531118 - jit performance is 3x slower than interp on this testcase

Status: RESOLVED WORKSFORME

(Core :: JavaScript Engine, defect)

Description

[Boris Zbarsky [:bzbarsky]]

Attachment: Testcase

Silly testcase that just triggers us to fail shape-guards over and over, hit MAX_BRANCHES, and continue with interp, but after paying all the trace-entry/exit and guard-jumping costs.  Clearly contrived, at least in terms of the loop body being tiny, but not that dissimilar from some things I've looked at where we were slow in the recent past (apart from being a bit more egregious).

Comment 1

[Jan de Mooij [:jandem]]

-m: 33 ms
v8: 33 ms
interp: 79 ms
-m -j: 108 ms
-j: 111 ms

This should not trace; will retest after bug 580468 lands.

Comment 2

[Ryan VanderMeulen [:RyanVM]]

Interp: 203
TM: 205
JM: 35
JM+TI: 28
d8: 27

Looks like JM+TI got the attached testcase to parity with v8.