diff options
Diffstat (limited to '')
-rw-r--r-- | compiler/rustc_mir_transform/src/dest_prop.rs | 917 |
1 files changed, 917 insertions, 0 deletions
diff --git a/compiler/rustc_mir_transform/src/dest_prop.rs b/compiler/rustc_mir_transform/src/dest_prop.rs new file mode 100644 index 000000000..33572068f --- /dev/null +++ b/compiler/rustc_mir_transform/src/dest_prop.rs @@ -0,0 +1,917 @@ +//! Propagates assignment destinations backwards in the CFG to eliminate redundant assignments. +//! +//! # Motivation +//! +//! MIR building can insert a lot of redundant copies, and Rust code in general often tends to move +//! values around a lot. The result is a lot of assignments of the form `dest = {move} src;` in MIR. +//! MIR building for constants in particular tends to create additional locals that are only used +//! inside a single block to shuffle a value around unnecessarily. +//! +//! LLVM by itself is not good enough at eliminating these redundant copies (eg. see +//! <https://github.com/rust-lang/rust/issues/32966>), so this leaves some performance on the table +//! that we can regain by implementing an optimization for removing these assign statements in rustc +//! itself. When this optimization runs fast enough, it can also speed up the constant evaluation +//! and code generation phases of rustc due to the reduced number of statements and locals. +//! +//! # The Optimization +//! +//! Conceptually, this optimization is "destination propagation". It is similar to the Named Return +//! Value Optimization, or NRVO, known from the C++ world, except that it isn't limited to return +//! values or the return place `_0`. On a very high level, independent of the actual implementation +//! details, it does the following: +//! +//! 1) Identify `dest = src;` statements that can be soundly eliminated. +//! 2) Replace all mentions of `src` with `dest` ("unifying" them and propagating the destination +//! backwards). +//! 3) Delete the `dest = src;` statement (by making it a `nop`). +//! +//! Step 1) is by far the hardest, so it is explained in more detail below. +//! +//! ## Soundness +//! +//! Given an `Assign` statement `dest = src;`, where `dest` is a `Place` and `src` is an `Rvalue`, +//! there are a few requirements that must hold for the optimization to be sound: +//! +//! * `dest` must not contain any *indirection* through a pointer. It must access part of the base +//! local. Otherwise it might point to arbitrary memory that is hard to track. +//! +//! It must also not contain any indexing projections, since those take an arbitrary `Local` as +//! the index, and that local might only be initialized shortly before `dest` is used. +//! +//! * `src` must be a bare `Local` without any indirections or field projections (FIXME: Is this a +//! fundamental restriction or just current impl state?). It can be copied or moved by the +//! assignment. +//! +//! * The `dest` and `src` locals must never be [*live*][liveness] at the same time. If they are, it +//! means that they both hold a (potentially different) value that is needed by a future use of +//! the locals. Unifying them would overwrite one of the values. +//! +//! Note that computing liveness of locals that have had their address taken is more difficult: +//! Short of doing full escape analysis on the address/pointer/reference, the pass would need to +//! assume that any operation that can potentially involve opaque user code (such as function +//! calls, destructors, and inline assembly) may access any local that had its address taken +//! before that point. +//! +//! Here, the first two conditions are simple structural requirements on the `Assign` statements +//! that can be trivially checked. The liveness requirement however is more difficult and costly to +//! check. +//! +//! ## Previous Work +//! +//! A [previous attempt] at implementing an optimization like this turned out to be a significant +//! regression in compiler performance. Fixing the regressions introduced a lot of undesirable +//! complexity to the implementation. +//! +//! A [subsequent approach] tried to avoid the costly computation by limiting itself to acyclic +//! CFGs, but still turned out to be far too costly to run due to suboptimal performance within +//! individual basic blocks, requiring a walk across the entire block for every assignment found +//! within the block. For the `tuple-stress` benchmark, which has 458745 statements in a single +//! block, this proved to be far too costly. +//! +//! Since the first attempt at this, the compiler has improved dramatically, and new analysis +//! frameworks have been added that should make this approach viable without requiring a limited +//! approach that only works for some classes of CFGs: +//! - rustc now has a powerful dataflow analysis framework that can handle forwards and backwards +//! analyses efficiently. +//! - Layout optimizations for generators have been added to improve code generation for +//! async/await, which are very similar in spirit to what this optimization does. Both walk the +//! MIR and record conflicting uses of locals in a `BitMatrix`. +//! +//! Also, rustc now has a simple NRVO pass (see `nrvo.rs`), which handles a subset of the cases that +//! this destination propagation pass handles, proving that similar optimizations can be performed +//! on MIR. +//! +//! ## Pre/Post Optimization +//! +//! It is recommended to run `SimplifyCfg` and then `SimplifyLocals` some time after this pass, as +//! it replaces the eliminated assign statements with `nop`s and leaves unused locals behind. +//! +//! [liveness]: https://en.wikipedia.org/wiki/Live_variable_analysis +//! [previous attempt]: https://github.com/rust-lang/rust/pull/47954 +//! [subsequent approach]: https://github.com/rust-lang/rust/pull/71003 + +use crate::MirPass; +use itertools::Itertools; +use rustc_data_structures::unify::{InPlaceUnificationTable, UnifyKey}; +use rustc_index::{ + bit_set::{BitMatrix, BitSet}, + vec::IndexVec, +}; +use rustc_middle::mir::visit::{MutVisitor, PlaceContext, Visitor}; +use rustc_middle::mir::{dump_mir, PassWhere}; +use rustc_middle::mir::{ + traversal, Body, InlineAsmOperand, Local, LocalKind, Location, Operand, Place, PlaceElem, + Rvalue, Statement, StatementKind, Terminator, TerminatorKind, +}; +use rustc_middle::ty::TyCtxt; +use rustc_mir_dataflow::impls::{borrowed_locals, MaybeInitializedLocals, MaybeLiveLocals}; +use rustc_mir_dataflow::Analysis; + +// Empirical measurements have resulted in some observations: +// - Running on a body with a single block and 500 locals takes barely any time +// - Running on a body with ~400 blocks and ~300 relevant locals takes "too long" +// ...so we just limit both to somewhat reasonable-ish looking values. +const MAX_LOCALS: usize = 500; +const MAX_BLOCKS: usize = 250; + +pub struct DestinationPropagation; + +impl<'tcx> MirPass<'tcx> for DestinationPropagation { + fn is_enabled(&self, sess: &rustc_session::Session) -> bool { + // FIXME(#79191, #82678): This is unsound. + // + // Only run at mir-opt-level=3 or higher for now (we don't fix up debuginfo and remove + // storage statements at the moment). + sess.opts.unstable_opts.unsound_mir_opts && sess.mir_opt_level() >= 3 + } + + fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) { + let def_id = body.source.def_id(); + + let candidates = find_candidates(body); + if candidates.is_empty() { + debug!("{:?