[robobun]

Problem

• bun build prints a legacy alias before the property it aliases: a{gap:9px;grid-gap:1px} becomes a{grid-gap:1px;gap:9px}, and 9px applies. Bootstrap 5.0.2 .form-select loses -moz-padding-start the same way.
• DeclarationBlock::minify (src/css/declaration.rs:82) writes an unknown name such as grid-gap at once, while a handler still holds gap.

Fix

• alias: rows in the name table (PropertyId::from_name_and_prefix) list 54 names that a current Blink, WebKit or Gecko table aliases to a held property. Each parses as CustomPropertyName::Alias(name, property).
• The Owner::Custom arm of css: dispatch each declaration to the handler that owns its property #45747 calls flush_before_alias: a handler that holds what the alias can override flushes first.
• Verified: test/js/bun/css/css.test.ts (228 new rows, 148 fail without the change). Also test/js/bun/css/ and test/bundler/css/.
• Self-reviewed: 54 concerns raised, 49 addressed, 5 rejected (reasons in Notes).

Background

• The base branch merges css: dispatch each declaration to the handler that owns its property #45747 and css: compare custom property names in PropertyId equality and importance in duplicate-rule detection #33684. Merge those first, then retarget this PR to main.
• A handler holds one shorthand family (margin-*) and writes it merged at a flush.
• Considered a name list in each handler (css: keep font/transition sub-properties on the side of the shorthand they were written on #41990) and a flush with no guards (24 sheets change, not 9). Notes ask a maintainer two questions.

Downsides

• 9 of 1,189 stylesheets change (40 rules, +397 B). A handler that flushes for an alias cannot merge across it.
• +0.027% instructions on the 15 in-tree sheets, +4,608 B of .text. With 289 name table keys (was 251), overflow-x resolves in 359 instructions (was 258).
• Under the default targets, a background after an alias gets a generated fallback also after an author fallback (background:red;-webkit-background-size:2px;background:lab(50% 10 10) gets background:#8c7166). In vertical writing modes, 8 of 40,000 random rules newly get a wrong border colour.

Notes

Reproduce.

echo 'a{gap:9px;grid-gap:1px}' > a.css
bun build a.css --minify
# 1.4.2:   a{grid-gap:1px;gap:9px}
# this PR: a{gap:9px;grid-gap:1px}

The reorder also occurs without --minify, in nested rules, in @media, and when two rules with the same selector merge.

Base branch. robobun/b77e2224/css-alias-stack-base is a merge of the branches of #45747 (owner dispatch) and #33684 (name and importance in the duplicate-rule check). This PR needs both:

• The alias test sits in the Owner::Custom arm that css: dispatch each declaration to the handler that owns its property #45747 adds.
• An alias that stays in its place can line up with another name in a later rule with the same selector. StyleRule::is_duplicate then dropped the earlier rule, because PropertyId equality ignored custom names. css: compare custom property names in PropertyId equality and importance in duplicate-rule detection #33684 compares the names and the importance. Seven rows of the new test block fail without it, for example a{gap:1px;grid-gap:2px;zoom:1}b{color:red}a{gap:inherit;grid-gap:3px;cursor:pointer}.

Which names have a row. A name has a row when all of these are true:

• The current Blink (css_properties.json5, alias_for), WebKit (CSSProperties.json, aliases) or Gecko (longhands.toml and shorthands.toml, aliases) property table lists it as another name of a property.
• Bun has that property as a typed property, and does not parse the name.
• A handler holds that property.
• The flush that the row causes prints valid CSS.

61 pairs pass the first two tests. 54 have a row:

• gap: grid-gap, grid-row-gap, grid-column-gap, -webkit-column-gap
• -webkit-justify-items
• margin and padding: -webkit-margin-start/-end/-before/-after, -moz-margin-start/-end, and the same six for padding
• scroll-snap-margin and its four sides
• size: -webkit-logical-width/-height and the min- and max- names
• -webkit-background-size, -webkit-background-origin
• border: -webkit-border-start/-end/-before/-after and -moz-border-start/-end, each with -color, -style and -width

The other 7 have no row:

• 6 have no handler, so they already keep their place: -epub-text-emphasis-color, -webkit-mask-composite, -webkit-mask-position-x, -webkit-mask-position-y, -webkit-opacity, supported-color-schemes.
• font-width (WebKit and Gecko alias font-stretch to it). A flush for it can print font:75% 12px serif, which browsers drop, because FontHandler::flush folds a percentage stretch into font. css: keep css-wide keywords and percentage stretch out of the font and transition shorthands #42022 fixes that fold, and css: keep font/transition sub-properties on the side of the shorthand they were written on #41990 has font-width in its font name list.

