perf: hot-path throughput, locking, and allocation optimizations#5
Merged
Conversation
Route Screen::print's narrow/wide/spacer writes and promote_cluster_to_wide's spacer write through Cell::set_from instead of struct-literal assignment, so a repeatedly overwritten cell reuses its combining String allocation. Interleaved same-harness A/B on a combining-heavy overwrite workload: +45% (author) / +69% (independent rerun) median; end-to-end bench_throughput unchanged (<1%).
rebuild_row_instances now fills caller-owned in/out Vecs (RowInstanceBuffers) instead of returning fresh ones, so dirty-row rebuilds in rebuild_pane_cached reuse the PaneRenderCache row slots' capacity from the previous frame — same recycle pattern as FrameSnapshot::from_terminal_recycled.
focused_cursor_wants_blink ran on every blink-interval wake and took the focused pane's terminal lock just to read cursor visibility/style/ viewport. Capture that state in redraw() under the lock it already holds per visible pane (same no-lock-on-poll pattern as kitty_animation_flag) and let the timer gate read the cache; at worst one redraw stale, and occlusion/focus gates veto independently.
feed_terminal_batch held one Terminal lock across an entire drained batch (up to 1MiB, a few ms of parsing), starving the main thread's snapshot lock during bulk output. Feed each reader chunk (<=64KiB) under its own lock and release it with MutexGuard::unlock_fair so a waiting main thread actually gets the handoff — a plain drop lets the io thread barge straight back in, and parking_lot's fair unlock costs the same as a normal unlock when nobody is waiting. Batch-tail work (overview/sidebar/ipc publish, pending-writes drain) still runs once per batch. Throughput unchanged (interleaved A/B within noise: author -0.5%, verifier +1.3%).
Re-profiling bench_throughput showed mode lookups as the largest untouched cost: ModeState backed its (value, ansi) pairs with a HashSet, and Terminal::print_str queries autowrap/grapheme_clustering once per bulk-print run while execute_c0 queries linefeed_newline on every CR/LF — millions of SipHash keyings to look up a set that holds about a dozen entries in a real session (~8% of bench self-time). Compare the pairs directly in a Vec instead: insert stays dedup-guarded, reset uses retain, and the mouse-format exclusivity retain matches the old per-mode removes. Interleaved same-path A/B medians: +6.0% (author) / +11.9% (independent verifier) end-to-end; both runs clear the gate, absolute baseline varies with machine load. Known trade-off (tracked): a stream that sets thousands of distinct unknown private modes degrades lookups to O(n) where the HashSet stayed O(1); bounded by the u16 mode space, and a candidate follow-up is dedicated flags for known modes with a side table for the rest.
Ground-state printable runs were scanned twice: a per-byte closure walk to find the control/DEL boundary, then core::str::from_utf8 over the same range. Re-profiling put that loop at 22% of bench_throughput self-time. Replace it with scan_run: an 8-byte u64 SWAR scan (underflow trick for <0x20, xor trick for 0x7f) that finds the boundary and tracks whether the run is pure ASCII in one pass; an all-ASCII run then skips validation via from_utf8_unchecked (SAFETY: scan_run only reports ascii for runs it proved all <0x80). Non-ASCII runs keep the existing from_utf8 + valid_up_to recovery path unchanged. Verified +7.7% end-to-end (177.5 -> 191.2 MiB/s best-median; hub arbitration and independent verifier agree, author's initial +15% reading did not reproduce). Soundness of the unsafe: exhaustive 256-value-at-every-position sweep, adjacent-pair cross-lane sweep (256x256x7), and a 2M-case seeded fuzz against a naive oracle — zero mismatches — plus chunk-boundary sweeps for split UTF-8, overlong sequences, and lone C1 bytes.
The erase family (erase_line, erase_display, erase_chars, insert_blank_chars, delete_chars) blanked cells with *c = blank.clone(), dropping and reallocating each destination cell's combining String where Row::clear and the print paths already reuse it through Cell::set_from. Route all five through set_from — same seven fields written, blank's combining is always empty, so output bytes are identical; only the allocation behavior changes. Interleaved same-path A/B on bench_throughput (its TUI-style lines hit EL on ~20% of rows): +4.0% (author, 207.1 -> 215.4 MiB/s) / +3.5% (independent verifier, 191.5 -> 198.2) — relative gains agree, absolute baselines vary with machine load.
Externally-triggered redraws (focus change, cursor blink, a sibling pane's redraw in the same window) never checked DECSET 2026, so they could snapshot a terminal mid-update and present a torn frame — and the io thread's chunked locking (33621c3) made that window easy to hit. Each pane now holds the last snapshot captured while synchronized output is active and redraws from it instead of re-reading the terminal, gated by sync_output_snapshot_decision: a resize always forces a fresh read (an old-sized snapshot is worse than a tear), the hold expires after the same SYNCHRONIZED_OUTPUT_MAX_SUPPRESSION cap the io thread applies to redraw pacing (a runaway mode 2026 cannot freeze a pane), and the held snapshot is released as soon as sync ends so panes that used mode 2026 once do not retain a full-grid snapshot for life. Dirty bits are not consumed while reusing, so the first redraw after sync releases renders the complete final state. Known residual (documented at the decision fn): the first externally-triggered redraw of a sync block has no held snapshot yet and can still tear once; holding one continuously would charge every pane that never uses mode 2026 a permanent full-grid snapshot, so the fix narrows per-block tearing to at most one at its start instead.
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
1 participant
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.
Summary
Measurement-driven optimization pass over the pty → vt → grid → render hot path, plus a synchronized-output fix the locking work exposed. Every shipped change was gated on interleaved same-path A/B benchmarks (≥2% end-to-end on
bench_throughput) and independently re-verified; rejected attempts and their numbers are recorded in.agents/nexus.md.perf(grid): scalar print overwrites reuse combining-String buffers (+45–69% on a combining-heavy overwrite microbench)perf(render): dirty-row rebuilds reuse row instance buffer capacityperf(app): cursor-blink gate reads a cached state instead of locking the terminal on every wakeperf(app): io thread feeds the terminal per 64KiB chunk withunlock_fair(worst-case lock hold was a full 1MiB batch)perf(grid): terminal modes stored in a linear-scanVec(+6–12% e2e — per-print_str/C0 SipHash lookups eliminated)perf(vt): SWAR word scan + ASCII fast path for printable runs (+7.7% e2e; theunsafeis validated by a 2.5M-case adversarial fuzz against a naive oracle)perf(grid): erase family blanks cells in place viaCell::set_from(+3.5–4% e2e)fix(app): externally-triggered redraws no longer tear a pane mid-update under synchronized output (DECSET 2026)bench_throughput: ~169 → ~198–215 MiB/s across the campaign (absolute numbers vary with machine load; every step verified by interleaved A/B).Known trade-offs (documented in code and commit bodies)
Test plan
cargo test --workspaceon real hardware (openpty, loopback sockets, Metal GPU pipeline tests): green. Two noa-pty tests are flaky under full-workspace parallel load but pass in isolation; noa-pty is untouched by this diff (depends only on noa-core + portable-pty).