Makefile, header dependencies

问题:

Let\'s say I have a makefile with the rule

%.o: %.c
 gcc -Wall -Iinclude ...

I want *.o to be rebuilt whenever a header file changes. Rather than work out a list of dependencies, whenever any header file in /include changes, then all objects in the dir must be rebuilt.

I can\'t think of a nice way to change the rule to accomodate this, I\'m open to suggestions. Bonus points if the list of headers doesn\'t have to be hard-coded

回答1:

If you are using a GNU compiler, the compiler can assemble a list of dependencies for you. Makefile fragment:

depend: .depend

.depend: $(SRCS)
        rm -f ./.depend
        $(CC) $(CFLAGS) -MM $^ -MF  ./.depend;

include .depend

or

depend: .depend

.depend: $(SRCS)
        rm -f ./.depend
        $(CC) $(CFLAGS) -MM $^ > ./.depend;

include .depend

where SRCS is a variable pointing to your entire list of source files.

There is also the tool makedepend, but I never liked it as much as gcc -MM

回答2:

Most answers are surprisingly complicated or erroneous. However simple and robust examples have been posted elsewhere [codereview]. Admittedly the options provided by the gnu preprocessor are a bit confusing. However, the removal of all directories from the build target with -MM is documented and not a bug [gpp]:

By default CPP takes the name of the main input file, deletes any directory components and any file suffix such as ‘.c’, and appends the platform\'s usual object suffix.

The (somewhat newer) -MMD option is probably what you want. For completeness an example of a makefile that supports multiple src dirs and build dirs with some comments. For a simple version without build dirs see [codereview].

CXX = clang++
CXX_FLAGS = -Wfatal-errors -Wall -Wextra -Wpedantic -Wconversion -Wshadow

# Final binary
BIN = mybin
# Put all auto generated stuff to this build dir.
BUILD_DIR = ./build

# List of all .cpp source files.
CPP = main.cpp $(wildcard dir1/*.cpp) $(wildcard dir2/*.cpp)

# All .o files go to build dir.
OBJ = $(CPP:%.cpp=$(BUILD_DIR)/%.o)
# Gcc/Clang will create these .d files containing dependencies.
DEP = $(OBJ:%.o=%.d)

# Default target named after the binary.
$(BIN) : $(BUILD_DIR)/$(BIN)

# Actual target of the binary - depends on all .o files.
$(BUILD_DIR)/$(BIN) : $(OBJ)
    # Create build directories - same structure as sources.
    mkdir -p $(@D)
    # Just link all the object files.
    $(CXX) $(CXX_FLAGS) $^ -o $@

# Include all .d files
-include $(DEP)

# Build target for every single object file.
# The potential dependency on header files is covered
# by calling `-include $(DEP)`.
$(BUILD_DIR)/%.o : %.cpp
    mkdir -p $(@D)
    # The -MMD flags additionaly creates a .d file with
    # the same name as the .o file.
    $(CXX) $(CXX_FLAGS) -MMD -c $< -o $@

.PHONY : clean
clean :
    # This should remove all generated files.
    -rm $(BUILD_DIR)/$(BIN) $(OBJ) $(DEP)

This method works because if there are multiple dependency lines for a single target, the dependencies are simply joined, e.g.:

a.o: a.h
a.o: a.c
    ./cmd

is equivalent to:

a.o: a.c a.h
    ./cmd

as mentioned at: Makefile multiple dependency lines for a single target?

回答3:

As I posted here gcc can create dependencies and compile at the same time:

DEPS := $(OBJS:.o=.d)

-include $(DEPS)

%.o: %.c
    $(CC) $(CFLAGS) -MM -MF $(patsubst %.o,%.d,$@) -o $@ $<

The \'-MF\' parameter specifies a file to store the dependencies in.

The dash at the start of \'-include\' tells Make to continue when the .d file doesn\'t exist (e.g. on first compilation).

Note there seems to be a bug in gcc regarding the -o option. If you set the object filename to say obj/_file__c.o then the generated file.d will still contain file.o, not obj/_file__c.o.

回答4:

How about something like:

includes = $(wildcard include/*.h)

%.o: %.c ${includes}
    gcc -Wall -Iinclude ...

You could also use the wildcards directly, but I tend to find I need them in more than one place.

Note that this only works well on small projects, since it assumes that every object file depends on every header file.

回答5:

Martin\'s solution above works great, but does not handle .o files that reside in subdirectories. Godric points out that the -MT flag takes care of that problem, but it simultaneously prevents the .o file from being written correctly. The following will take care of both of those problems:

DEPS := $(OBJS:.o=.d)

-include $(DEPS)

%.o: %.c
    $(CC) $(CFLAGS) -MM -MT $@ -MF $(patsubst %.o,%.d,$@) $<
    $(CC) $(CFLAGS) -o $@ $<

回答6:

This will do the job just fine , and even handle subdirs being specified:

    $(CC) $(CFLAGS) -MD -o $@ $<

tested it with gcc 4.8.3

回答7:

I prefer this solution, over the accepted answer by Michael Williamson, it catches changes to sources+inline files, then sources+headers, and finally sources only. Advantage here is that the whole library is not recompiled if only a a few changes are made. Not a huge consideration for a project with a couple of files, bur if you have 10 or a 100 sources, you will notice the difference.

COMMAND= gcc -Wall -Iinclude ...

%.o: %.cpp %.inl
    $(COMMAND)

%.o: %.cpp %.hpp
    $(COMMAND)

%.o: %.cpp
    $(COMMAND)

回答8:

The following works for me:

DEPS := $(OBJS:.o=.d)

-include $(DEPS)

%.o: %.cpp
    $(CXX) $(CFLAGS) -MMD -c -o $@ $<

回答9:

Here\'s a two-liner:

CPPFLAGS = -MMD
-include $(OBJS:.c=.d)

This works with the default make recipe, as long as you have a list of all your object files in OBJS.

回答10:

A slightly modified version of Sophie\'s answer which allows to output the *.d files to a different folder (I will only paste the interesting part that generates the dependency files):

$(OBJDIR)/%.o: %.cpp
# Generate dependency file
    mkdir -p $(@D:$(OBJDIR)%=$(DEPDIR)%)
    $(CXX) $(CXXFLAGS) $(CPPFLAGS) -MM -MT $@ $< -MF $(@:$(OBJDIR)/%.o=$(DEPDIR)/%.d)
# Generate object file
    mkdir -p $(@D)
    $(CXX) $(CXXFLAGS) $(CPPFLAGS) -c $< -o $@

Note that the parameter

-MT $@

is used to ensure that the targets (i.e. the object file names) in the generated *.d files contain the full path to the *.o files and not just the file name.

I don\'t know why this parameter is NOT needed when using -MMD in combination with -c (as in Sophie\'s version). In this combination it seems to write the full path of the *.o files into the *.d files. Without this combination, -MMD also writes only the pure file names without any directory components into the *.d files. Maybe somebody knows why -MMD writes the full path when combined with -c. I have not found any hint in the g++ man page.