url: speed up URLSearchParams

[anonrig]

This speeds up WHATWG URLSearchParams without changing observable behavior.

Independent of the new URL() parse PR.

What changed

• Parse: walk & / = with indexOf instead of a per-character state machine. When the input has no + or %, each component is a slice of the original string. Percent-decoding still runs only when a complete %HH sequence exists, so lone % and fake sequences like %© stay intact.
• USVString: skip `${value}` when the value is already a string (constructor, append/get/set/has/delete).
• Serialize: cache toString() until the list mutates; join encoded pairs instead of repeated +=.
• Tests:test/parallel/test-whatwg-url-searchparams-fast-path.js covers leading ?, empty pairs, + / percent-decoding, invalid %, mutation cache invalidation, copy constructor isolation, record/sequence init, unpaired surrogates, and fake percent-encoding.

Tests

• All test/parallel/test-whatwg-url-custom-searchparams*.js plus the new fast-path file
• WPT test/wpt/test-url.js: 5107 passed, 0 unexpected failures

Local benches

Same binary family, both run with --no-node-snapshot (new JS is not in the V8 snapshot). Rates in ops/s:

Benchmark	Before	After
new URLSearchParams(string) noencode	~6.9M	~7.3M
new URLSearchParams(string) encodemany	6.9M	7.5M
new URLSearchParams(iterable)	9.9M	10.3M
toString() noencode (repeated)	7.7M	~213M
toString() encodemany (repeated)	4.9M	~200M
get()	~75M	~85M
has()	~52M	~59M

The large toString() jump is the serialization cache: the common “build params, stringify many times / read URL.href” path no longer re-encodes an unchanged list.

Ada already has a C url_search_params API. This keeps the implementation in JS to avoid a JS/C++ call on every get/append.

Assisted-by: Cursor

[anonrig]

This speeds up WHATWG URLSearchParams without changing observable behavior.

Independent of the new URL() parse PR.

What changed

• Parse: walk & / = with indexOf instead of a per-character state machine. When the input has no + or %, each component is a slice of the original string. Percent-decoding still runs only when a complete %HH sequence exists, so lone % and fake sequences like %© stay intact.
• USVString: skip `${value}` when the value is already a string (constructor, append/get/set/has/delete).
• Serialize: cache toString() until the list mutates; join encoded pairs instead of repeated +=.
• Tests:test/parallel/test-whatwg-url-searchparams-fast-path.js covers leading ?, empty pairs, + / percent-decoding, invalid %, mutation cache invalidation, copy constructor isolation, record/sequence init, unpaired surrogates, and fake percent-encoding.

Tests

• All test/parallel/test-whatwg-url-custom-searchparams*.js plus the new fast-path file
• WPT test/wpt/test-url.js: 5107 passed, 0 unexpected failures

Local benches

Same binary family, both run with --no-node-snapshot (new JS is not in the V8 snapshot). Rates in ops/s:

Benchmark	Before	After
new URLSearchParams(string) noencode	~6.9M	~7.3M
new URLSearchParams(string) encodemany	6.9M	7.5M
new URLSearchParams(iterable)	9.9M	10.3M
toString() noencode (repeated)	7.7M	~213M
toString() encodemany (repeated)	4.9M	~200M
get()	~75M	~85M
has()	~52M	~59M

The large toString() jump is the serialization cache: the common “build params, stringify many times / read URL.href” path no longer re-encodes an unchanged list.

Ada already has a C url_search_params API. This keeps the implementation in JS to avoid a JS/C++ call on every get/append.

Assisted-by: Cursor

[anonrig]

This speeds up WHATWG URLSearchParams without changing observable behavior.

Independent of the new URL() parse PR.

What changed

• Parse: walk & / = with indexOf instead of a per-character state machine. When the input has no + or %, each component is a slice of the original string. Percent-decoding still runs only when a complete %HH sequence exists, so lone % and fake sequences like %© stay intact.
• USVString: skip `${value}` when the value is already a string (constructor, append/get/set/has/delete).
• Serialize: cache toString() until the list mutates; join encoded pairs instead of repeated +=.
• Tests:test/parallel/test-whatwg-url-searchparams-fast-path.js covers leading ?, empty pairs, + / percent-decoding, invalid %, mutation cache invalidation, copy constructor isolation, record/sequence init, unpaired surrogates, and fake percent-encoding.

