+//! A constant propagation optimization pass based on dataflow analysis.
+//!
+//! Currently, this pass only propagates scalar values.
+
use rustc_const_eval::interpret::{ConstValue, ImmTy, Immediate, InterpCx, Scalar};
use rustc_data_structures::fx::FxHashMap;
use rustc_middle::mir::visit::{MutVisitor, Visitor};
use rustc_middle::mir::*;
use rustc_middle::ty::{self, Ty, TyCtxt};
-use rustc_mir_dataflow::value_analysis::{
- Map, ProjElem, State, ValueAnalysis, ValueOrPlace, ValueOrPlaceOrRef,
-};
+use rustc_mir_dataflow::value_analysis::{Map, State, TrackElem, ValueAnalysis, ValueOrPlace};
use rustc_mir_dataflow::{lattice::FlatSet, Analysis, ResultsVisitor, SwitchIntEdgeEffects};
use rustc_span::DUMMY_SP;
use crate::MirPass;
+// These constants are somewhat random guesses and have not been optimized.
+// If `tcx.sess.mir_opt_level() >= 4`, we ignore the limits (this can become very expensive).
+const BLOCK_LIMIT: usize = 100;
+const PLACE_LIMIT: usize = 100;
+
pub struct DataflowConstProp;
impl<'tcx> MirPass<'tcx> for DataflowConstProp {
fn is_enabled(&self, sess: &rustc_session::Session) -> bool {
- // Choose different minimum level?
- sess.mir_opt_level() >= 4
+ sess.mir_opt_level() >= 3
}
+ #[instrument(skip_all level = "debug")]
fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
+ if tcx.sess.mir_opt_level() < 4 && body.basic_blocks.len() > BLOCK_LIMIT {
+ debug!("aborted dataflow const prop due too many basic blocks");
+ return;
+ }
+
// Decide which places to track during the analysis.
- let mut map = Map::new();
- map.register_with_filter(tcx, body, 3, |ty| ty.is_scalar() && !ty.is_unsafe_ptr());
+ let map = Map::from_filter(tcx, body, Ty::is_scalar);
+
+ // We want to have a somewhat linear runtime w.r.t. the number of statements/terminators.
+ // Let's call this number `n`. Dataflow analysis has `O(h*n)` transfer function
+ // applications, where `h` is the height of the lattice. Because the height of our lattice
+ // is linear w.r.t. the number of tracked places, this is `O(tracked_places * n)`. However,
+ // because every transfer function application could traverse the whole map, this becomes
+ // `O(num_nodes * tracked_places * n)` in terms of time complexity. Since the number of
+ // map nodes is strongly correlated to the number of tracked places, this becomes more or
+ // less `O(n)` if we place a constant limit on the number of tracked places.
+ if tcx.sess.mir_opt_level() < 4 && map.tracked_places() > PLACE_LIMIT {
+ debug!("aborted dataflow const prop due to too many tracked places");
+ return;
+ }
// Perform the actual dataflow analysis.
let analysis = ConstAnalysis::new(tcx, body, map);
- let results = analysis.wrap().into_engine(tcx, body).iterate_to_fixpoint();
+ let results = debug_span!("analyze")
+ .in_scope(|| analysis.wrap().into_engine(tcx, body).iterate_to_fixpoint());
// Collect results and patch the body afterwards.
