[llvm-dev] RFC: APIs for bitcode files containing multiple modules

[Will Dietz via llvm-dev]

On Fri, Oct 28, 2016 at 2:06 PM, Peter Collingbourne <peter at pcc.me.uk> wrote:
> On Fri, Oct 28, 2016 at 6:11 AM, Will Dietz <willdtz at gmail.com> wrote:
>>
>> On Wed, Oct 26, 2016 at 2:04 PM, Peter Collingbourne via llvm-dev
>> <llvm-dev at lists.llvm.org> wrote:
>> > On Tue, Oct 25, 2016 at 8:36 PM, Mehdi Amini <mehdi.amini at apple.com>
>> > wrote:
>> >>
>> >>
>> >> On Oct 25, 2016, at 6:28 PM, Peter Collingbourne <peter at pcc.me.uk>
>> >> wrote:
>> >>
>> >> Hi all,
>> >>
>> >> As mentioned in my recent RFC entitled "RFC: a more detailed design for
>> >> ThinLTO + vcall CFI" I would like to introduce the ability for bitcode
>> >> files
>> >> to contain multiple modules. In https://reviews.llvm.org/D24786 I took
>> >> a
>> >> step towards that by proposing a change to the module format so that
>> >> the
>> >> block info block is stored at the top level. The next step is to think
>> >> about
>> >> what the API would look like for reading and writing multiple modules.
>> >>
>> >> Here's what I have in mind. To create a multi-module bitcode file, you
>> >> would create a BitcodeWriter object and add modules to it:
>> >>
>> >> BitcodeWriter W(OS);
>> >> W.addModule(M1);
>> >> W.addModule(M2);
>> >> W.write();
>> >>
>> >>
>> >> That requires the two modules to lives longer than the bitcode write,
>> >> the
>> >> API could be:
>> >>
>> >> BitcodeWriter W(OS);
>> >> W.writeModule(M1);
>> >> // delete M1
>> >> // ...
>> >> // create M2
>> >> W.writeModule(M2);
>> >>
>> >> (Maybe you had this in mind, but the API naming didn’t reflect it so
>> >> I’m
>> >> not sure).
>> >
>> >
>> > In the API I prototyped, I took the maximum BitsRequiredForTypeIndices
>> > value
>> > from all the modules, and used it to produce the abbreviations for the
>> > top/
>> > level block info block (without this I was seeing "Unexpected abbrev
>> > ordering!" errors in the bitcode writer as a result of emitting the
>> > "same"
>> > abbreviation multiple times). That would have required us to keep the
>> > modules around until the call to write(). However, let me revisit this,
>> > because it does not seem necessary (i.e. we can just continue to emit
>> > block
>> > info blocks within the module block except with different abbreviation
>> > numbers for each module).
>> >>
>> >> Reading a multi-module bitcode file would be supported with a
>> >> BitcodeReader class. Each of the functional reader APIs in
>> >> ReaderWriter.h
>> >> would have a member function on BitcodeReader. We would also have a
>> >> next()
>> >> member function which would move to the next module in the file. For
>> >> example:
>> >>
>> >> BitcodeReader R(MBRef);
>> >> Expected<bool> B = R.hasGlobalValueSummary();
>>
>> What's this used for?
>
>
> This would be the equivalent to the existing llvm::hasGlobalValueSummary()
> function, which currently controls whether we compile a module with regular
> LTO or with ThinLTO.
>
>> Would there be a "readGlobalValueSummary()"
>> similar to function summaries?
>
>
> There would be a getModuleSummaryIndex() which again would be similar to
> llvm::getModuleSummaryIndex(). Note that the module summary already covers
> all global values, not just functions.
>
>> >> std::unique_ptr<Module> M1 = R.getLazyModule(Ctx); // lazily load the
>> >> first module
>> >> R.next();
>> >> std::unique_ptr<Module> M2 = R.parseBitcodeFile(Ctx); // eagerly load
>> >> the
>> >> second module
>>
>> I'm very excited about the idea of storing multiple modules in a
>> bitcode file, and the (thin)LTO and CFI goodness you're building using
>> it.
>>
>> I have a few questions about where you're going if you don't mind--and
>> it's related to the API in that it's awfully hard to judge an API
>> without knowing what it's expected to be used for or what the
>> underlying data represents.
>>
>> On that-- I'm sorry if I've missed this information, but reading
>> through your RFC's and posts I'm not finding the answer.
>> Is there a definition/explanation of what it means to have a bitcode
>> file containing multiple modules?
>>
>> Is this a storage optimization where each module is what today is an
>> "llvm::Module" but we're encoding them into a single file for
>> efficiency/convenience reasons?
>
>
> Yes, each module would be an llvm::Module. This is more for convenience
> reasons -- it's the simplest way to split modules that use CFI into a
> regular LTO part and a ThinLTO part (as described in the RFC entitled "RFC:
> a more detailed design for ThinLTO + vcall CFI") while storing the entire
> compiled translation unit in a single file.
>

