Avoiding duplicate symbol when compiling to multiple instruction sets - c++

I am in the process of using CPU dispatch based on processor features to switch implementation of a complicated numerical algorithm. I want to include the two versions (an sse2 and sse3 version for arguments sake) I am compiling in the same dynamic library.
The approach taken so far is to wrap all architecture specific code into a namespace e.g. namespace sse2 and namespace sse3 and thus avoiding duplicate symbol names when linking into the final dynamic library.
However, what happens if I use some code outside my control (e.g. a std::vector<int>) in both the sse2 and ss3 version. As far as I can see, the std::vector implementation will be present in both the sse2 and sse3 object files, but could in theory contain different instructions depending on the optimizations performed by the compiler. When I link these object files into the dynamic library, one of them will be used, and I risk potentially trying to run an sse3 instruction on a cpu only supporting sse2.
Aside from compiling to two separate dynamic libraries, what can be done to get around this problem? I need a solution working with both Visual Studio and clang on windows, mac os x and linux.

One approach would be to dispatch at the shared-library level instead of the object-file level. This would require compiling the entire library multiple times with different instruction set support, then dispatching to the appropriate shared library at runtime based on the CPU capabilities that you detect. I detail an approach that can work for this on OS X and Linux in this previous answer. I have not tried to implement this on Windows (yet), however.

This scenario is fully supported in the language and should not require any explicit handling. Your dynamic dispatch scenario doesn't have much to do with it - it is very typical for a project to instantiate std::vector in multiple translation units, and still ODR is not violated. In general, inline functions - and in particular template instantiations - are just not visible to the linker (i.e. do not appear as 'External' at the obj file tables).
If for some exotic reasons you would need to explicitly control this linkage type you'd have to resort to compiler-specific apparati. The MSVC apparatus is selectany, gcc has some other devices. Don't know about clang - but the point is you'd be hard pressed to come up with a reason to use any of them.

Related

What affects generated machine code at each step of the compilation process?

I am almost certain this question has been asked before, but I can not seem to find the right keywords to search for to get an answer. My apologies if this is a duplicate.
I am better trying to understand the compilation process of say a C++ file as it goes from the C++ syntax to the binary machine code. In addition I am trying to understand what influences the resulting machine code.
First, I am nearly certain that the following are the only factors (for most systems) that dictate the final machine code (please correct me if I am wrong here)
The tools used to compile, assemble, and link.
Things like gnu c compiler, clang, visual studio, nasm, ect.
The kernel of the system being used.
Whether its a specific version of the linux kernel, windows microkernel, or some other kernel like a mac os x one.
The operating system being used.
This one I am less clear about. I am unsure if machines running the same linux kernel, but different os, in this case let's say debian vs centos, will they produce different binaries.
Lastly the hardware architecture.
Different cpu architectures like arm 64, x86, power pc, ect. take different op codes so obviously the machine code should be different.
So with that being said here is my understanding of the compilation process and where each of these dependencies show up.
I write a C++ file and use code that my system can understand. A good example might be using <winsock.h> on windows and <sys/socket.h> on linux.
The preprocessor runs and executes any preprocessor macros.
Here I know that different preprocessors will define different macros but for now I will assume this is not too machine dependent. (This might be wrong to assume).
The compiler tools run to produce assembly file outputs.
Here the assembly produced depends on the compiler and what optimizations or choices it makes.
It also depends on the kernel because different kernels have different system calls and store files in different locations. This means the assembly might make changes such as different branching when calling functions specific to that kernel.
The operating system? Still unsure how the operating system fits in to this. If two machines have the same kernel, what does the operating system do to the binaries?
Finally the assembly code depends on the cpu architecture. I think that is a pretty obvious statement.
Once the compiler produces an assembly. We can then invoke the assembler to turn our assembly code into almost complete machine code. (I think machine code is identical to binary opcodes a cpu manual lists but this might be wrong).
The corresponding machine code files (often called object files I think) contain nearly all the instructions needed to run or reference other machine code files which will be linked in the next step.
This machine code usually has some format (I think ELF is a popular format for linux) and this format is dependent on the linker for sure.
I don't think the kernel, operating system, or hardware affect the layout/format of the object file but this is probably wrong. If they do please correct this.
The hardware will affect the actual machine code produced because again I think it is a 1 to 1 mapping of machine code instructions to opcodes for a cpu.
I am unsure if the kernel or operating system affect the linking process because I thought their changes were already incorporated in the compiling step.
Finally the linking step occurs.
I think this is as simple as the linker looking for all the referenced machine code and injecting it into one complete machine code file which can be executed.
I have no clue what affects this besides the linker tool itself.
So with all that, I need help identifying inaccuracies with the procedure I described above, and any dependencies I might have missed whether it be cpu, os, kernel, or tool ones.
Thank you and sorry for the long winded question. This probably should have been broken up into multiple questions but I am too far in. If this does not go well I may ask each part in individual questions.
EDIT:
Questions with more focus.
What components of a machine affect the machine code produced given a C++ file input?
Actually that is a lot of questions and usually you're question would be much too broad for SO (as you managed to recognize by yourself). But on the other hand you showed a deep interest (just by writing such a long and profound question) and also a lot of correct understanding of the process of compiling a program. The things you are missing or not understanding correctly (and you are probably the most interested in) are those things, that I myself found hard to learn. Thus I will provide you with some important points, that I think you are missing in the big picture.
