=head1 NAME
makepp_build_check -- How makepp decides to rebuild files
=for vc $Id: makepp_build_check.pod,v 1.14 2012/02/07 22:26:15 pfeiffer Exp $
=head1 DESCRIPTION
=for genindex '\w+' makepp_build_check.pod
B<A:>E<nbsp>L</architecture_independent>,E<nbsp>
B<E:>E<nbsp>L</exact_match>,E<nbsp>
B<I:>E<nbsp>L</ignore_action>,E<nbsp>
B<O:>E<nbsp>L</only_action>,E<nbsp>
B<T:>E<nbsp>L</target_newer>
Makepp stores a variety of information about how any given file was
built the last time. This information includes the build command, the
architecture, and the signatures of all the file's dependencies. (All
the stored information is in the subdirectory F<.makepp> of each
directory.) If any of this information has changed, makepp usually
decides to rebuild the file. The build check method is what controls
makepp's decision to rebuild. It decides which information to look at,
and which to ignore.
Makepp usually picks the correct build check method automatically. However,
you can change the signature method for an individual rule by using
L<:build_check|makepp_rules/build_check_build_check_method> modifier on the
rule, or for all rules in a makefile by using the
L<build_check|makepp_statements/build_check_build_check_method> statement, or
for all makefiles at once using the L<-m or
--build-check-method|makepp_command/build_check_method> command line option.
The data used to decide about a rebuild or a repository or build cache import
is stored in the internal build info file. You can display it with
L<makeppinfo, mppi|makeppinfo>. Below each method gives an example of how to
see its keys.
=head2 Build check methods included in the distribution
At present, there are five build check methods included in the distribution:
=over 4
=item exact_match
This method uses the modification dates on the file as signatures. It
rebuilds the targets unless all of the following conditions are true:
=over 4
=item *
The signature of each dependency is the same as it was on the
last build.
=item *
The signature of each target is the same as it was on the last
build (i.e., no one has messed with the targets since makepp built
them).
=item *
The build command has not changed.
=item *
The machine architecture (or what Perl thinks it is) has not changed.
=back
C<exact_match> is the default method unless you are rebuilding a
Makefile (see below). This is a highly reliable way of ensuring correct
builds, and is almost always what you want. However, it does have a few
side effects that may be surprising:
=over 4
=item *
If you've been compiling with the traditional make, and then switch to
makepp, everything is recompiled the first time you run makepp.
=item *
If you damage makepp's information about what happened on the last build
(e.g., you delete the subdirectory C<.makepp>, or don't copy it when you
copy everything else), then a rebuild is triggered.
=item *
If you replace a file with an older version, a rebuild is triggered.
This is normally what you want, but it might be surprising.
=item *
If you modify a file outside of the control of makepp (e.g., you run
the compilation command yourself), then makepp will rebuild the file
next time. (If you want to avoid this, check out the C<--dont-build>
command line option.)
=item *
Architecture-independent files are rebuilt when you switch to a
different architecture. This is usually not a problem, because they
often don't take long to build. The reason why all files are tagged
with the architecture, instead of just binary files, is that often times
even ASCII files are architecture-dependent. For example, output from
the Solaris C<lex> program won't compile on Linux (or at least this used
to be true the last time I tried it).
=back
Concretely, a file will not be rebuilt, or can be fetched from repository or
build cache, if the following command output stays the same, i.e. matches the
signatures of the dependencies:
mppi -k'COMMAND ARCH SORTED_DEPS DEP_SIGS ENV_{DEP,VAL}S' file
=item architecture_independent
The C<architecture_independent> method is the same as C<exact_match>
except that it does not check the architecture. This can be useful for
architecture-independent files, that don't need to be rebuilt when you
switch to a different architecture. For example, you probably don't
need to rerun C<bison> on Solaris if you already ran it on Linux.
The C<architecture_independent> method is best used by specifying it
using the S<C<:build_check architecture_independent>> modifier to the
each rule that produces architecture independent files. Makepp by
default never assumes any files are architecture independent, because
even F<.c> files can be architecture dependent. For example, the output
of Solaris lex will not compile under Linux, or at least it wouldn't
last time I tried. So you must manually specify this build check method
for any files which are truly architecture-independent.
Concretely, a file will not be rebuilt, or can be fetched from repository or
build cache, if the following command output stays the same, i.e. matches the
signatures of the dependencies:
mppi -k'COMMAND SORTED_DEPS DEP_SIGS ENV_{DEP,VAL}S' file
=item ignore_action
The C<ignore_action> method is the same as C<exact_match> except that it does
not check the action string (the command). Sometimes a command can change and
you don't want to force a rebuild.
For example, you might want to explicitly put a date into your command to log
when the build was done, but you don't want to force a rebuild every time the
command is executed. For example,
BUILD_DATE := $(shell date)
my_program : $(MODULES).o
$(CXX) $(inputs) -DBUILD_DATE="\"$(BUILD_DATE)\"" date_stamp.c -o $(output)
This will compile F<date_stamp.c> with the last build date stamp, but won't
force a recompile when the date changes. Unfortunately, if something else
about the link command changes (e.g., you change linker options), it also
won't trigger a rebuild.
