Allowing users to override CFLAGS, CXXFLAGS and friends - makefile

Typical makefiles often use the built-in variables CFLAGS, CXXFLAGS, CPPFLAGS and so on1 to set the flags passed to the C, C++ or other compilers/tools. In principle, this sometimes even lets you avoid writing a compilation recipe entirely since the various built-in rules use these flags.
In general, a makefile might add things to the FLAGS variables that are required for the code to compile, such as include directories, arguments indicating which language standard to use and so on. The variables might also include "optional" or "default" arguments, such as optimization level, warning level and other settings that might validly be altered or removed.
Since CFLAGS and fields are "well known" variables, they are also apparently a configuration point for end users. For example, if a project compiles without debug information by default, it is expected that CFLAGS=-g on the make command line causes -g to be added to the $(CC) compiler command line and hence cause debug info to be produced. Similarly for other options the end user might want to control, such as the optimization level, the -march setting on gcc, and so on.
However, these two uses seem incompatible to me. If the user overrides $(CFLAGS) they will obliterate any internal "required" flags as described above, and the project either may not compile or may compile incorrectly.
Is there a best practice for handling this? The same problem doesn't really arise for "single value" variables like $(CC) since they generally have exactly one value: in this example, the C compiler to use. If the user overrides it, you use their value. Things like $(CFLAGS) are in principle a list of values, some of which are internal and shouldn't be overridden, an others which a user may want to override.
Intuitively, a solution seems to be to leave $(CFLAGS) and friends empty and unused in your makefile, preferring say CFLAGS_INTERNAL for in-makefile arguments, and then put both on the command line. I'm curious, however, if there is a best practice around this or if I'm missing something obvious.
1 For the rest of this question I will often simply refer to $(CFLAGS) with the understanding that this is simply a convenient representative of the whole family of well known compiler flag variables such as $(CPPFLAGS), $(CXXFLAGS) and so on.

I am faced with the same problem. For the time being my solution is to provide "non-standard" flags such as OPTIMS, WARNINGS, MODENV which will be appended to the "standard" CXXFLAGS internally.
If the user defines CXXFLAGS from the command-line it is assumed that he wants to override it, and if that's what he wants, that's what he should get: an override. Ironically this means I'm not using override CXXFLAGS += ... in the Makefile.
I don't want advanced users to pull their hairs out because I insist on appending/prepending my stuff to their flags, so in my opinion the final situation is like this:
GOOD: require advanced users to pass intricate custom flags
BAD: require advanced users to patch the Makefile

Just stumbled upon the same question while building an RPM with debuginfo package.
The requirement for debuginfo generation is to pass -g in CFLAGS while preserving whatever CFLAGS the software has in its Makefile.
So if you want to add some extra bits to CFLAGS, without overwriting the ones present in Makefile, you can simply use CFLAGS as an environment variable. But only as long as the Makefile in question uses CFLAGS += ... notation.
Example, suppose that you have software with Makefile having:
CFLAGS += $(ARCH) -O3 -std=gnu11 -Wall ...
To have it build with all those flags and -g, you will do:
CFLAGS='-g' make
Note that passing it as an argument to make won't work, as in: make CFLAGS='-g' is wrong, because it will overwrite internal CFLAGS.
More on the solution to pass -g for building debuginfo packages properly
Here's reference on make: appending to variables.

The approach I prefer is to provide sensible default values to these common variables, but let users provide their own - overriding the default values.
include $(wildcard makefile.in Makefile.in)
BUILD ?= build
CFLAGS ?= -O2 -fPIC -pedantic -Wall -Wextra -Wconversion
This can be done by either environment variables, command line parameters like make CFLAGS=-g or persistently in a makefile.in.
I am aware that this doesn't exactly pick up the issue you described in the questions, but I found use cases in which users want to compile a project with non-default flags should be able to
Define these variables to their needs
Check their defaults, preferably at the top of the makefile
Maybe adjust the definitions in accordance to the defaults
If someone wants to build with some special flags and is incapable of these steps, there will be some more serious problems anyhow.
This approach will not scale well when the build becomes more involved and the defaults are set across a larger makefile and dependent on other conditions.