Note that I am very much used to Linux, so I will mostly describe how things work on Linux. But I believe that most things also happen in a similar way on other operating systems.
Let's begin with the hardware. A modern computer has a CPU of some architecture. There are lots of different of CPU architectures. You mentioned some of them like arm, x86, etc. which are families of similar CPUs and can be divided into smaller groups by bit width and/or supported extensions. Ultimately your processor has a specified instruction set that defines which opcodes it supports and what those opcodes do. If a native (compiled) program runs, there are raw opcodes in the memory and the CPU directly executes them following its architecture specification.
Aside from the CPU there is a lot more hardware connected to your computer. Usually communicating with this hardware is complicated and not standardized. If a user program for example gets input keystrokes from the keyboard, in does not have to directly communicate with the keyboard, but rather does this via the operating system kernel. This works by a mechanism called syscall interrupt. The kernel installs an handler routine, that is called if a user program triggers such an interrupt with a special CPU instruction. You can think of it like a language agnostic function call from the program into the kernel. For example for Linux you can find a list of all syscalls at the syscall(2) man page. The syscalls form the kernel's Application Binary Interface (kernel ABI). Reading and writing from a terminal or using a filesystem are examples for syscall functionality.
As you can see, there are already very high level functions, that are implemented in the kernel. However the functionality is still quite limited for most typical applications. To encapsulate the syscalls and provide functions for memory management, utility functions, mathematical functions and many other things you probably use in your daily programs, there is usually another layer between the program and the kernel. This thing is called the C standard library, and it is a shared library (we will cover what exactly this is in a moment). On GNU/Linux it is the glibc which is the single most important library on a GNU/Linux system (and notably not part of the kernel 1). While it implements all the features that are required by the C standard (for example functions like malloc() or strcpy()), it also ships a lot of additional functions which are a superset of the ISO C standard library, the POSIX standard and some extensions. This interface is usually called the Application Programming Interface (API) of the operating system. While it is in principle possible to bypass the API and directly use the syscalls, almost all programs (even when written in other languages than C or C++) use the C library.
Now get yourself a coffee and a few minutes of rest. We now have enough background information to look at how a C++ program is transformed into a binary, and how exactly this binary is executed.
A C++ program consists of different compilation units (usually each different source file is a compilation unit). Each compilation unit undergoes the following steps
The preprocessor is run on the file. It includes header, expands macros and does some other stuff. As you wrote in your question this is rather platform independent. The preprocessor actions are standardized in the C++ standard.
The resulting code is compiled. That means C++ code is translated into assembly code. Because assembly code directly reflects the CPU instructions, this step is dependent on the target CPU architecture, that the compiler was configured for (usually the host CPU). The compiler is allowed to optimize and translate the program in any way it wants, as long as it follows the as-if rule. Thus this step is also higly dependent on the compiler you are using.
Note: Symbols (especially functions) that are not defined, are left undefined. If you say call the malloc() function, this will not be compiled, but left unevaluated until later. Thus this step is also not much dependent on the operating system.
Assembling takes place. This is very straightforward. The assembly code usually can be converted directly into binary CPU instructions. Local symbols (such as goto labels etc.) are resolved and replaced by their corresponding addresses. Unknown external symbols such as the mentioned malloc() call still are left unevaluated and are stored in the object file's symbol table. Because most of the syscalls are wrapped in library functions, the assembly code will usually not directly contain syscall code. Thus this step is depended on the CPU architecture. It is however dependent on the ABI2, which in term is dependent on the compiler and the OS.
Linking takes place. The different compilation units are combined into a single executable binary in an OS-dependent format (e.g. GNU/Linux uses ELF). Here yet more symbols are resolved. For example if one compilation calls a function in another compilation unit, this call is resolved and the symbol is replaced by the function address. If you link to a library statically, this is just treated like another compilation unit and included into the executable with its symbols resolved.
Shared libraries are checked for the needed symbols, but not linked yet. For example in case of the malloc() call, the linker checks, that there is a malloc symbol in the glibc, but the symbol in the executable still remains unresolved.
At this point you have a executable binary. As you might noticed, there might still be unresolved symbols in that binary. Thus you cannot just load that binary into RAM and let the CPU execute it. A final step called dynamic linking is needed. On Linux the program that performs this step is called the dynamic linker/loader. Its task is to load the executable ELF file into memory, look up all the needed dynamic libraries, load them into memory as well (a list is stored in the ELF file) and resolve the remaining symbols. This last step happens each time the program is executed. Now finally the malloc() symbol is resolved with the address in the glibc shared library.
You have pure CPU instructions in memory, the CPU's program counter register (the one that tracks the next instruction) is set to the entry point, and the program can begin to run. Every now and then it is interrupted either because it makes a syscall, or because it is interrupted by the kernel scheduler to let another program run on that CPU core.
I hope I could answer some of your questions and satisfy your curiosity. I think the most important part you were missing, was how dynamic linking happens. This is a very interesting topic which is related to concepts like position independent code. I wish you could luck learning.
1 this is also one reason why some people insist on calling Linux based systems GNU/Linux. The glibc library (together with many other GNU programs) defines much of the operating system structure, interacts with supplementary programs and configuration files etc. There are however Linux based systems without glibc. One of them is Android, using Googles bionic libc.