let mut visitor = CollectAndPatch::new(tcx, &results.analysis.0.map);
- results.visit_reachable_with(body, &mut visitor);
- visitor.visit_body(body);
+ debug_span!("collect").in_scope(|| results.visit_reachable_with(body, &mut visitor));
+ debug_span!("patch").in_scope(|| visitor.visit_body(body));
}
}
state.flood_idx(target, self.map());
}
- let value_target = target.and_then(|target| {
- self.map().apply_elem(target, ProjElem::Field(0_u32.into()))
- });
- let overflow_target = target.and_then(|target| {
- self.map().apply_elem(target, ProjElem::Field(1_u32.into()))
- });
+ let value_target = target
+ .and_then(|target| self.map().apply(target, TrackElem::Field(0_u32.into())));
+ let overflow_target = target
+ .and_then(|target| self.map().apply(target, TrackElem::Field(1_u32.into())));
if value_target.is_some() || overflow_target.is_some() {
let (val, overflow) = self.binary_op(state, *op, left, right);
if let Some(value_target) = value_target {
- state.assign_idx(value_target, ValueOrPlaceOrRef::Value(val), self.map());
+ state.assign_idx(value_target, ValueOrPlace::Value(val), self.map());
}
if let Some(overflow_target) = overflow_target {
+ let overflow = match overflow {
+ FlatSet::Top => FlatSet::Top,
+ FlatSet::Elem(overflow) => {
+ if overflow {
+ // Overflow cannot be reliably propagated. See: https://github.com/rust-lang/rust/pull/101168#issuecomment-1288091446
+ FlatSet::Top
+ } else {
+ self.wrap_scalar(Scalar::from_bool(false), self.tcx.types.bool)
+ }
+ }
+ FlatSet::Bottom => FlatSet::Bottom,
+ };
state.assign_idx(
overflow_target,
- ValueOrPlaceOrRef::Value(overflow),
+ ValueOrPlace::Value(overflow),
self.map(),
);
}
&self,
rvalue: &Rvalue<'tcx>,
state: &mut State<Self::Value>,
- ) -> ValueOrPlaceOrRef<Self::Value> {
+ ) -> ValueOrPlace<Self::Value> {
match rvalue {
- Rvalue::Cast(CastKind::Misc, operand, ty) => {
- let operand = self.eval_operand(operand, state);
- match operand {
- FlatSet::Elem(operand) => self
- .ecx
- .misc_cast(&operand, *ty)
- .map(|result| ValueOrPlaceOrRef::Value(self.wrap_immediate(result, *ty)))
- .unwrap_or(ValueOrPlaceOrRef::Unknown),
- _ => ValueOrPlaceOrRef::Unknown,
+ Rvalue::Cast(
+ kind @ (CastKind::IntToInt
+ | CastKind::FloatToInt
+ | CastKind::FloatToFloat
+ | CastKind::IntToFloat),
+ operand,
+ ty,
+ ) => match self.eval_operand(operand, state) {
+ FlatSet::Elem(op) => match kind {
+ CastKind::IntToInt | CastKind::IntToFloat => {
+ self.ecx.int_to_int_or_float(&op, *ty)
+ }
+ CastKind::FloatToInt | CastKind::FloatToFloat => {
+ self.ecx.float_to_float_or_int(&op, *ty)
+ }
+ _ => unreachable!(),
}
- }
+ .map(|result| ValueOrPlace::Value(self.wrap_immediate(result, *ty)))
+ .unwrap_or(ValueOrPlace::top()),
+ _ => ValueOrPlace::top(),
+ },
Rvalue::BinaryOp(op, box (left, right)) => {
+ // Overflows must be ignored here.
let (val, _overflow) = self.binary_op(state, *op, left, right);
- // FIXME: Just ignore overflow here?
- ValueOrPlaceOrRef::Value(val)
+ ValueOrPlace::Value(val)
}
Rvalue::UnaryOp(op, operand) => match self.eval_operand(operand, state) {
FlatSet::Elem(value) => self
.ecx
.unary_op(*op, &value)
- .map(|val| ValueOrPlaceOrRef::Value(self.wrap_immty(val)))
- .unwrap_or(ValueOrPlaceOrRef::Value(FlatSet::Top)),
- FlatSet::Bottom => ValueOrPlaceOrRef::Value(FlatSet::Bottom),
- FlatSet::Top => ValueOrPlaceOrRef::Value(FlatSet::Top),
+ .map(|val| ValueOrPlace::Value(self.wrap_immty(val)))
+ .unwrap_or(ValueOrPlace::Value(FlatSet::Top)),
+ FlatSet::Bottom => ValueOrPlace::Value(FlatSet::Bottom),
+ FlatSet::Top => ValueOrPlace::Value(FlatSet::Top),
},
_ => self.super_rvalue(rvalue, state),
}
let value = match self.handle_operand(discr, state) {
ValueOrPlace::Value(value) => value,
ValueOrPlace::Place(place) => state.get_idx(place, self.map()),
- ValueOrPlace::Unknown => FlatSet::Top,
};
let result = match value {
FlatSet::Top => FlatSet::Top,
impl<'tcx> std::fmt::Debug for ScalarTy<'tcx> {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+ // This is used for dataflow visualization, so we return something more concise.
