read: cache abbreviations at offset 0

src/read/abbrev.rs

@@ -1,6 +1,7 @@
 //! Functions for parsing DWARF debugging abbreviations.
 
 use alloc::collections::btree_map;
+use alloc::sync::Arc;
 use alloc::vec::Vec;
 use core::convert::TryFrom;
 use core::fmt::{self, Debug};
@@ -10,7 +11,8 @@ use core::ops::Deref;
 use crate::common::{DebugAbbrevOffset, Encoding, SectionId};
 use crate::constants;
 use crate::endianity::Endianity;
-use crate::read::{EndianSlice, Error, Reader, Result, Section, UnitHeader};
+use crate::read::lazy::LazyArc;
+use crate::read::{EndianSlice, Error, Reader, ReaderOffset, Result, Section, UnitHeader};
 
 /// The `DebugAbbrev` struct represents the abbreviations describing
 /// `DebuggingInformationEntry`s' attribute names and forms found in the
@@ -100,6 +102,38 @@ impl<R> From<R> for DebugAbbrev<R> {
     }
 }
 
+/// A cache of previously parsed `Abbreviations`.
+///
+/// Currently this only caches the abbreviations for offset 0,
+/// since this is a common case in which abbreviations are reused.
+/// This strategy may change in future if there is sufficient need.
+#[derive(Debug, Default)]
+pub struct AbbreviationsCache {
+    abbreviations: LazyArc<Abbreviations>,
+}
+
+impl AbbreviationsCache {
+    /// Create an empty abbreviations cache.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Parse the abbreviations at the given offset.
+    ///
+    /// This uses or updates the cache as required.
+    pub fn get<R: Reader>(
+        &self,
+        debug_abbrev: &DebugAbbrev<R>,
+        offset: DebugAbbrevOffset<R::Offset>,
+    ) -> Result<Arc<Abbreviations>> {
+        if offset.0 != R::Offset::from_u8(0) {
+            return debug_abbrev.abbreviations(offset).map(Arc::new);
+        }
+        self.abbreviations
+            .get(|| debug_abbrev.abbreviations(offset))
+    }
+}
+
 /// A set of type abbreviations.
 ///
 /// Construct an `Abbreviations` instance with the
@@ -993,4 +1027,64 @@ pub mod tests {
             .unwrap();
         assert!(abbrevs.get(0).is_none());
     }
+
+    #[test]
+    fn abbreviations_cache() {
+        #[rustfmt::skip]
+        let buf = Section::new()
+            .abbrev(1, constants::DW_TAG_subprogram, constants::DW_CHILDREN_no)
+                .abbrev_attr(constants::DW_AT_name, constants::DW_FORM_string)
+                .abbrev_attr_null()
+            .abbrev_null()
+            .abbrev(1, constants::DW_TAG_compile_unit, constants::DW_CHILDREN_yes)
+                .abbrev_attr(constants::DW_AT_producer, constants::DW_FORM_strp)
+                .abbrev_attr(constants::DW_AT_language, constants::DW_FORM_data2)
+                .abbrev_attr_null()
+            .abbrev_null()
+            .get_contents()
+            .unwrap();
+
+        let abbrev1 = Abbreviation::new(
+            1,
+            constants::DW_TAG_subprogram,
+            constants::DW_CHILDREN_no,
+            vec![AttributeSpecification::new(
+                constants::DW_AT_name,
+                constants::DW_FORM_string,
+                None,
+            )]
+            .into(),
+        );
+
+        let abbrev2 = Abbreviation::new(
+            1,
+            constants::DW_TAG_compile_unit,
+            constants::DW_CHILDREN_yes,
+            vec![
+                AttributeSpecification::new(
+                    constants::DW_AT_producer,
+                    constants::DW_FORM_strp,
+                    None,
+                ),
+                AttributeSpecification::new(
+                    constants::DW_AT_language,
+                    constants::DW_FORM_data2,
+                    None,
+                ),
+            ]
+            .into(),
+        );
+
+        let debug_abbrev = DebugAbbrev::new(&buf, LittleEndian);
+        let cache = AbbreviationsCache::new();
+        let abbrevs1 = cache.get(&debug_abbrev, DebugAbbrevOffset(0)).unwrap();
+        assert_eq!(abbrevs1.get(1), Some(&abbrev1));
+        let abbrevs2 = cache.get(&debug_abbrev, DebugAbbrevOffset(8)).unwrap();
+        assert_eq!(abbrevs2.get(1), Some(&abbrev2));
+        let abbrevs3 = cache.get(&debug_abbrev, DebugAbbrevOffset(0)).unwrap();
+        assert_eq!(abbrevs3.get(1), Some(&abbrev1));
+
+        assert!(!Arc::ptr_eq(&abbrevs1, &abbrevs2));
+        assert!(Arc::ptr_eq(&abbrevs1, &abbrevs3));
+    }
 }

