Should dramatically reduce the memory spike on M-oth.

Comment on attachment 8396771 [details] [diff] [review] start_memory_uncommitted-v0.diff Review of attachment 8396771 [details] [diff] [review]: ----------------------------------------------------------------- ::: js/src/gc/Memory.cpp @@ +78,5 @@ > +void * > +gc::RetainAlignedAddressRange(JSRuntime *rt, size_t size, size_t alignment) > +{ > + int type = MEM_RESERVE; > + VirtualAllocAligned(rt, size, alignment, ); Seems like this won't compile - did you forget to qref? ::: js/src/gc/Memory.h @@ +20,5 @@ > InitMemorySubsystem(JSRuntime *rt); > > +// Add a mapping for an aligned address range, but do not commit the memory. > +void * > +RetainAlignedAddressRange(JSRuntime *rt, size_t size, size_t alignment); I'm not sure about the name "retain". To me this sounds like the address range was already present somehow. Maybe we could just go with AllocateAlignedAddressRange?

No, even dumber: I copied after roughing out the idea and not after fixing the bugs in the linux impl.

Comment on attachment 8397062 [details] [diff] [review] start_memory_uncommitted-v1.diff Review of attachment 8397062 [details] [diff] [review]: ----------------------------------------------------------------- Looks good! ::: js/src/gc/Nursery.h @@ +207,5 @@ > + // Ensure our usable range is committed. > + if (numActiveChunks_ > 0) { > + bool result = gc::CommitAddressRange((void *)start(), gc::ChunkSize * nchunks); > + if (!result) > + return false; If we put off setting numActiveChunks_ until after this point, I think we can avoid crashing if this fails when we are trying to grow. We just won't grow the nursery and will continue with the previous size.

Rebased and applied review comments.

I removed JS_GC_ZEAL from around POISON, so the poison patterns need to be moved under JS_CRASH_DIAGNOSTICS.

Moving the assignment down means we don't have the right value yet.

Hi terrence, sorry had to backout this change for bustage like https://tbpl.mozilla.org/php/getParsedLog.php?id=36861808&tree=Mozilla-Inbound

The M-oth spike was unrelated -- our existing decommit logic works well enough on all platforms to effectively reduce RSS by the amount we save from GGC: 10-15% across the board. Given that RSS is already down, it's not clear that this actually does anything other than complexify the logic.