Merge pull request #2443 from miDeb/sort/data-oriented

sort: separate additional data from the Line struct

src/uu/sort/src/check.rs

@@ -8,7 +8,7 @@
 //! Check if a file is ordered
 
 use crate::{
-    chunks::{self, Chunk},
+    chunks::{self, Chunk, RecycledChunk},
     compare_by, open, GlobalSettings,
 };
 use itertools::Itertools;
@@ -34,7 +34,7 @@ pub fn check(path: &str, settings: &GlobalSettings) -> i32 {
         move || reader(file, recycled_receiver, loaded_sender, &settings)
     });
     for _ in 0..2 {
-        let _ = recycled_sender.send(Chunk::new(vec![0; 100 * 1024], |_| Vec::new()));
+        let _ = recycled_sender.send(RecycledChunk::new(100 * 1024));
     }
 
     let mut prev_chunk: Option<Chunk> = None;
@@ -44,21 +44,29 @@ pub fn check(path: &str, settings: &GlobalSettings) -> i32 {
         if let Some(prev_chunk) = prev_chunk.take() {
             // Check if the first element of the new chunk is greater than the last
             // element from the previous chunk
-            let prev_last = prev_chunk.borrow_lines().last().unwrap();
+            let prev_last = prev_chunk.lines().last().unwrap();
-            let new_first = chunk.borrow_lines().first().unwrap();
+            let new_first = chunk.lines().first().unwrap();
 
-            if compare_by(prev_last, new_first, settings) == Ordering::Greater {
+            if compare_by(
+                prev_last,
+                new_first,
+                settings,
+                prev_chunk.line_data(),
+                chunk.line_data(),
+            ) == Ordering::Greater
+            {
                 if !settings.check_silent {
                     println!("sort: {}:{}: disorder: {}", path, line_idx, new_first.line);
                 }
                 return 1;
             }
-            let _ = recycled_sender.send(prev_chunk);
+            let _ = recycled_sender.send(prev_chunk.recycle());
         }
 
-        for (a, b) in chunk.borrow_lines().iter().tuple_windows() {
+        for (a, b) in chunk.lines().iter().tuple_windows() {
             line_idx += 1;
-            if compare_by(a, b, settings) == Ordering::Greater {
+            if compare_by(a, b, settings, chunk.line_data(), chunk.line_data()) == Ordering::Greater
+            {
                 if !settings.check_silent {
                     println!("sort: {}:{}: disorder: {}", path, line_idx, b.line);
                 }
@@ -74,16 +82,15 @@ pub fn check(path: &str, settings: &GlobalSettings) -> i32 {
 /// The function running on the reader thread.
 fn reader(
     mut file: Box<dyn Read + Send>,
-    receiver: Receiver<Chunk>,
+    receiver: Receiver<RecycledChunk>,
     sender: SyncSender<Chunk>,
     settings: &GlobalSettings,
 ) {
     let mut carry_over = vec![];
-    for chunk in receiver.iter() {
+    for recycled_chunk in receiver.iter() {
-        let (recycled_lines, recycled_buffer) = chunk.recycle();
         let should_continue = chunks::read(
             &sender,
-            recycled_buffer,
+            recycled_chunk,
             None,
             &mut carry_over,
             &mut file,
@@ -93,7 +100,6 @@ fn reader(
             } else {
                 b'\n'
             },
-            recycled_lines,
             settings,
         );
         if !should_continue {

src/uu/sort/src/chunks.rs

@@ -15,7 +15,7 @@ use std::{
 use memchr::memchr_iter;
 use ouroboros::self_referencing;
 
-use crate::{GlobalSettings, Line};
+use crate::{numeric_str_cmp::NumInfo, GeneralF64ParseResult, GlobalSettings, Line};
 
 /// The chunk that is passed around between threads.
 /// `lines` consist of slices into `buffer`.
@@ -25,28 +25,87 @@ pub struct Chunk {
     pub buffer: Vec<u8>,
     #[borrows(buffer)]
     #[covariant]
-    pub lines: Vec<Line<'this>>,
+    pub contents: ChunkContents<'this>,
+}
+
+#[derive(Debug)]
+pub struct ChunkContents<'a> {
+    pub lines: Vec<Line<'a>>,
+    pub line_data: LineData<'a>,
+}
+
+#[derive(Debug)]
+pub struct LineData<'a> {
+    pub selections: Vec<&'a str>,
+    pub num_infos: Vec<NumInfo>,
+    pub parsed_floats: Vec<GeneralF64ParseResult>,
 }
 