2 The ABI is related to the calling convention. This is a mixture of operating system, programming language and compiler specification. It is one of the reasons (besides name mangling, see the comment of PeterCordes below) you need those extern "C" {...} scopes in C++ header files, that declare C functions in shared libraries. It basically is a convention on how to pass parameters and return values between functions.
Neither operating system nor kernel are directly involved in any of this.
Their limited involvement is in that if you want to build Linux 64 bit binaries for x86 using gnu tools then you need to in some way (download and install or build yourself) build the gnu tools themselves for that target processor and that operating system. As system calls are specific to the operating system and target, and also the binaries supported by that operating system. Not strictly just the elf file format, that is just a container, but the linking and possibly bootstrap is also specific to the operating systems loader. (or if building something for the kernel that would have other rules). For example, does the application loader initialize .bss and .data for you from specific information in the .elf file, or like on an mcu does the bootstrap code itself have to do this?
The builder for gnu tools for a target like linux and ideally a pre-built binary for your os and target, would have paths setup in some way. The c library would have a default linker script and its intimate partner the bootstrap.
After that point, it is just a dumb toolchain. Include files be they at the system level, compiler level, or programmer level are just includes in the C language. The default paths and gcc knows where it was executed from so it knows where in a normal build the gcc and other libraries live.
gcc itself is not a compiler actually it calls other programs like the preprocessor, the compiler itself, the assembler and linker.
The preprocessor is going to do the search and replace for includes and defines and end up with one great big cpp file, then pass that to the compiler.
The compiler front end (C++ language for gcc for example) turns that into an internal language, allocate an int with this name, and another add the two and blah. A pseudo code if you will. This gets a lot of the optimization work done on it then eventually the back end (which for gnu could be x86, mips, arm, etc independent to some extent of the front and middle). The LLVM tools, are at least capable of exposing that middle, internal, language to external files (external to the memory used by the compiler to do the compilation) and you can combine and optimize those bytecode files and then convert them to assembly or direct to object in the llvm world. I think this is an exception not a rule, others just use internal tables.
While I think it is wise and sane to use an assembly language step. Not all compilers do and do not assume that all compilers do. Some output objects.
Yes that assembly is naturally partial, external functions (labels) and variables (labels) cannot be resolved at the object level. The linker has to do that.
So the target (x86, arm, etc) does affect the construction of the elf file as
there are certain items, magic numbers specific to the target. As mentioned the operating system and or kernel do affect the elf in that there are rules for construction of the binary for that kernel or operating system. Remember that elf is just a container like tar or zip or mkv etc. Do not assume that the operating system can handle every possible choice you want to make with the contents that the linker will allow (the tools are dumb, do what they are told).
So your source.
All the relevant sources that go with it including system includes, compiler includes and your includes.
gcc/g++ is a wrapper program that manages the steps.
calls the pre-processor expands includes and defines into one file (no magic here)
call the compiler to parse that one file into internal tables, think pseudo code and data
many, many possible optimizers that operate on these structures
backend, including peephole optimizer, turns the tables into assembly language (for gnu at least)
assembler is called to turn the asm into an object
If all the objects are specified and gcc is told to link, then...
Linker combines all the objects for the binary, including the bootstrap, including already built libraries, stubs, etc, and command line or more likely a linker script (linker script and bootstrap have an intimate relationship they are not assumed to be separable and not part of the compiler they are part of a C library, etc).
Kernel module loader or operating system application loader fed the file and per the rules of that loader loads and runs the program.

Inlining behavior when doing unity build (clang)

To optimize compile time I would like to enable unity build (via CMake) for my embedded system (C++) project. Works great, but as i see it there are some side effects.
One major thing that i observed is that the linked binary has a different size (bigger) than compared to the "normal" build. Looking at the elf file I noticed that there are fewer symbols in the unity built binary compared to the other one. As I see it there is some inlining happening at compile time (initially I thought inlining happens at link time?) and therefore binary size grows, when inlined functions are used multiple times.
Because of the inlining that happens with unity build, the runtime is also slightly shorter.
The concern i have right now is that with growing source code, i get different unity buckets and therefore inlining is not really deterministic.
If my assumption is right, is there a way to counteract to this problem?
First to this aspect
(initially I thought inlining happens at link time?)
Is compiler and context dependend but in general, compiler and linker are involved here both.
For the general question:
How do you define Unity buckets for you? I guess you mean effectively multiple translation units. If that's the case, nothing is really different to "common" building schemes (each .cpp leads to a single translation unit), so your concerns would have nothing to do with UnityBuild in detail. As a best practice: If you use UnityBuilds, make sure your project/solution is always fine for "common" builds and for UnityBuild likewise. That means for instance: Avoid global using namespaces always (in .cpp and .h), make sure you do not run into static initialization order fiasco (very relevant to your question here), avoid a lot of anonymous namespace usage cases (sad I know...), use a strict include guard scheme and so on...

Choose assembly implementation to use based on supported instructions

I am working on a C library which compiles/links to a .a file that users can statically link into their code. The library's performance is very important, so I am writing performance-critical routines in x86-64 assembly to optimize performance.
