1 //! A number of passes which remove various redundancies in the CFG.
3 //! The `SimplifyCfg` pass gets rid of unnecessary blocks in the CFG, whereas the `SimplifyLocals`
4 //! gets rid of all the unnecessary local variable declarations.
6 //! The `SimplifyLocals` pass is kinda expensive and therefore not very suitable to be run often.
7 //! Most of the passes should not care or be impacted in meaningful ways due to extra locals
8 //! either, so running the pass once, right before codegen, should suffice.
10 //! On the other side of the spectrum, the `SimplifyCfg` pass is considerably cheap to run, thus
11 //! one should run it after every pass which may modify CFG in significant ways. This pass must
12 //! also be run before any analysis passes because it removes dead blocks, and some of these can be
15 //! The cause of this typing issue is typeck allowing most blocks whose end is not reachable have
16 //! an arbitrary return type, rather than having the usual () return type (as a note, typeck's
17 //! notion of reachability is in fact slightly weaker than MIR CFG reachability - see #31617). A
18 //! standard example of the situation is:
22 //! let _a: char = { return; };
26 //! Here the block (`{ return; }`) has the return type `char`, rather than `()`, but the MIR we
27 //! naively generate still contains the `_a = ()` write in the unreachable block "after" the
31 use rustc_data_structures::fx::FxHashSet;
32 use rustc_index::vec::{Idx, IndexVec};
33 use rustc_middle::mir::coverage::*;
34 use rustc_middle::mir::visit::{MutVisitor, MutatingUseContext, PlaceContext, Visitor};
35 use rustc_middle::mir::*;
36 use rustc_middle::ty::TyCtxt;
37 use smallvec::SmallVec;
39 pub struct SimplifyCfg {
44 pub fn new(label: &str) -> Self {
45 SimplifyCfg { label: format!("SimplifyCfg-{}", label) }
49 pub fn simplify_cfg<'tcx>(tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
50 CfgSimplifier::new(body).simplify();
51 remove_dead_blocks(tcx, body);
53 // FIXME: Should probably be moved into some kind of pass manager
54 body.basic_blocks_mut().raw.shrink_to_fit();
57 impl<'tcx> MirPass<'tcx> for SimplifyCfg {
58 fn name(&self) -> &str {
62 fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
63 debug!("SimplifyCfg({:?}) - simplifying {:?}", self.label, body.source);
64 simplify_cfg(tcx, body);
68 pub struct CfgSimplifier<'a, 'tcx> {
69 basic_blocks: &'a mut IndexVec<BasicBlock, BasicBlockData<'tcx>>,
70 pred_count: IndexVec<BasicBlock, u32>,
73 impl<'a, 'tcx> CfgSimplifier<'a, 'tcx> {
74 pub fn new(body: &'a mut Body<'tcx>) -> Self {
75 let mut pred_count = IndexVec::from_elem(0u32, &body.basic_blocks);
77 // we can't use mir.predecessors() here because that counts
78 // dead blocks, which we don't want to.
79 pred_count[START_BLOCK] = 1;
81 for (_, data) in traversal::preorder(body) {
82 if let Some(ref term) = data.terminator {
83 for tgt in term.successors() {
89 let basic_blocks = body.basic_blocks_mut();
91 CfgSimplifier { basic_blocks, pred_count }
94 pub fn simplify(mut self) {
97 // Vec of the blocks that should be merged. We store the indices here, instead of the
98 // statements itself to avoid moving the (relatively) large statements twice.