Tests

• All test/parallel/test-whatwg-url-custom-searchparams*.js plus the new fast-path file
• WPT test/wpt/test-url.js: 5107 passed, 0 unexpected failures

Local benches

Same binary family, both run with --no-node-snapshot (new JS is not in the V8 snapshot). Rates in ops/s:

Benchmark	Before	After
new URLSearchParams(string) noencode	~6.9M	~7.3M
new URLSearchParams(string) encodemany	6.9M	7.5M
new URLSearchParams(iterable)	9.9M	10.3M
toString() noencode (repeated)	7.7M	~213M
toString() encodemany (repeated)	4.9M	~200M
get()	~75M	~85M
has()	~52M	~59M

The large toString() jump is the serialization cache: the common “build params, stringify many times / read URL.href” path no longer re-encodes an unchanged list.

Ada already has a C url_search_params API. This keeps the implementation in JS to avoid a JS/C++ call on every get/append.

Assisted-by: Cursor

[anonrig]

This speeds up WHATWG URLSearchParams without changing observable behavior.

Independent of the new URL() parse PR.

What changed

• Parse: walk & / = with indexOf instead of a per-character state machine. When the input has no + or %, each component is a slice of the original string. Percent-decoding still runs only when a complete %HH sequence exists, so lone % and fake sequences like %© stay intact.
• USVString: skip `${value}` when the value is already a string (constructor, append/get/set/has/delete).
• Serialize: cache toString() until the list mutates; join encoded pairs instead of repeated +=.
• Tests:test/parallel/test-whatwg-url-searchparams-fast-path.js covers leading ?, empty pairs, + / percent-decoding, invalid %, mutation cache invalidation, copy constructor isolation, record/sequence init, unpaired surrogates, and fake percent-encoding.

Tests

• All test/parallel/test-whatwg-url-custom-searchparams*.js plus the new fast-path file
• WPT test/wpt/test-url.js: 5107 passed, 0 unexpected failures

Local benches

Same binary family, both run with --no-node-snapshot (new JS is not in the V8 snapshot). Rates in ops/s:

Benchmark	Before	After
new URLSearchParams(string) noencode	~6.9M	~7.3M
new URLSearchParams(string) encodemany	6.9M	7.5M
new URLSearchParams(iterable)	9.9M	10.3M
toString() noencode (repeated)	7.7M	~213M
toString() encodemany (repeated)	4.9M	~200M
get()	~75M	~85M
has()	~52M	~59M

The large toString() jump is the serialization cache: the common “build params, stringify many times / read URL.href” path no longer re-encodes an unchanged list.

Ada already has a C url_search_params API. This keeps the implementation in JS to avoid a JS/C++ call on every get/append.

Assisted-by: Cursor

[anonrig]

This speeds up WHATWG URLSearchParams without changing observable behavior.

Independent of the new URL() parse PR.

What changed

• Parse: walk & / = with indexOf instead of a per-character state machine. When the input has no + or %, each component is a slice of the original string. Percent-decoding still runs only when a complete %HH sequence exists, so lone % and fake sequences like %© stay intact.
• USVString: skip `${value}` when the value is already a string (constructor, append/get/set/has/delete).
• Serialize: cache toString() until the list mutates; join encoded pairs instead of repeated +=.
• Tests:test/parallel/test-whatwg-url-searchparams-fast-path.js covers leading ?, empty pairs, + / percent-decoding, invalid %, mutation cache invalidation, copy constructor isolation, record/sequence init, unpaired surrogates, and fake percent-encoding.

Tests

• All test/parallel/test-whatwg-url-custom-searchparams*.js plus the new fast-path file
• WPT test/wpt/test-url.js: 5107 passed, 0 unexpected failures

Local benches

Same binary family, both run with --no-node-snapshot (new JS is not in the V8 snapshot). Rates in ops/s:

Benchmark	Before	After
new URLSearchParams(string) noencode	~6.9M	~7.3M
new URLSearchParams(string) encodemany	6.9M	7.5M
new URLSearchParams(iterable)	9.9M	10.3M
toString() noencode (repeated)	7.7M	~213M
toString() encodemany (repeated)	4.9M	~200M
get()	~75M	~85M
has()	~52M	~59M

The large toString() jump is the serialization cache: the common “build params, stringify many times / read URL.href” path no longer re-encodes an unchanged list.

Ada already has a C url_search_params API. This keeps the implementation in JS to avoid a JS/C++ call on every get/append.

Assisted-by: Cursor

[anonrig]

This speeds up WHATWG URLSearchParams without changing observable behavior.

Independent of the new URL() parse PR.

What changed

• Parse: walk & / = with indexOf instead of a per-character state machine. When the input has no + or %, each component is a slice of the original string. Percent-decoding still runs only when a complete %HH sequence exists, so lone % and fake sequences like %© stay intact.
• USVString: skip `${value}` when the value is already a string (constructor, append/get/set/has/delete).
• Serialize: cache toString() until the list mutates; join encoded pairs instead of repeated +=.
• Tests:test/parallel/test-whatwg-url-searchparams-fast-path.js covers leading ?, empty pairs, + / percent-decoding, invalid %, mutation cache invalidation, copy constructor isolation, record/sequence init, unpaired surrogates, and fake percent-encoding.