For some routines, I can get significantly better performance if I use BMI2 instructions than if I stick to the "standard" x86-64 instruction set. Trouble is, BMI2 was introduced fairly recently and some of my users use processors that do not support those instructions.
So, I've written optimized the routines twice, once using BMI2 instructions and once without using them. In my current setup, I would distribute two versions of the .a file: a "fast" one that requires support for BMI2 instructions, and a "slow" one that does not require support for BMI2 instructions.
I am asking if there's a way to simplify this by distributing a single .a file that will dynamically choose the correct implementation based on whether the CPU on which the final application runs supports BMI2 instructions.
Unlike similar questions on StackOverflow, there are two peculiarities here:
The technique to choose the function needs to have particularly low overhead in the critical path. The routines in question, after assembly-optimization, run in ~10 ns, so even a single if statement could be significant.
The function that needs to be chosen "dynamically" is chosen once at the beginning, and then remains fixed for the duration of the program. I'm hoping that this will offer a faster solution than the one suggested in this question: Choosing method implementation at runtime
The fastest solution I've come up with so far is to do the following:
Check whether the CPU supports BMI2 instructions using the cpuid instruction.
Set a global variable true or false depending on the result.
Branch on the value of this global variable on every function invocation.
I'm not satisfied with this approach because it has two drawbacks:
I'm not sure how I can automatically run cpuid and set a global variable at the beginning of the program, given that I'm distributing a .a file and don't have control over the main function in the final binary. I'm happy to use C++ here if it offers a better solution, as long as the final library can still be linked with and called from a C program.
This incurs overhead on every function call, when ideally the only overhead would be on program startup.
Are there any solutions that are more efficient than the one I've detailed above?
x264 uses an init function (which users of the library are required to call before calling anything else, or something like that) to set up a struct of function pointers based on CPUID results. Including taking into account that pshufb is slow on some early CPUs that support it.
If your functions depend on pdep / pext, you probably want to detect AMD vs. Intel, because AMD's pdep/pext is very slow and probably not worth using on Ryzen, even though it is available. (See https://agner.org/optimize/ for instruction tables.)
Function pointers are fairly low overhead, about the same as calling a function in a shared library or DLL. call [rel funcptr] instead of call func. (In the compiler-generated asm that calls your functions).
CPU dependent code: how to avoid function pointers? shows a very simple example of it in C, and is asking for ways to avoid it. With dynamic linking, you can do CPU detection at dynamic link time so the dynamic-linking indirection becomes your CPU-dispatch indirection as well (like glibc does for selecting an optimized memcpy implementation.)
But with static linking for a .a, just make function pointers that are statically initialized to the baseline versions, and your CPU init function (which hopefully runs before any of the function pointers are dereferenced) rewrites them to point at the best version for the current CPU.
If you are using gcc, you can get the compiler to implement all the boiler plate code automatically. gcc manual page on function multiversioning

standard way to put static objects together

I want to put some objects with static linkage together, without any other data interleaved. The objects are defined in separate source files. The order doesn't matter, as long as they're contiguous (excepting minimal/standard padding).
In gcc I can achieve this with __attribute__((section("mydata"))), but AFAIK that's GCC specific and only many compilers support it (gcc, llvm and a few others, but I think not msvc), for many targets (elf yes, others I'm not sure, I doubt it).
Q1: is there a standard way to achieve this?
Q2: does GCC guarantee that the objects will be contiguous with padding? i.e. the objects can be enumerated by a linear scan with a constant stride.
Q3: is there a difference in behavior between GCC and other compilers that support this attribute?
Q4: which targets and major compilers wouldn't support this?
It sounds like you're trying to produce platform independent code (hence the need for a standard mechanism) and while you have access to the source, you only want to make limited changes to it. The solution you are proposing involves accessing datastructures defined in multiple translation units as though they were elments of a single array.
Q1. Neither the C standard nor the C++ standard contain the concept of linker sections, so, no, there is no truly platform independent way to specify output section names.
You would need to check for each compiler that you want to target. You have already found the incantation for GCC and Clang. Visual C++ has #pragma section and __declspec(allocate) which together can achieve the same thing.
Q2. GCC uses the platform's linker to build an executable. Assuming a GNU environment, then that linker will be the binutils ld or gold. With these linkers, each output section is contiguous, and alignment of the input sections within the output section can be specified or allowed to default. According to the documentation the default will work for you assuming all of your object files were compiled with the same version of GCC, using the same compiler options. Different compiler options may result in different default alignments.
Q3. Regarding differences in behavior on platforms: your use of the linker is fairly unusual. You ought to carefully test each platform you target. I'd suggest building a small binary that puts object files together in the way you want and (programatically) checks that the datastructures are accessible in the way you desire.
Q4. It looks like your approach will work for GCC/Clang/binutils on Linux and for Visual C++/LINK on Windows. It would take quite some effort to learn each platform's idosyncracies, but I think you could make your approach work on just about any post-1995 platform with a real keyboard. You may have difficulty on older machines and smaller embedded systems. You can probably get it going on Android. I'm not sure about iOS.
By way of general advice, accessing datastructures by assuming they are contiguous across translation unit boundaries is unusual. You are likely to run into strange problems and find it difficult to get help. Here are some alternate mechansims that might get you what you want, with less per-platform effort:
Put the source files together before compilation and compile them as one unit. If it's not possible to do by hand, consider using the pre-processor or building a small Perl or Python script to do it for you at build time.
