blop

CHANGELOG.md

@@ -1,8 +1,6 @@
 Tantivy 0.13.0
 ======================
 - Bugfix in `FuzzyTermQuery` not matching terms by prefix when it should (@Peachball)
-- Relaxed constraints on the custom/tweak score functions. At the segment level, they can be mut, and they are not required to be Sync + Send.
-- `MMapDirectory::open` does not return a `Result` anymore.
 
 Tantivy 0.12.0
 ======================

Cargo.toml

@@ -18,7 +18,7 @@ byteorder = "1.0"
 crc32fast = "1.2.0"
 once_cell = "1.0"
 regex ={version = "1.3.0", default-features = false, features = ["std"]}
-tantivy-fst = "0.3"
+tantivy-fst = {path="../tantivy-fst", version="0.3"}
 memmap = {version = "0.7", optional=true}
 lz4 = {version="1.20", optional=true}
 snap = "1"
@@ -45,12 +45,11 @@ fnv = "1.0.6"
 owned-read = "0.4"
 failure = "0.1"
 htmlescape = "0.3.1"
-fail = "0.4"
+fail = "0.3"
 murmurhash32 = "0.2"
 chrono = "0.4"
 smallvec = "1.0"
 rayon = "1"
-# ordered-float = "1"
 
 [target.'cfg(windows)'.dependencies]
 winapi = "0.3"
@@ -59,11 +58,9 @@ winapi = "0.3"
 rand = "0.7"
 maplit = "1"
 matches = "0.1.8"
-proptest = "0.9"
-float-cmp = "0.6"
 
 [dev-dependencies.fail]
-version = "0.4"
+version = "0.3"
 features = ["failpoints"]
 
 [profile.release]

README.md

@@ -31,16 +31,12 @@ Tantivy is, in fact, strongly inspired by Lucene's design.
 
 # Benchmark
 
+Tantivy is typically faster than Lucene, but the results depend on 
+the nature of the queries in your workload.
+
 The following [benchmark](https://tantivy-search.github.io/bench/) break downs 
 performance for different type of queries / collection.
 
-
-In general, Tantivy tends to be 
-- slower than Lucene on union with a Top-K due to Block-WAND optimization.
-- faster than Lucene on intersection and phrase queries. 
-
-Your mileage WILL vary depending on the nature of queries and their load.
-
 # Features
 
 - Full-text search

examples/faceted_search_with_tweaked_score.rs

@@ -1,98 +0,0 @@
-use std::collections::HashSet;
-use tantivy::collector::TopDocs;
-use tantivy::doc;
-use tantivy::query::BooleanQuery;
-use tantivy::schema::*;
-use tantivy::{DocId, Index, Score, SegmentReader};
-
-fn main() -> tantivy::Result<()> {
-    let mut schema_builder = Schema::builder();
-
-    let title = schema_builder.add_text_field("title", STORED);
-    let ingredient = schema_builder.add_facet_field("ingredient");
-
-    let schema = schema_builder.build();
-    let index = Index::create_in_ram(schema.clone());
-
-    let mut index_writer = index.writer(30_000_000)?;
-
-    index_writer.add_document(doc!(
-        title => "Fried egg",
-        ingredient => Facet::from("/ingredient/egg"),
-        ingredient => Facet::from("/ingredient/oil"),
-    ));
-    index_writer.add_document(doc!(
-        title => "Scrambled egg",
-        ingredient => Facet::from("/ingredient/egg"),
-        ingredient => Facet::from("/ingredient/butter"),
-        ingredient => Facet::from("/ingredient/milk"),
-        ingredient => Facet::from("/ingredient/salt"),
-    ));
-    index_writer.add_document(doc!(
-        title => "Egg rolls",
-        ingredient => Facet::from("/ingredient/egg"),
-        ingredient => Facet::from("/ingredient/garlic"),
-        ingredient => Facet::from("/ingredient/salt"),
-        ingredient => Facet::from("/ingredient/oil"),
-        ingredient => Facet::from("/ingredient/tortilla-wrap"),
-        ingredient => Facet::from("/ingredient/mushroom"),
-    ));
-    index_writer.commit()?;
-
-    let reader = index.reader()?;
-    let searcher = reader.searcher();
-    {
-        let facets = vec![
-            Facet::from("/ingredient/egg"),
-            Facet::from("/ingredient/oil"),
-            Facet::from("/ingredient/garlic"),
-            Facet::from("/ingredient/mushroom"),
-        ];
-        let query = BooleanQuery::new_multiterms_query(
-            facets
-                .iter()
-                .map(|key| Term::from_facet(ingredient, &key))
-                .collect(),
-        );
-        let top_docs_by_custom_score =
-            TopDocs::with_limit(2).tweak_score(move |segment_reader: &SegmentReader| {
-                let mut ingredient_reader = segment_reader.facet_reader(ingredient).unwrap();
-                let facet_dict = ingredient_reader.facet_dict();
-
-                let query_ords: HashSet<u64> = facets
-                    .iter()
-                    .filter_map(|key| facet_dict.term_ord(key.encoded_str()))
-                    .collect();
-
-                let mut facet_ords_buffer: Vec<u64> = Vec::with_capacity(20);
-
-                move |doc: DocId, original_score: Score| {
-                    ingredient_reader.facet_ords(doc, &mut facet_ords_buffer);
-                    let missing_ingredients = facet_ords_buffer
-                        .iter()
-                        .filter(|ord| !query_ords.contains(ord))
-                        .count();
-                    let tweak = 1.0 / 4_f32.powi(missing_ingredients as i32);
-
-                    original_score * tweak
-                }
-            });
-        let top_docs = searcher.search(&query, &top_docs_by_custom_score)?;
-
-        let titles: Vec<String> = top_docs
-            .iter()
-            .map(|(_, doc_id)| {
-                searcher
-                    .doc(*doc_id)
-                    .unwrap()
-                    .get_first(title)
-                    .unwrap()
-                    .text()
-                    .unwrap()
-                    .to_owned()
-            })
-            .collect();
-        assert_eq!(titles, vec!["Fried egg", "Egg rolls"]);
-    }
-    Ok(())
-}

src/collector/custom_score_top_collector.rs

@@ -28,7 +28,7 @@ where
 /// It is the segment local version of the [`CustomScorer`](./trait.CustomScorer.html).
 pub trait CustomSegmentScorer<TScore>: 'static {
     /// Computes the score of a specific `doc`.
-    fn score(&mut self, doc: DocId) -> TScore;
+    fn score(&self, doc: DocId) -> TScore;
 }
 
 /// `CustomScorer` makes it possible to define any kind of score.
@@ -117,9 +117,9 @@ where
 
 impl<F, TScore> CustomSegmentScorer<TScore> for F
 where
-    F: 'static + FnMut(DocId) -> TScore,
+    F: 'static + Sync + Send + Fn(DocId) -> TScore,
 {
-    fn score(&mut self, doc: DocId) -> TScore {
+    fn score(&self, doc: DocId) -> TScore {
         (self)(doc)
     }
 }

src/collector/mod.rs

@@ -84,7 +84,7 @@ See the `custom_collector` example.
 
 */
 
-use crate::{DocId, Searcher, Executor};
+use crate::DocId;
 use crate::Score;
 use crate::SegmentLocalId;
 use crate::SegmentReader;
@@ -100,9 +100,6 @@ mod top_collector;
 
 mod top_score_collector;
 pub use self::top_score_collector::TopDocs;
-#[cfg(test)]
-pub(crate) use self::top_score_collector::TopScoreSegmentCollector;
-
 
 mod custom_score_top_collector;
 pub use self::custom_score_top_collector::{CustomScorer, CustomSegmentScorer};
@@ -112,9 +109,6 @@ pub use self::tweak_score_top_collector::{ScoreSegmentTweaker, ScoreTweaker};
 
 mod facet_collector;
 pub use self::facet_collector::FacetCollector;
-use crate::fastfield::DeleteBitSet;
-use crate::query::{Scorer, Weight};
-use std::borrow::BorrowMut;
 
 /// `Fruit` is the type for the result of our collection.
 /// e.g. `usize` for the `Count` collector.
@@ -122,8 +116,6 @@ pub trait Fruit: Send + downcast_rs::Downcast {}
 
 impl<T> Fruit for T where T: Send + downcast_rs::Downcast {}
 
-
-
 /// Collectors are in charge of collecting and retaining relevant
 /// information from the document found and scored by the query.
 ///
@@ -162,20 +154,6 @@ pub trait Collector: Sync {
     /// Combines the fruit associated to the collection of each segments
     /// into one fruit.
     fn merge_fruits(&self, segment_fruits: Vec<Self::Fruit>) -> crate::Result<Self::Fruit>;
-
-    fn collect_weight(&self, searcher: &Searcher, weight: &dyn Weight, executor: &Executor) -> crate::Result<Self::Fruit> {
-        let segment_readers = searcher.segment_readers();
-        let fruits = executor.map(
-            |(segment_ord, segment_reader)| {
-                let mut scorer = weight.scorer(segment_reader, 1.0f32)?;
-                let segment_collector =
-                    self.for_segment(segment_ord as u32, segment_reader)?;
-                Ok(segment_collector.collect_scorer(scorer.borrow_mut(), segment_reader.delete_bitset()))
-            },
-            segment_readers.iter().enumerate(),
-        )?;
-        self.merge_fruits(fruits)
-    }
 }
 
 /// The `SegmentCollector` is the trait in charge of defining the
@@ -183,7 +161,7 @@ pub trait Collector: Sync {
 ///
 /// `.collect(doc, score)` will be called for every documents
 /// matching the query.
-pub trait SegmentCollector: 'static + Sized {
+pub trait SegmentCollector: 'static {
     /// `Fruit` is the type for the result of our collection.
     /// e.g. `usize` for the `Count` collector.
     type Fruit: Fruit;
@@ -193,19 +171,6 @@ pub trait SegmentCollector: 'static + Sized {
 
     /// Extract the fruit of the collection from the `SegmentCollector`.
     fn harvest(self) -> Self::Fruit;
-
-    fn collect_scorer(mut self, scorer: &mut dyn Scorer, delete_bitset: Option<&DeleteBitSet>) -> Self::Fruit {
-        if let Some(delete_bitset) = delete_bitset {
-            scorer.for_each(&mut |doc, score| {
-                if delete_bitset.is_alive(doc) {
-                    self.collect(doc, score);
-                }
-            });
-        } else {
-            scorer.for_each(&mut |doc, score| self.collect(doc, score));
-        }
-        self.harvest()
-    }
 }
 
 // -----------------------------------------------

src/collector/top_collector.rs

@@ -56,7 +56,7 @@ impl<T: PartialOrd, D: PartialOrd> PartialEq for ComparableDoc<T, D> {
 impl<T: PartialOrd, D: PartialOrd> Eq for ComparableDoc<T, D> {}
 
 pub(crate) struct TopCollector<T> {
-    pub limit: usize,
+    limit: usize,
     _marker: PhantomData<T>,
 }
 
@@ -69,7 +69,9 @@ where
     /// # Panics
     /// The method panics if limit is 0
     pub fn with_limit(limit: usize) -> TopCollector<T> {
-        assert!(limit > 0, "Limit must be strictly greater than 0.");
+        if limit < 1 {
+            panic!("Limit must be strictly greater than 0.");
+        }
         TopCollector {
             limit,
             _marker: PhantomData,
@@ -122,13 +124,13 @@ where
 /// The theorical complexity for collecting the top `K` out of `n` documents
 /// is `O(n log K)`.
 pub(crate) struct TopSegmentCollector<T> {
-    pub limit: usize,
+    limit: usize,
     heap: BinaryHeap<ComparableDoc<T, DocId>>,
     segment_id: u32,
 }
 
 impl<T: PartialOrd> TopSegmentCollector<T> {
-    pub fn new(segment_id: SegmentLocalId, limit: usize) -> TopSegmentCollector<T> {
+    fn new(segment_id: SegmentLocalId, limit: usize) -> TopSegmentCollector<T> {
         TopSegmentCollector {
             limit,
             heap: BinaryHeap::with_capacity(limit),
@@ -159,10 +161,6 @@ impl<T: PartialOrd + Clone> TopSegmentCollector<T> {
         self.heap.len() >= self.limit
     }
 
-    pub fn pruning_score(&self) -> Option<T> {
-        self.heap.peek().map(|head| head.feature.clone())
-    }
-
     /// Collects a document scored by the given feature
     ///
     /// It collects documents until it has reached the max capacity. Once it reaches capacity, it

src/collector/top_score_collector.rs

@@ -8,13 +8,12 @@ use crate::collector::{
 };
 use crate::fastfield::FastFieldReader;
 use crate::schema::Field;
-use crate::{DocAddress, Executor, Searcher};
+use crate::DocAddress;
 use crate::DocId;
 use crate::Score;
 use crate::SegmentLocalId;
 use crate::SegmentReader;
 use std::fmt;
-use crate::query::{Weight, PruningScorerIfPossible};
 