The override directive may be what you are looking for:
$ cat Makefile
override CFLAGS += -foobar
all:
$(info CFLAGS = $(CFLAGS))
$ make
CFLAGS = -foobar
make: 'all' is up to date.
$ make CFLAGS=-g
CFLAGS = -g -foobar
make: 'all' is up to date.
Note that you can also use:
$ make CFLAGS+=-g
on the command line but it behaves just like:
$ make CFLAGS=-g

Related

Potential Makefile bug with Target-specific Variable

I recently discovered that
setting a Target-specific Variable
using a conditional assignment (?=)
has the effect of unexporting the global variable using the same name.
For example:
target: CFLAGS ?= -O2
If this statement is anywhere in the Makefile, it has the same impact as unexport CFLAGS for the global variable.
It means that the CFLAGS passed as environment variable to the Makefile will not be passed as environment variable to any sub-makefile, as if it was never set.
Could it be a make bug ?
I couldn't find any mention of this side effect in the documentation.
Example : root Makefile
target:
$(MAKE) -C $(DIR) target
disruptor: CFLAGS ?= -O1
disruptor:
#echo CFLAGS = $(CFLAGS)
and then into $DIR/Makefile:
target:
#echo target CFLAGS = $(CFLAGS)
Now :
make will display target CFLAGS =
make CFLAGS=-Os will display target CFLAGS = -Os
but CFLAGS=-Os make will display target CFLAGS =
after commenting the first disruptor line (CFLAGS ?= -O1), then CFLAGS=-Os make will display target CFLAGS = -Os as expected.
Other mitigations that work :
adding export CFLAGS after the first disruptor line
replacing the ?= assignment by =, := or +=. None of them produce the "implicit unexport" side effect (of course, it also changes the assignment meaning, this is just for test).
I haven't tested with other variable names yet, but I presume it's not specific to CFLAGS.
I reproduce your observed behavior with GNU make 4.0. I concur with your characterization that the effect seems to be as if the variable in question had been unexported, and I confirm that the same effect is observed with other variable names, including names that are without any special significance to make.
This effect is undocumented as far as I can tell, and unexpected. It seems to conflict with the manual, in that the manual describes target-specific variable values as causing a separate instance of the affected variable to be created, so as to avoid affecting the global one, yet we do see the global one being affected.
Could it be a make bug ?
It indeed does look like a bug to me. Evidently to other people, too, as it appears that the issue has already been reported.

Should CXXFLAGS, CPPFLAGS and LDFLAGS be overridden in a Makefile?

Situation
I am using a handwritten GNUmakefile in which CXXFLAGS, CPPFLAGS and LDFLAGS are appended to by the += assignment, as in:
CXXFLAGS += -std=c++11 $(MODENV) $(WARNINGS) $(OPTIMS)
CPPFLAGS += $(DMACROS) $(INCDIRS)
LDFLAGS += $(MODENV) $(LIBDIRS) $(EXTRA_LIBS)
Problem
When the user defines his own flags at the command-line, the appending in the Makefile will be ignored. This leaves the variables to exactly what the user set them. (And in my case, the build will fail.) The generic solution for this problem is to override the variables, as in:
override CXXFLAGS += -std=c++11 $(MODENV) $(WARNINGS) $(OPTIMS)
override CPPFLAGS += $(DMACROS) $(INCDIRS)
override LDFLAGS += $(MODENV) $(LIBDIRS) $(EXTRA_LIBS)
This way, the necessary content will be appended to the user's variable.
Questions
Is overriding variables considered bad practice?
Is setting the above flags inside the Makefile considered bad practice?
If "yes" to both questions above, then where do I put -std=c++11, if not in CXXFLAGS?
If a user overrides the variables, assume the user knows what they're doing. The system they're working on may have very different requirements and need to override the variables. Don't make the difficult impossible.
I'm not going to comment on good/bad practice, as it depends on the use case (who are the users of this makefile? Do they all work with you, so you could just tell them how it's meant to be used? Can you just document your conventions, in the makefile comments or a project README?)
But ...
where do I put -std=c++11, if not in CXXFLAGS?
If you have flags that are essential and must not be overridden by the user, you can find another way to put them in the compilation command.
For example:
CXX := g++ -std=c++11
Or:
foo.o: foo.cc
$(CXX) $(CPPFLAGS) $(CXXFLAGS) -std=c++11 -o $# $^
Even better might be to use a variable so that it can still be set by the user (e.g. to use -std=c++14 instead) but isn't set by CXXFLAGS:
STD := -std=c++11
foo.o: foo.cc
$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(STD) -o $# $^
(Or just put the -std option before CXXFLAGS so that a different std option in CXXFLAGS will take precedence).
As the use cases for build management with make are so diverse, there simply is no ground to objectively judge this as "bad practice". If you are writing OSS for a multitude of platforms and for an unknown audience and timeframe, the POLA/POLS should be applied. That said, a user who is surprised to find more flags than the ones she gave on the command line is a rather unrealistic corner case, so -override has its place. In the end the values you add are absolutely necessary for the build, aren't they?
PS: the POLA should be applied to all engineering activities of course - its just that the definition of "surprise" shifts with the task in question.

