perf: cached-window render fast path for wrapped-line navigation (stacked on #2784) - #2813
Open
sinelaw wants to merge 2 commits into
Open
perf: cached-window render fast path for wrapped-line navigation (stacked on #2784)#2813sinelaw wants to merge 2 commits into
sinelaw wants to merge 2 commits into
Conversation
…on bump Navigating a very long soft-wrapped line cost hundreds of milliseconds per keypress. This lands the algorithmic fixes (a follow-up PR adds a render-side cache on top): - Grapheme boundary helpers (next/prev_grapheme_boundary, grapheme_at, snap_to_grapheme_boundary) iterated grapheme_indices from the start of the string on every call, making callers' walks quadratic — the reference-highlight word scan alone cost ~600 ms/keypress inside a giant word. Rewritten on GraphemeCursor, which resolves boundaries locally; prev_grapheme_boundary now also returns the containing cluster's start for a position inside a multi-byte code point instead of overshooting one cluster left. - indent_folding line-boundary scans (find_line_start_byte, find_line_end_byte, find_next_line_start_byte) read the buffer one byte at a time via slice_bytes; they now scan in 4 KiB blocks. These run per rendered frame for the cursor line. - extend_streaming appends content without bumping buffer.version(), leaving version-keyed layout caches (visual-row index) reachable with pre-append state. It now bumps the version (without marking the buffer dirty — content still matches the on-disk stream). Measured per keypress at 80x24 (release; the test harness renders twice per press): 200 KB single wrapped line ~340 ms -> ~140 ms, 50 KB ~168 ms -> ~37 ms, 10 KB ~120 ms -> ~10 ms; short lines unchanged (~2.8 ms). The follow-up cache PR takes the long-line cases to ~5-30 ms. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01J3uiv3CFZEAFARmTkqxpaQ
…igation Builds on the algorithmic fixes in the base PR. The renderer re-ran the tokenise -> wrap pipeline for every visible logical line on every frame — O(line length) per keypress inside a long soft-wrapped line. This adds a fast path that assembles the visible window straight from per-line entries in the line-wrap cache: - try_cached_window (view_data.rs): walks logical-line boundaries from the viewport top, row-bounded (stops once accumulated entries cover the scroll offset plus two screens), byte-bounded (never reads more than the pipeline would tokenise), validates each entry's completeness EXACTLY per line ending, and falls back to the full pipeline on any miss — whose writeback repopulates the cache for the next frame. Eligibility: no view transform, folds, virtual text, soft breaks, conceals, grid wrap, binary content, or pending scroll-to-end sync. - Cursor-independent cache keys: when no soft breaks / conceals exist the pipeline never reads cursor positions, so the writeback and layout_for_line key with cursor_sig 0 (layout_depends_on_cursors), keeping the cursor line's entry valid as the cursor moves within it. - tab_size joins LineWrapKey / VisualRowIndexKey / WrapGeometry: it is a layout input; without it a tab-size change served stale layouts. - Generation-aware eviction (begin_frame / get_touch): entries written or touched in the current frame are spared while the total stays under 2x the byte budget, so a huge line's entry and its window neighbours stop evicting each other every frame. Measured per keypress at 80x24 (release; the test harness renders twice per press), on top of the base PR's numbers: 200 KB single wrapped line ~140 ms -> ~25 ms, 50 KB ~37 ms -> ~13 ms, 10 KB ~10 ms -> ~5 ms; short lines unchanged. Tests: warm-vs-cold cache equivalence (LF and CRLF, including the trailing-EOF-row key), up/down round-trip stability, edit invalidation, black-box mid-file-vs-EOF navigation throughput, and the completeness-check contract test. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01J3uiv3CFZEAFARmTkqxpaQ
sinelaw
changed the title
sinelaw
force-pushed
the
claude/wrapped-line-nav-cache-wqwlh7
branch
from
July 28, 2026 18:38
e5ca213 to
c05d02c
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #2784 (algorithmic fixes). This PR adds the cache layer that takes long-wrapped-line navigation from sluggish to smooth. Merge #2784 first; this diff shows only the cache addition.
What
The renderer re-ran the tokenise-then-wrap pipeline for every visible logical line on every frame, which is O(line length) per keypress inside a long soft-wrapped line. This adds a fast path that assembles the visible window straight from per-line entries in the line-wrap cache:
try_cached_window(view_data.rs): walks logical-line boundaries from the viewport top, row-bounded (stops once accumulated entries cover the scroll offset plus two screens), byte-bounded (never reads more than the pipeline would tokenise), validates each entry for completeness exactly per line ending, and falls back to the full pipeline on any miss. The fallback writeback repopulates the cache so the next frame hits. Eligibility: no plugin view transform, folds, virtual text, soft breaks, conceals, terminal grid wrap, binary content, or pending scroll-to-end sync.layout_for_linekey entries withcursor_sig: 0(layout_depends_on_cursors). The cursor line entry stays valid as the cursor moves within it — previously every cursor move invalidated the very line being navigated.tab_sizejoins the cache keys (LineWrapKey,VisualRowIndexKey,WrapGeometry). Tab width is a layout input; without it a tab-size change served stale layouts.begin_frame,get_touch). Entries written or touched in the current frame are spared while the total stays under twice the byte budget, so a huge line entry and its window neighbours stop evicting each other every frame.Performance
Per keypress at 80x24 (release build; the test harness renders twice per press), on top of the base PR numbers:
Correctness
The design is fail-safe: any doubt (missing, stale, or truncated entry; ineligible feature) falls back to the full pipeline, so a wrong answer requires a keyed entry that is both reachable and wrong. An adversarial review of the cache-invalidation surface was run over this design; all confirmed findings are fixed and pinned by regression tests (two in this PR, two in #2784).
Tests: warm-vs-cold cache equivalence (LF and CRLF, including the trailing-EOF-row key), up/down round-trip stability, edit invalidation, black-box mid-file vs EOF navigation throughput, and the completeness-check contract test — plus the full lib, scroll, and wrap suites.
🤖 Generated with Claude Code
https://claude.ai/code/session_01J3uiv3CFZEAFARmTkqxpaQ