[PATCH] D11722: [ThinLTO] Bitcode reading/writing support for ThinLTO function summary/index

[Teresa Johnson via llvm-commits]

On Tue, Aug 18, 2015 at 7:58 PM, Duncan P. N. Exon Smith
<dexonsmith at apple.com> wrote:
>
>> On 2015-Aug-13, at 10:22, Teresa Johnson <tejohnson at google.com> wrote:
>>
>>
>>
>> On Thu, Aug 13, 2015 at 9:42 AM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
>>
>> > On 2015-Aug-12, at 21:23, Teresa Johnson <tejohnson at google.com> wrote:
>> >
>> > tejohnson updated the summary for this revision.
>> > tejohnson updated this revision to Diff 32027.
>> > tejohnson added a comment.
>> >
>> > Removed native object wrapper support from this patch.
>> >
>> >
>> > http://reviews.llvm.org/D11722
>> >
>> > Files:
>> >  include/llvm/Bitcode/LLVMBitCodes.h
>> >  include/llvm/Bitcode/ReaderWriter.h
>> >  lib/Bitcode/Reader/BitcodeReader.cpp
>> >  lib/Bitcode/Writer/BitcodeWriter.cpp
>> >  tools/llvm-bcanalyzer/llvm-bcanalyzer.cpp
>> >
>> > <D11722.32027.patch>
>>
>> (Sorry I'm so late getting to this patch.)
>>
>> Hi Duncan,
>> Thanks for the comments.
>>
>>
>> I'm concerned about the direction.  IIUC, the purpose of this is faster
>> lazy-loading.  Why do this in a side-channel?  Why not optimize the
>> existing lazy-loader?  I imagine the LTO model implemented by the gold
>> plugin would benefit from this, for example.  If we need an extra index
>> to make lazy-loading "fast enough", it seems like the index should be
>> available (as an opt-in for time/space tradeoff) for other consumers as
>> well.  Alternatively, maybe there's something we can change about how
>> the existing lazy-loader works (or the layout of bitcode) to be more
>> amenable to the partial loading ThinLTO needs.
>>
>> Only one part of the index (the function's bitcode offset) is used to do the fast loading of the function from a given module. Much of the other information in the ThinLTO sections is to identify/locate the module to import from (via the module strtab index which is populated when creating the combined (global) index), and information used in making importing decisions from another module (like the function's size, hotness when there is profile, etc, which we don't have since we haven't yet parsed that other module).
>>
>> The other issue with using the existing lazy loading support is that we may import from a number of other modules in some interleaved fashion, so we may open/import a function/close a module multiple times. My  understanding is that the lazy loading is more for the case of loading a few different functions in sequence.
>>
>> So for example, if we have:
>>
>> A.cc:
>>
>> a() {
>>    b1();
>>    ...
>>    c1();
>>    ...
>>    b2();
>>    ...
>>    c2();
>> }
>>
>> B.cc:
>>
>> b1() { d1(); }
>> b2() { ... }
>>
>> C.cc:
>>
>> c1() { d2(); }
>> c2() { ... }
>>
>> D.cc:
>>
>> d1() { ... }
>> d2() { ... }
>>
>>
>> When compiling A.cc through the parallel backend's importing pass, we may have the following sequence of events:
>>
>> Consider importing b1  -> Decide to import b1  (exposes call to d1)
>> Consider importing d1  -> Decide to import d1
>> Consider importing c1  -> Decide to import c1  (exposes call to d2)
>> Consider importing d2  -> Decide to import d2
>> Consider importing b2  -> Decide *not* to import b2
>> Consider importing c2  -> Decide to import c2
>>
>> For each of the inter-module calls considered in some priority based order, we would look up the callee in the combined index, which provides some information such as the callee's size, hotness, etc. A decision on whether to import is made based on that information (without doing any parsing of the callee's module). If we decide to import, then the combined index entry for the callee provides the callee module path and the bitcode index, which is used to open the module and import just that function, then the callee's module is closed. Some of the lazy loading support could potentially be used in place of the bitcode index, although I am not convinced it is suited for that purpose. In any case, that is just a small part of what is contained in the function summary/index as described earlier.
>>
>> On the flip side, the bitcode index could be used by the lazy loader, in place of building the index on the fly. I'm not sure how much benefit there is to lazy loading, which eventually needs to read the entire module once a function is materialized.
>
> It seems generally useful to be able to cherry-pick functions out of a
> large bitcode file without scanning the whole thing.  It's really cool
> that you can teardown and buildup the BitcodeReader, swapping different
> modules in and out.
>
> I don't like this being tied to ThinLTO, since I think it's useful for
> other consumers as well.  To start, the old/normal lazy loader should
> use the index when it's available.  I'm sure other people will think of
> other crazy uses.

I've been thinking this morning about the best way to leverage the
ThinLTO index in the existing lazy loader and have some ideas. But I
think that just applies to using the bitcode offset of the function
body block.

Do you think that it is interesting to put the rest of the summary
information I am going to use for importing decisions like the
instruction count, profile hotness, other function characteristics
(queried from other modules, via the combined index) in the bitcode in
a non-ThinLTO-specific way? Perhaps they could still be optional and
currently only produced with -fthinlto, but the block/record names not
contain "ThinLTO".

