Introduce -Z pack-coroutine-layout flag

compiler/rustc_abi/src/layout.rs

@@ -4,6 +4,7 @@ use std::ops::Deref;
 use std::range::{RangeFrom, RangeInclusive, RangeToInclusive};
 use std::{cmp, iter};
 
+pub use coroutine::PackCoroutineLayout;
 use rustc_hashes::Hash64;
 use rustc_index::Idx;
 use rustc_index::bit_set::BitMatrix;
@@ -241,24 +242,26 @@ impl<Cx: HasDataLayout> LayoutCalculator<Cx> {
     /// fields may be shared between multiple variants (see the [`coroutine`] module for details).
     pub fn coroutine<
         'a,
-        F: Deref<Target = &'a LayoutData<FieldIdx, VariantIdx>> + fmt::Debug + Copy,
         VariantIdx: Idx,
         FieldIdx: Idx,
         LocalIdx: Idx,
+        F: Deref<Target = &'a LayoutData<FieldIdx, VariantIdx>> + fmt::Debug + Copy,
     >(
         &self,
         local_layouts: &IndexSlice<LocalIdx, F>,
-        prefix_layouts: IndexVec<FieldIdx, F>,
+        upvar_layouts: IndexVec<FieldIdx, F>,
         variant_fields: &IndexSlice<VariantIdx, IndexVec<FieldIdx, LocalIdx>>,
         storage_conflicts: &BitMatrix<LocalIdx, LocalIdx>,
+        pack: PackCoroutineLayout,
         tag_to_layout: impl Fn(Scalar) -> F,
     ) -> LayoutCalculatorResult<FieldIdx, VariantIdx, F> {
         coroutine::layout(
             self,
             local_layouts,
-            prefix_layouts,
+            upvar_layouts,
             variant_fields,
             storage_conflicts,
+            pack,
             tag_to_layout,
         )
     }

compiler/rustc_abi/src/layout/coroutine.rs

@@ -30,6 +30,17 @@ use crate::{
     StructKind, TagEncoding, VariantLayout, Variants, WrappingRange,
 };
 
+/// This option controls how coroutine saved locals are packed
+/// into the coroutine state data
+#[derive(Debug, Clone, Copy)]
+pub enum PackCoroutineLayout {
+    /// The classic layout where captures are always promoted to coroutine state prefix
+    Classic,
+    /// Captures are first saved into the `UNRESUMED` state and promoted
+    /// when they are used across more than one suspension
+    CapturesOnly,
+}
+
 /// Overlap eligibility and variant assignment for each CoroutineSavedLocal.
 #[derive(Clone, Debug, PartialEq)]
 enum SavedLocalEligibility<VariantIdx, FieldIdx> {
@@ -74,6 +85,7 @@ fn coroutine_saved_local_eligibility<VariantIdx: Idx, FieldIdx: Idx, LocalIdx: I
             }
         }
     }
+    debug!(?ineligible_locals, "after counting variants containing a saved local");
 
     // Next, check every pair of eligible locals to see if they
     // conflict.
@@ -103,6 +115,7 @@ fn coroutine_saved_local_eligibility<VariantIdx: Idx, FieldIdx: Idx, LocalIdx: I
             trace!("removing local {:?} due to conflict with {:?}", remove, other);
         }
     }
+    debug!(?ineligible_locals, "after checking conflicts");
 
     // Count the number of variants in use. If only one of them, then it is
     // impossible to overlap any locals in our layout. In this case it's
@@ -122,6 +135,7 @@ fn coroutine_saved_local_eligibility<VariantIdx: Idx, FieldIdx: Idx, LocalIdx: I
             }
             ineligible_locals.insert_all();
         }
+        debug!(?ineligible_locals, "after checking used variants");
     }
 
     // Write down the order of our locals that will be promoted to the prefix.
@@ -145,20 +159,23 @@ pub(super) fn layout<
 >(
     calc: &super::LayoutCalculator<impl HasDataLayout>,
     local_layouts: &IndexSlice<LocalIdx, F>,
-    mut prefix_layouts: IndexVec<FieldIdx, F>,
+    upvar_layouts: IndexVec<FieldIdx, F>,
     variant_fields: &IndexSlice<VariantIdx, IndexVec<FieldIdx, LocalIdx>>,
     storage_conflicts: &BitMatrix<LocalIdx, LocalIdx>,
+    pack: PackCoroutineLayout,
     tag_to_layout: impl Fn(Scalar) -> F,
 ) -> super::LayoutCalculatorResult<FieldIdx, VariantIdx, F> {
     use SavedLocalEligibility::*;
 
     let (ineligible_locals, assignments) =
         coroutine_saved_local_eligibility(local_layouts.len(), variant_fields, storage_conflicts);
+    debug!(?ineligible_locals);
 
