2001-06-29 Raja R Harinath <harinath@cs.umn.edu>

[automake.git] / TODO

the new YFLAGS code doesn't correctly handle srcdir

allow foo_NAME to rename an object (library or program)
at build/install time

remove _LTLIBRARIES and just use _LIBRARIES
then use this for zip/jar as well

consider `sub am_unique_error', which only prints a given error
message once.  Use eg in lang_c_rewrite for ansi2knr error

for 1.5
investigate problems with conditionally defined libraries

add an error if the user makefile.am violates our
   namespace rules

have 'make check' print tests which are skipped

we need a document describing automake from the end user's point of view
eg describe INSTALL_HEADER there, among other things

* maintainer-clean

Akim:
> @@ -31,5 +31,9 @@
>  DISTCLEAN    -test -z "$(DISTCLEANFILES)" || rm -f $(DISTCLEANFILES)
>
>  maintainer-clean-generic:
> +## FIXME: shouldn't we really print these messages before running
> +## the dependencies?
> +     @echo "This command is intended for maintainers to use"
> +     @echo "it deletes files that may require special tools to rebuild."
>         -rm -f Makefile.in

Tom:
> I'd like to eventually fix the FIXME comment by having
> maintainer-clean look like:
>
>     maintainer-clean:
>         @echo ...
>         $(MAKE) whatever
>
> We're left with the question of whether we should repeat them in every
> subdir.

*
Alexandre Oliva:
> Hmm...  Interesting.  It must have been a side effect of the enabling
> of forced `relink' on GNU/Linux/x86.  Anyway, on platforms that
> actually require relinking, this problem remains, and I see no way to
> overcome it other than arranging for automake to install libraries
> before executables, as you suggest.  This shouldn't be a big problem,
> anyway.
>
> A bigger problem could show up if two libraries in the same directory,
> one dependent on the other, are installed concurrently.  If relinking
> is needed for the dependent library, we have a problem.  It appears to
> me that user will have to live without `make -j install', in this
> case.

Alex Hornby
> Here's an Automake patch and changelog entry allow make -j install on
> such degenerate systems (and Linux with buggy libtool <g>)
>
> If you install to locations other that bin_ and lib_ then a larger fix
> is necessary, but this should fix the 90% case.

* think about how per-object flags should work.  in particular:
  * how should they be specified?
    using the object name is confusing when .lo/.obj in use
    however, the object name provides a nice interaction with
    per-exe flags
  * how should they interact with per-executable flags?
  [ this is probably a feature in search of a problem ]

* cross-compilation support:
  programs built and used by the build process need to be
  built for CC_FOR_BUILD
  introduce a new prefxi for this, e.g. `build_PROGRAMS'
  [ we can do this in an automatic way I think.
    unfortunately it isn't that useful until autoconf has support
    for this sort of thing as well ]

* distcheck should make sure that each file that uses _() is
  listed in POTFILES.in
  From Jim Meyering:
    # Verify that all source files using _() are listed in po/POTFILES.in.
    po-check:
            grep -E -v '^(#|$$)' po/POTFILES.in | sort > $@-1
            grep -E -l '\b_\(' lib/*.c src/*.c | sort > $@-2
            diff -u $@-1 $@-2
            rm -f $@-1 $@-2

* one performance enhancement would be to have autoconf write
  a single file containing all the macro assignments.
  then read this file via `include'
  unfortunately this can't be done because of conditionals

* support prog_LIBS as override for LIBS

* Scan configure.in using traces as autoheader does.
  This will be much more reliable.

* Test subdir-objects option with yacc, lex, ansi2knr
  Our locking scheme won't prevent a parallel make from losing
  if there are two `bar.o' files and the timing is just right
  This only happens with parallel make and no-`-c -o' compiler,
  so it probably isn't very important
  `-c -o' when doing libtool
  try to find a losing compiler and see if it really works.
  (actually: hack config.cache and do it)

* per-exe flags
** We're using `$<' in explicit rules when using per-exe flags
** per-exe flags don't work for CPPFLAGS/YFLAGS/LFLAGS.  Fix.
** LIBOBJS shouldn't be used when there are per-exe flags (?)

* Test nodist_SOURCES with lex, yacc, etc.