Names of removed aliases have no row. -moz-column-gap was an alias in Gecko through Firefox 73. Firefox 74 put it behind the pref layout.css.prefixes.columns (default off), and Firefox 102 has no such prefix. The oldest Firefox in the default targets is 78. No row comes from a prefix alone: -webkit-grid-gap and -ms-gap stay unknown.

What a handler does at an alias. DeclarationHandler::flush_before_alias asks the handlers in turn. A handler returns false for a property of another handler. A debug build asserts that one handler knows the property.

• gap: writes the held row-gap and column-gap. The rest of the align family stays held. -webkit-justify-items flushes the align family when justify-items is held.
• margin, padding, scroll-margin: flush when the slot of the alias is held. The writing mode decides which physical side a logical side is. So a logical alias also flushes for a held physical side, and a physical alias for a held logical side.
• size: writes only the sizes that the alias can override: the logical size and the width and height of the same kind (min-, max-). The other sizes stay held.
• border: flushes when the side of the alias or a physical side is held.
• background: flushes when a size or an origin is held, or when a prefixed background shorthand waits in decls for the end of the rule. The prefixed declarations are written first.

Except for gap and size, a flush writes all that the handler holds, also what the alias cannot override.

Generated fallbacks after an alias. With targets, a handler generates a fallback for a value: a hex colour for lab(), -moz-fit-content for fit-content. It does not when flushed_properties says that an earlier flush already wrote that property. The earlier value is then the fallback of the author. An alias between the two values can override the earlier one. The rule: a later declaration that can override the alias gets its own generated fallbacks. A later declaration that cannot override the alias keeps the earlier value as its fallback.

• size: after the write, the marks of the sizes that the alias can override are removed. This is done at each alias, also when nothing is held. The result is then the same when the rule merges with another rule and is minified again.
• border: the marks of the components that the alias can override are removed. For a width alias these are the widths of its own side and of the four physical sides. The other marks stay.
• background: a background shorthand resets the size and the origin, so it can override the alias. After an alias it gets its fallbacks. The marks from before the alias still count for background-color and background-image.
• Consequence: a{background:red;-webkit-background-size:2px;background:lab(50% 10 10)} gets generated fallbacks after the alias, where main keeps only red. a{width:100px;-webkit-logical-width:5px;width:stretch} gets width:-webkit-fill-available;width:-moz-available before stretch.
• a{background-size:1px;background-color:red;-webkit-background-size:2px;background-color:lab(50% 10 10)} keeps only the author's red, as on main.
• Without this, a{background-size:1px;-webkit-background-size:2px;background:lab(50% 10 10)} loses the hex fallback that main prints.

Measurements. Base is the base branch (7944935). Both are release builds. valgrind, bloaty and perf are not installed on the build machine. The counts come from qemu-x86_64 -one-insn-per-tb -d exec,nochain -dfilter, from a hardware instruction counter (perf_event_open), from llvm-size and from nm.

