7 files changed, 936 insertions, 0 deletions
diff --git a/helix-vcs/src/diff.rs b/helix-vcs/src/diff.rs
new file mode 100644
index 00000000..b1acd1f2
--- /dev/null
+++ b/helix-vcs/src/diff.rs
@@ -0,0 +1,198 @@
+use std::ops::Range;
+use std::sync::Arc;
+
+use helix_core::Rope;
+use imara_diff::Algorithm;
+use parking_lot::{Mutex, MutexGuard};
+use tokio::sync::mpsc::{unbounded_channel, UnboundedSender};
+use tokio::sync::{Notify, OwnedRwLockReadGuard, RwLock};
+use tokio::task::JoinHandle;
+use tokio::time::Instant;
+
+use crate::diff::worker::DiffWorker;
+
+mod line_cache;
+mod worker;
+
+type RedrawHandle = (Arc<Notify>, Arc<RwLock<()>>);
+
+/// A rendering lock passed to the differ the prevents redraws from occurring
+struct RenderLock {
+    pub lock: OwnedRwLockReadGuard<()>,
+    pub timeout: Option<Instant>,
+}
+
+struct Event {
+    text: Rope,
+    is_base: bool,
+    render_lock: Option<RenderLock>,
+}
+
+#[derive(Clone, Debug)]
+pub struct DiffHandle {
+    channel: UnboundedSender<Event>,
+    render_lock: Arc<RwLock<()>>,
+    hunks: Arc<Mutex<Vec<Hunk>>>,
+    inverted: bool,
+}
+
+impl DiffHandle {
+    pub fn new(diff_base: Rope, doc: Rope, redraw_handle: RedrawHandle) -> DiffHandle {
+        DiffHandle::new_with_handle(diff_base, doc, redraw_handle).0
+    }
+
+    fn new_with_handle(
+        diff_base: Rope,
+        doc: Rope,
+        redraw_handle: RedrawHandle,
+    ) -> (DiffHandle, JoinHandle<()>) {
+        let (sender, receiver) = unbounded_channel();
+        let hunks: Arc<Mutex<Vec<Hunk>>> = Arc::default();
+        let worker = DiffWorker {
+            channel: receiver,
+            hunks: hunks.clone(),
+            new_hunks: Vec::default(),
+            redraw_notify: redraw_handle.0,
+            diff_finished_notify: Arc::default(),
+        };
+        let handle = tokio::spawn(worker.run(diff_base, doc));
+        let differ = DiffHandle {
+            channel: sender,
+            hunks,
+            inverted: false,
+            render_lock: redraw_handle.1,
+        };
+        (differ, handle)
+    }
+
+    pub fn invert(&mut self) {
+        self.inverted = !self.inverted;
+    }
+
+    pub fn hunks(&self) -> FileHunks {
+        FileHunks {
+            hunks: self.hunks.lock(),
+            inverted: self.inverted,
+        }
+    }
+
+    /// Updates the document associated with this redraw handle
+    /// This function is only intended to be called from within the rendering loop
+    /// if called from elsewhere it may fail to acquire the render lock and panic
+    pub fn update_document(&self, doc: Rope, block: bool) -> bool {
+        // unwrap is ok here because the rendering lock is
+        // only exclusively locked during redraw.
+        // This function is only intended to be called
+        // from the core rendering loop where no redraw can happen in parallel
+        let lock = self.render_lock.clone().try_read_owned().unwrap();
+        let timeout = if block {
+            None
+        } else {
+            Some(Instant::now() + tokio::time::Duration::from_millis(SYNC_DIFF_TIMEOUT))
+        };
+        self.update_document_impl(doc, self.inverted, Some(RenderLock { lock, timeout }))
+    }
+
+    pub fn update_diff_base(&self, diff_base: Rope) -> bool {
+        self.update_document_impl(diff_base, !self.inverted, None)
+    }
+
+    fn update_document_impl(
+        &self,
+        text: Rope,
+        is_base: bool,
+        render_lock: Option<RenderLock>,
+    ) -> bool {
+        let event = Event {
+            text,
+            is_base,
+            render_lock,
+        };
+        self.channel.send(event).is_ok()
+    }
+}
+
+/// synchronous debounce value should be low
+/// so we can update synchronously most of the time
+const DIFF_DEBOUNCE_TIME_SYNC: u64 = 1;
+/// maximum time that rendering should be blocked until the diff finishes
+const SYNC_DIFF_TIMEOUT: u64 = 12;
+const DIFF_DEBOUNCE_TIME_ASYNC: u64 = 96;
+const ALGORITHM: Algorithm = Algorithm::Histogram;
+const MAX_DIFF_LINES: usize = 64 * u16::MAX as usize;
+// cap average line length to 128 for files with MAX_DIFF_LINES
+const MAX_DIFF_BYTES: usize = MAX_DIFF_LINES * 128;
+
+/// A single change in a file potentially spanning multiple lines
+/// Hunks produced by the differs are always ordered by their position
+/// in the file and non-overlapping.
