470133 - TM: fails to trace condswitch with type mismatch

Status: VERIFIED FIXED

(Core :: JavaScript Engine, defect)

Description

[Jesse Ruderman]

20% slower with -j than without:

three=3; for(i=0;i<500000;++i) { switch("") { case three: } }

Comment 1

[Jesse Ruderman]

Running with TRACEMONKEY=verbose, it says:

recording starting from /Users/jruderman/e.js:1@20
globalObj=0x24f000, shape=143
import vp=0x811334 name=$global0 type=int flags=0
abort: 4590: unsupported type for cmp vs string
Abort recording (line 1, pc 27): JSOP_CASE.

I believe switch uses === instead of ==, so why is this using a cmp that becomes unhappy when the types don't match?

Comment 2

[Jeff Walden [:Waldo]]

Attachment: Patch

I can't figure out how to demonstrate misbehavior caused by using == instead of === in comparing a case value with the switch value.  I think this might be caused by the way that currently case returns from ::equality before getting to the fusing/setting code, but I don't know that for sure.

Comment 3

[Jeff Walden [:Waldo]]

This should block, I think, because it makes us fall off trace and because the wrong comparison function is used; I was unsuccessful at getting semantically different behavior with the wrong comparison function, but that doesn't mean it wasn't present.  Better not to depend on being lucky somehow, methinks.

Comment 4

[Jeff Walden [:Waldo]]

Hmm, forgot it wasn't blocking yet, but fixed in TM:

http://hg.mozilla.org/tracemonkey/rev/cc3cbf3ea676

Comment 5

[Robert Sayre]

merged to mc

Comment 6

[Robert Sayre]

http://hg.mozilla.org/releases/mozilla-1.9.1/rev/180f5d4290f4

Comment 7

[Bob Clary [:bc] (inactive)]

test included in js1_8_1/trace/trace-test.js 
http://hg.mozilla.org/mozilla-central/rev/8f967a7729e2

Comment 8

[Bob Clary [:bc] (inactive)]

v  1.9.1, 1.9.2