777564 - IonMonkey: fix math-partial-sums performance

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

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

Attachment: benchmark

Benchmark that runs for 100X the time, shows that IonMonkey is indeed slower. Without this change, we're about 3ms slower on my machine (11ms to 14ms).

JM+TI: 826ms
JM:    869ms
V8:    618ms

Comment 1

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

Err, that "JM:" should read "Ion:"

Comment 2

[Sean Stangl [:sstangl]]

Resolving Bug 777570 brings the numbers to (local for me):

JM+TI: 803ms
Ion:   770ms
V8:    734ms

Comment 3

[Paul Wright]

Are the numbers from comment 0 and comment 2 from different platforms?  (I'm only asking because the scales are different.)

Comment 4

[David Mandelin [:dmandelin]]

AWFY does show IM now faster than JM+TI. V8 is still 2ms faster. Any idea why?

Comment 5

[Sean Stangl [:sstangl]]

Results after landing Bug 772892 have us being faster than JM_TI. On the benchmark dvander attached in Comment 0, results on x86_64 are as follows:

JM+TI: 805ms
Ion:   718ms
V8:    727ms

Numbers are rough averages from a bunch of runs. We're still slower than V8 on the actual benchmark, but the numbers suggest that this may be because we spend time running in JM. We could make up time by attempting to further rice Ion, but the generated code already looks very good for this benchmark.

Comment 6

[David Mandelin [:dmandelin]]

Nice.

Comment 7

[Jan de Mooij [:jandem]]

We're now indeed faster than JM+TI and V8 on AWFY.