+/// Specifically for any two hunks `x` and `y` the following properties hold:
+///
+/// ``` no_compile
+/// assert!(x.before.end <= y.before.start);
+/// assert!(x.after.end <= y.after.start);
+/// ```
+#[derive(PartialEq, Eq, Clone, Debug)]
+pub struct Hunk {
+    pub before: Range<u32>,
+    pub after: Range<u32>,
+}
+
+impl Hunk {
+    /// Can be used instead of `Option::None` for better performance
+    /// because lines larger then `i32::MAX` are not supported by `imara-diff` anyways.
+    /// Has some nice properties where it usually is not necessary to check for `None` separately:
+    /// Empty ranges fail contains checks and also fails smaller then checks.
+    pub const NONE: Hunk = Hunk {
+        before: u32::MAX..u32::MAX,
+        after: u32::MAX..u32::MAX,
+    };
+
+    /// Inverts a change so that `before`
+    pub fn invert(&self) -> Hunk {
+        Hunk {
+            before: self.after.clone(),
+            after: self.before.clone(),
+        }
+    }
+
+    pub fn is_pure_insertion(&self) -> bool {
+        self.before.is_empty()
+    }
+
+    pub fn is_pure_removal(&self) -> bool {
+        self.after.is_empty()
+    }
+}
+
+/// A list of changes in a file sorted in ascending
+/// non-overlapping order
+#[derive(Debug)]
+pub struct FileHunks<'a> {
+    hunks: MutexGuard<'a, Vec<Hunk>>,
+    inverted: bool,
+}
+
+impl FileHunks<'_> {
+    pub fn is_inverted(&self) -> bool {
+        self.inverted
+    }
+
+    /// Returns the `Hunk` for the `n`th change in this file.
+    /// if there is no `n`th change  `Hunk::NONE` is returned instead.
+    pub fn nth_hunk(&self, n: u32) -> Hunk {
+        match self.hunks.get(n as usize) {
+            Some(hunk) if self.inverted => hunk.invert(),
+            Some(hunk) => hunk.clone(),
+            None => Hunk::NONE,
+        }
+    }
+
+    pub fn len(&self) -> u32 {
+        self.hunks.len() as u32
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+}
diff --git a/helix-vcs/src/diff/line_cache.rs b/helix-vcs/src/diff/line_cache.rs
new file mode 100644
index 00000000..c3ee5daa
--- /dev/null
+++ b/helix-vcs/src/diff/line_cache.rs
@@ -0,0 +1,130 @@
+//! This modules encapsulates a tiny bit of unsafe code that
+//! makes diffing significantly faster and more ergonomic to implement.
+//! This code is necessary because diffing requires quick random
+//! access to the lines of the text that is being diffed.
+//!
+//! Therefore it is best to collect the `Rope::lines` iterator into a vec
+//! first because access to the vec is `O(1)` where `Rope::line` is `O(log N)`.
+//! However this process can allocate a (potentially quite large) vector.
+//!
+//! To avoid reallocation for every diff, the vector is reused.
+//! However the RopeSlice references the original rope and therefore forms a self-referential data structure.
+//! A transmute is used to change the lifetime of the slice to static to circumvent that project.
+use std::mem::transmute;
+
+use helix_core::{Rope, RopeSlice};
+use imara_diff::intern::{InternedInput, Interner};
+
+use super::{MAX_DIFF_BYTES, MAX_DIFF_LINES};
+
+/// A cache that stores the `lines` of a rope as a vector.
