Build tools

Compiling your Fortran projects by hand can become quite complicated depending on the number of source files and the interdependencies through the module. Supporting different compilers and linkers or different platforms can become increasingly complicated unless the right tools are used to automatically perform those tasks.

Depending on the size of your project and the purpose of project different options for the build automation can be used.

First, your integrated development environment probably provides a way to build your program. A popular cross-platform tool is Microsoft's Visual Studio Code, but others exist, such as Atom, Eclipse Photran, and Code::Blocks. They offer a graphical user-interface, but are often very specific for the compiler and platform.

For smaller projects, the rule based build system make is a common choice. Based on the rules defined it can perform task like (re)compiling object files from updated source files, creating libraries and linking executables. To use make for your project you have to encode those rules in Makefile, which defines the interdependencies of all the final program, the intermediary object files or libraries and the actual source files. For a short introduction see the guide on make.

Maintenance tools like autotools and CMake can generate Makefiles or Visual Studio project files via a high-level description. They abstract away from the compiler and platform specifics.

Which of those tools are the best choice for your projects depends on many factors. Choose a build tool you are comfortable working with, it should not get in your way while developing. Spending more time on working against your build tools than doing actual development work can quickly become frustrating.

Also, consider the accessibility of your build tools. If it is restricted to a specific integrated development environment, can all developers on your project access it? If you are using a specific build system, does it work on all platforms you are developing for? How large is the entry barrier of your build tools? Consider the learning curve for the build tools, the perfect build tool will be of no use, if you have to learn a complex programming language first to add a new source file. Finally, consider what other project are using, those you are depending on and those that use (or will use) your project as dependency.

Using make as build tool

The most well-known and commonly used build system is called make. It performs actions following rules defined in a configuration file called Makefile or makefile, which usually leads to compiling a program from the provided source code.

Tip

For an in-depth make tutorial lookup its info page. There is an online version of this info page, available.

We will start with the basics from your clean source directory. Create and open the file Makefile, we start with a simple rule called all:

all:
	echo "$@"

After saving the Makefile run it by executing make in the same directory. You should see the following output:

echo "all"
all

First, we note that make is substituting $@ for the name of the rule, the second thing to note is that make is always printing the command it is running, finally, we see the result of running echo "all".

注釈

We call the entry point of our Makefile always all by convention, but you can choose whatever name you like.

注釈

You should not have noticed it if your editor is working correctly, but you have to indent the content of a rule with a tab character. In case you have problems running the above Makefile and see an error like

Makefile:2: *** missing separator.  Stop.

The indentation is probably not correct. In this case replace the indentation in the second line with a tab character.

Now we want to make our rules more complicated, therefore we add another rule:

PROG := my_prog

all: $(PROG)
	echo "$@ depends on $^"

$(PROG):
	echo "$@"

Note how we declare variables in make, you should always declare your local variables with :=. To access the content of a variable we use the $(...), note that we have to enclose the variable name in parenthesis.

注釈

The declaration of variables is usually done with :=, but make does support recursively expanded variables as well with =. Normally, the first kind of declaration is wanted, as they are more predictable and do not have a runtime overhead from the recursive expansion.

We introduced a dependency of the rule all, namely the content of the variable PROG, also we modified the printout, we want to see all the dependencies of this rule, which are stored in the variable $^. Now for the new rule which we name after the value of the variable PROG, it does the same thing we did before for the rule all, note how the value of $@ is dependent on the rule it is used in.

Again check by running the make, you should see:

echo "my_prog"
my_prog
echo "all depends on my_prog"
all depends on my_prog

The dependency has been correctly resolved and evaluated before performing any action on the rule all. Let's run only the second rule: type make my_prog and you will only find the first two lines in your terminal.

The next step is to perform some real actions with make, we take the source code from the previous chapter here and add new rules to our Makefile:

OBJS := tabulate.o functions.o
PROG := my_prog

all: $(PROG)

$(PROG): $(OBJS)
	gfortran -o $@ $^

$(OBJS): %.o: %.f90
	gfortran -c -o $@ $<

We define OBJS which stands for object files, our program depends on those OBJS and for each object file we create a rule to make them from a source file. The last rule we introduced is a pattern matching rule, % is the common pattern between tabulate.o and tabulate.f90, which connects our object file tabulate.o to the source file tabulate.f90. With this set, we run our compiler, here gfortran and translate the source file into an object file, we do not create an executable yet due to the -c flag. Note the usage of the $< for the first element of the dependencies here.

After compiling all the object files we attempt to link the program, we do not use a linker directly, but gfortran to produce the executable.