• Minify, instructions per declaration (qemu, handler and DeclarationBlock::minify symbols): --x 60 (was 63), cursor 60 (was 63), display 57 (was 62), color 86 (was 83), width 102 (was 98), margin 212 (was 211), grid-gap 99 (was 63), -webkit-margin-start 191 (was 63).
• Name resolution, instructions per declaration (qemu): cursor 240 (was 240), display 299 (was 296), margin 293 (was 290), text-align 340 (was 308), transition 579 (was 510), overflow-x 359 (was 258), --x 278 (was 278), grid-gap 205 (was 248). One lookup per name before and after.
• The table is a binary search per name length. 38 more keys move 65 typed names 1 to 4 steps deeper and 48 names 1 to 3 steps nearer. Over the 1,299,452 declarations of the corpus (counted with a regular expression) the average is +0.099 steps.
• Parse, minify and print of the 15 sheets in test/js/bun/css/files (standalone driver over the bun_css crate, hardware counter, minimum of 5 runs): 896,501,407 to 896,739,750 instructions with no targets (+0.027%), 918,039,672 to 918,281,339 with browser targets (+0.026%). Two copies of the base differ by 0.0002%.
• Type sizes: CustomPropertyName 24 B (was 24), PropertyId 24 B (was 24), Property 216 B (was 216).
• Release bun: .text 81,072,325 to 81,076,933 B (+4,608 B). .rodata grows by 4,096 B: 684 B of table data, and the rest is alignment. DeclarationHandler::flush_before_alias is 2,379 B and is marked cold.
• Allocations (an LD_PRELOAD counter on the driver): 1,003 per 1,000 cursor, --x or grid-gap declarations before and after.

Output on real stylesheets. 1,254 sheets: the 15 in-tree sheets and 1,239 sheets from npm packages. 65 do not parse, with or without the change.

• No targets: 9 sheets change, 40 rules, +397 B. Default browser targets: the same 9 sheets, 42 rules, +431 B.
• The sheets: amazeui 2.7.2 (2 files), bonsai.css 1.2.2, bootstrap 5.0.2 (2 files), fundamental-styles 0.41.9 (3 files), material-dashboard 3.0.9.
• The changed rules have a margin alias (24), a padding alias (14) or -webkit-background-size (2). No rule changes without an alias in it. No rule changes for the gap names.

Browser check. Headless Chrome 153.0.8010.52 computed the style of each source sheet and of its output, in four writing modes. The page renames the names that Chrome does not know (-moz-*, scroll-snap-margin*) to the property they alias.

• Input space: one rule, 2 to 6 declarations from one alias family, with px values, three colours and three line styles. Only alias names get a var() value.
• 40,000 such sheets, no targets: the computed style differs for 7,598 sheets on the base and for 0 with the change.
• The same sheets, default browser targets: 7,841 differ on the base and 324 with the change. All 324 are border rules in a vertical writing mode. There, Bun writes an equal border-inline-* pair as left and right for a target that has no such shorthand. 316 of them also differ on the base. In the other 8 the flush runs that lowering before a later declaration can merge with the pair.
• 10,000 sheets of the shape .t{..} .o{color:red} .t{..}, no targets: 1,041 differ on the base and 0 with the change.
• 10,000 one-rule sheets where typed names also get var(): 1,731 differ on the base and 21 with the change. The 21 also differ on the base. A typed value that does not parse is css: keep an unparsed width/height/color after the declarations written before it #42101.

Fallback check. Two scripts model the cascade of Chrome 80 and Firefox 78 for the size properties and for border and background colours. They compare the old browser on the output with a current browser on the source, over 20,000 random rules each, alone and merged with a second rule.

• Sizes: the result is the same in 16,237 rules for Chrome 80 (base 13,048) and in 17,546 for Firefox 78 (base 16,717). Merged: 16,151 (base 12,614) and 17,455 (base 16,588).
• Colours, Chrome 80: 16,095 (base 15,264). Merged: 16,097 (base 15,202).
• The base is right and the change is wrong in 44 and 16 size rules and in 11 colour rules. In those, the base is right by accident: it prints a held value after a var() value or after the alias. Or the alias itself has a value that the old browser does not know.

Not in this PR.

• An unknown name that is not an engine alias, after a shorthand that resets it: css: keep font/transition sub-properties on the side of the shorthand they were written on #41990 (font and transition). Its is_unknown_any_of needs an arm for CustomPropertyName::Alias when it meets this PR.
• A shorthand that Bun does not parse, after an earlier declaration for its longhands: no PR. See the next section.
• font-width: see above.
• A typed property with a value that does not parse: css: keep an unparsed width/height/color after the declarations written before it #42101.
• The all shorthand: css: make the all shorthand reset the declarations before it #42042.
• BackgroundHandler keeps prefixed background declarations in decls until the end of the rule. After a flush for a var() value they still print last. That output is as on main.
• The lowering of an equal border-inline-* pair to left and right is wrong in vertical writing modes, with or without an alias.
• Targets that compile logical properties (chrome before 69, firefox before 41): the compiled value goes to generated rules after the main rule. bun build has no such target. css: keep later physical declarations ahead of compiled logical ones #38623 works on that order.