 /// The `TopDocs` collector keeps track of the top `K` documents
 /// sorted by their score.
@@ -67,7 +66,7 @@ struct ScorerByFastFieldReader {
 }
 
 impl CustomSegmentScorer<u64> for ScorerByFastFieldReader {
-    fn score(&mut self, doc: DocId) -> u64 {
+    fn score(&self, doc: DocId) -> u64 {
         self.ff_reader.get_u64(u64::from(doc))
     }
 }
@@ -418,42 +417,6 @@ impl Collector for TopDocs {
         true
     }
 
-    fn collect_weight(&self, searcher: &Searcher, weight: &dyn Weight, executor: &Executor) -> crate::Result<Self::Fruit> {
-        let segment_readers = searcher.segment_readers();
-        let fruits = executor.map(
-            |(segment_ord, segment_reader)| {
-                match weight.pruning_scorer(segment_reader, 1.0f32)? {
-                    PruningScorerIfPossible::NonPruning(mut scorer) => {
-                        let segment_collector =
-                            self.for_segment(segment_ord as u32, segment_reader)?;
-                        let fruit =
-                            segment_collector.collect_scorer(scorer.as_mut(), segment_reader.delete_bitset());
-                        Ok(fruit)
-                    }
-                    PruningScorerIfPossible::Pruning(mut pruning_scorer) => {
-                        let limit = self.0.limit;
-                        let mut segment_collector =
-                            self.for_segment(segment_ord as u32, segment_reader)?;
-                        for _ in 0..limit {
-                            if !pruning_scorer.advance() {
-                                return Ok(segment_collector.harvest());
-                            }
-                            segment_collector.collect(pruning_scorer.doc(), pruning_scorer.score());
-                        }
-                        let mut pruning_score = segment_collector.0.pruning_score().unwrap_or(0.0f32);
-                        while pruning_scorer.advance_with_pruning(pruning_score) {
-                            segment_collector.0.collect(pruning_scorer.doc(), pruning_scorer.score());
-                            pruning_score = segment_collector.0.pruning_score().unwrap_or(0.0f32);
-                        }
-                        Ok(segment_collector.harvest())
-                    }
-                }
-            },
-            segment_readers.iter().enumerate(),
-        )?;
-        self.merge_fruits(fruits)
-    }
-
     fn merge_fruits(
         &self,
         child_fruits: Vec<Vec<(Score, DocAddress)>>,
@@ -465,12 +428,6 @@ impl Collector for TopDocs {
 /// Segment Collector associated to `TopDocs`.
 pub struct TopScoreSegmentCollector(TopSegmentCollector<Score>);
 
-impl TopScoreSegmentCollector {
-    pub fn new(segment_id: SegmentLocalId, limit: usize) -> Self {
-        TopScoreSegmentCollector(TopSegmentCollector::new(segment_id, limit))
-    }
-}
-
 impl SegmentCollector for TopScoreSegmentCollector {
     type Fruit = Vec<(Score, DocAddress)>;
 

src/collector/tweak_score_top_collector.rs

@@ -29,7 +29,7 @@ where
 /// It is the segment local version of the [`ScoreTweaker`](./trait.ScoreTweaker.html).
 pub trait ScoreSegmentTweaker<TScore>: 'static {
     /// Tweak the given `score` for the document `doc`.
-    fn score(&mut self, doc: DocId, score: Score) -> TScore;
+    fn score(&self, doc: DocId, score: Score) -> TScore;
 }
 
 /// `ScoreTweaker` makes it possible to tweak the score
@@ -121,9 +121,9 @@ where
 
 impl<F, TScore> ScoreSegmentTweaker<TScore> for F
 where
-    F: 'static + FnMut(DocId, Score) -> TScore,
+    F: 'static + Sync + Send + Fn(DocId, Score) -> TScore,
 {
-    fn score(&mut self, doc: DocId, score: Score) -> TScore {
+    fn score(&self, doc: DocId, score: Score) -> TScore {
         (self)(doc, score)
     }
 }

src/core/index.rs

@@ -21,6 +21,7 @@ use crate::schema::FieldType;
 use crate::schema::Schema;
 use crate::tokenizer::{TextAnalyzer, TokenizerManager};
 use crate::IndexWriter;
+use num_cpus;
 use std::borrow::BorrowMut;
 use std::collections::HashSet;
 use std::fmt;

src/core/index_meta.rs

@@ -3,7 +3,9 @@ use crate::core::SegmentId;
 use crate::schema::Schema;
 use crate::Opstamp;
 use census::{Inventory, TrackedObject};
+use serde;
 use serde::{Deserialize, Serialize};
+use serde_json;
 use std::collections::HashSet;
 use std::fmt;
 use std::path::PathBuf;

src/core/searcher.rs

@@ -4,6 +4,7 @@ use crate::core::Executor;
 use crate::core::InvertedIndexReader;
 use crate::core::SegmentReader;
 use crate::query::Query;
+use crate::query::Scorer;
 use crate::query::Weight;
 use crate::schema::Document;
 use crate::schema::Schema;
@@ -23,9 +24,17 @@ fn collect_segment<C: Collector>(
     segment_reader: &SegmentReader,
 ) -> crate::Result<C::Fruit> {
     let mut scorer = weight.scorer(segment_reader, 1.0f32)?;
-    let segment_collector =
-        collector.for_segment(segment_ord as u32, segment_reader)?;
-    Ok(segment_collector.collect_scorer(&mut scorer, segment_reader.delete_bitset()))
+    let mut segment_collector = collector.for_segment(segment_ord as u32, segment_reader)?;
+    if let Some(delete_bitset) = segment_reader.delete_bitset() {
+        scorer.for_each(&mut |doc, score| {
+            if delete_bitset.is_alive(doc) {
+                segment_collector.collect(doc, score);
+            }
+        });
+    } else {
+        scorer.for_each(&mut |doc, score| segment_collector.collect(doc, score));
+    }
+    Ok(segment_collector.harvest())
 }
 
 /// Holds a list of `SegmentReader`s ready for search.

src/directory/managed_directory.rs

@@ -11,6 +11,7 @@ use crate::error::DataCorruption;
 use crate::Directory;
 
 use crc32fast::Hasher;
+use serde_json;
 use std::collections::HashSet;
 use std::io;
 use std::io::Write;

src/directory/mmap_directory.rs

@@ -1,3 +1,10 @@
+use fs2;
+use notify;
+
+use self::fs2::FileExt;
+use self::notify::RawEvent;
+use self::notify::RecursiveMode;
+use self::notify::Watcher;
 use crate::core::META_FILEPATH;
 use crate::directory::error::LockError;
 use crate::directory::error::{
@@ -13,11 +20,8 @@ use crate::directory::WatchCallback;
 use crate::directory::WatchCallbackList;
 use crate::directory::WatchHandle;
 use crate::directory::{TerminatingWrite, WritePtr};
-use fs2::FileExt;
+use atomicwrites;
 use memmap::Mmap;
-use notify::RawEvent;
-use notify::RecursiveMode;
-use notify::Watcher;
 use serde::{Deserialize, Serialize};
 use std::collections::HashMap;
 use std::convert::From;
@@ -220,13 +224,17 @@ struct MmapDirectoryInner {
 }
 
 impl MmapDirectoryInner {
-    fn new(root_path: PathBuf, temp_directory: Option<TempDir>) -> MmapDirectoryInner {
-        MmapDirectoryInner {
+    fn new(
+        root_path: PathBuf,
+        temp_directory: Option<TempDir>,
+    ) -> Result<MmapDirectoryInner, OpenDirectoryError> {
+        let mmap_directory_inner = MmapDirectoryInner {
             root_path,
             mmap_cache: Default::default(),
             _temp_directory: temp_directory,
             watcher: RwLock::new(None),
-        }
+        };
+        Ok(mmap_directory_inner)
     }
 
     fn watch(&self, watch_callback: WatchCallback) -> crate::Result<WatchHandle> {
@@ -260,11 +268,14 @@ impl fmt::Debug for MmapDirectory {
 }
 
 impl MmapDirectory {
-    fn new(root_path: PathBuf, temp_directory: Option<TempDir>) -> MmapDirectory {
-        let inner = MmapDirectoryInner::new(root_path, temp_directory);
-        MmapDirectory {
+    fn new(
+        root_path: PathBuf,
+        temp_directory: Option<TempDir>,
+    ) -> Result<MmapDirectory, OpenDirectoryError> {
+        let inner = MmapDirectoryInner::new(root_path, temp_directory)?;
+        Ok(MmapDirectory {
             inner: Arc::new(inner),
-        }
+        })
     }
 
     /// Creates a new MmapDirectory in a temporary directory.
@@ -274,7 +285,7 @@ impl MmapDirectory {
     pub fn create_from_tempdir() -> Result<MmapDirectory, OpenDirectoryError> {
         let tempdir = TempDir::new().map_err(OpenDirectoryError::IoError)?;
         let tempdir_path = PathBuf::from(tempdir.path());
-        Ok(MmapDirectory::new(tempdir_path, Some(tempdir)))
+        MmapDirectory::new(tempdir_path, Some(tempdir))
     }
 
     /// Opens a MmapDirectory in a directory.
@@ -292,7 +303,7 @@ impl MmapDirectory {
                 directory_path,
             )))
         } else {
-            Ok(MmapDirectory::new(PathBuf::from(directory_path), None))
+            Ok(MmapDirectory::new(PathBuf::from(directory_path), None)?)
         }
     }
 

src/error.rs

@@ -7,6 +7,7 @@ use crate::directory::error::{Incompatibility, LockError};
 use crate::fastfield::FastFieldNotAvailableError;
 use crate::query;
 use crate::schema;
+use serde_json;
 use std::fmt;
 use std::path::PathBuf;
 use std::sync::PoisonError;

src/indexer/segment_updater.rs

@@ -23,6 +23,7 @@ use futures::channel::oneshot;
 use futures::executor::{ThreadPool, ThreadPoolBuilder};
 use futures::future::Future;
 use futures::future::TryFutureExt;
+use serde_json;
 use std::borrow::BorrowMut;
 use std::collections::HashSet;
 use std::io::Write;