At run-time, dynamically allocate a contiguous section of memory and copy the datastructures to it. If the datastructures are small, this won't use much extra memory. If the datastructures are large, the memory containing the original structures will be paged out and this won't use much extra memory.
Use indirection: have your program access the datastructures through an array of pointers to the data.

What techniques can be used to speed up C++ compilation times?

What techniques can be used to speed up C++ compilation times?
This question came up in some comments to Stack Overflow question C++ programming style, and I'm interested to hear what ideas there are.
I've seen a related question, Why does C++ compilation take so long?, but that doesn't provide many solutions.
Language techniques
Pimpl Idiom
Take a look at the Pimpl idiom here, and here, also known as an opaque pointer or handle classes. Not only does it speed up compilation, it also increases exception safety when combined with a non-throwing swap function. The Pimpl idiom lets you reduce the dependencies between headers and reduces the amount of recompilation that needs to be done.
Forward Declarations
Wherever possible, use forward declarations. If the compiler only needs to know that SomeIdentifier is a struct or a pointer or whatever, don't include the entire definition, forcing the compiler to do more work than it needs to. This can have a cascading effect, making this way slower than they need to be.
The I/O streams are particularly known for slowing down builds. If you need them in a header file, try #including <iosfwd> instead of <iostream> and #include the <iostream> header in the implementation file only. The <iosfwd> header holds forward declarations only. Unfortunately the other standard headers don't have a respective declarations header.
Prefer pass-by-reference to pass-by-value in function signatures. This will eliminate the need to #include the respective type definitions in the header file and you will only need to forward-declare the type. Of course, prefer const references to non-const references to avoid obscure bugs, but this is an issue for another question.
Guard Conditions
Use guard conditions to keep header files from being included more than once in a single translation unit.
#pragma once
#ifndef filename_h
#define filename_h
// Header declarations / definitions
#endif
By using both the pragma and the ifndef, you get the portability of the plain macro solution, as well as the compilation speed optimization that some compilers can do in the presence of the pragma once directive.
Reduce interdependency
The more modular and less interdependent your code design is in general, the less often you will have to recompile everything. You can also end up reducing the amount of work the compiler has to do on any individual block at the same time, by virtue of the fact that it has less to keep track of.
Compiler options
Precompiled Headers
These are used to compile a common section of included headers once for many translation units. The compiler compiles it once, and saves its internal state. That state can then be loaded quickly to get a head start in compiling another file with that same set of headers.
Be careful that you only include rarely changed stuff in the precompiled headers, or you could end up doing full rebuilds more often than necessary. This is a good place for STL headers and other library include files.
ccache is another utility that takes advantage of caching techniques to speed things up.
Use Parallelism
Many compilers / IDEs support using multiple cores/CPUs to do compilation simultaneously. In GNU Make (usually used with GCC), use the -j [N] option. In Visual Studio, there's an option under preferences to allow it to build multiple projects in parallel. You can also use the /MP option for file-level paralellism, instead of just project-level paralellism.
Other parallel utilities:
Incredibuild
Unity Build
distcc
Use a Lower Optimization Level
The more the compiler tries to optimize, the harder it has to work.
Shared Libraries
Moving your less frequently modified code into libraries can reduce compile time. By using shared libraries (.so or .dll), you can reduce linking time as well.
Get a Faster Computer
More RAM, faster hard drives (including SSDs), and more CPUs/cores will all make a difference in compilation speed.
I work on the STAPL project which is a heavily-templated C++ library. Once in a while, we have to revisit all the techniques to reduce compilation time. In here, I have summarized the techniques we use. Some of these techniques are already listed above:
Finding the most time-consuming sections
Although there is no proven correlation between the symbol lengths and compilation time, we have observed that smaller average symbol sizes can improve compilation time on all compilers. So your first goals it to find the largest symbols in your code.
Method 1 - Sort symbols based on size
You can use the nm command to list the symbols based on their sizes:
nm --print-size --size-sort --radix=d YOUR_BINARY
In this command the --radix=d lets you see the sizes in decimal numbers (default is hex). Now by looking at the largest symbol, identify if you can break the corresponding class and try to redesign it by factoring the non-templated parts in a base class, or by splitting the class into multiple classes.
Method 2 - Sort symbols based on length
You can run the regular nm command and pipe it to your favorite script (AWK, Python, etc.) to sort the symbols based on their length. Based on our experience, this method identifies the largest trouble making candidates better than method 1.
Method 3 - Use Templight
"Templight is a Clang-based tool to profile the time and memory consumption of template instantiations and to perform interactive debugging sessions to gain introspection into the template instantiation process".
You can install Templight by checking out LLVM and Clang (instructions) and applying the Templight patch on it. The default setting for LLVM and Clang is on debug and assertions, and these can impact your compilation time significantly. It does seem like Templight needs both, so you have to use the default settings. The process of installing LLVM and Clang should take about an hour or so.
After applying the patch you can use templight++ located in the build folder you specified upon installation to compile your code.