Shorthands that Bun does not parse. 28 names in the current engine tables are shorthands of longhands that a handler holds. Each prints ahead of an earlier held declaration, with and without this PR.

• a{border-radius:4px;border-top-radius:8px} prints a{border-top-radius:8px;border-radius:4px}. a{width:10px;size:20px} prints a{size:20px;width:10px}.
• css-borders-4 has 25 of the names: 8 side shorthands such as border-top-radius, and 17 corner shorthands. css-sizing-4 has 3: size, min-size and max-size.
• WebKit main enables the side shorthands and the size shorthands by default. Blink has the css-borders-4 names behind experimental flags, and no size shorthands. Gecko has none of the 28 shorthands.
• They are shorthands, not aliases, so they have no row.

Alternatives.

• A name list in each handler, as css: keep font/transition sub-properties on the side of the shorthand they were written on #41990 has for font and transition. The minify pass then compares each unknown name with each list. The name table does one lookup, where the parser already resolves the name.
• A flush of the whole family for each alias, with no guards (scratch build, same corpus): 24 sheets change, 111 rules, +1,556 B with no targets.
• A dispatch of the alias on target.owner(). The 15 in-tree sheets then take 0.093% more instructions, because the compiler stops inlining owner() into handle_property.

Questions for a maintainer.

• Is the property name table the right home for engine aliases? The other home is a name list in each handler, as in css: keep font/transition sub-properties on the side of the shorthand they were written on #41990. The two PRs should end with one mechanism. The answer also decides where the 28 shorthands above go.
• Is the fallback rule after an alias right (see "Generated fallbacks after an alias")? The other choice keeps flushed_properties as the flush sets it. Then the author's earlier value stays the only fallback, and the background:lab() example above loses its generated fallback.

Self-review, rejected concerns.

• A cascade model in the test file that computes the expected order for each row. It is about 230 lines of test code for one block. The rows pin each guard clause, and the browser check covers the semantics.
• Dispatch on target.owner() with a flush of the whole family: measured slower and it changes more sheets, see Alternatives. The handlers now return whether they know the property, and a debug assert covers a row that no handler knows.
• A plain flush at the alias, with no change to flushed_properties: it loses fallbacks that main generates.
• Rows for the 15 names of css: keep font/transition sub-properties on the side of the shorthand they were written on #41990 in this PR: that waits for the answer to the first question.
• Issues for the leftovers under "Not in this PR": they are listed here.

Tests. bun bd test test/js/bun/css/css.test.ts: 1,901 pass, 0 fail. The other files in test/js/bun/css/, the 9 files in test/bundler/css/, test/bundler/esbuild/css.test.ts and test/regression/issue/44704.test.ts pass in the debug build. nested-selector-expansion.test.ts needs a longer time limit there. css-fuzz.test.ts has 10 s limits that the debug build on the build machine does not meet, with and without the change. It passes on the release build.

[robobun]

Status

Reproduced on Bun 1.4.2 and on the base branch of this PR:

echo 'a{gap:9px;grid-gap:1px}' > a.css
bun build a.css --minify
# a{grid-gap:1px;gap:9px}

With this PR the output is a{gap:9px;grid-gap:1px}.

bun bd test test/js/bun/css/css.test.ts -t "legacy name aliases" runs the 228 new rows. 148 fail on the base branch. All 228 pass with the change.

CI on d0e476a (build 128472): no CSS test fails on any lane. One test is red, test/cli/install/bun-install-windows-rename-retry.test.ts on Windows 11 aarch64. This PR does not touch bun install, and the same test is red on the builds of #45747 and on the build before this push. The build is finished: 180 of 181 jobs passed.

