Comment Re:Hmmm (Score 5, Informative) 134
I've done some work before that required concatenating very large strings in Javascript, and the code was certain to be run in IE by our users. It was clear to me from the observed performance, that when you use the "+" or "+=" operators to build a large string from many smaller strings, IE does not attempt to optimize things under the hood with a String Buffer. In fact, *every* "+/+=" operation seems to allocate an entirely new String object to contain the result. Problem is, most of those Strings are intermediate results that will just end up get tossed by the garbage collector, and hence their allocation was for naught. A quick analysis of this means that we're looking at a total memory allocation cost on the order of N*N, which leads to very unreasonable performance when you're looking at concatenating 1000+ strings. From scanning the SunSpider string benchmarks, it looks like the specific benchmarks that IE has trouble with are indeed using the +/+= ops to build big Strings.
Fortunately, there are workarounds. The easiest is put all the component strings you wish to add together into an array and then call join('') on the array to build the result string. This executes much, much faster in IE and should be fast in the other browsers as well. Before I figured out the join('') "trick", I implemented a scheme to add the component strings in a much more optimal order...the N^2 cost is only if you add them in a naive, left-to-right manner. By pairing up adjacent components and adding, you get N/2 intermediate results. Repeating the pairing process on these intermediates, and over again until you have your 1 final result, the allocation cost is only N*Log(N), which is far FAR better than N*N. Incidentally, this is similar to the "Russian Peasant Algorithm" technique used to optimize the addition of N numbers on massively parallel processors...whereby you can reduce the cost of N additions from N to Log(N) time assuming there are enough processors (N/2) on hand. When I learned about the join(), I compared its performance to the Russian Peasant technique and got very comparable results in IE. Weired, since if join() were using a StringBuffer under the hood (like it should), then its performance should be O(N), not N*Log(N)...
Anyways, in summary - yeah it sucks hard that IE does that, but it's easy to workaround once you know what's up (gee, that's true in so many cases with IE quirks).
Fortunately, there are workarounds. The easiest is put all the component strings you wish to add together into an array and then call join('') on the array to build the result string. This executes much, much faster in IE and should be fast in the other browsers as well. Before I figured out the join('') "trick", I implemented a scheme to add the component strings in a much more optimal order...the N^2 cost is only if you add them in a naive, left-to-right manner. By pairing up adjacent components and adding, you get N/2 intermediate results. Repeating the pairing process on these intermediates, and over again until you have your 1 final result, the allocation cost is only N*Log(N), which is far FAR better than N*N. Incidentally, this is similar to the "Russian Peasant Algorithm" technique used to optimize the addition of N numbers on massively parallel processors...whereby you can reduce the cost of N additions from N to Log(N) time assuming there are enough processors (N/2) on hand. When I learned about the join(), I compared its performance to the Russian Peasant technique and got very comparable results in IE. Weired, since if join() were using a StringBuffer under the hood (like it should), then its performance should be O(N), not N*Log(N)...
Anyways, in summary - yeah it sucks hard that IE does that, but it's easy to workaround once you know what's up (gee, that's true in so many cases with IE quirks).