>
> One possible use: could llvm-ar use this index (when available) instead
> of walking through all the functions?

I couldn't find where llvm-ar walks through the functions. What ar
operation is doing that?

>
>> Not included in this patch is support for lazy-loading of metadata,
>> something you suggested you had support for in your prototype (and I
>> assume ThinLTO relies on it).  I think it's particularly important that
>> whatever you do there isn't ThinLTO-specific.
>>
>>
>> Right, that is needed when we do the importing. This initial patch is just to write/read the function summary/index. I had briefly outlined the way I was hoping to stage patches earlier today (http://lists.llvm.org/pipermail/llvm-dev/2015-August/089222.html). I was planning to do the importing as a separate set of patches once I have the infrastructure for generating/reading the function indexes as that is needed first to drive the importing. Let me know if the staging of work I describe there makes sense.
>
> (What's difficult for me with the staging is how many threads are going
> on in parallel.  This stuff is all related, and some of the patches
> don't make sense outside the context of the others, and if we need
> changes in one, that'll influence the design of the others.  I don't
> really see why the review isn't all in one place.  Maybe I just need to
> stop worrying and learn to love my inbox, though... and I guess if I'd
> responded more promptly to the RFCs you wouldn't have unleashed all the
> threads.)

Yea, I definitely see your point. I thought breaking it up into
multiple smaller patches would make it easier to review. But it
certainly would be easier for me as well to have a single larger patch
as I keep having to update multiple patches when I rename things or
make other data structure changes in one patch.

Do you think I should kill these individual patches off and combine
into one larger patch when I make the next round of changes?

>
>> To facilitate the process of reading/materializing a single function each time a module is imported from, as described above, I we be parsing the module level metadata once as a post-pass. Otherwise supporting the interleaved importing from multiple modules is probably very difficult.
>
> I seem to remember that you only pull in the necessary metadata?  I'm
> wondering how you do this, particularly for debug info, which has all
> sorts of circular references.  In particular, !llvm.dbg.cu seems like it
> could pull in all the subprograms (and their variables) even if you've
> just cherry-picked a couple of functions.  How do you get around that?

So there are a lot of circular references and therefore there is still
a fair amount pulled in (not sure how much has changed in this since I
did my prototype work very early this year but this was an issue for
types if I remember). But the llvm.dbg.cu is handled (specifically,
the list of subprograms hanging off of the DICompileUnit). When doing
the post-importing-pass to link the needed metadata from a module we
imported from, I have a list of all functions that were imported, and
I also recursively find any other DISubProgram functions reached from
MD in the imported function. This is passed to the ValueMapper and it
only maps in DISubProgram MD nodes and descendants for the referenced
functions.

>
> (I've been planning some debug info schema changes to make this sort of
> thing easier, but were you able to do this without them?  How?
> (Frankly, it'd be helpful to look at the complete prototype (warts and
> all) if it's available somewhere.))

It isn't available anywhere, and I'm very reluctant to do so as it was
very much a prototype and my first LLVM work. It has a lot of cruft
like tons of debugging/tracing code, stuff that isn't needed anymore
(some of it commented out but left in with big notes about why I went
with a different approach), and I didn't put much thought into nice
data structures.

Perhaps I should just tackle the implementation of the importing
mechanics next. Unfortunately, getting anything that is testable in a
good way requires implementing the linkage changes/static promotion
support and the importing pass and other things. But perhaps I can
come up with a simple way to test/demonstrate the importing mechanics
without all of that.

>
>> This is again different than the existing lazy loading, which is parsing/materializing all necessary functions and metadata out of each module in a single pass.
>
> Right, but during "normal" lazy loading, function bodies can be deleted
> before they're loaded, so that they're never loaded at all.  It would be
> awesome to delay loading metadata until we know what is actually
> needed/used, so it's not all leaked onto the LLVMContext.

Ok, it seems like part of this (the DISubProgram handling) could be
used for lazy loading as well then. But it wouldn't need to be done as
a post pass, which makes a few other things easier than in the ThinLTO
case (fixing up temp MD).

>
> ----
>
> Stepping back a little, I have felt (and, sadly, have probably been)
> confrontational about the ThinLTO work.  It's important to say: this is
> a really cool approach, and I want it in tree (somehow).

Thanks!

>
> But at a high level, it feels like you've examined how LLVM does all
> these things, determined (correctly!) that it's not good enough for the
> ThinLTO use case, and created a new approach for ThinLTO.
>
> I'd rather see patches that fix the current LLVM approach.  Even when
> there's a tradeoff (for example, the index takes extra disk space, so
> that should be opt-in (via a -cc1 flag like
> -emit-llvm-function-index?)), the new awesome shouldn't be ThinLTO-only.

Ok, I have some thoughts on combining at least the function bitcode
offset so that the lazy loading of functions and ThinLTO function
importing can be unified. Let me put some thoughts down and send that
in a follow up.

Teresa

>
>> Let me know if this seems reasonable given the above example and description.
>>
>> Thanks,
>> Teresa
>>
>>
>> --
>> Teresa Johnson |       Software Engineer |     tejohnson at google.com |  408-460-2413
>



-- 
Teresa Johnson | Software Engineer | tejohnson at google.com | 408-460-2413