}: no dest prop candidates, done", def_id); + return; + } + + // Collect all locals we care about. We only compute conflicts for these to save time. + let mut relevant_locals = BitSet::new_empty(body.local_decls.len()); + for CandidateAssignment { dest, src, loc: _ } in &candidates { + relevant_locals.insert(dest.local); + relevant_locals.insert(*src); + } + + // This pass unfortunately has `O(l² * s)` performance, where `l` is the number of locals + // and `s` is the number of statements and terminators in the function. + // To prevent blowing up compile times too much, we bail out when there are too many locals. + let relevant = relevant_locals.count(); + debug!( + "{:?}: {} locals ({} relevant), {} blocks", + def_id, + body.local_decls.len(), + relevant, + body.basic_blocks().len() + ); + if relevant > MAX_LOCALS { + warn!( + "too many candidate locals in {:?} ({}, max is {}), not optimizing", + def_id, relevant, MAX_LOCALS + ); + return; + } + if body.basic_blocks().len() > MAX_BLOCKS { + warn!( + "too many blocks in {:?} ({}, max is {}), not optimizing", + def_id, + body.basic_blocks().len(), + MAX_BLOCKS + ); + return; + } + + let mut conflicts = Conflicts::build(tcx, body, &relevant_locals); + + let mut replacements = Replacements::new(body.local_decls.len()); + for candidate @ CandidateAssignment { dest, src, loc } in candidates { + // Merge locals that don't conflict. + if !conflicts.can_unify(dest.local, src) { + debug!("at assignment {:?}, conflict {:?} vs. {:?}", loc, dest.local, src); + continue; + } + + if replacements.for_src(candidate.src).is_some() { + debug!("src {:?} already has replacement", candidate.src); + continue; + } + + if !tcx.consider_optimizing(|| { + format!("DestinationPropagation {:?} {:?}", def_id, candidate) + }) { + break; + } + + replacements.push(candidate); + conflicts.unify(candidate.src, candidate.dest.local); + } + + replacements.flatten(tcx); + + debug!("replacements {:?}", replacements.map); + + Replacer { tcx, replacements, place_elem_cache: Vec::new() }.visit_body(body); + + // FIXME fix debug info + } +} + +#[derive(Debug, Eq, PartialEq, Copy, Clone)] +struct UnifyLocal(Local); + +impl From<Local> for UnifyLocal { + fn from(l: Local) -> Self { + Self(l) + } +} + +impl UnifyKey for UnifyLocal { + type Value = (); + #[inline] + fn index(&self) -> u32 { + self.0.as_u32() + } + #[inline] + fn from_index(u: u32) -> Self { + Self(Local::from_u32(u)) + } + fn tag() -> &'static str { + "UnifyLocal" + } +} + +struct Replacements<'tcx> { + /// Maps locals to their replacement. + map: IndexVec<Local, Option<Place<'tcx>>>, + + /// Whose locals' live ranges to kill. + kill: BitSet<Local>, +} + +impl<'tcx> Replacements<'tcx> { + fn new(locals: usize) -> Self { + Self { map: IndexVec::from_elem_n(None, locals), kill: BitSet::new_empty(locals) } + } + + fn push(&mut self, candidate: CandidateAssignment<'tcx>) { + trace!("Replacements::push({:?})", candidate); + let entry = &mut self.map[candidate.src]; + assert!(entry.is_none()); + + *entry = Some(candidate.dest); + self.kill.insert(candidate.src); + self.kill.insert(candidate.dest.local); + } + + /// Applies the stored replacements to all replacements, until no replacements would result in + /// locals that need further replacements when applied. + fn flatten(&mut self, tcx: TyCtxt<'tcx>) { + // Note: This assumes that there are no cycles in the replacements, which is enforced via + // `self.unified_locals`. Otherwise this can cause an infinite loop. + + for local in self.map.indices() { + if let Some(replacement) = self.map[local] { + // Substitute the base local of `replacement` until fixpoint. + let mut base = replacement.local; + let mut reversed_projection_slices = Vec::with_capacity(1); + while let Some(replacement_for_replacement) = self.map[base] { + base = replacement_for_replacement.local; + reversed_projection_slices.push(replacement_for_replacement.projection); + } + + let projection: Vec<_> = reversed_projection_slices + .iter() + .rev() + .flat_map(|projs| projs.iter()) + .chain(replacement.projection.iter()) + .collect(); + let projection = tcx.intern_place_elems(&projection); + + // Replace with the final `Place`. + self.map[local] = Some(Place { local: base, projection }); + } + } + } + + fn for_src(&self, src: Local) -> Option<Place<'tcx>> { + self.map[src] + } +} + +struct Replacer<'tcx> { + tcx: TyCtxt<'tcx>, + replacements: Replacements<'tcx>, + place_elem_cache: Vec<PlaceElem<'tcx>>, +} + +impl<'tcx> MutVisitor<'tcx> for Replacer<'tcx> { + fn tcx(&self) -> TyCtxt<'tcx> { + self.tcx + } + + fn visit_local(&mut self, local: &mut Local, context: PlaceContext, location: Location) { + if context.is_use() && self.replacements.for_src(*local).is_some() { + bug!( + "use of local {:?} should have been replaced by visit_place; context={:?}, loc={:?}", + local, + context, + location, + ); + } + } + + fn visit_place(&mut self, place: &mut Place<'tcx>, context: PlaceContext, location: Location) { + if let Some(replacement) = self.replacements.for_src(place.local) { + // Rebase `place`s projections onto `replacement`'s. + self.place_elem_cache.clear(); + self.place_elem_cache.extend(replacement.projection.iter().chain(place.projection)); + let projection = self.tcx.intern_place_elems(&self.place_elem_cache); + let new_place = Place { local: replacement.local, projection }; + + debug!("Replacer: {:?} -> {:?}", place, new_place); + *place = new_place; + } + + self.super_place(place, context, location); + } + + fn visit_statement(&mut self, statement: &mut Statement<'tcx>, location: Location) { + self.super_statement(statement, location); + + match &statement.kind { + // FIXME: Don't delete storage statements, merge the live ranges instead + StatementKind::StorageDead(local) | StatementKind::StorageLive(local) + if self.replacements.kill.contains(*local) => + { + statement.make_nop() + } + + StatementKind::Assign(box (dest, rvalue)) => { + match rvalue { + Rvalue::Use(Operand::Copy(place) | Operand::Move(place)) => { + // These might've been turned into self-assignments by the replacement + // (this includes the original statement we wanted to eliminate). + if dest == place { + debug!("{:?} turned into self-assignment, deleting", location); + statement.make_nop(); + } + } + _ => {} + } + } + + _ => {} + } + } +} + +struct Conflicts<'a> { + relevant_locals: &'a BitSet<Local>, + + /// The conflict matrix. It is always symmetric and the adjacency matrix of the corresponding + /// conflict graph. + matrix: BitMatrix<Local, Local>, + + /// Preallocated `BitSet` used by `unify`. + unify_cache: BitSet<Local>, + + /// Tracks locals that have been merged together to prevent cycles and propagate conflicts. + unified_locals: InPlaceUnificationTable<UnifyLocal>, +} + +impl<'a> Conflicts<'a> { + fn build<'tcx>( + tcx: TyCtxt<'tcx>, + body: &'_ Body<'tcx>, + relevant_locals: &'a BitSet<Local>, + ) -> Self { + // We don't have to look out for locals that have their address taken, since + // `find_candidates` already takes care of that. + + let conflicts = BitMatrix::from_row_n( + &BitSet::new_empty(body.local_decls.len()), + body.local_decls.len(), + ); + + let mut init = MaybeInitializedLocals + .into_engine(tcx, body) + .iterate_to_fixpoint() + .into_results_cursor(body); + let mut live = + MaybeLiveLocals.into_engine(tcx, body).iterate_to_fixpoint().into_results_cursor(body); + + let mut reachable = None; + dump_mir(tcx, None, "DestinationPropagation-dataflow", &"", body, |pass_where, w| { + let reachable = reachable.get_or_insert_with(|| traversal::reachable_as_bitset(body)); + + match pass_where { + PassWhere::BeforeLocation(loc) if reachable.contains(loc.block) => { + init.seek_before_primary_effect(loc); + live.seek_after_primary_effect(loc); + + writeln!(w, " // init: {:?}", init.get())?; + writeln!(w, " // live: {:?