+/// It allows safely reusing the allocation of the vec when updating the rope
+pub(crate) struct InternedRopeLines {
+    diff_base: Rope,
+    doc: Rope,
+    num_tokens_diff_base: u32,
+    interned: InternedInput<RopeSlice<'static>>,
+}
+
+impl InternedRopeLines {
+    pub fn new(diff_base: Rope, doc: Rope) -> InternedRopeLines {
+        let mut res = InternedRopeLines {
+            interned: InternedInput {
+                before: Vec::with_capacity(diff_base.len_lines()),
+                after: Vec::with_capacity(doc.len_lines()),
+                interner: Interner::new(diff_base.len_lines() + doc.len_lines()),
+            },
+            diff_base,
+            doc,
+            // will be populated by update_diff_base_impl
+            num_tokens_diff_base: 0,
+        };
+        res.update_diff_base_impl();
+        res
+    }
+
+    /// Updates the `diff_base` and optionally the document if `doc` is not None
+    pub fn update_diff_base(&mut self, diff_base: Rope, doc: Option<Rope>) {
+        self.interned.clear();
+        self.diff_base = diff_base;
+        if let Some(doc) = doc {
+            self.doc = doc
+        }
+        if !self.is_too_large() {
+            self.update_diff_base_impl();
+        }
+    }
+
+    /// Updates the `doc` without reinterning the `diff_base`, this function
+    /// is therefore significantly faster than `update_diff_base` when only the document changes.
+    pub fn update_doc(&mut self, doc: Rope) {
+        // Safety: we clear any tokens that were added after
+        // the interning of `self.diff_base` finished so
+        // all lines that refer to `self.doc` have been purged.
+
+        self.interned
+            .interner
+            .erase_tokens_after(self.num_tokens_diff_base.into());
+
+        self.doc = doc;
+        if self.is_too_large() {
+            self.interned.after.clear();
+        } else {
+            self.update_doc_impl();
+        }
+    }
+
+    fn update_diff_base_impl(&mut self) {
+        // Safety: This transmute is safe because it only transmutes a lifetime, which has no effect.
+        // The backing storage for the RopeSlices referred to by the lifetime is stored in `self.diff_base`.
+        // Therefore as long as `self.diff_base` is not dropped/replaced this memory remains valid.
+        // `self.diff_base` is only changed in `self.update_diff_base`, which clears the interner.
+        // When the interned lines are exposed to consumer in `self.diff_input`, the lifetime is bounded to a reference to self.
+        // That means that on calls to update there exist no references to `self.interned`.
+        let before = self
+            .diff_base
+            .lines()
+            .map(|line: RopeSlice| -> RopeSlice<'static> { unsafe { transmute(line) } });
+        self.interned.update_before(before);
+        self.num_tokens_diff_base = self.interned.interner.num_tokens();
+        // the has to be interned again because the interner was fully cleared
+        self.update_doc_impl()
+    }
+
+    fn update_doc_impl(&mut self) {
+        // Safety: This transmute is save because it only transmutes a lifetime, which has no effect.
+        // The backing storage for the RopeSlices referred to by the lifetime is stored in `self.doc`.
+        // Therefore as long as `self.doc` is not dropped/replaced this memory remains valid.
+        // `self.doc` is only changed in `self.update_doc`, which clears the interner.
+        // When the interned lines are exposed to consumer in `self.diff_input`, the lifetime is bounded to a reference to self.
+        // That means that on calls to update there exist no references to `self.interned`.
+        let after = self
+            .doc
+            .lines()
+            .map(|line: RopeSlice| -> RopeSlice<'static> { unsafe { transmute(line) } });
+        self.interned.update_after(after);
+    }
+
+    fn is_too_large(&self) -> bool {
+        // bound both lines and bytes to avoid huge files with few (but huge) lines
+        // or huge file with tiny lines. While this makes no difference to
+        // diff itself (the diff performance only depends on the number of tokens)
+        // the interning runtime depends mostly on filesize and is actually dominant
+        // for large files
+        self.doc.len_lines() > MAX_DIFF_LINES
+            || self.diff_base.len_lines() > MAX_DIFF_LINES
+            || self.doc.len_bytes() > MAX_DIFF_BYTES
+            || self.diff_base.len_bytes() > MAX_DIFF_BYTES
+    }
+
+    /// Returns the `InternedInput` for performing the diff.
+    /// If `diff_base` or `doc` is so large that performing a diff could slow the editor
+    /// this function returns `None`.