 impl Chunk {
     /// Destroy this chunk and return its components to be reused.
-    ///
+    pub fn recycle(mut self) -> RecycledChunk {
-    /// # Returns
+        let recycled_contents = self.with_contents_mut(|contents| {
-    ///
+            contents.lines.clear();
-    /// * The `lines` vector, emptied
+            contents.line_data.selections.clear();
-    /// * The `buffer` vector, **not** emptied
+            contents.line_data.num_infos.clear();
-    pub fn recycle(mut self) -> (Vec<Line<'static>>, Vec<u8>) {
+            contents.line_data.parsed_floats.clear();
-        let recycled_lines = self.with_lines_mut(|lines| {
+            let lines = unsafe {
-            lines.clear();
-            unsafe {
                 // SAFETY: It is safe to (temporarily) transmute to a vector of lines with a longer lifetime,
                 // because the vector is empty.
                 // Transmuting is necessary to make recycling possible. See https://github.com/rust-lang/rfcs/pull/2802
                 // for a rfc to make this unnecessary. Its example is similar to the code here.
-                std::mem::transmute::<Vec<Line<'_>>, Vec<Line<'static>>>(std::mem::take(lines))
+                std::mem::transmute::<Vec<Line<'_>>, Vec<Line<'static>>>(std::mem::take(
-            }
+                    &mut contents.lines,
+                ))
+            };
+            let selections = unsafe {
+                // SAFETY: (same as above) It is safe to (temporarily) transmute to a vector of &str with a longer lifetime,
+                // because the vector is empty.
+                std::mem::transmute::<Vec<&'_ str>, Vec<&'static str>>(std::mem::take(
+                    &mut contents.line_data.selections,
+                ))
+            };
+            (
+                lines,
+                selections,
+                std::mem::take(&mut contents.line_data.num_infos),
+                std::mem::take(&mut contents.line_data.parsed_floats),
+            )
         });
-        (recycled_lines, self.into_heads().buffer)
+        RecycledChunk {
+            lines: recycled_contents.0,
+            selections: recycled_contents.1,
+            num_infos: recycled_contents.2,
+            parsed_floats: recycled_contents.3,
+            buffer: self.into_heads().buffer,
+        }
+    }
+
+    pub fn lines(&self) -> &Vec<Line> {
+        &self.borrow_contents().lines
+    }
+    pub fn line_data(&self) -> &LineData {
+        &self.borrow_contents().line_data
+    }
+}
+
+pub struct RecycledChunk {
+    lines: Vec<Line<'static>>,
+    selections: Vec<&'static str>,
+    num_infos: Vec<NumInfo>,
+    parsed_floats: Vec<GeneralF64ParseResult>,
+    buffer: Vec<u8>,
+}
+
+impl RecycledChunk {
+    pub fn new(capacity: usize) -> Self {
+        RecycledChunk {
+            lines: Vec::new(),
+            selections: Vec::new(),
+            num_infos: Vec::new(),
+            parsed_floats: Vec::new(),
+            buffer: vec![0; capacity],
+        }
     }
 }
 
@@ -63,28 +122,32 @@ impl Chunk {
 /// (see also `read_to_chunk` for a more detailed documentation)
 ///
 /// * `sender`: The sender to send the lines to the sorter.
-/// * `buffer`: The recycled buffer. All contents will be overwritten, but it must already be filled.
+/// * `recycled_chunk`: The recycled chunk, as returned by `Chunk::recycle`.
 ///   (i.e. `buffer.len()` should be equal to `buffer.capacity()`)
 /// * `max_buffer_size`: How big `buffer` can be.
 /// * `carry_over`: The bytes that must be carried over in between invocations.
 /// * `file`: The current file.
 /// * `next_files`: What `file` should be updated to next.
 /// * `separator`: The line separator.
-/// * `lines`: The recycled vector to fill with lines. Must be empty.
 /// * `settings`: The global settings.
 #[allow(clippy::too_many_arguments)]
 pub fn read<T: Read>(
     sender: &SyncSender<Chunk>,
-    mut buffer: Vec<u8>,
+    recycled_chunk: RecycledChunk,
     max_buffer_size: Option<usize>,
     carry_over: &mut Vec<u8>,
     file: &mut T,
     next_files: &mut impl Iterator<Item = T>,
     separator: u8,
-    lines: Vec<Line<'static>>,
     settings: &GlobalSettings,
 ) -> bool {
-    assert!(lines.is_empty());
+    let RecycledChunk {
+        lines,
+        selections,
+        num_infos,
+        parsed_floats,
+        mut buffer,
+    } = recycled_chunk;
     if buffer.len() < carry_over.len() {
         buffer.resize(carry_over.len() + 10 * 1024, 0);
     }
@@ -101,15 +164,25 @@ pub fn read<T: Read>(
     carry_over.extend_from_slice(&buffer[read..]);
 
     if read != 0 {
-        let payload = Chunk::new(buffer, |buf| {
+        let payload = Chunk::new(buffer, |buffer| {
+            let selections = unsafe {
+                // SAFETY: It is safe to transmute to an empty vector of selections with shorter lifetime.
+                // It was only temporarily transmuted to a Vec<Line<'static>> to make recycling possible.
+                std::mem::transmute::<Vec<&'static str>, Vec<&'_ str>>(selections)
+            };
             let mut lines = unsafe {
-                // SAFETY: It is safe to transmute to a vector of lines with shorter lifetime,
+                // SAFETY: (same as above) It is safe to transmute to a vector of lines with shorter lifetime,
                 // because it was only temporarily transmuted to a Vec<Line<'static>> to make recycling possible.
                 std::mem::transmute::<Vec<Line<'static>>, Vec<Line<'_>>>(lines)
             };
-            let read = crash_if_err!(1, std::str::from_utf8(&buf[..read]));
+            let read = crash_if_err!(1, std::str::from_utf8(&buffer[..read]));
-            parse_lines(read, &mut lines, separator, settings);
+            let mut line_data = LineData {
-            lines
+                selections,
+                num_infos,
+                parsed_floats,
+            };
+            parse_lines(read, &mut lines, &mut line_data, separator, settings);
+            ChunkContents { lines, line_data }
         });
         sender.send(payload).unwrap();
     }
@@ -120,6 +193,7 @@ pub fn read<T: Read>(
 fn parse_lines<'a>(
     mut read: &'a str,
     lines: &mut Vec<Line<'a>>,
+    line_data: &mut LineData<'a>,
     separator: u8,
     settings: &GlobalSettings,
 ) {
@@ -128,9 +202,15 @@ fn parse_lines<'a>(
         read = &read[..read.len() - 1];
     }
 
