[LLVMdev] Proposal: type uniquing of debug info for LTO

[David Blaikie]

On Thu, Jun 20, 2013 at 5:13 PM, Manman Ren <mren at apple.com> wrote:
>
> On Jun 20, 2013, at 4:52 PM, David Blaikie wrote:
>
>> On Thu, Jun 20, 2013 at 4:45 PM, Manman Ren <mren at apple.com> wrote:
>>>
>>> On Jun 20, 2013, at 3:55 PM, Manman Ren <mren at apple.com> wrote:
>>>
>>>
>>> On Jun 20, 2013, at 2:58 PM, Eric Christopher wrote:
>>>
>>> Hi Manman,
>>>
>>> On Thu, Jun 20, 2013 at 2:51 PM, Manman Ren <mren at apple.com> wrote:
>>>
>>>
>>> The intent of this proposal is to speedup compilation of "-flto -g" for c++
>>> programs.
>>> This is based on discussions with Adrian, David and Eric.
>>>
>>>
>>> Thanks for bringing this back to the list. The original thread was
>>> getting quite long.
>>>
>>> ---------------------------
>>> Problem:
>>> A single class can be used in multiple source files and the DI (Debug Info)
>>> class is included in multiple bc files. The duplication of
>>> class definitions causes blow-up in # of MDNodes, # of DIEs, leading to
>>> large memory requirement.
>>>
>>> As an example, SPEC xalancbmk requires 7GB of memory when compiled with
>>> -flto -g.
>>> With a preliminary implementation of type uniquing, the memory usage will be
>>> down to 2GB.
>>>
>>> In order to unique types, we have to break cycles in the MDNodes.
>>>
>>> A simple struct definition
>>> struct Base {
>>> int a;
>>> };
>>> can cause cycles in MDNodes:
>>> !12 = metadata !{i32 786451, metadata !13, null, metadata !"Base", i32 1,
>>> i64 32, i64 32, i32 0, i32 0, null, metadata !14, i32 0, null, null} ; [
>>> DW_TAG_structure_type ] [Base] [line 1, size 32, align 32, offset 0] [from ]
>>> !14 = metadata !{metadata !15, metadata !16}
>>> !15 = metadata !{i32 786445, metadata !13, metadata !12, metadata !"a", i32
>>> 2, i64 32, i64 32, i64 0, i32 0, metadata !8} ; [ DW_TAG_member ] [a] [line
>>> 2, size 32, align 32, offset 0] [from int]
>>> !16 = metadata !{i32 786478, metadata !13, metadata !12, metadata !"Base",
>>> metadata !"Base", metadata !"", i32 1, metadata !17, i1 false, i1 false, i32
>>> 0, i32 0, null, i32 320, i1 false, null, null, i32 0, metadata !20, i32 1} ;
>>> [ DW_TAG_subprogram ] [line 1] [Base]
>>>
>>> Cycles: !12 -- !14 -- !15 -- !12
>>>          !12 -- !14 -- !16 -- !12
>>>
>>> These cycles make it hard to unique the same struct used in two bc files.
>>>
>>> ---------------------------
>>> How to fix:
>>>
>>> We attach a hash value to types to help type uniquing and we also replace
>>> references to types with their hash values.
>>> For the above struct "Base", we now have the following MDNodes:
>>> !4 = metadata !{i32 786451, metadata !5, null, metadata !"Base", i32 1, i64
>>> 32, i64 32, i32 0, i32 0, null, metadata !6, i32 0, i32 0, null, i32
>>> 915398439} ; [ DW_TAG_structure_type ] [Base] [line 1, size 32, align 32,
>>> offset 0] [from ]
>>> !6 = metadata !{metadata !7, metadata !9}
>>> !7 = metadata !{i32 786445, metadata !5, i32 915398439, metadata !"a", i32
>>> 2, i64 32, i64 32, i64 0, i32 0, metadata !8} ; [ DW_TAG_member ] [a] [line
>>> 2, size 32, align 32, offset 0] [from int]
>>> !9 = metadata !{i32 786478, metadata !5, i32 915398439, metadata !"Base",
>>> metadata !"Base", metadata !"", i32 1, metadata !10, i1 false, i1 false, i32
>>> 0, i32 0, null, i32 320, i1 false, null, null, i32 0, metadata !13, i32 1} ;
>>> [ DW_TAG_subprogram ] [line 1] [Base]
>>>
>>> Note that the cycles are gone and !4 has a hash value of 915398439, and the
>>> references to !4 are replaced with 915398439.
>>> Thanks Eric for suggesting replacing MD reference with a hash value.
>>>
>>>
>>> In particular I recommended this:
>>>
>>> a) For C++ odr replace it with the "hash" that's just a string
>>> representing the type name.
>>> b) For Internal C++ types and all C types replace it with a string
>>> that's a concatenation of the type name and the name of the compile
>>> unit.
>>>
>>> Yes, that is what we agreed on over email.