+    pub fn interned_lines(&self) -> Option<&InternedInput<RopeSlice>> {
+        if self.is_too_large() {
+            None
+        } else {
+            Some(&self.interned)
+        }
+    }
+}
diff --git a/helix-vcs/src/diff/worker.rs b/helix-vcs/src/diff/worker.rs
new file mode 100644
index 00000000..b8659c9b
--- /dev/null
+++ b/helix-vcs/src/diff/worker.rs
@@ -0,0 +1,207 @@
+use std::mem::swap;
+use std::ops::Range;
+use std::sync::Arc;
+
+use helix_core::{Rope, RopeSlice};
+use imara_diff::intern::InternedInput;
+use parking_lot::Mutex;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio::sync::Notify;
+use tokio::time::{timeout, timeout_at, Duration};
+
+use crate::diff::{
+    Event, RenderLock, ALGORITHM, DIFF_DEBOUNCE_TIME_ASYNC, DIFF_DEBOUNCE_TIME_SYNC,
+};
+
+use super::line_cache::InternedRopeLines;
+use super::Hunk;
+
+#[cfg(test)]
+mod test;
+
+pub(super) struct DiffWorker {
+    pub channel: UnboundedReceiver<Event>,
+    pub hunks: Arc<Mutex<Vec<Hunk>>>,
+    pub new_hunks: Vec<Hunk>,
+    pub redraw_notify: Arc<Notify>,
+    pub diff_finished_notify: Arc<Notify>,
+}
+
+impl DiffWorker {
+    async fn accumulate_events(&mut self, event: Event) -> (Option<Rope>, Option<Rope>) {
+        let mut accumulator = EventAccumulator::new();
+        accumulator.handle_event(event).await;
+        accumulator
+            .accumulate_debounced_events(
+                &mut self.channel,
+                self.redraw_notify.clone(),
+                self.diff_finished_notify.clone(),
+            )
+            .await;
+        (accumulator.doc, accumulator.diff_base)
+    }
+
+    pub async fn run(mut self, diff_base: Rope, doc: Rope) {
+        let mut interner = InternedRopeLines::new(diff_base, doc);
+        if let Some(lines) = interner.interned_lines() {
+            self.perform_diff(lines);
+        }
+        self.apply_hunks();
+        while let Some(event) = self.channel.recv().await {
+            let (doc, diff_base) = self.accumulate_events(event).await;
+
+            let process_accumulated_events = || {
+                if let Some(new_base) = diff_base {
+                    interner.update_diff_base(new_base, doc)
+                } else {
+                    interner.update_doc(doc.unwrap())
+                }
+
+                if let Some(lines) = interner.interned_lines() {
+                    self.perform_diff(lines)
+                }
+            };
+
+            // Calculating diffs is computationally expensive and should
+            // not run inside an async function to avoid blocking other futures.
+            // Note: tokio::task::block_in_place does not work during tests
+            #[cfg(test)]
+            process_accumulated_events();
+            #[cfg(not(test))]
+            tokio::task::block_in_place(process_accumulated_events);
+
+            self.apply_hunks();
+        }
+    }
+
+    /// update the hunks (used by the gutter) by replacing it with `self.new_hunks`.
+    /// `self.new_hunks` is always empty after this function runs.
+    /// To improve performance this function tries to reuse the allocation of the old diff previously stored in `self.line_diffs`
+    fn apply_hunks(&mut self) {
+        swap(&mut *self.hunks.lock(), &mut self.new_hunks);
+        self.diff_finished_notify.notify_waiters();
+        self.new_hunks.clear();
+    }
+
+    fn perform_diff(&mut self, input: &InternedInput<RopeSlice>) {
+        imara_diff::diff(ALGORITHM, input, |before: Range<u32>, after: Range<u32>| {
+            self.new_hunks.push(Hunk { before, after })
+        })
+    }
+}
+
+struct EventAccumulator {
+    diff_base: Option<Rope>,
+    doc: Option<Rope>,
+    render_lock: Option<RenderLock>,
+}
+
+impl<'a> EventAccumulator {
+    fn new() -> EventAccumulator {
+        EventAccumulator {
+            diff_base: None,
+            doc: None,
+            render_lock: None,
+        }
+    }
+
+    async fn handle_event(&mut self, event: Event) {
+        let dst = if event.is_base {
+            &mut self.diff_base
+        } else {
+            &mut self.doc
+        };
+
+        *dst = Some(event.text);
+
+        // always prefer the most synchronous requested render mode
+        if let Some(render_lock) = event.render_lock {
+            match &mut self.render_lock {
+                Some(RenderLock { timeout, .. }) => {
+                    // A timeout of `None` means that the render should
+                    // always wait for the diff to complete (so no timeout)
+                    // remove the existing timeout, otherwise keep the previous timeout
+                    // because it will be shorter then the current timeout
+                    if render_lock.timeout.is_none() {
+                        timeout.take();
+                    }
+                }
