Advanced Optimization and New Capabilities of GCC 12

The first release of the GNU Compiler Collection (GCC) with the major version 12, GCC 12.1, took place in May 2022. Later that month, the entire openSUSE Tumbleweed Linux distribution was rebuilt with it and shipped to users. GCC 12.2, with fixes to over 71 bugs, was released in August of the same year. Subsequently, it has replaced the compiler in the SUSE Linux Enterprise (SLE) Development Tools Module. GCC 12.3 followed in May 2023. Apart from further bug fixes, it also introduced support for Zen 4 based CPUs. GCC 12 comes with many new features, such as implementing parts of the most recent versions of specifications of various languages (especially C2X, C++20, C++23) and their extensions (OpenMP, OpenACC), supporting new capabilities of a wide range of computer architectures and numerous generic optimization improvements.

This document gives an overview of GCC 12. It focuses on selecting appropriate optimization options for your application and stresses the benefits of advanced modes of compilation. First, we describe the optimization levels the compiler offers, and other important options developers often use. We explain when and how you can benefit from using Link Time Optimization (LTO) and Profile Guided Optimization (PGO) builds. We also detail their effects when building a set of well-known CPU intensive benchmarks. Finally, we look at how these perform on AMD Zen 4 based EPYC 9004 Series Processors.

The major version of the system compiler in SUSE Linux Enterprise 15 remains to be GCC 7, regardless of the service pack level. This is to minimize the danger of any unintended changes over the entire life time of the product.

sles15: # gcc --version
gcc (SUSE Linux) 7.5.0
Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

That does not mean that, as a user of SUSE Linux Enterprise 15, you are forced to use a compiler with features frozen in 2016. You can install an add-on module called Development Tools Module. This module is included in the SUSE Linux Enterprise Server 15 subscription and contains a much newer compiler.

At the time of writing this document, the compiler included in the Development Tools Module is GCC 12.3. Nevertheless, it is important to stress that, unlike the system compiler, the major version of the most recent GCC from the module will change a few months after the upstream release of GCC 13.2 (which is planned for summer 2023), GCC 14.2 (summer 2024) and so forth. Note that only the most recent compiler in the Development Tools Module is supported at any time, except for a six months overlap period after an upgrade happened. Developers on a SUSE Linux Enterprise Server 15 system therefore have always access to two supported GCC versions: the almost unchanging system compiler and the most recent compiler from the Development Tools Module.

Programs and libraries built with the compiler from the Development Tools Module can run on computers running SUSE Linux Enterprise Server 15 which do not have the module installed. All necessary runtime libraries are available from the main repositories of the operating system itself, and new ones are added through the standard update mechanism. In this document, we use the term GCC 12 as synonym for any minor version of the major version 12 and GCC 12.3, to refer to specifically that version. In practice they should be interchangeable except when we discuss targeting AMD Zen 4 based CPUs which is only supported in 12.3 and newer versions.

2.1 When to use compilers from the Development Tools Module #

  

Often you will find that the system compiler perfectly satisfies your needs. After all, it is the compiler used to build the vast majority of packages and their updates in the system itself. On the other hand, there are situations where a newer compiler is necessary, or where you want to consider using a newer compiler to get some benefits of its ongoing development.

If the program or library you are building uses language features which are not supported by GCC 7, you cannot use the system compiler. However, the compiler from the Development Tools Module will usually be sufficiently new. The most obvious case is C++. GCC 12 has a mature implementation of C++17 features, whereas the one in GCC 7 is only experimental and incomplete. The GNU C++ Library which accompanies GCC 12 is also C++17 feature-complete.

Important: Code using C++17 features

Code using C++17 features should always be compiled with the compiler from the Development Tools Module. Linking two objects, such as an application and a shared library, which both use C++17, where one was built with g++ 8 or earlier and the other with g++ 9 or later, is particularly dangerous. This is because C++ STL objects instantiated by the experimental code may provide implementation and even ABI that is different from what the mature implementation expects and vice versa. Issues caused by such a mismatch are difficult to predict and may include silent data corruption.

Most of C++20 features are implemented in GCC 12 as experimental features. Try them out with appropriate caution and avoid linking together code that uses them and is produced by different compilers. Modules are only partially implemented ^[1] and require that the source file is compiled with -⁠fmodules-ts option. Similarly, coroutines ^[2] are also implemented but require that the source file is compiled with the -⁠fcoroutines switch. GCC 12 also experimentally implements many C++23 features. If you are interested in the implementation status of any particular C++ feature in the compiler or the standard library, consult the following pages:

• C++ Standards Support in GCC, and

• The GNU C++ Library Manual.

Advances in supporting new language specifications are not limited to C++. GCC 12 supports several new features from the ISO 202X C standard draft, and the Fortran compiler has also seen many improvements. And if you use OpenMP or OpenACC extensions for parallel programming, you will find that the compiler supports a lot of features of new versions of these standards. For more details, visit the links at the end of this section.

In addition to new supported language constructs, GCC 12 offers improved diagnostics when it reports errors and warnings to the user so that they are easier to understand and to be acted upon. This is particularly useful when dealing with issues in templated C++ code. Furthermore, there are several new warnings which help to avoid common programming mistakes.

Because GCC 12 is newer, it can generate code for many recent processors not supported by GCC 7. Such a list of processors would be too large to be displayed here. Nevertheless, in Section 7, “Performance evaluation: SPEC CPU 2017” we specifically look at optimizing code for an AMD EPYC 9004 Series Processor which is based on AMD Zen 4 cores. The system compiler does not know this kind of core and therefore cannot optimize for it. On the other hand, Zen 4 support has been backported to GCC 12.3 and thus it can often produce significantly faster code for it.

Finally, the general optimization pipeline of the compiler has also significantly improved over the years. To find out more about improvements in versions of GCC 8 through 12, visit their respective “changes” pages:

• GCC 8 Release Series Changes, New Features, and Fixes,

• GCC 9 Release Series Changes, New Features, and Fixes,

• GCC 10 Release Series Changes, New Features, and Fixes, and

• GCC 11 Release Series Changes, New Features, and Fixes.

• GCC 12 Release Series Changes, New Features, and Fixes.

2.3 Installing GCC 12 from the Development Tools Module #

  

Similar to other modules and extensions for SUSE Linux Enterprise Server 15, you can activate the Development Tools Module using either the command line tool SUSEConnect or the YaST setup and configuration tool. To use the former, carry out the following steps:

1. As root, start by listing the available and activated modules and extensions:

   sles15: # SUSEConnect --list-extensions

2. In the computer output, look for “Development Tools Module”:

               Development Tools Module 15 SP4 x86_64
               Activate with: SUSEConnect -p sle-module-development-tools/15.4/x86_64

   If you see the text (Activated) next to the module name, the module is already ready to be used. You can safely proceed to the installation of the compiler packages.

