[subdirmk.git] / README

subdirmk - assistance for non-recursive use of make
===================================================

Introduction
------------

Peter Miller's 1997 essay _Recursive Make Considered Harmful_
persuasively argues that it is better to arrange to have a single
make invocation with the project's complete dependency tree, rather
than the currently conventional `$(MAKE) -C subdirectory' approach.

However, actually writing a project's build system in a non-recursive
style is not very ergonomic.  The main difficulties are:
  - constantly having to write out long file and directory names
  - the lack of a per-directory make variable namespace means
    long make variables (or namespace clashes)
  - it is difficult to arrange that one can cd to a subdirectory
    and say `make all' and have something reasonable happen
    (to wit, build an appropriate subset)

`subdirmk' is an attempt to solve these problems (and it also slightly
alleviates some of the boilerplate needed to support out-of-tree
builds well).

Basic approach
--------------

The developer is expected to write a makefile fragment, in each
relevant subdirectory, called `Subdir.sd.mk'.

These fragments may contain ordinary make language.

However, the sigil & is treated specially.  By and large, it refers to
`the current directory'.  There are a variety of convenient
constructions.

The result is that to a large extent, the Subdir.sd.mk has an easy way
to namespace its "local" make variables, and an easy way to refer to
its "local" filenames.

The Subdir.sd.mk's are filtered, fed through autoconf in the usual way
(for @..@-substitutions) and included by one autogenerated toplevel
makefile.

So all of the input is combined and passed to one make invocation.
(A corollary is that there is no enforcement of the namespacing:
discipline is required to prefix relevant variable names with &, etc.)

Each subdirectory is also provided with an autogenerated `Makefile'
which exists purely to capture ordinary make invocations and arrange
for something suitable to happen.

Where there are dependencies between subdirectories, each Subdir.sd.mk
can simply refer to files in other subdirectories directly.

Invocation, "recursive" per-directory targets
---------------------------------------------

Arrangements are made so that when you run `make foo' in a
subdirectory, it is like running the whole toplevel makefile, from the
toplevel, as `make subdir/foo'.  If `subdir/foo' is a file that might
be built, that builds it.

But `foo' can also be a conventional target like `all'.

Each subdirectory has its own `all' target.  For example a
subdirectory `src' has a target `src/all'.  The rules for these are
automatically generated from the settings of the per-directory
&TARGETS variables.  &TARGETS is magic in this way.  (In
src/Subdir.sd.mk, &TARGETS of course refers to a make variable called
src_TARGETS.)

The `all' target in a parent directory is taken to imply the `all'
targets in all of its subdirectories, recursively.  And in the
autogenerated stub Makefiles, `all' is the default target.  So if you
just type `make' in the toplevel, you are asking for `&all'
(<subdir>/all) for every directory in the project.

In a parallel build, the rules for all these various subdirectory
targets may be in run in parallel: there is only one `make' invocation
at a time.  There is no sequencing between subdirectories, only been
individual targets (as specified according to their dependencies).

You can define other per-directory recursive targets too: simply
mention (usually, by setting) the variable &TARGETS_zonk, or whatever.
This will create a src/zonk target (for appropriate value of src/).
Unlike `all', these other targets only exist in areas of the project
where at least something mentions them.  So for example, if
&TARGETS_zonk is mentioned in src but not lib, `make zonk' in
lib will fail.  If you want to make a target exist everywhere,
mention its name in Perdir.sd.mk (see below).

Perdir.sd.mk, inclusion
-----------------------

The file Perdir.sd.mk in the toplevel of the source is automatically
processed after each individual directory's Subdir.sd.mk, and the
&-substituted contents therefore appear once for each subdirectory.

This lets you do per-directory boilerplate.  Some useful boilerplate
is already provided in subdirmk, for you to reference like this:
  &:include subdirmk/cdeps.sd.mk
  &:include subdirmk/clean.sd.mk
For example you could put that in Perdir.sd.mk.

Global definitions
------------------

If want to set global variables, such as CC, that should only be done
once.  You can put them in your top-level Subdir.sd.mk, or a separate
file you `include' and declare using SUBDIRMK_MAKEFILES.

If you need different settings of variables like CC for different
subdirectories, you should probably do that with target-specific
variable settings.  See the info node `(make) Target-specific'.

Subdirectory templates `.sd.mk' vs plain autoconf templates `.mk.in'
--------------------------------------------------------------------

There are two kinds of template files.

 Filename                 .sd.mk                  .mk.in

 Processed by             &-substitution,         autoconf only
                          then autoconf

 Instantiated             Usu. once per subdir    Once only

 Need to be mentioned     No, but Subdir.sd.mk    All not in subdirmk/
 in configure.ac?         via SUBDIRMK_SUBDIRS    via SUBDIRMK_MAKEFILES

 How to include           `&:include foo.sd.mk'   `include foo.mk'
                          in all relevant .sd.mk  in only one
                          (but not needed for     Subdir.sd.mk
                           Subdir and Perdir)

If you `include subdirmk/regen.mk', dependency management and
automatic regeneration for all of this template substitution, and for
config.status etc. is done for you.