std::fmt::Display::fmt(&ConstantKind::Val(ConstValue::Scalar(self.0), self.1), f)
}
}
impl<'tcx> ConstAnalysis<'tcx> {
pub fn new(tcx: TyCtxt<'tcx>, body: &Body<'tcx>, map: Map) -> Self {
+ let param_env = tcx.param_env(body.source.def_id());
Self {
map,
tcx,
- ecx: InterpCx::new(tcx, DUMMY_SP, ty::ParamEnv::empty(), DummyMachine),
- param_env: tcx.param_env(body.source.def_id()),
+ ecx: InterpCx::new(tcx, DUMMY_SP, param_env, DummyMachine),
+ param_env: param_env,
}
}
op: BinOp,
left: &Operand<'tcx>,
right: &Operand<'tcx>,
- ) -> (FlatSet<ScalarTy<'tcx>>, FlatSet<ScalarTy<'tcx>>) {
+ ) -> (FlatSet<ScalarTy<'tcx>>, FlatSet<bool>) {
let left = self.eval_operand(left, state);
let right = self.eval_operand(right, state);
match (left, right) {
(FlatSet::Elem(left), FlatSet::Elem(right)) => {
match self.ecx.overflowing_binary_op(op, &left, &right) {
- Ok((val, overflow, ty)) => (
- self.wrap_scalar(val, ty),
- self.wrap_scalar(Scalar::from_bool(overflow), self.tcx.types.bool),
- ),
+ Ok((val, overflow, ty)) => (self.wrap_scalar(val, ty), FlatSet::Elem(overflow)),
_ => (FlatSet::Top, FlatSet::Top),
}
}
let value = match self.handle_operand(op, state) {
ValueOrPlace::Value(value) => value,
ValueOrPlace::Place(place) => state.get_idx(place, &self.map),
- ValueOrPlace::Unknown => FlatSet::Top,
};
match value {
FlatSet::Top => FlatSet::Top,
- FlatSet::Elem(ScalarTy(scalar, ty)) => {
- let layout = self
- .tcx
- .layout_of(ty::ParamEnv::empty().and(ty))
- .expect("this should not happen"); // FIXME
- FlatSet::Elem(ImmTy::from_scalar(scalar, layout))
- }
+ FlatSet::Elem(ScalarTy(scalar, ty)) => self
+ .tcx
+ .layout_of(self.param_env.and(ty))
+ .map(|layout| FlatSet::Elem(ImmTy::from_scalar(scalar, layout)))
+ .unwrap_or(FlatSet::Top),
FlatSet::Bottom => FlatSet::Bottom,
}
}
struct CollectAndPatch<'tcx, 'map> {
tcx: TyCtxt<'tcx>,
map: &'map Map,
+
+ /// For a given MIR location, this stores the values of the operands used by that location. In
+ /// particular, this is before the effect, such that the operands of `_1 = _1 + _2` are
+ /// properly captured. (This may become UB soon, but it is currently emitted even by safe code.)
before_effect: FxHashMap<(Location, Place<'tcx>), ScalarTy<'tcx>>,
+
+ /// Stores the assigned values for assignments where the Rvalue is constant.
assignments: FxHashMap<Location, ScalarTy<'tcx>>,
}
location: Location,
) {
match statement.kind {
+ StatementKind::Assign(box (_, Rvalue::Use(Operand::Constant(_)))) => {
+ // Don't overwrite the assignment if it already uses a constant (to keep the span).
+ }
StatementKind::Assign(box (place, _)) => match state.get(place.as_ref(), self.map) {
FlatSet::Top => (),
FlatSet::Elem(value) => {
self.assignments.insert(location, value);
}
FlatSet::Bottom => {
- // This statement is not reachable. Do nothing, it will (hopefully) be removed.
+ // This assignment is either unreachable, or an uninitialized value is assigned.
}
},
_ => (),
FlatSet::Elem(value) => {
self.visitor.before_effect.insert((location, *place), value);
}
- FlatSet::Bottom => {
- // This only happens if this location is unreachable.
- }
+ FlatSet::Bottom => (),
}
}
_ => (),
_left: &rustc_const_eval::interpret::ImmTy<'tcx, Self::Provenance>,
_right: &rustc_const_eval::interpret::ImmTy<'tcx, Self::Provenance>,
) -> interpret::InterpResult<'tcx, (interpret::Scalar<Self::Provenance>, bool, Ty<'tcx>)> {
- unimplemented!()
+ throw_unsup!(Unsupported("".into()))
}
fn expose_ptr(