rust/compiler/rustc_mir_transform/src/simplify.rs
Tomasz Miąsko 4d614e1d1f Do not collapse goto chains beginning with the start block
If any block on a goto chain has more than one predecessor, then the new
start block would have basic block predecessors.

Skip the transformation for the start block altogether, to avoid
violating the new invariant that the start block does not have any basic
block predecessors.
2021-09-18 07:28:55 +02:00

568 lines
21 KiB
Rust

//! A number of passes which remove various redundancies in the CFG.
//!
//! The `SimplifyCfg` pass gets rid of unnecessary blocks in the CFG, whereas the `SimplifyLocals`
//! gets rid of all the unnecessary local variable declarations.
//!
//! The `SimplifyLocals` pass is kinda expensive and therefore not very suitable to be run often.
//! Most of the passes should not care or be impacted in meaningful ways due to extra locals
//! either, so running the pass once, right before codegen, should suffice.
//!
//! On the other side of the spectrum, the `SimplifyCfg` pass is considerably cheap to run, thus
//! one should run it after every pass which may modify CFG in significant ways. This pass must
//! also be run before any analysis passes because it removes dead blocks, and some of these can be
//! ill-typed.
//!
//! The cause of this typing issue is typeck allowing most blocks whose end is not reachable have
//! an arbitrary return type, rather than having the usual () return type (as a note, typeck's
//! notion of reachability is in fact slightly weaker than MIR CFG reachability - see #31617). A
//! standard example of the situation is:
//!
//! ```rust
//! fn example() {
//! let _a: char = { return; };
//! }
//! ```
//!
//! Here the block (`{ return; }`) has the return type `char`, rather than `()`, but the MIR we
//! naively generate still contains the `_a = ()` write in the unreachable block "after" the
//! return.
use crate::MirPass;
use rustc_index::vec::{Idx, IndexVec};
use rustc_middle::mir::coverage::*;
use rustc_middle::mir::visit::{MutVisitor, MutatingUseContext, PlaceContext, Visitor};
use rustc_middle::mir::*;
use rustc_middle::ty::TyCtxt;
use smallvec::SmallVec;
use std::borrow::Cow;
use std::convert::TryInto;
pub struct SimplifyCfg {
label: String,
}
impl SimplifyCfg {
pub fn new(label: &str) -> Self {
SimplifyCfg { label: format!("SimplifyCfg-{}", label) }
}
}
pub fn simplify_cfg(tcx: TyCtxt<'tcx>, body: &mut Body<'_>) {
CfgSimplifier::new(body).simplify();
remove_dead_blocks(tcx, body);
// FIXME: Should probably be moved into some kind of pass manager
body.basic_blocks_mut().raw.shrink_to_fit();
}
impl<'tcx> MirPass<'tcx> for SimplifyCfg {
fn name(&self) -> Cow<'_, str> {
Cow::Borrowed(&self.label)
}
fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
debug!("SimplifyCfg({:?}) - simplifying {:?}", self.label, body.source);
simplify_cfg(tcx, body);
}
}
pub struct CfgSimplifier<'a, 'tcx> {
basic_blocks: &'a mut IndexVec<BasicBlock, BasicBlockData<'tcx>>,
pred_count: IndexVec<BasicBlock, u32>,
}
impl<'a, 'tcx> CfgSimplifier<'a, 'tcx> {
pub fn new(body: &'a mut Body<'tcx>) -> Self {
let mut pred_count = IndexVec::from_elem(0u32, body.basic_blocks());
// we can't use mir.predecessors() here because that counts
// dead blocks, which we don't want to.
pred_count[START_BLOCK] = 1;
for (_, data) in traversal::preorder(body) {
if let Some(ref term) = data.terminator {
for &tgt in term.successors() {
pred_count[tgt] += 1;
}
}
}
let basic_blocks = body.basic_blocks_mut();
CfgSimplifier { basic_blocks, pred_count }
}
pub fn simplify(mut self) {
self.strip_nops();
// Vec of the blocks that should be merged. We store the indices here, instead of the
// statements itself to avoid moving the (relatively) large statements twice.