So what "is" AM_(F)CFLAGS?

The autotools documentation is very confusing. I am writng Fortran, so the AM_CFLAGS equivalent is AM_FCFLAGS. The two work exactly the same way (presumably).
First off, what actually "is" AM_CFLAGS, conceptually? Clearly, the "CFLAGS" bit is to do with setting compiler flags. But what does the "AM_" part mean?
There seems to be conflicting advice as to how to use it. Some say don't put it in Makefile.am, and some say don't put it in configure.ac. Who is right?
Here is my current Makefile.am:
AM_FCFLAGS = -Wall -O0 -C -fbacktrace
.f90.o:
$(FC) -c $(AM_FCFLAGS) $<
What I want to happen is to compile with "-Wall -O0 -C -fbacktrace" by default if I'm compiling with gfortran. However, a user might want to use a different compiler, eg FC=ifort, in which case they'll probably have to pass in FCFLAGS="whatever" and completely scrap AM_FCFLAGS
Can the user also override the default AM_FCFLAGS from the configure option if they're still using gfortran?
Basically, WTF?
AM_FCFLAGS (and similarly AM_CFLAGS and similar) are designed to not be user-overridable, so you should not put those options there unless you want them to always be present.
Users can pass their own FCFLAGS as part of their ./configure call — what you can do, if you want to default to those rather than what autoconf will default by itself, is to change configure.ac and compare the default flags (which to be honest I don't know for Fortran) to the current FCFLAGS, and if they match, replace FCFLAGS with your defaults.
In Makefile.am I had
AM_CFCFLAGS = -Wall -O0 -C -fbacktrace
Bad idea! It assumes that folks are using gfortran and/or won't want to override those defaults. So I deleted that line.
Instead, I now have the following lines in configure.ac:
AC_PROG_FC([gfortran], [Fortran 90]) # we need a Fortran 90 compiler
AS_IF([test x$FC = xgfortran -a x$ac_cv_env_FCFLAGS_set = x],
AC_SUBST([FCFLAGS], ["-Wall -O0 -C -fbacktrace"])
[Set some flags automatically if using gfortran])
AC_PROG_FC checks for gfortran that meets with Fortran 90 standards, and automatically sets FC and FCFLAGS.
The last 3 lines set sensible defaults if the user is using gfortran, but hasn't set FCFLAGS.
I discovered about ac_cv_env_FCFLAGS_set when I looked at config.log. It is set to "set" if the user sets their own FCFLAGS.
In Makefile.am I now have rules like:
.f90.o:
$(FC) -c $(FCFLAGS) $<
datetime_module.mod : datetime.o
datetime.o : datetime.f90 mod_clock.o mod_datetime.o mod_strftime.o mod_timedelta.o
mod_clock.o: mod_clock.f90 mod_datetime.o mod_timedelta.o
mod_datetime.o: mod_datetime.f90 mod_constants.o mod_strftime.o mod_timedelta.o
mod_timedelta.o: mod_timedelta.f90
It's starting to make sense now.

