[kartikkabadi]

Summary

Moving the caret or scrolling a long teleprompter script repeatedly scans the full string, rebuilds the interaction layout during flex measurement, and enters glyph loops for offscreen paragraphs. This keeps the warm interaction path short while preserving existing text geometry and editing behavior.

 warm text-field interaction
- scan graphemes from a string endpoint
+ find the adjacent boundary from the caret with full Unicode context
- replace the active row layout during width measurement
+ retain the active layout and up to four alternate width measurements
- enter every paragraph's glyph paint loop
+ skip fully clipped unmarked paragraphs

One file, +487 / -72, including five focused regressions. Production code accounts for +192 / -70; tests account for +295 / -2. The cache tracks the complete inherited font, color, text, composition range, font size, line height and active wrap width. Deferred measurement captures the inherited font. Marked text keeps its existing painting path, and paragraph clipping preserves GPUI's fractional row coordinates.

Evidence

Fresh timing comparison against main ec15b9840344 on Apple M1 Pro, macOS 27.0.1, Rust 1.95.0. The exact committed suite was subsequently rebuilt at 900dd827b962. Last-checked main 100f43a0d359 added web/CLI-only changes; its GPUI workspace, Rust crates and build configuration are identical to the tested tip. Both measured text-input sources remain unchanged throughout. The title's rounded range covers the ~100KB key-dispatch/CPU-draw and scroll-redraw cases across ASCII, multilingual and emoji scripts.

For the 100KB ASCII fixture with the viewport near the middle:

Operation	Before per step	After per step	Median paired speedup	Across the three pairs
Left key + CPU redraw	11,453.125 µs	48.454 µs	237.73×	233.59–285.29×
Right key + CPU redraw	12,101.652 µs	53.878 µs	224.90×	213.59–272.20×
Shift-left key + CPU redraw	11,669.375 µs	61.855 µs	188.66×	179.98–257.02×
Shift-right key + CPU redraw	12,029.483 µs	61.584 µs	195.81×	191.12–993.73×
Scroll redraw	10,298.966 µs	44.803 µs	230.81×	229.87–359.78×

Ranges of case-level median paired speedups across the full text/position matrix:

Target script size	Key dispatch + CPU redraw	Scroll CPU redraw
10KB	9.97–21.47×	11.76–23.68×
100KB	105.33–237.73×	142.89–347.73×

These measure elapsed time for warm, synchronous CPU operations in GPUI TestPlatform/NoopTextSystem. Key measurements include binding dispatch and CPU scene preparation; scroll measurements include offset changes and CPU redraw. Native font shaping, OS input latency, GPU presentation, FPS, release-profile performance and export throughput remain unmeasured. The native checks below validate rendering correctness separately.

Immutable evidence snapshot includes all raw samples, fixture sources, compiler profiles, build/test logs, native images, analysis scripts, source hashes and a per-file checksum manifest. The two earlier audit archives are preserved there with their original manifests. Evidence files are outside this PR's code diff.

Benchmark method and reproduction

• Three alternating process pairs, seven samples per case and 120 operations per sample. All 2,772 raw samples / 66 cases independently reproduce the reported medians and ratios. Each process must produce the same fixture/viewport/caret case keys.
• ASCII, multilingual and emoji scripts targeting 10,000 and 100,000 bytes. Actual byte/grapheme counts are in the CSV. Navigation covers middle positions with left/right and Shift-left/Shift-right handlers and binding dispatch; scrolling covers start/middle/end positions.
• Window 560×320, scroll viewport 560×240, text width 496, font size 30, line height 45, centered text. A two-row scroll round trip and the navigation viewport are prepared outside the timed loop.
• Each process contributes the median of seven samples. Speedup is the median of the three paired before/after ratios. The displayed before/after times are medians of process medians; their quotient need not equal the median paired ratio. All per-pair ranges are retained.
• App opt-level 3, debug info disabled, debug assertions and overflow checks enabled. This is an optimized test build, with dependencies using the repository's development settings. All 1,018 compiler-artifact feature/profile records match between the two builds. The exact committed suite also passed at the app's normal opt-level 0.
• Shared 8-logical-CPU, 16GB host, without CPU-affinity or power-state control. No build or other audit test ran concurrently with these final measurements. Other host workloads were uncontrolled.
• Earlier minimization checks include a five-pair identical-binary control. That control showed apparent differences approaching 1.9× despite identical executables. The original-versus-minimized repeat showed no consistent regression attributable to minimization. Those records are retained; the headline numbers above come from the fresh current-main comparison.

Raw summary CSV, process metadata, independent analysis, compiler profile, source/patch integrity and reproduction notes.

