http://allievi.sssup.it/jacopone/cnontetris/ is a demo of Duetto, a C++ to js compiler. The demo spikes over 1GB of RAM once, but then looks ok. Note that it does not happen during startup, but a little after it is already running, you can sometimes see the bottom of the piece show up. I don't see the huge spike when I have the profiler running (maybe opening devtools disables something?). I also don't see it on chrome (but chrome looks less smooth when running it).

cc'ing luke who I hope hates memory spikes as much as me ;)

Probably related: while FF release (24) runs reliably smoothly, Nightly and Aurora often (1/2 the time I load the URL) get into a cycle of long (1-2s) pauses every few seconds. The profile of one of these pauses on Nightly shows all the time in IonBuilder: http://pastebin.mozilla.org/3202491. There are also GCs right around the time of the pause, which suggests some bad cycle of GC discarding JIT code. Firefox Beta is weirder: half the time, it's smooth, half the time it completely hangs when the first block becomes visible. Notes: - This is duetto (http://leaningtech.com/duetto)-generated code which means it's not like Emscripten-generated code; it's mapping C++ objects to JS objects. - The demo is using requestAnimationFrame so JIT-code discarding should be inhibited, but perhaps we discard JIT-code in some severe situations? Could we perhaps be getting into some pathological case where the type analysis is re-analyzing loops which causes us to take a lot of time and consume a lot of LifoAlloc storage allocating new MIR nodes?

njn has some techniques for investigating memory spikes that might be useful

> njn has some techniques for investigating memory spikes that might be useful There's a patch in bug 842800. It may have rotted somewhat.

WFM? The game uses almost no memory.

Yes, looks ok now.