-    // Build a prefix layout, including "promoting" all ineligible
-    // locals as part of the prefix. We compute the layout of all of
-    // these fields at once to get optimal packing.
-    let tag_index = prefix_layouts.next_index();
+    // Build a prefix layout, consisting of only the state tag and, as per request, upvars
+    let tag_index = match pack {
+        PackCoroutineLayout::CapturesOnly => FieldIdx::new(0),
+        PackCoroutineLayout::Classic => upvar_layouts.next_index(),
+    };
 
     // `variant_fields` already accounts for the reserved variants, so no need to add them.
     let max_discr = (variant_fields.len() - 1) as u128;
@@ -169,18 +186,29 @@ pub(super) fn layout<
     };
 
     let promoted_layouts = ineligible_locals.iter().map(|local| local_layouts[local]);
-    prefix_layouts.push(tag_to_layout(tag));
-    prefix_layouts.extend(promoted_layouts);
+    // FIXME: when we introduce more pack scheme, we need to change the prefix layout here
+    let prefix_layouts: IndexVec<_, _> = match pack {
+        PackCoroutineLayout::Classic => {
+            // Classic scheme packs the states as follows
+            // [ <upvars>.. , <state tag>, <promoted ineligibles>] ++ <variant data>
+            // In addition, UNRESUMED overlaps with the <upvars> part
+            upvar_layouts.into_iter().chain([tag_to_layout(tag)]).chain(promoted_layouts).collect()
+        }
+        PackCoroutineLayout::CapturesOnly => {
+            [tag_to_layout(tag)].into_iter().chain(promoted_layouts).collect()
+        }
+    };
+    debug!(?pack, "prefix_layouts={prefix_layouts:#?}");
     let prefix =
         calc.univariant(&prefix_layouts, &ReprOptions::default(), StructKind::AlwaysSized)?;
 
     let (prefix_size, prefix_align) = (prefix.size, prefix.align);
 
-    // Split the prefix layout into the "outer" fields (upvars and
-    // discriminant) and the "promoted" fields. Promoted fields will
-    // get included in each variant that requested them in
-    // CoroutineLayout.
-    debug!("prefix = {:#?}", prefix);
+    // Split the prefix layout into the discriminant and
+    // the "promoted" fields.
+    // Promoted fields will get included in each variant
+    // that requested them in CoroutineLayout.
+    debug!("prefix={prefix:#?}");
     let (outer_fields, promoted_offsets, promoted_memory_index) = match prefix.fields {
         FieldsShape::Arbitrary { mut offsets, in_memory_order } => {
             // "a" (`0..b_start`) and "b" (`b_start..`) correspond to
@@ -209,6 +237,7 @@ pub(super) fn layout<
         _ => unreachable!(),
     };
 
+    // Here we start to compute layout of each state variant
     let mut size = prefix.size;
     let mut align = prefix.align;
     let variants = variant_fields

compiler/rustc_abi/src/lib.rs

@@ -74,7 +74,7 @@ pub use extern_abi::CVariadicStatus;
 pub use extern_abi::{ExternAbi, all_names};
 pub use layout::{FIRST_VARIANT, FieldIdx, LayoutCalculator, LayoutCalculatorError, VariantIdx};
 #[cfg(feature = "nightly")]
-pub use layout::{Layout, TyAbiInterface, TyAndLayout};
+pub use layout::{Layout, PackCoroutineLayout, TyAbiInterface, TyAndLayout};
 pub use wrapping_range::WrappingRange;
 
 #[derive(Clone, Copy, PartialEq, Eq, Default)]

compiler/rustc_middle/src/mir/pretty.rs

@@ -566,8 +566,9 @@ fn write_coroutine_layout<'tcx>(
     w: &mut dyn io::Write,
     options: PrettyPrintMirOptions,
 ) -> io::Result<()> {
-    let CoroutineLayout { field_tys, variant_fields, variant_source_info, storage_conflicts } =
-        layout;
+    let CoroutineLayout {
+        field_tys, variant_fields, variant_source_info, storage_conflicts, ..
+    } = layout;
 
     writeln!(w, "{INDENT}coroutine layout {{")?;
 

compiler/rustc_middle/src/mir/query.rs

@@ -7,6 +7,7 @@ use rustc_errors::ErrorGuaranteed;
 use rustc_index::IndexVec;
 use rustc_index::bit_set::BitMatrix;
 use rustc_macros::{StableHash, TyDecodable, TyEncodable, TypeFoldable, TypeVisitable};
+use rustc_session::config::PackCoroutineLayout;
 use rustc_span::{Span, Symbol};
 