+    assert!(lines.is_empty());
+    assert!(line_data.selections.is_empty());
+    assert!(line_data.num_infos.is_empty());
+    assert!(line_data.parsed_floats.is_empty());
+    let mut token_buffer = vec![];
     lines.extend(
         read.split(separator as char)
-            .map(|line| Line::create(line, settings)),
+            .enumerate()
+            .map(|(index, line)| Line::create(line, index, line_data, &mut token_buffer, settings)),
     );
 }
 

src/uu/sort/src/ext_sort.rs

@@ -23,15 +23,16 @@ use std::{
 
 use itertools::Itertools;
 
+use crate::chunks::RecycledChunk;
 use crate::merge::ClosedTmpFile;
 use crate::merge::WriteableCompressedTmpFile;
 use crate::merge::WriteablePlainTmpFile;
 use crate::merge::WriteableTmpFile;
-use crate::Line;
 use crate::{
     chunks::{self, Chunk},
-    compare_by, merge, output_sorted_lines, sort_by, GlobalSettings,
+    compare_by, merge, sort_by, GlobalSettings,
 };
+use crate::{print_sorted, Line};
 use tempfile::TempDir;
 
 const START_BUFFER_SIZE: usize = 8_000;
@@ -98,16 +99,39 @@ fn reader_writer<F: Iterator<Item = Box<dyn Read + Send>>, Tmp: WriteableTmpFile
             merger.write_all(settings);
         }
         ReadResult::SortedSingleChunk(chunk) => {
-            output_sorted_lines(chunk.borrow_lines().iter(), settings);
+            if settings.unique {
+                print_sorted(
+                    chunk.lines().iter().dedup_by(|a, b| {
+                        compare_by(a, b, settings, chunk.line_data(), chunk.line_data())
+                            == Ordering::Equal
+                    }),
+                    settings,
+                );
+            } else {
+                print_sorted(chunk.lines().iter(), settings);
+            }
         }
         ReadResult::SortedTwoChunks([a, b]) => {
-            let merged_iter = a
+            let merged_iter = a.lines().iter().map(|line| (line, &a)).merge_by(
-                .borrow_lines()
+                b.lines().iter().map(|line| (line, &b)),
-                .iter()
+                |(line_a, a), (line_b, b)| {
-                .merge_by(b.borrow_lines().iter(), |line_a, line_b| {
+                    compare_by(line_a, line_b, settings, a.line_data(), b.line_data())
-                    compare_by(line_a, line_b, settings) != Ordering::Greater
+                        != Ordering::Greater
-                });
+                },
-            output_sorted_lines(merged_iter, settings);
+            );
+            if settings.unique {
+                print_sorted(
+                    merged_iter
+                        .dedup_by(|(line_a, a), (line_b, b)| {
+                            compare_by(line_a, line_b, settings, a.line_data(), b.line_data())
+                                == Ordering::Equal
+                        })
+                        .map(|(line, _)| line),
+                    settings,
+                );
+            } else {
+                print_sorted(merged_iter.map(|(line, _)| line), settings);
+            }
         }
         ReadResult::EmptyInput => {
             // don't output anything
@@ -118,7 +142,9 @@ fn reader_writer<F: Iterator<Item = Box<dyn Read + Send>>, Tmp: WriteableTmpFile
 /// The function that is executed on the sorter thread.
 fn sorter(receiver: Receiver<Chunk>, sender: SyncSender<Chunk>, settings: GlobalSettings) {
     while let Ok(mut payload) = receiver.recv() {
-        payload.with_lines_mut(|lines| sort_by(lines, &settings));
+        payload.with_contents_mut(|contents| {
+            sort_by(&mut contents.lines, &settings, &contents.line_data)
+        });
         sender.send(payload).unwrap();
     }
 }
@@ -154,20 +180,16 @@ fn read_write_loop<I: WriteableTmpFile>(
     for _ in 0..2 {
         let should_continue = chunks::read(
             &sender,
-            vec![
+            RecycledChunk::new(if START_BUFFER_SIZE < buffer_size {
-                0;
+                START_BUFFER_SIZE
-                if START_BUFFER_SIZE < buffer_size {
+            } else {
-                    START_BUFFER_SIZE
+                buffer_size
-                } else {
+            }),
-                    buffer_size
-                }
-            ],
             Some(buffer_size),
             &mut carry_over,
             &mut file,
             &mut files,
             separator,
-            Vec::new(),
             settings,
         );
 
@@ -216,18 +238,17 @@ fn read_write_loop<I: WriteableTmpFile>(
 
         file_number += 1;
 