After the repository's normal native setup, build two disposable checkouts at the recorded baseline and this PR's code commit. The evidence includes exact pure source snapshots and the external audit fixtures. Append the viewport fixture under cfg(test) to each source, and use:

CARGO_INCREMENTAL=0 CARGO_BUILD_JOBS=2 cargo +1.95.0 test \
  --manifest-path apps/desktop-gpui/Cargo.toml --locked --offline \
  --bin cap-gpui --no-run \
  --config profile.dev.package.cap-desktop-gpui.opt-level=3 \
  --config profile.dev.package.cap-desktop-gpui.debug=0 \
  --config 'profile.dev.package."*".debug=0' --message-format=json

python3 run-viewport-pairs.py --repo <checkout> \
  --baseline <baseline-test-binary> --candidate <candidate-test-binary> \
  --build-profile <candidate-profile.json> --pairs 3 --matrix core --output <results>

The retained build scripts show the exact commands and native runtime environment used. Adjust their recorded workspace paths for reproduction. Source snapshots and executable hashes identify the audited builds. The native example is built separately with gpui_platform/test-support and captures both builds at the same display scale. Executables, app profiles and dependency caches are excluded from the evidence snapshot.

All 66 benchmark cases

Times are µs per step. Handler-only rows exclude drawing and are kept separate from the title's key/scroll CPU range.