3. Otherwise, issue the activation command that is shown in the command output above:

   sles15: # SUSEConnect -p sle-module-development-tools/15.4/x86_64
   Registering system to SUSE Customer Center
   
   Updating system details on https://scc.suse.com ...
   
   Activating sle-module-development-tools 15.4 x86_64 ...
   -> Adding service to system ...
   -> Installing release package ...
   
   Successfully registered system

If you prefer to use YaST, the procedure is also straightforward. Run YaST as root and go to the Add-On Products menu in the Software section. If “Development Tools Module” is among the listed installed modules, you already have the module activated and can proceed with installing individual compiler packages. If not, click the Add button, select Select Extensions and Modules from Registration Server, and YaST will guide you through a simple procedure to add the module.

When you have the Development Tools Module installed, you can verify that the GCC 12 packages are available to be installed on your system:.

sles15: # zypper search gcc12
Refreshing service 'Basesystem_Module_15_SP4_x86_64'.
Refreshing service 'Certifications_Module_15_SP4_x86_64'.
Refreshing service 'Containers_Module_15_SP4_x86_64'.
Refreshing service 'Desktop_Applications_Module_15_SP4_x86_64'.
Refreshing service 'Development_Tools_Module_15_SP4_x86_64'.
Refreshing service 'Server_Applications_Module_15_SP4_x86_64'.
Refreshing service 'Web_and_Scripting_Module_15_SP4_x86_64'.
Loading repository data...
Reading installed packages...

S | Name                         | Summary
--+------------------------------+-------------------------------------------------------
  | gcc12                        | The GNU C Compiler and Support Files
  | gcc12-32bit                  | The GNU C Compiler 32bit support
  | gcc12-ada                    | GNU Ada Compiler Based on GCC (GNAT)
  | gcc12-ada-32bit              | GNU Ada Compiler Based on GCC (GNAT)
  | gcc12-c++                    | The GNU C++ Compiler
  | gcc12-c++-32bit              | The GNU C++ Compiler
  | gcc12-d                      | GNU D Compiler
  | gcc12-d-32bit                | GNU D Compiler
  | gcc12-fortran                | The GNU Fortran Compiler and Support Files
  | gcc12-fortran-32bit          | The GNU Fortran Compiler and Support Files
  | gcc12-go                     | GNU Go Compiler
  | gcc12-go-32bit               | GNU Go Compiler
  | gcc12-info                   | Documentation for the GNU compiler collection
  | gcc12-locale                 | Locale Data for the GNU Compiler Collection
  | gcc12-obj-c++                | GNU Objective C++ Compiler
  | gcc12-obj-c++-32bit          | GNU Objective C++ Compiler
  | gcc12-objc                   | GNU Objective C Compiler
  | gcc12-objc-32bit             | GNU Objective C Compiler
  | gcc12-PIE                    | A default configuration to build binaries in PIE mode
  | gcc12-testresults            | Testsuite results
  | libstdc++6-devel-gcc12       | Include Files and Libraries mandatory for Development
  | libstdc++6-devel-gcc12-32bit | Include Files and Libraries mandatory for Development

Now you can simply install the compilers for the programming languages you use with zypper:

sles15: # zypper install gcc12 gcc12-c++ gcc12-fortran

The compilers are installed on your system, the executables are called gcc-12, g++-12, gfortran-12 and so forth. It is also possible to install the packages in YaST. To do so, simply enter the “Software Management” menu in the Software section and search for “gcc12”. Then select the packages you want to install. Finally, click the Accept button.

Note: Newer compilers on openSUSE Leap 15.4

The community distribution openSUSE Leap 15.4 shares the base packages with SUSE Linux Enterprise Server 15 SP4. The system compiler on systems running openSUSE Leap 15.4 is also GCC 7.5. There is no Development Tools Module for the community distribution available, but a newer compiler is provided. Simply install the packages gcc12, gcc12-c++, gcc12-fortran, and the like.

GCC has a rich optimization pipeline that is controlled by approximately a hundred of command line options. It would be impractical to force users to decide about each one of them whether they want to have it enabled when compiling their code. Like all other modern compilers, GCC therefore introduces the concept of optimization levels which allow the user to pick a configuration from a few common ones. Optionally, the user can tweak the selected level, but that does not happen frequently.

The default is to not optimize. You can specify this optimization level on the command line as -⁠O0. It is often used when developing and debugging a project. This means it is usually accompanied with the command line switch -g so that debug information is emitted. As no optimizations take place, no information is lost because of it. No variables are optimized away, the compiler only inlines functions with special attributes that require it, and so forth. As a consequence, the debugger can almost always find everything it searches for in the running program and report on its state very well. On the other hand, the resulting code is big and slow. Thus this optimization level should not be used for release builds.

The most common optimization level for release builds is -⁠O2 which attempts to optimize the code aggressively but avoids large compile times and excessive code growth. Optimization level -⁠O3 instructs GCC to simply optimize as much as possible, even if the resulting code might be considerably bigger and the compilation can take longer. Note that neither -⁠O2 nor -⁠O3 imply anything about the precision and semantics of floating-point operations. Even at the optimization level -⁠O3 GCC implements math operations and functions so that they follow the respective IEEE and/or ISO rules^[3] with the exception of allowing floating-point expression contraction, for example when fusing an addition and a multiplication into one operation^[4]. This often means that the compiled programs run markedly slower than necessary if such strict adherence is not required. The command line switch -⁠ffast-math is a common way to relax rules governing floating-point operations. It is out of scope of this document to provide a list of the fine-grained options it enables and their meaning. However, if your software crunches floating-point numbers and its runtime is a priority, you can look them up in the GCC manual and review what semantics of floating-point operations you need.

The most aggressive optimization level is -⁠Ofast which does imply -⁠ffast-math along with a few options that disregard strict standard compliance. In GCC 12 this level also means the optimizers may introduce data races when moving memory stores which may not be safe for multithreaded applications and disregards the possibility of ELF symbol interposition happening at runtime. Additionally, the Fortran compiler can take advantage of associativity of math operations even across parentheses and convert big memory allocations on the heap to allocations on stack. The last mentioned transformation may cause the code to violate maximum stack size allowed by ulimit which is then reported to the user as a segmentation fault. We often use level -⁠Ofast to build benchmarks. It is a shorthand for the options on top of -⁠O3 which often make them run faster. Most benchmarks are intentionally written in a way that they run correctly even when these rules are relaxed.

If you feed the compiler with huge machine-generated input, especially if individual functions happen to be extremely large, the compile time can become an issue even when using -⁠O2. In such cases, use the most lightweight optimization level -⁠O1 that avoids running almost all optimizations with quadratic complexity. Finally, the -⁠Os level directs the compiler to aggressively optimize for the size of the binary.