Make sure that templight++ is in your PATH. Now to compile add the following switches to your CXXFLAGS in your Makefile or to your command line options:
CXXFLAGS+=-Xtemplight -profiler -Xtemplight -memory -Xtemplight -ignore-system
Or
templight++ -Xtemplight -profiler -Xtemplight -memory -Xtemplight -ignore-system
After compilation is done, you will have a .trace.memory.pbf and .trace.pbf generated in the same folder. To visualize these traces, you can use the Templight Tools that can convert these to other formats. Follow these instructions to install templight-convert. We usually use the callgrind output. You can also use the GraphViz output if your project is small:
$ templight-convert --format callgrind YOUR_BINARY --output YOUR_BINARY.trace
$ templight-convert --format graphviz YOUR_BINARY --output YOUR_BINARY.dot
The callgrind file generated can be opened using kcachegrind in which you can trace the most time/memory consuming instantiation.
Reducing the number of template instantiations
Although there are no exact solution for reducing the number of template instantiations, there are a few guidelines that can help:
Refactor classes with more than one template arguments
For example, if you have a class,
template <typename T, typename U>
struct foo { };
and both of T and U can have 10 different options, you have increased the possible template instantiations of this class to 100. One way to resolve this is to abstract the common part of the code to a different class. The other method is to use inheritance inversion (reversing the class hierarchy), but make sure that your design goals are not compromised before using this technique.
Refactor non-templated code to individual translation units
Using this technique, you can compile the common section once and link it with your other TUs (translation units) later on.
Use extern template instantiations (since C++11)
If you know all the possible instantiations of a class you can use this technique to compile all cases in a different translation unit.
For example, in:
enum class PossibleChoices = {Option1, Option2, Option3}
template <PossibleChoices pc>
struct foo { };
We know that this class can have three possible instantiations:
template class foo<PossibleChoices::Option1>;
template class foo<PossibleChoices::Option2>;
template class foo<PossibleChoices::Option3>;
Put the above in a translation unit and use the extern keyword in your header file, below the class definition:
extern template class foo<PossibleChoices::Option1>;
extern template class foo<PossibleChoices::Option2>;
extern template class foo<PossibleChoices::Option3>;
This technique can save you time if you are compiling different tests with a common set of instantiations.
NOTE : MPICH2 ignores the explicit instantiation at this point and always compiles the instantiated classes in all compilation units.
Use unity builds
The whole idea behind unity builds is to include all the .cc files that you use in one file and compile that file only once. Using this method, you can avoid reinstantiating common sections of different files and if your project includes a lot of common files, you probably would save on disk accesses as well.
As an example, let's assume you have three files foo1.cc, foo2.cc, foo3.cc and they all include tuple from STL. You can create a foo-all.cc that looks like:
#include "foo1.cc"
#include "foo2.cc"
#include "foo3.cc"
You compile this file only once and potentially reduce the common instantiations among the three files. It is hard to generally predict if the improvement can be significant or not. But one evident fact is that you would lose parallelism in your builds (you can no longer compile the three files at the same time).
Further, if any of these files happen to take a lot of memory, you might actually run out of memory before the compilation is over. On some compilers, such as GCC, this might ICE (Internal Compiler Error) your compiler for lack of memory. So don't use this technique unless you know all the pros and cons.
Precompiled headers
Precompiled headers (PCHs) can save you a lot of time in compilation by compiling your header files to an intermediate representation recognizable by a compiler. To generate precompiled header files, you only need to compile your header file with your regular compilation command. For example, on GCC:
$ g++ YOUR_HEADER.hpp
This will generate a YOUR_HEADER.hpp.gch file (.gch is the extension for PCH files in GCC) in the same folder. This means that if you include YOUR_HEADER.hpp in some other file, the compiler will use your YOUR_HEADER.hpp.gch instead of YOUR_HEADER.hpp in the same folder before.
There are two issues with this technique:
You have to make sure that the header files being precompiled is stable and is not going to change (you can always change your makefile)
You can only include one PCH per compilation unit (on most of compilers). This means that if you have more than one header file to be precompiled, you have to include them in one file (e.g., all-my-headers.hpp). But that means that you have to include the new file in all places. Fortunately, GCC has a solution for this problem. Use -include and give it the new header file. You can comma separate different files using this technique.
For example:
g++ foo.cc -include all-my-headers.hpp
Use unnamed or anonymous namespaces
Unnamed namespaces (a.k.a. anonymous namespaces) can reduce the generated binary sizes significantly. Unnamed namespaces use internal linkage, meaning that the symbols generated in those namespaces will not be visible to other TU (translation or compilation units). Compilers usually generate unique names for unnamed namespaces. This means that if you have a file foo.hpp:
namespace {
template <typename T>
struct foo { };
} // Anonymous namespace
using A = foo<int>;
And you happen to include this file in two TUs (two .cc files and compile them separately). The two foo template instances will not be the same. This violates the One Definition Rule (ODR). For the same reason, using unnamed namespaces is discouraged in the header files. Feel free to use them in your .cc files to avoid symbols showing up in your binary files. In some cases, changing all the internal details for a .cc file showed a 10% reduction in the generated binary sizes.
Changing visibility options
In newer compilers you can select your symbols to be either visible or invisible in the Dynamic Shared Objects (DSOs). Ideally, changing the visibility can improve compiler performance, link time optimizations (LTOs), and generated binary sizes. If you look at the STL header files in GCC you can see that it is widely used. To enable visibility choices, you need to change your code per function, per class, per variable and more importantly per compiler.
