Next: , Previous: Plugins, Up: Top

24 Link Time Optimization

24.1 Design Overview

Link time optimization is implemented as a GCC front end for a bytecode representation of GIMPLE that is emitted in special sections of .o files. Currently, LTO support is enabled in most ELF-based systems, as well as darwin, cygwin and mingw systems.

Since GIMPLE bytecode is saved alongside final object code, object files generated with LTO support are larger than regular object files. This “fat” object format makes it easy to integrate LTO into existing build systems, as one can, for instance, produce archives of the files. Additionally, one might be able to ship one set of fat objects which could be used both for development and the production of optimized builds. A, perhaps surprising, side effect of this feature is that any mistake in the toolchain that leads to LTO information not being used (e.g. an older libtool calling ld directly). This is both an advantage, as the system is more robust, and a disadvantage, as the user is not informed that the optimization has been disabled.

The current implementation only produces “fat” objects, effectively doubling compilation time and increasing file sizes up to 5x the original size. This hides the problem that some tools, such as ar and nm, need to understand symbol tables of LTO sections. These tools were extended to use the plugin infrastructure, and with these problems solved, GCC will also support “slim” objects consisting of the intermediate code alone.

At the highest level, LTO splits the compiler in two. The first half (the “writer”) produces a streaming representation of all the internal data structures needed to optimize and generate code. This includes declarations, types, the callgraph and the GIMPLE representation of function bodies.

When -flto is given during compilation of a source file, the pass manager executes all the passes in all_lto_gen_passes. Currently, this phase is composed of two IPA passes:

The second half of LTO support is the “reader”. This is implemented as the GCC front end lto1 in lto/lto.c. When collect2 detects a link set of .o/.a files with LTO information and the -flto is enabled, it invokes lto1 which reads the set of files and aggregates them into a single translation unit for optimization. The main entry point for the reader is lto/lto.c:lto_main.

24.1.1 LTO modes of operation

One of the main goals of the GCC link-time infrastructure was to allow effective compilation of large programs. For this reason GCC implements two link-time compilation modes.

  1. LTO mode, in which the whole program is read into the compiler at link-time and optimized in a similar way as if it were a single source-level compilation unit.
  2. WHOPR or partitioned mode, designed to utilize multiple CPUs and/or a distributed compilation environment to quickly link large applications. WHOPR stands for WHOle Program optimizeR (not to be confused with the semantics of -fwhole-program). It partitions the aggregated callgraph from many different .o files and distributes the compilation of the sub-graphs to different CPUs.

    Note that distributed compilation is not implemented yet, but since the parallelism is facilitated via generating a Makefile, it would be easy to implement.

WHOPR splits LTO into three main stages:

  1. Local generation (LGEN) This stage executes in parallel. Every file in the program is compiled into the intermediate language and packaged together with the local call-graph and summary information. This stage is the same for both the LTO and WHOPR compilation mode.
  2. Whole Program Analysis (WPA) WPA is performed sequentially. The global call-graph is generated, and a global analysis procedure makes transformation decisions. The global call-graph is partitioned to facilitate parallel optimization during phase 3. The results of the WPA stage are stored into new object files which contain the partitions of program expressed in the intermediate language and the optimization decisions.
  3. Local transformations (LTRANS) This stage executes in parallel. All the decisions made during phase 2 are implemented locally in each partitioned object file, and the final object code is generated. Optimizations which cannot be decided efficiently during the phase 2 may be performed on the local call-graph partitions.

WHOPR can be seen as an extension of the usual LTO mode of compilation. In LTO, WPA and LTRANS are executed within a single execution of the compiler, after the whole program has been read into memory.

When compiling in WHOPR mode, the callgraph is partitioned during the WPA stage. The whole program is split into a given number of partitions of roughly the same size. The compiler tries to minimize the number of references which cross partition boundaries. The main advantage of WHOPR is to allow the parallel execution of LTRANS stages, which are the most time-consuming part of the compilation process. Additionally, it avoids the need to load the whole program into memory.