Note: Optimization level recommendation

Usually we recommend using -⁠O2. This is the optimization level we use to build most SUSE and openSUSE packages, because at this level the compiler makes balanced size and speed trade-offs when building a general-purpose operating system. However, we suggest using -⁠O3 if you know that your project is compute-intensive and is either small or an important part of your actual workload. Moreover, if the compiled code contains performance-critical floating-point operations, we strongly advise that you investigate whether -⁠ffast-math or any of the fine-grained options it implies can be safely used.

If your project and the techniques you use to debug or instrument it do not depend on ELF symbol interposition, you may consider trying to speed it up by using -⁠fno-semantic-interposition. This allows the compiler to inline calls and propagate information even when it would be illegal if a symbol changed during dynamic linking. Using this option to signal to the compiler that interposition is not going to happen is known to significantly boost performance of some projects, most notably the Python interpreter.

Some projects use -⁠fno-strict-aliasing to work around type punning problems in the source code. This is not recommended except for very low-level hand-optimized code such as the Linux kernel. Type-based alias analysis is a very powerful tool. It is used to enable other transformations, such as store-to-load propagation that in turn enables other high level optimizations, such as aggressive inlining, vectorization and others.

With the -g switch GCC tries hard to generate useful debug information even when optimizing. However, a lot of information is irrecoverably lost in the process. Debuggers also often struggle to present the user with a view of the state of a program in which statements are not necessarily executed in the original order. Debugging optimized code can therefore be a challenging task but usually is still somewhat possible.

The complete list of optimization and other command line switches is available in the compiler manual. The manual is provided in the info format in the package gcc12-info or online at the GCC project Web site.

Bear in mind that although almost all optimizing compilers have the concept of optimization levels and their optimization levels often have the same names as those in GCC, they do not necessarily mean to make the same trade-offs. Famously, GCC's -⁠Os optimizes for size much more aggressively than LLVM/Clang's level with the same name. Therefore, it often produces slower code; the more equivalent option in Clang is -⁠Oz. Similarly, -⁠O2 can have different meanings for different compilers. For example, the difference between -⁠O2 and -⁠O3 is much bigger in GCC than in LLVM/Clang.

Note: Changing the optimization level with cmake

If you use cmake to configure and set up builds of your application, be aware that its release optimization level defaults to -⁠O3 which might not be what you want. To change it, you must modify the CMAKE_C_FLAGS_RELEASE, CMAKE_CXX_FLAGS_RELEASE and/or CMAKE_Fortran_FLAGS_RELEASE variables. Since they are appended at the end of the compilation command lines, they are overwriting any level set in the variables CMAKE_C_FLAGS, CMAKE_CXX_FLAGS, and the like.

By default GCC assumes that you want to run the compiled program on a wide variety of CPUs, including fairly old ones, regardless of the selected optimization level. On architectures like x86_64 and aarch64 the generated code will only contain instructions available on every CPU model of the architecture, including the earliest ones. On x86_64 in particular this means that the programs will use the SSE and SSE2 instruction sets for floating-point and vector operations but not any more recent ones.

If you know that the generated binary will run only on machines supporting newer instruction set extensions, you can specify it on the command line. Their complete list is available in the manual, but the most prominent one is -⁠march which lets you select a CPU model to generate code for. For example, if you know that your program will only be executed on AMD EPYC 9004 Series Processors based on AMD Zen 4 cores or processors that are compatible with it, you can instruct GCC to take advantage of all the instructions the CPU supports with option -⁠march=znver4. Note that on SUSE Linux Enterprise Server 15, the system compiler does not know this particular value of the switch; you need to use GCC 12 from the Development Tools Module to optimize code for these processors.

To run the program on the machine on which you are compiling it, you can have the compiler auto-detect the target CPU model for you with the option -⁠march=native. This only works if the compiler is new enough. The system compiler of SUSE Linux Enterprise Server, for example, misidentifies AMD EPYC 9004 Series Processors as being based on the AMD Zen 1 core. Among other things, this means that it only emits 128 bit vector instructions, even though the CPU has data-paths wide enough to efficiently process 512 bit ones. Again, the easy solution is to use the compiler from the Development Tools Module when targeting recent processors.

Note: Running 32-bit code

SUSE Linux Enterprise Server does not support compilation of 32-bit applications, it only offers runtime support for 32-bit binaries. To do so, you will need 32-bit libraries your binary depends on which likely include at least glibc which can be found in package glibc-32bit. See chapter 20 (32-bit and 64-bit applications in a 64-bit system environment) of the Administration Guide for more information.

Figure 1 outlines the classic mode of operation of a compiler and a linker. Pieces of a program are compiled and optimized in chunks defined by the user called compilation units to produce so-called object files. These object files already contain binary machine instructions and are combined together by a linker. Because the linker works at such low level, it cannot perform much optimization and the division of the program into compilation units thus presents a profound barrier to optimization.

Figure 1: Traditional program build #

  

This limitation can be overcome by rearranging the process so that the linker does not receive as its input the almost finished object files containing machine instructions, but is invoked on files containing so called intermediate language (IL). This is a much richer representation of each original compilation unit (see figure figure 2). The linker identifies the input as not yet entirely compiled and invokes a linker plugin which in turn runs the compiler again. But this time it has at its disposal the representation of the entire program or library that is being built. The compiler makes decisions about what optimizations across function and compilation unit boundaries will be carried out and then divides the program into a set of partitions. Each of the partitions is further optimized independently, and machine code is emitted for it, which is finally linked the traditional way. Processing of the partitions is performed in parallel.

Figure 2: Building a program with GCC using Link Time Optimization (LTO) #

  

To use Link Time Optimization, all you need do is to add the -⁠flto switch to the compilation command line. The vast majority of packages in the Linux distribution openSUSE Tumbleweed has been built with LTO for over three years without any major problems. A lot of work has been put into emitting good debug information when building with LTO too. Thus the debugging experience is not severely limited anymore as it was five years ago.

LTO in GCC always consists of a whole program analysis (WPA) stage followed by the majority of the compilation process performed in parallel, which greatly reduces the build times of most projects. To control the parallelism, you can explicitly cap the number of parallel compilation processes by n if you specify -⁠flto=n at linker command line. Alternatively, it is possible to use the GNU make jobserver with -⁠flto=jobserv while also prepending the makefile rule invoking link step with character + to instruct GNU make to keep the jobserver available to the linker process. You can also use -⁠flto=auto which instructs GCC to search for the jobserver and if it is not found, use all available CPU threads.