-        let (recycled_lines, recycled_buffer) = chunk.recycle();
+        let recycled_chunk = chunk.recycle();
 
         if let Some(sender) = &sender_option {
             let should_continue = chunks::read(
                 sender,
-                recycled_buffer,
+                recycled_chunk,
                 None,
                 &mut carry_over,
                 &mut file,
                 &mut files,
                 separator,
-                recycled_lines,
                 settings,
             );
             if !should_continue {
@@ -245,12 +266,9 @@ fn write<I: WriteableTmpFile>(
     compress_prog: Option<&str>,
     separator: u8,
 ) -> I::Closed {
-    chunk.with_lines_mut(|lines| {
+    let mut tmp_file = I::create(file, compress_prog);
-        // Write the lines to the file
+    write_lines(chunk.lines(), tmp_file.as_write(), separator);
-        let mut tmp_file = I::create(file, compress_prog);
+    tmp_file.finished_writing()
-        write_lines(lines, tmp_file.as_write(), separator);
-        tmp_file.finished_writing()
-    })
 }
 
 fn write_lines<'a, T: Write>(lines: &[Line<'a>], writer: &mut T, separator: u8) {

src/uu/sort/src/merge.rs

@@ -24,7 +24,7 @@ use itertools::Itertools;
 use tempfile::TempDir;
 
 use crate::{
-    chunks::{self, Chunk},
+    chunks::{self, Chunk, RecycledChunk},
     compare_by, GlobalSettings,
 };
 
@@ -125,14 +125,14 @@ fn merge_without_limit<M: MergeInput + 'static, F: Iterator<Item = M>>(
         }));
         // Send the initial chunk to trigger a read for each file
         request_sender
-            .send((file_number, Chunk::new(vec![0; 8 * 1024], |_| Vec::new())))
+            .send((file_number, RecycledChunk::new(8 * 1024)))
             .unwrap();
     }
 
     // Send the second chunk for each file
     for file_number in 0..reader_files.len() {
         request_sender
-            .send((file_number, Chunk::new(vec![0; 8 * 1024], |_| Vec::new())))
+            .send((file_number, RecycledChunk::new(8 * 1024)))
             .unwrap();
     }
 
@@ -181,13 +181,12 @@ struct ReaderFile<M: MergeInput> {
 
 /// The function running on the reader thread.
 fn reader(
-    recycled_receiver: Receiver<(usize, Chunk)>,
+    recycled_receiver: Receiver<(usize, RecycledChunk)>,
     files: &mut [Option<ReaderFile<impl MergeInput>>],
     settings: &GlobalSettings,
     separator: u8,
 ) {
-    for (file_idx, chunk) in recycled_receiver.iter() {
+    for (file_idx, recycled_chunk) in recycled_receiver.iter() {
-        let (recycled_lines, recycled_buffer) = chunk.recycle();
         if let Some(ReaderFile {
             file,
             sender,
@@ -196,13 +195,12 @@ fn reader(
         {
             let should_continue = chunks::read(
                 sender,
-                recycled_buffer,
+                recycled_chunk,
                 None,
                 carry_over,
                 file.as_read(),
                 &mut iter::empty(),
                 separator,
-                recycled_lines,
                 settings,
             );
             if !should_continue {
@@ -234,7 +232,7 @@ struct PreviousLine {
 /// Merges files together. This is **not** an iterator because of lifetime problems.
 pub struct FileMerger<'a> {
     heap: binary_heap_plus::BinaryHeap<MergeableFile, FileComparator<'a>>,
-    request_sender: Sender<(usize, Chunk)>,
+    request_sender: Sender<(usize, RecycledChunk)>,
     prev: Option<PreviousLine>,
 }
 
@@ -257,14 +255,16 @@ impl<'a> FileMerger<'a> {
                 file_number: file.file_number,
             });
 
-            file.current_chunk.with_lines(|lines| {
+            file.current_chunk.with_contents(|contents| {
-                let current_line = &lines[file.line_idx];
+                let current_line = &contents.lines[file.line_idx];
                 if settings.unique {
                     if let Some(prev) = &prev {
                         let cmp = compare_by(
-                            &prev.chunk.borrow_lines()[prev.line_idx],
+                            &prev.chunk.lines()[prev.line_idx],
                             current_line,
                             settings,
+                            prev.chunk.line_data(),
+                            file.current_chunk.line_data(),
                         );
                         if cmp == Ordering::Equal {
                             return;
@@ -274,8 +274,7 @@ impl<'a> FileMerger<'a> {
                 current_line.print(out, settings);
             });
 
-            let was_last_line_for_file =
+            let was_last_line_for_file = file.current_chunk.lines().len() == file.line_idx + 1;
-                file.current_chunk.borrow_lines().len() == file.line_idx + 1;
 
             if was_last_line_for_file {
                 if let Ok(next_chunk) = file.receiver.recv() {
@@ -295,7 +294,7 @@ impl<'a> FileMerger<'a> {
                     // If nothing is referencing the previous chunk anymore, this means that the previous line
                     // was the last line of the chunk. We can recycle the chunk.
                     self.request_sender
-                        .send((prev.file_number, prev_chunk))
+                        .send((prev.file_number, prev_chunk.recycle()))
                         .ok();
                 }
             }
@@ -312,9 +311,11 @@ struct FileComparator<'a> {
 impl<'a> Compare<MergeableFile> for FileComparator<'a> {
     fn compare(&self, a: &MergeableFile, b: &MergeableFile) -> Ordering {
         let mut cmp = compare_by(
-            &a.current_chunk.borrow_lines()[a.line_idx],
+            &a.current_chunk.lines()[a.line_idx],
-            &b.current_chunk.borrow_lines()[b.line_idx],
+            &b.current_chunk.lines()[b.line_idx],
             self.settings,
+            a.current_chunk.line_data(),
+            b.current_chunk.line_data(),
         );
         if cmp == Ordering::Equal {
             // To make sorting stable, we need to consider the file number as well,

src/uu/sort/src/sort.rs

@@ -23,11 +23,11 @@ mod ext_sort;
 mod merge;
 mod numeric_str_cmp;
 