Summary of recommended directory reference syntaxes
---------------------------------------------------

Path construction &-expansions, meanings summary:

                      In build tree             In source tree
  This directory        just &                    &,
  Top level             &.  implies absolute      &;

Adding `@' means "absolute path".  This is not needed with &. because
there is never any need to use &. since it would expand to nothing.
`/' terminates the escape (needed if the next thing is not a lowercase
character, or space).  `=' means "just the value, no /".  Space starts
multi-word processing.

      Recommended     In build tree             In source tree
             when     Relative  Absolute        Relative     Absolute
                                          
  This       lc       &file     &@file          &,file       &@,file
  directory  any      &/file    &@/file         &,/file      &@,/file
             several  & f g h   &@ f g h        &, f g h     &@, f g h

  Top        lc                 &.file          &;file       &@;file
  level      any      file      &./file         &;/file      &@;/file
             several  f g h     &. f g h        &; f g h     &@; f g h
             .mk.in   file      $(abs)/file     $(src)/file  $(abs_src)/file

Substitution syntax
-------------------

In general & expands to the subdirectory name when used for a
filename, and to the subdirectory name with / replaced with _ for
variable names.

Note that & is processed *even in makefile comments*.  The substitutor
does not understand make syntax, or shell syntax, at all.  However,
the substitution rules are chosen to work well with constructs which
are common in makefiles.

In the notation below, we suppose that the substitution is being in
done in a subdirectory sub/dir of the source tree.  In the RH column
we describe the expansion at the top level, which is often a special
case (in general in variable names we call that TOP rather than the
empty string).

&CAPS           =>      sub_dir_CAPS                    or TOP_CAPS
&lc             =>      sub/dir/lc                      or lc
        Here CAPS is any ASCII letter A-Z and lc is a-z.
        The assumption is that filenames are usually lowercase and
        variables usually uppercase.  Otherwise, use another syntax:

&_              =>      sub_dir_                        or TOP_
&=_             =>      sub_dir                         or TOP

&/              =>      sub/dir/                        or nothing
&=/             =>      sub/dir                         or .

&,lc    => $(top_srcdir)/sub/dir/lc       &,/   => $(top_srcdir)/sub/dir/
&;lc    => $(top_srcdir)/lc               &;/   => $(top_srcdir)/

&@lc    => $(PWD)/sub/dir/lc              &@/   => $(PWD)/sub/dir/
&.lc    => $(PWD)/lc                      &./   => $(PWD)/
&@,lc   => $(abs_top_srcdir)/sub/dir/lc   &@,/  => $(abs_top_srcdir)/sub/dir/
&@;lc   => $(abs_top_srcdir)/lc           &@;/  => $(abs_top_srcdir)/

In general:
    =   return subdir without delimiter (not allowed with `,' `;' `@')
    ,   pathname of this subdirectory in source tree
    ;   pathname of top level of source tree
    .   pathname of this directory in build tree, implies absolute pathnames
    @   absolute pathnames (forbidden with `.', must come first)

So pathname syntax is a subset of:
    '&' [ '@' ] [ ',' | ';' | '.' ] [ lc | '/' ]

    To avoid incomprehensible .sd.mk files, some combinations are not
    allowed.  For example `&@=./' would mean `$(PWD)/sub/dir' but can
    be spelled `$(PWD)/&=/', but more normally the trailing / can be
    tolerated, so use `&@/'.

&&              =>      &&              for convenience in shell runes
\&              =>      &               general escaping mechanism

& thing thing... &      &@ thing thing... &
                        &. thing thing... &
&, thing thing... &     &@, thing thing... &
&; thing thing... &     &@; thing thing... &
        Convenience syntax for prefixing multiple filenames.
        Introduced by & followed by lwsp where lc could go.
        Each lwsp-separated non-ws word is prefixed by &/ etc.
        etc. respectively.  No other & escapes are recognised.
        This processing continues until & preceded by lwsp,
        or until EOL (the end of the line), or \ then EOL.

&:<directive> <args>....
        recognised at start of line only (possibly after lwsp)
        args are processed for &

&:include filename              filename should usually be foo.sd.mk
&:-include filename             tolerate nonexistent file
        filenames are relative to $(top_srcdir)

&!<lwsp>        disables & until EOL (and then disappears)

&#      delete everything to end of line
        (useful if the RHS contains unrecognised & constructions)

&!STUFF
        changes the escape sequence from & to literally STUFF
        STUFF may be any series of of non-whitespace characters,
        and is terminated by EOL or lwsp.  &!STUFF and the lwsp
        are discarded.

