759978 - Inline property assignment after normal property assignment in constructor (much) slower on TI/JM

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Shu-yu Guo [:shu]]

Attachment: test case

At a glance the definite property analysis fails for |C|, but it's unclear why removing

  this.cval = 0;

from the constructor speeds it up.

Comment 1

[Shu-yu Guo [:shu]]

5 seems to be the magic number of properties that triggers this.

Comment 2

[Brian Hackett [Laid off!]]

This test is just allocating objects in a loop so the main bottleneck is the giant number of page faults we will take.  The objects have 4 properties without the .cval assign, and 5 with it.  That bumps the objects into a higher size class and they will increase from 12 words (on 32 bit, 4 header + 8 fixed slots) to 20 words (4 header + 16 fixed slots) and we will take a lot more faults.

I get about the same times for C() and D().  What engine(s) and running modes are you looking at, and what are the times you get?

Comment 3

[Shu-yu Guo [:shu]]

64bit debug build on Linux:

$ ./js -n -m objwtf.js 
481
12

Comment 4

[Shu-yu Guo [:shu]]

err, 64bit debug JM/TI on m-i tip.

Comment 5

[Brian Hackett [Laid off!]]

Hmm, what do you get with an opt build?  (debug builds can be pretty flaky for doing perf comparions.)  And is this on the version of the test where C() and D() both define just four properties?

Comment 6

[Shu-yu Guo [:shu]]

Sorry I'm being sloppy. I get similar slowdown in both opt and debug builds for the version with *5* properties on both C and D.

Comment 7

[Shu-yu Guo [:shu]]

Attachment: WIP

It is definitely caused in part by |AnalyzeNewScriptProperties| not understanding inline assignments like |this.x = this.y = 0|, as the 'this' values are pushed but not used immediately, which causes the function to bail out and mark nothing as definite.

Callgrind shows most of the time being spent in the SetProp IC stub due to properties not being marked definite for both the 5 and 4-properties versions. The reason for the 5-properties version being much slower than the 4-properties version even when the IC stub keeps getting hit must be the faults that bhackett mentioned.

This WIP patch tries to teach |AnalyzeNewScriptProperties| to understand multiple inline property assignments on 'this'.

Comment 8

[Shu-yu Guo [:shu]]

Attachment: WIP

Reorder functions to make patch smaller.

Comment 9

[Shu-yu Guo [:shu]]

Attachment: patch

new version

Comment 10

[Shu-yu Guo [:shu]]

Brian,

Why do non-SETPROP uses of 'this' bail out the entire |AnalyzeNewScriptProperties| function instead of be ignored? Why should something like:

  this.x = 0;
  this.y = this.x;

not mark this.y as a definite property? I feel like I'm missing some soundness issue.

Comment 11

[Brian Hackett [Laid off!]]

Definite properties need to describe not just properties that an object will eventually have, but those it will have before anything actually tries to access the property.  So AnalyzeNewScriptProperties bails out when it finds a use of 'this' which could escape and cause properties of the 'this' object to be accessed in unknown ways.

This shouldn't conflict with this patch, you just need to make sure that when the constructor accesses this.f that this.f is definitely a property of the object by that point.

Comment 12

[Shu-yu Guo [:shu]]

Attachment: patch

Final-ish version.

This patch refactors |AnalyzeNewScriptProperties| to process popped 'this' values in FIFO (stack) order instead of linear PC order.

The main refactoring is that there is a "producer" loop in |AnalyzeNewScriptProperties| that pushes 'this' uses onto a stack and a "consumer" loop in |AnalyzePoppedThis| that processes them all in FIFO order, which results in appending SETPROPs to |initializerList| in PC order.

Re: complexity concerns, I don't think this adds any more complexity than the previous case, and processes the 'this' in the natural order that it's used in the bytecode anyways.

The ordering constraints that check for 'this' escaping via eval and the like are intact -- the notion of |lastThisPopped| now generalizes to the offset of the bottom of the 'this' stack everytime that stack is processed, which is (assuming non-crazy malformed bytecode) the largest 'this' pop offset that we've processed.

More could be done in general to analyze definite properties, of course, but this patch just fixes this single thing.

Comment 13

[Brian Hackett [Laid off!]]

Comment on attachment 629899 [details] [diff] [review]
patch

Review of attachment 629899 [details] [diff] [review]:
-----------------------------------------------------------------

Looks great!

Comment 14

[Shu-yu Guo [:shu]]

http://hg.mozilla.org/integration/mozilla-inbound/rev/f56e2197d9cd

Comment 15

[Paul Wright]

This should not be marked RESOLVED FIXED, as it has not been merged to M-C yet.  Also, this patch significantly regressed 64-bit Kraken Gaussian Blur on AWFY.

Comment 16

[Paul Wright]

Correction:  The patch actually has been merged to M-C.  It still regresses Kraken, though.

Comment 17

[Geoff Lankow (:darktrojan)]

https://hg.mozilla.org/mozilla-central/rev/f56e2197d9cd

Comment 18

[Shu-yu Guo [:shu]]

I can't reproduce the slowdown locally for 64-bit imaging-gaussian-blur. That benchmark doesn't even call the function that this patch changed.

Any suggestions?

Comment 19

[Paul Wright]

I have no suggestions for you on this one, but the regression remains and is consistent on AWFY.

Comment 20

[Brian Hackett [Laid off!]]

Several kraken tests and at least one SS test are very sensitive to cache behavior and can change in performance dramatically for no good reason.  Since the shell is deterministic that change will persist across runs until some other random commit changes the behavior again.  This stuff generally won't reproduce on other machines or in the browser.

Comment 21

[Paul Wright]

Filed Bug 762022 and marked blocking this one.