+use chunks::LineData;
 use clap::{crate_version, App, Arg};
 use custom_str_cmp::custom_str_cmp;
 use ext_sort::ext_sort;
 use fnv::FnvHasher;
-use itertools::Itertools;
 use numeric_str_cmp::{numeric_str_cmp, NumInfo, NumInfoParseSettings};
 use rand::distributions::Alphanumeric;
 use rand::{thread_rng, Rng};
@@ -170,6 +170,17 @@ pub struct GlobalSettings {
     tmp_dir: PathBuf,
     compress_prog: Option<String>,
     merge_batch_size: usize,
+    precomputed: Precomputed,
+}
+
+/// Data needed for sorting. Should be computed once before starting to sort
+/// by calling `GlobalSettings::init_precomputed`.
+#[derive(Clone, Debug)]
+struct Precomputed {
+    needs_tokens: bool,
+    num_infos_per_line: usize,
+    floats_per_line: usize,
+    selections_per_line: usize,
 }
 
 impl GlobalSettings {
@@ -210,6 +221,25 @@ impl GlobalSettings {
             None => BufWriter::new(Box::new(stdout()) as Box<dyn Write>),
         }
     }
+
+    /// Precompute some data needed for sorting.
+    /// This function **must** be called before starting to sort, and `GlobalSettings` may not be altered
+    /// afterwards.
+    fn init_precomputed(&mut self) {
+        self.precomputed.needs_tokens = self.selectors.iter().any(|s| s.needs_tokens);
+        self.precomputed.selections_per_line =
+            self.selectors.iter().filter(|s| s.needs_selection).count();
+        self.precomputed.num_infos_per_line = self
+            .selectors
+            .iter()
+            .filter(|s| matches!(s.settings.mode, SortMode::Numeric | SortMode::HumanNumeric))
+            .count();
+        self.precomputed.floats_per_line = self
+            .selectors
+            .iter()
+            .filter(|s| matches!(s.settings.mode, SortMode::GeneralNumeric))
+            .count();
+    }
 }
 
 impl Default for GlobalSettings {
@@ -237,9 +267,16 @@ impl Default for GlobalSettings {
             tmp_dir: PathBuf::new(),
             compress_prog: None,
             merge_batch_size: 32,
+            precomputed: Precomputed {
+                num_infos_per_line: 0,
+                floats_per_line: 0,
+                selections_per_line: 0,
+                needs_tokens: false,
+            },
         }
     }
 }
+
 #[derive(Clone, PartialEq, Debug)]
 struct KeySettings {
     mode: SortMode,
@@ -322,32 +359,10 @@ impl Default for KeySettings {
         Self::from(&GlobalSettings::default())
     }
 }
-
+enum Selection<'a> {
-#[derive(Clone, Debug)]
-enum NumCache {
     AsF64(GeneralF64ParseResult),
-    WithInfo(NumInfo),
+    WithNumInfo(&'a str, NumInfo),
-}
+    Str(&'a str),
-
-impl NumCache {
-    fn as_f64(&self) -> GeneralF64ParseResult {
-        match self {
-            NumCache::AsF64(n) => *n,
-            _ => unreachable!(),
-        }
-    }
-    fn as_num_info(&self) -> &NumInfo {
-        match self {
-            NumCache::WithInfo(n) => n,
-            _ => unreachable!(),
-        }
-    }
-}
-
-#[derive(Clone, Debug)]
-struct Selection<'a> {
-    slice: &'a str,
-    num_cache: Option<Box<NumCache>>,
 }
 
 type Field = Range<usize>;
@@ -355,31 +370,44 @@ type Field = Range<usize>;
 #[derive(Clone, Debug)]
 pub struct Line<'a> {
     line: &'a str,
-    selections: Box<[Selection<'a>]>,
+    index: usize,
 }
 
 impl<'a> Line<'a> {
-    fn create(string: &'a str, settings: &GlobalSettings) -> Self {
+    /// Creates a new `Line`.
-        let fields = if settings
+    ///
+    /// If additional data is needed for sorting it is added to `line_data`.
+    /// `token_buffer` allows to reuse the allocation for tokens.
+    fn create(
+        line: &'a str,
+        index: usize,
+        line_data: &mut LineData<'a>,
+        token_buffer: &mut Vec<Field>,
+        settings: &GlobalSettings,
+    ) -> Self {
+        token_buffer.clear();
+        if settings.precomputed.needs_tokens {
+            tokenize(line, settings.separator, token_buffer);
+        }
+        for (selector, selection) in settings
             .selectors
             .iter()
-            .any(|selector| selector.needs_tokens)
+            .map(|selector| (selector, selector.get_selection(line, token_buffer)))
         {
-            // Only tokenize if we will need tokens.
+            match selection {
-            Some(tokenize(string, settings.separator))
+                Selection::AsF64(parsed_float) => line_data.parsed_floats.push(parsed_float),
-        } else {
+                Selection::WithNumInfo(str, num_info) => {
-            None
+                    line_data.num_infos.push(num_info);
-        };
+                    line_data.selections.push(str);
-
+                }
-        Line {
+                Selection::Str(str) => {
-            line: string,
+                    if selector.needs_selection {
-            selections: settings
+                        line_data.selections.push(str)
-                .selectors
+                    }
-                .iter()
+                }
-                .filter(|selector| !selector.is_default_selection)
+            }
-                .map(|selector| selector.get_selection(string, fields.as_deref()))
-                .collect(),
         }
+        Self { line, index }
     }
 
     fn print(&self, writer: &mut impl Write, settings: &GlobalSettings) {
@@ -408,7 +436,8 @@ impl<'a> Line<'a> {
         let line = self.line.replace('\t', ">");
         writeln!(writer, "{}", line)?;
 