24.2 LTO file sections

LTO information is stored in several ELF sections inside object files. Data structures and enum codes for sections are defined in lto-streamer.h.

These sections are emitted from lto-streamer-out.c and mapped in all at once from lto/lto.c:lto_file_read. The individual functions dealing with the reading/writing of each section are described below.

24.3 Using summary information in IPA passes

Programs are represented internally as a callgraph (a multi-graph where nodes are functions and edges are call sites) and a varpool (a list of static and external variables in the program).

The inter-procedural optimization is organized as a sequence of individual passes, which operate on the callgraph and the varpool. To make the implementation of WHOPR possible, every inter-procedural optimization pass is split into several stages that are executed at different times during WHOPR compilation:

The implementation of the inter-procedural passes are shared between LTO, WHOPR and classic non-LTO compilation.

To simplify development, the GCC pass manager differentiates between normal inter-procedural passes and small inter-procedural passes. A small inter-procedural pass (SIMPLE_IPA_PASS) is a pass that does everything at once and thus it can not be executed during WPA in WHOPR mode. It defines only the Execute stage and during this stage it accesses and modifies the function bodies. Such passes are useful for optimization at LGEN or LTRANS time and are used, for example, to implement early optimization before writing object files. The simple inter-procedural passes can also be used for easier prototyping and development of a new inter-procedural pass.

24.3.1 Virtual clones

One of the main challenges of introducing the WHOPR compilation mode was addressing the interactions between optimization passes. In LTO compilation mode, the passes are executed in a sequence, each of which consists of analysis (or Generate summary), propagation (or Execute) and Transform stages. Once the work of one pass is finished, the next pass sees the updated program representation and can execute. This makes the individual passes dependent on each other.

In WHOPR mode all passes first execute their Generate summary stage. Then summary writing marks the end of the LGEN stage. At WPA time, the summaries are read back into memory and all passes run the Execute stage. Optimization summaries are streamed and sent to LTRANS, where all the passes execute the Transform stage.

Most optimization passes split naturally into analysis, propagation and transformation stages. But some do not. The main problem arises when one pass performs changes and the following pass gets confused by seeing different callgraphs between the Transform stage and the Generate summary or Execute stage. This means that the passes are required to communicate their decisions with each other.

To facilitate this communication, the GCC callgraph infrastructure implements virtual clones, a method of representing the changes performed by the optimization passes in the callgraph without needing to update function bodies.

A virtual clone in the callgraph is a function that has no associated body, just a description of how to create its body based on a different function (which itself may be a virtual clone).

The description of function modifications includes adjustments to the function's signature (which allows, for example, removing or adding function arguments), substitutions to perform on the function body, and, for inlined functions, a pointer to the function that it will be inlined into.

It is also possible to redirect any edge of the callgraph from a function to its virtual clone. This implies updating of the call site to adjust for the new function signature.

Most of the transformations performed by inter-procedural optimizations can be represented via virtual clones. For instance, a constant propagation pass can produce a virtual clone of the function which replaces one of its arguments by a constant. The inliner can represent its decisions by producing a clone of a function whose body will be later integrated into a given function.

Using virtual clones, the program can be easily updated during the Execute stage, solving most of pass interactions problems that would otherwise occur during Transform.

Virtual clones are later materialized in the LTRANS stage and turned into real functions. Passes executed after the virtual clone were introduced also perform their Transform stage on new functions, so for a pass there is no significant difference between operating on a real function or a virtual clone introduced before its Execute stage.

Optimization passes then work on virtual clones introduced before their Execute stage as if they were real functions. The only difference is that clones are not visible during the Generate Summary stage.

To keep function summaries updated, the callgraph interface allows an optimizer to register a callback that is called every time a new clone is introduced as well as when the actual function or variable is generated or when a function or variable is removed. These hooks are registered in the Generate summary stage and allow the pass to keep its information intact until the Execute stage. The same hooks can also be registered during the Execute stage to keep the optimization summaries updated for the Transform stage.