+                None => self.render_lock = Some(render_lock),
+            }
+        }
+    }
+
+    async fn accumulate_debounced_events(
+        &mut self,
+        channel: &mut UnboundedReceiver<Event>,
+        redraw_notify: Arc<Notify>,
+        diff_finished_notify: Arc<Notify>,
+    ) {
+        let async_debounce = Duration::from_millis(DIFF_DEBOUNCE_TIME_ASYNC);
+        let sync_debounce = Duration::from_millis(DIFF_DEBOUNCE_TIME_SYNC);
+        loop {
+            // if we are not blocking rendering use a much longer timeout
+            let debounce = if self.render_lock.is_none() {
+                async_debounce
+            } else {
+                sync_debounce
+            };
+
+            if let Ok(Some(event)) = timeout(debounce, channel.recv()).await {
+                self.handle_event(event).await;
+            } else {
+                break;
+            }
+        }
+
+        // setup task to trigger the rendering
+        match self.render_lock.take() {
+            // diff is performed outside of the rendering loop
+            // request a redraw after the diff is done
+            None => {
+                tokio::spawn(async move {
+                    diff_finished_notify.notified().await;
+                    redraw_notify.notify_one();
+                });
+            }
+            // diff is performed inside the rendering loop
+            // block redraw until the diff is done or the timeout is expired
+            Some(RenderLock {
+                lock,
+                timeout: Some(timeout),
+            }) => {
+                tokio::spawn(async move {
+                    let res = {
+                        // Acquire a lock on the redraw handle.
+                        // The lock will block the rendering from occurring while held.
+                        // The rendering waits for the diff if it doesn't time out
+                        timeout_at(timeout, diff_finished_notify.notified()).await
+                    };
+                    // we either reached the timeout or the diff is finished, release the render lock
+                    drop(lock);
+                    if res.is_ok() {
+                        // Diff finished in time we are done.
+                        return;
+                    }
+                    // Diff failed to complete in time log the event
+                    // and wait until the diff occurs to trigger an async redraw
+                    log::warn!("Diff computation timed out, update of diffs might appear delayed");
+                    diff_finished_notify.notified().await;
+                    redraw_notify.notify_one();
+                });
+            }
+            // a blocking diff is performed inside the rendering loop
+            // block redraw until the diff is done
+            Some(RenderLock {
+                lock,
+                timeout: None,
+            }) => {
+                tokio::spawn(async move {
+                    diff_finished_notify.notified().await;
+                    // diff is done release the lock
+                    drop(lock)
+                });
+            }
+        };
+    }
+}
diff --git a/helix-vcs/src/diff/worker/test.rs b/helix-vcs/src/diff/worker/test.rs
new file mode 100644
index 00000000..14442426
--- /dev/null
+++ b/helix-vcs/src/diff/worker/test.rs
@@ -0,0 +1,149 @@
+use helix_core::Rope;
+use tokio::task::JoinHandle;
+
+use crate::diff::{DiffHandle, Hunk};
+
+impl DiffHandle {
+    fn new_test(diff_base: &str, doc: &str) -> (DiffHandle, JoinHandle<()>) {
+        DiffHandle::new_with_handle(
+            Rope::from_str(diff_base),
+            Rope::from_str(doc),
+            Default::default(),
+        )
+    }
+    async fn into_diff(self, handle: JoinHandle<()>) -> Vec<Hunk> {
+        let hunks = self.hunks;
+        // dropping the channel terminates the task
+        drop(self.channel);
+        handle.await.unwrap();
+        let hunks = hunks.lock();
+        Vec::clone(&*hunks)
+    }
+}
+
+#[tokio::test]
+async fn append_line() {
+    let (differ, handle) = DiffHandle::new_test("foo\n", "foo\nbar\n");
+    let line_diffs = differ.into_diff(handle).await;
+    assert_eq!(
+        &line_diffs,
+        &[Hunk {
+            before: 1..1,
+            after: 1..2
+        }]
+    )
+}
+
+#[tokio::test]
+async fn prepend_line() {
+    let (differ, handle) = DiffHandle::new_test("foo\n", "bar\nfoo\n");
+    let line_diffs = differ.into_diff(handle).await;
+    assert_eq!(
+        &line_diffs,
+        &[Hunk {
+            before: 0..0,
+            after: 0..1
+        }]
+    )
+}
+
+#[tokio::test]
+async fn modify() {
+    let (differ, handle) = DiffHandle::new_test("foo\nbar\n", "foo bar\nbar\n");
+    let line_diffs = differ.into_diff(handle).await;
+    assert_eq!(