Text	Size	Position	Motion	Scope	Before µs	After µs	Median paired	Pair range
ascii	10KB	middle	left	Key + CPU draw	1,062.0014	47.9951	21.47×	16.23–23.50×
ascii	10KB	middle	left	Handler only	61.2753	0.0767	802.12×	783.83–819.49×
ascii	10KB	middle	right	Key + CPU draw	1,089.2920	52.0896	21.41×	13.92–21.50×
ascii	10KB	middle	right	Handler only	112.4538	0.0851	1321.94×	1308.39–1365.07×
ascii	10KB	middle	select_left	Key + CPU draw	1,076.0861	57.1007	17.91×	16.31–19.62×
ascii	10KB	middle	select_left	Handler only	61.1260	0.0781	784.37×	782.41–981.72×
ascii	10KB	middle	select_right	Key + CPU draw	1,132.7073	56.5052	19.45×	14.96–20.05×
ascii	10KB	middle	select_right	Handler only	113.1562	0.0844	1348.67×	1341.11–1442.46×
ascii	10KB	start	none	Scroll CPU draw	962.8306	41.2184	23.29×	21.13–23.39×
ascii	10KB	middle	none	Scroll CPU draw	959.6101	42.4420	22.61×	20.61–23.00×
ascii	10KB	end	none	Scroll CPU draw	965.6875	41.0566	23.68×	22.68–23.76×
multilingual	10KB	middle	left	Key + CPU draw	599.8476	51.9722	11.59×	11.13–12.06×
multilingual	10KB	middle	left	Handler only	51.6920	0.0976	537.43×	525.64–549.59×
multilingual	10KB	middle	right	Key + CPU draw	646.7691	56.5438	11.36×	11.14–12.05×
multilingual	10KB	middle	right	Handler only	82.8677	0.1153	717.87×	701.92–779.59×
multilingual	10KB	middle	select_left	Key + CPU draw	600.3396	60.1938	9.97×	8.98–10.10×
multilingual	10KB	middle	select_left	Handler only	51.7351	0.0990	524.61×	482.98–542.61×
multilingual	10KB	middle	select_right	Key + CPU draw	618.3878	60.2125	10.09×	8.90–11.99×
multilingual	10KB	middle	select_right	Handler only	81.1778	0.1156	687.61×	686.02–769.16×
multilingual	10KB	start	none	Scroll CPU draw	532.0486	41.3871	12.86×	12.19–14.04×
multilingual	10KB	middle	none	Scroll CPU draw	546.7656	46.6632	11.76×	10.58–12.92×
multilingual	10KB	end	none	Scroll CPU draw	534.9163	41.9601	12.75×	12.41–13.10×
emoji	10KB	middle	left	Key + CPU draw	969.8663	47.1774	20.56×	20.14–20.79×
emoji	10KB	middle	left	Handler only	57.2267	0.0781	737.40×	690.51–739.05×
emoji	10KB	middle	right	Key + CPU draw	1,000.6222	51.9559	19.25×	19.23–20.79×
emoji	10KB	middle	right	Handler only	101.8740	0.0882	1159.62×	1051.60–1251.28×
emoji	10KB	middle	select_left	Key + CPU draw	970.4059	56.0108	17.23×	16.85–18.40×
emoji	10KB	middle	select_left	Handler only	57.4476	0.0819	701.08×	698.80–708.18×
emoji	10KB	middle	select_right	Key + CPU draw	1,025.8646	56.1701	18.33×	18.08–18.42×
emoji	10KB	middle	select_right	Handler only	102.7424	0.0906	1134.34×	1124.02–1164.99×
emoji	10KB	start	none	Scroll CPU draw	911.9719	41.2413	22.26×	21.69–23.17×
emoji	10KB	middle	none	Scroll CPU draw	913.6128	42.8073	21.37×	20.98–22.36×
emoji	10KB	end	none	Scroll CPU draw	900.8705	41.6746	21.62×	21.29–21.84×
ascii	100KB	middle	left	Key + CPU draw	11,453.1253	48.4542	237.73×	233.59–285.29×
ascii	100KB	middle	left	Handler only	613.8167	0.0757	8146.21×	7964.26–8180.65×
ascii	100KB	middle	right	Key + CPU draw	12,101.6524	53.8781	224.90×	213.59–272.20×
ascii	100KB	middle	right	Handler only	1,195.4670	0.0840	14226.16×	13234.19–16610.83×
ascii	100KB	middle	select_left	Key + CPU draw	11,669.3750	61.8545	188.66×	179.98–257.02×
ascii	100KB	middle	select_left	Handler only	610.4674	0.0792	7848.34×	7570.21–10733.40×
ascii	100KB	middle	select_right	Key + CPU draw	12,029.4830	61.5837	195.81×	191.12–993.73×
ascii	100KB	middle	select_right	Handler only	1,145.2601	0.0844	13687.00×	13122.90–58962.18×
ascii	100KB	start	none	Scroll CPU draw	14,640.8660	42.1042	347.73×	281.29–391.76×
ascii	100KB	middle	none	Scroll CPU draw	10,298.9656	44.8035	230.81×	229.87–359.78×
ascii	100KB	end	none	Scroll CPU draw	13,395.3389	44.2097	290.46×	235.76–303.00×
multilingual	100KB	middle	left	Key + CPU draw	6,250.3764	50.2705	124.33×	122.93–208.92×
multilingual	100KB	middle	left	Handler only	544.6726	0.1215	4537.90×	4481.98–6467.08×
multilingual	100KB	middle	right	Key + CPU draw	6,272.0035	53.3767	118.17×	115.54–188.74×
multilingual	100KB	middle	right	Handler only	776.5788	0.0868	9000.57×	8945.94–12724.62×
multilingual	100KB	middle	select_left	Key + CPU draw	6,354.5337	60.3924	105.33×	98.34–150.64×
multilingual	100KB	middle	select_left	Handler only	559.3913	0.1208	4760.43×	4629.46–5928.14×
multilingual	100KB	middle	select_right	Key + CPU draw	6,410.3969	60.2306	106.43×	103.44–125.82×
multilingual	100KB	middle	select_right	Handler only	804.1306	0.0851	9491.07×	8987.61–10794.09×
multilingual	100KB	start	none	Scroll CPU draw	6,224.7684	42.1677	142.89×	131.81–150.60×
multilingual	100KB	middle	none	Scroll CPU draw	6,711.8236	44.4753	151.86×	119.31–182.35×
multilingual	100KB	end	none	Scroll CPU draw	6,102.3285	42.7024	144.22×	129.37–144.41×
emoji	100KB	middle	left	Key + CPU draw	10,563.7198	49.9507	205.61×	200.34–332.77×
emoji	100KB	middle	left	Handler only	594.1542	0.0823	6924.40×	6900.10–7706.05×
emoji	100KB	middle	right	Key + CPU draw	11,431.4035	53.2903	193.63×	190.71–283.96×
emoji	100KB	middle	right	Handler only	1,122.8167	0.0868	12935.68×	11401.02–13860.51×
emoji	100KB	middle	select_left	Key + CPU draw	11,057.9194	61.5580	177.74×	162.25–189.11×
emoji	100KB	middle	select_left	Handler only	592.4993	0.0840	7352.45×	6774.57–7387.03×
emoji	100KB	middle	select_right	Key + CPU draw	11,428.2972	60.9476	188.38×	169.03–213.14×
emoji	100KB	middle	select_right	Handler only	1,047.4684	0.0910	11513.80×	11016.30–12313.91×
emoji	100KB	start	none	Scroll CPU draw	9,728.7066	42.5514	225.93×	220.94–255.19×
emoji	100KB	middle	none	Scroll CPU draw	11,396.9778	44.3861	246.87×	216.41–281.18×
emoji	100KB	end	none	Scroll CPU draw	9,670.4868	43.1924	221.82×	214.29–232.55×