pass variable to make file

I read documents about CPPFLAGS. Shortly, I understand that CPPFLAGS used for pass parameter to compiler. Sample usage CPPFLAGS in makefile is below.
gcc $(CPPFLAGS) main.c -o main.o
Execute make
make CPPFLAGS=-I../header
What is special CPPFLAGS text? It can be interchangable any other text like "FOO". What is the differences between FOO variable and CPPFLAGS variable? Replace all CPPFLAGS text with FOO text build is success again, nothing changes.
Main problem that actually need to solve. There are lots of makefiles. There is no include CPPFLAGS variable in these makefiles. Is there a way to pass compiler options without change makefiles.
Thanks.
What is special CPPFLAGS text? It can be interchangable any other text like "FOO"
Three things are special about CPPFLAGS:
It is a convention that many tools follow. Most notably GNU autoconf/automake.
GNU Make provides implicit rules to build many target types. These implicit rules use CPPFLAGS variable when compiling .o from C and C++ sources. These rules can be replaced with one's own rules if necessary.
When you use CPPFLAGS for preprocessor flags you follow the principle of least astonishment.
There is no include CPPFLAGS variable in these makefiles
If you meant that there are no occurrences of CPPFLAGS in the makefiles that may be because the implicit rules are used which I mentioned above.

