[pathanin]

Follow-up to pathanin/paint#19, found while measuring it in WebKit.

The bug

drawCells (assets.js) found the fill cells a tile meets with an inclusive far edge. A tile that ends exactly where a cell starts therefore drew that cell too. So a fill that hadn't moved drew up to four cells per tile instead of one.

• Pixels were unaffected: the extra cells are clipped away.
• The work was up to 4×. The new objects.js section 32b measures it: repainting a 2048 × 1024 fill (8 tiles) called fill() 21 times before this fix, and 8 after.

The fix makes the far edges exclusive. A tile that partly overlaps a cell, as with a moved or scaled fill, still draws it.

A/B for PR 19 + this fix

Pre-PR master (2315e42) against after, at 4000², on a software GPU. Run in-page with the same script in Playwright's Chromium and in WebKitGTK 2.52's MiniBrowser. MiniBrowser came from Ubuntu's webkit2gtk-driver, run under Xvfb and driven over WebDriver, because Playwright's WebKit download is blocked in that environment.

Pencil scribble (300 frames, 4 moves each), main-thread ms per frame:

	Before: first 50 → last 50	After: first 50 → last 50
Chromium	0.97 → 2.30	0.93 → 0.79
WebKit	5.4 → 27.2, worst 66	3.7 → 3.5, worst 11

Chromium's wall time for the scribble dropped from 99 s to 37 s.

Background fill around 12 rings, medians of 3 runs:

	Before	After
Chromium: total	1909 ms	1348 ms
Chromium: longest frame gap	611 ms	236 ms
Chromium: heap growth	131 MB	3 MB
WebKit: total	1035 ms	920 ms
WebKit: pause when the fill lands	144–172 ms	80–94 ms

Before this fix, PR 19 alone measured 1023 ms total in WebKit.

Picker: about 1–2 ms in both engines before and after, when the GPU is idle.

WebKit functional pass (WebKitGTK 2.52)

These all check out:

• Fills: Transparent, soft-edged over white, soft-edged over transparency, and dense (noise). All are paths except the noise fill's one dense cell, which is pixels.
• Pencil, brush and eraser strokes.
• The cache equals a fresh render.
• Undo and redo are pixel-exact through every step.
• Save and reopen brings the objects back.

Save and reopen in WebKit, compared with the picture before saving:

	Pre-PR master	After
Differing pixels	385,387 (the Transparent fill came back uncleared, the bug PR 19 fixed)	3,757
Largest channel difference	255	63

All 3,757 are translucent pixels, changed by WebKit's premultiplied-alpha rounding. That rounding predates this work and also shows on master; Chrome's round trip is exact. task.md lists it as open.

Testing

• Unit tests: 87 pass.
• e2e, locally in Playwright's Chromium on a software GPU: objects 200/0, including the new 32b; smoke 313/0; alpha 105/0.
• Same failures as master: libs and filledge each fail one check, identically on master (2315e42) here, and both passed in CI on the merged PR 19.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Bjaj6ac9XGsjJWsPrMECjo