+        &line_diffs,
+        &[Hunk {
+            before: 0..1,
+            after: 0..1
+        }]
+    )
+}
+
+#[tokio::test]
+async fn delete_line() {
+    let (differ, handle) = DiffHandle::new_test("foo\nfoo bar\nbar\n", "foo\nbar\n");
+    let line_diffs = differ.into_diff(handle).await;
+    assert_eq!(
+        &line_diffs,
+        &[Hunk {
+            before: 1..2,
+            after: 1..1
+        }]
+    )
+}
+
+#[tokio::test]
+async fn delete_line_and_modify() {
+    let (differ, handle) = DiffHandle::new_test("foo\nbar\ntest\nfoo", "foo\ntest\nfoo bar");
+    let line_diffs = differ.into_diff(handle).await;
+    assert_eq!(
+        &line_diffs,
+        &[
+            Hunk {
+                before: 1..2,
+                after: 1..1
+            },
+            Hunk {
+                before: 3..4,
+                after: 2..3
+            },
+        ]
+    )
+}
+
+#[tokio::test]
+async fn add_use() {
+    let (differ, handle) = DiffHandle::new_test(
+        "use ropey::Rope;\nuse tokio::task::JoinHandle;\n",
+        "use ropey::Rope;\nuse ropey::RopeSlice;\nuse tokio::task::JoinHandle;\n",
+    );
+    let line_diffs = differ.into_diff(handle).await;
+    assert_eq!(
+        &line_diffs,
+        &[Hunk {
+            before: 1..1,
+            after: 1..2
+        },]
+    )
+}
+
+#[tokio::test]
+async fn update_document() {
+    let (differ, handle) = DiffHandle::new_test("foo\nbar\ntest\nfoo", "foo\nbar\ntest\nfoo");
+    differ.update_document(Rope::from_str("foo\ntest\nfoo bar"), false);
+    let line_diffs = differ.into_diff(handle).await;
+    assert_eq!(
+        &line_diffs,
+        &[
+            Hunk {
+                before: 1..2,
+                after: 1..1
+            },
+            Hunk {
+                before: 3..4,
+                after: 2..3
+            },
+        ]
+    )
+}
+
+#[tokio::test]
+async fn update_base() {
+    let (differ, handle) = DiffHandle::new_test("foo\ntest\nfoo bar", "foo\ntest\nfoo bar");
+    differ.update_diff_base(Rope::from_str("foo\nbar\ntest\nfoo"));
+    let line_diffs = differ.into_diff(handle).await;
+    assert_eq!(
+        &line_diffs,
+        &[
+            Hunk {
+                before: 1..2,
+                after: 1..1
+            },
+            Hunk {
+                before: 3..4,
+                after: 2..3
+            },
+        ]
+    )
+}
diff --git a/helix-vcs/src/git.rs b/helix-vcs/src/git.rs
new file mode 100644
index 00000000..82b2b558
--- /dev/null
+++ b/helix-vcs/src/git.rs
@@ -0,0 +1,80 @@
+use std::path::Path;
+
+use git::objs::tree::EntryMode;
+use git::sec::trust::DefaultForLevel;
+use git::{Commit, ObjectId, Repository, ThreadSafeRepository};
+use git_repository as git;
+
+use crate::DiffProvider;
+
+#[cfg(test)]
+mod test;
+
+pub struct Git;
+
+impl Git {
+    fn open_repo(path: &Path, ceiling_dir: Option<&Path>) -> Option<ThreadSafeRepository> {
+        // custom open options
+        let mut git_open_opts_map = git::sec::trust::Mapping::<git::open::Options>::default();
+
+        // don't use the global git configs (not needed)
+        let config = git::permissions::Config {
+            system: false,
+            git: false,
+            user: false,
+            env: true,
+            includes: true,
+            git_binary: false,
+        };
+        // change options for config permissions without touching anything else
+        git_open_opts_map.reduced = git_open_opts_map.reduced.permissions(git::Permissions {
+            config,
+            ..git::Permissions::default_for_level(git::sec::Trust::Reduced)
+        });
+        git_open_opts_map.full = git_open_opts_map.full.permissions(git::Permissions {
+            config,
+            ..git::Permissions::default_for_level(git::sec::Trust::Full)
+        });
+
+        let mut open_options = git::discover::upwards::Options::default();
+        if let Some(ceiling_dir) = ceiling_dir {
+            open_options.ceiling_dirs = vec![ceiling_dir.to_owned()];
+        }
+
+        ThreadSafeRepository::discover_with_environment_overrides_opts(
+            path,
+            open_options,
+            git_open_opts_map,
+        )
+        .ok()
+    }
+}
+
+impl DiffProvider for Git {
+    fn get_diff_base(&self, file: &Path) -> Option<Vec<u8>> {
+        debug_assert!(!file.exists() || file.is_file());
+        debug_assert!(file.is_absolute());
+
+        // TODO cache repository lookup
+        let repo = Git::open_repo(file.parent()?, None)?.to_thread_local();
+        let head = repo.head_commit().ok()?;
+        let file_oid = find_file_in_commit(&repo, &head, file)?;
+
+        let file_object = repo.find_object(file_oid).ok()?;
+        Some(file_object.detach().data)
+    }
+}
+
+/// Finds the object that contains the contents of a file at a specific commit.