Hmm, interesting.  Thank you for the explanation.

This seems to be closer to partitioning a single Module than
supporting multiple modules (at least not yet).
Does that seem accurate?
If so maybe the API should be geared towards that--allow
"partition-aware" clients to read the pieces individually while
transparently treating the overall file as a single Module for
existing clients.
Just a thought, perhaps this wouldn't work for your use case?

Anyway I actually am very interested in support for multiple modules,
my use case being for use in shipping software in IR form as part of
the ALLVM project.  Hence questions about things like linker semantics
and such.

Don't mean to burden you with accommodating the use-cases of everyone
else (like myself),
I guess I was just was surprised to see the bitcode format extended in
this way without an explicit discussion of the bigger picture--
what this was intended to be used for or why it was necessary, where
it was going... :).  Mostly because as you say it seems rather useful
for other parties (heterogeneous, for example) but I suppose we/they
can chime in and help refine the details later on once these bits are
committed :).

Thank you for your explanation, very much appreciated :).

>> If so, can these modules have different triples?
>
>
> That would certainly be possible in principle, but it's not part of my use
> case. I'd imagine that another potential use case for this could be to allow
> for LTO when targeting heterogeneous architectures (e.g. CUDA/OpenMP), but
> I'm not sure about the specifics of how that could work.
>
>>
>> Different ("conflicting") definitions for a global?
>
>
> In principle such inputs would be rejected by the linker with a duplicate
> symbol error. That might not be the appropriate thing to do in the
> heterogeneous case though.

Yeah, it seemed unclear what this would "mean" and I suppose for now
is simply something folks can interpret/handle however makes sense for
their use case :).

>
>> There are also multiple tools that take bitcode as input, and
>> currently expect a single module.
>> Will these be made to reject multiple-module bitcode, and if not is
>> the plan to extend tools to handle multiple-module files?
>
>
> For testing purposes I was planning to extend llvm-dis (and possibly opt) to
> take a flag specifying a module index, and introduce an llvm-join tool which
> could be used to create a bitcode from multiple inputs.
>

Awesome! I'm not sure how important it is but it seems that it should
be made an error to ignore part of a bitcode file?
(Shouldn't llvm-nm print vtable bits?)

> The other tools probably don't need to know about this and could just read
> the first module.
>
>> Beyond the random access suggestion (+1) and lifetime comments, it
>> seems like there should be a way to reason about the contents of these
>> modules--names, identifiers, flags, *something* so that "load the
>> first module lazily and the second eagerly" can become "load the
>> module containing my CFI information eagerly but the rest lazily" or
>> something, or at least to check that this file was created using
>> -fsanitize=cfi and not something else.
>
>
> Right, this is the sort of functionality that would be provided by functions
> such as hasGlobalValueSummary().

Ah, neat.  I'll look into that, since apparently it answers many of my
questions :D.  Sorry for the trouble :).

Thanks again, happy LLVM'ing...

~Will

>
> Thanks,
> --
> Peter