1664786 - Warp: tweak inlining heuristics

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, task, P2)

Description

[Jan de Mooij [:jandem]]

I have some patches to improve trial inlining.

Comment 1

[Jan de Mooij [:jandem]]

Attachment: Bug 1664786 part 1 - Be more conservative with trial inlining depth and recursion. r?iain!

Comment 2

[Jan de Mooij [:jandem]]

Attachment: Bug 1664786 part 2 - Don't trigger Ion compilation or OSR in trial-inlined scripts. r?iain!

Depends on D90224

Comment 3

[Jan de Mooij [:jandem]]

Attachment: Bug 1664786 part 3 - Don't trial inline scripts that entered Warp via OSR. r?iain!

To prevent getting stuck in Baseline because we can't tier up in trial-inlined
scripts.

Depends on D90225

Comment 4

[Jan de Mooij [:jandem]]

Attachment: Bug 1664786 part 4 - Set initial warm-up threshold for trial-inlined scripts. r?iain!

This makes inlining more effective because we can inline more before Warp compilation
happens.

Depends on D90226

Comment 5

[Jan de Mooij [:jandem]]

Attachment: Bug 1664786 part 5 - Reset warm-up counts in inlined ICScripts too. r?iain!

Depends on D90227

Comment 6

[Jan de Mooij [:jandem]]

Attachment: Bug 1664786 part 6 - Limit total bytecode size when trial inlining. r?iain!

To help prevent badness (security, perf, memory usage) in pathological cases.

Depends on D90228

Comment 7

[Jan de Mooij [:jandem]]

Attachment: Bug 1664786 part 7 - Iterate over IC entries list. r?iain!

Later on we could remove the BytecodeLocation/JSOp use too and replace it with a
tag/flag on trial-inliable stubs.

Depends on D90229

Comment 8

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/e124e2c9e4ca
part 1 - Be more conservative with trial inlining depth and recursion. r=iain
https://hg.mozilla.org/integration/autoland/rev/c4379a121a32
part 2 - Don't trigger Ion compilation or OSR in trial-inlined scripts. r=iain
https://hg.mozilla.org/integration/autoland/rev/149d67f2d77c
part 3 - Don't trial inline scripts that entered Warp via OSR. r=iain
https://hg.mozilla.org/integration/autoland/rev/cfb6ef22174c
part 4 - Set initial warm-up threshold for trial-inlined scripts. r=iain
https://hg.mozilla.org/integration/autoland/rev/84cc1114fc4f
part 5 - Reset warm-up counts in inlined ICScripts too. r=iain
https://hg.mozilla.org/integration/autoland/rev/f5f7119a80ac
part 6 - Limit total bytecode size when trial inlining. r=iain
https://hg.mozilla.org/integration/autoland/rev/d8078e1dea25
part 7 - Iterate over IC entries list. r=iain

Comment 9

[Jan de Mooij [:jandem]]

Attachment: Bug 1664786 part 8 - Add TrialInliningState to the fallback stub. r?iain!

This patch adds an enum class to track the inlining state per fallback stub and
uses it for the following:

• Don't scan CacheIR of stubs we definitely can't inline (will matter more for getters/setters).
• Similarly, optimize the check for trial-inlined stubs in WarpOracle.
• Assert in WarpOracle that we didn't trial-inline a stub the transpiler doesn't support.
• Ensure we never attempt a second trial-inlining at the same location.

Comment 10

[Cristina Coroiu [:ccoroiu]]

https://hg.mozilla.org/mozilla-central/rev/e124e2c9e4ca
https://hg.mozilla.org/mozilla-central/rev/c4379a121a32
https://hg.mozilla.org/mozilla-central/rev/149d67f2d77c
https://hg.mozilla.org/mozilla-central/rev/cfb6ef22174c
https://hg.mozilla.org/mozilla-central/rev/84cc1114fc4f
https://hg.mozilla.org/mozilla-central/rev/f5f7119a80ac
https://hg.mozilla.org/mozilla-central/rev/d8078e1dea25

Comment 11

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/924265487b20
part 8 - Add TrialInliningState to the fallback stub. r=iain

Comment 12

[Alexandru Michis [:malexandru]]

https://hg.mozilla.org/mozilla-central/rev/924265487b20