1562816 - Investigate pretenuring strings by allocation site

Status: NEW

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

Description

[Steve Fink [:sfink] [:s:]]

Currently, string pretenuring is done on a whole-Zone basis (if too many strings are getting tenured, we permanently stop allocating them from the nursery and allocate directly in the tenured heap). This is a very coarse heuristic. It would be far better to be able to make the decision per allocation site. We can only do that for objects right now, because their ObjectGroup is tied to an allocation site, but strings do not have anything analogous.

https://ai.google/research/pubs/pub43823 describes V8's implementation of this. They associate a "memento" with nursery-allocated items created by the baseline compiler (only), and accumulate statistics during tenuring (at which time the memento is discarded.)

We are already carrying extra data with JSStrings (the Zone*), so it would be fairly straightforward to add a memento. One option for the memento would be script pointer + bytecode offset, or a hash derived from that. Another option would be to just generate sequential or random ids and embed them in our Baseline-compiled code only, and propagate them through to the Ion-compiled code (to choose where to allocate; Ion presumably would not continue to collect allocate site-specific counts.)

The V8 system appears to record both total allocations at a site, as well as counting up how many were tenured, so they have a true percentage tenured per allocation site.

Note that if we do deduplication during string tenuring, then we really want an allocation site that produces lots of duplicates to allocate in the nursery. We might want to factor that into the heuristic, in addition to the allocation count and tenure percentage. But that's a lesser-order consideration that only matters for moderately high tenuring percentages.

Comment 1

[Jon Coppeard (:jonco)]

I've implemented most of this but don't yet have data to show that it's a win over what we already have. I'll post the patches I have so far.

Comment 2

[Jon Coppeard (:jonco)]

Attachment: WIP: Bug 1562816 - Add a test for the different kinds of pretenuring

This tests whether certain kinds of allocation produce tenured or nursery
allocations (whereas existing tests check the number of major/minor GCs).

Comment 3

[Jon Coppeard (:jonco)]

Attachment: WIP: Bug 1562816 - Add allocation site parameter to GC string allocation APIs

Depends on D121949

Comment 4

[Jon Coppeard (:jonco)]

Attachment: WIP: Bug 1562816 - Add allocation site parameter to string concatenation APIs

Depends on D121950

Comment 5

[Jon Coppeard (:jonco)]

Attachment: WIP: Bug 1562816 - Add allocation stie parameter to macroassembler string allocation APIs

Depends on D121951

Comment 6

[Jon Coppeard (:jonco)]

Attachment: WIP: Bug 1562816 - Make MaybeCreateAllocSite be a method on IRGenerator

Depends on D121952

Comment 7

[Jon Coppeard (:jonco)]

Attachment: WIP: Bug 1562816 - Pretenure string concatenation (WIP)

Depends on D121954

Comment 8

[Jon Coppeard (:jonco)]

Attachment: WIP: Bug 1562816 - Create a separate stub for tenured string concatentation (WIP)

Depends on D121955

Comment 9

[Jon Coppeard (:jonco)]

Attachment: WIP: Bug 1562816 - Pretenure the result of the String constructor

Depends on D121956

Comment 10

[Jon Coppeard (:jonco)]

Attachment: WIP: Bug 1562816 - Pretenure the result of number to string conversions

Depends on D121957