203151 - array performance test problem using reals as indices

Status: VERIFIED INVALID

(Core :: JavaScript Engine, defect)

Description

[ash]

User-Agent:       Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.3) Gecko/20030313
Build Identifier: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.3) Gecko/20030313

As i mentioned in bug 171262, I'm doing some performance testing of spidermonkey
1.5 rc5.  One of my tests creates a 10*1024*1024 element array with integer
indices.  It uses 400M, takes minutes, but eventually completes successfully.  A
similar test uses real indices, i.e., numbers like i + .1, where i is an
integer, and creates an array with only 1024*1024 elements.  It says:
[ash@localhost hash-time]$ ~/code/js/src/Linux_All_DBG.OBJ/js test.js
test.js:4: out of memory
after using only about 100M on the same system, i.e., when it's not true that
there's no more memory.

It seems like this test case should complete successfully, since it has fewer
elements, so i think this is a bug.

Reproducible: Always

Steps to Reproduce:
1.run attachment in the js shell
2.
3.

Actual Results:  
you get the error message "out of memory" when you're not out of memory.

Comment 1

[ash]

Attachment: js shell test case that causes an inappropriate out of memory error

Comment 2

[ash]

Comment on attachment 121501 [details]
js shell test case that causes an inappropriate out of memory error

sorry, wrong file

Comment 3

[ash]

Attachment: a js shell test case that causes an inappropriate out of memory error

Comment 4

[ash]

cc'ing brendan since he seemed to express some interest in my other experiments

Comment 5

[Phil Schwartau]

This testcase thrashes my WinNT box around quite a bit. I actually got
the Windows warning "Your system is running low on memory", after about
three minutes, before getting any JS "out of memory" error.

Comment 6

[Brendan Eich [:brendan]]

This isn't a valid bug.  First, when JS says "out of memory", it has hit a null
failure from malloc.  Your OS may limit your process size to less than available
physical memory.

Second, using real numbers (IEEE doubles) as array indices, per ECMA-262,
requires turning those indices into strings, which chews up lots more memory
than is saved by the factor of 10 savings in number of elements.

We're not going to optimize things here, so this could be WONTFIX or even
WORKSFORME, but I think the bug is invalid as stated in comment 0.

/be

Comment 7

[Phil Schwartau]

Marking Verified.

[e-mail]: thank you for this report and your testcases;
this is an interesting question -