[LLVMdev] RFC: ThinLTO Impementation Plan

[Teresa Johnson]

Thanks for all the feedback and questions, answers below.
Teresa

On Thu, May 14, 2015 at 4:29 PM, Duncan P. N. Exon Smith
<dexonsmith at apple.com> wrote:
>
>> On 2015-May-13, at 11:44, Teresa Johnson <tejohnson at google.com> wrote:
>>
>> I've included below an RFC for implementing ThinLTO in LLVM, looking
>> forward to feedback and questions.
>> Thanks!
>> Teresa
>>
>>
>>
>> RFC to discuss plans for implementing ThinLTO upstream. Background can
>> be found in slides from EuroLLVM 2015:
>>   https://drive.google.com/open?id=0B036uwnWM6RWWER1ZEl5SUNENjQ&authuser=0)
>> As described in the talk, we have a prototype implementation, and
>> would like to start staging patches upstream. This RFC describes a
>> breakdown of the major pieces. We would like to commit upstream
>> gradually in several stages, with all functionality off by default.
>> The core ThinLTO importing support and tuning will require frequent
>> change and iteration during testing and tuning, and for that part we
>> would like to commit rapidly (off by default). See the proposed staged
>> implementation described in the Implementation Plan section.
>>
>>
>> ThinLTO Overview
>> ==============
>>
>> See the talk slides linked above for more details. The following is a
>> high-level overview of the motivation.
>>
>> Cross Module Optimization (CMO) is an effective means for improving
>> runtime performance, by extending the scope of optimizations across
>> source module boundaries. Without CMO, the compiler is limited to
>> optimizing within the scope of single source modules. Two solutions
>> for enabling CMO are Link-Time Optimization (LTO), which is currently
>> supported in LLVM and GCC, and Lightweight-Interprocedural
>> Optimization (LIPO). However, each of these solutions has limitations
>> that prevent it from being enabled by default. ThinLTO is a new
>> approach that attempts to address these limitations, with a goal of
>> being enabled more broadly. ThinLTO is designed with many of the same
>> principals as LIPO, and therefore its advantages, without any of its
>> inherent weakness. Unlike in LIPO where the module group decision is
>> made at profile training runtime, ThinLTO makes the decision at
>> compile time, but in a lazy mode that facilitates large scale
>> parallelism. The serial linker plugin phase is designed to be razor
>> thin and blazingly fast. By default this step only does minimal
>> preparation work to enable the parallel lazy importing performed
>> later. ThinLTO aims to be scalable like a regular O2 build, enabling
>> CMO on machines without large memory configurations, while also
>> integrating well with distributed build systems. Results from early
>> prototyping on SPEC cpu2006 C++ benchmarks are in line with
>> expectations that ThinLTO can scale like O2 while enabling much of the
>> CMO performed during a full LTO build.
>>
>>
>> A ThinLTO build is divided into 3 phases, which are referred to in the
>> following implementation plan:
>>
>> phase-1: IR and Function Summary Generation (-c compile)
>> phase-2: Thin Linker Plugin Layer (thin archive linker step)
>> phase-3: Parallel Backend with Demand-Driven Importing
>>
>>
>> Implementation Plan
>> ================
>>
>> This section gives a high-level breakdown of the ThinLTO support that
>> will be added, in roughly the order that the patches would be staged.
>> The patches are divided into three stages. The first stage contains a
>> minimal amount of preparation work that is not ThinLTO-specific. The
>> second stage contains most of the infrastructure for ThinLTO, which
>> will be off by default. The third stage includes
>> enhancements/improvements/tunings that can be performed after the main
>> ThinLTO infrastructure is in.
>>
>> The second and third implementation stages will initially be very
>> volatile, requiring a lot of iterations and tuning with large apps to
>> get stabilized. Therefore it will be important to do fast commits for
>> these implementation stages.
>>
>>
>> 1. Stage 1: Preparation
>> -------------------------------
>>
>> The first planned sets of patches are enablers for ThinLTO work:
>>
>>
>> a. LTO directory structure:
>>
>> Restructure the LTO directory to remove circular dependence when
>> ThinLTO pass added. Because ThinLTO is being implemented as a SCC pass
>> within Transforms/IPO, and leverages the LTOModule class for linking
>> in functions from modules, IPO then requires the LTO library. This
>> creates a circular dependence between LTO and IPO. To break that, we
>> need to split the lib/LTO directory/library into lib/LTO/CodeGen and
>> lib/LTO/Module, containing LTOCodeGenerator and LTOModule,
>> respectively. Only LTOCodeGenerator has a dependence on IPO, removing
>> the circular dependence.
>>
>
> I wonder whether LTOModule is a good fit (it might be; I'm not sure).
> We still use it in libLTO, but gold-plugin.cpp no longer uses it,
> instead using lib/Object and lib/Linker directly.
>
>> b. ELF wrapper generation support:
>
> (From elsewhere in the thread, it looks like you're just using ELF
> as a short-hand for "native".)