}", live.get())?; + } + PassWhere::AfterTerminator(bb) if reachable.contains(bb) => { + let loc = body.terminator_loc(bb); + init.seek_after_primary_effect(loc); + live.seek_before_primary_effect(loc); + + writeln!(w, " // init: {:?}", init.get())?; + writeln!(w, " // live: {:?}", live.get())?; + } + + PassWhere::BeforeBlock(bb) if reachable.contains(bb) => { + init.seek_to_block_start(bb); + live.seek_to_block_start(bb); + + writeln!(w, " // init: {:?}", init.get())?; + writeln!(w, " // live: {:?}", live.get())?; + } + + PassWhere::BeforeCFG | PassWhere::AfterCFG | PassWhere::AfterLocation(_) => {} + + PassWhere::BeforeLocation(_) | PassWhere::AfterTerminator(_) => { + writeln!(w, " // init: <unreachable>")?; + writeln!(w, " // live: <unreachable>")?; + } + + PassWhere::BeforeBlock(_) => { + writeln!(w, " // init: <unreachable>")?; + writeln!(w, " // live: <unreachable>")?; + } + } + + Ok(()) + }); + + let mut this = Self { + relevant_locals, + matrix: conflicts, + unify_cache: BitSet::new_empty(body.local_decls.len()), + unified_locals: { + let mut table = InPlaceUnificationTable::new(); + // Pre-fill table with all locals (this creates N nodes / "connected" components, + // "graph"-ically speaking). + for local in 0..body.local_decls.len() { + assert_eq!(table.new_key(()), UnifyLocal(Local::from_usize(local))); + } + table + }, + }; + + let mut live_and_init_locals = Vec::new(); + + // Visit only reachable basic blocks. The exact order is not important. + for (block, data) in traversal::preorder(body) { + // We need to observe the dataflow state *before* all possible locations (statement or + // terminator) in each basic block, and then observe the state *after* the terminator + // effect is applied. As long as neither `init` nor `borrowed` has a "before" effect, + // we will observe all possible dataflow states. + + // Since liveness is a backwards analysis, we need to walk the results backwards. To do + // that, we first collect in the `MaybeInitializedLocals` results in a forwards + // traversal. + + live_and_init_locals.resize_with(data.statements.len() + 1, || { + BitSet::new_empty(body.local_decls.len()) + }); + + // First, go forwards for `MaybeInitializedLocals` and apply intra-statement/terminator + // conflicts. + for (i, statement) in data.statements.iter().enumerate() { + this.record_statement_conflicts(statement); + + let loc = Location { block, statement_index: i }; + init.seek_before_primary_effect(loc); + + live_and_init_locals[i].clone_from(init.get()); + } + + this.record_terminator_conflicts(data.terminator()); + let term_loc = Location { block, statement_index: data.statements.len() }; + init.seek_before_primary_effect(term_loc); + live_and_init_locals[term_loc.statement_index].clone_from(init.get()); + + // Now, go backwards and union with the liveness results. + for statement_index in (0..=data.statements.len()).rev() { + let loc = Location { block, statement_index }; + live.seek_after_primary_effect(loc); + + live_and_init_locals[statement_index].intersect(live.get()); + + trace!("record conflicts at {:?}", loc); + + this.record_dataflow_conflicts(&mut live_and_init_locals[statement_index]); + } + + init.seek_to_block_end(block); + live.seek_to_block_end(block); + let mut conflicts = init.get().clone(); + conflicts.intersect(live.get()); + trace!("record conflicts at end of {:?}", block); + + this.record_dataflow_conflicts(&mut conflicts); + } + + this + } + + fn record_dataflow_conflicts(&mut self, new_conflicts: &mut BitSet<Local>) { + // Remove all locals that are not candidates. + new_conflicts.intersect(self.relevant_locals); + + for local in new_conflicts.iter() { + self.matrix.union_row_with(&new_conflicts, local); + } + } + + fn record_local_conflict(&mut self, a: Local, b: Local, why: &str) { + trace!("conflict {:?} <-> {:?