CFLAGS vs CPPFLAGS

I understand that CFLAGS (or CXXFLAGS for C++) are for the compiler, whereas CPPFLAGS is used by the preprocessor.
But I still don't understand the difference.
I need to specify an include path for a header file that is included with #include -- because #include is a preprocessor directive, is the preprocessor (CPPFLAGS) the only thing I care about?
Under what circumstances do I need to give the compiler an extra include path?
In general, if the preprocessor finds and includes needed header files, why does it ever need to be told about extra include directories? What use is CFLAGS at all?
(In my case, I actually found that BOTH of these allow me to compile my program, which adds to the confusion... I can use CFLAGS OR CPPFLAGS to accomplish my goal (in autoconf context at least). What gives?)
The implicit make rule for compiling a C program is
%.o:%.c
$(CC) $(CPPFLAGS) $(CFLAGS) -c -o $# $<
where the $() syntax expands the variables. As both CPPFLAGS and CFLAGS are used in the compiler call, which you use to define include paths is a matter of personal taste. For instance if foo.c is a file in the current directory
make foo.o CPPFLAGS="-I/usr/include"
make foo.o CFLAGS="-I/usr/include"
will both call your compiler in exactly the same way, namely
gcc -I/usr/include -c -o foo.o foo.c
The difference between the two comes into play when you have multiple languages which need the same include path, for instance if you have bar.cpp then try
make bar.o CPPFLAGS="-I/usr/include"
make bar.o CFLAGS="-I/usr/include"
then the compilations will be
g++ -I/usr/include -c -o bar.o bar.cpp
g++ -c -o bar.o bar.cpp
as the C++ implicit rule also uses the CPPFLAGS variable.
This difference gives you a good guide for which to use - if you want the flag to be used for all languages put it in CPPFLAGS, if it's for a specific language put it in CFLAGS, CXXFLAGS etc. Examples of the latter type include standard compliance or warning flags - you wouldn't want to pass -std=c99 to your C++ compiler!
You might then end up with something like this in your makefile
CPPFLAGS=-I/usr/include
CFLAGS=-std=c99
CXXFLAGS=-Weffc++
The CPPFLAGS macro is the one to use to specify #include directories.
Both CPPFLAGS and CFLAGS work in your case because the make(1) rule combines both preprocessing and compiling in one command (so both macros are used in the command).
You don't need to specify . as an include-directory if you use the form #include "...". You also don't need to specify the standard compiler include directory. You do need to specify all other include-directories.
You are after implicit make rules.
To add to those who have mentioned the implicit rules, it's best to see what make has defined implicitly and for your env using:
make -p
For instance:
%.o: %.c
$(COMPILE.c) $(OUTPUT_OPTION) $<
which expands
COMPILE.c = $(CXX) $(CXXFLAGS) $(CPPFLAGS) $(TARGET_ARCH) -c
This will also print # environment data. Here, you will find GCC's include path among other useful info.
C_INCLUDE_PATH=/usr/include
In make, when it comes to search, the paths are many, the light is one... or something to that effect.
C_INCLUDE_PATH is system-wide, set it in your shell's *.rc.
$(CPPFLAGS) is for the preprocessor include path.
If you need to add a general search path for make, use:
VPATH = my_dir_to_search
... or even more specific
vpath %.c src
vpath %.h include
make uses VPATH as a general search path so use cautiously. If a file exists in more than one location listed in VPATH, make will take the first occurrence in the list.
I installed httpd on Ubuntu 18.04 using the CPPFLAGS variable for the -DLINUX flag. When run, CPPFLAGS scans the code from top to bottom, file by file, looking for directives before compiling, and will not be extended by other meaningful things like size optimization, flags that do not increase the size of the output file; under the type of processor; to reduce the size of the code and speed up the program; disable all variables except case. The only difference between CPPFLAGS and CFLAGS is that CFLAGS can be set to specify additional switches to be passed to the compiler. That is, the CFLAGS environment variable creates a directory in the installation path (eg CFLAGS=-i/opt/include) to add debugging information to the executable target's path: include general alarm messages; turning off alarm information; independent location generation; display compiler driver, preprocessor, compiler version number.
Standard way to set CPPFLAGS:
sudo ./configure --enable-unixd=DLINUX #for example
list of some known variables:
CPPFLAGS - is the variable name for flags to the C preprocessor.
CXXFLAGS - is the standard variable name for flags to the C++ compiler.
CFLAGS is - the standard name for a variable with compilation flags.
LDFLAGS - should be used for search flags/paths (-L) - i.e. -L/usr/lib (/usr/lib are library binaries).
LDLIBS - for linking libraries.
CPPFLAGS seems to be an invention of GNU Make, referenced in some of its built-in recipes.
If your program is built by some Free software distributions, you may find that some of them require packages to interpolate this variable, using CPPFLAGS for passing down options like -D_WHATEVER=1 for passing down a macro definition.
This separation is a poor idea and completely unnecessary in the GNU environment because:
There is a way to run gcc to do preprocessing only (while ignoring compiler options unrelated to preprocessing).
The stand-alone GNU cpp is tolerant to compiler options, such as -W warnings that do not pertain to preprocessing and even code generation options like -fstrict-aliasing and the linker-pass through like -Wl,--whatever.
So generally speaking, build systems that need to call the stand-alone preprocessor for whatever reason can just pass it $(CFLAGS).
As an application developer writing a Makefile, you cannot rely on the existence of CPPFLAGS. Users who are not insider experts in open source building won't know about CPPFLAGS, and will do things like make CFLAGS=-Dfoo=bar when building your program. If that doesn't work, they will be annoyed.
As a distro maintainer, you cannot rely on programs to pull in CPPFLAGS; even otherwise well-behaved ones that pull in CFLAGS, LDFLAGS and LDLIBS.
It's easy enough for the application developers to write GNU Make code to separate preprocessor flags out of $(CFLAGS):
cpp_only_flags := $(foreach arg, \
$(CFLAGS), \
$(or $(filter -D%,$(arg)), \
$(filter -U%,$(arg)), \
$(filter -I%,$(arg)), \
$(filter -iquote%,$(arg)), \
$(filter -W%,$(arg)), \
$(filter -M%,$(arg)))) \
$(CPPFLAGS) # also pull this in
all:
#echo cpp_only_flags == $(cpp_only_flags)
Demo:
$ make CFLAGS="-Wall -I/path/to/include -W -UMAC -DFOO=bar -o foo.o -lm"
cpp_only_flags == -Wall -I/path/to/include -W -UMAC -DFOO=bar
In the case of the GNU compiler and preprocessor, this is probably unnnecessary; but it illustrates a technique that could be used for non-GNU compilers and preprocessors, in a build system based on GNU Make.

Resources