Right, I should have written this as native object wrapper. I had
focused on ELF since that was what I have been looking at most
closely, but the support can be more general.

>
>>
>> Implement ELF wrapped bitcode writer. In order to more easily interact
>> with tools such as $AR, $NM, and “$LD -r” we plan to emit the phase-1
>> bitcode wrapped in ELF via the .llvmbc section, along with a symbol
>> table. The goal is both to interact with these tools without requiring
>> a plugin, and also to avoid doing partial LTO/ThinLTO across files
>> linked with “$LD -r” (i.e. the resulting object file should still
>> contain ELF-wrapped bitcode to enable ThinLTO at the full link step).
>
> Shouldn't `ld -r` change symbol visibility and such?  How do you plan
> to handle that when you concatenate sections?

If we use native object wrapped bitcode, ld -r would not do any
changing of symbols or merging. It would be more like an archive in
that it packages the bitcode and delays merging until the backend.
That way it's constituents are still bitcode available for importing
into other modules.

For the non-wrapped bitcode option, using the gold plugin, we would
want to change the behavior for ld -r to be similar to what you are
describing for ld64, i.e. emit bitcode.

>
> For reference, ld64 (through libLTO) merges all the bitcode together
> with lib/Linker, gives all "hidden" symbols local linkage (by running
> -internalize with OnlyHidden=1), and writes out a new bitcode file.
>
>> I will send a separate design document for these changes, but the
>> following is a high-level overview.
>>
>> Support was added to LLVM for reading ELF-wrapped bitcode
>> (http://reviews.llvm.org/rL218078), but there does not yet exist
>> support in LLVM/Clang for emitting bitcode wrapped in ELF. I plan to
>> add support for optionally generating bitcode in an ELF file
>> containing a single .llvmbc section holding the bitcode. Specifically,
>> the patch would add new options “emit-llvm-bc-elf” (object file) and
>> corresponding “emit-llvm-elf” (textual assembly code equivalent).
>
> If we decide to go this way -- wrapping the bitcode in the native
> object format -- wouldn't emit-llvm-native or emit-llvm-object be
> better?  The native object format is implied by the triple.

Yes, that is better.

>
>> Eventually these would be automatically triggered under “-fthinlto -c”
>> and “-fthinlto -S”, respectively.
>>
>> Additionally, a symbol table will be generated in the ELF file,
>> holding the function symbols within the bitcode. This facilitates
>> handling archives of the ELF-wrapped bitcode created with $AR, since
>> the archive will have a symbol table as well. The archive symbol table
>> enables gold to extract and pass to the plugin the constituent
>> ELF-wrapped bitcode files. To support the concatenated llvmbc section
>> generated by “$LD -r”, some handling needs to be added to gold and to
>> the backend driver to process each original module’s bitcode.
>>
>> The function index/summary will later be added as a special ELF
>> section alongside the .llvmbc sections.
>>
>>
>> 2. Stage 2: ThinLTO Infrastructure
>> ----------------------------------------------
>>
>> The next set of patches adds the base implementation of the ThinLTO
>> infrastructure, specifically those required to make ThinLTO functional
>> and generate correct but not necessarily high-performing binaries. It
>> also does not include support to make debug support under -g efficient
>> with ThinLTO.
>
> I think we should at least have a vague plan...