src/read/dwarf.rs

@@ -9,11 +9,11 @@ use crate::common::{
 };
 use crate::constants;
 use crate::read::{
-    Abbreviations, AttributeValue, DebugAbbrev, DebugAddr, DebugAranges, DebugCuIndex, DebugInfo,
-    DebugInfoUnitHeadersIter, DebugLine, DebugLineStr, DebugLoc, DebugLocLists, DebugRngLists,
-    DebugStr, DebugStrOffsets, DebugTuIndex, DebugTypes, DebugTypesUnitHeadersIter,
-    DebuggingInformationEntry, EntriesCursor, EntriesRaw, EntriesTree, Error,
-    IncompleteLineProgram, LocListIter, LocationLists, Range, RangeLists, RawLocListIter,
+    Abbreviations, AbbreviationsCache, AttributeValue, DebugAbbrev, DebugAddr, DebugAranges,
+    DebugCuIndex, DebugInfo, DebugInfoUnitHeadersIter, DebugLine, DebugLineStr, DebugLoc,
+    DebugLocLists, DebugRngLists, DebugStr, DebugStrOffsets, DebugTuIndex, DebugTypes,
+    DebugTypesUnitHeadersIter, DebuggingInformationEntry, EntriesCursor, EntriesRaw, EntriesTree,
+    Error, IncompleteLineProgram, LocListIter, LocationLists, Range, RangeLists, RawLocListIter,
     RawRngListIter, Reader, ReaderOffset, ReaderOffsetId, Result, RngListIter, Section, UnitHeader,
     UnitIndex, UnitIndexSectionIterator, UnitOffset, UnitType,
 };
@@ -59,6 +59,9 @@ pub struct Dwarf<R> {
 
     /// The DWARF sections for a supplementary object file.
     pub sup: Option<Arc<Dwarf<R>>>,
+
+    /// A cache of previously parsed abbreviations for units in this file.
+    pub abbreviations_cache: AbbreviationsCache,
 }
 
 impl<T> Dwarf<T> {
@@ -96,6 +99,7 @@ impl<T> Dwarf<T> {
             ranges: RangeLists::new(debug_ranges, debug_rnglists),
             file_type: DwarfFileType::Main,
             sup: None,
+            abbreviations_cache: AbbreviationsCache::new(),
         })
     }
 
@@ -157,6 +161,7 @@ impl<T> Dwarf<T> {
             ranges: self.ranges.borrow(&mut borrow),
             file_type: self.file_type,
             sup: self.sup().map(|sup| Arc::new(sup.borrow(borrow))),
+            abbreviations_cache: AbbreviationsCache::new(),
         }
     }
 
@@ -192,10 +197,10 @@ impl<R: Reader> Dwarf<R> {
     }
 
     /// Parse the abbreviations for a compilation unit.
-    // TODO: provide caching of abbreviations
     #[inline]
-    pub fn abbreviations(&self, unit: &UnitHeader<R>) -> Result<Abbreviations> {
-        unit.abbreviations(&self.debug_abbrev)
+    pub fn abbreviations(&self, unit: &UnitHeader<R>) -> Result<Arc<Abbreviations>> {
+        self.abbreviations_cache
+            .get(&self.debug_abbrev, unit.debug_abbrev_offset())
     }
 
