When the cycle collector decides an object is live due to a marked JS object, it is useful to be able to see why the GC decided the marked JS object is alive. We can do this by taking a JS heap dump at the same time as the CC heap dump.

There's something weird about this patch. It works fine for normal CCs, but hangs the browser when you manually invoke a CC using the button in about:memory. Another problem with this patch is that it dumps the GC heap every time there is a CC, where we really only want it when we do a CC dump at the same time.

Comment on attachment 576507 [details] [diff] [review] Patch Looks reasonable to me. It would be kind of nice to have a way to dump the CC graph without dumping a GC graph, but I guess either way people are going to have tons of log files to delete, and any solutions I can think of would be annoying. Thanks for picking this up!

Review ping

Unbitrotted.

Comment on attachment 588053 [details] [diff] [review] Patch Andrew, want to steal this review?

Sure I can look at it. At a glance here, your inclusion of xpcprivate in the cycle collector seems sketchy. Is that going to make Ms2ger sad?

Yes, it will make him sad.

Yes. Just add a xpc::DumpGCHeap to xpcpublic.h. And use nsXPConnect::GetXPConnect instead of do_GetService("@mozilla.org/js/xpc/ContextStack;1"); inside xpconnect

Comment on attachment 588053 [details] [diff] [review] Patch Yeah, Ms2ger's idea sounds better. I can look into this if you don't have the time, Kyle. Or are on vacation. ;)

If you could pick this up and run with it it would be greatly appreciated.

Thanks for the suggestion, Ms2ger, it works a lot better than various hacky things I tried before to make this work!

Comment on attachment 594011 [details] [diff] [review] add xpc::DumpJSHeap(), call it when doing a CC dump bholley for the XPConnect changes, smaug for the cycle collector changes. All of these things have to be wrapped in #ifdef DEBUG because js::DumpHeapComplete is likewise wrapped. The ++gLogCounter had to be hoisted out of the sprintf to ensure it is always called with or without the DEBUG. I pushed a build to try on Linux to see if it builds with opt.