99 // We do not push the statements directly into the target block (`bb`) as that is slower
100 // due to additional reallocations
101 let mut merged_blocks = Vec::new();
103 let mut changed = false;
105 for bb in self.basic_blocks.indices() {
106 if self.pred_count[bb] == 0 {
110 debug!("simplifying {:?}", bb);
113 self.basic_blocks[bb].terminator.take().expect("invalid terminator state");
115 for successor in terminator.successors_mut() {
116 self.collapse_goto_chain(successor, &mut changed);
119 let mut inner_changed = true;
120 merged_blocks.clear();
121 while inner_changed {
122 inner_changed = false;
123 inner_changed |= self.simplify_branch(&mut terminator);
124 inner_changed |= self.merge_successor(&mut merged_blocks, &mut terminator);
125 changed |= inner_changed;
128 let statements_to_merge =
129 merged_blocks.iter().map(|&i| self.basic_blocks[i].statements.len()).sum();
131 if statements_to_merge > 0 {
132 let mut statements = std::mem::take(&mut self.basic_blocks[bb].statements);
133 statements.reserve(statements_to_merge);
134 for &from in &merged_blocks {
135 statements.append(&mut self.basic_blocks[from].statements);
137 self.basic_blocks[bb].statements = statements;
140 self.basic_blocks[bb].terminator = Some(terminator);
149 /// This function will return `None` if
150 /// * the block has statements
151 /// * the block has a terminator other than `goto`
152 /// * the block has no terminator (meaning some other part of the current optimization stole it)
153 fn take_terminator_if_simple_goto(&mut self, bb: BasicBlock) -> Option<Terminator<'tcx>> {
154 match self.basic_blocks[bb] {
158 ref mut terminator @ Some(Terminator { kind: TerminatorKind::Goto { .. }, .. }),
160 } if statements.is_empty() => terminator.take(),
161 // if `terminator` is None, this means we are in a loop. In that
162 // case, let all the loop collapse to its entry.
167 /// Collapse a goto chain starting from `start`
168 fn collapse_goto_chain(&mut self, start: &mut BasicBlock, changed: &mut bool) {
169 // Using `SmallVec` here, because in some logs on libcore oli-obk saw many single-element
170 // goto chains. We should probably benchmark different sizes.
171 let mut terminators: SmallVec<[_; 1]> = Default::default();
172 let mut current = *start;
173 while let Some(terminator) = self.take_terminator_if_simple_goto(current) {
174 let Terminator { kind: TerminatorKind::Goto { target }, .. } = terminator else {
177 terminators.push((current, terminator));
182 while let Some((current, mut terminator)) = terminators.pop() {
183 let Terminator { kind: TerminatorKind::Goto { ref mut target }, .. } = terminator else {
186 *changed |= *target != last;
188 debug!("collapsing goto chain from {:?} to {:?}", current, target);
190 if self.pred_count[current] == 1 {
191 // This is the last reference to current, so the pred-count to
192 // to target is moved into the current block.
193 self.pred_count[current] = 0;
195 self.pred_count[*target] += 1;
196 self.pred_count[current] -= 1;
198 self.basic_blocks[current].terminator = Some(terminator);
202 // merge a block with 1 `goto` predecessor to its parent
205 merged_blocks: &mut Vec<BasicBlock>,
206 terminator: &mut Terminator<'tcx>,
208 let target = match terminator.kind {
209 TerminatorKind::Goto { target } if self.pred_count[target] == 1 => target,
213 debug!("merging block {:?} into {:?}", target, terminator);
214 *terminator = match self.basic_blocks[target].terminator.take() {
215 Some(terminator) => terminator,
217 // unreachable loop - this should not be possible, as we
218 // don't strand blocks, but handle it correctly.