Tests

• All test/parallel/test-whatwg-url-custom-searchparams*.js plus the new fast-path file
• WPT test/wpt/test-url.js: 5107 passed, 0 unexpected failures

Local benches

Same binary family, both run with --no-node-snapshot (new JS is not in the V8 snapshot). Rates in ops/s:

Benchmark	Before	After
new URLSearchParams(string) noencode	~6.9M	~7.3M
new URLSearchParams(string) encodemany	6.9M	7.5M
new URLSearchParams(iterable)	9.9M	10.3M
toString() noencode (repeated)	7.7M	~213M
toString() encodemany (repeated)	4.9M	~200M
get()	~75M	~85M
has()	~52M	~59M

The large toString() jump is the serialization cache: the common “build params, stringify many times / read URL.href” path no longer re-encodes an unchanged list.

Ada already has a C url_search_params API. This keeps the implementation in JS to avoid a JS/C++ call on every get/append.

Assisted-by: Cursor

[anonrig]

This speeds up WHATWG URLSearchParams without changing observable behavior.

Independent of the new URL() parse PR.

What changed

• Parse: walk & / = with indexOf instead of a per-character state machine. When the input has no + or %, each component is a slice of the original string. Percent-decoding still runs only when a complete %HH sequence exists, so lone % and fake sequences like %© stay intact.
• USVString: skip `${value}` when the value is already a string (constructor, append/get/set/has/delete).
• Serialize: cache toString() until the list mutates; join encoded pairs instead of repeated +=.
• Tests:test/parallel/test-whatwg-url-searchparams-fast-path.js covers leading ?, empty pairs, + / percent-decoding, invalid %, mutation cache invalidation, copy constructor isolation, record/sequence init, unpaired surrogates, and fake percent-encoding.

Tests

• All test/parallel/test-whatwg-url-custom-searchparams*.js plus the new fast-path file
• WPT test/wpt/test-url.js: 5107 passed, 0 unexpected failures

Local benches

Same binary family, both run with --no-node-snapshot (new JS is not in the V8 snapshot). Rates in ops/s:

Benchmark	Before	After
new URLSearchParams(string) noencode	~6.9M	~7.3M
new URLSearchParams(string) encodemany	6.9M	7.5M
new URLSearchParams(iterable)	9.9M	10.3M
toString() noencode (repeated)	7.7M	~213M
toString() encodemany (repeated)	4.9M	~200M
get()	~75M	~85M
has()	~52M	~59M

The large toString() jump is the serialization cache: the common “build params, stringify many times / read URL.href” path no longer re-encodes an unchanged list.

Ada already has a C url_search_params API. This keeps the implementation in JS to avoid a JS/C++ call on every get/append.

Assisted-by: Cursor

[anonrig]

This speeds up WHATWG URLSearchParams without changing observable behavior.

Independent of the new URL() parse PR.

What changed

• Parse: walk & / = with indexOf instead of a per-character state machine. When the input has no + or %, each component is a slice of the original string. Percent-decoding still runs only when a complete %HH sequence exists, so lone % and fake sequences like %© stay intact.
• USVString: skip `${value}` when the value is already a string (constructor, append/get/set/has/delete).
• Serialize: cache toString() until the list mutates; join encoded pairs instead of repeated +=.
• Tests:test/parallel/test-whatwg-url-searchparams-fast-path.js covers leading ?, empty pairs, + / percent-decoding, invalid %, mutation cache invalidation, copy constructor isolation, record/sequence init, unpaired surrogates, and fake percent-encoding.

Tests

• All test/parallel/test-whatwg-url-custom-searchparams*.js plus the new fast-path file
• WPT test/wpt/test-url.js: 5107 passed, 0 unexpected failures

Local benches

Same binary family, both run with --no-node-snapshot (new JS is not in the V8 snapshot). Rates in ops/s:

Benchmark	Before	After
new URLSearchParams(string) noencode	~6.9M	~7.3M
new URLSearchParams(string) encodemany	6.9M	7.5M
new URLSearchParams(iterable)	9.9M	10.3M
toString() noencode (repeated)	7.7M	~213M
toString() encodemany (repeated)	4.9M	~200M
get()	~75M	~85M
has()	~52M	~59M

The large toString() jump is the serialization cache: the common “build params, stringify many times / read URL.href” path no longer re-encodes an unchanged list.

Ada already has a C url_search_params API. This keeps the implementation in JS to avoid a JS/C++ call on every get/append.

Assisted-by: Cursor