// We do not push the statements directly into the target block (`bb`) as that is slower
// due to additional reallocations
let mut merged_blocks = Vec::new();
loop {
let mut changed = false;
for bb in self.basic_blocks.indices() {
if self.pred_count[bb] == 0 {
continue;
}
debug!("simplifying {:?}", bb);
let mut terminator =
self.basic_blocks[bb].terminator.take().expect("invalid terminator state");
for successor in terminator.successors_mut() {
self.collapse_goto_chain(successor, &mut changed);
}
let mut inner_changed = true;
merged_blocks.clear();
while inner_changed {
inner_changed = false;
inner_changed |= self.simplify_branch(&mut terminator);
inner_changed |= self.merge_successor(&mut merged_blocks, &mut terminator);
changed |= inner_changed;
}
let statements_to_merge =
merged_blocks.iter().map(|&i| self.basic_blocks[i].statements.len()).sum();
if statements_to_merge > 0 {
let mut statements = std::mem::take(&mut self.basic_blocks[bb].statements);
statements.reserve(statements_to_merge);
for &from in &merged_blocks {
statements.append(&mut self.basic_blocks[from].statements);
}
self.basic_blocks[bb].statements = statements;
}
self.basic_blocks[bb].terminator = Some(terminator);
}
if !changed {
break;
}
}
}
/// This function will return `None` if
/// * the block has statements
/// * the block has a terminator other than `goto`
/// * the block has no terminator (meaning some other part of the current optimization stole it)
fn take_terminator_if_simple_goto(&mut self, bb: BasicBlock) -> Option<Terminator<'tcx>> {
match self.basic_blocks[bb] {
BasicBlockData {
ref statements,
terminator:
ref mut terminator @ Some(Terminator { kind: TerminatorKind::Goto { .. }, .. }),
..
} if statements.is_empty() => terminator.take(),
// if `terminator` is None, this means we are in a loop. In that
// case, let all the loop collapse to its entry.
_ => None,
}
}
/// Collapse a goto chain starting from `start`
fn collapse_goto_chain(&mut self, start: &mut BasicBlock, changed: &mut bool) {
// Using `SmallVec` here, because in some logs on libcore oli-obk saw many single-element
// goto chains. We should probably benchmark different sizes.
let mut terminators: SmallVec<[_; 1]> = Default::default();
let mut current = *start;
while let Some(terminator) = self.take_terminator_if_simple_goto(current) {
let target = match terminator {
Terminator { kind: TerminatorKind::Goto { target }, .. } => target,
_ => unreachable!(),
};
terminators.push((current, terminator));
current = target;
}
let last = current;
*start = last;
while let Some((current, mut terminator)) = terminators.pop() {
let target = match terminator {
Terminator { kind: TerminatorKind::Goto { ref mut target }, .. } => target,
_ => unreachable!(),
};
*changed |= *target != last;
*target = last;
debug!("collapsing goto chain from {:?} to {:?}", current, target);
if self.pred_count[current] == 1 {
// This is the last reference to current, so the pred-count to
// to target is moved into the current block.
self.pred_count[current] = 0;
} else {
self.pred_count[*target] += 1;
self.pred_count[current] -= 1;
}
self.basic_blocks[current].terminator = Some(terminator);
}
}
// merge a block with 1 `goto` predecessor to its parent
fn merge_successor(
&mut self,
merged_blocks: &mut Vec<BasicBlock>,
terminator: &mut Terminator<'tcx>,
) -> bool {
let target = match terminator.kind {
TerminatorKind::Goto { target } if self.pred_count[target] == 1 => target,
_ => return false,
};
debug!("merging block {:?} into {:?}", target, terminator);
*terminator = match self.basic_blocks[target].terminator.take() {
Some(terminator) => terminator,
None => {
// unreachable loop - this should not be possible, as we
// don't strand blocks, but handle it correctly.
return false;
}
};
merged_blocks.push(target);
self.pred_count[target] = 0;
true
}
// turn a branch with all successors identical to a goto
fn simplify_branch(&mut self, terminator: &mut Terminator<'tcx>) -> bool {
match terminator.kind {
TerminatorKind::SwitchInt { .. } => {}
_ => return false,
};
let first_succ = {
if let Some(&first_succ) = terminator.successors().next() {
if terminator.successors().all(|s| *s == first_succ) {
let count = terminator.successors().count();
self.pred_count[first_succ] -= (count - 1) as u32;
first_succ
} else {
return false;
}
} else {
return false;
}
};
debug!("simplifying branch {:?}", terminator);
terminator.kind = TerminatorKind::Goto { target: first_succ };
true
}
fn strip_nops(&mut self) {
for blk in self.basic_blocks.iter_mut() {
blk.statements.retain(|stmt| !matches!(stmt.kind, StatementKind::Nop))
}
}
}
pub fn remove_dead_blocks(tcx: TyCtxt<'tcx>, body: &mut Body<'_>) {
let reachable = traversal::reachable_as_bitset(body);
let num_blocks = body.basic_blocks().len();
if num_blocks == reachable.count() {
return;
}
let basic_blocks = body.basic_blocks_mut();
let mut replacements: Vec<_> = (0..num_blocks).map(BasicBlock::new).collect();
let mut used_blocks = 0;
for alive_index in reachable.iter() {
let alive_index = alive_index.index();
replacements[alive_index] = BasicBlock::new(used_blocks);
if alive_index != used_blocks {
// Swap the next alive block data with the current available slot. Since
// alive_index is non-decreasing this is a valid operation.
basic_blocks.raw.swap(alive_index, used_blocks);
}
used_blocks += 1;
}
if tcx.sess.instrument_coverage() {
save_unreachable_coverage(basic_blocks, used_blocks);
}
basic_blocks.raw.truncate(used_blocks);
for block in basic_blocks {
for target in block.terminator_mut().successors_mut() {
*target = replacements[target.index()];
}
}
}
/// Some MIR transforms can determine at compile time that a sequences of
/// statements will never be executed, so they can be dropped from the MIR.