+fn find_file_in_commit(repo: &Repository, commit: &Commit, file: &Path) -> Option<ObjectId> {
+    let repo_dir = repo.work_dir()?;
+    let rel_path = file.strip_prefix(repo_dir).ok()?;
+    let tree = commit.tree().ok()?;
+    let tree_entry = tree.lookup_entry_by_path(rel_path).ok()??;
+    match tree_entry.mode() {
+        // not a file, everything is new, do not show diff
+        EntryMode::Tree | EntryMode::Commit | EntryMode::Link => None,
+        // found a file
+        EntryMode::Blob | EntryMode::BlobExecutable => Some(tree_entry.object_id()),
+    }
+}
diff --git a/helix-vcs/src/git/test.rs b/helix-vcs/src/git/test.rs
new file mode 100644
index 00000000..d6e9af08
--- /dev/null
+++ b/helix-vcs/src/git/test.rs
@@ -0,0 +1,121 @@
+use std::{fs::File, io::Write, path::Path, process::Command};
+
+use tempfile::TempDir;
+
+use crate::{DiffProvider, Git};
+
+fn exec_git_cmd(args: &str, git_dir: &Path) {
+    let res = Command::new("git")
+        .arg("-C")
+        .arg(git_dir) // execute the git command in this directory
+        .args(args.split_whitespace())
+        .env_remove("GIT_DIR")
+        .env_remove("GIT_ASKPASS")
+        .env_remove("SSH_ASKPASS")
+        .env("GIT_TERMINAL_PROMPT", "false")
+        .env("GIT_AUTHOR_DATE", "2000-01-01 00:00:00 +0000")
+        .env("GIT_AUTHOR_EMAIL", "author@example.com")
+        .env("GIT_AUTHOR_NAME", "author")
+        .env("GIT_COMMITTER_DATE", "2000-01-02 00:00:00 +0000")
+        .env("GIT_COMMITTER_EMAIL", "committer@example.com")
+        .env("GIT_COMMITTER_NAME", "committer")
+        .env("GIT_CONFIG_COUNT", "2")
+        .env("GIT_CONFIG_KEY_0", "commit.gpgsign")
+        .env("GIT_CONFIG_VALUE_0", "false")
+        .env("GIT_CONFIG_KEY_1", "init.defaultBranch")
+        .env("GIT_CONFIG_VALUE_1", "main")
+        .output()
+        .unwrap_or_else(|_| panic!("`git {args}` failed"));
+    if !res.status.success() {
+        println!("{}", String::from_utf8_lossy(&res.stdout));
+        eprintln!("{}", String::from_utf8_lossy(&res.stderr));
+        panic!("`git {args}` failed (see output above)")
+    }
+}
+
+fn create_commit(repo: &Path, add_modified: bool) {
+    if add_modified {
+        exec_git_cmd("add -A", repo);
+    }
+    exec_git_cmd("commit -m message", repo);
+}
+
+fn empty_git_repo() -> TempDir {
+    let tmp = tempfile::tempdir().expect("create temp dir for git testing");
+    exec_git_cmd("init", tmp.path());
+    exec_git_cmd("config user.email test@helix.org", tmp.path());
+    exec_git_cmd("config user.name helix-test", tmp.path());
+    tmp
+}
+
+#[test]
+fn missing_file() {
+    let temp_git = empty_git_repo();
+    let file = temp_git.path().join("file.txt");
+    File::create(&file).unwrap().write_all(b"foo").unwrap();
+
+    assert_eq!(Git.get_diff_base(&file), None);
+}
+
+#[test]
+fn unmodified_file() {
+    let temp_git = empty_git_repo();
+    let file = temp_git.path().join("file.txt");
+    let contents = b"foo".as_slice();
+    File::create(&file).unwrap().write_all(contents).unwrap();
+    create_commit(temp_git.path(), true);
+    assert_eq!(Git.get_diff_base(&file), Some(Vec::from(contents)));
+}
+
+#[test]
+fn modified_file() {
+    let temp_git = empty_git_repo();
+    let file = temp_git.path().join("file.txt");
+    let contents = b"foo".as_slice();
+    File::create(&file).unwrap().write_all(contents).unwrap();
+    create_commit(temp_git.path(), true);
+    File::create(&file).unwrap().write_all(b"bar").unwrap();
+
+    assert_eq!(Git.get_diff_base(&file), Some(Vec::from(contents)));
+}
+
+/// Test that `get_file_head` does not return content for a directory.