* Support subdir-objects with fortran

* Allow creation of Java .zip/.jar files in natural way
  If you are building a compiled Java library, then the .zip/.jar
  ought to be made automatically.

* Run automake before libtool.  It will report an error but
  still won't put the file into the disty.  This is wrong.
  From Mark H Wilkinson <mhw@kremvax.demon.co.uk>

* CFLAGS only defined if C source seen
  but really it should be a configure variable, shouldn't it?
  There are other examples of this
  [ moving to autoconf --trace ought to fix this ]

* in gnu/gnits mode, give error if Makefile.am overrides a user
  variable like CFLAGS.
  [ this is low priority because the package author can always
    circumvent our check by redefining in configure.in
    plus it is probably better to encourage good behavior than to
    punish bad ]

* If we see `foo.o' in LIBOBJS, and we've seen AC_OBJEXT, then complain.
  [ how will we know that?  it is better to handle this automatically
    via an autoconf hook ]

* examine possibility of using any character in a macro name
  and rewriting names automatically.  this means we must rewrite
  all references as well.
  [ this is a 2.0-style feature ]

* AM_CONFIG_HEADER might generate the wrong stamp file names
  when given multiple headers.  Write a test.

* Currently don't correctly handle multiple inputs to a config header.
  [ this should no matter in the future as acconfig.h and so on are
    obsoleted by the AH series of macros.]

* header stamp files still in wrong dirs.
  stamp-h.in must be in dir with h.in file
  stamp-h must be in dir with output file

* conditionals and macros
  Our current scheme cause combinatoric explosion.

  In fact, to be honest, I no longer understand very well why we perform
  such a closure.  I mean, as is, Automake transforms (this is
  cond3.test)

    | bin_PROGRAMS = targ
    |
    | if ONE
    | SONE = one.c
    | else
    | SONE =
    | endif
    |
    | if TWO
    | STWO = two.c
    | else
    | STWO =
    | endif
    |
    | if THREE
    | STHREE = three.c
    | else
    | STHREE =
    | endif
    |
    | targ_SOURCES = $(SONE) $(STWO) $(STHREE)

  into

    | @ONE_FALSE@@THREE_FALSE@@TWO_TRUE@am_targ_OBJECTS = two.$(OBJEXT)
    | @ONE_FALSE@@THREE_FALSE@@TWO_FALSE@am_targ_OBJECTS =
    | @ONE_FALSE@@THREE_TRUE@@TWO_TRUE@am_targ_OBJECTS = two.$(OBJEXT) \
    | @ONE_FALSE@@THREE_TRUE@@TWO_TRUE@ three.$(OBJEXT)
    | @ONE_FALSE@@THREE_TRUE@@TWO_FALSE@am_targ_OBJECTS = three.$(OBJEXT)
    | @ONE_TRUE@@THREE_FALSE@@TWO_TRUE@am_targ_OBJECTS = one.$(OBJEXT) \
    | @ONE_TRUE@@THREE_FALSE@@TWO_TRUE@ two.$(OBJEXT)
    | @ONE_TRUE@@THREE_FALSE@@TWO_FALSE@am_targ_OBJECTS = one.$(OBJEXT)
    | @ONE_TRUE@@THREE_TRUE@@TWO_TRUE@am_targ_OBJECTS = one.$(OBJEXT) \
    | @ONE_TRUE@@THREE_TRUE@@TWO_TRUE@ two.$(OBJEXT) three.$(OBJEXT)
    | @ONE_TRUE@@THREE_TRUE@@TWO_FALSE@am_targ_OBJECTS = one.$(OBJEXT) \
    | @ONE_TRUE@@THREE_TRUE@@TWO_FALSE@ three.$(OBJEXT)

  why don't we just output

    | @ONE_TRUE@am_SONE_OBJECTS = one.$(OBJEXT)
    | @ONE_FALSE@am_SONE_OBJECTS =
    |
    | @TWO_TRUE@am_STWO_OBJECTS = two.$(OBJEXT)
    | @TWO_FALSE@am_STWO_OBJECTS =
    |
    | @THREE_TRUE@am_STHREE_OBJECTS = three.$(OBJEXT)
    | @THREE_FALSE@am_STHREE_OBJECTS =
    |
    | am_targ_OBJECTS = $(am_SONE_OBJECTS) $(am_STWO_OBJECTS) $(am_STHREE_OBJECTS)

  which means also, why do we look for the closure of PROGRAMS, instead
  of just adding $(EXEEXT) to all its components and sub components
  (i.e., inside sub vars such as $(SONE) above being a sub var of
  targ_SOURCES)?