Note that there is a principal architectural difference in how GCC and LLVM/Clang approach LTO. Clang provides two LTO mechanisms, so-called thin LTO and full LTO. In full LTO, LLVM processes the whole program as if it was a single translation unit which does not allow for any parallelism. GCC can be configured to operate in this way with the option -⁠flto-partition=one. LLVM in thin LTO mode can compile different compilation units in parallel and makes possible inlining across compilation unit boundaries, but not most other types of cross-module optimizations. This mechanism therefore has inherently higher code quality penalty than full LTO or the approach of GCC.

5.1 Most notable benefits of LTO #

  

Applications built with LTO are often faster, mainly because the compiler can inline calls to functions in another compilation unit. This possibility also allows programmers to structure their code according to its logical division because they are not forced to put function definitions into header files to enable their inlining. Because the compiler cannot inline all calls conveying information known at compilation time, GCC tracks and propagates constants, value ranges and devirtualization contexts from the call sites to the callees, often even when passed in an aggregate or by reference. These can then subsequently save unnecessary computations. LTO allows such propagation across compilation unit boundaries, too.

Link Time Optimization with whole program analysis also offers many opportunities to shrink the code size of the built project. Thanks to symbol promotion and inter-procedural unreachable code elimination, functions and their parts which are not necessary in any particular project can be removed even when they are not declared static and are not defined in an anonymous namespace. Automatic attribute discovery can identify C++ functions that do not throw exceptions which allows the compiler to avoid generating a lot of code in exception cleanup regions. Identical code folding can find functions with the same semantics and remove all but one of them. The code size savings are often very significant and a compelling reason to use LTO even for applications which are not CPU-bound.

Note: Building libraries with LTO

The symbol promotion is controlled by resolution information given to the linker and depends on type of the DSO build. When producing a dynamically loaded shared library, all symbols with default visibility can be overwritten by the dynamic linker. This blocks the promotion of all functions not declared inline, thus it is necessary to use the hidden visibility wherever possible to achieve best results. Similar problems happen even when building static libraries with -rdynamic.

__attribute__ ((__symver__ ("foo@VERS_1")))
int foo_v1 (void)
{
}

Optimizing compilers frequently make decisions that depend on which path through the code they consider most likely to be executed, how many times a loop is expected to iterate, and similar estimates. They also often face trade-offs between potential runtime benefits and code size growth. Ideally, they would optimize only frequently executed (also called hot) bits of a program for speed and everything else for size to reduce strain on caches and make the distribution of the built software cheaper. Unfortunately, guessing which parts of a program are the hot ones is difficult, and even sophisticated estimation algorithms implemented in GCC are no match for a measurement.

If you do not mind adding an extra level of complexity to the build system of your project, you can make such measurement part of the process. The makefile (or any other) build script needs to compile the project twice. The first time it needs to compile with the -⁠fprofile-generate option and then execute the resulting binary in one or multiple train runs during which it will save information about the behavior of the program to special files. Afterward, the project needs to be rebuilt again, this time with the -⁠fprofile-use option. This instructs the compiler to look for the files with the measurements and use them when making optimization decisions, a process called Profile-Guided Optimization (PGO).

It is important that the train run exhibits the same characteristics as the real workload. Unless you use the option -⁠fprofile-partial-training in the second build, it needs to exercise the code that is also the most frequently executed in real use, otherwise it will be optimized for size and PGO would make more harm than good. With the option, GCC reverts to guessing properties of portions of the projects not exercised in the train run, as if they were compiled without profile feedback. This however also means that this code will not perform better or shrink as much as one would expect from a PGO build.

On the other hand, train runs do not need to be a perfect simulation of the real workload. For example, even though a test suite should not be a very good train run in theory because it disproportionally often tests various corner cases, in practice many projects use it as a train run and achieve significant runtime improvements with real workloads, too.

Profiles collected using an instrumented binary for multithreaded programs may be inconsistent because of missed counter updates. You can use -⁠fprofile-correction in addition to -⁠fprofile-use so that GCC uses heuristics to correct or smooth out such inconsistencies instead of emitting an error.

Profile-Guided Optimization can be combined and is complimentary to Link Time Optimization. While LTO expands what the compiler can do, PGO informs it about which parts of the program are the important ones and should be focused on. The case study in the following section shows how the two techniques work with each other on a well-known set of benchmarks.

Standard Performance Evaluation Corporation (SPEC) is a non-profit corporation that publishes a variety of industry standard benchmarks to evaluate performance and other characteristics of computer systems. Its latest suite of CPU intensive workloads, SPEC CPU 2017, is often used to compare compilers and how well they optimize code with different settings. This is because the included benchmarks are well known and represent a wide variety of computation-heavy programs. The following section highlights selected results of a GCC 12 evaluation using the suite.

Note that when we use SPEC to perform compiler comparisons, we are lenient toward some official SPEC rules which system manufacturers need to observe to claim an official score for their system. We disregard the concepts of base and peak metrics and simply focus on results of compilations using a particular set of options. We even patched several benchmarks:

For these and other reasons, all the results in this document are non-reportable. Finally, SPEC 2017 CPU offers so-called speed and rate metrics. For our purposes, we mostly ignore the differences and simply run the benchmarks configured for rate metrics (mainly because the runtimes are smaller) but we always run all benchmarks single-threaded.

SPEC specifies a base runtime for each benchmark and defines a rate as the ratio of the base runtime and the median measured runtime (this rate is a separate concept from the rate metrics). The overall suite score is then calculated as geometric mean of these ratios. The bigger the rate or score, the better it is. In the remainder of this section, we report runtimes using relative rates and their geometric means as they were measured on an AMD EPYC 9654 Processor running SUSE Linux Enterprise Server 15 SP4.

7.1 Benefits of LTO and PGO #

  

In Section 3, “Optimization levels and related options” we recommend that HPC workloads are compiled with -⁠O3 and benchmarks with -⁠Ofast. But it is still interesting to look at integer crunching benchmarks built with only -⁠O2 because that is how Linux distributions often build the programs from which they were extracted. We have already mentioned that almost the whole openSUSE Tumbleweed distribution is now built with LTO, and selected packages with PGO, and the following paragraphs demonstrate why.

Figure 3: Overall performance (bigger is better) of SPEC INTrate 2017 built with GCC 12.3 and -⁠O2 #

  

Figure 3 shows the overall performance effect on the whole integer benchmark suite as captured by the geometric mean of all individual benchmark rates. The relative uplift is no longer as remarkable as with the previous versions of GCC because GCC 12 can conservatively vectorize code in 525.x264_r also at plain -⁠O2. As a consequence, the benchmark, which in practice is usually compiled with -⁠O3, runs 37% faster than when compiled with GCC 11 and the same optimization level. Nevertheless, it still benefits from the more advanced modes of compilation a lot, together with several other benchmarks which are derived from programs that are typically compiled with -⁠O2. This is illustrated in figure 4.

