1874373 - Mark String.prototype.{codePointAt,at} and StringIterator.prototype.next as always inlinable

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, enhancement, P1)

Description

[André Bargull [:anba]]

• String.prototype.codePointAt and String.prototype.at are slightly too large to be inlinable, so they're noticeably slower when compared to String.prototype.charCodeAt resp. String.prototype.charAt.
• StringIteratorNext is also too large to be inlined by default, which means scalar replacement of the iterator and the result object doesn't work.

Comment 1

[André Bargull [:anba]]

Attachment: Bug 1874373 - Part 1: Mark StringIterator.p.next as inlinable. r=jandem!

• String.prototype.codePointAt and String.prototype.at are slightly too large
  to be inlinable, so they're noticeably slower when compared to
  String.prototype.charCodeAt resp. String.prototype.charAt.
• StringIteratorNext is also too large to be inlined by default, which means
  scalar replacement of the iterator and the result object doesn't work.

Comment 2

[André Bargull [:anba]]

Attachment: Bug 1874373 - Part 2: Use faster code point conversion. r=jandem!

Reorder the operations to match js::unicode::UTF16Decode, which allows the
frontend to constant fold the right most operand into a single constant. This
also saves two subtractions:

  (first - 0xd800) * 0x400 + (second - 0xdc00) + 0x10000
= (first * 0x400) - (0xd800 * 0x400) + (second - 0xdc00) + 0x10000
= (first * 0x400) + second - (0xd800 * 0x400) - 0xdc00 + 0x10000
= (first << 10) + second + (0x10000 - (0xd800 << 10) - 0xdc00)

Depends on D198360

Comment 3

[André Bargull [:anba]]

Attachment: Bug 1874373 - Part 2: Call String.p.codePointAt in StringIteratorNext. r=jandem!

Directly call String.p.codePointAt instead of manually computing the code
point in StringIteratorNext.

Depends on D198360

Comment 4

[André Bargull [:anba]]

Attachment: Bug 1874373 - Part 3: Call String.p.codePointAt in AdvanceStringIndex. r=jandem!

Directly call String.p.codePointAt instead of manually computing the code
point in AdvanceStringIndex. This also makes AdvanceStringIndex small
enough to be inlined by default.

Depends on D199350

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/26beade9b70b
Part 1: Mark StringIterator.p.next as inlinable. r=jandem
https://hg.mozilla.org/integration/autoland/rev/43b98212a7d0
Part 2: Call String.p.codePointAt in StringIteratorNext. r=jandem
https://hg.mozilla.org/integration/autoland/rev/e48b625ffa25
Part 3: Call String.p.codePointAt in AdvanceStringIndex. r=jandem

Comment 6

[Narcis Beleuzu [:NarcisB]]

https://hg.mozilla.org/mozilla-central/rev/26beade9b70b
https://hg.mozilla.org/mozilla-central/rev/43b98212a7d0
https://hg.mozilla.org/mozilla-central/rev/e48b625ffa25