src/lib.rs

@@ -269,6 +269,144 @@ impl DocAddress {
     }
 }
 
+
+
+mod sse2 {
+    use crate::postings::compression::{AlignedBuffer, COMPRESSION_BLOCK_SIZE};
+    use std::arch::x86_64::__m128i as DataType;
+    use std::arch::x86_64::_mm_add_epi32 as op_add;
+    use std::arch::x86_64::_mm_cmplt_epi32 as op_lt;
+    use std::arch::x86_64::_mm_load_si128 as op_load;
+    // requires 128-bits alignment
+    use std::arch::x86_64::_mm_set1_epi32 as set1;
+    use std::arch::x86_64::_mm_setzero_si128 as set0;
+    use std::arch::x86_64::_mm_sub_epi32 as op_sub;
+    use std::arch::x86_64::{_mm_cvtsi128_si32, _mm_shuffle_epi32};
+
+    const MASK1: i32 = 78;
+    const MASK2: i32 = 177;
+
+    /// Performs an exhaustive linear search over the
+    ///
+    /// There is no early exit here. We simply count the
+    /// number of elements that are `< target`.
+    pub unsafe fn linear_search_sse2_128(arr: &AlignedBuffer, target: u32) -> usize {
+        let ptr = arr as *const AlignedBuffer as *const DataType;
+        let vkey = set1(target as i32);
+        let mut cnt = set0();
+        // We work over 4 `__m128i` at a time.
+        // A single `__m128i` actual contains 4 `u32`.
+        for i in 0..(COMPRESSION_BLOCK_SIZE as isize) / (4 * 4) {
+            let cmp1 = op_lt(op_load(ptr.offset(i * 4)), vkey);
+            let cmp2 = op_lt(op_load(ptr.offset(i * 4 + 1)), vkey);
+            let cmp3 = op_lt(op_load(ptr.offset(i * 4 + 2)), vkey);
+            let cmp4 = op_lt(op_load(ptr.offset(i * 4 + 3)), vkey);
+            let sum = op_add(op_add(cmp1, cmp2), op_add(cmp3, cmp4));
+            cnt = op_sub(cnt, sum);
+        }
+        cnt = op_add(cnt, _mm_shuffle_epi32(cnt, MASK1));
+        cnt = op_add(cnt, _mm_shuffle_epi32(cnt, MASK2));
+        _mm_cvtsi128_si32(cnt) as usize
+    }
+}
+
+
+const MAX_DOC: u32 = 2_000_000_000u32;
+
+use sse2::linear_search_sse2_128;
+use crate::postings::BlockSegmentPostings;
+use crate::schema::IndexRecordOption;
+
+struct DocCursor {
+    block_segment_postings: BlockSegmentPostings,
+    cursor: usize,
+}
+
+impl DocCursor {
+    fn new(mut block_segment_postings: BlockSegmentPostings) -> DocCursor {
+        block_segment_postings.advance();
+        DocCursor {
+            block_segment_postings,
+            cursor: 0,
+        }
+    }
+    fn doc(&self,) -> DocId {
+        self.block_segment_postings.doc(self.cursor)
+    }
+
+    fn len(&self,) -> usize {
+        self.block_segment_postings.doc_freq()
+    }
+
+    fn advance(&mut self) {
+        if self.cursor == 127 {
+            self.block_segment_postings.advance();
+            self.cursor = 0;
+        } else {
+            self.cursor += 1;
+        }
+    }
+
+    fn skip_to(&mut self, target: DocId) {
+        if self.block_segment_postings.skip_reader.doc() >= target {
+            unsafe {
+                let mut ptr = self.block_segment_postings.docs_aligned_b().0.get_unchecked(self.cursor) as *const u32;
+                while *ptr < target {
+                    self.cursor += 1;
+                    ptr = ptr.offset(1);
+                }
+            }
+            return;
+        }
+        self.block_segment_postings.skip_to_b(target);
+        let block= self.block_segment_postings.docs_aligned_b();
+        self.cursor = unsafe { linear_search_sse2_128(&block, target) };
+
+    }
+}
+
+pub fn intersection(idx: &InvertedIndexReader, terms: &[Term]) -> crate::Result<u32> {
+    let mut posting_lists: Vec<DocCursor> = Vec::with_capacity(terms.len());
+    for term in terms {
+        if let Some(mut block_postings) = idx.read_block_postings(term, IndexRecordOption::Basic) {
+            posting_lists.push(DocCursor::new(block_postings ));
+        } else {
+            return Ok(0);
+        }
+    }
+    posting_lists.sort_by_key(|posting_list| posting_list.len());
+    Ok({ intersection_idx(&mut posting_lists[..]) })
+}
+
+fn intersection_idx(posting_lists: &mut [DocCursor]) -> DocId {
+    let mut candidate = posting_lists[0].doc();
+    let mut i = 1;
+
+    let mut count = 0u32;
+
+    let num_posting_lists = posting_lists.len();
+    loop {
+        while i < num_posting_lists {
+            posting_lists[i].skip_to(candidate);
+            if posting_lists[i].doc() != candidate {
+                candidate = posting_lists[i].doc();
+                i = 0;
+                continue;
+            }
+            i += 1;
+        }
+        count += 1;
+        if candidate == MAX_DOC {
+            break;
+        }
+        posting_lists[0].advance();
+        candidate = posting_lists[0].doc();
+        i = 1;
+    }
+    count - 1u32
+}
+
+
 /// `DocAddress` contains all the necessary information
 /// to identify a document given a `Searcher` object.
 ///

src/postings/block_max_postings.rs

@@ -1,16 +0,0 @@
-use crate::postings::Postings;
-use crate::DocId;
-
-/// Inverted list with additional information about the maximum term frequency
-/// within a block, as well as globally within the list.
-pub trait BlockMaxPostings: Postings {
-    /// Returns the maximum frequency in the entire list.
-    fn max_term_freq(&self) -> u32;
-    /// Returns the maximum frequency in the current block.
-    fn block_max_term_freq(&mut self) -> u32;
-    /// Returns the document with the largest frequency.
-    fn max_doc(&self) -> DocId;
-    /// Returns the document with the largest frequency within the current
-    /// block.
-    fn block_max_doc(&self) -> DocId;
-}

src/postings/block_max_segment_postings.rs

@@ -1,76 +0,0 @@
-use crate::postings::{BlockMaxPostings, Postings, SegmentPostings};
-use crate::{DocId, DocSet, SkipResult};
-
-/// A wrapper over [`SegmentPostings`](./struct.SegmentPostings.html)
-/// with max block frequencies.
-pub struct BlockMaxSegmentPostings {
-    postings: SegmentPostings,
-    max_blocks: SegmentPostings,
-    doc_with_max_term_freq: DocId,
-    max_term_freq: u32,
-}
-
-impl BlockMaxSegmentPostings {
-    /// Constructs a new segment postings with block-max information.
-    pub fn new(
-        postings: SegmentPostings,
-        max_blocks: SegmentPostings,
-        doc_with_max_term_freq: DocId,
-        max_term_freq: u32,
-    ) -> Self {
-        Self {
-            postings,
-            max_blocks,
-            doc_with_max_term_freq,
-            max_term_freq,
-        }
-    }
-}
-
-impl DocSet for BlockMaxSegmentPostings {
-    fn advance(&mut self) -> bool {
-        self.postings.advance()
-    }
-
-    fn doc(&self) -> DocId {
-        self.postings.doc()
-    }
-
-    fn size_hint(&self) -> u32 {
-        self.postings.size_hint()
-    }
-
-    fn skip_next(&mut self, target: DocId) -> SkipResult {
-        self.postings.skip_next(target)
-    }
-}
-
-impl Postings for BlockMaxSegmentPostings {
-    fn term_freq(&self) -> u32 {
-        self.postings.term_freq()
-    }
-    fn positions_with_offset(&mut self, offset: u32, output: &mut Vec<u32>) {
-        self.postings.positions_with_offset(offset, output);
-    }
-    fn positions(&mut self, output: &mut Vec<u32>) {
-        self.postings.positions(output);
-    }
-}
-
-impl BlockMaxPostings for BlockMaxSegmentPostings {
-    fn max_term_freq(&self) -> u32 {
-        self.max_term_freq
-    }
-    fn block_max_term_freq(&mut self) -> u32 {
-        if let SkipResult::End = self.max_blocks.skip_next(self.doc()) {
-            panic!("Max blocks corrupted: reached end of max block");
-        }
-        self.max_blocks.term_freq()
-    }
-    fn max_doc(&self) -> DocId {
-        self.doc_with_max_term_freq
-    }
-    fn block_max_doc(&self) -> DocId {
-        self.max_blocks.doc()
-    }
-}

src/postings/block_search.rs

@@ -25,7 +25,7 @@ mod sse2 {
     ///
     /// There is no early exit here. We simply count the
     /// number of elements that are `< target`.
-    pub(crate) fn linear_search_sse2_128(arr: &AlignedBuffer, target: u32) -> usize {
+    pub fn linear_search_sse2_128(arr: &AlignedBuffer, target: u32) -> usize {
         unsafe {
             let ptr = arr as *const AlignedBuffer as *const DataType;
             let vkey = set1(target as i32);
@@ -63,6 +63,8 @@ mod sse2 {
     }
 }
 
+pub use self::sse2::linear_search_sse2_128;
+
 /// This `linear search` browser exhaustively through the array.
 /// but the early exit is very difficult to predict.
 ///