Sorry, I should have been clearer here. I do have a plan for this and
know how to do it (it is implemented in my prototype). It's discussed
below under Stage 3. I was debating whether to put the metadata
handling under Stage 2, but it isn't strictly necessary to get the
ThinLTO pipeline working. You just end up with a lot of duplicate
metadata/debug as you have to import it multiple times. But really the
metadata (incl debug) handling should be the next thing after the
basic ThinLTO pipeline is done.

>
>> a. Clang/LLVM/gold linker options:
>>
>> An early set of clang/llvm patches is needed to provide options to
>> enable ThinLTO (off by default), so that the rest of the
>> implementation can be disabled by default as it is added.
>> Specifically, clang options -fthinlto (used instead of -flto) will
>> cause clang to invoke the phase-1 emission of LLVM bitcode and
>> function summary/index on a compile step, and pass the appropriate
>> option to the gold plugin on a link step. The -thinlto option will be
>> added to the gold plugin and llvm-lto tool to launch the phase-2 thin
>> archive step. The -thinlto option will also be added to the ‘opt’ tool
>> to invoke it as a phase-3 parallel backend instance.
>
> I'm not sure I follow the `opt` part of this.  That's a developer
> tool, not something we ship.  It also doesn't have a backend (doesn't
> do CodeGen).  What am I missing?

For the prototype I was using llvm-lto as my backend driver. I
realized that this was probably not the best option as we don't need
all of the LTO handling built into that driver, and it isn't listed as
a tool on http://llvm.org/docs/CommandGuide/, so my feeling was that
'opt' was better supported and a better alternative. Unfortunately
when I was writing this up I forgot that 'opt' generates bitcode not
an object file.

Another option would be to use clang and allow it to accept bitcode
and bypass parsing under an appropriate ThinLTO option. AFAICT there
isn't currently an option for clang to accept bitcode. Do you think
this is the right approach?


>
>> b. Thin-archive linking support in Gold plugin and llvm-lto:
>>
>> Under the new plugin option (see above), the plugin needs to perform
>> the phase-2 (thin archive) link which simply emits a combined function
>> map from the linked modules, without actually performing the normal
>> link. Corresponding support should be added to the standalone llvm-lto
>> tool to enable testing/debugging without involving the linker and
>> plugin.
>>
>>
>> c. ThinLTO backend support:
>>
>> Support for invoking a phase-3 backend invocation (including
>> importing) on a module should be added to the ‘opt’ tool under the new
>> option. The main change under the option is to instantiate a Linker
>> object used to manage the process of linking imported functions into
>> the module, efficient read of the combined function map, and enable
>> the ThinLTO import pass.
>>
>>
>> d. Function index/summary support:
>>
>> This includes infrastructure for writing and reading the function
>> index/summary section. As noted earlier this will be encoded in a
>> special ELF section within the module, alongside the .llvmbc section
>> containing the bitcode. The thin archive generated by phase-2 of
>> ThinLTO simply contains all of the function index/summary sections
>> across the linked modules, organized for efficient function lookup.
>>
>> Each function available for importing from the module contains an
>> entry in the module’s function index/summary section and in the
>> resulting combined function map. Each function entry contains that
>> function’s offset within the bitcode file, used to efficiently locate
>> and quickly import just that function.
>
> I don't think you'll actually buy anything here over the lazy-loading
> feature in the BitcodeReader (although perhaps you can help improve
> it if you have some ideas).  In practice, to correctly load a
> Function you need to load constants (include declarations for other
> GlobalValues) and metadata that it references.