} due to {}", a, b, why); + self.matrix.insert(a, b); + self.matrix.insert(b, a); + } + + /// Records locals that must not overlap during the evaluation of `stmt`. These locals conflict + /// and must not be merged. + fn record_statement_conflicts(&mut self, stmt: &Statement<'_>) { + match &stmt.kind { + // While the left and right sides of an assignment must not overlap, we do not mark + // conflicts here as that would make this optimization useless. When we optimize, we + // eliminate the resulting self-assignments automatically. + StatementKind::Assign(_) => {} + + StatementKind::SetDiscriminant { .. } + | StatementKind::Deinit(..) + | StatementKind::StorageLive(..) + | StatementKind::StorageDead(..) + | StatementKind::Retag(..) + | StatementKind::FakeRead(..) + | StatementKind::AscribeUserType(..) + | StatementKind::Coverage(..) + | StatementKind::CopyNonOverlapping(..) + | StatementKind::Nop => {} + } + } + + fn record_terminator_conflicts(&mut self, term: &Terminator<'_>) { + match &term.kind { + TerminatorKind::DropAndReplace { + place: dropped_place, + value, + target: _, + unwind: _, + } => { + if let Some(place) = value.place() + && !place.is_indirect() + && !dropped_place.is_indirect() + { + self.record_local_conflict( + place.local, + dropped_place.local, + "DropAndReplace operand overlap", + ); + } + } + TerminatorKind::Yield { value, resume: _, resume_arg, drop: _ } => { + if let Some(place) = value.place() { + if !place.is_indirect() && !resume_arg.is_indirect() { + self.record_local_conflict( + place.local, + resume_arg.local, + "Yield operand overlap", + ); + } + } + } + TerminatorKind::Call { + func, + args, + destination, + target: _, + cleanup: _, + from_hir_call: _, + fn_span: _, + } => { + // No arguments may overlap with the destination. + for arg in args.iter().chain(Some(func)) { + if let Some(place) = arg.place() { + if !place.is_indirect() && !destination.is_indirect() { + self.record_local_conflict( + destination.local, + place.local, + "call dest/arg overlap", + ); + } + } + } + } + TerminatorKind::InlineAsm { + template: _, + operands, + options: _, + line_spans: _, + destination: _, + cleanup: _, + } => { + // The intended semantics here aren't documented, we just assume that nothing that + // could be written to by the assembly may overlap with any other operands. + for op in operands { + match op { + InlineAsmOperand::Out { reg: _, late: _, place: Some(dest_place) } + | InlineAsmOperand::InOut { + reg: _, + late: _, + in_value: _, + out_place: Some(dest_place), + } => { + // For output place `place`, add all places accessed by the inline asm. + for op in operands { + match op { + InlineAsmOperand::In { reg: _, value } => { + if let Some(p) = value.place() + && !p.is_indirect() + && !dest_place.is_indirect() + { + self.record_local_conflict( + p.local, + dest_place.local, + "asm! operand overlap", + ); + } + } + InlineAsmOperand::Out { + reg: _, + late: _, + place: Some(place), + } => { + if !place.is_indirect() && !dest_place.is_indirect() { + self.record_local_conflict( + place.local, + dest_place.local, + "asm! operand overlap", + ); + } + } + InlineAsmOperand::InOut { + reg: _, + late: _, + in_value, + out_place, + } => { + if let Some(place) = in_value.place() + && !place.is_indirect() + && !dest_place.is_indirect() + { + self.record_local_conflict( + place.local, + dest_place.local, + "asm! operand overlap", + ); + } + + if let Some(place) = out_place + && !place.is_indirect() + && !dest_place.is_indirect() + { + self.record_local_conflict( + place.local, + dest_place.local, + "asm! operand overlap", + ); + } + } + InlineAsmOperand::Out { reg: _, late: _, place: None } + | InlineAsmOperand::Const { value: _ } + | InlineAsmOperand::SymFn { value: _ } + | InlineAsmOperand::SymStatic { def_id: _ } => {} + } + } + } + InlineAsmOperand::InOut { + reg: _, + late: _, + in_value: _, + out_place: None, + } + | InlineAsmOperand::In { reg: _, value: _ } + | InlineAsmOperand::Out { reg: _, late: _, place: None } + | InlineAsmOperand::Const { value: _ } + | InlineAsmOperand::SymFn { value: _ } + | InlineAsmOperand::SymStatic { def_id: _ } => {} + } + } + } + + TerminatorKind::Goto { .. } + | TerminatorKind::SwitchInt { .. } + | TerminatorKind::Resume + | TerminatorKind::Abort + | TerminatorKind::Return + | TerminatorKind::Unreachable + | TerminatorKind::Drop { .. } + | TerminatorKind::Assert { .. } + | TerminatorKind::GeneratorDrop + | TerminatorKind::FalseEdge { .. } + | TerminatorKind::FalseUnwind { .. } => {} + } + } + + /// Checks whether `a` and `b` may be merged. Returns `false` if there's a conflict. + fn can_unify(&mut self, a: Local, b: Local) -> bool { + // After some locals have been unified, their conflicts are only tracked in the root key, + // so look that up. + let a = self.unified_locals.find(a).0; + let b = self.unified_locals.find(b).0; + + if a == b { + // Already merged (part of the same connected component). + return false; + } + + if self.matrix.contains(a, b) { + // Conflict (derived via dataflow, intra-statement conflicts, or inherited from another + // local during unification). + return false; + } + + true + } + + /// Merges the conflicts of `a` and `b`, so that each one inherits all conflicts of the other. + /// + /// `can_unify` must have returned `true` for the same locals, or this may panic or lead to + /// miscompiles. + /// + /// This is called when the pass makes the decision to unify `a` and `b` (or parts of `a` and + /// `b`) and is needed to ensure that future unification decisions take potentially newly + /// introduced conflicts into account. + /// + /// For an example, assume we have locals `_0`, `_1`, `_2`, and `_3`. There are these conflicts: + /// + /// * `_0` <-> `_1` + /// * `_1` <-> `_2` + /// * `_3` <-> `_0` + /// + /// We then decide to merge `_2` with `_3` since they don't conflict. Then we decide to merge + /// `_2` with `_0`, which also doesn't have a conflict in the above list. However `_2` is now + /// `_3`, which does conflict with `_0`. + fn unify(&mut self, a: Local, b: Local) { + trace!("unify({:?}, {:?})", a, b); + + // Get the root local of the connected components. The root local stores the conflicts of + // all locals in the connected component (and *is stored* as the conflicting local of other + // locals). + let a = self.unified_locals.find(a).0; + let b = self.unified_locals.find(b).0; + assert_ne!(a, b); + + trace!("roots: a={:?}, b={:?}", a, b); + trace!("{:?} conflicts: {:?}", a, self.matrix.iter(a).format(", ")); + trace!("{:?} conflicts: {:?}", b, self.matrix.iter(b).format(", ")); + + self.unified_locals.union(a, b); + + let root = self.unified_locals.find(a).0; + assert!(root == a || root == b); + + // Make all locals that conflict with `a` also conflict with `b`, and vice versa. + self.unify_cache.clear(); + for conflicts_with_a in self.matrix.iter(a) { + self.unify_cache.insert(conflicts_with_a); + } + for conflicts_with_b in self.matrix.iter(b) { + self.unify_cache.insert(conflicts_with_b); + } + for conflicts_with_a_or_b in self.unify_cache.iter() { + // Set both `a` and `b` for this local's row. + self.matrix.insert(conflicts_with_a_or_b, a); + self.matrix.insert(conflicts_with_a_or_b, b); + } + + // Write the locals `a` conflicts with to `b`'s row. + self.matrix.union_rows(a, b); + // Write the locals `b` conflicts with to `a`'s row. + self.matrix.union_rows(b, a); + } +} + +/// A `dest = {move} src;` statement at `loc`. +/// +/// We want to consider merging `dest` and `src` due to this assignment. +#[derive(Debug, Copy, Clone)] +struct CandidateAssignment<'tcx> { + /// Does not contain indirection