Correctness, regression tests and final source review

Check	Result
Exact one-file committed source, normal app opt-level 0	1,430 passed, 0 failed, 24 pre-existing opt-in tests ignored
Extended audit, app opt-level 3	1,439 passed, 0 failed, 26 ignored; cache diagnostic and viewport benchmark also executed explicitly
Existing/focused component tests in the PR	All 25 existing tests retained, plus five focused regressions
Geometry, selection and hit testing	480 exact baseline/candidate matches, including cold/warm layout, width changes, empty/Unicode text, alignments, single/multiline and fit-content modes
Native font/Metal rendering	108 byte-identical PNG/geometry pairs across ASCII, Unicode, empty lines and one long paragraph, fractional line heights, scroll positions, selection and composition
Native font/color invalidation	Four cached-versus-fresh font/color transitions pass in each build
Unicode	374,742 offsets in both directions; official Unicode 17 expectations cover 766 cases / 7,801 offsets
Layout-cache regression	Baseline probe fails on right, Shift-right and scroll; candidate retains the layout for all three
Scoped binary/test Clippy	Same 49 pre-existing app diagnostics, zero new, none in the changed file
Formatting	Changed-file rustfmt and patch whitespace pass; full-app rustfmt reports four untouched upstream files, verified byte-identical to main
Patch integrity	Independently applies to a temporary index at checked main and produces the exact final source; no audit hooks or temporary native example in the PR

The five repository regressions cover Unicode navigation at every byte offset in explicit fixtures, fractional wrapped-row clipping, width-probe/cache retention, resize/text/style/composition invalidation, and inherited fonts across deferred layout. The seven original exhaustive audit regressions and two independent geometry/composition checks were preserved and rerun outside the code diff.

Unicode verification uses direct old/new evaluation for short inputs and sampled long offsets, with the old forward/reverse iterator boundary oracle for all long offsets. Official Unicode expected chunks provide an independent conformance check. Both production helper bodies exactly match the freshly executed verification sources.

The native fixture uses the real production component, platform fonts and Metal with a fixed test theme. Its current baseline and candidate captures are both 560×320 at the same backing scale. Earlier display scales are retained in separate historical snapshots. Composition/editing/undo/redo checks pass, but these are not real system IME event entry or full signed-in app acceptance.

The other 24 existing ignored tests are opt-in audio/device/GPU/fixture benchmarks and real-app recording/upload integrations. They remain unexecuted; the logs list every ignored test and any recorded reason. No check was weakened or production failure waived.

Committed test output, extended test output, cache diagnostic, validation record, native captures, Unicode verification, Clippy output, formatting baseline comparison and final source review.

The final review traced settings/search, renaming, editor fields, screenshot/annotation inputs and teleprompter editing/scrolling. It checked cache ownership and invalidation, Unicode context, measurement geometry, clipping/decorations, the deferred font lifetime and shared public contracts. No unresolved correctness finding remained.

Latest main and duplicate PR check

Rechecked upstream main and the changed-file lists of all 154 open PRs, plus 100 recently merged PRs and historical text-input/grapheme/teleprompter searches. No published duplicate found in that audit.

The only open PR touching this file is #2587; its patch changes theme styling and removes selection_tint. It does not implement grapheme cursor navigation, measurement retention or paragraph culling. Recently merged #3125 only adds the focus accessor in this file, which this patch preserves. #2773 handles platform editing shortcuts, #2524/#2739 concern renderer text-overlay caches, and #3104 concerns platform frame scheduling.

Compared with the original audit base, current main's only change in this file is is_focused. It is preserved. The exact suite was rebuilt and rerun after the unrelated Windows-picker change at 900dd827b962. The PR branch starts at last-checked main 100f43a0d359; its later web/CLI-only commits leave all GPUI workspace, Rust crate and build-configuration trees identical to that tested tip. Both measured text-input sources remain byte-identical. Source-tree identity record, duplicate audit record.

Merge danger

Door: two-way. Reverting this one-file change restores the previous implementation. There are no dependency, persistence, migration, recording or export format changes.

Blast radius: shared text fields. Native Windows/Linux execution, system IME event entry and full signed-in app acceptance were not performed locally. Cold shaping, text edits, style changes and resizing can still scale with document length. Some Unicode boundaries require scanning surrounding context, and marked composition deliberately bypasses paragraph culling. This is a scoped improvement to warmed input and CPU redraw work.

Confidence Score: 5/5

The reviewed changes appear safe to merge; no actionable issue was found.

Summary

Speeds up shared GPUI text fields by finding nearby grapheme boundaries, keeping the active layout during width measurements, and skipping fully clipped paragraphs.

• Text fields keep their active layout while flex checks different widths.
• Caret movement finds nearby Unicode character boundaries from the current position.
• Scrolling text fields skip painting paragraphs outside the visible area.

_{Reviews (1) · Last reviewed commit: "perf(gpui): reuse text layouts during lo..." · Reviewed by Greptile}

[vercel]

@kartikkabadi is attempting to deploy a commit to the Cap Software Inc Team on Vercel.

A member of the Team first needs to authorize it.