Figure 4: Runtime performance (bigger is better) of individual integer benchmarks built with GCC 12.3 and -⁠O2 #

  

Figure 5 shows another important advantage of LTO and PGO which is significant reduction of the size of the binaries (measured without debug info). Note that it does not depict that the size of benchmark 548.exchange2_r grew to 290% and 200% of the original size when built with PGO or both PGO and LTO respectively, which looks huge but the growth is from a particularly small base. It is the only Fortran benchmark in the integer suite and, most importantly, the size penalty is offset by significant speed-up, making the trade-off reasonable. For completeness, we show this result in figure 6

Figure 5: Binary size (smaller is better) of individual integer benchmarks built with GCC 12.3 and -⁠O2 #

  

Figure 6: Binary size (smaller is better) of 548.exchange2_r built with GCC 12.3 and -⁠O2 #

  

The runtime benefits and binary size savings are also substantial when using the optimization level -⁠Ofast and option -⁠march=native to allow the compiler to take full advantage of all instructions that the AMD EPYC 9654 Processor supports. Figure 7 shows the respective geometric means, and figure 8 shows how rates improve for individual benchmarks. Moreover, even though optimization levels -⁠O3 and -⁠Ofast are permitted to be relaxed about the final binary size, PGO and especially LTO can bring it nicely down at these levels, too. Figure 9 depicts the relative binary sizes of all integer benchmarks.

Figure 7: Overall performance (bigger is better) of SPEC INTrate 2017 built with GCC 12.3 using -⁠Ofast and -⁠march=native #

  

Figure 8: Runtime performance (bigger is better) of individual integer benchmarks built with GCC 12.3 using -⁠Ofast and -⁠march=native #

  

Figure 9: Binary size (smaller is better) of SPEC INTrate 2017 built with GCC 12.3 using -⁠Ofast and -⁠march=native #

  

Many of the SPEC 2017 floating-point benchmarks measure how well a given system can optimize and execute a handful of number crunching loops. They often come from performance sensitive programs written with traditional compilation method in mind. Consequently there are fewer cross-module dependencies, identifying hot paths is less crucial and the overall effect of LTO and PGO suite only improves by 5% (see figure 11). Nevertheless, there are important cases when these modes of compilation also bring about significant performance increases. Figure 11 shows the effect of these methods on individual benchmarks when compiled at -⁠Ofast and targeting the full ISA of the AMD EPYC 9654 Processor. Furthermore, binary size savings of PGO and LTO are sometimes even bigger than those achieved on integer benchmarks, as can be seen on figure 12

Figure 10: Overall performance (bigger is better) of SPEC FPrate 2017 built with GCC 12.3 and -⁠Ofast #

  

Figure 11: Runtime performance (bigger is better) of individual floating-point benchmarks built with GCC 12.3 using -⁠Ofast and -⁠march=native #

  

Figure 12: Binary size (smaller is better) of SPEC FPrate 2017 built with GCC 12.3 using -⁠Ofast and -⁠march=native #

  

7.2 GCC 12.3 compared to GCC 7.5 #

  

In previous sections we have recommended the use of GCC 12.3 from the Development Tools Module over the system compiler. Among other reasons, we did so because of its more powerful optimization pipeline and its support for newer CPUs. This section compares SPEC CPU 2017 obtained with GCC 7.5, which corresponds to the system compiler in SUSE Linux Enterprise Server 15, and GCC 12.3 on an AMD EPYC 9654 Processor, when all benchmarks are compiled with -⁠Ofast and -⁠march=native. Note that the latter option means that both compilers differ in their CPU targets because GCC 7.5 does not know the Zen 4 core. This in turn means that in large part the optimization benefits presented here exist because the old compiler only issues 128bit (AVX2) vector operations whereas the newer one can take full advantage of AVX512. Nevertheless, be aware that simply using wider vectors everywhere often backfires. GCC has made substantial advancements over the recent years to avoid such issues, both in its vectorizer and other optimizers. It is therefore much better placed to use the extra vector width appropriately and produce code which utilizes the processor better in general.

Figure 13: Overall performance (bigger is better) of SPEC INTrate 2017 built with GCC 7.5 and 12.3 (-⁠Ofast -⁠march=native) #

  

Figure 13 captures the benefits of using the modern compiler with integer workloads in the form of relative improvements of the geometric mean of the whole SPEC INTrate 2017 suite. Figure 14 dives deeper and shows which particular benchmarks gained most in terms of performance. It was already mentioned that 525.x264_r especially benefits from vectorization and therefore it is not surprising it has improved a lot. 531.deepsjeng_r is faster chiefly because it can emit better code for count trailing zeros (CTZ) operation which it performs frequently. Finally, modern GCC can optimize 548.exchange2_r particularly well by specializing different invocations of the hottest recursive function and it also clearly shows in the picture.

Figure 14: Runtime performance (bigger is better) of selected integer benchmarks built with GCC 7.5 and 12.3 (-⁠Ofast -⁠march=native) #

  

Floating-point computations tend to particularly benefit from vectorization advancements. Thus it should be no surprise that the FPrate benchmarks also improve substantially when compiled with GCC 12.3, which also emits AVX512 instructions for a Zen 4 based CPU. The overall boost is shown in figure 15 whereas figure 16 provides a detailed look at which benchmarks contributed most to the overall score difference.

Figure 15: Overall performance (bigger is better) of SPEC FPrate 2017 built with GCC 7.5 and 12.3 (-⁠Ofast -⁠march=native) #

  

Figure 16: Runtime performance (bigger is better) of selected floating-point benchmarks built with GCC 7.5 and 12.3 (-⁠Ofast -⁠march=native) #

  

7.3 Effects of -⁠ffast-math on floating-point performance #

  

In Section 3, “Optimization levels and related options” we pointed out that, if you do not relax the semantics of floating-point math functions even though you do not need strict adherence to all respective IEEE and/or ISO rules, you are likely to be leaving some performance on the table. This section uses the SPEC FPrate 2017 test suite to illustrate how much performance that might be.

We have built the benchmarking suite using optimization level -⁠O3, LTO (though without PGO) and -⁠march=native to target the native ISA of our AMD EPYC 9654 Processor. Then we compared its runtime score against the suite built with these options and -⁠ffast-math. As you can see in figure 17, the geometric mean grew by over 13%. But a quick look at figure 18 will tell you that there are four benchmarks with scores which improved by more than 20% and that of 510.parest_r grew by over 76%.