     /// Return the string offset at the given index.
@@ -783,6 +788,7 @@ impl<R: Reader> DwarfPackage<R> {
             ranges: RangeLists::new(debug_ranges, debug_rnglists),
             file_type: DwarfFileType::Dwo,
             sup: None,
+            abbreviations_cache: AbbreviationsCache::new(),
         })
     }
 }
@@ -799,7 +805,7 @@ where
     pub header: UnitHeader<R, Offset>,
 
     /// The parsed abbreviations for the unit.
-    pub abbreviations: Abbreviations,
+    pub abbreviations: Arc<Abbreviations>,
 
     /// The `DW_AT_name` attribute of the unit.
     pub name: Option<R>,
@@ -833,7 +839,7 @@ impl<R: Reader> Unit<R> {
     /// Construct a new `Unit` from the given unit header.
     #[inline]
     pub fn new(dwarf: &Dwarf<R>, header: UnitHeader<R>) -> Result<Self> {
-        let abbreviations = header.abbreviations(&dwarf.debug_abbrev)?;
+        let abbreviations = dwarf.abbreviations(&header)?;
         let mut unit = Unit {
             abbreviations,
             name: None,

src/read/lazy.rs

@@ -0,0 +1,115 @@
+pub(crate) use imp::*;
+
+#[cfg(not(feature = "std"))]
+mod imp {
+    use alloc::sync::Arc;
+    use core::sync::atomic::{AtomicPtr, Ordering};
+    use core::{mem, ptr};
+
+    #[derive(Debug, Default)]
+    pub(crate) struct LazyArc<T> {
+        // Only written once with a value obtained from `Arc<T>::into_raw`.
+        // This holds a ref count for the `Arc`, so it is always safe to
+        // clone the `Arc` given a reference to the `LazyArc`.
+        value: AtomicPtr<T>,
+    }
+
+    impl<T> Drop for LazyArc<T> {
+        fn drop(&mut self) {
+            let value_ptr = self.value.load(Ordering::Acquire);
+            if !value_ptr.is_null() {
+                // SAFETY: all writes to `self.value` are pointers obtained from `Arc::into_raw`.
+                drop(unsafe { Arc::from_raw(value_ptr) });
+            }
+        }
+    }
+
+    impl<T> LazyArc<T> {
+        pub(crate) fn get<E, F: FnOnce() -> Result<T, E>>(&self, f: F) -> Result<Arc<T>, E> {
+            // Clone an `Arc` given a pointer obtained from `Arc::into_raw`.
+            // SAFETY: `value_ptr` must be a valid pointer obtained from `Arc<T>::into_raw`.
+            unsafe fn clone_arc_ptr<T>(value_ptr: *const T) -> Arc<T> {
+                let value = Arc::from_raw(value_ptr);
+                let clone = Arc::clone(&value);
+                mem::forget(value);
+                clone
+            }
+
+            // Return the existing value if already computed.
+            // `Ordering::Acquire` is needed so that the content of the loaded `Arc` is
+            // visible to this thread.
+            let value_ptr = self.value.load(Ordering::Acquire);
+            if !value_ptr.is_null() {
+                // SAFETY: all writes to `self.value` are pointers obtained from `Arc::into_raw`.
+                return Ok(unsafe { clone_arc_ptr(value_ptr) });
+            }
+
+            // Race to compute and set the value.
+            let value = f().map(Arc::new)?;
+            let value_ptr = Arc::into_raw(value);
+            match self.value.compare_exchange(
+                ptr::null_mut(),
+                value_ptr as *mut T,
+                // Success: `Ordering::Release` is needed so that the content of the stored `Arc`
+                // is visible to other threads. No ordering is required for the null ptr that is
+                // loaded.
+                Ordering::Release,
+                // Failure: `Ordering::Acquire` is needed so that the content of the loaded `Arc`
+                // is visible to this thread.
+                Ordering::Acquire,
+            ) {
+                Ok(_) => {
+                    // Return the value we computed.
+                    // SAFETY: `value_ptr` was obtained from `Arc::into_raw`.
+                    Ok(unsafe { clone_arc_ptr(value_ptr) })
+                }
+                Err(existing_value_ptr) => {
+                    // We lost the race, drop unneeded `value_ptr`.
+                    // SAFETY: `value_ptr` was obtained from `Arc::into_raw`.
+                    drop(unsafe { Arc::from_raw(value_ptr) });
+                    // Return the existing value.
+                    // SAFETY: all writes to `self.value` are pointers obtained from `Arc::into_raw`.
+                    Ok(unsafe { clone_arc_ptr(existing_value_ptr) })
+                }
+            }
+        }
+    }
+}
+
+#[cfg(feature = "std")]
+mod imp {
+    use std::sync::{Arc, Mutex};
+
+    #[derive(Debug, Default)]
+    pub(crate) struct LazyArc<T> {
+        value: Mutex<Option<Arc<T>>>,
+    }
+
+    impl<T> LazyArc<T> {
+        pub(crate) fn get<E, F: FnOnce() -> Result<T, E>>(&self, f: F) -> Result<Arc<T>, E> {
+            let mut lock = self.value.lock().unwrap();
+            if let Some(value) = &*lock {
+                return Ok(value.clone());
+            }
+            let value = f().map(Arc::new)?;
+            *lock = Some(value.clone());
+            Ok(value)
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn lazy_arc() {
+        let lazy = LazyArc::default();
+        let value = lazy.get(|| Err(()));
+        assert_eq!(value, Err(()));
+        let value = lazy.get(|| Ok::<i32, ()>(3)).unwrap();
+        assert_eq!(*value, 3);
+        let value = lazy.get(|| Err(())).unwrap();
+        assert_eq!(*value, 3);
+    }
+}

