Proposal: type uniquing of debug info for LTO

[David Blaikie]

+1 to roughly everything Eric said.

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.
>
> ---------------------------
> 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.
>
> 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.
>
> 2> Make sure the streamers (bit code or ll writer) will output all required MDNodes.
> For MDNodes that are only used through hash values, we have to make sure they are dumped.

should read "to make sure they are not dumped" - just to clarify for
other readers: unnamed metadata is kept alive by being reference
(directly or indirectly) by named metadata. If we break the reference
by using a string identifier pair, the type metadata wouldn't exist as
it would be unreferenced in the IR sense.

> The solution currently is to add these MDNodes to retained types of the compile unit.

So we can just keep real metadata (rather than string/hash/thingy)
references to these nodes in a flat list to ensure they don't get
dropped.

> Thanks David for the suggestion.
>
> 3> To actually access the MDNode referenced via the hash value, we need to perform a lookup from the hash value to find the corresponding MDNode.
> The questions are where to store this map and how to keep it up-to-date when a MDNode is replaced.
> Each hash value should have at most one corresponding definition and at most one declaration.
>
> We need the map when building DI for verification purposes, when writing the ll file to dump the comments, at llvm linking time to update the map,
> at backend when generating the actual dwarf.
>
> 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 we've already discussed off list, both Eric & I believe this isn't
an appropriate layering & that LLVM's core IR metadata APIs cannot
reasonably have special hooks for this debug info feature.

> 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.

I'm not sure what you mean by "in sync with the MDNodeSet".

> b> Generalize MDNode to be aware of hash (David can provide more details)

My suggestion here, which I think is the only reason to really discuss
this broadly (because it would be an LLVM metadata API feature) is
roughly as follows:

An LLVM IR feature, completely independent from debug info that would
work something like:
a) a blessed named MD node root that would contain string+node pairs
(alternating elements probably, for space conservation)
b) a means to add an operand to an MD node with an "indirect"
reference - this would cause the string to be used as the value (with
some flag that it's a 'special' string) but any access of the node
through the operand list would return the underlying MDNode from the
list (lookup via a map in LLVMContext, essentially).

I'm not at all deeply familiar with the LLVM IR metadata APIs - so
this is a strawman at best but an idea about how we could provide the
functionality that debug info could benefit from in an isolated, not
debug-specific, feature. If we're going to touch metadata APIs I think
something like this should be how we do it.

If anyone more familiar with these APIs & the direction we should and
shouldn't take them could clarify if this is
possible/reasonable/acceptable/the right tradeoff compared to (a) that
would be much appreciated. That's what I'd like to get out of this
thread.

(for some more context: (a) will be isolated to debug info code, but
will require (in my naive estimation) a fair amount of mechanical work
to perform manual lookups on types whenever we access them - at some
point we may have a patch that demonstrates just how much work that
is, but until then we can only make vague statements, and by the time
we do this by any substantial amount it will've been a fair bit of
work only to throw it away in favor of (b) if it looks too unweildy.)

> 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.

As discussed offline, Eric & I really don't think this is acceptable.
Again, those who own/maintain/have opinions on LLVM metadata APIs can
weigh in if you think we're off-base in that assessment.

> Other suggestions are welcome.
>
> ---------------------------
> Preliminary Results:
> SPEC xalancbmk: ld time down from 20+ mins to 4+ mins on my Mac Air
>
> -----------------------------
> 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.
>
> -----------------------------
> Patches:
> Expect the following patches:
> 1> add flag -gtype-hashing
> 2> add hash field to DI types

If we can avoid changing the schema of DI types it's really best that
way - especially adding (or removing) fields, moreso while that's
under a flag. It'll be a real pain to update test cases & diverge the
schema making it harder to maintain good test coverage.

Essentially what we want to do is a single pass (as Eric said) - where
before there were MDNodes, there would now be strings. That means
updating DIBuilder and DwarfDebug, etc is one go, basically.

> 3> modify DIBuilder to use hash instead of MD reference
> 4> related to issue 3
> 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
> All changes should be local to debug info classes except patch #4.