Figure 17: Overall performance (bigger is better) of SPEC FPrate 2017 built with GCC 12.3 and -⁠O3 -⁠flto -⁠march=native, without and with -⁠ffast-math #

  

Figure 18: Runtime performance (bigger is better) of selected floating-point benchmarks built with GCC 12.3 and -⁠O3 -⁠flto -⁠march=native, without and with -⁠ffast-math #

  

7.4 Comparison with other compilers #

  

The toolchain team at SUSE regularly uses the SPEC CPU 2017 suite to compare the optimization capabilities of GCC with other compilers, mainly LLVM/Clang and ICC and ICX from Intel. In the final section of this case study we will share how the Development Module compiler stands compared to these competitors on SUSE Linux Enterprise Server 15 SP4. Before we start, we should emphasize that the comparison has been carried out by people who have much better knowledge of GCC than of the other compilers and are not “unbiased”. Also, keep in mind that everything we explained previously about how we carry out the measurements and patch the benchmarks also applies to this section. On the other hand, the results often guide our own work and therefore we strive to be accurate.

LLVM/Clang 16.0 now comes with a new Fortran front-end called flang-new which is capable of compiling SPEC, but we were not able to successfully run 527.cam4_r benchmark compiled with it and LTO. Comparison with LLVM in this report is therefore incomplete but for the first time we were able to include the rest of the benchmarks using Fortran in our comparison with LLVM/Clang.

We have built the clang and clang++ compilers from sources obtained from the official git repository (tag llvmorg-16.0.1), used it to compile the SPEC CPU 2017 suite with -⁠Ofast and -⁠march=native and compared the performance against the suites built with GCC 12.3 with the same options. When using Clang's LTO to compile SPEC, we selected the full variant.

Figure 19: Overall performance (bigger is better) of C/C++ integer benchmarks built with Clang 16 and GCC 12.3 #

  

Figure 19 shows that the geometric mean of the whole SPEC INTrate 2017 suite is quite substantially better when the benchmarks are compiled with GCC. To be fair, a disproportionate amount of the difference is because GNU Fortran can optimize 548.exchange2_r much better than LLVM. Given that the LLVM Fortran front-end is very new and the optimization opportunities in this particular benchmark are quite specific, the result may not be important for many users.

Figure 20: Runtime performance (bigger is better) of 548.exchange2_r benchmarks built with Clang 16 and GCC 12.3 #

  

Figure 21: Runtime performance (bigger is better) of C/C++ integer benchmarks built with Clang 16 and GCC 12.3 #

  

Figure 21 shows relative rates of integer benchmarks written in C/C++ and the compilers perform fairly similarly there. GCC wins by a large margin on 500.perlbench_r but loses significantly when compiling 525.x264_r. This is because the compiler chooses a vectorizing factor that is too large for the important loops in this video encoder. It is possible to mitigate the problem using compiler option -⁠mprefer-⁠vector-⁠width=128, with which it is again competitive, as you can see in figure 22. This problem is being actively worked on by the upstream GCC community. We plan to use masked vectorized epilogues to minimize the fallout of choosing a large vectorizing factor for the principal vector loop. Note that PGO can substantially help in this case too.

Figure 22: Runtime performance (bigger is better) of 525.x264_r benchmark built with Clang 16 and with GCC 12.3 using -mprefer-vector-width=128 #

  

Because we were not able to successfully run 527.cam4_r benchmark compiled with LLVM with LTO, we have excluded the benchmark in our comparison of geometric mean of SPEC FPrate 2017 suite depicted in figure 23. The floating point benchmark suite contains many more Fortran benchmarks. It can be seen that GCC has advantage in having a mature optimization pipeline for this language as well, especially when compiling 503.bwaves_r, 510.parest_r, 549.fotonik3d_r, 554.roms_r (see figure 24) and the already mentioned 527.cam4_r (see figure 25). The comparison also shows that the performance of 538.imagick_r when compiled with GCC 12.3 is substantially smaller. This is caused by store-to-load forwarding stall issues, which can be mitigated by relaxing inlining limits, something that GCC 13 does automatically.

Figure 23: Overall performance (bigger is better) of SPEC FPrate 2017 excluding 527.cam4_r built with Clang 16 and GCC 12.3 #

  

Figure 24: Runtime performance (bigger is better) of floating point benchmarks built with Clang 16 and GCC 12.3 #

  

Figure 25: Runtime performance (bigger is better) of 527.cam4_r benchmark built with Clang 16 and GCC 12.3 #

  

Even though ICC is not intended as a compiler for AMD processors, it is known for its high-level optimization capabilities, especially when it comes to vectorization. Therefore we have traditionally included it our comparisons of compilers. Recently, however, Intel has decided to abandon this compiler and is directing its users toward ICX, a new one built on top of LLVM. This year we have therefore included not just ICC 2021.9.0 (20230302) but also ICX 2023.1.0 in our comparison. To keep the amount of presented data in the rest of this section reasonable, we only compare binaries built with -⁠Ofast and LTO. We have simply passed -⁠march=native GCC and ICX. On the other hand, we have used -⁠march=core-avx2 option to specify the target ISA for the old ICC because it is unclear which option is the most appropriate for AMD EPYC 9654 Processor. This puts this compiler at a disadvantage because it can only emit AVX256 instructions while the other two can, and GCC does, make use of AVX512. We believe that the comparison is still useful as ICC serves mainly as a base and the focus now shifts to ICX but keep this in mind when looking at the results below.

Figure 26: Overall performance (bigger is better) of SPEC INTrate 2017 built with ICC 2021.9.0, ICX 2023.1.0 and GCC 12.3 #

  

Figure 26 shows that the new ICX compiler takes the lead in overall SPEC INTrate assessment. The results of individual benchmarks however quickly show that the majority of the lead is due to one benchmark, 525.x264_r, and for the same reasons we outlined when discussing LLVM/Clang results. GCC picks too large vectorizing factor and the mitigation is again using -⁠mprefer-⁠vector-⁠width=128 which leads to a much narrower gap (see figure 28). When looking at the other benchmarks, GCC achieves comparable results.

Figure 27: Runtime performance (bigger is better) of individual integer benchmarks built with ICC 2021.9.0, ICX 2023.1.0 and GCC 12.3 #

  

Figure 28: Runtime performance (bigger is better) of 525.x264_r benchmark built with ICC 2021.9.0, ICX 2023.1.0 and with GCC 12.3 using -mprefer-vector-width=128 #

  

Comparison with ICX on SPEC FPrate suite has been hampered by the fact that again there is a benchmark which did not run correctly, this time it was 521.wrf_r. Therefore we have calculated the geometric means of rates for figure 29 excluding it.