 use super::{ConstValue, SourceInfo};
@@ -52,6 +53,29 @@ pub struct CoroutineLayout<'tcx> {
     #[type_foldable(identity)]
     #[type_visitable(ignore)]
     pub storage_conflicts: BitMatrix<CoroutineSavedLocal, CoroutineSavedLocal>,
+
+    /// This map `A -> B` allows later MIR passes, error reporters
+    /// and layout calculator to relate saved locals `A` sourced from upvars
+    /// and locals `B` that upvars are moved into.
+    ///
+    /// For instance, an upvar `_1.0` is assigned saved local `_s12`,
+    /// see notation of [`CoroutineSavedLocal`], in the UNRESUMED state and
+    /// further moved into the internal saved local `_s13`.
+    /// This map, therefore, establishes the mapping from `_s12` to `_s13`,
+    /// so that their memory layout within the coroutine should be overlapped.
+    #[type_foldable(identity)]
+    #[type_visitable(ignore)]
+    pub relocated_upvars: IndexVec<CoroutineSavedLocal, Option<CoroutineSavedLocal>>,
+
+    /// Coroutine layout packing
+    #[type_foldable(identity)]
+    #[type_visitable(ignore)]
+    pub pack: PackCoroutineLayout,
+}
+
+impl<'tcx> CoroutineLayout<'tcx> {
+    /// The initial state of a coroutine
+    pub const UNRESUMED: VariantIdx = VariantIdx::ZERO;
 }
 
 impl Debug for CoroutineLayout<'_> {
@@ -77,6 +101,7 @@ impl Debug for CoroutineLayout<'_> {
                 map.finish()
             })
             .field("storage_conflicts", &self.storage_conflicts)
+            .field("relocated_upvars", &self.relocated_upvars.debug_map_view())
             .finish()
     }
 }

compiler/rustc_middle/src/ty/mod.rs

@@ -2021,6 +2021,8 @@ impl<'tcx> TyCtxt<'tcx> {
                     variant_fields,
                     variant_source_info,
                     storage_conflicts: BitMatrix::new(0, 0),
+                    relocated_upvars: IndexVec::new(),
+                    pack: rustc_session::config::PackCoroutineLayout::No,
                 };
                 return Ok(self.arena.alloc(proxy_layout));
             } else {

compiler/rustc_mir_transform/src/coroutine/layout.rs

@@ -40,6 +40,7 @@ use rustc_mir_dataflow::impls::{
     always_storage_live_locals,
 };
 use rustc_mir_dataflow::{Analysis, Results, ResultsCursor, ResultsVisitor, visit_results};
+use rustc_session::config::PackCoroutineLayout;
 use rustc_span::Span;
 use rustc_span::def_id::{DefId, LocalDefId};
 use rustc_trait_selection::error_reporting::InferCtxtErrorExt;
@@ -433,8 +434,14 @@ pub(super) fn compute_layout<'tcx>(
         tys[saved_local].debuginfo_name.get_or_insert(var.name);
     }
 
-    let layout =
-        CoroutineLayout { field_tys: tys, variant_fields, variant_source_info, storage_conflicts };
+    let layout = CoroutineLayout {
+        field_tys: tys,
+        variant_fields,
+        variant_source_info,
+        storage_conflicts,
+        relocated_upvars: IndexVec::new(),
+        pack: PackCoroutineLayout::No,
+    };
     debug!(?remap);
     debug!(?layout);
     debug!(?storage_liveness);

compiler/rustc_session/src/config.rs

@@ -3346,9 +3346,9 @@ pub(crate) mod dep_tracking {
         FunctionReturn, InliningThreshold, InstrumentCoverage, InstrumentMcount,
         InstrumentMcountOpts, InstrumentXRay, LinkerPluginLto, LocationDetail, LtoCli,
         MirStripDebugInfo, NextSolverConfig, Offload, OptLevel, OutFileName, OutputType,
-        OutputTypes, PatchableFunctionEntry, PointerAuthOption, Polonius, ResolveDocLinks,
-        SourceFileHashAlgorithm, SplitDwarfKind, SwitchWithOptPath, SymbolManglingVersion,
-        WasiExecModel,
+        OutputTypes, PackCoroutineLayout, PatchableFunctionEntry, PointerAuthOption, Polonius,
+        ResolveDocLinks, SourceFileHashAlgorithm, SplitDwarfKind, SwitchWithOptPath,
+        SymbolManglingVersion, WasiExecModel,
     };
     use crate::lint;
     use crate::utils::NativeLib;
@@ -3453,6 +3453,7 @@ pub(crate) mod dep_tracking {
         Polonius,
         InliningThreshold,
         FunctionReturn,
+        PackCoroutineLayout,
         Align,
         CodegenRetagOptions,
         RustcVersion,
@@ -3687,6 +3688,19 @@ pub enum FunctionReturn {
     ThunkExtern,
 }
 