src/postings/block_segment_postings.rs

@@ -1,316 +0,0 @@
-use crate::DocId;
-use tantivy_fst::Streamer;
-use crate::postings::{SkipReader, FreqReadingOption, USE_SKIP_INFO_LIMIT};
-use owned_read::OwnedRead;
-use crate::postings::compression::{BlockDecoder, COMPRESSION_BLOCK_SIZE, VIntDecoder, compressed_block_size, AlignedBuffer};
-use crate::schema::IndexRecordOption;
-use crate::common::{VInt, BinarySerializable};
-
-
-fn split_into_skips_and_postings(
-    doc_freq: u32,
-    mut data: OwnedRead,
-) -> (Option<OwnedRead>, OwnedRead) {
-    if doc_freq >= USE_SKIP_INFO_LIMIT {
-        let skip_len = VInt::deserialize(&mut data).expect("Data corrupted").0 as usize;
-        let mut postings_data = data.clone();
-        postings_data.advance(skip_len);
-        data.clip(skip_len);
-        (Some(data), postings_data)
-    } else {
-        (None, data)
-    }
-}
-
-/// `BlockSegmentPostings` is a cursor iterating over blocks
-/// of documents.
-///
-/// # Warning
-///
-/// While it is useful for some very specific high-performance
-/// use cases, you should prefer using `SegmentPostings` for most usage.
-pub struct BlockSegmentPostings {
-    doc_decoder: BlockDecoder,
-    freq_decoder: BlockDecoder,
-    freq_reading_option: FreqReadingOption,
-
-    doc_freq: usize,
-    doc_offset: DocId,
-
-    num_vint_docs: usize,
-
-    remaining_data: OwnedRead,
-    skip_reader: SkipReader,
-}
-
-
-#[derive(Debug, Eq, PartialEq)]
-pub enum BlockSegmentPostingsSkipResult {
-    Terminated,
-    Success(u32), //< number of term freqs to skip
-}
-
-impl BlockSegmentPostings {
-    pub(crate) fn from_data(
-        doc_freq: u32,
-        data: OwnedRead,
-        record_option: IndexRecordOption,
-        requested_option: IndexRecordOption,
-    ) -> BlockSegmentPostings {
-        let freq_reading_option = match (record_option, requested_option) {
-            (IndexRecordOption::Basic, _) => FreqReadingOption::NoFreq,
-            (_, IndexRecordOption::Basic) => FreqReadingOption::SkipFreq,
-            (_, _) => FreqReadingOption::ReadFreq,
-        };
-
-        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, data);
-        let skip_reader = match skip_data_opt {
-            Some(skip_data) => SkipReader::new(skip_data, record_option),
-            None => SkipReader::new(OwnedRead::new(&[][..]), record_option),
-        };
-        let doc_freq = doc_freq as usize;
-        let num_vint_docs = doc_freq % COMPRESSION_BLOCK_SIZE;
-        BlockSegmentPostings {
-            num_vint_docs,
-            doc_decoder: BlockDecoder::new(),
-            freq_decoder: BlockDecoder::with_val(1),
-            freq_reading_option,
-            doc_offset: 0,
-            doc_freq,
-            remaining_data: postings_data,
-            skip_reader,
-        }
-    }
-
-    // Resets the block segment postings on another position
-    // in the postings file.
-    //
-    // This is useful for enumerating through a list of terms,
-    // and consuming the associated posting lists while avoiding
-    // reallocating a `BlockSegmentPostings`.
-    //
-    // # Warning
-    //
-    // This does not reset the positions list.
-    pub(crate) fn reset(&mut self, doc_freq: u32, postings_data: OwnedRead) {
-        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, postings_data);
-        let num_vint_docs = (doc_freq as usize) & (COMPRESSION_BLOCK_SIZE - 1);
-        self.num_vint_docs = num_vint_docs;
-        self.remaining_data = postings_data;
-        if let Some(skip_data) = skip_data_opt {
-            self.skip_reader.reset(skip_data);
-        } else {
-            self.skip_reader.reset(OwnedRead::new(&[][..]))
-        }
-        self.doc_offset = 0;
-        self.doc_freq = doc_freq as usize;
-    }
-
-    /// Returns the document frequency associated to this block postings.
-    ///
-    /// This `doc_freq` is simply the sum of the length of all of the blocks
-    /// length, and it does not take in account deleted documents.
-    pub fn doc_freq(&self) -> usize {
-        self.doc_freq
-    }
-
-    /// Returns the array of docs in the current block.
-    ///
-    /// Before the first call to `.advance()`, the block
-    /// returned by `.docs()` is empty.
-    #[inline]
-    pub fn docs(&self) -> &[DocId] {
-        self.doc_decoder.output_array()
-    }
-
-    pub(crate) fn docs_aligned(&self) -> (&AlignedBuffer, usize) {
-        self.doc_decoder.output_aligned()
-    }
-
-    /// Return the document at index `idx` of the block.
-    #[inline]
-    pub fn doc(&self, idx: usize) -> u32 {
-        self.doc_decoder.output(idx)
-    }
-
-    /// Return the array of `term freq` in the block.
-    #[inline]
-    pub fn freqs(&self) -> &[u32] {
-        self.freq_decoder.output_array()
-    }
-
-    /// Return the frequency at index `idx` of the block.
-    #[inline]
-    pub fn freq(&self, idx: usize) -> u32 {
-        self.freq_decoder.output(idx)
-    }
-
-    /// Returns the length of the current block.
-    ///
-    /// All blocks have a length of `NUM_DOCS_PER_BLOCK`,
-    /// except the last block that may have a length
-    /// of any number between 1 and `NUM_DOCS_PER_BLOCK - 1`
-    #[inline]
-    pub(crate) fn block_len(&self) -> usize {
-        self.doc_decoder.output_len
-    }
-
-    /// position on a block that may contains `doc_id`.
-    /// Always advance the current block.
-    ///
-    /// Returns true if a block that has an element greater or equal to the target is found.
-    /// Returning true does not guarantee that the smallest element of the block is smaller
-    /// than the target. It only guarantees that the last element is greater or equal.
-    ///
-    /// Returns false iff all of the document remaining are smaller than
-    /// `doc_id`. In that case, all of these document are consumed.
-    ///
-    pub fn skip_to(&mut self, target_doc: DocId) -> BlockSegmentPostingsSkipResult {
-        let mut skip_freqs = 0u32;
-        while self.skip_reader.advance() {
-            if self.skip_reader.doc() >= target_doc {
-                // the last document of the current block is larger
-                // than the target.
-                //
-                // We found our block!
-                let num_bits = self.skip_reader.doc_num_bits();
-                let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
-                    self.remaining_data.as_ref(),
-                    self.doc_offset,
-                    num_bits,
-                );
-                self.remaining_data.advance(num_consumed_bytes);
-                let tf_num_bits = self.skip_reader.tf_num_bits();
-                match self.freq_reading_option {
-                    FreqReadingOption::NoFreq => {}
-                    FreqReadingOption::SkipFreq => {
-                        let num_bytes_to_skip = compressed_block_size(tf_num_bits);
-                        self.remaining_data.advance(num_bytes_to_skip);
-                    }
-                    FreqReadingOption::ReadFreq => {
-                        let num_consumed_bytes = self
-                            .freq_decoder
-                            .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
-                        self.remaining_data.advance(num_consumed_bytes);
-                    }
-                }
-                self.doc_offset = self.skip_reader.doc();
-                return BlockSegmentPostingsSkipResult::Success(skip_freqs);
-            } else {
-                skip_freqs += self.skip_reader.tf_sum();
-                let advance_len = self.skip_reader.total_block_len();
-                self.doc_offset = self.skip_reader.doc();
-                self.remaining_data.advance(advance_len);
-            }
-        }
-
-        // we are now on the last, incomplete, variable encoded block.
-        if self.num_vint_docs > 0 {
-            let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
-                self.remaining_data.as_ref(),
-                self.doc_offset,
-                self.num_vint_docs,
-            );
-            self.remaining_data.advance(num_compressed_bytes);
-            match self.freq_reading_option {
-                FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
-                FreqReadingOption::ReadFreq => {
-                    self.freq_decoder
-                        .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
-                }
-            }
-            self.num_vint_docs = 0;
-            return self
-                .docs()
-                .last()
-                .map(|last_doc| {
-                    if *last_doc >= target_doc {
-                        BlockSegmentPostingsSkipResult::Success(skip_freqs)
-                    } else {
-                        BlockSegmentPostingsSkipResult::Terminated
-                    }
-                })
-                .unwrap_or(BlockSegmentPostingsSkipResult::Terminated);
-        }
-        BlockSegmentPostingsSkipResult::Terminated
-    }
-
-    /// Advance to the next block.
-    ///
-    /// Returns false iff there was no remaining blocks.
-    pub fn advance(&mut self) -> bool {
-        if self.skip_reader.advance() {
-            let num_bits = self.skip_reader.doc_num_bits();
-            let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
-                self.remaining_data.as_ref(),
-                self.doc_offset,
-                num_bits,
-            );
-            self.remaining_data.advance(num_consumed_bytes);
-            let tf_num_bits = self.skip_reader.tf_num_bits();
-            match self.freq_reading_option {
-                FreqReadingOption::NoFreq => {}
-                FreqReadingOption::SkipFreq => {
-                    let num_bytes_to_skip = compressed_block_size(tf_num_bits);
-                    self.remaining_data.advance(num_bytes_to_skip);
-                }
-                FreqReadingOption::ReadFreq => {
-                    let num_consumed_bytes = self
-                        .freq_decoder
-                        .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
-                    self.remaining_data.advance(num_consumed_bytes);
-                }
-            }
-            // it will be used as the next offset.
-            self.doc_offset = self.doc_decoder.output(COMPRESSION_BLOCK_SIZE - 1);
-            true
-        } else if self.num_vint_docs > 0 {
-            let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
-                self.remaining_data.as_ref(),
-                self.doc_offset,
-                self.num_vint_docs,
-            );
-            self.remaining_data.advance(num_compressed_bytes);
-            match self.freq_reading_option {
-                FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
-                FreqReadingOption::ReadFreq => {
-                    self.freq_decoder
-                        .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
-                }
-            }
-            self.num_vint_docs = 0;
-            true
-        } else {
-            false
-        }
-    }
-
-    /// Returns an empty segment postings object
-    pub fn empty() -> BlockSegmentPostings {
-        BlockSegmentPostings {
-            num_vint_docs: 0,
-
-            doc_decoder: BlockDecoder::new(),
-            freq_decoder: BlockDecoder::with_val(1),
-            freq_reading_option: FreqReadingOption::NoFreq,
-
-            doc_offset: 0,
-            doc_freq: 0,
-
-            remaining_data: OwnedRead::new(vec![]),
-            skip_reader: SkipReader::new(OwnedRead::new(vec![]), IndexRecordOption::Basic),
-        }
-    }
-}
-
-impl<'a> Streamer<'a> for BlockSegmentPostings {
-    type Item = &'a [DocId];
-
-    fn next(&'a mut self) -> Option<&'a [DocId]> {
-        if self.advance() {
-            Some(self.docs())
-        } else {
-            None
-        }
-    }
-}

src/postings/compression/mod.rs

@@ -46,7 +46,7 @@ impl BlockEncoder {
 /// We ensure that the OutputBuffer is align on 128 bits
 /// in order to run SSE2 linear search on it.
 #[repr(align(128))]
-pub(crate) struct AlignedBuffer(pub [u32; COMPRESSION_BLOCK_SIZE]);
+pub struct AlignedBuffer(pub [u32; COMPRESSION_BLOCK_SIZE]);
 
 pub struct BlockDecoder {
     bitpacker: BitPacker4x,
@@ -67,6 +67,18 @@ impl BlockDecoder {
         }
     }
 
+    pub fn uncompress_vint_sorted_b<'a>(
+        &mut self,
+        compressed_data: &'a [u8],
+        offset: u32,
+        num_els: usize,
+    ) {
+        if num_els > 0 {
+            vint::uncompress_sorted(compressed_data, &mut self.output.0[..num_els], offset);
+        }
+        self.output.0[num_els..].iter_mut().for_each(|val| *val = 2_000_000_000u32);
+    }
+
     pub fn uncompress_block_sorted(
         &mut self,
         compressed_data: &[u8],
@@ -94,6 +106,11 @@ impl BlockDecoder {
         (&self.output, self.output_len)
     }
 
+    #[inline]
+    pub(crate) fn output_aligned_2(&self) -> &AlignedBuffer {
+        &self.output
+    }
+
     #[inline]
     pub fn output(&self, idx: usize) -> u32 {
         self.output.0[idx]

src/postings/mod.rs

@@ -2,10 +2,8 @@
 Postings module (also called inverted index)
 */
 
-mod block_max_postings;
-mod block_max_segment_postings;
 mod block_search;
-pub(crate) mod compression;
+pub mod compression;
 /// Postings module
 ///
 /// Postings, also called inverted lists, is the key datastructure
@@ -14,7 +12,6 @@ mod postings;
 mod postings_writer;
 mod recorder;
 mod segment_postings;
-mod block_segment_postings;
 mod serializer;
 mod skip;
 mod stacker;
@@ -30,11 +27,7 @@ pub use self::postings::Postings;
 pub(crate) use self::skip::SkipReader;
 pub use self::term_info::TermInfo;
 
-pub use self::segment_postings::SegmentPostings;
-
-pub use self::block_segment_postings::BlockSegmentPostings;
-pub use self::block_max_postings::BlockMaxPostings;
-pub use self::block_max_segment_postings::BlockMaxSegmentPostings;
+pub use self::segment_postings::{BlockSegmentPostings, SegmentPostings};
 
 pub(crate) use self::stacker::compute_table_size;
 

src/postings/segment_postings.rs

@@ -1,16 +1,21 @@
 use crate::common::BitSet;
 use crate::common::HasLen;
+use crate::common::{BinarySerializable, VInt};
 use crate::docset::{DocSet, SkipResult};
 use crate::positions::PositionReader;
-use crate::postings::compression::COMPRESSION_BLOCK_SIZE;
+use crate::postings::compression::{compressed_block_size, AlignedBuffer};
+use crate::postings::compression::{BlockDecoder, VIntDecoder, COMPRESSION_BLOCK_SIZE};
 use crate::postings::serializer::PostingsSerializer;
-use crate::postings::{BlockSearcher, BlockSegmentPostings};
+use crate::postings::BlockSearcher;
+use crate::postings::FreqReadingOption;
 use crate::postings::Postings;
+use crate::postings::SkipReader;
+use crate::postings::USE_SKIP_INFO_LIMIT;
 use crate::schema::IndexRecordOption;
 use crate::DocId;
 use owned_read::OwnedRead;
 use std::cmp::Ordering;
-use crate::postings::block_segment_postings::BlockSegmentPostingsSkipResult;
+use tantivy_fst::Streamer;
 
 struct PositionComputer {
     // store the amount of position int
@@ -294,6 +299,354 @@ impl Postings for SegmentPostings {
     }
 }
 