Now we run the build script with make:

gfortran -c -o tabulate.o tabulate.f90
tabulate.f90:2:7:

    2 |    use user_functions
      |       1
Fatal Error: Cannot open module file ‘user_functions.mod’ for reading at (1): No such file or directory
compilation terminated.
make: *** [Makefile:10: tabulate.f90.o] Error 1

We remember that we have dependencies between our source files, therefore we add this dependency explicitly to the Makefile with

tabulate.o: functions.o

Now we can retry and find that the build is working correctly. The output should look like

gfortran -c -o functions.o functions.f90
gfortran -c -o tabulate.o tabulate.f90
gfortran -o my_prog tabulate.o functions.o

You should find four new files in the directory now. Run my_prog to make sure everything works as expected. Let's run make again:

make: Nothing to be done for 'all'.

Using the timestamps of the executable make was able to determine, it is newer than both tabulate.o and functions.o, which in turn are newer than tabulate.f90 and functions.f90. Therefore, the program is already up-to-date with the latest code and no action has to be performed.

In the end, we will have a look at a complete Makefile.

# Disable all of make's built-in rules (similar to Fortran's implicit none)
MAKEFLAGS += --no-builtin-rules --no-builtin-variables
# configuration
FC := gfortran
LD := $(FC)
RM := rm -f
# list of all source files
SRCS := tabulate.f90 functions.f90
PROG := my_prog

OBJS := $(addsuffix .o, $(SRCS))

.PHONY: all clean
all: $(PROG)

$(PROG): $(OBJS)
	$(LD) -o $@ $^

$(OBJS): %.o: %
	$(FC) -c -o $@ $<

# define dependencies between object files
tabulate.f90.o: functions.f90.o user_functions.mod

# rebuild all object files in case this Makefile changes
$(OBJS): $(MAKEFILE_LIST)

clean:
	$(RM) $(filter %.o, $(OBJS)) $(wildcard *.mod) $(PROG)

Since you are starting with make we highly recommend to always include the first line, like with Fortran's implicit none we do not want to have implicit rules messing up our Makefile in surprising and harmful ways.

Next, we have a configuration section where we define variables, in case you want to switch out your compiler, it can be easily done here. We also introduced the SRCS variable to hold all source files, which is more intuitive than specifying object files. We can easily create the object files by appending a .o suffix using the functions addsuffix. The .PHONY is a special rule, which should be used for all entry points of your Makefile, here we define two entry point, we already know all, the new clean rule deletes all the build artifacts again such that we indeed start with a clean directory.

Also, we slightly changed the build rule for the object files to account for appending the .o suffix instead of substituting it. Notice that we still need to explicitly define the interdependencies in the Makefile. We also added a dependency for the object files on the Makefile itself, in case you change the compiler, this will allow you to safely rebuild.

Now you know enough about make to use it for building small projects. If you plan to use make more extensively, we have compiled a few tips for you as well.

Tip

In this guide, we avoided and disabled a lot of the commonly used make features that can be particularly troublesome if not used correctly, we highly recommend staying away from the builtin rules and variables if you do not feel confident working with make, but explicitly declare all variables and rules.

You will find that make is capable tool to automate short interdependent workflows and to build small projects. But for larger projects, you will probably soon run against some of it limitations. Usually, make is therefore not used alone but combined with other tools to generate the Makefile completely or in parts.

Recursively expanded variables

Commonly seen in many projects are recursively expanded variables (declared with = instead of :=). Recursive expansion of your variables allows out-of-order declaration and other neat tricks with make, since they are defined as rules, which are expanded at runtime, rather than being defined while parsing.

For example, declaring and using your Fortran flags with this snippet will work completely fine:

all:
	echo $(FFLAGS)

FFLAGS = $(include_dirs) -O
include_dirs += -I./include
include_dirs += -I/opt/some_dep/include

You should find the expected (or maybe unexpected) printout after running make

echo -I./include -I/opt/some_dep/include -O
-I./include -I/opt/some_dep/include -O

Tip

appending with += to an undefined variable will produce a recursively expanded variable with this state being inherited for all further appending.

While, it seems like an interesting feature to use, it tends to lead to surprising and unexpected outcomes. Usually, when defining variables like your compiler, there is little reason to actually use the recursive expansion at all.

The same can easily be achieved using the := declaration:

all:
	echo $(FFLAGS)

include_dirs := -I./include
include_dirs += -I/opt/some_dep/include
FFLAGS := $(include_dirs) -O

重要

always think of a Makefile as a whole set of rules, it must be parsed completely before any rule can be evaluated.

You can use whatever kind of variables you like most, mixing them should be done carefully, of course. It is important to be aware of the differences between the two kinds and the respective implications.

Comments and whitespace

There are some caveats with whitespace and comments, which might pop up from time to time when using make. First, make does not know of any data type except for strings and the default separator is just a space. This means make will give a hard time trying to build a project which has spaces in file names. If you encounter such case, renaming the file is possibly the easiest solution at hand.