-        let fields = tokenize(self.line, settings.separator);
+        let mut fields = vec![];
+        tokenize(self.line, settings.separator, &mut fields);
         for selector in settings.selectors.iter() {
             let mut selection = selector.get_range(self.line, Some(&fields));
             match selector.settings.mode {
@@ -539,51 +568,51 @@ impl<'a> Line<'a> {
     }
 }
 
-/// Tokenize a line into fields.
+/// Tokenize a line into fields. The result is stored into `token_buffer`.
-fn tokenize(line: &str, separator: Option<char>) -> Vec<Field> {
+fn tokenize(line: &str, separator: Option<char>, token_buffer: &mut Vec<Field>) {
+    assert!(token_buffer.is_empty());
     if let Some(separator) = separator {
-        tokenize_with_separator(line, separator)
+        tokenize_with_separator(line, separator, token_buffer)
     } else {
-        tokenize_default(line)
+        tokenize_default(line, token_buffer)
     }
 }
 
 /// By default fields are separated by the first whitespace after non-whitespace.
 /// Whitespace is included in fields at the start.
-fn tokenize_default(line: &str) -> Vec<Field> {
+/// The result is stored into `token_buffer`.
-    let mut tokens = vec![0..0];
+fn tokenize_default(line: &str, token_buffer: &mut Vec<Field>) {
+    token_buffer.push(0..0);
     // pretend that there was whitespace in front of the line
     let mut previous_was_whitespace = true;
     for (idx, char) in line.char_indices() {
         if char.is_whitespace() {
             if !previous_was_whitespace {
-                tokens.last_mut().unwrap().end = idx;
+                token_buffer.last_mut().unwrap().end = idx;
-                tokens.push(idx..0);
+                token_buffer.push(idx..0);
             }
             previous_was_whitespace = true;
         } else {
             previous_was_whitespace = false;
         }
     }
-    tokens.last_mut().unwrap().end = line.len();
+    token_buffer.last_mut().unwrap().end = line.len();
-    tokens
 }
 
 /// Split between separators. These separators are not included in fields.
-fn tokenize_with_separator(line: &str, separator: char) -> Vec<Field> {
+/// The result is stored into `token_buffer`.
-    let mut tokens = vec![];
+fn tokenize_with_separator(line: &str, separator: char, token_buffer: &mut Vec<Field>) {
     let separator_indices =
         line.char_indices()
             .filter_map(|(i, c)| if c == separator { Some(i) } else { None });
     let mut start = 0;
     for sep_idx in separator_indices {
-        tokens.push(start..sep_idx);
+        token_buffer.push(start..sep_idx);
         start = sep_idx + 1;
     }
     if start < line.len() {
-        tokens.push(start..line.len());
+        token_buffer.push(start..line.len());
     }
-    tokens
 }
 
 #[derive(Clone, PartialEq, Debug)]
@@ -640,8 +669,10 @@ struct FieldSelector {
     to: Option<KeyPosition>,
     settings: KeySettings,
     needs_tokens: bool,
-    // Whether the selection for each line is going to be the whole line with no NumCache
+    // Whether this selector operates on a sub-slice of a line.
-    is_default_selection: bool,
+    // Selections are therefore not needed when this selector matches the whole line
+    // or the sort mode is general-numeric.
+    needs_selection: bool,
 }
 
 impl Default for FieldSelector {
@@ -651,7 +682,7 @@ impl Default for FieldSelector {
             to: None,
             settings: Default::default(),
             needs_tokens: false,
-            is_default_selection: true,
+            needs_selection: false,
         }
     }
 }
@@ -747,14 +778,12 @@ impl FieldSelector {
             Err("invalid character index 0 for the start position of a field".to_string())
         } else {
             Ok(Self {
-                is_default_selection: from.field == 1
+                needs_selection: (from.field != 1
-                    && from.char == 1
+                    || from.char != 1
-                    && to.is_none()
+                    || to.is_some()
-                    && !matches!(
+                    || matches!(settings.mode, SortMode::Numeric | SortMode::HumanNumeric)
-                        settings.mode,
+                    || from.ignore_blanks)
-                        SortMode::Numeric | SortMode::GeneralNumeric | SortMode::HumanNumeric
+                    && !matches!(settings.mode, SortMode::GeneralNumeric),
-                    )
-                    && !from.ignore_blanks,
                 needs_tokens: from.field != 1 || from.char == 0 || to.is_some(),
                 from,
                 to,
@@ -764,12 +793,16 @@ impl FieldSelector {
     }
 