With the help of visibility you can hide the symbols that you consider them private from the generated shared objects. On GCC you can control the visibility of symbols by passing default or hidden to the -visibility option of your compiler. This is in some sense similar to the unnamed namespace but in a more elaborate and intrusive way.
If you would like to specify the visibilities per case, you have to add the following attributes to your functions, variables, and classes:
__attribute__((visibility("default"))) void foo1() { }
__attribute__((visibility("hidden"))) void foo2() { }
__attribute__((visibility("hidden"))) class foo3 { };
void foo4() { }
The default visibility in GCC is default (public), meaning that if you compile the above as a shared library (-shared) method, foo2 and class foo3 will not be visible in other TUs (foo1 and foo4 will be visible). If you compile with -visibility=hidden then only foo1 will be visible. Even foo4 would be hidden.
You can read more about visibility on GCC wiki.
I'd recommend these articles from "Games from Within, Indie Game Design And Programming":
Physical Structure and C++ – Part 1: A First Look
Physical Structure and C++ – Part 2: Build Times
Even More Experiments with Includes
How Incredible Is Incredibuild?
The Care and Feeding of Pre-Compiled Headers
The Quest for the Perfect Build System
The Quest for the Perfect Build System (Part 2)
Granted, they are pretty old - you'll have to re-test everything with the latest versions (or versions available to you), to get realistic results. Either way, it is a good source for ideas.
One technique which worked quite well for me in the past: don't compile multiple C++ source files independently, but rather generate one C++ file which includes all the other files, like this:
// myproject_all.cpp
// Automatically generated file - don't edit this by hand!
#include "main.cpp"
#include "mainwindow.cpp"
#include "filterdialog.cpp"
#include "database.cpp"
Of course this means you have to recompile all of the included source code in case any of the sources changes, so the dependency tree gets worse. However, compiling multiple source files as one translation unit is faster (at least in my experiments with MSVC and GCC) and generates smaller binaries. I also suspect that the compiler is given more potential for optimizations (since it can see more code at once).
This technique breaks in various cases; for instance, the compiler will bail out in case two or more source files declare a global function with the same name. I couldn't find this technique described in any of the other answers though, that's why I'm mentioning it here.
For what it's worth, the KDE Project used this exact same technique since 1999 to build optimized binaries (possibly for a release). The switch to the build configure script was called --enable-final. Out of archaeological interest I dug up the posting which announced this feature: http://lists.kde.org/?l=kde-devel&m=92722836009368&w=2
I will just link to my other answer: How do YOU reduce compile time, and linking time for Visual C++ projects (native C++)?. Another point I want to add, but which causes often problems is to use precompiled headers. But please, only use them for parts which hardly ever change (like GUI toolkit headers). Otherwise, they will cost you more time than they save you in the end.
Another option is, when you work with GNU make, to turn on -j<N> option:
-j [N], --jobs[=N] Allow N jobs at once; infinite jobs with no arg.
I usually have it at 3 since I've got a dual core here. It will then run compilers in parallel for different translation units, provided there are no dependencies between them. Linking cannot be done in parallel, since there is only one linker process linking together all object files.
But the linker itself can be threaded, and this is what the GNU gold ELF linker does. It's optimized threaded C++ code which is said to link ELF object files a magnitude faster than the old ld (and was actually included into binutils).
There's an entire book on this topic, which is titled Large-Scale C++ Software Design (written by John Lakos).
The book pre-dates templates, so to the contents of that book add "using templates, too, can make the compiler slower".
Once you have applied all the code tricks above (forward declarations, reducing header inclusion to the minimum in public headers, pushing most details inside the implementation file with Pimpl...) and nothing else can be gained language-wise, consider your build system. If you use Linux, consider using distcc (distributed compiler) and ccache (cache compiler).
The first one, distcc, executes the preprocessor step locally and then sends the output to the first available compiler in the network. It requires the same compiler and library versions in all the configured nodes in the network.
The latter, ccache, is a compiler cache. It again executes the preprocessor and then check with an internal database (held in a local directory) if that preprocessor file has already been compiled with the same compiler parameters. If it does, it just pops up the binary and output from the first run of the compiler.
Both can be used at the same time, so that if ccache does not have a local copy it can send it trough the net to another node with distcc, or else it can just inject the solution without further processing.
Here are some:
Use all processor cores by starting a multiple-compile job (make -j2 is a good example).
Turn off or lower optimizations (for example, GCC is much faster with -O1 than -O2 or -O3).
Use precompiled headers.
When I came out of college, the first real production-worthy C++ code I saw had these arcane #ifndef ... #endif directives in between them where the headers were defined. I asked the guy who was writing the code about these overarching things in a very naive fashion and was introduced to world of large-scale programming.
Coming back to the point, using directives to prevent duplicate header definitions was the first thing I learned when it came to reducing compiling times.
More RAM.
Someone talked about RAM drives in another answer. I did this with a 80286 and Turbo C++ (shows age) and the results were phenomenal. As was the loss of data when the machine crashed.
You could use Unity Builds.
​​
Use
#pragma once
at the top of header files, so if they're included more than once in a translation unit, the text of the header will only get included and parsed once.
Use forward declarations where you can. If a class declaration only uses a pointer or reference to a type, you can just forward declare it and include the header for the type in the implementation file.
