move(p) => self.add_place(*p),
665 Operand::Copy(_) | Operand::Constant(_) => (),
669 fn reset(&mut self) {
671 self.skip_pair = None;
675 /////////////////////////////////////////////////////
676 // Candidate accumulation
678 /// If the pair of places is being considered for merging, returns the candidate which would be
679 /// merged in order to accomplish this.
681 /// The contract here is in one direction - there is a guarantee that merging the locals that are
682 /// outputted by this function would result in an assignment between the inputs becoming a
683 /// self-assignment. However, there is no guarantee that the returned pair is actually suitable for
684 /// merging - candidate collection must still check this independently.
686 /// This output is unique for each unordered pair of input places.
687 fn places_to_candidate_pair<'tcx>(
691 ) -> Option<(Local, Local)> {
692 let (mut a, mut b) = if a.projection.len() == 0 && b.projection.len() == 0 {
698 // By sorting, we make sure we're input order independent
700 std::mem::swap(&mut a, &mut b);
703 // We could now return `(a, b)`, but then we miss some candidates in the case where `a` can't be
705 if is_local_required(a, body) {
706 std::mem::swap(&mut a, &mut b);
708 // We could check `is_local_required` again here, but there's no need - after all, we make no
709 // promise that the candidate pair is actually valid
713 /// Collects the candidates for merging
715 /// This is responsible for enforcing the first and third bullet point.
716 fn find_candidates<'alloc, 'tcx>(
718 borrowed: &BitSet<Local>,
719 candidates: &'alloc mut FxHashMap<Local, Vec<Local>>,
720 candidates_reverse: &'alloc mut FxHashMap<Local, Vec<Local>>,
721 ) -> Candidates<'alloc> {
723 candidates_reverse.clear();
724 let mut visitor = FindAssignments { body, candidates, borrowed };
725 visitor.visit_body(body);
726 // Deduplicate candidates
727 for (_, cands) in candidates.iter_mut() {
731 // Generate the reverse map
732 for (src, cands) in candidates.iter() {
733 for dest in cands.iter().copied() {
734 candidates_reverse.entry(dest).or_default().push(*src);
737 Candidates { c: candidates, reverse: candidates_reverse }
740 struct FindAssignments<'a, 'alloc, 'tcx> {
741 body: &'a Body<'tcx>,
742 candidates: &'alloc mut FxHashMap<Local, Vec<Local>>,
743 borrowed: &'a BitSet<Local>,
746 impl<'tcx> Visitor<'tcx> for FindAssignments<'_, '_, 'tcx> {
747 fn visit_statement(&mut self, statement: &Statement<'tcx>, _: Location) {
748 if let StatementKind::Assign(box (
750 Rvalue::Use(Operand::Copy(rhs) | Operand::Move(rhs)),
753 let Some((src, dest)) = places_to_candidate_pair(*lhs, *rhs, self.body) else {
757 // As described at the top of the file, we do not go near things that have their address
759 if self.borrowed.contains(src) || self.borrowed.contains(dest) {
763 // Also, we need to make sure that MIR actually allows the `src` to be removed
764 if is_local_required(src, self.body) {
768 // We may insert duplicates here, but that's fine
769 self.candidates.entry(src).or_default().push(dest);
774 /// Some locals are part of the function's interface and can not be removed.
776 /// Note that these locals *can* still be merged with non-required locals by removing that other
778 fn is_local_required(local: Local, body: &Body<'_>) -> bool {
779 match body.local_kind(local) {
780 LocalKind::Arg | LocalKind::ReturnPointer => true,
781 LocalKind::Var | LocalKind::Temp => false,
785 /////////////////////////////////////////////////////////
788 fn dest_prop_mir_dump<'body, 'tcx>(
790 body: &'body Body<'tcx>,
791 live: &mut ResultsCursor<'body, 'tcx, MaybeLiveLocals>,
794 let mut reachable = None;
795 dump_mir(tcx, false, "DestinationPropagation-dataflow", &round, body, |pass_where, w| {
796 let reachable = reachable.get_or_insert_with(|| traversal::reachable_as_bitset(body));
799 PassWhere::BeforeLocation(loc) if reachable.contains(loc.block) => {
800 live.seek_after_primary_effect(loc);
801 writeln!(w, " // live: {:?}", live.get())?;
803 PassWhere::AfterTerminator(bb) if reachable.contains(bb) => {
804 let loc = body.terminator_loc(bb);
805 live.seek_before_primary_effect(loc);
806 writeln!(w, " // live: {:?}", live.get())?;
809 PassWhere::BeforeBlock(bb) if reachable.contains(bb) => {
810 live.seek_to_block_start(bb);
811 writeln!(w, " // live: {:?}", live.get())?;
814 PassWhere::BeforeCFG | PassWhere::AfterCFG | PassWhere::AfterLocation(_) => {}
816 PassWhere::BeforeLocation(_) | PassWhere::AfterTerminator(_) => {
817 writeln!(w, " // live: <unreachable>")?;
820 PassWhere::BeforeBlock(_) => {
821 writeln!(w, " // live: <unreachable>")?;