Oh, such magnificent deletion.

Comment on attachment 400632 [details] [diff] [review] Chop, chop, chop. This makes the tempAlloc even more important. Working on it.

Backed out, perf regression. Will hunt tomorrow.

This update corrects the leak that was (as far as I can tell) the source of the regression. Tracker pages weren't being allocated in the Allocator.

Comment on attachment 400927 [details] [diff] [review] Update the patch This is moving a ton of data into the allocator. We really need that tempAlloc. Feel free to steal that bug from me. I was thinking codeAlloc/dataAlloc and then tempAlloc which is a member of TraceRecorder and NativeRegExpCompiler. For that those would have to stay new/delete allocated. Or we create the temp allocator dynamically and throw it away in the old destructor/now thrashTrees().

http://hg.mozilla.org/tracemonkey/rev/c256a60dd22c Landed a more-conservative subset of the patch that only removes MMgc-ness, doesn't move anything into allocators. Uses operator new and calloc explicitly. Can play with allocators later.

valgrind complains about a mismatched calloc / operator delete. I think technically it's justified, since operator delete is not required to just call std::free. The patch adds an explicit delete that calls free.

Luke: Jesse already filed, I'm fixing pending r? from gal, see bug 517299.

Comment on attachment 401904 [details] [diff] [review] match user-defined operator new with delete Landed elsewhere (bug 517299), thanks though!