Another common problem is leading and trailing whitespace, once introduced, make will happily carry it along and it does in fact make a difference when comparing strings in make.

Those can be introduced by comments like

prefix := /usr  # path to install location
install:
	echo "$(prefix)/lib"

While the comment will be correctly removed by make, the trailing two spaces are now part of the variable content. Run make and check that this is indeed the case:

echo "/usr  /lib"
/usr  /lib

To solve this issue, you can either move the comment, or strip the whitespace with the strip function instead. Alternatively, you could try to join the strings.

prefix := /usr  # path to install location
install:
	echo "$(strip $(prefix))/lib"
	echo "$(join $(join $(prefix), /), lib)"

All in all, none of this solutions will make your Makefile more readable, therefore, it is prudent to pay extra attention to whitespace and comments when writing and using make.

The meson build system

After you have learned the basics of make, which we call a low-level build system, we will introduce meson, a high-level build system. While you specify in a low-level build system how to build your program, you can use a high-level build system to specify what to build. A high-level build system will deal for you with how and generate build files for a low-level build system.

There are plenty of high-level build systems available, but we will focus on meson because it is constructed to be particularly user friendly. The default low-level build-system of meson is called ninja.

Let's have a look at a complete meson.build file:

project('my_proj', 'fortran', meson_version: '>=0.49')
executable('my_prog', files('tabulate.f90', 'functions.f90'))

And we are already done, the next step is to configure our low-level build system with meson setup build, you should see output somewhat similar to this

The Meson build system
Version: 0.53.2
Source dir: /home/awvwgk/Examples
Build dir: /home/awvwgk/Examples/build
Build type: native build
Project name: my_proj
Project version: undefined
Fortran compiler for the host machine: gfortran (gcc 9.2.1 "GNU Fortran (Arch Linux 9.2.1+20200130-2) 9.2.1 20200130")
Fortran linker for the host machine: gfortran ld.bfd 2.34
Host machine cpu family: x86_64
Host machine cpu: x86_64
Build targets in project: 1

Found ninja-1.10.0 at /usr/bin/ninja

The provided information at this point is already more detailed than anything we could have provided in a Makefile, let's run the build with ninja -C build, which should show something like

[1/4] Compiling Fortran object 'my_prog@exe/functions.f90.o'.
[2/4] Dep hack
[3/4] Compiling Fortran object 'my_prog@exe/tabulate.f90.o'.
[4/4] Linking target my_prog.

Find and test your program at build/my_prog to ensure it works correctly. We note the steps ninja performed are the same we would have coded up in a Makefile (including the dependency), yet we did not have to specify them, have a look at your meson.build file again:

project('my_proj', 'fortran', meson_version: '>=0.49')
executable('my_prog', files('tabulate.f90', 'functions.f90'))

We only specified that we have a Fortran project (which happens to require a certain version of meson for the Fortran support) and told meson to build an executable my_prog from the files tabulate.f90 and functions.f90. We had not to tell meson how to build the project, it figured this out by itself.

注釈

meson is a cross-platform build system, the project you just specified for your program can be used to compile binaries for your native operating system or to cross-compile your project for other platforms. Similarly, the meson.build file is portable and will work on different platforms as well.

The documentation of meson can be found at the meson-build webpage.

Creating a CMake project

Similar to meson CMake is a high-level build system as well and commonly used to build Fortran projects.

注釈

CMake follows a slightly different strategy and provides you with a complete programming language to create your build files. This is has the advantage that you can do almost everything with CMake, but your CMake build files can also become as complex as the program you are building.

Start by creating the file CMakeLists.txt with the content

cmake_minimum_required(VERSION 3.7)
project("my_proj" LANGUAGES "Fortran")
add_executable("my_prog" "tabulate.f90" "functions.f90")

Similar to meson we are already done with our CMake build file. We configure our low-level build files with cmake -B build -G Ninja, you should see output similar to this

-- The Fortran compiler identification is GNU 10.2.0
-- Detecting Fortran compiler ABI info
-- Detecting Fortran compiler ABI info - done
-- Check for working Fortran compiler: /usr/bin/f95 - skipped
-- Checking whether /usr/bin/f95 supports Fortran 90
-- Checking whether /usr/bin/f95 supports Fortran 90 - yes
-- Configuring done
-- Generating done
-- Build files have been written to: /home/awvwgk/Examples/build

You might be surprised that CMake tries to use the compiler f95, fortunately this is just a symbolic link to gfortran on most systems and not the actual f95 compiler. To give CMake a better hint you can export the environment variable FC=gfortran rerunning should show the correct compiler name now

-- The Fortran compiler identification is GNU 10.2.0
-- Detecting Fortran compiler ABI info
-- Detecting Fortran compiler ABI info - done
-- Check for working Fortran compiler: /usr/bin/gfortran - skipped
-- Checking whether /usr/bin/gfortran supports Fortran 90
-- Checking whether /usr/bin/gfortran supports Fortran 90 - yes
-- Configuring done
-- Generating done
-- Build files have been written to: /home/awvwgk/Example/build

In a similar manner you could use your Intel Fortran compiler instead to build your project (set FC=ifort).