>>>
>>>
>>>
>>> There are a few issues:
>>> 1> How to generate the hash for a given type?
>>> With C++'s ODR, it should be enough by using the context and the name for
>>> non-internal c++ types.
>>> For internal c++ types and types of other languages, hash will not be used.
>>>
>>>
>>> Explain this?
>>>
>>>
>>> For a while, I am going to support both hash and MD reference, once
>>> everything is working with hash,
>>> I will update all debug info testing cases, turn -gtype-uniquing on, and
>>> remove the other path.
>>>
>>> For internal c++ types, initially they will follow the path of using MD
>>> references without a hash.
>>>
>>>
>>> My current implementation is to add a few static member functions in MDNode
>>> to profile DI nodes differently.
>>> + /// If the array of Vals is for debug info, profile it specially and
>>> return true.
>>> + /// If the DI node has a hash value, generate the profile using only the
>>> hash value and the declaration flag.
>>> +  static bool profileDebugInfoNode(ArrayRef<Value*> Vals, FoldingSetNodeID
>>> &ID);
>>>
>>> + /// If the MDNode is for debug info, profile it specially and return true.
>>> + /// If the DI node has a hash value, generate the profile using only the
>>> hash value and the declaration flag.
>>> +  static bool profileDebugInfoNode(const MDNode *M, FoldingSetNodeID &ID);
>>>
>>> + /// Given a hash value and a flag, generate the profile for later lookup.
>>> +  static bool profileDebugInfoNode(unsigned Hash, bool Declaration,
>>> FoldingSetNodeID &ID);
>>>
>>> These static functions are called in Metadata.cpp:
>>> void MDNode::Profile(FoldingSetNodeID &ID) const {
>>> +  if (profileDebugInfoNode(this, ID))
>>> +    return;
>>> +
>>>
>>> There are other examples of these in MDNode for handling of specific
>>> metadata.
>>> /// Methods for metadata merging.
>>> static MDNode *getMostGenericTBAA(MDNode *A, MDNode *B);
>>> static MDNode *getMostGenericFPMath(MDNode *A, MDNode *B);
>>> static MDNode *getMostGenericRange(MDNode *A, MDNode *B);
>>>
>>> Comments are welcome on whether this violates any layering rule.
>>>
>>>
>>> As I've said many times in email, I don't think this is a good idea
>>> and would prefer either a or b below. a) is a much simpler solution.
>>>
>>> Any reason that why it is not a good idea?
>>>
>>>
>>> Other choices are:
>>> a> Keep a map in DwarfDebug
>>>  Keep in mind that the map is used at many stages, and it has to be in sync
>>> with MDNodeSet.
>>> b> Generalize MDNode to be aware of hash (David can provide more details)
>>> c> Extend MDNode to DINode and modify streamers (bitcode reader|writer, ll
>>> reader|writer) to be aware of DINode
>>>  We can provide DINode::get(…) to create a DINode. DINode can have its own
>>> Profile function.
>>> Other suggestions are welcome.
>>>
>>>
>>> a or b please.
>>>
>>> Option a> will require a DwarfDebug pointer in every stage of the compiler,
>>> and passing the map to the DI classes.
>>> A rough estimation is around 100 places.
>>> Is it reasonable to pass a DwarfDebug pointer to DIBuilder and llvm linker?
>>> Also the map needs to be in sync with MDNodeSet, maybe using ValueHandle can
>>> solve the problem.
>>>
>>>
>>> What about putting the map in LLVMContextImpl?
>>> It already has a few things specifically for debug info:
>>> std::vector<DebugRecVH> ScopeRecords;
>>> DenseMap<std::pair<MDNode*, MDNode*>, int> ScopeInlinedAt;
>>> …
>>>
>>> I remember David mentioned it once and I forgot about the conclusion.
>>
>> I mentioned it only as speculation as to how you were implementing it
>> already (but you were doing the profile-changing stuff).
>>
>> I don't think it should be necessary to have the map (in option (a))
>> in such a central location as LLVMContext. It should be usable just
>> from DwarfDebug for generation, and DIBuilder can have its own,
>> separate map to do similar things during DI building.
>
> We also need the map during llvm linking since linking will create new MDNodes.

