61897 - Space pigs: String.prototype.charAt and str_getProperty for s[i]

Status: VERIFIED FIXED

(Core :: JavaScript Engine, defect, P3)

Description

[Brendan Eich [:brendan]]

JS users should not be writing too many char-by-char loops over strings (they
should use regexps instead!) but if any are, then the engine will malloc 4-byte
string vectors (two chars, the indexed char and NUL) to beat the band.  Ideally,
we'd make a JSString that shared space with its super-string until the latter
needed to be finalized, then copied (if the substring was still alive too, of
course).

See bug 54743 for some results from kandrot.

/be

Comment 1

[Brendan Eich [:brendan]]

Moving out, but may fix at the same time as I fix bug 56940.

/be

Comment 2

[Brendan Eich [:brendan]]

Slipping.

Comment 3

[Edward Kandrot]

Putting comments here since bug #54742 is closed.  The speed of JS appears to be
much faster now!  The looping and math routines are faster than 4.7.6 (2x - 3x),
under Linux.  The test for credit card validation is still half the speed of
4.7.6, due to the charAt code that is a part of this bug.  There is a sample js
called Luhn in the original bug that demonstrates this problem.

If there is any need to further track down this problem, let me know, and I will
add the scripts/profiles to this bug.

Comment 4

[Brendan Eich [:brendan]]

Load-balancing.

/be

Comment 5

[Brendan Eich [:brendan]]

Not clear that this matters much to real-world JS.  Anyone have data?

/be

Comment 6

[Brendan Eich [:brendan]]

Fixed by bug 56940's patches.

/be

Comment 7

[Brendan Eich [:brendan]]

I claim this is fixed by the fix for bug 56940.  The only wrinkle left, which I
think should result in a new bug, is that if you repeated fetch s[i] for a given
value of s and i, you'll create a new dependent string for each access.  Note
well that this is no worse than what happens with the ECMA-conformant charAt
method, used instead of indexing (s.charAt(i)).

We might be able to avoid that without requiring a real property and slot to be
allocated for each such i, perhaps by atomizing the alphabet.  But I'm not going
to go there without compelling, real-world examples that need it.  Given the
non-ECMA status of s[i], I don't expect such examples to exist.

/be

Comment 8

[Phil Schwartau]

Marking Verified -