1499684 - Accelerate text node and attribute value serialization using SIMD

Status: RESOLVED FIXED

(Core :: DOM: Core & HTML, enhancement, P3)

Description

[Henri Sivonen (:hsivonen)]

When we serialize an attribute value in the innerHTML getter, we skip past Latin1 or UTF-16 code units that aren't any of ", & or the no-break space.

When we serialize a text node, we skip past code units other than <, >, & and the no-break space.

We should add a SIMD code paths for skipping over code points except the above 3 or 4, both as 8-bit units and as 16-bit units. The code paths should probably use the pcmpestri instruction when SSE 4.2 is available.

(See https://github.com/BurntSushi/rust-memchr/blob/master/src/x86/sse42.rs for inspiration. Per https://github.com/BurntSushi/rust-memchr/issues/38, adding all the variants we need is out of scope for the memchr crate itself.)

Comment 1

[Henri Sivonen (:hsivonen)]

The code added in bug 1499682 has the right shape for this.

Comment 2

[Henri Sivonen (:hsivonen)]

Attachment: Bug 1499684 - Accelerate text node and attribute value serialization using SIMD.

Comment 3

[Henri Sivonen (:hsivonen)]

https://treeherder.mozilla.org/jobs?repo=try&landoCommitID=162300

Comment 4

[Henri Sivonen (:hsivonen)]

Days since char being signed caused a bug: 0

Let's try again:
https://treeherder.mozilla.org/jobs?repo=try&landoCommitID=162307

Comment 5

[Henri Sivonen (:hsivonen)]

Benchmark: https://hsivonen.com/test/moz/serializer_bench.html

Comment 6

[Henri Sivonen (:hsivonen)]

Mac build: https://treeherder.mozilla.org/jobs?repo=try&landoCommitID=162322

Comment 7

[Henri Sivonen (:hsivonen)]

On Skylake, I see a 23% reduction in execution time for attribute value or text node that does not need any escapes and a 41% reduction in execution time for an attribute value or text node that needs one escape.

Comment 8

[Henri Sivonen (:hsivonen)]

NOTE TO SHERIFF: This landing may change the exact error message for x86 Windows noopt linking failure that's bug 1998532. I'll fix that in bug 1998532.

Comment 9

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/c553709e5cfb
https://hg.mozilla.org/integration/autoland/rev/d7607ac34210
Accelerate text node and attribute value serialization using SIMD. r=smaug

Comment 10

[Henri Sivonen (:hsivonen)]

M3 Pro
Attribute value or Unicode text node without escape: 13% reduction
Latin1 text node without escape: 24% reduction
Attribute value with single escape: 59% reduction
Text node with single escape: 63% reduction

Might be worthwhile to complicate the counting loop on aarch64 by reducing on every 15th iteration (to avoid overflow) instead of reducing on every iteration.

Comment 11

[Henri Sivonen (:hsivonen)]

(In reply to Henri Sivonen (:hsivonen) from comment #10)

Might be worthwhile to complicate the counting loop on aarch64 by reducing on every 15th iteration (to avoid overflow) instead of reducing on every iteration.

Bug 1998635.

Comment 12

[Cosmin Sabou [:CosminS]]

https://hg.mozilla.org/mozilla-central/rev/d7607ac34210