223 merged_blocks.push(target);
224 self.pred_count[target] = 0;
229 // turn a branch with all successors identical to a goto
230 fn simplify_branch(&mut self, terminator: &mut Terminator<'tcx>) -> bool {
231 match terminator.kind {
232 TerminatorKind::SwitchInt { .. } => {}
237 if let Some(first_succ) = terminator.successors().next() {
238 if terminator.successors().all(|s| s == first_succ) {
239 let count = terminator.successors().count();
240 self.pred_count[first_succ] -= (count - 1) as u32;
250 debug!("simplifying branch {:?}", terminator);
251 terminator.kind = TerminatorKind::Goto { target: first_succ };
255 fn strip_nops(&mut self) {
256 for blk in self.basic_blocks.iter_mut() {
257 blk.statements.retain(|stmt| !matches!(stmt.kind, StatementKind::Nop))
262 pub fn remove_dead_blocks<'tcx>(tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
263 let reachable = traversal::reachable_as_bitset(body);
264 let num_blocks = body.basic_blocks.len();
265 if num_blocks == reachable.count() {
269 let basic_blocks = body.basic_blocks.as_mut();
270 let source_scopes = &body.source_scopes;
271 let mut replacements: Vec<_> = (0..num_blocks).map(BasicBlock::new).collect();
272 let mut used_blocks = 0;
273 for alive_index in reachable.iter() {
274 let alive_index = alive_index.index();
275 replacements[alive_index] = BasicBlock::new(used_blocks);
276 if alive_index != used_blocks {
277 // Swap the next alive block data with the current available slot. Since
278 // alive_index is non-decreasing this is a valid operation.
279 basic_blocks.raw.swap(alive_index, used_blocks);
284 if tcx.sess.instrument_coverage() {
285 save_unreachable_coverage(basic_blocks, source_scopes, used_blocks);
288 basic_blocks.raw.truncate(used_blocks);
290 for block in basic_blocks {
291 for target in block.terminator_mut().successors_mut() {
292 *target = replacements[target.index()];
297 /// Some MIR transforms can determine at compile time that a sequences of
298 /// statements will never be executed, so they can be dropped from the MIR.
299 /// For example, an `if` or `else` block that is guaranteed to never be executed
300 /// because its condition can be evaluated at compile time, such as by const
301 /// evaluation: `if false { ... }`.
303 /// Those statements are bypassed by redirecting paths in the CFG around the
304 /// `dead blocks`; but with `-C instrument-coverage`, the dead blocks usually
305 /// include `Coverage` statements representing the Rust source code regions to
306 /// be counted at runtime. Without these `Coverage` statements, the regions are
307 /// lost, and the Rust source code will show no coverage information.
309 /// What we want to show in a coverage report is the dead code with coverage
310 /// counts of `0`. To do this, we need to save the code regions, by injecting
311 /// `Unreachable` coverage statements. These are non-executable statements whose
312 /// code regions are still recorded in the coverage map, representing regions
313 /// with `0` executions.
315 /// If there are no live `Counter` `Coverage` statements remaining, we remove
316 /// `Coverage` statements along with the dead blocks. Since at least one
317 /// counter per function is required by LLVM (and necessary, to add the
318 /// `function_hash` to the counter's call to the LLVM intrinsic
319 /// `instrprof.increment()`).
321 /// The `generator::StateTransform` MIR pass and MIR inlining can create
322 /// atypical conditions, where all live `Counter`s are dropped from the MIR.
324 /// With MIR inlining we can have coverage counters belonging to different
325 /// instances in a single body, so the strategy described above is applied to
326 /// coverage counters from each instance individually.
327 fn save_unreachable_coverage(
328 basic_blocks: &mut IndexVec<BasicBlock, BasicBlockData<'_>>,
329 source_scopes: &IndexVec<SourceScope, SourceScopeData<'_>>,
330 first_dead_block: usize,
332 // Identify instances that still have some live coverage counters left.
333 let mut live = FxHashSet::default();
334 for basic_block in &basic_blocks.raw[0..first_dead_block] {
335 for statement in &basic_block.statements {
336 let StatementKind::Coverage(coverage) = &statement.kind else { continue };
337 let CoverageKind::Counter { .. } = coverage.kind else { continue };
338 let instance = statement.source_info.scope.inlined_instance(source_scopes);
339 live.insert(instance);
343 for block in &mut basic_blocks.raw[..first_dead_block] {
344 for statement in &mut block.statements {
345 let StatementKind::Coverage(_) = &statement.kind else { continue };
346 let instance = statement.source_info.scope.inlined_instance(source_scopes);
347 if !live.contains(&instance) {
348 statement.make_nop();
357 // Retain coverage for instances that still have some live counters left.
