Auto merge of rust-lang#71956 - ecstatic-morse:remove-requires-storage-analysis, r=tmandry

Clean up logic around live locals in generator analysis

Resolves rust-lang#69902. Requires rust-lang#71893.

I've found it difficult to make changes in the logic around live locals in `generator/transform.rs`. It uses a custom dataflow analysis, `MaybeRequiresStorage`, that AFAICT computes whether a local is either initialized or borrowed. That analysis is using `before` effects, which we should try to avoid if possible because they are harder to reason about than ones only using the unprefixed effects. @pnkfelix has suggested removing "before" effects entirely to simplify the dataflow framework, which I might pursue someday.

This PR replaces `MaybeRequiresStorage` with a combination of the existing `MaybeBorrowedLocals` and a new `MaybeInitializedLocals`. `MaybeInitializedLocals` is just `MaybeInitializedPlaces` with a coarser resolution: it works on whole locals instead of move paths. As a result, I was able to simplify the logic in `compute_storage_conflicts` and `locals_live_across_suspend_points`.

This is not exactly equivalent to the old logic; some generators are now smaller than before. I believe this was because the old logic was too conservative, but I'm not as familiar with the constraints as the original implementers were, so I could be wrong. For example, I don't see a reason the size of the `mixed_sizes` future couldn't be 5K. It went from 7K to 6K in this PR.

r? @jonas-schievink @tmandry

Author: bors

src/librustc_mir/dataflow/impls/borrowed_locals.rs

@@ -99,6 +99,9 @@ impl<K> GenKillAnalysis<'tcx> for MaybeBorrowedLocals<K>
 where
     K: BorrowAnalysisKind<'tcx>,
 {
+    // The generator transform relies on the fact that this analysis does **not** use "before"
+    // effects.
+
     fn statement_effect(
         &self,
         trans: &mut impl GenKill<Self::Idx>,

src/librustc_mir/dataflow/impls/init_locals.rs

@@ -0,0 +1,118 @@
+//! A less precise version of `MaybeInitializedPlaces` whose domain is entire locals.
+//!
+//! A local will be maybe initialized if *any* projections of that local might be initialized.
+
+use crate::dataflow::{self, BottomValue, GenKill};
+
+use rustc_index::bit_set::BitSet;
+use rustc_middle::mir::visit::{PlaceContext, Visitor};
+use rustc_middle::mir::{self, BasicBlock, Local, Location};
+
+pub struct MaybeInitializedLocals;
+
+impl BottomValue for MaybeInitializedLocals {
+    /// bottom = uninit
+    const BOTTOM_VALUE: bool = false;
+}
+
+impl dataflow::AnalysisDomain<'tcx> for MaybeInitializedLocals {
+    type Idx = Local;
+
+    const NAME: &'static str = "maybe_init_locals";
+
+    fn bits_per_block(&self, body: &mir::Body<'tcx>) -> usize {
+        body.local_decls.len()
+    }
+
+    fn initialize_start_block(&self, body: &mir::Body<'tcx>, entry_set: &mut BitSet<Self::Idx>) {
+        // Function arguments are initialized to begin with.
+        for arg in body.args_iter() {
+            entry_set.insert(arg);
+        }
+    }
+}
+
+impl dataflow::GenKillAnalysis<'tcx> for MaybeInitializedLocals {
+    // The generator transform relies on the fact that this analysis does **not** use "before"
+    // effects.
+
+    fn statement_effect(
+        &self,
+        trans: &mut impl GenKill<Self::Idx>,
+        statement: &mir::Statement<'tcx>,
+        loc: Location,
+    ) {
+        TransferFunction { trans }.visit_statement(statement, loc)
+    }
+
+    fn terminator_effect(
+        &self,
+        trans: &mut impl GenKill<Self::Idx>,
+        terminator: &mir::Terminator<'tcx>,
+        loc: Location,
+    ) {
+        TransferFunction { trans }.visit_terminator(terminator, loc)
+    }
+
+    fn call_return_effect(
+        &self,
+        trans: &mut impl GenKill<Self::Idx>,
+        _block: BasicBlock,
+        _func: &mir::Operand<'tcx>,
+        _args: &[mir::Operand<'tcx>],
+        return_place: mir::Place<'tcx>,
+    ) {
+        trans.gen(return_place.local)
+    }
+
+    /// See `Analysis::apply_yield_resume_effect`.
+    fn yield_resume_effect(
+        &self,
+        trans: &mut impl GenKill<Self::Idx>,
+        _resume_block: BasicBlock,
+        resume_place: mir::Place<'tcx>,
+    ) {
+        trans.gen(resume_place.local)
+    }
+}
+
+struct TransferFunction<'a, T> {
+    trans: &'a mut T,
+}
+
+impl<T> Visitor<'tcx> for TransferFunction<'a, T>
+where
+    T: GenKill<Local>,
+{
+    fn visit_local(&mut self, &local: &Local, context: PlaceContext, _: Location) {
+        use rustc_middle::mir::visit::{MutatingUseContext, NonMutatingUseContext, NonUseContext};
+        match context {
+            // These are handled specially in `call_return_effect` and `yield_resume_effect`.
+            PlaceContext::MutatingUse(MutatingUseContext::Call | MutatingUseContext::Yield) => {}
+
+            // Otherwise, when a place is mutated, we must consider it possibly initialized.
+            PlaceContext::MutatingUse(_) => self.trans.gen(local),
+
+            // If the local is moved out of, or if it gets marked `StorageDead`, consider it no
+            // longer initialized.
+            PlaceContext::NonUse(NonUseContext::StorageDead)
+            | PlaceContext::NonMutatingUse(NonMutatingUseContext::Move) => self.trans.kill(local),
+
+            // All other uses do not affect this analysis.
+            PlaceContext::NonUse(
+                NonUseContext::StorageLive
+                | NonUseContext::AscribeUserTy
+                | NonUseContext::VarDebugInfo,
+            )
+            | PlaceContext::NonMutatingUse(
+                NonMutatingUseContext::Inspect
+                | NonMutatingUseContext::Copy
+                | NonMutatingUseContext::SharedBorrow
+                | NonMutatingUseContext::ShallowBorrow
+                | NonMutatingUseContext::UniqueBorrow
+                | NonMutatingUseContext::AddressOf
+                | NonMutatingUseContext::Projection,
+            ) => {}
+        }
+    }
+}

