1841620 - Improve inlined nursery allocation path

Status: RESOLVED FIXED

(Core :: JavaScript: GC, enhancement, P3)

Description

[Jon Coppeard (:jonco)]

This bug is for improvements that follow on from the work in bug 1838629.

Comment 1

[Jon Coppeard (:jonco)]

Attachment: Bug 1841620 - Part 1: Teach the compiler that successful nursery allocation cannot result in a nullptr r?sfink

This removes two conditional branches per inlined allocation from the generated
code as best I can make out.

Comment 2

[Jon Coppeard (:jonco)]

Attachment: Bug 1841620 - Part 3: Avoid loading thing size at runtime where possible r?sfink

We don't need to load the size from an array if we have it available at compile
time, like we do for strings and big ints.

Comment 3

[Julian Seward [:jseward]]

Attachment: Some perf numbers

This is a clear win for wasm-gc struct/array allocations. +1 to land.

Comment 4

[Jon Coppeard (:jonco)]

Attachment: Bug 1841620 - Part 2: Make AllocKind functions constexpr where possible r?sfink

Not related to this bug (and doesn't make any difference to the generated
code), but the next patch requires IsObjectAllocKind to be constexpr.

Depends on D182723

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/73e84f16c18d
Part 1: Teach the compiler that successful nursery allocation cannot result in a nullptr r=sfink
https://hg.mozilla.org/integration/autoland/rev/7eba8591b7e7
Part 2: Make AllocKind functions constexpr where possible r=sfink
https://hg.mozilla.org/integration/autoland/rev/53a47bf7757c
Part 3: Avoid loading thing size at runtime where possible r=sfink

Comment 6

[Noemi Erli[:noemi_erli]]

https://hg.mozilla.org/mozilla-central/rev/73e84f16c18d
https://hg.mozilla.org/mozilla-central/rev/7eba8591b7e7
https://hg.mozilla.org/mozilla-central/rev/53a47bf7757c