24.3.2 IPA references

GCC represents IPA references in the callgraph. For a function or variable A, the IPA reference is a list of all locations where the address of A is taken and, when A is a variable, a list of all direct stores and reads to/from A. References represent an oriented multi-graph on the union of nodes of the callgraph and the varpool. See ipa-reference.c:ipa_reference_write_optimization_summary and ipa-reference.c:ipa_reference_read_optimization_summary for details.

24.3.3 Jump functions

Suppose that an optimization pass sees a function A and it knows the values of (some of) its arguments. The jump function describes the value of a parameter of a given function call in function A based on this knowledge.

Jump functions are used by several optimizations, such as the inter-procedural constant propagation pass and the devirtualization pass. The inliner also uses jump functions to perform inlining of callbacks.

24.4 Whole program assumptions, linker plugin and symbol visibilities

Link-time optimization gives relatively minor benefits when used alone. The problem is that propagation of inter-procedural information does not work well across functions and variables that are called or referenced by other compilation units (such as from a dynamically linked library). We say that such functions are variables are externally visible.

To make the situation even more difficult, many applications organize themselves as a set of shared libraries, and the default ELF visibility rules allow one to overwrite any externally visible symbol with a different symbol at runtime. This basically disables any optimizations across such functions and variables, because the compiler cannot be sure that the function body it is seeing is the same function body that will be used at runtime. Any function or variable not declared static in the sources degrades the quality of inter-procedural optimization.

To avoid this problem the compiler must assume that it sees the whole program when doing link-time optimization. Strictly speaking, the whole program is rarely visible even at link-time. Standard system libraries are usually linked dynamically or not provided with the link-time information. In GCC, the whole program option (-fwhole-program) asserts that every function and variable defined in the current compilation unit is static, except for function main (note: at link time, the current unit is the union of all objects compiled with LTO). Since some functions and variables need to be referenced externally, for example by another DSO or from an assembler file, GCC also provides the function and variable attribute externally_visible which can be used to disable the effect of -fwhole-program on a specific symbol.

The whole program mode assumptions are slightly more complex in C++, where inline functions in headers are put into COMDAT sections. COMDAT function and variables can be defined by multiple object files and their bodies are unified at link-time and dynamic link-time. COMDAT functions are changed to local only when their address is not taken and thus un-sharing them with a library is not harmful. COMDAT variables always remain externally visible, however for readonly variables it is assumed that their initializers cannot be overwritten by a different value.

GCC provides the function and variable attribute visibility that can be used to specify the visibility of externally visible symbols (or alternatively an -fdefault-visibility command line option). ELF defines the default, protected, hidden and internal visibilities.

The most commonly used is visibility is hidden. It specifies that the symbol cannot be referenced from outside of the current shared library. Unfortunately, this information cannot be used directly by the link-time optimization in the compiler since the whole shared library also might contain non-LTO objects and those are not visible to the compiler.

GCC solves this problem using linker plugins. A linker plugin is an interface to the linker that allows an external program to claim the ownership of a given object file. The linker then performs the linking procedure by querying the plugin about the symbol table of the claimed objects and once the linking decisions are complete, the plugin is allowed to provide the final object file before the actual linking is made. The linker plugin obtains the symbol resolution information which specifies which symbols provided by the claimed objects are bound from the rest of a binary being linked.

Currently, the linker plugin works only in combination with the Gold linker, but a GNU ld implementation is under development.

GCC is designed to be independent of the rest of the toolchain and aims to support linkers without plugin support. For this reason it does not use the linker plugin by default. Instead, the object files are examined by collect2 before being passed to the linker and objects found to have LTO sections are passed to lto1 first. This mode does not work for library archives. The decision on what object files from the archive are needed depends on the actual linking and thus GCC would have to implement the linker itself. The resolution information is missing too and thus GCC needs to make an educated guess based on -fwhole-program. Without the linker plugin GCC also assumes that symbols are declared hidden and not referred by non-LTO code by default.

24.5 Internal flags controlling lto1

The following flags are passed into lto1 and are not meant to be used directly from the command line.