[llvm-dev] [RFC] Lazy-loading of debug info metadata

[Duncan P. N. Exon Smith via llvm-dev]

> On 2016-Mar-22, at 20:11, Eric Christopher <echristo at gmail.com> wrote:
> 
>> On Tue, Mar 22, 2016 at 8:04 PM David Blaikie <dblaikie at gmail.com> wrote:
>> +pcc, who had some other ideas/patch out for improving memory usage of debug info
>> +Reid, who's responsible for the windows/CodeView/PDB debug info which is motivating some of the ideas about changes to type emission
> 
> So I discussed this with Adrian and Mehdi at the social last Thursday and I'm getting set to finish the write up. I think it'll have some bearing on this proposal as I think it'll change how we want to take a look at the format of DISubprogram metadata a bit more.

(The interesting bit here is to make a clearer split between
DISubprogram declarations (part of the type hierarchY) and
DISubprogram definitions (part of the code/line table/variable
locations).  I think that'll end up being mostly orthogonal to what
I'm trying to do.)

> That said, most of it is orthogonal to the changes Duncan is talking about here. Just puts the pressure on to get the other proposal written up.

Which is now here:
http://lists.llvm.org/pipermail/llvm-dev/2016-March/097773.html

>> Baking into the IR more about types as units has pretty direct overlap with Reid/CodeView/etc - so, yeah, that'll takes ome discussion (but, as you say, it's not in your immediate plan anyway, so we can come back to that - but would be good for whoever gets there first to discuss it with the others)

After thinking for a few days, I don't think this will bake anything
new into the IR.  If anything it removes a few special cases.

More at the bottom.

>>> Motivation
>>> ==========
>>> 
>>> Based on some analysis Mehdi ran (ping him for details), there are three
>>> (related) compile-time bottlenecks we're seeing with `-flto=thin -g`:
>>> 
>>>  a) Reading the large number of Metadata bitcode records in the global
>>>     metadata block.  I'm talking about raw `BitStreamer` calls here.
>>> 
>>>  b) Creating unnecessary `DI*` instances (that aren't relevant to code).
>>> 
>>>  c) Emitting unnecessary `DI*` instances (that aren't relevant to code).
>>> 
>>> Here is my recollection of some peak memory stats on a small testcase
>>> during thin-LTO, which should be a decent indicator of (b):
>>> 
>>>   - ~150MB: DILocation
>>>   - ~100MB: DISubprogram
>>>   - ~70MB: DILocalVariable
>>>   - ~50MB: (cumulative) DIType descendents
>>> 
>>> It looks, suprisingly, like types are not the primary bottleneck.

(Probably wrong for (a), BTW.  Caveats matter.)

>>> There are caveats:
>>> 
>>>   - `DISubprogram` declarations -- member function descriptors -- are
>>>     part of the type hierarchy.
>>>   - Most of the type hierarchy gets uniqued at parse time.
>>>   - As a result, these data are a poor indicator for (a).

((a) is the main bottleneck for compile-time of -flto=thin (since it's
quadratic in the number of files).  (b) only affects memory.  Also
important, but at least it scales linearly.)

>>> Even so, non-types are substantial.
>>> 
>>> Proposal
>>> ========
>>> 
>>> Short version
>>> -------------
>>> 
>>>  4. Remove `DICompositeType`s from `retainedTypes:`, similar to (2).

This is the part that's relevant to the new RFC Eric just posted.

>>> Long version
>>> -------------
>>> 
>>>  4. Implement my proposal to remove the `DICompositeType` name map from
>>>     `retainedTypes:`.
>>> 
>>>     http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20160125/327936.html
>>> 
>>>     Similar to (2) above, this will naturally filter the types that get
>>>     linked in to the ones actually used by the code being linked.
>>> 
>>>     It should also allow the reader to skip records for types that have
>>>     already been loaded in the main module.

The essential things I want to accomplish with this part:

  - Make `type:` operands less special: instead of referencing types
    indirectly through MDString, point directly at the type node.

  - Stop using `retainedTypes:` by default (only for -gfull, etc.).

  - Avoid blowing up memory in -flto=full (which converting MDString
    references back to pointers would do naively, through
    re-introducing cycles).  Note that this needs to be handled
    somehow at BitcodeReader time.

After chatting with Eric, I don't think this conflicts at all with the
other RFC.  Unifying the `type:` operands might actually help both.

One good point David mentioned last week was that we don't want to
teach the IR any more about types.  Rather than inventing some new
context (as I suggested originally), I figure LTO plugins can just
pass a (StringRef -> DIType*) map to the BitcodeReader, and the Module
itself doesn't need to know anything about it.