src/librustc_mir/dataflow/impls/mod.rs

@@ -22,13 +22,15 @@ use crate::dataflow::drop_flag_effects;
 
 mod borrowed_locals;
 pub(super) mod borrows;
+mod init_locals;
 mod liveness;
 mod storage_liveness;
 
 pub use self::borrowed_locals::{MaybeBorrowedLocals, MaybeMutBorrowedLocals};
 pub use self::borrows::Borrows;
+pub use self::init_locals::MaybeInitializedLocals;
 pub use self::liveness::MaybeLiveLocals;
-pub use self::storage_liveness::{MaybeRequiresStorage, MaybeStorageLive};
+pub use self::storage_liveness::MaybeStorageLive;
 
 /// `MaybeInitializedPlaces` tracks all places that might be
 /// initialized upon reaching a particular point in the control flow

src/librustc_mir/dataflow/impls/storage_liveness.rs

@@ -1,11 +1,9 @@
 pub use super::*;
 
 use crate::dataflow::BottomValue;
-use crate::dataflow::{self, GenKill, Results, ResultsRefCursor};
+use crate::dataflow::{self, GenKill};
 use crate::util::storage::AlwaysLiveLocals;
-use rustc_middle::mir::visit::{NonMutatingUseContext, PlaceContext, Visitor};
 use rustc_middle::mir::*;
-use std::cell::RefCell;
 