+/// `BlockSegmentPostings` is a cursor iterating over blocks
+/// of documents.
+///
+/// # Warning
+///
+/// While it is useful for some very specific high-performance
+/// use cases, you should prefer using `SegmentPostings` for most usage.
+pub struct BlockSegmentPostings {
+    doc_decoder: BlockDecoder,
+    freq_decoder: BlockDecoder,
+    freq_reading_option: FreqReadingOption,
+
+    doc_freq: usize,
+    doc_offset: DocId,
+
+    num_vint_docs: usize,
+
+    remaining_data: OwnedRead,
+    pub skip_reader: SkipReader,
+}
+
+fn split_into_skips_and_postings(
+    doc_freq: u32,
+    mut data: OwnedRead,
+) -> (Option<OwnedRead>, OwnedRead) {
+    if doc_freq >= USE_SKIP_INFO_LIMIT {
+        let skip_len = VInt::deserialize(&mut data).expect("Data corrupted").0 as usize;
+        let mut postings_data = data.clone();
+        postings_data.advance(skip_len);
+        data.clip(skip_len);
+        (Some(data), postings_data)
+    } else {
+        (None, data)
+    }
+}
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum BlockSegmentPostingsSkipResult {
+    Terminated,
+    Success(u32), //< number of term freqs to skip
+}
+
+impl BlockSegmentPostings {
+    pub(crate) fn from_data(
+        doc_freq: u32,
+        data: OwnedRead,
+        record_option: IndexRecordOption,
+        requested_option: IndexRecordOption,
+    ) -> BlockSegmentPostings {
+        let freq_reading_option = match (record_option, requested_option) {
+            (IndexRecordOption::Basic, _) => FreqReadingOption::NoFreq,
+            (_, IndexRecordOption::Basic) => FreqReadingOption::SkipFreq,
+            (_, _) => FreqReadingOption::ReadFreq,
+        };
+
+        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, data);
+        let skip_reader = match skip_data_opt {
+            Some(skip_data) => SkipReader::new(skip_data, record_option),
+            None => SkipReader::new(OwnedRead::new(&[][..]), record_option),
+        };
+        let doc_freq = doc_freq as usize;
+        let num_vint_docs = doc_freq % COMPRESSION_BLOCK_SIZE;
+        BlockSegmentPostings {
+            num_vint_docs,
+            doc_decoder: BlockDecoder::new(),
+            freq_decoder: BlockDecoder::with_val(1),
+            freq_reading_option,
+            doc_offset: 0,
+            doc_freq,
+            remaining_data: postings_data,
+            skip_reader,
+        }
+    }
+
+    // Resets the block segment postings on another position
+    // in the postings file.
+    //
+    // This is useful for enumerating through a list of terms,
+    // and consuming the associated posting lists while avoiding
+    // reallocating a `BlockSegmentPostings`.
+    //
+    // # Warning
+    //
+    // This does not reset the positions list.
+    pub(crate) fn reset(&mut self, doc_freq: u32, postings_data: OwnedRead) {
+        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, postings_data);
+        let num_vint_docs = (doc_freq as usize) & (COMPRESSION_BLOCK_SIZE - 1);
+        self.num_vint_docs = num_vint_docs;
+        self.remaining_data = postings_data;
+        if let Some(skip_data) = skip_data_opt {
+            self.skip_reader.reset(skip_data);
+        } else {
+            self.skip_reader.reset(OwnedRead::new(&[][..]))
+        }
+        self.doc_offset = 0;
+        self.doc_freq = doc_freq as usize;
+    }
+
+    /// Returns the document frequency associated to this block postings.
+    ///
+    /// This `doc_freq` is simply the sum of the length of all of the blocks
+    /// length, and it does not take in account deleted documents.
+    pub fn doc_freq(&self) -> usize {
+        self.doc_freq
+    }
+
+    /// Returns the array of docs in the current block.
+    ///
+    /// Before the first call to `.advance()`, the block
+    /// returned by `.docs()` is empty.
+    #[inline]
+    pub fn docs(&self) -> &[DocId] {
+        self.doc_decoder.output_array()
+    }
+
+    #[inline(always)]
+    pub fn docs_aligned_b(&self) -> &AlignedBuffer {
+        self.doc_decoder.output_aligned_2()
+    }
+
+    pub fn docs_aligned(&self) -> (&AlignedBuffer, usize) {
+        self.doc_decoder.output_aligned()
+    }
+
+    /// Return the document at index `idx` of the block.
+    #[inline]
+    pub fn doc(&self, idx: usize) -> u32 {
+        self.doc_decoder.output(idx)
+    }
+
+    /// Return the array of `term freq` in the block.
+    #[inline]
+    pub fn freqs(&self) -> &[u32] {
+        self.freq_decoder.output_array()
+    }
+
+    /// Return the frequency at index `idx` of the block.
+    #[inline]
+    pub fn freq(&self, idx: usize) -> u32 {
+        self.freq_decoder.output(idx)
+    }
+
+    /// Returns the length of the current block.
+    ///
+    /// All blocks have a length of `NUM_DOCS_PER_BLOCK`,
+    /// except the last block that may have a length
+    /// of any number between 1 and `NUM_DOCS_PER_BLOCK - 1`
+    #[inline]
+    fn block_len(&self) -> usize {
+        self.doc_decoder.output_len
+    }
+
+    /// position on a block that may contains `doc_id`.
+    /// Always advance the current block.
+    ///
+    /// Returns true if a block that has an element greater or equal to the target is found.
+    /// Returning true does not guarantee that the smallest element of the block is smaller
+    /// than the target. It only guarantees that the last element is greater or equal.
+    ///
+    /// Returns false iff all of the document remaining are smaller than
+    /// `doc_id`. In that case, all of these document are consumed.
+    ///
+    pub fn skip_to(&mut self, target_doc: DocId) -> BlockSegmentPostingsSkipResult {
+        let mut skip_freqs = 0u32;
+        while self.skip_reader.advance() {
+            if self.skip_reader.doc() >= target_doc {
+                // the last document of the current block is larger
+                // than the target.
+                //
+                // We found our block!
+                let num_bits = self.skip_reader.doc_num_bits();
+                let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
+                    self.remaining_data.as_ref(),
+                    self.doc_offset,
+                    num_bits,
+                );
+                self.remaining_data.advance(num_consumed_bytes);
+                let tf_num_bits = self.skip_reader.tf_num_bits();
+                match self.freq_reading_option {
+                    FreqReadingOption::NoFreq => {}
+                    FreqReadingOption::SkipFreq => {
+                        let num_bytes_to_skip = compressed_block_size(tf_num_bits);
+                        self.remaining_data.advance(num_bytes_to_skip);
+                    }
+                    FreqReadingOption::ReadFreq => {
+                        let num_consumed_bytes = self
+                            .freq_decoder
+                            .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
+                        self.remaining_data.advance(num_consumed_bytes);
+                    }
+                }
+                self.doc_offset = self.skip_reader.doc();
+                unsafe { core::arch::x86_64::_mm_prefetch(self.remaining_data.as_ref() as *mut i8, _MM_HINT_T1) };
+                return BlockSegmentPostingsSkipResult::Success(skip_freqs);
+            } else {
+                skip_freqs += self.skip_reader.tf_sum();
+                let advance_len = self.skip_reader.total_block_len();
+                self.doc_offset = self.skip_reader.doc();
+                self.remaining_data.advance(advance_len);
+            }
+        }
+
+        // we are now on the last, incomplete, variable encoded block.
+        if self.num_vint_docs > 0 {
+            let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
+                self.remaining_data.as_ref(),
+                self.doc_offset,
+                self.num_vint_docs,
+            );
+            self.remaining_data.advance(num_compressed_bytes);
+            match self.freq_reading_option {
+                FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
+                FreqReadingOption::ReadFreq => {
+                    self.freq_decoder
+                        .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
+                }
+            }
+            self.num_vint_docs = 0;
+            return self
+                .docs()
+                .last()
+                .map(|last_doc| {
+                    if *last_doc >= target_doc {
+                        BlockSegmentPostingsSkipResult::Success(skip_freqs)
+                    } else {
+                        BlockSegmentPostingsSkipResult::Terminated
+                    }
+                })
+                .unwrap_or(BlockSegmentPostingsSkipResult::Terminated);
+        }
+        BlockSegmentPostingsSkipResult::Terminated
+    }
+
+    /// Ensure we are located on the right block.
+    #[inline(never)]
+    pub fn skip_to_b(&mut self, target_doc: DocId) {
+       while self.skip_reader.advance() {
+            if self.skip_reader.doc() >= target_doc {
+                // the last document of the current block is larger
+                // than the target.
+                //
+                // We found our block!
+                let num_bits = self.skip_reader.doc_num_bits();
+                let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
+                    self.remaining_data.as_ref(),
+                    self.doc_offset,
+                    num_bits,
+                );
+                let tf_num_bits = self.skip_reader.tf_num_bits();
+                let num_bytes_to_skip = compressed_block_size(tf_num_bits);
+                self.remaining_data.advance(num_consumed_bytes + num_bytes_to_skip);
+                self.doc_offset = self.skip_reader.doc();
+                return;
+            } else {
+                let advance_len = self.skip_reader.total_block_len();
+                self.doc_offset = self.skip_reader.doc();
+                self.remaining_data.advance(advance_len);
+            }
+        }
+
+        // we are now on the last, incomplete, variable encoded block.
+        self.doc_decoder.uncompress_vint_sorted_b(
+           self.remaining_data.as_ref(),
+           self.doc_offset,
+          self.num_vint_docs
+        );
+    }
+
+    /// Advance to the next block.
+    ///
+    /// Returns false iff there was no remaining blocks.
+    pub fn advance(&mut self) -> bool {
+        if self.skip_reader.advance() {
+            let num_bits = self.skip_reader.doc_num_bits();
+            let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
+                self.remaining_data.as_ref(),
+                self.doc_offset,
+                num_bits,
+            );
+            self.remaining_data.advance(num_consumed_bytes);
+            let tf_num_bits = self.skip_reader.tf_num_bits();
+            match self.freq_reading_option {
+                FreqReadingOption::NoFreq => {}
+                FreqReadingOption::SkipFreq => {
+                    let num_bytes_to_skip = compressed_block_size(tf_num_bits);
+                    self.remaining_data.advance(num_bytes_to_skip);
+                }
+                FreqReadingOption::ReadFreq => {
+                    let num_consumed_bytes = self
+                        .freq_decoder
+                        .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
+                    self.remaining_data.advance(num_consumed_bytes);
+                }
+            }
+            // it will be used as the next offset.
+            self.doc_offset = self.doc_decoder.output(COMPRESSION_BLOCK_SIZE - 1);
+            true
+        } else if self.num_vint_docs > 0 {
+            let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
+                self.remaining_data.as_ref(),
+                self.doc_offset,
+                self.num_vint_docs,
+            );
+            self.remaining_data.advance(num_compressed_bytes);
+            match self.freq_reading_option {
+                FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
+                FreqReadingOption::ReadFreq => {
+                    self.freq_decoder
+                        .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
+                }
+            }
+            self.num_vint_docs = 0;
+            true
+        } else {
+            false
+        }
+    }
+
+    /// Returns an empty segment postings object
+    pub fn empty() -> BlockSegmentPostings {
+        BlockSegmentPostings {
+            num_vint_docs: 0,
+
+            doc_decoder: BlockDecoder::new(),
+            freq_decoder: BlockDecoder::with_val(1),
+            freq_reading_option: FreqReadingOption::NoFreq,
+
+            doc_offset: 0,
+            doc_freq: 0,
+
+            remaining_data: OwnedRead::new(vec![]),
+            skip_reader: SkipReader::new(OwnedRead::new(vec![]), IndexRecordOption::Basic),
+        }
+    }
+}
+
+impl<'b> Streamer<'b> for BlockSegmentPostings {
+    type Item = &'b [DocId];
+
+    fn next(&'b mut self) -> Option<&'b [DocId]> {
+        if self.advance() {
+            Some(self.docs())
+        } else {
+            None
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::BlockSegmentPostings;

src/postings/skip.rs

@@ -49,7 +49,7 @@ impl SkipSerializer {
     }
 }
 
-pub(crate) struct SkipReader {
+pub struct SkipReader {
     doc: DocId,
     owned_read: OwnedRead,
     doc_num_bits: u8,

src/postings/stacker/term_hashmap.rs

@@ -1,4 +1,6 @@
-use murmurhash32::murmurhash2;
+use murmurhash32;
+
+use self::murmurhash32::murmurhash2;
 
 use super::{Addr, MemoryArena};
 use crate::postings::stacker::memory_arena::store;

src/query/block_max_scorer.rs

@@ -1,31 +0,0 @@
-use crate::docset::DocSet;
-use crate::DocId;
-use crate::Score;
-use downcast_rs::impl_downcast;
-use crate::query::Scorer;
-
-/// A set of documents matching a query within a specific segment
-/// and having a maximum score within certain blocks.
-///
-/// See [`Query`](./trait.Query.html) and [`Scorer`](./trait.Scorer.html).
-pub trait BlockMaxScorer: downcast_rs::Downcast + DocSet + Scorer + 'static {
-    /// Returns the maximum score within the current block.
-    ///
-    /// The blocks are defined when indexing. For example, blocks can be
-    /// have a specific number postings each, or can be optimized for
-    /// retrieval speed. Read more in
-    /// [Faster BlockMax WAND with Variable-sized Blocks][vbmw]
-    ///
-    /// This method will perform a bit of computation and is not cached.
-    ///
-    /// [vbmw]: https://dl.acm.org/doi/abs/10.1145/3077136.3080780
-    fn block_max_score(&mut self) -> Score;
-
-    /// Returns the last document in the current block.
-    fn block_max_doc(&mut self) -> DocId;
-
-    /// Returns the maximum possible score within the entire document set.
-    fn max_score(&self) -> Score;
-}
-
-impl_downcast!(BlockMaxScorer);