I don't understand what you mean by this. IR linking shouldn't need to
do anything debug-info-specific, it should just be the normal IR
approach.

The declaration-v-definition resolution can be done at codegen-time.
We'll have to walk all the lists of retained types anyway to build the
map, so we can do declaration-v-definition (keep definitions over
declarations when we see both) at that point.

> So any other opinion on putting it in LLVMContext other than it being central?
>
> Thanks,
> Manman
>
>>
>>>
>>> Thanks,
>>> Manman
>>>
>>>
>>> More details for option b from David
>>>
>>> < The alternative I have in mind is a more complete version of what
>>> < you're proposing - a full MD feature, not an MD feature that's just
>>> < barely enough to support the needs of debug info. What we could do is
>>> < allow the insertion of these MDHash things you spoke about but take it
>>> < a step further and have MDNode::getOperand walk through the hash &
>>> < give the value (in this way, DebugInfo wouldn't have to change at all
>>> < to handle hashes - if the Metadata APIs are going to be aware of the
>>> < hashes anyway, they might as well provide this convenience
>>> < functionality) the metadata feature would also have to have some
>>> < blessed top-level named metadata that would have a list of hash+MDNode
>>> < to keep those MDNodes alive (so you wouldn't have to stuff all the
>>> < types in the retained types list - metadata would provide the full
>>> < support, not just half of it).
>>>
>>>
>>>
>>> Transition from current DI Metadata:
>>> To have a smooth transition, I will add a flag "-gtype-hashing" for the type
>>> uniquing work and turn it on by default when we are ready.
>>>
>>>
>>> I'd prefer just make the change to have the front end emit the "hash"
>>> (it's not really a hash, it's just a string).
>>>
>>> Are you saying no transition period? A single patch to have correct handling
>>> of "hash" and to update all existing testing cases?
>>>
>>>
>>> -----------------------------
>>> Patches:
>>> Expect the following patches:
>>> 1> add flag -gtype-hashing
>>> 2> add hash field to DI types
>>> 3> modify DIBuilder to use hash instead of MD reference
>>> 4> related to issue 3
>>>
>>>
>>> These can all be a single patch since it shouldn't be very large if we
>>> go with a) above. If we go with b) then the MDNode work should be done
>>> in isolation first and then the debug info on top of it.
>>>
>>> What is wrong with smaller patches?
>>> My estimation for all the above with a) is about 30K + testing cases.
>>>
>>>
>>> 5> backend change (in DwarfDebug|CompileUnit) to support types shared among
>>> compile units
>>>  requires gdwarf-2 gdwarf-3 gdwarf-4 support for issues related to ref_addr
>>>
>>>
>>> #5 can and should be done before the rest of them.
>>>
>>> I prefer to submit patches according to the flow of the compiler, starting
>>> from the frontend, then IR, then backend.
>>> The testing cases will be added for front end, llvm-link and backend.
>>> Any reason why #5 should be done first?
>>>
>>>
>>> All changes should be local to debug info classes except patch #4.
>>>
>>>
>>> What's patch #4?
>>>
>>> Patch #4 above: related to issue 3 (changes corresponding to how to solve
>>> issue #3)
>>>
>>> -Manman
>>>
>>>
>>> -eric
>>>
>>>
>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>>
>>>
>>>
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>
>