 #[derive(Clone)]
 pub struct MaybeStorageLive {
@@ -78,233 +76,3 @@ impl BottomValue for MaybeStorageLive {
     /// bottom = dead
     const BOTTOM_VALUE: bool = false;
 }
-
-type BorrowedLocalsResults<'a, 'tcx> = ResultsRefCursor<'a, 'a, 'tcx, MaybeBorrowedLocals>;
-
-/// Dataflow analysis that determines whether each local requires storage at a
-/// given location; i.e. whether its storage can go away without being observed.
-pub struct MaybeRequiresStorage<'mir, 'tcx> {
-    body: &'mir Body<'tcx>,
-    borrowed_locals: RefCell<BorrowedLocalsResults<'mir, 'tcx>>,
-}
-
-impl<'mir, 'tcx> MaybeRequiresStorage<'mir, 'tcx> {
-    pub fn new(
-        body: &'mir Body<'tcx>,
-        borrowed_locals: &'mir Results<'tcx, MaybeBorrowedLocals>,
-    ) -> Self {
-        MaybeRequiresStorage {
-            body,
-            borrowed_locals: RefCell::new(ResultsRefCursor::new(&body, borrowed_locals)),
-        }
-    }
-}
-
-impl<'mir, 'tcx> dataflow::AnalysisDomain<'tcx> for MaybeRequiresStorage<'mir, 'tcx> {
-    type Idx = Local;
-
-    const NAME: &'static str = "requires_storage";
-
-    fn bits_per_block(&self, body: &mir::Body<'tcx>) -> usize {
-        body.local_decls.len()
-    }
-
-    fn initialize_start_block(&self, body: &mir::Body<'tcx>, on_entry: &mut BitSet<Self::Idx>) {
-        // The resume argument is live on function entry (we don't care about
-        // the `self` argument)
-        for arg in body.args_iter().skip(1) {
-            on_entry.insert(arg);
-        }
-    }
-}
-
-impl<'mir, 'tcx> dataflow::GenKillAnalysis<'tcx> for MaybeRequiresStorage<'mir, 'tcx> {
-    fn before_statement_effect(
-        &self,
-        trans: &mut impl GenKill<Self::Idx>,
-        stmt: &mir::Statement<'tcx>,
-        loc: Location,
-    ) {
-        // If a place is borrowed in a statement, it needs storage for that statement.
-        self.borrowed_locals.borrow().analysis().statement_effect(trans, stmt, loc);
-
-        match &stmt.kind {
-            StatementKind::StorageDead(l) => trans.kill(*l),
-
-            // If a place is assigned to in a statement, it needs storage for that statement.
-            StatementKind::Assign(box (place, _))
-            | StatementKind::SetDiscriminant { box place, .. } => {
-                trans.gen(place.local);
-            }
-            StatementKind::LlvmInlineAsm(asm) => {
-                for place in &*asm.outputs {
-                    trans.gen(place.local);
-                }
-            }
-
-            // Nothing to do for these. Match exhaustively so this fails to compile when new
-            // variants are added.
-            StatementKind::AscribeUserType(..)
-            | StatementKind::FakeRead(..)
-            | StatementKind::Nop
-            | StatementKind::Retag(..)
-            | StatementKind::StorageLive(..) => {}
-        }
-    }
-
-    fn statement_effect(
-        &self,
-        trans: &mut impl GenKill<Self::Idx>,
-        _: &mir::Statement<'tcx>,
-        loc: Location,
-    ) {
-        // If we move from a place then only stops needing storage *after*
-        // that statement.
-        self.check_for_move(trans, loc);
-    }
-
-    fn before_terminator_effect(
-        &self,
-        trans: &mut impl GenKill<Self::Idx>,
-        terminator: &mir::Terminator<'tcx>,
-        loc: Location,
-    ) {
-        // If a place is borrowed in a terminator, it needs storage for that terminator.
-        self.borrowed_locals.borrow().analysis().terminator_effect(trans, terminator, loc);
-
-        match &terminator.kind {
-            TerminatorKind::Call { destination: Some((place, _)), .. } => {
-                trans.gen(place.local);
-            }
-
-            // Note that we do *not* gen the `resume_arg` of `Yield` terminators. The reason for
-            // that is that a `yield` will return from the function, and `resume_arg` is written
-            // only when the generator is later resumed. Unlike `Call`, this doesn't require the
-            // place to have storage *before* the yield, only after.
-            TerminatorKind::Yield { .. } => {}
-
-            TerminatorKind::InlineAsm { operands, .. } => {
-                for op in operands {
-                    match op {
-                        InlineAsmOperand::Out { place, .. }
-                        | InlineAsmOperand::InOut { out_place: place, .. } => {
-                            if let Some(place) = place {