src/query/block_max_wand.rs

@@ -1,613 +0,0 @@
-use crate::docset::{DocSet, SkipResult};
-use crate::query::score_combiner::ScoreCombiner;
-use crate::query::{BlockMaxScorer, Scorer};
-use crate::DocId;
-use crate::Score;
-use crate::query::weight::PruningScorer;
-
-#[derive(Debug, Copy, Clone, PartialEq, Eq)]
-struct Pivot {
-    position: usize,
-    first_occurrence: usize,
-    doc: DocId,
-}
-
-
-/// Find the position in the sorted list of posting lists of the **pivot**.
-///
-/// docsets need to be advanced, and are required to be sorted by the doc they point to.
-///
-/// The pivot is then defined as the lowest DocId that has a chance of matching our condition.
-fn find_pivot_position<'a, TScorer>(
-    mut docsets: impl Iterator<Item = &'a TScorer>,
-    lower_bound_score: Score,
-) -> Option<Pivot>
-    where TScorer: BlockMaxScorer
-{
-    let mut position = 0;
-    let mut upper_bound = Score::default();
-    while let Some(docset) = docsets.next() {
-        upper_bound += docset.max_score();
-        if lower_bound_score < upper_bound {
-            let pivot_doc = docset.doc();
-            let first_occurrence = position;
-            while let Some(docset) = docsets.next() {
-                if docset.doc() != pivot_doc {
-                    break;
-                } else {
-                    position += 1;
-                }
-            }
-            return Some(Pivot {
-                position,
-                doc: pivot_doc,
-                first_occurrence,
-            });
-        }
-        position += 1;
-    }
-    None
-}
-
-/// Given an iterator over all ordered lists up to the pivot (inclusive) and the following list (if
-/// exists), it returns the next document ID that can be possibly relevant, based on the block max
-/// scores.
-fn find_next_relevant_doc<T, TScorer>(
-    docsets_up_to_pivot: &mut [T],
-    pivot_docset: &mut T,
-    docset_after_pivot: Option<&mut T>,
-) -> DocId
-    where
-        T: AsMut<TScorer>,
-        TScorer: BlockMaxScorer + Scorer,
-{
-    let pivot_docset = pivot_docset.as_mut();
-    let mut next_doc = 1 + docsets_up_to_pivot
-        .iter_mut()
-        .map(|docset| docset.as_mut().block_max_doc())
-        .chain(std::iter::once(pivot_docset.block_max_doc()))
-        .min()
-        .unwrap();
-    if let Some(docset) = docset_after_pivot {
-        let doc = docset.as_mut().doc();
-        if doc < next_doc {
-            next_doc = doc;
-        }
-    }
-    if next_doc <= pivot_docset.doc() {
-        pivot_docset.doc() + 1
-    } else {
-        next_doc
-    }
-}
-
-/// Sifts down the first element of the slice.
-///
-/// `docsets[1..]` are assumed sorted.
-/// This function swaps `docsets[0]` with its right
-/// neighbor successively -bubble sort style- until it reaches the first
-/// position such that `docsets` is sorted.
-fn sift_down<T, TScorer>(docsets: &mut [T])
-    where
-        T: AsRef<TScorer>,
-        TScorer: BlockMaxScorer + Scorer,
-{
-    for idx in 1..docsets.len() {
-        if docsets[idx].as_ref().doc() >= docsets[idx - 1].as_ref().doc() {
-            return;
-        }
-        docsets.swap(idx, idx - 1);
-    }
-}
-
-/// Creates a `DocSet` that iterates through the union of two or more `DocSet`s,
-/// applying [BlockMaxWand] dynamic pruning.
-///
-/// [BlockMaxWand]: https://dl.acm.org/doi/10.1145/2009916.2010048
-pub struct BlockMaxWand<TScorer, TScoreCombiner> {
-    docsets: Vec<Box<TScorer>>,
-    doc: DocId,
-    score: Score,
-    combiner: TScoreCombiner,
-}
-
-impl<TScorer, TScoreCombiner> BlockMaxWand<TScorer, TScoreCombiner>
-    where
-        TScoreCombiner: ScoreCombiner,
-        TScorer: BlockMaxScorer + Scorer,
-{
-    fn new(
-        docsets: Vec<TScorer>,
-        combiner: TScoreCombiner,
-    ) -> BlockMaxWand<TScorer, TScoreCombiner> {
-        let mut non_empty_docsets: Vec<_> = docsets
-            .into_iter()
-            .flat_map(|mut docset| {
-                if docset.advance() {
-                    Some(Box::new(docset))
-                } else {
-                    None
-                }
-            })
-            .collect();
-        non_empty_docsets.sort_by_key(Box::<TScorer>::doc);
-        BlockMaxWand {
-            docsets: non_empty_docsets,
-            combiner,
-            doc: 0u32,
-            score: 0f32
-        }
-    }
-
-    /// Find the position in the sorted list of posting lists of the **pivot**.
-    fn find_pivot_position(&self, lower_bound_score: Score) -> Option<Pivot> {
-        find_pivot_position(
-            self.docsets.iter().map(|docset| docset.as_ref()),
-            lower_bound_score)
-    }
-
-    fn advance_with_pivot(&mut self, pivot: Pivot, lower_bound_score: Score) -> SkipResult {
-        let block_upper_bound: Score = self.docsets[..=pivot.position]
-            .iter_mut()
-            .map(|docset| docset.block_max_score())
-            .sum();
-        if block_upper_bound > lower_bound_score {
-            if pivot.doc == self.docsets[0].doc() {
-                // Since self.docsets is sorted by their current doc, in this branch, all
-                // docsets in [0..=pivot] are positioned on pivot.doc.
-                //
-                // Lets compute the actual score for this doc.
-                //
-                // NOTE(elshize): One additional check needs to be done to improve performance:
-                // update block-wise bound while accumulating score with the actual score,
-                // and check each time if still above threshold.
-                self.combiner.clear();
-                for idx in (0..=pivot.position).rev() {
-                    self.combiner.update(self.docsets[idx].as_mut());
-                    if !self.docsets[idx].advance() {
-                        self.docsets.swap_remove(idx);
-                    }
-                }
-                self.score = self.combiner.score();
-                self.doc = pivot.doc;
-                self.docsets.sort_by_key(Box::<TScorer>::doc);
-                SkipResult::Reached
-            } else {
-                // The substraction does not underflow because otherwise we would go to the other
-                // branch.
-                //
-                // `advanced_idx` is the last idx that is not positionned on the pivot yet.
-                let advanced_idx = pivot.first_occurrence - 1;
-                if !self.docsets[advanced_idx].advance() {
-                    self.docsets.swap_remove(advanced_idx);
-                }
-                if self.docsets.is_empty() {
-                    return SkipResult::End;
-                }
-                sift_down(&mut self.docsets[advanced_idx..]);
-                SkipResult::OverStep
-            }
-        } else {
-            let (up_to_pivot, pivot_and_rest) = self.docsets.split_at_mut(pivot.position as usize);
-            let (pivot, after_pivot) = pivot_and_rest.split_first_mut().unwrap();
-            let next_doc = find_next_relevant_doc(up_to_pivot, pivot, after_pivot.first_mut());
-            // NOTE(elshize): It might be more efficient to advance the list with the higher
-            // max score, but let's advance the first one for now for simplicity.
-            if self.docsets[0].skip_next(next_doc) == SkipResult::End {
-                self.docsets.swap_remove(0);
-            }
-            if self.docsets.is_empty() {
-                return SkipResult::End;
-            }
-            sift_down(&mut self.docsets[..]);
-            SkipResult::OverStep
-        }
-    }
-}
-
-impl<TScorer, TScoreCombiner> PruningScorer
-for BlockMaxWand<TScorer, TScoreCombiner>
-    where
-        TScoreCombiner: ScoreCombiner,
-        TScorer: Scorer + BlockMaxScorer,
-{
-    fn doc(&self) -> DocId {
-        self.doc
-    }
-
-    fn score(&self) -> Score {
-        self.score
-    }
-
-    fn advance_with_pruning(&mut self, lower_bound_score: f32) -> bool {
-        while let Some(pivot) = self.find_pivot_position(lower_bound_score) {
-            match self.advance_with_pivot(pivot, lower_bound_score) {
-                SkipResult::End => { return false },
-                SkipResult::Reached=> { return true; }
-                SkipResult::OverStep => {}
-            }
-        }
-        false
-    }
-}
-
-
-#[cfg(test)]
-mod tests {
-
-    use super::*;
-    use crate::common::HasLen;
-    use crate::docset::DocSet;
-    use crate::query::score_combiner::SumCombiner;
-    use crate::query::Union;
-    use crate::query::{BlockMaxScorer, Scorer};
-    use crate::{DocId, Score};
-    use float_cmp::approx_eq;
-    use proptest::strategy::Strategy;
-    use std::cmp::Ordering;
-    use std::num::Wrapping;
-    use crate::collector::{SegmentCollector, TopScoreSegmentCollector};
-
-    /*
-    #[derive(Debug, Clone)]
-    pub struct VecDocSet {
-        postings: Vec<(DocId, Score)>,
-        cursor: Wrapping<usize>,
-        block_max_scores: Vec<(DocId, Score)>,
-        max_score: Score,
-        block_size: usize,
-    }
-
-    impl VecDocSet {
-        fn new(postings: Vec<(DocId, Score)>, block_size: usize) -> VecDocSet {
-            let block_max_scores: Vec<(DocId, f32)> = postings
-                .chunks(block_size)
-                .into_iter()
-                .map(|block| {
-                    (
-                        block.iter().last().unwrap().0,
-                        block
-                            .iter()
-                            .map(|(_, s)| *s)
-                            .fold(-f32::INFINITY, |left, right| left.max(right))
-                    )
-                })
-                .collect();
-            let max_score = block_max_scores
-                .iter()
-                .copied()
-                .map(|(_, s)| s)
-                .fold(-f32::INFINITY, |left, right| left.max(right));
-            VecDocSet {
-                postings,
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores,
-                max_score,
-                block_size,
-            }
-        }
-        /// Constructs a new set and advances it.
-        fn started(postings: Vec<(DocId, Score)>, block_size: usize) -> VecDocSet {
-            let mut docset = VecDocSet::new(postings, block_size);
-            docset.advance();
-            docset
-        }
-    }
-
-    impl DocSet for VecDocSet {
-        fn advance(&mut self) -> bool {
-            self.cursor += Wrapping(1);
-            self.postings.len() > self.cursor.0
-        }
-
-        fn doc(&self) -> DocId {
-            self.postings[self.cursor.0].0
-        }
-
-        fn size_hint(&self) -> u32 {
-            self.len() as u32
-        }
-    }
-
-    impl HasLen for VecDocSet {
-        fn len(&self) -> usize {
-            self.postings.len()
-        }
-    }
-
-    impl BlockMaxScorer for VecDocSet {
-        fn max_score(&self) -> Score {
-            self.max_score
-        }
-        fn block_max_score(&mut self) -> Score {
-            self.block_max_scores[self.cursor.0 / self.block_size].1
-        }
-        fn block_max_doc(&mut self) -> DocId {
-            self.block_max_scores[self.cursor.0 / self.block_size].0
-        }
-    }
-
-    impl Scorer for VecDocSet {
-        fn score(&mut self) -> Score {
-            self.postings[self.cursor.0].1
-        }
-    }
-
-    #[derive(Debug)]
-    struct ComparableDoc<T, D> {
-        feature: T,
-        doc: D,
-    }
-
-    impl<T: PartialOrd, D: PartialOrd> PartialOrd for ComparableDoc<T, D> {
-        fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
-            Some(self.cmp(other))
-        }
-    }
-
-    impl<T: PartialOrd, D: PartialOrd> Ord for ComparableDoc<T, D> {
-        #[inline]
-        fn cmp(&self, other: &Self) -> Ordering {
-            // Reversed to make BinaryHeap work as a min-heap
-            let by_feature = other
-                .feature
-                .partial_cmp(&self.feature)
-                .unwrap_or(Ordering::Equal);
-
-            let lazy_by_doc_address =
-                || self.doc.partial_cmp(&other.doc).unwrap_or(Ordering::Equal);
-
-            // In case of a tie on the feature, we sort by ascending
-            // `DocAddress` in order to ensure a stable sorting of the
-            // documents.
-            by_feature.then_with(lazy_by_doc_address)
-        }
-    }
-
-    impl<T: PartialOrd, D: PartialOrd> PartialEq for ComparableDoc<T, D> {
-        fn eq(&self, other: &Self) -> bool {
-            self.cmp(other) == Ordering::Equal
-        }
-    }
-
-    impl<T: PartialOrd, D: PartialOrd> Eq for ComparableDoc<T, D> {}
-
-    fn union_vs_bmw(posting_lists: Vec<VecDocSet>) {
-        let mut union = Union::<VecDocSet, SumCombiner>::from(posting_lists.clone());
-        let mut top_union = TopScoreSegmentCollector::new(0, 10);
-        while union.advance() {
-            top_union.collect(union.doc(), union.score());
-        }
-        let top_bmw = TopScoreSegmentCollector::new(0, 10 );
-        let mut bmw = BlockMaxWand::new(posting_lists, SumCombiner::default());
-        let top_docs_bnw = top_bmw.collect_scorer(&mut bmw, None);
-        for ((expected_score, expected_doc), (actual_score, actual_doc)) in
-        top_union.harvest().into_iter().zip( top_docs_bnw )
-        {
-            assert!(approx_eq!(
-                f32,
-                expected_score,
-                actual_score,
-                epsilon = 0.0001
-            ));
-            assert_eq!(expected_doc, actual_doc);
-        }
-    }
-
-    #[test]
-    fn test_bmw_0() {
-        union_vs_bmw(vec![
-            VecDocSet {
-                postings: vec![
-                    (0, 1.0),
-                    (23, 1.0),
-                    (28, 1.0),
-                    (56, 1.0),
-                    (59, 1.0),
-                    (66, 1.0),
-                    (93, 1.0),
-                ],
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores: vec![(93, 1.0)],
-                max_score: 1.0,
-                block_size: 16,
-            },
-            VecDocSet {
-                postings: vec![
-                    (2, 1.6549665),
-                    (43, 2.6958032),
-                    (53, 3.5309567),
-                    (71, 2.7688136),
-                    (87, 3.4279852),
-                    (96, 3.9028034),
-                ],
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores: vec![(96, 3.9028034)],
-                max_score: 3.9028034,
-                block_size: 16,
-            },
-        ])
-    }
-
-    #[test]
-    fn test_bmw_1() {
-        union_vs_bmw(vec![
-            VecDocSet {
-                postings: vec![(73, 1.0), (82, 1.0)],
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores: vec![(82, 1.0)],
-                max_score: 1.0,
-                block_size: 16,
-            },
-            VecDocSet {
-                postings: vec![
-                    (21, 3.582513),
-                    (23, 1.6928024),
-                    (27, 3.887647),
-                    (42, 1.5469292),
-                    (61, 1.7317574),
-                    (62, 1.2968783),
-                    (82, 2.4040694),
-                    (85, 3.1487892),
-                ],
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores: vec![(85, 3.887647)],
-                max_score: 3.887647,
-                block_size: 16,
-            },
-        ])
-    }
-
-    proptest::proptest! {
-        #[test]
-        fn test_union_vs_bmw(postings in proptest::collection::vec(
-            proptest::collection::vec(0_u32..100, 1..10)
-                .prop_flat_map(|v| {
-                    let scores = proptest::collection::vec(1_f32..4_f32, v.len()..=v.len());
-                    scores.prop_map(move |s| {
-                        let mut postings: Vec<_> = v.iter().copied().zip(s.iter().copied()).collect();
-                        postings.sort_by_key(|p| p.0);
-                        postings.dedup_by_key(|p| p.0);
-                        VecDocSet::new(postings, 16)
-                    })
-                }),
-            2..5)
-        ) {
-            union_vs_bmw(postings);
-        }
-    }
-
-    #[test]
-    fn test_find_pivot_position() {
-        let postings = vec![
-            VecDocSet::started(vec![(0, 2.0)], 1),
-            VecDocSet::started(vec![(1, 3.0)], 1),
-            VecDocSet::started(vec![(2, 4.0)], 1),
-            VecDocSet::started(vec![(3, 5.0)], 1),
-            VecDocSet::started(vec![(3, 6.0)], 1),
-        ];
-        assert_eq!(
-            find_pivot_position(postings.iter(), 2.0f32),
-            Some(Pivot {
-                position: 1,
-                doc: 1,
-                first_occurrence: 1,
-            })
-        );
-        assert_eq!(
-            find_pivot_position(postings.iter(), 5.0f32),
-            Some(Pivot {
-                position: 2,
-                doc: 2,
-                first_occurrence: 2,
-            })
-        );
-        assert_eq!(
-            find_pivot_position(postings.iter(), 9.0f32),
-            Some(Pivot {
-                position: 4,
-                doc: 3,
-                first_occurrence: 3,
-            })
-        );
-        assert_eq!(
-            find_pivot_position(postings.iter(), 20.0f32),
-            None
-        );
-    }
-
-    #[test]
-    fn test_find_next_relevant_doc_before_pivot() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (3, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (4, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(2, 0.0), (6, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)),
-        ];
-        let (up_to_pivot, rest) = postings.split_at_mut(2);
-        let (pivot, after_pivot) = rest.split_first_mut().unwrap();
-        let next_doc = find_next_relevant_doc(up_to_pivot, pivot, Some(&mut after_pivot[0]));
-        assert_eq!(next_doc, 4);
-    }
-
-    #[test]
-    fn test_find_next_relevant_doc_prefix_smaller_than_pivot() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (3, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (4, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(5, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)),
-        ];
-        let (up_to_pivot, rest) = postings.split_at_mut(2);
-        let (pivot, after_pivot) = rest.split_first_mut().unwrap();
-        let next_doc = find_next_relevant_doc(up_to_pivot, pivot, Some(&mut after_pivot[0]));
-        assert_eq!(next_doc, 6);
-    }
-
-    #[test]
-    fn test_find_next_relevant_doc_after_pivot() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(2, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(5, 0.0), (7, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (7, 0.0)], 2)),
-        ];
-        let (up_to_pivot, rest) = postings.split_at_mut(2);
-        let (pivot, after_pivot) = rest.split_first_mut().unwrap();
-        let next_doc = find_next_relevant_doc(up_to_pivot, pivot, Some(&mut after_pivot[0]));
-        assert_eq!(next_doc, 5);
-    }
-
-    #[test]
-    fn test_sift_down_already_sifted() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(2, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(5, 0.0), (7, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (7, 0.0)], 2)),
-        ];
-        sift_down(&mut postings[2..]);
-        assert_eq!(
-            postings.into_iter().map(|p| p.doc()).collect::<Vec<_>>(),
-            vec![0, 1, 2, 5, 6]
-        );
-    }
-
-    #[test]
-    fn test_sift_down_sift_one_down() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(5, 0.0), (7, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(7, 0.0), (7, 0.0)], 2)),
-        ];
-        sift_down(&mut postings[2..]);
-        assert_eq!(
-            postings.into_iter().map(|p| p.doc()).collect::<Vec<_>>(),
-            vec![0, 1, 5, 6, 7]
-        );
-    }
-
-    #[test]
-    fn test_sift_down_to_bottom() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(7, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(5, 0.0), (7, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (7, 0.0)], 2)),
-        ];
-        sift_down(&mut postings[2..]);
-        assert_eq!(
-            postings.into_iter().map(|p| p.doc()).collect::<Vec<_>>(),
-            vec![0, 1, 5, 6, 7]
-        );
-    }
-    
-     */
-}

