517535 - Wow, the Oracle is bad!

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: ss with isGlobalSlotUndemotable always true

The Oracle is a sight for sore eyes. It's non-deterministic, it can prevent nested trees from building (see bug 459301), and as Luke pointed out this morning, not even thread safe.

But what's really annoying is that it actually improves SunSpider in most cases, so we've never had lots of motivation to sit down and write something better.

I ran some tests with parts of the Oracle disabled to see how SunSpider is affected.

Comment 1

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: ss with isStackSlotUndemotable always true

Comment 2

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: ss with isInstructionUndemotable always true

Comment 3

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: oracle completely disabled

Comment 4

[David Mandelin [:dmandelin]]

Thanks for doing this. Well, at least if it gives only a 5% speedup then if we need to get rid of it it will only cost that much. I see that disabling the oracle speeds up certain benchmarks, maybe by 1% total, so we know we could win at least 1% with a magically better algorithm.

What I would like to see is a limit study that shows the maximum performance we get on selected programs by making "perfect" demotion decisions. I'm not sure which programs to select, but bits-in-byte, bitwise-and, cube, aes, tofte, and partial-sums looks like an interesting set.

Comment 5

[Mike Shaver (:shaver -- probably not reading bugmail closely)]

What does disabling do to v8 numbers?

Comment 6

[Nicholas Nethercote [inactive]]

(In reply to comment #0)
> 
> The Oracle is a sight for sore eyes.

A "sight for sore eyes" is something good, something that you are happy to see.  I think you meant the opposite.  This confused me for a minute.

Also, attachments 2 and 3 are identical, presumably you accidentally attached one of them twice.

(I'll go back into my hole now.)

Comment 7

[David Anderson [:dvander] - inactive, e-mail if emergency]

Yup, you're right. I was looking for some other idiom but I forget what. 

Will test v8 numbers.

Comment 8

[David Anderson [:dvander] - inactive, e-mail if emergency]

v8 with 10 runs. First column is with the Oracle, second column is without.



TEST              COMPARISON            FROM                 TO             DETAILS

=============================================================================

** TOTAL **:      -                 14091.0ms +/- 1.1%   13959.4ms +/- 1.0% 

=============================================================================

  v8:             -                 14091.0ms +/- 1.1%   13959.4ms +/- 1.0% 
    crypto:       1.070x as fast      695.2ms +/- 2.0%     649.9ms +/- 0.7%     significant
    deltablue:    -                  7587.1ms +/- 2.1%    7489.8ms +/- 1.8% 
    earley-boyer: *1.012x as slow*   1839.2ms +/- 0.3%    1861.6ms +/- 1.3%     significant
    raytrace:     *1.029x as slow*   1147.2ms +/- 1.0%    1180.1ms +/- 0.7%     significant
    regexp:       1.012x as fast      946.0ms +/- 0.2%     934.9ms +/- 0.1%     significant
    richards:     1.053x as fast      777.9ms +/- 0.3%     738.7ms +/- 0.3%     significant
    splay:        ??                 1098.4ms +/- 0.5%    1104.4ms +/- 0.6%     not conclusive: might be *1.005x as slow*

Comment 9

[Luke Wagner [:luke]]

Tracer was removed