CMake provides support for several low-level build files, since the default is platform specific, we will just use ninja since we already used it together with meson. As before, build your project with ninja -C build:

[1/6] Building Fortran preprocessed CMakeFiles/my_prog.dir/functions.f90-pp.f90
[2/6] Building Fortran preprocessed CMakeFiles/my_prog.dir/tabulate.f90-pp.f90
[3/6] Generating Fortran dyndep file CMakeFiles/my_prog.dir/Fortran.dd
[4/6] Building Fortran object CMakeFiles/my_prog.dir/functions.f90.o
[5/6] Building Fortran object CMakeFiles/my_prog.dir/tabulate.f90.o
[6/6] Linking Fortran executable my_prog

Find and test your program at build/my_prog to ensure it works correctly. The steps ninja performed are somewhat different, because there is usually more than one way to write the low-level build files to accomplish the task of building a project. Fortunately, we do not have to concern ourselves but have our build system handle those details for us.

Finally, we will shortly recap on our complete CMakeLists.txt to specify our project:

cmake_minimum_required(VERSION 3.7)
project("my_proj" LANGUAGES "Fortran")
add_executable("my_prog" "tabulate.f90" "functions.f90")

We specified that we have a Fortran project and told CMake to create an executable my_prog from the files tabulate.f90 and functions.f90. CMake knows the details how to build the executable from the specified sources, so we do not have to worry about the actual steps in the build process.

Tip

CMake's official reference can be found at the CMake webpage. It is organised in manpages, which are also available with your local CMake installation as well using man cmake. While it covers all functionality of CMake, it sometimes covers them only very briefly.

The SCons build system

SCons is one another high-level cross-platform build system with automatic dependency analysis that supports Fortran projects. SCons configuration files are executed as Python scripts but it could be successfully used without knowledge about Python programming. The using of Python scripting if needed allows to handle build process and file naming in more sophisticated manner.

注釈

SCons doesn't automatically passes the external environment variables such as PATH therefore it will not find programs and tools that installed into non-standard locations unless are specified or passed via appropriate variables. That guaranties that build isn't affected by external (especially user's) environment variables and build is repeatable. The most of such variables and compiler options and flags are required to be configured within special "isolated" Environments (please refer to User Guide for additional information).

The SCons doesn't use external low-level build systems and relies on own build system. The ninja support as external tool to generate ninja.build file is highly experimental (available since scons 4.2) and required to be enabled explicitly by additional configuration.

The simple SCons project is SConstruct file that contains:

Program('my_proj', ['tabulate.f90', 'functions.f90'])

The next step is to build our project with command scons:

scons: Reading SConscript files ...
scons: done reading SConscript files.
scons: Building targets ...
gfortran -o functions.o -c functions.f90
gfortran -o tabulate.o -c tabulate.f90
gfortran -o my_proj tabulate.o functions.o
scons: done building targets.

or with scons -Q to disable extended output:

gfortran -o functions.o -c functions.f90
gfortran -o tabulate.o -c tabulate.f90
gfortran -o my_proj tabulate.o functions.o

Find and test your program my_prog at the same directory as source files (by default) to ensure it works correctly.

To cleanup the build artifacts run scons -c (or scons -Qc):

scons: Reading SConscript files ...
scons: done reading SConscript files.
scons: Cleaning targets ...
Removed functions.o
Removed user_functions.mod
Removed tabulate.o
Removed my_proj
scons: done cleaning targets.

In our SCons SConstruct file

Program('my_proj', ['tabulate.f90', 'functions.f90'])

we specified executable target name my_proj (optional, if omitted the first source file name is used) and list of source files ['tabulate.f90', 'functions.f90']. There is no need to specify project source files language -- it's detected automatically by SCons for supported languages.

The list of source files could be specified with SCons Glob function:

Program('my_proj', Glob('*.f90'))

or using SCons Split function:

Program('my_proj', Split('tabulate.f90 functions.f90'))

or in more readable form by assigning variable:

src_files = Split('tabulate.f90 functions.f90')
Program('my_proj', src_files)

In case of Split function the multiple lines could be used with Python "triple-quote" syntax:

src_files = Split("""tabulate.f90
                     functions.f90""")
Program('my_proj', src_files)

Tip

SCons official webpage provides: User Guide with extended description of various aspects how to handle the build process, Frequently Asked Questions page and man page.