     /// Get the selection that corresponds to this selector for the line.
-    /// If needs_fields returned false, tokens may be None.
+    /// If needs_fields returned false, tokens may be empty.
-    fn get_selection<'a>(&self, line: &'a str, tokens: Option<&[Field]>) -> Selection<'a> {
+    fn get_selection<'a>(&self, line: &'a str, tokens: &[Field]) -> Selection<'a> {
+        // `get_range` expects `None` when we don't need tokens and would get confused by an empty vector.
+        let tokens = if self.needs_tokens {
+            Some(tokens)
+        } else {
+            None
+        };
         let mut range = &line[self.get_range(line, tokens)];
-        let num_cache = if self.settings.mode == SortMode::Numeric
+        if self.settings.mode == SortMode::Numeric || self.settings.mode == SortMode::HumanNumeric {
-            || self.settings.mode == SortMode::HumanNumeric
-        {
             // Parse NumInfo for this number.
             let (info, num_range) = NumInfo::parse(
                 range,
@@ -780,24 +813,18 @@ impl FieldSelector {
             );
             // Shorten the range to what we need to pass to numeric_str_cmp later.
             range = &range[num_range];
-            Some(Box::new(NumCache::WithInfo(info)))
+            Selection::WithNumInfo(range, info)
         } else if self.settings.mode == SortMode::GeneralNumeric {
             // Parse this number as f64, as this is the requirement for general numeric sorting.
-            Some(Box::new(NumCache::AsF64(general_f64_parse(
+            Selection::AsF64(general_f64_parse(&range[get_leading_gen(range)]))
-                &range[get_leading_gen(range)],
-            ))))
         } else {
             // This is not a numeric sort, so we don't need a NumCache.
-            None
+            Selection::Str(range)
-        };
-        Selection {
-            slice: range,
-            num_cache,
         }
     }
 
     /// Look up the range in the line that corresponds to this selector.
-    /// If needs_fields returned false, tokens may be None.
+    /// If needs_fields returned false, tokens must be None.
     fn get_range<'a>(&self, line: &'a str, tokens: Option<&[Field]>) -> Range<usize> {
         enum Resolution {
             // The start index of the resolved character, inclusive
@@ -1297,18 +1324,9 @@ pub fn uumain(args: impl uucore::Args) -> i32 {
         );
     }
 
-    exec(&files, &settings)
+    settings.init_precomputed();
-}
 
-fn output_sorted_lines<'a>(iter: impl Iterator<Item = &'a Line<'a>>, settings: &GlobalSettings) {
+    exec(&files, &settings)
-    if settings.unique {
-        print_sorted(
-            iter.dedup_by(|a, b| compare_by(a, b, settings) == Ordering::Equal),
-            settings,
-        );
-    } else {
-        print_sorted(iter, settings);
-    }
 }
 
 fn exec(files: &[String], settings: &GlobalSettings) -> i32 {
@@ -1328,55 +1346,59 @@ fn exec(files: &[String], settings: &GlobalSettings) -> i32 {
     0
 }
 
-fn sort_by<'a>(unsorted: &mut Vec<Line<'a>>, settings: &GlobalSettings) {
+fn sort_by<'a>(unsorted: &mut Vec<Line<'a>>, settings: &GlobalSettings, line_data: &LineData<'a>) {
     if settings.stable || settings.unique {
-        unsorted.par_sort_by(|a, b| compare_by(a, b, settings))
+        unsorted.par_sort_by(|a, b| compare_by(a, b, settings, line_data, line_data))
     } else {
-        unsorted.par_sort_unstable_by(|a, b| compare_by(a, b, settings))
+        unsorted.par_sort_unstable_by(|a, b| compare_by(a, b, settings, line_data, line_data))
     }
 }
 
-fn compare_by<'a>(a: &Line<'a>, b: &Line<'a>, global_settings: &GlobalSettings) -> Ordering {
+fn compare_by<'a>(
-    let mut idx = 0;
+    a: &Line<'a>,
+    b: &Line<'a>,
+    global_settings: &GlobalSettings,
+    a_line_data: &LineData<'a>,
+    b_line_data: &LineData<'a>,
+) -> Ordering {
+    let mut selection_index = 0;
+    let mut num_info_index = 0;
+    let mut parsed_float_index = 0;
     for selector in &global_settings.selectors {
-        let mut _selections = None;
+        let (a_str, b_str) = if !selector.needs_selection {
-        let (a_selection, b_selection) = if selector.is_default_selection {
             // We can select the whole line.
-            // We have to store the selections outside of the if-block so that they live long enough.
+            (a.line, b.line)
-            _selections = Some((
-                Selection {
-                    slice: a.line,
-                    num_cache: None,
-                },
-                Selection {
-                    slice: b.line,
-                    num_cache: None,
-                },
-            ));
-            // Unwrap the selections again, and return references to them.
-            (
-                &_selections.as_ref().unwrap().0,
-                &_selections.as_ref().unwrap().1,
-            )
         } else {
-            let selections = (&a.selections[idx], &b.selections[idx]);
+            let selections = (
-            idx += 1;
+                a_line_data.selections
+                    [a.index * global_settings.precomputed.selections_per_line + selection_index],
+                b_line_data.selections
+                    [b.index * global_settings.precomputed.selections_per_line + selection_index],
+            );
+            selection_index += 1;
             selections
         };
-        let a_str = a_selection.slice;
+
-        let b_str = b_selection.slice;
         let settings = &selector.settings;
 