/// For example, an `if` or `else` block that is guaranteed to never be executed
/// because its condition can be evaluated at compile time, such as by const
/// evaluation: `if false { ... }`.
///
/// Those statements are bypassed by redirecting paths in the CFG around the
/// `dead blocks`; but with `-Z instrument-coverage`, the dead blocks usually
/// include `Coverage` statements representing the Rust source code regions to
/// be counted at runtime. Without these `Coverage` statements, the regions are
/// lost, and the Rust source code will show no coverage information.
///
/// What we want to show in a coverage report is the dead code with coverage
/// counts of `0`. To do this, we need to save the code regions, by injecting
/// `Unreachable` coverage statements. These are non-executable statements whose
/// code regions are still recorded in the coverage map, representing regions
/// with `0` executions.
fn save_unreachable_coverage(
basic_blocks: &mut IndexVec<BasicBlock, BasicBlockData<'_>>,
first_dead_block: usize,
) {
let has_live_counters = basic_blocks.raw[0..first_dead_block].iter().any(|live_block| {
live_block.statements.iter().any(|statement| {
if let StatementKind::Coverage(coverage) = &statement.kind {
matches!(coverage.kind, CoverageKind::Counter { .. })
} else {
false
}
})
});
if !has_live_counters {
// If there are no live `Counter` `Coverage` statements anymore, don't
// move dead coverage to the `START_BLOCK`. Just allow the dead
// `Coverage` statements to be dropped with the dead blocks.
//
// The `generator::StateTransform` MIR pass can create atypical
// conditions, where all live `Counter`s are dropped from the MIR.
//
// At least one Counter per function is required by LLVM (and necessary,
// to add the `function_hash` to the counter's call to the LLVM
// intrinsic `instrprof.increment()`).
return;
}
// Retain coverage info for dead blocks, so coverage reports will still
// report `0` executions for the uncovered code regions.
let mut dropped_coverage = Vec::new();
for dead_block in basic_blocks.raw[first_dead_block..].iter() {
for statement in dead_block.statements.iter() {
if let StatementKind::Coverage(coverage) = &statement.kind {
if let Some(code_region) = &coverage.code_region {
dropped_coverage.push((statement.source_info, code_region.clone()));
}
}
}
}
let start_block = &mut basic_blocks[START_BLOCK];
for (source_info, code_region) in dropped_coverage {
start_block.statements.push(Statement {
source_info,
kind: StatementKind::Coverage(Box::new(Coverage {
kind: CoverageKind::Unreachable,
code_region: Some(code_region),
})),
})
}
}
pub struct SimplifyLocals;
impl<'tcx> MirPass<'tcx> for SimplifyLocals {
fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
trace!("running SimplifyLocals on {:?}", body.source);
simplify_locals(body, tcx);
}
}
pub fn simplify_locals<'tcx>(body: &mut Body<'tcx>, tcx: TyCtxt<'tcx>) {
// First, we're going to get a count of *actual* uses for every `Local`.
let mut used_locals = UsedLocals::new(body);
// Next, we're going to remove any `Local` with zero actual uses. When we remove those
// `Locals`, we're also going to subtract any uses of other `Locals` from the `used_locals`
// count. For example, if we removed `_2 = discriminant(_1)`, then we'll subtract one from
// `use_counts[_1]`. That in turn might make `_1` unused, so we loop until we hit a
// fixedpoint where there are no more unused locals.
remove_unused_definitions(&mut used_locals, body);
// Finally, we'll actually do the work of shrinking `body.local_decls` and remapping the `Local`s.
let map = make_local_map(&mut body.local_decls, &used_locals);
// Only bother running the `LocalUpdater` if we actually found locals to remove.
if map.iter().any(Option::is_none) {
// Update references to all vars and tmps now
let mut updater = LocalUpdater { map, tcx };
updater.visit_body(body);
body.local_decls.shrink_to_fit();
}
}
/// Construct the mapping while swapping out unused stuff out from the `vec`.