For example:
// T.h
class Class2; // Forward declaration
class T {
public:
void doSomething(Class2 &c2);
private:
Class2 *m_Class2Ptr;
};
// T.cpp
#include "Class2.h"
void Class2::doSomething(Class2 &c2) {
// Whatever you want here
}
Fewer includes means far less work for the preprocessor if you do it enough.
Just for completeness: a build might be slow because the build system is being stupid as well as because the compiler is taking a long time to do its work.
Read Recursive Make Considered Harmful (PDF) for a discussion of this topic in Unix environments.
Not about the compilation time, but about the build time:
Use ccache if you have to rebuild the same files when you are working
on your buildfiles
Use ninja-build instead of make. I am currently compiling a project
with ~100 source files and everything is cached by ccache. make needs
5 minutes, ninja less than 1.
You can generate your ninja files from cmake with -GNinja.
Upgrade your computer
Get a quad core (or a dual-quad system)
Get LOTS of RAM.
Use a RAM drive to drastically reduce file I/O delays. (There are companies that make IDE and SATA RAM drives that act like hard drives).
Then you have all your other typical suggestions
Use precompiled headers if available.
Reduce the amount of coupling between parts of your project. Changing one header file usually shouldn't require recompiling your entire project.
I had an idea about using a RAM drive. It turned out that for my projects it doesn't make that much of a difference after all. But then they are pretty small still. Try it! I'd be interested in hearing how much it helped.
Dynamic linking (.so) can be much much faster than static linking (.a). Especially when you have a slow network drive. This is since you have all of the code in the .a file which needs to be processed and written out. In addition, a much larger executable file needs to be written out to the disk.
Where are you spending your time? Are you CPU bound? Memory bound? Disk bound? Can you use more cores? More RAM? Do you need RAID? Do you simply want to improve the efficiency of your current system?
Under gcc/g++, have you looked at ccache? It can be helpful if you are doing make clean; make a lot.
Starting with Visual Studio 2017 you have the capability to have some compiler metrics about what takes time.
Add those parameters to C/C++ -> Command line (Additional Options) in the project properties window:
/Bt+ /d2cgsummary /d1reportTime
You can have more informations in this post.
Faster hard disks.
Compilers write many (and possibly huge) files to disk. Work with SSD instead of typical hard disk and compilation times are much lower.
On Linux (and maybe some other *NIXes), you can really speed the compilation by NOT STARING at the output and changing to another TTY.
Here is the experiment: printf slows down my program
Networks shares will drastically slow down your build, as the seek latency is high. For something like Boost, it made a huge difference for me, even though our network share drive is pretty fast. Time to compile a toy Boost program went from about 1 minute to 1 second when I switched from a network share to a local SSD.
If you have a multicore processor, both Visual Studio (2005 and later) as well as GCC support multi-processor compiles. It is something to enable if you have the hardware, for sure.
First of all, we have to understand what so different about C++ that sets it apart from other languages.
Some people say it's that C++ has many too features. But hey, there are languages that have a lot more features and they are nowhere near that slow.
Some people say it's the size of a file that matters. Nope, source lines of code don't correlate with compile times.
But wait, how can it be? More lines of code should mean longer compile times, what's the sorcery?
The trick is that a lot of lines of code is hidden in preprocessor directives. Yes. Just one #include can ruin your module's compilation performance.
You see, C++ doesn't have a module system. All *.cpp files are compiled from scratch. So having 1000 *.cpp files means compiling your project a thousand times. You have more than that? Too bad.
That's why C++ developers hesitate to split classes into multiple files. All those headers are tedious to maintain.
So what can we do other than using precompiled headers, merging all the cpp files into one, and keeping the number of headers minimal?
C++20 brings us preliminary support of modules! Eventually, you'll be able to forget about #include and the horrible compile performance that header files bring with them. Touched one file? Recompile only that file! Need to compile a fresh checkout? Compile in seconds rather than minutes and hours.
The C++ community should move to C++20 as soon as possible. C++ compiler developers should put more focus on this, C++ developers should start testing preliminary support in various compilers and use those compilers that support modules. This is the most important moment in C++ history!
Although not a "technique", I couldn't figure out how Win32 projects with many source files compiled faster than my "Hello World" empty project. Thus, I hope this helps someone like it did me.
In Visual Studio, one option to increase compile times is Incremental Linking (/INCREMENTAL). It's incompatible with Link-time Code Generation (/LTCG) so remember to disable incremental linking when doing release builds.
Using dynamic linking instead of static one make you compiler faster that can feel.
If you use t Cmake, active the property:
set(BUILD_SHARED_LIBS ON)
Build Release, using static linking can get more optimize.
From Microsoft: https://devblogs.microsoft.com/cppblog/recommendations-to-speed-c-builds-in-visual-studio/
Specific recommendations include:
DO USE PCH for projects
DO include commonly used system, runtime and third party headers in
PCH
DO include rarely changing project specific headers in PCH
DO NOT include headers that change frequently
DO audit PCH regularly to keep it up to date with product churn
DO USE /MP
DO Remove /Gm in favor of /MP
DO resolve conflict with #import and use /MP
DO USE linker switch /incremental
DO USE linker switch /debug:fastlink
DO consider using a third party build accelerator