This is also useful in conjunction with the $(changed_inputs) or $? variable
for actions that merely update a target, rather than rebuilding it from
scratch. For example, you could update a .a file like this:
libmine.a : *.o : build_check ignore_action
$(AR) ru $(output) $(changed_inputs)
This will still mostly work if you forget to specify the
C<: build_check ignore_action>. However, suppose that none of the
.o files have changed. The command will now be S<C<ar ru libmine.a>> which
is probably different from what it was last time (e.g.,
S<C<ar ru libmine.a buggy_module.o>>), so makepp will run the command. In
this case, the command won't do anything except waste time.
Building .a files like this is discouraged, because it can leave stale
.o files inside the archive. If you delete a source file, the .o file
is still inside the .a file, and this can lead to incorrect builds.
It's better to build a .a file like this:
libmine.a : *.o
&rm $(output)
$(AR) ru $(output) $(inputs)
Concretely, a file will not be rebuilt, or can be fetched from repository or
build cache, if the following command output stays the same, i.e. matches the
signatures of the dependencies:
mppi -k'ARCH SORTED_DEPS DEP_SIGS ENV_{DEP,VAL}S' file
=item target_newer
The C<target_newer> method looks only at the file date. If any
dependency is more recent than the target, the target is rebuilt. This
is the algorithm that the traditional Unix I<make> utility uses.
The C<target_newer> method isn't as safe as the C<exact_match> method because
it won't trigger a rebuild if you change the build command, or if you replace
a file with an older version. Sometimes also it can get confused if clocks
are not properly synchronized. For example, if a file somehow gets a date of
June 4, 2048, then between now and 2048, every file that depends on that file
will be rebuilt even though the file doesn't change. Also switching to a
different architecture won't trigger a rebuild. It prevents fetching a rule's
target from a build cache, because there is no unique signature that can be
associated to the endless set of pairs fulfilling the relationship newer than.
But there are a few cases where you may want to use the C<target_newer>
method:
=over 4
=item *
When it is reasonable for a user to build a file outside of the control
of makepp. Perhaps the most common example are the commands that
generate the makefile itself, i.e., the autoconfigure procedure. Users
commonly issue the configure command manually, but makefiles often have
a way to update themselves automatically. In this case, we don't want
to force the makefile to rebuild itself if the user typed the command in
manually, so the C<target_newer> method is more appropriate than the
C<exact_match> method. In fact, if makepp is trying to build a
makefile, it makes C<target_newer> the default method until it has
finished building the makefile.
=item *
When it is reasonable for a user to modify a file after makepp has built
it. For example, if a file does not exist, you may want to copy it from
a central location, or check it out from a repository; but the user
should be allowed to modify it. If you use the default C<exact_match>
build check method, makepp will detect that the user has changed the
file and so it will force a fresh copy from the central location or a
fresh checkout, wiping out the user's changes.
=back
If you need to manually check the timestamps, see L<makeppinfo
examples|makeppinfo/EXAMPLES> for how to get the path of each dependency.
=item only_action
The very specific C<only_action> method will only execute the action if the
command string differs from the last time it was executed. For example,
$(ROOT)/include/%.h : %.h
&ln -fr $(input) $(output)
publishes a file, but does not repeat this when the file changes. Note that
the C<&ln> command is builtin and thus cheap, but makepp still has to fork off
and monitor a process to perform the whole action. So if you have lots of
files to publish, there is still a benefit. Actually we did not specify the
method, because, when the target is a symbolic link, this build check gets
used automatically. You only need to specify it for other commands that
depend solely on the command (which usually contains the names of the inputs):
%.list : %.x : build_check only_action
&echo $(inputs) -o $(output)
Concretely, a file will not be rebuilt, or can be fetched from repository or
build cache, if the following command output stays the same, i.e. matches the
signatures of the dependencies:
mppi -kCOMMAND file
=back
Other build check methods are possible. You can write your own build
check method by creating a module C<Mpp::BuildCheck::MyMethod>. Read the
documentation in F<Mpp/BuildCheck.pm> in the makepp distribution. Most
likely, you will want your build check method to inherit from
C<Mpp::BuildCheck::exact_match>, so read its documentation too.
It's more commonly useful modify the signature mechanism than to modify
the build check mechanism directly. Before you change the build check
mechanism, see if your problem is better served by changing signatures
(see L<makepp_signatures> for details).
Here are some reasons why a custom build check method might be useful:
=over 4
=item *
If you want makepp to ignore part of the command. For example, if you
have commands in your makefile like this:
x.o : x.c
ssh $(REMOTE_MACHINE) cc $< -o $@
you might want makepp not to force a rebuild if C<$(REMOTE_MACHINE)> changes.
You could modify the C<exact_match> method so it knows about ssh commands and
ignores the machine name. Check L<:dispatch|makepp_rules/dispatch_command>
for another way to achieve that.
=back