* fix: always use set for private identifiers
* we can simplify this further - no need to check the value was transformed, since the outcome of not returning immediately is the same but with extra steps
* add explanatory note
---------
Co-authored-by: Rich Harris <rich.harris@vercel.com>
through #14363 I noticed our `export default` validation wasn't quite right:
- missed checking for derived/state exports
- the "cannot have a default export" error was only thrown if you did `export default` from the instance script, but it shouldn't matter from which component part you export it; it's never ok
* fix: ensure internal cloning can work circular values
* better fixc
* 'original' feels slightly clearer than 'json_instance'
* use an optional parameter, so we can omit it in most cases
* Update packages/svelte/src/internal/shared/clone.js
Co-authored-by: Rich Harris <rich.harris@vercel.com>
---------
Co-authored-by: Rich Harris <rich.harris@vercel.com>
The detection whether or not props are writable is buggy; it doesn't take into account the case when instantiating components via `mount` or legacy-`new`
Fixes#14161
This posed an interesting question: What to do about (non-)`bind:`able properties? The answer I arrived on was: Treat it as if everything that _can_ be bound _is_ treated as bound, and everything else as readonly. In other words, if you're reassigning a prop, it will diverge from the passed in props if it's not bindable or not set in the parent, otherwise it will mutate the passed in props. I think that makes the most sense, given that you can't control this behavior from the outside.
* fix: avoid marking subtree as dynamic for inlined attributes
* fix: i'm a silly goose 🪿
* chore: refactor `is_inlinable_expression` to accept the attribute
* feat: inline dom expression too
* fix: special case literals with `"` in it and fix standalone case
* chore: simpler check first
Co-authored-by: Ben McCann <322311+benmccann@users.noreply.github.com>
* typo
* add more stuff to snapshot test
* simplify/speedup by doing the work once, during analysis
* simplify
* simplify - no reason these cases should prevent inlining
* return template
* name is incorrect
* name is incorrect
* fix escaping
* no longer necessary
* remove obsolete description
* better concatenation
* fix test
* do the work at runtime
* fix another thing
* tidy
* tidy up
* simplify
* simplify
* fix
* note to self
* another
* simplify
* more accurate name
* simplify
* simplify
* explain what is happening
* tidy up
* simplify
* better inlining
* update test
* colocate some code
* better inlining
* use attribute metadata
* Update packages/svelte/src/compiler/phases/2-analyze/visitors/Identifier.js
Co-authored-by: Ben McCann <322311+benmccann@users.noreply.github.com>
* Apply suggestions from code review
---------
Co-authored-by: Ben McCann <322311+benmccann@users.noreply.github.com>
Co-authored-by: Rich Harris <rich.harris@vercel.com>
We previously marked all `:root` selectors as global-like, which excempted them from further analysis. This causes problems:
- things like `:not(...)` are never visited and therefore never marked as used -> we gotta do that directly when coming across this
- `:has(...)` was never visited, too. Just marking it as used is not enough though, because we might need to scope its contents
Therefore the logic is enhanced to account for these special cases. Fixes#14118
While fixing this I cleaned up some inconsistencies in what we mark as global. This simplified code and fixed some adjacent bugs, which conindicentally also fixes#14189
* fix: consider static attributes that are inlined in the template
* fix: use `is_inlinable_expression`
* fix: move check for inlinable expression as last
* fix: simplify and correct
* chore: accept single node in `is_inlinable_expression`
* chore: update comment
* chore: add snapshots for non static nodes
Previously, we were applying an explicit nesting selector to the start of a relative selector chain only when starting the traversal. Prepending the selector is important because it ensures we traverse upwards to the parent rule when the current selectors all matched and there's still more to do. But we forgot to do the prepend for parent rules, which meant that if we were nested two levels deep, we would stop too early. This fix ensures we prepend in that case, too.
Fixes#14178
fixes#14168
This reverts the whole "selectors inside `:not` are scoped" logic. Scoping is done so that styles don't bleed. But within `:not`,everything is reversed, which means scoping the selectors now means they are more likely to bleed. That is the opposite of what we want to achieve, therefore we should just leave those selectors alone.
The exception are `:not` selectors with descendant selectors, as that means "look up the tree" and we need to scope all ancestor elements in that case.
If the contents of a `:not` selector don't match, then it's actually a match for `:not` because it's inverted. Therefore, we need to scope such elements. We're also making sure that contents of `:not` that never match actually count as a used (because the result is negated), and as such the contents of `:not` that always match are actually marked as unused.
Fixes#13974
`<a>` tags are valid in both the SVG and HTML namespace. If there's no parent, we therefore have to look downwards to see if it's the parent of a SVG or HTML element.
fixes#7807fixes#13793
We didn't account for the `$props` rune being writtin in a way that makes some props unknown, and they would only be visible through the `customElement.props` definition. This changes the iteration to account for that and also adds a note to the documentation that you need to list out the properties explicitly.
fixes#13785
Fixes#13848.
When we set custom element attributes/props, we should be doing so without the current effect/reaction active. Otherwise, the custom element lifecycle might attach effects/dependencies to the wrong reaction and all manner of things can incorrectly occur
---------
Co-authored-by: Simon Holthausen <simon.holthausen@vercel.com>
We originally didn't extend from `Error` anymore because its fields are of no real value to us, and has problems with serialization in a worker context.
Turns out this was a mistake, because various build tools rely on errors being thrown as something that extends Error, else they try to wrap it in their own error.
We therefore revert that change while still trying to preserve most of the advantages of not extending `Error`, namely nuking the useless stack trace and making sure the message is enumerable.
When CSS is externalized we rightfully rely on the following tooling chain to properly minify CSS. When we inject the CSS however, that tooling won't be able to do that, so we gotta do it ourselves.
This PR brings back most of that logic that existed in Svelte 4. Fixes#13716
Only count down after timeout, else we would reach 0 before our own render effect reruns, but reach 1 again when the tick callback of the prior teardown runs. That would mean we re-subcribe unnecessarily and create a memory leak because the old subscription is never cleaned up.
Dominic Gannaway
Rich Harris
Theodor Steiner
github-actions[bot]
Paolo Ricciuti
Frédéric Crozatier
Simon H
Santiago Cézar
Torsten Dittmann
Caique Torres
Gonzalo Ruiz
wackbyte
FoHoOV
Ahmad S.
Ben McCann