Stupid, stupid makedepend

One of the most frustrating, time-wasting debugging experiences you can have is when the code you thought you compiled didn't really get compiled. Or something you didn't even think about because someone else updated it didn't get compiled.

This is why the utility makedepend was introduced, so that compile dependencies get taken care of automatically, every time. So why is the Linux makedepend so stupid that it doesn't work every time?!?!

In the interest of speed, the Linux makedepend was written to parse include files only once per run (usually this is per directory). That's usually OK, but when the include file is preprocessed differently by different files (or multiple times in one file), it doesn't work. I thought the first rule of optimization was: make the common case fast, but make every case work.

Moral: use the dependency scheme described in the GNU make info pages. It goes like this:

# In the variables area:
sources = foo.cpp bar.cpp ...
OBJS = $(sources:.cpp=.o)

# In the targets area, somewhere after "all:"
ifneq ($(MAKECMDGOALS),clean)
include $(sources:.cpp=.d)
endif

%.d: %.cpp
  @(set -o pipefail; \
    $(CXX) -MM $(CXXFLAGS) $< \
     | sed 's,\($*\)\.o[ :]*,\1.o $@ : ,g' \
      > $@;) || $(RM) $@

Of course this assumes that CXX = g++. By the way, that little gem about not building the dependencies during "clean" is in the info pages, but not the page that describes the generic dependency rules. Yes, I have read every make info page.

The "pipefail" bit is my own little contribution; it saves you some headaches when the g++ command fails when you're not watching, because the sed command will succeed and allow make to continue ignorantly on its way. That's a bash option, so it might not work on platforms other than Linux.