-                                trans.gen(place.local);
-                            }
-                        }
-                        InlineAsmOperand::In { .. }
-                        | InlineAsmOperand::Const { .. }
-                        | InlineAsmOperand::SymFn { .. }
-                        | InlineAsmOperand::SymStatic { .. } => {}
-                    }
-                }
-            }
-
-            // Nothing to do for these. Match exhaustively so this fails to compile when new
-            // variants are added.
-            TerminatorKind::Call { destination: None, .. }
-            | TerminatorKind::Abort
-            | TerminatorKind::Assert { .. }
-            | TerminatorKind::Drop { .. }
-            | TerminatorKind::DropAndReplace { .. }
-            | TerminatorKind::FalseEdges { .. }
-            | TerminatorKind::FalseUnwind { .. }
-            | TerminatorKind::GeneratorDrop
-            | TerminatorKind::Goto { .. }
-            | TerminatorKind::Resume
-            | TerminatorKind::Return
-            | TerminatorKind::SwitchInt { .. }
-            | TerminatorKind::Unreachable => {}
-        }
-    }
-
-    fn terminator_effect(
-        &self,
-        trans: &mut impl GenKill<Self::Idx>,
-        terminator: &mir::Terminator<'tcx>,
-        loc: Location,
-    ) {
-        match &terminator.kind {
-            // For call terminators the destination requires storage for the call
-            // and after the call returns successfully, but not after a panic.
-            // Since `propagate_call_unwind` doesn't exist, we have to kill the
-            // destination here, and then gen it again in `call_return_effect`.
-            TerminatorKind::Call { destination: Some((place, _)), .. } => {
-                trans.kill(place.local);
-            }
-
-            // Nothing to do for these. Match exhaustively so this fails to compile when new
-            // variants are added.
-            TerminatorKind::Call { destination: None, .. }
-            | TerminatorKind::Yield { .. }
-            | TerminatorKind::Abort
-            | TerminatorKind::Assert { .. }
-            | TerminatorKind::Drop { .. }
-            | TerminatorKind::DropAndReplace { .. }
-            | TerminatorKind::FalseEdges { .. }
-            | TerminatorKind::FalseUnwind { .. }
-            | TerminatorKind::GeneratorDrop
-            | TerminatorKind::Goto { .. }
-            | TerminatorKind::InlineAsm { .. }
-            | TerminatorKind::Resume
-            | TerminatorKind::Return
-            | TerminatorKind::SwitchInt { .. }
-            | TerminatorKind::Unreachable => {}
-        }
-
-        self.check_for_move(trans, loc);
-    }
-
-    fn call_return_effect(
-        &self,
-        trans: &mut impl GenKill<Self::Idx>,
-        _block: BasicBlock,
-        _func: &mir::Operand<'tcx>,
-        _args: &[mir::Operand<'tcx>],
-        return_place: mir::Place<'tcx>,
-    ) {
-        trans.gen(return_place.local);
-    }
-
-    fn yield_resume_effect(
-        &self,
-        trans: &mut impl GenKill<Self::Idx>,
-        _resume_block: BasicBlock,
-        resume_place: mir::Place<'tcx>,
-    ) {
-        trans.gen(resume_place.local);
-    }
-}
-
-impl<'mir, 'tcx> MaybeRequiresStorage<'mir, 'tcx> {
-    /// Kill locals that are fully moved and have not been borrowed.
-    fn check_for_move(&self, trans: &mut impl GenKill<Local>, loc: Location) {
-        let mut visitor = MoveVisitor { trans, borrowed_locals: &self.borrowed_locals };
-        visitor.visit_location(&self.body, loc);
-    }
-}
-
-impl<'mir, 'tcx> BottomValue for MaybeRequiresStorage<'mir, 'tcx> {
-    /// bottom = dead
-    const BOTTOM_VALUE: bool = false;
-}
-
-struct MoveVisitor<'a, 'mir, 'tcx, T> {
-    borrowed_locals: &'a RefCell<BorrowedLocalsResults<'mir, 'tcx>>,
-    trans: &'a mut T,
-}
-
-impl<'a, 'mir, 'tcx, T> Visitor<'tcx> for MoveVisitor<'a, 'mir, 'tcx, T>
-where
-    T: GenKill<Local>,
-{
-    fn visit_local(&mut self, local: &Local, context: PlaceContext, loc: Location) {
-        if PlaceContext::NonMutatingUse(NonMutatingUseContext::Move) == context {
-            let mut borrowed_locals = self.borrowed_locals.borrow_mut();
-            borrowed_locals.seek_before_primary_effect(loc);
-            if !borrowed_locals.contains(*local) {
-                self.trans.kill(*local);
-            }
-        }
-    }
-}