Figure 29: Overall performance (bigger is better) of SPEC FPrate 2017 excluding 521.wrf_r built with ICC 2021.9.0, ICX 2023.1.0 and GCC 12.3 #

  

While GCC achieves the best geometric mean, it is important to look at individual results too. The overall picture is mixed (see figure 30), as each of the three compilers managed to be the fastest in at least one benchmark. We do not know the reason for rather poor performance of ICX on 554.roms_r. But we have seen a similar issue with the compiler on an Intel Cascadelake server machine too, so it is not a consequence of using an Intel compiler on an AMD platform. For completeness, 521.wrf_r results for ICC and ICX are provided in figure 31. In conclusion, GCC manages to perform consistently and competitively against these high-performance compilers.

Figure 30: Runtime performance (bigger is better) of individual floating point benchmarks built with ICC 2021.9.0, ICX 2023.1.0 and GCC 12.3 #

  

Figure 31: Runtime performance (bigger is better) of 521.wrf_r built with ICC 2021.9.0 and GCC 12.3 #

  

Copyright ©2006-2024 SUSE LLC and contributors. All rights reserved.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or (at your option) version 1.3; with the Invariant Section being this copyright notice and license. A copy of the license version 1.2 is included in the section entitled “GNU Free Documentation License”.

SUSE, the SUSE logo and YaST are registered trademarks of SUSE LLC in the United States and other countries. For SUSE trademarks, see http://www.suse.com/company/legal/. Linux is a registered trademark of Linus Torvalds. All other names or trademarks mentioned in this document may be trademarks or registered trademarks of their respective owners.

Documents published as part of the SUSE Best Practices series have been contributed voluntarily by SUSE employees and third parties. They are meant to serve as examples of how particular actions can be performed. They have been compiled with utmost attention to detail. However, this does not guarantee complete accuracy. SUSE cannot verify that actions described in these documents do what is claimed or whether actions described have unintended consequences. SUSE LLC, its affiliates, the authors, and the translators may not be held liable for possible errors or the consequences thereof.

Below we draw your attention to the license under which the articles are published.

Copyright (C) 2000, 2001, 2002 Free Software Foundation, Inc. 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed.

The purpose of this License is to make a manual, textbook, or other functional and useful document "free" in the sense of freedom: to assure everyone the effective freedom to copy and redistribute it, with or without modifying it, either commercially or non-commercially. Secondarily, this License preserves for the author and publisher a way to get credit for their work, while not being considered responsible for modifications made by others.

This License is a kind of "copyleft", which means that derivative works of the document must themselves be free in the same sense. It complements the GNU General Public License, which is a copyleft license designed for free software.

We have designed this License to use it for manuals for free software, because free software needs free documentation: a free program should come with manuals providing the same freedoms that the software does. But this License is not limited to software manuals; it can be used for any textual work, regardless of subject matter or whether it is published as a printed book. We recommend this License principally for works whose purpose is instruction or reference.

This License applies to any manual or other work, in any medium, that contains a notice placed by the copyright holder saying it can be distributed under the terms of this License. Such a notice grants a world-wide, royalty-free license, unlimited in duration, to use that work under the conditions stated herein. The "Document", below, refers to any such manual or work. Any member of the public is a licensee, and is addressed as "you". You accept the license if you copy, modify or distribute the work in a way requiring permission under copyright law.

A "Modified Version" of the Document means any work containing the Document or a portion of it, either copied verbatim, or with modifications and/or translated into another language.

A "Secondary Section" is a named appendix or a front-matter section of the Document that deals exclusively with the relationship of the publishers or authors of the Document to the Document's overall subject (or to related matters) and contains nothing that could fall directly within that overall subject. (Thus, if the Document is in part a textbook of mathematics, a Secondary Section may not explain any mathematics.) The relationship could be a matter of historical connection with the subject or with related matters, or of legal, commercial, philosophical, ethical or political position regarding them.

The "Invariant Sections" are certain Secondary Sections whose titles are designated, as being those of Invariant Sections, in the notice that says that the Document is released under this License. If a section does not fit the above definition of Secondary then it is not allowed to be designated as Invariant. The Document may contain zero Invariant Sections. If the Document does not identify any Invariant Sections then there are none.

The "Cover Texts" are certain short passages of text that are listed, as Front-Cover Texts or Back-Cover Texts, in the notice that says that the Document is released under this License. A Front-Cover Text may be at most 5 words, and a Back-Cover Text may be at most 25 words.

A "Transparent" copy of the Document means a machine-readable copy, represented in a format whose specification is available to the general public, that is suitable for revising the document straightforwardly with generic text editors or (for images composed of pixels) generic paint programs or (for drawings) some widely available drawing editor, and that is suitable for input to text formatters or for automatic translation to a variety of formats suitable for input to text formatters. A copy made in an otherwise Transparent file format whose markup, or absence of markup, has been arranged to thwart or discourage subsequent modification by readers is not Transparent. An image format is not Transparent if used for any substantial amount of text. A copy that is not "Transparent" is called "Opaque".

Examples of suitable formats for Transparent copies include plain ASCII without markup, Texinfo input format, LaTeX input format, SGML or XML using a publicly available DTD, and standard-conforming simple HTML, PostScript or PDF designed for human modification. Examples of transparent image formats include PNG, XCF and JPG. Opaque formats include proprietary formats that can be read and edited only by proprietary word processors, SGML or XML for which the DTD and/or processing tools are not generally available, and the machine-generated HTML, PostScript or PDF produced by some word processors for output purposes only.

The "Title Page" means, for a printed book, the title page itself, plus such following pages as are needed to hold, legibly, the material this License requires to appear in the title page. For works in formats which do not have any title page as such, "Title Page" means the text near the most prominent appearance of the work's title, preceding the beginning of the body of the text.

A section "Entitled XYZ" means a named subunit of the Document whose title either is precisely XYZ or contains XYZ in parentheses following text that translates XYZ in another language. (Here XYZ stands for a specific section name mentioned below, such as "Acknowledgements", "Dedications", "Endorsements", or "History".) To "Preserve the Title" of such a section when you modify the Document means that it remains a section "Entitled XYZ" according to this definition.

The Document may include Warranty Disclaimers next to the notice which states that this License applies to the Document. These Warranty Disclaimers are considered to be included by reference in this License, but only as regards disclaiming warranties: any other implication that these Warranty Disclaimers may have is void and has no effect on the meaning of this License.

You may copy and distribute the Document in any medium, either commercially or noncommercially, provided that this License, the copyright notices, and the license notice saying this License applies to the Document are reproduced in all copies, and that you add no other conditions whatsoever to those of this License. You may not use technical measures to obstruct or control the reading or further copying of the copies you make or distribute. However, you may accept compensation in exchange for copies. If you distribute a large enough number of copies you must also follow the conditions in section 3.

You may also lend copies, under the same conditions stated above, and you may publicly display copies.

