Dromaeo uses escaped closures, so we need to PIC these to get good performance on it. See also bug 576733. Shell testcase: var x = (function() { var ret; var str = "abc"; return function() { for (var j = 0; j < 5000; ++j) { ret = str.charAt(0) ret = str.charAt(str.length - 1) } } })() var start = new Date(); var i = 0; while ((new Date()) - start < 1000) { ++i; x(); } print((i / (new Date() - start)) * 1000);

As a note, currently TM is faster than V8 on this, but not a huge amount. It's definitely faster than nitro.

tm-tip score: 7553 jm-tip score: 395

http://hg.mozilla.org/users/danderson_mozilla.com/moo/rev/f801691c33ba ^-- PIC for reads from escaped call objects. New score: 830

http://hg.mozilla.org/users/danderson_mozilla.com/moo/rev/9889018ba3cf ^-- PIC for writes to escaped call objects. New score: 2130

Shark says 28% of time is now in SlowCall, 23% in str_charAt. In the tracer, 43% is in String_getelem - this might be completely inlined. v8's score is 10,000, and according to shark it looks like it inlines charAt as well.

On the dromaeo charAt test, this patch took the score from 181.82runs/s to 1191.00runs/s. So much better, for sure. TM is at 2715.00runs/s. Do we need followup bugs on the remaining slowdown vs v8 or TM?

(In reply to comment #6) Sure. I think most of it is "make calls fast" (bug 572275) - though maybe a bug to track this specific benchmark would be good.

Makes sense. I filed bug 577036.