         let cmp: Ordering = match settings.mode {
             SortMode::Random => random_shuffle(a_str, b_str, &global_settings.salt),
-            SortMode::Numeric | SortMode::HumanNumeric => numeric_str_cmp(
+            SortMode::Numeric | SortMode::HumanNumeric => {
-                (a_str, a_selection.num_cache.as_ref().unwrap().as_num_info()),
+                let a_num_info = &a_line_data.num_infos
-                (b_str, b_selection.num_cache.as_ref().unwrap().as_num_info()),
+                    [a.index * global_settings.precomputed.num_infos_per_line + num_info_index];
-            ),
+                let b_num_info = &b_line_data.num_infos
-            SortMode::GeneralNumeric => general_numeric_compare(
+                    [b.index * global_settings.precomputed.num_infos_per_line + num_info_index];
-                a_selection.num_cache.as_ref().unwrap().as_f64(),
+                num_info_index += 1;
-                b_selection.num_cache.as_ref().unwrap().as_f64(),
+                numeric_str_cmp((a_str, a_num_info), (b_str, b_num_info))
-            ),
+            }
+            SortMode::GeneralNumeric => {
+                let a_float = &a_line_data.parsed_floats
+                    [a.index * global_settings.precomputed.floats_per_line + parsed_float_index];
+                let b_float = &b_line_data.parsed_floats
+                    [b.index * global_settings.precomputed.floats_per_line + parsed_float_index];
+                parsed_float_index += 1;
+                general_numeric_compare(a_float, b_float)
+            }
             SortMode::Month => month_compare(a_str, b_str),
             SortMode::Version => version_compare(a_str, b_str),
             SortMode::Default => custom_str_cmp(
@@ -1470,7 +1492,7 @@ fn get_leading_gen(input: &str) -> Range<usize> {
 }
 
 #[derive(Copy, Clone, PartialEq, PartialOrd, Debug)]
-enum GeneralF64ParseResult {
+pub enum GeneralF64ParseResult {
     Invalid,
     NaN,
     NegInfinity,
@@ -1497,8 +1519,8 @@ fn general_f64_parse(a: &str) -> GeneralF64ParseResult {
 /// Compares two floats, with errors and non-numerics assumed to be -inf.
 /// Stops coercing at the first non-numeric char.
 /// We explicitly need to convert to f64 in this case.
-fn general_numeric_compare(a: GeneralF64ParseResult, b: GeneralF64ParseResult) -> Ordering {
+fn general_numeric_compare(a: &GeneralF64ParseResult, b: &GeneralF64ParseResult) -> Ordering {
-    a.partial_cmp(&b).unwrap()
+    a.partial_cmp(b).unwrap()
 }
 
 fn get_rand_string() -> String {
@@ -1646,6 +1668,12 @@ mod tests {
 
     use super::*;
 
+    fn tokenize_helper(line: &str, separator: Option<char>) -> Vec<Field> {
+        let mut buffer = vec![];
+        tokenize(line, separator, &mut buffer);
+        buffer
+    }
+
     #[test]
     fn test_get_hash() {
         let a = "Ted".to_string();
@@ -1689,20 +1717,23 @@ mod tests {
     #[test]
     fn test_tokenize_fields() {
         let line = "foo bar b    x";
-        assert_eq!(tokenize(line, None), vec![0..3, 3..7, 7..9, 9..14,],);
+        assert_eq!(tokenize_helper(line, None), vec![0..3, 3..7, 7..9, 9..14,],);
     }
 
     #[test]
     fn test_tokenize_fields_leading_whitespace() {
         let line = "    foo bar b    x";
-        assert_eq!(tokenize(line, None), vec![0..7, 7..11, 11..13, 13..18,]);
+        assert_eq!(
+            tokenize_helper(line, None),
+            vec![0..7, 7..11, 11..13, 13..18,]
+        );
     }
 
     #[test]
     fn test_tokenize_fields_custom_separator() {
         let line = "aaa foo bar b    x";
         assert_eq!(
-            tokenize(line, Some('a')),
+            tokenize_helper(line, Some('a')),
             vec![0..0, 1..1, 2..2, 3..9, 10..18,]
         );
     }
@@ -1710,11 +1741,11 @@ mod tests {
     #[test]
     fn test_tokenize_fields_trailing_custom_separator() {
         let line = "a";
-        assert_eq!(tokenize(line, Some('a')), vec![0..0]);
+        assert_eq!(tokenize_helper(line, Some('a')), vec![0..0]);
         let line = "aa";
-        assert_eq!(tokenize(line, Some('a')), vec![0..0, 1..1]);
+        assert_eq!(tokenize_helper(line, Some('a')), vec![0..0, 1..1]);
         let line = "..a..a";
-        assert_eq!(tokenize(line, Some('a')), vec![0..2, 3..5]);
+        assert_eq!(tokenize_helper(line, Some('a')), vec![0..2, 3..5]);
     }
 
     #[test]
@@ -1722,13 +1753,7 @@ mod tests {
     fn test_line_size() {
         // We should make sure to not regress the size of the Line struct because
         // it is unconditional overhead for every line we sort.
-        assert_eq!(std::mem::size_of::<Line>(), 32);
+        assert_eq!(std::mem::size_of::<Line>(), 24);
-        // These are the fields of Line:
-        assert_eq!(std::mem::size_of::<&str>(), 16);
-        assert_eq!(std::mem::size_of::<Box<[Selection]>>(), 16);
-
-        // How big is a selection? Constant cost all lines pay when we need selections.
-        assert_eq!(std::mem::size_of::<Selection>(), 24);
     }
 
     #[test]