As you saw below, it is leveraging the lazy loading support. The
metadata handling is discussed later on in 3a.

>
>> The entry also contains summary
>> information (e.g. basic information determined during parsing such as
>> the number of instructions in the function), that will be used to help
>> guide later import decisions. Because the contents of this section
>> will change frequently during ThinLTO tuning, it should also be marked
>> with a version id for backwards compatibility or version checking.
>>
>>
>> e. ThinLTO importing support:
>>
>> Support for the mechanics of importing functions from other modules,
>> which can go in gradually as a set of patches since it will be off by
>> default. Separate patches can include:
>>
>> - BitcodeReader changes to use function index to import/deserialize
>> single function of interest (small changes, leverages existing lazy
>> streamer support).
>
> Ah, here it is.  Should have read ahead.
>
> How do you plan to handle references to other GlobalValues (global
> variables, functions, and aliases)? If you're going to keep loading
> the symbol table (which I think you need to?), then the lazy loader
> already creates a function index.  Or do you have some other plan?

We do have to reload the declarations and other symbol table info.
Where it differs from the lazy loader is that we don't need to keep
parsing the module to build up the function index
(DeferredFunctionInfo), with repeated calls to
FindFunctionInStream/ParseModule. Once we hit the first function body
we stop, then when materializing we simply set up the
DeferredFunctionInfo entry from the bitcode index that was saved in
the ThinLTO function index.

>
> If an imported function references functions with internal linkage,
> will you pull in copies of those functions as well?

There are two possibilities in this case: promotion (along with
renaming to avoid name clashing with other modules), or force import.
As you note later on, I talk about promotion just below here. To limit
the required static promotions I have implemented a strategy where we
attempt to force import referenced functions that have internal
linkage. But we still must do static promotion if the local function
(or global) is potentially imported to another module (in the combined
function map) and is address exposed.

>
> If an imported function references global variables with internal
> linkage... actually, that doesn't seem legal.  Will you disallow
> importing such functions?  How will you mark them?

Static promotion handles this.

>
>>
>> - Minor LTOModule changes to pass the ThinLTO function to import and
>> its index into bitcode reader.
>>
>> - Marking of imported functions (for use in ThinLTO-specific symbol
>> linking and global DCE, for example).
>
> Marking how?  Do you mean giving them internal linkage, or something
> else?

Mentioned just after this: either an in-memory flag on the Function
class, or potentially in the IR. For the prototype I just had a flag
on the Function class.

>
> What's your plan for ThinLTO-specific symbol linking?

Mentioned just below here as you note.

>
>> This can be in-memory initially,
>> but IR support may be required in order to support streaming bitcode
>> out and back in again after importing.
>>
>> - ModuleLinker changes to do ThinLTO-specific symbol linking and
>> static promotion when necessary. The linkage type of imported
>> functions changes to AvailableExternallyLinkage, for example. Statics
>> must be promoted in certain cases, and renamed in consistent ways.
>
> Ah, could have read ahead again; this answers my questions about
> referencing global variables with local linkage.
>
> It also sounds pretty hairy.  Details welcome.

It has to be well thought out for sure. We had to do this for LIPO as
well so already knew what needed to be done here. I will put together
more details in a follow-on email.

>
>>
>> - GlobalDCE changes to support removing imported functions that were
>> not inlined (very small changes to existing pass logic).
>
> If you give them "available_externally" linkage, won't this already
> happen?

There were only a couple minor tweaks required here (under the flag I
added to the Function indicating that this was imported). Only
promoted statics are remarked available_externally. For a
non-discardable symbol that was imported, we can discard here since we
are done with inlining (it is non-discardable in its home module).