src/query/mod.rs

@@ -1,10 +1,10 @@
-/*! Query Module */
+/*!
+Query
+*/
 
 mod all_query;
 mod automaton_weight;
 mod bitset;
-mod block_max_scorer;
-mod block_max_wand;
 mod bm25;
 mod boolean_query;
 mod boost_query;
@@ -24,6 +24,7 @@ mod term_query;
 mod union;
 mod weight;
 
+
 #[cfg(test)]
 mod vec_docset;
 
@@ -37,15 +38,13 @@ pub use self::vec_docset::VecDocSet;
 pub use self::all_query::{AllQuery, AllScorer, AllWeight};
 pub use self::automaton_weight::AutomatonWeight;
 pub use self::bitset::BitSetDocSet;
-pub use self::block_max_scorer::BlockMaxScorer;
 pub use self::boolean_query::BooleanQuery;
 pub use self::boost_query::BoostQuery;
 pub use self::empty_query::{EmptyQuery, EmptyScorer, EmptyWeight};
 pub use self::exclude::Exclude;
 pub use self::explanation::Explanation;
-#[cfg(test)]
-pub(crate) use self::fuzzy_query::DFAWrapper;
 pub use self::fuzzy_query::FuzzyTermQuery;
+pub(crate) use self::fuzzy_query::DFAWrapper;
 pub use self::intersection::intersect_scorers;
 pub use self::phrase_query::PhraseQuery;
 pub use self::query::Query;
@@ -59,8 +58,6 @@ pub use self::scorer::Scorer;
 pub use self::term_query::TermQuery;
 pub use self::weight::Weight;
 pub use tantivy_query_grammar::Occur;
-pub use self::weight::PruningScorerIfPossible;
-
 
 #[cfg(test)]
 mod tests {

src/query/query.rs

@@ -1,11 +1,17 @@
 use super::Weight;
 use crate::core::searcher::Searcher;
 use crate::query::Explanation;
-use crate::DocAddress;
+use crate::{DocAddress, Score};
 use crate::Term;
 use downcast_rs::impl_downcast;
 use std::collections::BTreeSet;
 use std::fmt;
+use crate::collector::{TopDocs, Collector, SegmentCollector, Count};
+
+pub struct TopKResult {
+    pub count: u64,
+    docs: Vec<(Score, DocAddress)>
+}
 
 /// The `Query` trait defines a set of documents and a scoring method
 /// for those documents.
@@ -56,6 +62,27 @@ pub trait Query: QueryClone + downcast_rs::Downcast + fmt::Debug {
         weight.explain(reader, doc_address.doc())
     }
 
+    fn top_k(&self, searcher: &Searcher, num_hits: usize) -> crate::Result<TopKResult> {
+        let top_docs = TopDocs::with_limit(num_hits);
+        let collector = (Count, top_docs);
+        let weight = self.weight(searcher, false)?;
+        let mut count = 0u64;
+        let mut result = 0;
+        let mut child_fruits = Vec::with_capacity(searcher.segment_readers().len());
+        for (segment_ord, reader) in searcher.segment_readers().iter().enumerate() {
+            let mut child_top_k = collector.for_segment(segment_ord as u32, reader)?;
+            let mut scorer = weight.scorer(reader, 1.0f32)?;
+            // handle deletes
+            scorer.top_k(&mut child_top_k);
+            child_fruits.push(child_top_k.harvest());
+        }
+        let (count, docs) = collector.merge_fruits(child_fruits)?;
+        Ok(TopKResult {
+            count: count as u64,
+            docs
+        })
+    }
+
     /// Returns the number of documents matching the query.
     fn count(&self, searcher: &Searcher) -> crate::Result<usize> {
         let weight = self.weight(searcher, false)?;

src/query/scorer.rs

@@ -4,6 +4,8 @@ use crate::DocId;
 use crate::Score;
 use downcast_rs::impl_downcast;
 use std::ops::DerefMut;
+use crate::collector::{SegmentCollector, Collector, TopDocs, Count};
+
 
 /// Scored set of documents matching a query within a specific segment.
 ///
@@ -21,6 +23,12 @@ pub trait Scorer: downcast_rs::Downcast + DocSet + 'static {
             callback(self.doc(), self.score());
         }
     }
+
+    fn top_k(&mut self, collector: &mut <(Count, TopDocs) as Collector>::Child) {
+        while self.advance() {
+            collector.collect(self.doc(), self.score());
+        }
+    }
 }
 
 impl_downcast!(Scorer);

src/query/term_query/block_max_term_scorer.rs

@@ -1,98 +0,0 @@
-use crate::docset::{DocSet, SkipResult};
-use crate::query::{Explanation, Scorer};
-use crate::DocId;
-use crate::Score;
-
-use crate::fieldnorm::FieldNormReader;
-use crate::postings::Postings;
-use crate::postings::{BlockMaxPostings, BlockMaxSegmentPostings};
-use crate::query::bm25::BM25Weight;
-use crate::query::BlockMaxScorer;
-
-pub struct BlockMaxTermScorer {
-    postings: BlockMaxSegmentPostings,
-    fieldnorm_reader: FieldNormReader,
-    similarity_weight: BM25Weight,
-}
-
-impl BlockMaxTermScorer {
-    pub fn new(
-        postings: BlockMaxSegmentPostings,
-        fieldnorm_reader: FieldNormReader,
-        similarity_weight: BM25Weight,
-    ) -> Self {
-        Self {
-            postings,
-            fieldnorm_reader,
-            similarity_weight,
-        }
-    }
-}
-
-impl BlockMaxTermScorer {
-    fn _score(&self, fieldnorm_id: u8, term_freq: u32) -> Score {
-        self.similarity_weight.score(fieldnorm_id, term_freq)
-    }
-
-    pub fn term_freq(&self) -> u32 {
-        self.postings.term_freq()
-    }
-
-    pub fn fieldnorm_id(&self) -> u8 {
-        self.fieldnorm_reader.fieldnorm_id(self.doc())
-    }
-
-    pub fn explain(&self) -> Explanation {
-        let fieldnorm_id = self.fieldnorm_id();
-        let term_freq = self.term_freq();
-        self.similarity_weight.explain(fieldnorm_id, term_freq)
-    }
-}
-
-impl DocSet for BlockMaxTermScorer {
-    fn advance(&mut self) -> bool {
-        self.postings.advance()
-    }
-
-    fn skip_next(&mut self, target: DocId) -> SkipResult {
-        self.postings.skip_next(target)
-    }
-
-    fn doc(&self) -> DocId {
-        self.postings.doc()
-    }
-
-    fn size_hint(&self) -> u32 {
-        self.postings.size_hint()
-    }
-}
-
-impl Scorer for BlockMaxTermScorer {
-    fn score(&mut self) -> Score {
-        self._score(
-            self.fieldnorm_reader.fieldnorm_id(self.doc()),
-            self.postings.term_freq(),
-        )
-    }
-}
-
-impl BlockMaxScorer for BlockMaxTermScorer {
-    fn block_max_score(&mut self) -> Score {
-        self._score(
-            self.fieldnorm_reader
-                .fieldnorm_id(self.postings.block_max_doc()),
-            self.postings.term_freq(),
-        )
-    }
-
-    fn block_max_doc(&mut self) -> DocId {
-        self.postings.block_max_doc()
-    }
-
-    fn max_score(&self) -> Score {
-        self._score(
-            self.fieldnorm_reader.fieldnorm_id(self.postings.max_doc()),
-            self.postings.max_term_freq(),
-        )
-    }
-}

src/query/term_query/mod.rs

@@ -1,9 +1,7 @@
-mod block_max_term_scorer;
 mod term_query;
 mod term_scorer;
 mod term_weight;
 