        After this, write STUFF instead of &, everywhere.
        The effect is global and lasts until the next setting.
        It takes effect on &:include'd files too, so maybe set
        it back before using &:include.

        Notably
                STUFFSTUFF      =>      STUFFSTUFF
                \STUFF          =>      STUFF
                STUFF!&         set escape back to &

&TARGETS_things
        Handled specially.  If mentioned, declares that this
        subdir ought to have a target `things'.  The rule will be
                &/things:: $(&TARGETS_things)

        You may extend it by adding more :: rules for the target,
        but the preferred style is to do things like this:
                &TARGETS_check += & test-passed.stamp

        It is important to mention &TARGETS_things at least once in
        the context of each applicable directory, because doing so
        arranges that the *parent* will also have a `things' target
        which recursively implies this directory's `things'.

        Must be spelled exactly &TARGETS_things.  &_TARGETS_things,
        for example, is not magic.  But mentioning &TARGETS_things in
        a #-comment *does* work because the & filter does not care
        about comments.

        `all' is extra special: every directory has an `all'
        target, which corresponds to &TARGETS.

Subdirectory and variable naming
--------------------------------

The simple variable decoration scheme does not enforce a strict
namespace distinction between parts of variable names which come from
subdirectory names, and parts that mean something else.

So it is a good idea to be a bit careful with your directory naming.
`TOP', names that contain `_', and names that are similar to parts of
make variables (whether conventional ones, or ones used in your
project) are best avoided.

If you name your variables in ALL CAPS and your subdirectories in
lower case with `-' rather than `_', there will be no confusion.

Incorporating this into your project
------------------------------------

Use `git-subtree' to merge the subdirmk/ directory.  You may find it
useful to symlink the DEVELOPER-CERTIFICATE file (git can store
symlinks as symlinks - just `git add' the link).  And you probably
want to mention the situation in your top-level COPYING.

Symlink autogen.sh into your project toplevel.

In your configure.ac, say

  m4_include([subdirmk/subdirmk.ac])
  SUBDIRMK_SUBDIRS([...list of subdirectories in relative syntax...])

Write a Subdir.sd.mk in each directory.  The toplevel one should
probably contain:

  include subdirmk/usual.mk
  include subdirmk/regen.mk

Write a Perdir.sd.mk in the toplevel, if you want.  It should probably
have:

  &:include subdirmk/cdeps.sd.mk
  &:include subdirmk/clean.sd.mk

Hints
-----

You can convert your project incrementally.  Start with the top-level
Makefile.in and rename it to Subdir.sd.mk, and add the appropriate
stuff to configure.ac, and fix everything up.  Leave the existing
$(MAKE) -C for your existing subdirectories alone.  Then you can
convert individual subdirectories, or classes of subdirectories, at
your leisure.  (You must be /sure/ that each subdirectory will be
entered only once at a time, but your existing recursive make descent
system should already do that or you already have concurrency bugs.)

Aside from this, be very wary of any invocation of $(MAKE) anywhere.
This is a frequent source of concurrency bugs in recursive make build
systems.  When combined with nonrecursive make it's all in the same
directory and there is nothing stopping the different invocations
ending up trying to make the same targets at the same time. That
causes hideous racy lossage.  There are ways to get this to work
reliably but it is advanced stuff.

If you make syntax errors, or certain kinds of other errors, in your
makefiles, you may find that just `make' is broken now and cannot get
far enough to regenerate a working set of makefiles.  If this happens
just rerun ./config.status by hand.


Legal information
-----------------

subdirmk is
 Copyright 2019 Mark Wooding
 Copyright 2019 Ian Jackson

    subdirmk and its example is free software; you can redistribute it
    and/or modify it under the terms of the GNU Library General Public
    License as published by the Free Software Foundation; either
    version 2 of the License, or (at your option) any later version.

    This is distributed in the hope that it will be useful, but
    WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
    Library General Public License for more details.

    You should have received a copy of the GNU Library General Public
    License along with this library as the file LGPL-2.
    If not, see https://www.gnu.org/.

Individual files generally contain the following tag in the copyright
notice, instead of the full licence grant text:
  SPDX-License-Identifier: LGPL-2.0-or-later
As is conventional, this should be read as a licence grant.

Contributions are accepted based on the git commit Signed-off-by
convention, by which the contributors' certify their contributions
according to the Developer Certificate of Origin version 1.1 - see
the file DEVELOPER-CERTIFICATE.

Where subdirmk is used by and incorporated into another project (eg
via git subtree), the directory subdirmk/ is under GNU LGPL-2.0+, and
the rest of the project are under that other project's licence(s).
(The project's overall licence must be compatible with LGPL-2.0+.)