The diff is ready for a maintainer. The PR notes ask two questions, and the base branch needs #45747 and #33684 to merge first.

PR: #45825

[robobun]

^{Updated 3:11 PM PT - Oct 18th, 2026}

❌ @robobun, your commit d0e476a has 1 failures in Build #128472 (All Failures):

• test/cli/install/bun-install-windows-rename-retry.test.ts - code 1 on 🪟 11 aarch64

---

🧪   To try this PR locally:

bunx bun-pr 45825

That installs a local version of the PR into your bun-45825 executable, so you can run:

bun-45825 --bun

[claude]

Beyond the inline findings, I also checked the other places that match on CustomPropertyName (src/css/declaration.rs:336, src/css/properties/properties_generated.rs:2294, src/css/properties/ui.rs:162) for the new Alias variant — each only tests for Custom, so the new variant takes the existing unknown-name path there. I also confirmed properties_generated.rs is hand-maintained (header comment, no generator in the tree), so the alias rows will not be overwritten.

Extended reasoning...

The change adds an alias table to the CSS property-name lookup, a new CustomPropertyName::Alias variant, and per-handler flush_before_alias paths across eight handlers, with 237 lines of new tests. It touches no security-sensitive surface. Five verified findings (rule loss via is_duplicate, lost prefix fallback after an alias flush, and gaps in alias coverage) are posted inline, so this is not approvable as-is; the body records only what was additionally audited and ruled out.

[robobun]

The two blocking findings are fixed in 741e3f1. The three other findings have a reply in their thread.

• Rule loss in is_duplicate. The base branch is now a merge of css: dispatch each declaration to the handler that owns its property #45747 and css: compare custom property names in PropertyId equality and importance in duplicate-rule detection #33684. css: compare custom property names in PropertyId equality and importance in duplicate-rule detection #33684 compares custom names and importance, so the example keeps all three rules. Seven new rows fail without it.
• Lost fallback after an alias flush. A flush for an alias no longer counts as an earlier fallback. The size handler writes only the sizes that the alias can override. The row for background:lab() now expects the generated fallbacks.
• -moz-column-gap, logical scroll-snap-margin-*. No rows: no current Blink, WebKit or Gecko table lists these names.
• Compiled logical properties. Unchanged, and bun build has no such target.

This push also changes three things that the review did not name:

• font-width has no row. A flush for it can print a font shorthand with a percentage stretch (css: keep css-wide keywords and percentage stretch out of the font and transition shorthands #42022 fixes that fold).
• BackgroundHandler also flushes when a prefixed background waits in decls for the end of the rule. Before, -webkit-background-size could still print ahead of it.
• Each handler reports whether it knows the aliased property, and a debug build asserts that one does.

The PR body has the new numbers and two questions for a maintainer.

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

Beyond the inline finding, I also checked the AlignHandler::flush refactor in src/css/properties/align.rs — the gap pair was already the last thing flush wrote, so moving it into flush_gaps keeps the output order — and the slice bound in keeps_prefixed_shorthand (src/css/properties/background.rs), which cannot underflow because written is 0 whenever decls is empty.

Extended reasoning...

The second push restructures the alias flush in background.rs and border.rs to snapshot and restore flushed_properties, and adds the keeps_prefixed_shorthand check; the align.rs flush_gaps extraction and the background.rs slice arithmetic were the two refactor-shaped changes I examined and ruled out as behavior-preserving.

[robobun]

The optional finding is addressed in d0e476a.

• Border: a flush for an alias removes only the marks of what the alias can override. An earlier colour stays the fallback after a width alias.
• Background: the marks from before the alias still count for background-color and background-image. A background shorthand after the alias still gets its generated fallbacks, because it resets the size and the origin that the alias sets. The thread has the reason and the example.

The test block has rows for both directions (228 rows, 148 fail on the base branch). The PR body has the rule under "Generated fallbacks after an alias" and the new numbers.

[claude]

I re-reviewed the latest push (d0e476a) and found no bugs; the narrowed flushed_before_alias restore in background.rs addresses the point raised on the previous round. Because this changes cascade ordering across five property handlers and the description itself leaves two questions for a maintainer, a human look is still worthwhile.