-pub use self::block_max_term_scorer::BlockMaxTermScorer;
 pub use self::term_query::TermQuery;
 pub use self::term_scorer::TermScorer;
 pub use self::term_weight::TermWeight;

src/query/union.rs

@@ -1,7 +1,7 @@
 use crate::common::TinySet;
 use crate::docset::{DocSet, SkipResult};
 use crate::query::score_combiner::{DoNothingCombiner, ScoreCombiner};
-use crate::query::{Scorer, BlockMaxScorer};
+use crate::query::Scorer;
 use crate::DocId;
 use crate::Score;
 use std::cmp::Ordering;
@@ -9,99 +9,6 @@ use std::cmp::Ordering;
 const HORIZON_NUM_TINYBITSETS: usize = 64;
 const HORIZON: u32 = 64u32 * HORIZON_NUM_TINYBITSETS as u32;
 
-
-#[derive(Debug, Copy, Clone, PartialEq, Eq)]
-struct Pivot {
-    position: usize,
-    first_occurrence: usize,
-    doc: DocId,
-}
-
-/// Find the position in the sorted list of posting lists of the **pivot**.
-///
-/// docsets need to be advanced, and are required to be sorted by the doc they point to.
-///
-/// The pivot is then defined as the lowest DocId that has a chance of matching our condition.
-fn find_pivot_position<'a, TScorer>(
-    mut docsets: impl Iterator<Item = &'a TScorer>,
-    lower_bound_score: Score,
-) -> Option<Pivot>
-    where TScorer: BlockMaxScorer
-{
-    let mut position = 0;
-    let mut upper_bound = Score::default();
-    while let Some(docset) = docsets.next() {
-        upper_bound += docset.max_score();
-        if lower_bound_score < upper_bound {
-            let pivot_doc = docset.doc();
-            let first_occurrence = position;
-            while let Some(docset) = docsets.next() {
-                if docset.doc() != pivot_doc {
-                    break;
-                } else {
-                    position += 1;
-                }
-            }
-            return Some(Pivot {
-                position,
-                doc: pivot_doc,
-                first_occurrence,
-            });
-        }
-        position += 1;
-    }
-    None
-}
-
-/// Sifts down the first element of the slice.
-///
-/// `docsets[1..]` are assumed sorted.
-/// This function swaps `docsets[0]` with its right
-/// neighbor successively -bubble sort style- until it reaches the first
-/// position such that `docsets` is sorted.
-fn sift_down<TScorer>(docsets: &mut [TScorer])
-    where
-        TScorer: BlockMaxScorer + Scorer,
-{
-    for idx in 1..docsets.len() {
-        if docsets[idx].doc() >= docsets[idx - 1].doc() {
-            return;
-        }
-        docsets.swap(idx, idx - 1);
-    }
-}
-
-
-/// Given an iterator over all ordered lists up to the pivot (inclusive) and the following list (if
-/// exists), it returns the next document ID that can be possibly relevant, based on the block max
-/// scores.
-fn find_next_relevant_doc<TScorer>(
-    docsets_up_to_pivot: &mut [TScorer],
-    pivot_docset: &mut TScorer,
-    docset_after_pivot: Option<&mut TScorer>,
-) -> DocId
-    where
-        TScorer: BlockMaxScorer + Scorer,
-{
-    let mut next_doc = 1 + docsets_up_to_pivot
-        .iter_mut()
-        .map(|docset| docset.block_max_doc())
-        .chain(std::iter::once(pivot_docset.block_max_doc()))
-        .min()
-        .unwrap();
-    if let Some(docset) = docset_after_pivot {
-        let doc = docset.doc();
-        if doc < next_doc {
-            next_doc = doc;
-        }
-    }
-    if next_doc <= pivot_docset.doc() {
-        pivot_docset.doc() + 1
-    } else {
-        next_doc
-    }
-}
-
 // `drain_filter` is not stable yet.
 // This function is similar except that it does is not unstable, and
 // it does not keep the original vector ordering.
@@ -132,7 +39,6 @@ pub struct Union<TScorer, TScoreCombiner = DoNothingCombiner> {
     score: Score,
 }
 
-
 impl<TScorer, TScoreCombiner> From<Vec<TScorer>> for Union<TScorer, TScoreCombiner>
 where
     TScoreCombiner: ScoreCombiner,
@@ -220,76 +126,6 @@ impl<TScorer: Scorer, TScoreCombiner: ScoreCombiner> Union<TScorer, TScoreCombin
         }
         false
     }
-
-
-}
-
-
-impl<TScorer: BlockMaxScorer, TScoreCombiner: ScoreCombiner> Union<TScorer, TScoreCombiner> {
-    fn advance_with_pivot(&mut self, pivot: Pivot, lower_bound_score: Score) -> SkipResult {
-        let block_upper_bound: Score = self.docsets[..=pivot.position]
-            .iter_mut()
-            .map(|docset| docset.block_max_score())
-            .sum();
-        if block_upper_bound > lower_bound_score {
-            if pivot.doc == self.docsets[0].doc() {
-                // Since self.docsets is sorted by their current doc, in this branch, all
-                // docsets in [0..=pivot] are positioned on pivot.doc.
-                //
-                // Lets compute the actual score for this doc.
-                //
-                // NOTE(elshize): One additional check needs to be done to improve performance:
-                // update block-wise bound while accumulating score with the actual score,
-                // and check each time if still above threshold.
-                let mut combiner = TScoreCombiner::default();
-                for idx in (0..=pivot.position).rev() {
-                    combiner.update(&mut self.docsets[idx]);
-                    if !self.docsets[idx].advance() {
-                        self.docsets.swap_remove(idx);
-                    }
-                }
-                self.score = combiner.score();
-                self.doc = pivot.doc;
-                self.docsets.sort_by_key(TScorer::doc);
-                SkipResult::Reached
-            } else {
-                // The substraction does not underflow because otherwise we would go to the other
-                // branch.
-                //
-                // `advanced_idx` is the last idx that is not positionned on the pivot yet.
-                let advanced_idx = pivot.first_occurrence - 1;
-                if !self.docsets[advanced_idx].advance() {
-                    self.docsets.swap_remove(advanced_idx);
-                }
-                if self.docsets.is_empty() {
-                    return SkipResult::End;
-                }
-                sift_down(&mut self.docsets[advanced_idx..]);
-                SkipResult::OverStep
-            }
-        } else {
-            let (up_to_pivot, pivot_and_rest) = self.docsets.split_at_mut(pivot.position as usize);
-            let (pivot, after_pivot) = pivot_and_rest.split_first_mut().unwrap();
-            let next_doc = find_next_relevant_doc(up_to_pivot, pivot, after_pivot.first_mut());
-            // NOTE(elshize): It might be more efficient to advance the list with the higher
-            // max score, but let's advance the first one for now for simplicity.
-            if self.docsets[0].skip_next(next_doc) == SkipResult::End {
-                self.docsets.swap_remove(0);
-            }
-            if self.docsets.is_empty() {
-                return SkipResult::End;
-            }
-            sift_down(&mut self.docsets[..]);
-            SkipResult::OverStep
-        }
-    }
-
-    /// Find the position in the sorted list of posting lists of the **pivot**.
-    fn find_pivot_position(&self, lower_bound_score: Score) -> Option<Pivot> {
-        find_pivot_position(
-            self.docsets.iter().map(|docset| docset),
-            lower_bound_score)
-    }
 }
 
 impl<TScorer, TScoreCombiner> DocSet for Union<TScorer, TScoreCombiner>

src/query/weight.rs

@@ -1,26 +1,7 @@
 use super::Scorer;
 use crate::core::SegmentReader;
 use crate::query::Explanation;
-use crate::{DocId, Score};
-
-pub trait PruningScorer {
-
-    fn doc(&self) -> DocId;
-
-    fn score(&self) -> Score;
-
-    /// Advance to the next document that has a score strictly greater than
-    /// `lower_bound_score`.
-    fn advance_with_pruning(&mut self, score_lower_bound: f32) -> bool;
-    fn advance(&mut self) -> bool {
-        self.advance_with_pruning(std::f32::NEG_INFINITY)
-    }
-}
-
-pub enum PruningScorerIfPossible {
-    Pruning(Box<dyn PruningScorer>),
-    NonPruning(Box<dyn Scorer>)
-}
+use crate::DocId;
 
 /// A Weight is the specialization of a Query
 /// for a given set of segments.
@@ -34,11 +15,6 @@ pub trait Weight: Send + Sync + 'static {
     /// See [`Query`](./trait.Query.html).
     fn scorer(&self, reader: &SegmentReader, boost: f32) -> crate::Result<Box<dyn Scorer>>;
 
-    fn pruning_scorer(&self, reader: &SegmentReader, boost: f32) -> crate::Result<PruningScorerIfPossible> {
-        let scorer = self.scorer(reader, boost)?;
-        Ok(PruningScorerIfPossible::NonPruning(Box::new(scorer)))
-    }
-
     /// Returns an `Explanation` for the given document.
     fn explain(&self, reader: &SegmentReader, doc: DocId) -> crate::Result<Explanation>;
 

src/reader/mod.rs

@@ -9,7 +9,6 @@ use crate::directory::META_LOCK;
 use crate::Index;
 use crate::Searcher;
 use crate::SegmentReader;
-use std::convert::TryInto;
 use std::sync::Arc;
 
 /// Defines when a new version of the index should be reloaded.
@@ -61,6 +60,7 @@ impl IndexReaderBuilder {
     /// Building the reader is a non-trivial operation that requires
     /// to open different segment readers. It may take hundreds of milliseconds
     /// of time and it may return an error.
+    /// TODO(pmasurel) Use the `TryInto` trait once it is available in stable.
     pub fn try_into(self) -> crate::Result<IndexReader> {
         let inner_reader = InnerIndexReader {
             index: self.index,
@@ -113,14 +113,6 @@ impl IndexReaderBuilder {
     }
 }
 
-impl TryInto<IndexReader> for IndexReaderBuilder {
-    type Error = crate::TantivyError;
-
-    fn try_into(self) -> crate::Result<IndexReader> {
-        IndexReaderBuilder::try_into(self)
-    }
-}
-
 struct InnerIndexReader {
     num_searchers: usize,
     searcher_pool: Pool<Searcher>,

src/schema/document.rs

@@ -3,6 +3,7 @@ use crate::common::BinarySerializable;
 use crate::common::VInt;
 use crate::tokenizer::PreTokenizedString;
 use crate::DateTime;
+use serde;
 use std::io::{self, Read, Write};
 use std::mem;
 

src/schema/field.rs

@@ -1,4 +1,5 @@
 use crate::common::BinarySerializable;
+use serde;
 use std::io;
 use std::io::Read;
 use std::io::Write;

src/schema/field_value.rs

@@ -1,7 +1,10 @@
 use crate::common::BinarySerializable;
 use crate::schema::Field;
 use crate::schema::Value;
-use std::io::{self, Read, Write};
+use serde;
+use std::io;
+use std::io::Read;
+use std::io::Write;
 
 /// `FieldValue` holds together a `Field` and its `Value`.
 #[derive(Debug, Clone, Ord, PartialEq, Eq, PartialOrd, serde::Serialize, serde::Deserialize)]

src/store/compression_snap.rs

@@ -1,3 +1,5 @@
+use snap;
+
 use std::io::{self, Read, Write};
 
 /// Name of the compression scheme used in the doc store.

src/termdict/mod.rs

@@ -434,8 +434,8 @@ mod tests {
 
     #[test]
     fn test_automaton_search() {
-        use crate::query::DFAWrapper;
         use levenshtein_automata::LevenshteinAutomatonBuilder;
+        use crate::query::DFAWrapper;
 
         const COUNTRIES: [&'static str; 7] = [
             "San Marino",

src/termdict/termdict.rs

@@ -7,6 +7,7 @@ use crate::postings::TermInfo;
 use crate::termdict::TermOrdinal;
 use once_cell::sync::Lazy;
 use std::io::{self, Write};
+use tantivy_fst;
 use tantivy_fst::raw::Fst;
 use tantivy_fst::Automaton;