src/read/mod.rs

@@ -213,6 +213,9 @@ pub use self::aranges::*;
 mod index;
 pub use self::index::*;
 
+#[cfg(feature = "read")]
+mod lazy;
+
 #[cfg(feature = "read")]
 mod line;
 #[cfg(feature = "read")]

src/write/loc.rs

@@ -436,6 +436,7 @@ mod tests {
     };
     use crate::LittleEndian;
     use std::collections::HashMap;
+    use std::sync::Arc;
 
     #[test]
     fn test_loc_list() {
@@ -508,7 +509,7 @@ mod tests {
                             DebugInfoOffset(0).into(),
                             read::EndianSlice::default(),
                         ),
-                        abbreviations: read::Abbreviations::default(),
+                        abbreviations: Arc::new(read::Abbreviations::default()),
                         name: None,
                         comp_dir: None,
                         low_pc: 0,

src/write/op.rs

@@ -1071,6 +1071,7 @@ mod tests {
     };
     use crate::LittleEndian;
     use std::collections::HashMap;
+    use std::sync::Arc;
 
     #[test]
     fn test_operation() {
@@ -1578,7 +1579,7 @@ mod tests {
                             DebugInfoOffset(0).into(),
                             read::EndianSlice::new(&[], LittleEndian),
                         ),
-                        abbreviations: read::Abbreviations::default(),
+                        abbreviations: Arc::new(read::Abbreviations::default()),
                         name: None,
                         comp_dir: None,
                         low_pc: 0,

src/write/range.rs

@@ -315,6 +315,7 @@ mod tests {
     };
     use crate::LittleEndian;
     use std::collections::HashMap;
+    use std::sync::Arc;
 
     #[test]
     fn test_range() {
@@ -375,7 +376,7 @@ mod tests {
                             DebugInfoOffset(0).into(),
                             read::EndianSlice::default(),
                         ),
-                        abbreviations: read::Abbreviations::default(),
+                        abbreviations: Arc::new(read::Abbreviations::default()),
                         name: None,
                         comp_dir: None,
                         low_pc: 0,