What was reviewed:

• Every match on CustomPropertyName in src/css handles the new Alias variant (via | with Unknown, or a Custom(..) negation where Alias should behave like Unknown).
• keeps_prefixed_shorthand slice bound: written is 1 only when decls is non-empty, so no underflow.
• debug_assert!(known) in flush_before_alias compiles out in release, which is acceptable here since a missed handler only means the alias keeps today's passthrough behaviour.

Extended reasoning...

The branch adds a KnownName::AliasOf table in properties_generated.rs, a CustomPropertyName::Alias variant, and flush_before_alias hooks on the background, border, align, size, and margin/padding/scroll-margin handlers so legacy alias names print after the typed property they override. It touches no security-sensitive surface; it is a CSS minifier output-ordering change with +766/-38 lines and 228 new test rows. Deferring rather than approving because the writing-mode and fallback-retention heuristics are subtle, the PR depends on two other unmerged PRs as its base, and the author explicitly asks a maintainer to decide on design questions in the notes.

[robobun]

Check against current main

main has 109 commits that this branch does not have. Three of them change src/css/. I merged this branch (d0e476a) into main 5c6d4e2 and built a debug binary. The comparison is a debug build of main alone.

• a{gap:9px;grid-gap:1px} and Bootstrap 5.0.2 .form-select keep the source order. main prints the alias first.
• test/js/bun/css/css.test.ts: 1902 pass, 0 fail. main fails 157 tests of this file, 149 of them in the 228 alias rows.
• 45 more files pass: the rest of test/js/bun/css/ except css-fuzz.test.ts, test/bundler/css/ with wpt/, test/bundler/esbuild/css.test.ts, test/bundler/esbuild/loader.test.ts, test/bundler/bundler_html.test.ts, test/bundler/bundler_loader.test.ts, test/bake/dev/css.test.ts, and 13 CSS tests in test/regression/issue/.
• Time limits, on both debug builds: css-fuzz.test.ts exceeds its 10 s limits. Two tests of nested-selector-expansion.test.ts take 3.4 s to 4.6 s of a 5 s limit, and timed out in the first run after each build.

Output on 976 stylesheets

The sheets were on disk: WebKit Web Inspector, browserbench.org, webkit.org, the Rust docs and test/node_modules. Each was minified with no targets and with the default targets.

• 919 parse, on both builds. 914 are byte-identical. No build crashes.
• Alias order changes in 3 sheets: Bootstrap 5.3.7 (-webkit-margin-end after margin, 6 rules), the Next.js news site of Speedometer 3.0 (-webkit-margin-start), and webkit.org leaves.css (-webkit-background-size after background).
• 2 sheets keep rules that main drops, which is the fix of css: compare custom property names in PropertyId equality and importance in duplicate-rule detection #33684: Angular Material 20.1.6 (3 html{} rules) and the theme of The Rust Reference (1 :root{} rule).
• WebKit's CSSProperties.json lists 137 alias names. For a{property:v1;alias:v2}, main prints the alias first for 40 names that Bun does not parse. The merged build does so for none.

Landing order

• The base branch holds css: dispatch each declaration to the handler that owns its property #45747 at e3c8c10. css: dispatch each declaration to the handler that owns its property #45747 now has two more commits. Both change only test/js/bun/css/css.test.ts.
• After css: dispatch each declaration to the handler that owns its property #45747 and css: compare custom property names in PropertyId equality and importance in duplicate-rule detection #33684 are squash-merged, a retarget to main alone shows conflicts in 4 files of src/css/. git rebase --onto main 794493503b has no conflict today. Its src/ tree is the tree that I built and tested.

[Jarred-Sumner]

Not taken in #47025 (7942008), which consolidated the open CSS parser pull requests. It is right or harmless, but not worth its code, the bigger output or the behavior it changes.

[robobun]

Understood. This PR stays closed, and I plan no follow-up for the alias order.

For later readers: main at 7942008 still prints a{gap:9px;grid-gap:1px} as a{grid-gap:1px;gap:9px}. That is the decision above, not an open bug.