+/// Layout optimisation for Coroutines
+#[derive(Clone, Copy, PartialEq, Eq, Hash, StableHash, Debug, Default, Decodable, Encodable)]
+pub enum PackCoroutineLayout {
+    /// Keep coroutine captured variables throughout all states
+    #[default]
+    No,
+
+    /// Allow coroutine captured variables that are used only once
+    /// before the first suspension to be freed up for storage
+    /// in all other suspension states
+    CapturesOnly,
+}
+
 /// Whether extra span comments are included when dumping MIR, via the `-Z mir-include-spans` flag.
 /// By default, only enabled in the NLL MIR dumps, and disabled in all other passes.
 #[derive(Clone, Copy, Default, PartialEq, Debug)]

compiler/rustc_session/src/options.rs

@@ -832,6 +832,7 @@ mod desc {
     pub(crate) const parse_panic_strategy: &str = "either `unwind`, `abort`, or `immediate-abort`";
     pub(crate) const parse_on_broken_pipe: &str = "either `kill`, `error`, or `inherit`";
     pub(crate) const parse_patchable_function_entry: &str = "a comma separated list of (prefix_nops,total_nops,section_name), (prefix_nops,total_nops), or (total_nops). Where prefix_nops <= total_nops where 0 < total_nops <= 255 and prefix_nops <= total_nops";
+    pub(crate) const parse_pack_coroutine_layout: &str = "either `no` or `captures-only`";
     pub(crate) const parse_opt_panic_strategy: &str = parse_panic_strategy;
     pub(crate) const parse_relro_level: &str = "one of: `full`, `partial`, or `off`";
     pub(crate) const parse_sanitizers: &str = "comma separated list of sanitizers: `address`, `cfi`, `dataflow`, `hwaddress`, `kcfi`, `kernel-address`, `kernel-hwaddress`, `leak`, `memory`, `memtag`, `safestack`, `shadow-call-stack`, `thread`, or 'realtime'";
@@ -2085,6 +2086,18 @@ pub mod parse {
         true
     }
 
+    pub(crate) fn parse_pack_coroutine_layout(
+        slot: &mut PackCoroutineLayout,
+        v: Option<&str>,
+    ) -> bool {
+        *slot = match v {
+            Some("no") => PackCoroutineLayout::No,
+            Some("captures-only") => PackCoroutineLayout::CapturesOnly,
+            _ => return false,
+        };
+        true
+    }
+
     pub(crate) fn parse_inlining_threshold(slot: &mut InliningThreshold, v: Option<&str>) -> bool {
         match v {
             Some("always" | "yes") => {
@@ -2716,6 +2729,8 @@ options! {
         "behavior of std::io::ErrorKind::BrokenPipe (SIGPIPE)"),
     osx_rpath_install_name: bool = (false, parse_bool, [TRACKED],
         "pass `-install_name @rpath/...` to the macOS linker (default: no)"),
+    pack_coroutine_layout: PackCoroutineLayout = (PackCoroutineLayout::default(), parse_pack_coroutine_layout, [TRACKED],
+        "set strategy to pack coroutine state layout (default: no)"),
     packed_bundled_libs: bool = (false, parse_bool, [TRACKED],
         "change rlib format to store native libraries as archives"),
     packed_stack: bool = (false, parse_bool, [TRACKED],

compiler/rustc_ty_utils/src/layout.rs

@@ -24,6 +24,7 @@ use rustc_middle::ty::{
     self, AdtDef, CoroutineArgsExt, EarlyBinder, PseudoCanonicalInput, Ty, TyCtxt,
     TypeVisitableExt, Unnormalized,
 };
+use rustc_session::config::PackCoroutineLayout;
 use rustc_session::{DataTypeKind, FieldInfo, FieldKind, SizeKind, VariantInfo};
 use rustc_span::{Symbol, sym};
 use rustc_structures::Limit;
@@ -585,13 +586,19 @@ fn layout_of_uncached<'tcx>(
                 .map(|ty| cx.layout_of(ty))
                 .try_collect::<IndexVec<_, _>>()?;
 
+            let pack = match info.pack {
+                PackCoroutineLayout::No => rustc_abi::PackCoroutineLayout::Classic,
+                PackCoroutineLayout::CapturesOnly => rustc_abi::PackCoroutineLayout::CapturesOnly,
+            };
+
             let layout = cx
                 .calc
                 .coroutine(
                     &local_layouts,
                     prefix_layouts,
                     &info.variant_fields,
                     &info.storage_conflicts,
+                    pack,
                     |tag| TyAndLayout {
                         ty: tag.primitive().to_ty(tcx),
                         layout: tcx.mk_layout(LayoutData::scalar(cx, tag)),