358 let mut retained_coverage = Vec::new();
359 for dead_block in &basic_blocks.raw[first_dead_block..] {
360 for statement in &dead_block.statements {
361 let StatementKind::Coverage(coverage) = &statement.kind else { continue };
362 let Some(code_region) = &coverage.code_region else { continue };
363 let instance = statement.source_info.scope.inlined_instance(source_scopes);
364 if live.contains(&instance) {
365 retained_coverage.push((statement.source_info, code_region.clone()));
370 let start_block = &mut basic_blocks[START_BLOCK];
371 start_block.statements.extend(retained_coverage.into_iter().map(
372 |(source_info, code_region)| Statement {
374 kind: StatementKind::Coverage(Box::new(Coverage {
375 kind: CoverageKind::Unreachable,
376 code_region: Some(code_region),
382 pub struct SimplifyLocals {
386 impl SimplifyLocals {
387 pub fn new(label: &str) -> SimplifyLocals {
388 SimplifyLocals { label: format!("SimplifyLocals-{}", label) }
392 impl<'tcx> MirPass<'tcx> for SimplifyLocals {
393 fn name(&self) -> &str {
397 fn is_enabled(&self, sess: &rustc_session::Session) -> bool {
398 sess.mir_opt_level() > 0
401 fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
402 trace!("running SimplifyLocals on {:?}", body.source);
403 simplify_locals(body, tcx);
407 pub fn simplify_locals<'tcx>(body: &mut Body<'tcx>, tcx: TyCtxt<'tcx>) {
408 // First, we're going to get a count of *actual* uses for every `Local`.
409 let mut used_locals = UsedLocals::new(body);
411 // Next, we're going to remove any `Local` with zero actual uses. When we remove those
412 // `Locals`, we're also going to subtract any uses of other `Locals` from the `used_locals`
413 // count. For example, if we removed `_2 = discriminant(_1)`, then we'll subtract one from
414 // `use_counts[_1]`. That in turn might make `_1` unused, so we loop until we hit a
415 // fixedpoint where there are no more unused locals.
416 remove_unused_definitions(&mut used_locals, body);
418 // Finally, we'll actually do the work of shrinking `body.local_decls` and remapping the `Local`s.
419 let map = make_local_map(&mut body.local_decls, &used_locals);
421 // Only bother running the `LocalUpdater` if we actually found locals to remove.
422 if map.iter().any(Option::is_none) {
423 // Update references to all vars and tmps now
424 let mut updater = LocalUpdater { map, tcx };
425 updater.visit_body_preserves_cfg(body);
427 body.local_decls.shrink_to_fit();
431 /// Construct the mapping while swapping out unused stuff out from the `vec`.
432 fn make_local_map<V>(
433 local_decls: &mut IndexVec<Local, V>,
434 used_locals: &UsedLocals,
435 ) -> IndexVec<Local, Option<Local>> {
436 let mut map: IndexVec<Local, Option<Local>> = IndexVec::from_elem(None, &*local_decls);
437 let mut used = Local::new(0);
439 for alive_index in local_decls.indices() {
440 // `is_used` treats the `RETURN_PLACE` and arguments as used.
441 if !used_locals.is_used(alive_index) {
445 map[alive_index] = Some(used);
446 if alive_index != used {
447 local_decls.swap(alive_index, used);
449 used.increment_by(1);
451 local_decls.truncate(used.index());
455 /// Keeps track of used & unused locals.
459 use_count: IndexVec<Local, u32>,
463 /// Determines which locals are used & unused in the given body.
464 fn new(body: &Body<'_>) -> Self {
465 let mut this = Self {
467 arg_count: body.arg_count.try_into().unwrap(),
468 use_count: IndexVec::from_elem(0, &body.local_decls),
470 this.visit_body(body);
474 /// Checks if local is used.