If you publish printed copies (or copies in media that commonly have printed covers) of the Document, numbering more than 100, and the Document's license notice requires Cover Texts, you must enclose the copies in covers that carry, clearly and legibly, all these Cover Texts: Front-Cover Texts on the front cover, and Back-Cover Texts on the back cover. Both covers must also clearly and legibly identify you as the publisher of these copies. The front cover must present the full title with all words of the title equally prominent and visible. You may add other material on the covers in addition. Copying with changes limited to the covers, as long as they preserve the title of the Document and satisfy these conditions, can be treated as verbatim copying in other respects.

If the required texts for either cover are too voluminous to fit legibly, you should put the first ones listed (as many as fit reasonably) on the actual cover, and continue the rest onto adjacent pages.

If you publish or distribute Opaque copies of the Document numbering more than 100, you must either include a machine-readable Transparent copy along with each Opaque copy, or state in or with each Opaque copy a computer-network location from which the general network-using public has access to download using public-standard network protocols a complete Transparent copy of the Document, free of added material. If you use the latter option, you must take reasonably prudent steps, when you begin distribution of Opaque copies in quantity, to ensure that this Transparent copy will remain thus accessible at the stated location until at least one year after the last time you distribute an Opaque copy (directly or through your agents or retailers) of that edition to the public.

It is requested, but not required, that you contact the authors of the Document well before redistributing any large number of copies, to give them a chance to provide you with an updated version of the Document.

You may copy and distribute a Modified Version of the Document under the conditions of sections 2 and 3 above, provided that you release the Modified Version under precisely this License, with the Modified Version filling the role of the Document, thus licensing distribution and modification of the Modified Version to whoever possesses a copy of it. In addition, you must do these things in the Modified Version:

If the Modified Version includes new front-matter sections or appendices that qualify as Secondary Sections and contain no material copied from the Document, you may at your option designate some or all of these sections as invariant. To do this, add their titles to the list of Invariant Sections in the Modified Version's license notice. These titles must be distinct from any other section titles.

You may add a section Entitled "Endorsements", provided it contains nothing but endorsements of your Modified Version by various parties--for example, statements of peer review or that the text has been approved by an organization as the authoritative definition of a standard.

You may add a passage of up to five words as a Front-Cover Text, and a passage of up to 25 words as a Back-Cover Text, to the end of the list of Cover Texts in the Modified Version. Only one passage of Front-Cover Text and one of Back-Cover Text may be added by (or through arrangements made by) any one entity. If the Document already includes a cover text for the same cover, previously added by you or by arrangement made by the same entity you are acting on behalf of, you may not add another; but you may replace the old one, on explicit permission from the previous publisher that added the old one.

The author(s) and publisher(s) of the Document do not by this License give permission to use their names for publicity for or to assert or imply endorsement of any Modified Version.

You may combine the Document with other documents released under this License, under the terms defined in section 4 above for modified versions, provided that you include in the combination all of the Invariant Sections of all of the original documents, unmodified, and list them all as Invariant Sections of your combined work in its license notice, and that you preserve all their Warranty Disclaimers.

The combined work need only contain one copy of this License, and multiple identical Invariant Sections may be replaced with a single copy. If there are multiple Invariant Sections with the same name but different contents, make the title of each such section unique by adding at the end of it, in parentheses, the name of the original author or publisher of that section if known, or else a unique number. Make the same adjustment to the section titles in the list of Invariant Sections in the license notice of the combined work.

In the combination, you must combine any sections Entitled "History" in the various original documents, forming one section Entitled "History"; likewise combine any sections Entitled "Acknowledgements", and any sections Entitled "Dedications". You must delete all sections Entitled "Endorsements".

You may make a collection consisting of the Document and other documents released under this License, and replace the individual copies of this License in the various documents with a single copy that is included in the collection, provided that you follow the rules of this License for verbatim copying of each of the documents in all other respects.

You may extract a single document from such a collection, and distribute it individually under this License, provided you insert a copy of this License into the extracted document, and follow this License in all other respects regarding verbatim copying of that document.

A compilation of the Document or its derivatives with other separate and independent documents or works, in or on a volume of a storage or distribution medium, is called an "aggregate" if the copyright resulting from the compilation is not used to limit the legal rights of the compilation's users beyond what the individual works permit. When the Document is included in an aggregate, this License does not apply to the other works in the aggregate which are not themselves derivative works of the Document.

If the Cover Text requirement of section 3 is applicable to these copies of the Document, then if the Document is less than one half of the entire aggregate, the Document's Cover Texts may be placed on covers that bracket the Document within the aggregate, or the electronic equivalent of covers if the Document is in electronic form. Otherwise they must appear on printed covers that bracket the whole aggregate.

Translation is considered a kind of modification, so you may distribute translations of the Document under the terms of section 4. Replacing Invariant Sections with translations requires special permission from their copyright holders, but you may include translations of some or all Invariant Sections in addition to the original versions of these Invariant Sections. You may include a translation of this License, and all the license notices in the Document, and any Warranty Disclaimers, provided that you also include the original English version of this License and the original versions of those notices and disclaimers. In case of a disagreement between the translation and the original version of this License or a notice or disclaimer, the original version will prevail.

If a section in the Document is Entitled "Acknowledgements", "Dedications", or "History", the requirement (section 4) to Preserve its Title (section 1) will typically require changing the actual title.

You may not copy, modify, sublicense, or distribute the Document except as expressly provided for under this License. Any other attempt to copy, modify, sublicense or distribute the Document is void, and will automatically terminate your rights under this License. However, parties who have received copies, or rights, from you under this License will not have their licenses terminated so long as such parties remain in full compliance.

The Free Software Foundation may publish new, revised versions of the GNU Free Documentation License from time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address new problems or concerns. See http://www.gnu.org/copyleft/.

Each version of the License is given a distinguishing version number. If the Document specifies that a particular numbered version of this License "or any later version" applies to it, you have the option of following the terms and conditions either of that specified version or of any later version that has been published (not as a draft) by the Free Software Foundation. If the Document does not specify a version number of this License, you may choose any version ever published (not as a draft) by the Free Software Foundation.

Copyright (c) YEAR YOUR NAME.
Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License, Version 1.2
or any later version published by the Free Software Foundation;
with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts.
A copy of the license is included in the section entitled "GNU
 Free Documentation License".

If you have Invariant Sections, Front-Cover Texts and Back-Cover Texts, replace the "with...Texts". line with this:

with the Invariant Sections being LIST THEIR TITLES, with the
Front-Cover Texts being LIST, and with the Back-Cover Texts being LIST.

If you have Invariant Sections without Cover Texts, or some other combination of the three, merge those two alternatives to suit the situation.

If your document contains nontrivial examples of program code, we recommend releasing these examples in parallel under your choice of free software license, such as the GNU General Public License, to permit their use in free software.