>
>>
>>
>> f. ThinLTO Import Driver SCC pass:
>>
>> Adds Transforms/IPO/ThinLTO.cpp with framework for doing ThinLTO via
>> an SCC pass, enabled only under -fthinlto options. The pass includes
>> utilizing the thin archive (global function index/summary), import
>> decision heuristics, invocation of LTOModule/ModuleLinker routines
>> that perform the import, and any necessary callgraph updates and
>> verification.
>>
>>
>> g. Backend Driver:
>>
>> For a single node build, the gold plugin can simply write a makefile
>> and fork the parallel backend instances directly via parallel make.
>
> This doesn't seem like the way we'd want to test this, and it
> seems strange for the toolchain to require a build system...

The idea is to make this all transparent to the user. So you can just
do something like:
% clang -fthinlto -O2 *.cc -c
% clang -fthinlto -O2 *.o

the second command will do everything transparently (phase-2 thin
plugin later, launch parallel backend processes, hand back resulting
native object code to linker, produce a.out). So somehow the plugin
needs to launch the parallel backend processes.

>
>>
>>
>> 3. Stage 3: ThinLTO Tuning and Enhancements
>> ----------------------------------------------------------------
>>
>> This refers to the patches that are not required for ThinLTO to work,
>> but rather to improve compile time, memory, run-time performance and
>> usability.
>>
>>
>> a. Lazy Debug Metadata Linking:
>>
>> The prototype implementation included lazy importing of module-level
>> metadata during the ThinLTO pass finalization (i.e. after all function
>> importing is complete). This actually applies to all module-level
>> metadata, not just debug, although it is the largest. This can be
>> added as a separate set of patches. Changes to BitcodeReader,
>> ValueMapper, ModuleLinker
>
> It sounds like this would work well with the "full" LTO implemented
> by tools/gold-plugin right now.  What exactly did you do to improve
> this?

I don't think it will help with full LTO. The parsing of the metadata
is only delayed until the ThinLTO pass finalization, and the delayed
metadata import is necessary to avoid reading and linking in the
metadata multiple times (for each function imported from that module).
Coming out of the ThinLTO pass you still have all the metadata
necessary for each function that was imported. For a full LTO that
would end up being all of the metadata in the module.

The high level summary is that during the initial import it leaves the
temporary metadata on the instructions that were imported, but saves
the index used by the bitcode reader used to correlate with the
metadata when it is ready (i.e. the MDValuePtrs index), and skips the
metadata parsing. During finalization we parse just the metadata, and
suture it up during metadata value mapping using the saved index.

>
>>
>>
>> b. Import Tuning:
>>
>> Tuning the import strategy will be an iterative process that will
>> continue to be refined over time. It involves several different types
>> of changes: adding support for recording additional metrics in the
>> function summary, such as profile data and optional heavier-weight IPA
>> analyses, and tuning the import heuristics based on the summary and
>> callsite context.
>>
>>
>> c. Combined Function Map Pruning:
>>
>> The combined function map can be pruned of functions that are unlikely
>> to benefit from being imported. For example, during the phase-2 thin
>> archive plug step we can safely omit large and (with profile data)
>> cold functions, which are unlikely to benefit from being inlined.
>> Additionally, all but one copy of comdat functions can be suppressed.
>>
>>
>> d. Distributed Build System Integration:
>>
>> For a distributed build system, the gold plugin should write the
>> parallel backend invocations into a makefile, including the mapping
>> from the IR file to the real object file path, and exit. Additional
>> work needs to be done in the distributed build system itself to
>> distribute and dispatch the parallel backend jobs to the build
>> cluster.
>>
>>
>> e. Dependence Tracking and Incremental Compiles:
>>
>> In order to support build systems that stage from local disks or
>> network storage, the plugin will optionally support computation of
>> dependent sets of IR files that each module may import from. This can
>> be computed from profile data, if it exists, or from the symbol table
>> and heuristics if not. These dependence sets also enable support for
>> incremental backend compiles.
>>
>>
>>
>> --
>> Teresa Johnson | Software Engineer | tejohnson at google.com | 408-460-2413
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>



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