Auto merge of #127024 - cjgillot:jump-prof, r=oli-obk

Avoid cloning jump threading state when possible

The current implementation of jump threading passes most of its time cloning its state. This PR attempts to avoid such clones by special-casing the last predecessor when recursing through a terminator.

This is not optimal, but a first step while I refactor the state data structure to be sparse.

The two other commits are drive-by.
Fixes https://github.com/rust-lang/rust/issues/116721

r? `@oli-obk`
This commit is contained in:
bors 2024-06-30 11:09:53 +00:00
commit 2975a21b5d
3 changed files with 55 additions and 46 deletions

View File

@ -846,9 +846,10 @@ fn register_children<'tcx>(
if let ty::Ref(_, ref_ty, _) | ty::RawPtr(ref_ty, _) = ty.kind()
&& let ty::Slice(..) = ref_ty.kind()
// The user may have written a predicate like `[T]: Sized` in their where clauses,
// which makes slices scalars.
&& self.places[place].value_index.is_none()
{
assert!(self.places[place].value_index.is_none(), "slices are not scalars");
// Prepend new child to the linked list.
let len = self.places.push(PlaceInfo::new(Some(TrackElem::DerefLen)));
self.places[len].next_sibling = self.places[place].first_child;

View File

@ -91,43 +91,8 @@ fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) {
opportunities: Vec::new(),
};
for (bb, bbdata) in body.basic_blocks.iter_enumerated() {
debug!(?bb, term = ?bbdata.terminator());
if bbdata.is_cleanup || loop_headers.contains(bb) {
continue;
}
let Some((discr, targets)) = bbdata.terminator().kind.as_switch() else { continue };
let Some(discr) = discr.place() else { continue };
debug!(?discr, ?bb);
let discr_ty = discr.ty(body, tcx).ty;
let Ok(discr_layout) = finder.ecx.layout_of(discr_ty) else { continue };
let Some(discr) = finder.map.find(discr.as_ref()) else { continue };
debug!(?discr);
let cost = CostChecker::new(tcx, param_env, None, body);
let mut state = State::new(ConditionSet::default(), finder.map);
let conds = if let Some((value, then, else_)) = targets.as_static_if() {
let Some(value) = ScalarInt::try_from_uint(value, discr_layout.size) else {
continue;
};
arena.alloc_from_iter([
Condition { value, polarity: Polarity::Eq, target: then },
Condition { value, polarity: Polarity::Ne, target: else_ },
])
} else {
arena.alloc_from_iter(targets.iter().filter_map(|(value, target)| {
let value = ScalarInt::try_from_uint(value, discr_layout.size)?;
Some(Condition { value, polarity: Polarity::Eq, target })
}))
};
let conds = ConditionSet(conds);
state.insert_value_idx(discr, conds, finder.map);
finder.find_opportunity(bb, state, cost, 0);
for bb in body.basic_blocks.indices() {
finder.start_from_switch(bb);
}
let opportunities = finder.opportunities;
@ -216,6 +181,46 @@ fn is_empty(&self, state: &State<ConditionSet<'a>>) -> bool {
}
/// Recursion entry point to find threading opportunities.
#[instrument(level = "trace", skip(self))]
fn start_from_switch(&mut self, bb: BasicBlock) -> Option<!> {
let bbdata = &self.body[bb];
if bbdata.is_cleanup || self.loop_headers.contains(bb) {
return None;
}
let (discr, targets) = bbdata.terminator().kind.as_switch()?;
let discr = discr.place()?;
debug!(?discr, ?bb);
let discr_ty = discr.ty(self.body, self.tcx).ty;
let discr_layout = self.ecx.layout_of(discr_ty).ok()?;
let discr = self.map.find(discr.as_ref())?;
debug!(?discr);
let cost = CostChecker::new(self.tcx, self.param_env, None, self.body);
let mut state = State::new(ConditionSet::default(), self.map);
let conds = if let Some((value, then, else_)) = targets.as_static_if() {
let value = ScalarInt::try_from_uint(value, discr_layout.size)?;
self.arena.alloc_from_iter([
Condition { value, polarity: Polarity::Eq, target: then },
Condition { value, polarity: Polarity::Ne, target: else_ },
])
} else {
self.arena.alloc_from_iter(targets.iter().filter_map(|(value, target)| {
let value = ScalarInt::try_from_uint(value, discr_layout.size)?;
Some(Condition { value, polarity: Polarity::Eq, target })
}))
};
let conds = ConditionSet(conds);
state.insert_value_idx(discr, conds, self.map);
self.find_opportunity(bb, state, cost, 0);
None
}
/// Recursively walk statements backwards from this bb's terminator to find threading
/// opportunities.
#[instrument(level = "trace", skip(self, cost), ret)]
fn find_opportunity(
&mut self,
@ -270,12 +275,13 @@ fn find_opportunity(
self.process_switch_int(discr, targets, bb, &mut state);
self.find_opportunity(pred, state, cost, depth + 1);
}
_ => self.recurse_through_terminator(pred, &state, &cost, depth),
_ => self.recurse_through_terminator(pred, || state, &cost, depth),
}
} else {
} else if let &[ref predecessors @ .., last_pred] = &predecessors[..] {
for &pred in predecessors {
self.recurse_through_terminator(pred, &state, &cost, depth);
self.recurse_through_terminator(pred, || state.clone(), &cost, depth);
}
self.recurse_through_terminator(last_pred, || state, &cost, depth);
}
let new_tos = &mut self.opportunities[last_non_rec..];
@ -566,11 +572,12 @@ fn process_statement(
None
}
#[instrument(level = "trace", skip(self, cost))]
#[instrument(level = "trace", skip(self, state, cost))]
fn recurse_through_terminator(
&mut self,
bb: BasicBlock,
state: &State<ConditionSet<'a>>,
// Pass a closure that may clone the state, as we don't want to do it each time.
state: impl FnOnce() -> State<ConditionSet<'a>>,
cost: &CostChecker<'_, 'tcx>,
depth: usize,
) {
@ -600,7 +607,7 @@ fn recurse_through_terminator(
};
// We can recurse through this terminator.
let mut state = state.clone();
let mut state = state();
if let Some(place_to_flood) = place_to_flood {
state.flood_with(place_to_flood.as_ref(), self.map, ConditionSet::default());
}

View File

@ -1,5 +1,6 @@
//@ known-bug: #116721
//@ build-pass
//@ compile-flags: -Zmir-opt-level=3 --emit=mir
fn hey<T>(it: &[T])
where
[T]: Clone,