476 /// Return place and arguments are always considered used.
477 fn is_used(&self, local: Local) -> bool {
478 trace!("is_used({:?}): use_count: {:?}", local, self.use_count[local]);
479 local.as_u32() <= self.arg_count || self.use_count[local] != 0
482 /// Updates the use counts to reflect the removal of given statement.
483 fn statement_removed(&mut self, statement: &Statement<'_>) {
484 self.increment = false;
486 // The location of the statement is irrelevant.
487 let location = Location { block: START_BLOCK, statement_index: 0 };
488 self.visit_statement(statement, location);
491 /// Visits a left-hand side of an assignment.
492 fn visit_lhs(&mut self, place: &Place<'_>, location: Location) {
493 if place.is_indirect() {
494 // A use, not a definition.
495 self.visit_place(place, PlaceContext::MutatingUse(MutatingUseContext::Store), location);
497 // A definition. The base local itself is not visited, so this occurrence is not counted
498 // toward its use count. There might be other locals still, used in an indexing
500 self.super_projection(
502 PlaceContext::MutatingUse(MutatingUseContext::Projection),
509 impl<'tcx> Visitor<'tcx> for UsedLocals {
510 fn visit_statement(&mut self, statement: &Statement<'tcx>, location: Location) {
511 match statement.kind {
512 StatementKind::Intrinsic(..)
513 | StatementKind::Retag(..)
514 | StatementKind::Coverage(..)
515 | StatementKind::FakeRead(..)
516 | StatementKind::AscribeUserType(..) => {
517 self.super_statement(statement, location);
520 StatementKind::Nop => {}
522 StatementKind::StorageLive(_local) | StatementKind::StorageDead(_local) => {}
524 StatementKind::Assign(box (ref place, ref rvalue)) => {
525 if rvalue.is_safe_to_remove() {
526 self.visit_lhs(place, location);
527 self.visit_rvalue(rvalue, location);
529 self.super_statement(statement, location);
533 StatementKind::SetDiscriminant { ref place, variant_index: _ }
534 | StatementKind::Deinit(ref place) => {
535 self.visit_lhs(place, location);
540 fn visit_local(&mut self, local: Local, _ctx: PlaceContext, _location: Location) {
542 self.use_count[local] += 1;
544 assert_ne!(self.use_count[local], 0);
545 self.use_count[local] -= 1;
550 /// Removes unused definitions. Updates the used locals to reflect the changes made.
551 fn remove_unused_definitions(used_locals: &mut UsedLocals, body: &mut Body<'_>) {
552 // The use counts are updated as we remove the statements. A local might become unused
553 // during the retain operation, leading to a temporary inconsistency (storage statements or
554 // definitions referencing the local might remain). For correctness it is crucial that this
555 // computation reaches a fixed point.
557 let mut modified = true;
561 for data in body.basic_blocks.as_mut_preserves_cfg() {
562 // Remove unnecessary StorageLive and StorageDead annotations.
563 data.statements.retain(|statement| {
564 let keep = match &statement.kind {
565 StatementKind::StorageLive(local) | StatementKind::StorageDead(local) => {
566 used_locals.is_used(*local)
568 StatementKind::Assign(box (place, _)) => used_locals.is_used(place.local),
570 StatementKind::SetDiscriminant { ref place, .. }
571 | StatementKind::Deinit(ref place) => used_locals.is_used(place.local),
576 trace!("removing statement {:?}", statement);
578 used_locals.statement_removed(statement);
587 struct LocalUpdater<'tcx> {
588 map: IndexVec<Local, Option<Local>>,
592 impl<'tcx> MutVisitor<'tcx> for LocalUpdater<'tcx> {
593 fn tcx(&self) -> TyCtxt<'tcx> {
597 fn visit_local(&mut self, l: &mut Local, _: PlaceContext, _: Location) {
598 *l = self.map[*l].unwrap();