fn make_local_map<V>(
local_decls: &mut IndexVec<Local, V>,
used_locals: &UsedLocals,
) -> IndexVec<Local, Option<Local>> {
let mut map: IndexVec<Local, Option<Local>> = IndexVec::from_elem(None, &*local_decls);
let mut used = Local::new(0);
for alive_index in local_decls.indices() {
// `is_used` treats the `RETURN_PLACE` and arguments as used.
if !used_locals.is_used(alive_index) {
continue;
}
map[alive_index] = Some(used);
if alive_index != used {
local_decls.swap(alive_index, used);
}
used.increment_by(1);
}
local_decls.truncate(used.index());
map
}
/// Keeps track of used & unused locals.
struct UsedLocals {
increment: bool,
arg_count: u32,
use_count: IndexVec<Local, u32>,
}
impl UsedLocals {
/// Determines which locals are used & unused in the given body.
fn new(body: &Body<'_>) -> Self {
let mut this = Self {
increment: true,
arg_count: body.arg_count.try_into().unwrap(),
use_count: IndexVec::from_elem(0, &body.local_decls),
};
this.visit_body(body);
this
}
/// Checks if local is used.
///
/// Return place and arguments are always considered used.
fn is_used(&self, local: Local) -> bool {
trace!("is_used({:?}): use_count: {:?}", local, self.use_count[local]);
local.as_u32() <= self.arg_count || self.use_count[local] != 0
}
/// Updates the use counts to reflect the removal of given statement.
fn statement_removed(&mut self, statement: &Statement<'tcx>) {
self.increment = false;
// The location of the statement is irrelevant.
let location = Location { block: START_BLOCK, statement_index: 0 };
self.visit_statement(statement, location);
}
/// Visits a left-hand side of an assignment.
fn visit_lhs(&mut self, place: &Place<'tcx>, location: Location) {
if place.is_indirect() {
// A use, not a definition.
self.visit_place(place, PlaceContext::MutatingUse(MutatingUseContext::Store), location);
} else {
// A definition. The base local itself is not visited, so this occurrence is not counted
// toward its use count. There might be other locals still, used in an indexing
// projection.
self.super_projection(
place.as_ref(),
PlaceContext::MutatingUse(MutatingUseContext::Projection),
location,
);
}
}
}
impl Visitor<'_> for UsedLocals {
fn visit_statement(&mut self, statement: &Statement<'tcx>, location: Location) {
match statement.kind {
StatementKind::LlvmInlineAsm(..)
| StatementKind::CopyNonOverlapping(..)
| StatementKind::Retag(..)
| StatementKind::Coverage(..)
| StatementKind::FakeRead(..)
| StatementKind::AscribeUserType(..) => {
self.super_statement(statement, location);
}
StatementKind::Nop => {}
StatementKind::StorageLive(_local) | StatementKind::StorageDead(_local) => {}
StatementKind::Assign(box (ref place, ref rvalue)) => {
self.visit_lhs(place, location);
self.visit_rvalue(rvalue, location);
}
StatementKind::SetDiscriminant { ref place, variant_index: _ } => {
self.visit_lhs(place, location);
}
}
}
fn visit_local(&mut self, local: &Local, _ctx: PlaceContext, _location: Location) {
if self.increment {
self.use_count[*local] += 1;
} else {
assert_ne!(self.use_count[*local], 0);
self.use_count[*local] -= 1;
}
}
}
/// Removes unused definitions. Updates the used locals to reflect the changes made.
fn remove_unused_definitions<'a, 'tcx>(used_locals: &'a mut UsedLocals, body: &mut Body<'tcx>) {
// The use counts are updated as we remove the statements. A local might become unused
// during the retain operation, leading to a temporary inconsistency (storage statements or
// definitions referencing the local might remain). For correctness it is crucial that this
// computation reaches a fixed point.
let mut modified = true;
while modified {
modified = false;
for data in body.basic_blocks_mut() {
// Remove unnecessary StorageLive and StorageDead annotations.
data.statements.retain(|statement| {
let keep = match &statement.kind {
StatementKind::StorageLive(local) | StatementKind::StorageDead(local) => {
used_locals.is_used(*local)
}
StatementKind::Assign(box (place, _)) => used_locals.is_used(place.local),
StatementKind::SetDiscriminant { ref place, .. } => {
used_locals.is_used(place.local)
}
_ => true,
};
if !keep {
trace!("removing statement {:?}", statement);
modified = true;
used_locals.statement_removed(statement);
}
keep
});
}
}
}
struct LocalUpdater<'tcx> {
map: IndexVec<Local, Option<Local>>,
tcx: TyCtxt<'tcx>,
}
impl<'tcx> MutVisitor<'tcx> for LocalUpdater<'tcx> {
fn tcx(&self) -> TyCtxt<'tcx> {
self.tcx
}
fn visit_local(&mut self, l: &mut Local, _: PlaceContext, _: Location) {
*l = self.map[*l].unwrap();
}
}