or indexing (so the only local it contains is the place base). + dest: Place<'tcx>, + src: Local, + loc: Location, +} + +/// Scans the MIR for assignments between locals that we might want to consider merging. +/// +/// This will filter out assignments that do not match the right form (as described in the top-level +/// comment) and also throw out assignments that involve a local that has its address taken or is +/// otherwise ineligible (eg. locals used as array indices are ignored because we cannot propagate +/// arbitrary places into array indices). +fn find_candidates<'tcx>(body: &Body<'tcx>) -> Vec<CandidateAssignment<'tcx>> { + let mut visitor = FindAssignments { + body, + candidates: Vec::new(), + ever_borrowed_locals: borrowed_locals(body), + locals_used_as_array_index: locals_used_as_array_index(body), + }; + visitor.visit_body(body); + visitor.candidates +} + +struct FindAssignments<'a, 'tcx> { + body: &'a Body<'tcx>, + candidates: Vec<CandidateAssignment<'tcx>>, + ever_borrowed_locals: BitSet<Local>, + locals_used_as_array_index: BitSet<Local>, +} + +impl<'tcx> Visitor<'tcx> for FindAssignments<'_, 'tcx> { + fn visit_statement(&mut self, statement: &Statement<'tcx>, location: Location) { + if let StatementKind::Assign(box ( + dest, + Rvalue::Use(Operand::Copy(src) | Operand::Move(src)), + )) = &statement.kind + { + // `dest` must not have pointer indirection. + if dest.is_indirect() { + return; + } + + // `src` must be a plain local. + if !src.projection.is_empty() { + return; + } + + // Since we want to replace `src` with `dest`, `src` must not be required. + if is_local_required(src.local, self.body) { + return; + } + + // Can't optimize if either local ever has their address taken. This optimization does + // liveness analysis only based on assignments, and a local can be live even if its + // never assigned to again, because a reference to it might be live. + // FIXME: This can be smarter and take `StorageDead` into account (which invalidates + // borrows). + if self.ever_borrowed_locals.contains(dest.local) + || self.ever_borrowed_locals.contains(src.local) + { + return; + } + + assert_ne!(dest.local, src.local, "self-assignments are UB"); + + // We can't replace locals occurring in `PlaceElem::Index` for now. + if self.locals_used_as_array_index.contains(src.local) { + return; + } + + for elem in dest.projection { + if let PlaceElem::Index(_) = elem { + // `dest` contains an indexing projection. + return; + } + } + + self.candidates.push(CandidateAssignment { + dest: *dest, + src: src.local, + loc: location, + }); + } + } +} + +/// Some locals are part of the function's interface and can not be removed. +/// +/// Note that these locals *can* still be merged with non-required locals by removing that other +/// local. +fn is_local_required(local: Local, body: &Body<'_>) -> bool { + match body.local_kind(local) { + LocalKind::Arg | LocalKind::ReturnPointer => true, + LocalKind::Var | LocalKind::Temp => false, + } +} + +/// `PlaceElem::Index` only stores a `Local`, so we can't replace that with a full `Place`. +/// +/// Collect locals used as indices so we don't generate candidates that are impossible to apply +/// later. +fn locals_used_as_array_index(body: &Body<'_>) -> BitSet<Local> { + let mut visitor = IndexCollector { locals: BitSet::new_empty(body.local_decls.len()) }; + visitor.visit_body(body); + visitor.locals +} + +struct IndexCollector { + locals: BitSet<Local>, +} + +impl<'tcx> Visitor<'tcx> for IndexCollector { + fn visit_projection_elem( + &mut self, + local: Local, + proj_base: &[PlaceElem<'tcx>], + elem: PlaceElem<'tcx>, + context: PlaceContext, + location: Location, + ) { + if let PlaceElem::Index(i) = elem { + self.locals.insert(i); + } + self.super_projection_elem(local, proj_base, elem, context, location); + } +} |