+/// This is important to correctly cover cases where a directory is removed and replaced by a file.
+/// If the contents of the directory object were returned a diff between a path and the directory children would be produced.
+#[test]
+fn directory() {
+    let temp_git = empty_git_repo();
+    let dir = temp_git.path().join("file.txt");
+    std::fs::create_dir(&dir).expect("");
+    let file = dir.join("file.txt");
+    let contents = b"foo".as_slice();
+    File::create(&file).unwrap().write_all(contents).unwrap();
+
+    create_commit(temp_git.path(), true);
+
+    std::fs::remove_dir_all(&dir).unwrap();
+    File::create(&dir).unwrap().write_all(b"bar").unwrap();
+    assert_eq!(Git.get_diff_base(&dir), None);
+}
+
+/// Test that `get_file_head` does not return content for a symlink.
+/// This is important to correctly cover cases where a symlink is removed and replaced by a file.
+/// If the contents of the symlink object were returned a diff between a path and the actual file would be produced (bad ui).
+#[cfg(any(unix, windows))]
+#[test]
+fn symlink() {
+    #[cfg(unix)]
+    use std::os::unix::fs::symlink;
+    #[cfg(not(unix))]
+    use std::os::windows::fs::symlink_file as symlink;
+    let temp_git = empty_git_repo();
+    let file = temp_git.path().join("file.txt");
+    let contents = b"foo".as_slice();
+    File::create(&file).unwrap().write_all(contents).unwrap();
+    let file_link = temp_git.path().join("file_link.txt");
+    symlink("file.txt", &file_link).unwrap();
+
+    create_commit(temp_git.path(), true);
+    assert_eq!(Git.get_diff_base(&file_link), None);
+    assert_eq!(Git.get_diff_base(&file), Some(Vec::from(contents)));
+}
diff --git a/helix-vcs/src/lib.rs b/helix-vcs/src/lib.rs
new file mode 100644
index 00000000..97320d32
--- /dev/null
+++ b/helix-vcs/src/lib.rs
@@ -0,0 +1,51 @@
+use std::path::Path;
+
+#[cfg(feature = "git")]
+pub use git::Git;
+#[cfg(not(feature = "git"))]
+pub use Dummy as Git;
+
+#[cfg(feature = "git")]
+mod git;
+
+mod diff;
+
+pub use diff::{DiffHandle, Hunk};
+
+pub trait DiffProvider {
+    /// Returns the data that a diff should be computed against
+    /// if this provider is used.
+    /// The data is returned as raw byte without any decoding or encoding performed
+    /// to ensure all file encodings are handled correctly.
+    fn get_diff_base(&self, file: &Path) -> Option<Vec<u8>>;
+}
+
+#[doc(hidden)]
+pub struct Dummy;
+impl DiffProvider for Dummy {
+    fn get_diff_base(&self, _file: &Path) -> Option<Vec<u8>> {
+        None
+    }
+}
+
+pub struct DiffProviderRegistry {
+    providers: Vec<Box<dyn DiffProvider>>,
+}
+
+impl DiffProviderRegistry {
+    pub fn get_diff_base(&self, file: &Path) -> Option<Vec<u8>> {
+        self.providers
+            .iter()
+            .find_map(|provider| provider.get_diff_base(file))
+    }
+}
+
+impl Default for DiffProviderRegistry {
+    fn default() -> Self {
+        // currently only git is supported
+        // TODO make this configurable when more providers are added
+        let git: Box<dyn DiffProvider> = Box::new(Git);
+        let providers = vec![git];
+        DiffProviderRegistry { providers }
+    }
+}