New Solution for memory leak in tblgen

[wuhui1973]

Hope I can explain it clear enough :-)


> MemReclaimer has two containers to hold pointers of TreePatternNode & TreePattern which are allocated from heap.
> Only one global instance of MemReclaimer  is declared -- memReclaimer.
> operator new & operator delete are defined for TreePatternNode & TreePattern 
          new:    push the new allocated object (pointer) into containers in memReclaimer.
          delete: remove the pointer from the containers in memReclaimer.


          So we can guarantee that objects allocated from heap can always be recorded by memReclaimer. 
          And no double free may incur, as some places in the source delete some pointers manually (maybe we can remove these lines).
> So memReclaimer only keeps the objects that are leaked in current design.
> the author of CodeGenDAGPattern  create nodes, but destroy part of them (for TreePattern, the nodes destroyed manually account for more than 30%, 
          for some target even over 50%; while for TreePatternNode  only around 1%) and make the rest leaked deliberately.
> dtor of CodeGenDAGPattern now is the place to destroy all not freed objects
> move the functionality of MemReclaimer into CodeGenDAGPattern  is definitely feasible I think.


Regards
Hui Wu



At 2014-11-11 14:20:25, "David Blaikie" <dblaikie at gmail.com> wrote:

Could you describe the high level design here?

It looks like there's a static pool (CodeGenDAGPatterns.cpp::memReclaimer) of instances that's used as some kind of last-chance cleanup? Instances of TreePatternNode mostly manage their own lifetime but then if any haven't been destroyed by the time the CodeGenDAGPatterns dtor runs, the remaining elements are destroyed. This is to handle cycles, I take it?

It still seems like a bit of a GC-esque workaround to handle this case when there might be something better... but I don't know much about tablegen, perhaps there isn't.

How often are nodes destroyed by themselves? Should we just give up owning these anywhere else & move all ownership into CodeGenDAGPatterns and clean them up in CodeGenDAGPatterns' dtor? (just have a vector of unique_ptrs (maybe even a list or deque of nodes directly owned, rather than via unique_ptr), only create nodes, never destroy them, then destroy them all at the end)


On Mon, Nov 10, 2014 at 9:58 PM, wuhui1973 <wuhui1973 at 163.com> wrote:

Hi Andrew, David and Hal:


I have made a new solution for this memory leak issue, which is much simpler than previous one. 


I have tested it, it works well!


Please have a look, and appreciate any comment.


Thanks & Regards


Hui Wu






-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20141111/ae28dc8e/attachment.html>