  Aaaaaaaaaaah!  I think I know... Must be because of `+='.

  Hm... No.  Indeed we transform

    | FOO = foo
    | if BAR
    | FOO += BAR
    | endif

  into

    | @BAR_TRUE@FOO = foo bar
    | @BAR_FALSE@FOO = foo

  but this seems good to me too?

    | FOO = foo $(BAR_FOO)
    | @BAR_TRUE@BAR_FOO = bar
    | @BAR_FALSE@BAR_FOO =


* foo=bar
  if cond
  foo += joe
  endif
  ... this ought to work.  The fix is probably complicated, but might
  come for free when we rewrite the handling of conditionals.

* `distcheck' and `dist' should depend on `all'

* Add code to generate foo-config script like gnome, gtk

* `DEFS += foo' won't work.
  That's because DEFS is defined in header-vars.am, which is read
  after the user's Makefile.am.
  This will be a problem for any macro defined internally
    [ fixing this will probably fix the nasty `exeext redefines
      foo_PROGRAMS' hack that is in there right now ]
    [ we currently give an error when this occurs, so this is very low
      priority ]

* document user namespace for macro/target names
  adopt some conventions and use uniformly
    [ this is a good thing for the rewrite ]

* make distcheck uses directories like `=build'.
  Some (very rare) POSIX systems don't support `=' in filenames.
  If this ever becomes a problem, fix it

* distclean must remove config.status
  can't this cause problems for maintainer-clean?
  shouldn't maintainer-clean print the message before running
  any part of the make?  (just to slow things down long enough
  for the user to stop it)
  (maybe doesn't matter since people who even know about
  maintainer-clean already have a clue)

* reintroduce AM_FUNC_FNMATCH which sets LIBOBJS
  Then have automake know about fnmatch.h.
    [ probably should wait for autoconf to get right functionality ]

* "make diff" capability
  look at gcc's Makefile.in to see what to do
  or look at maint program

* in --cygnus, clean-info not generated at top level

* what if an element of a scanned variable looks like
        $(FOO).$(BAR)  ?
  or some other arbitrary thing?
  right now we try to cope, but not very well
    [ this is only of theoretical interest for now ]

* make sure every variable that is used is also defined
    [ we don't really look at variable uses in detail.
      2.0 thing ]

* make sure `missing' defines are generated

* missing should handle install -d and rmdir -p (for uninstall)

* notice when a .c file is a target somewhere, and auto-add it to
    BUILT_SOURCES

* NORMAL_INSTALL / NORMAL_UNINSTALL -vs- recursive rules
  [ requires changes to the standard ]

* copyrights on m4 files, aclocal output

* should not put texiname_TEXINFOS into distribution
  should rename this macro anyway, to foo_texi_DEPENDENCIES

* *all* installed scripts should support --version, --help

* For now I guess I'll just have automake give an error if it encounters
non-C source in a libtool library specification.

* must split $obj into two parts: one for libtool and one for
  deansification.  Otherwise .S files will be deansified!

* ansi2knr must currently appear in a directory that has some source

* if program has the same name as a target, do something sensible:
  - if the target is internal, rename it
  - if the target is mandated (eg, "info"), tell the user
    consider auto-modifying the program name to work around this

* should separate actual options from strictness levels
  strictness should only cover requirements
  You should be able to pick and choose options

* rewrite in guile (RMS request)
at the same time, consider adding a GUI
could use the same parsing code for the GUI and the standalone version
that means figuring out a better representation of internal state
[ that's easy -- anything is better than what we have now ]

having just one Makefile for a project would give a big speed increase
for a project with many directories, eg glibc.  ideally (?) you'd
still be able to have a Makefile.am in each directory somehow; this
might make editing conceptually easier.

* finish up TAGS work

* only remove libtool at top level?

* clean up source directory by moving stuff into subdirs

* consider adding pkglibexecdir, maybe others?
  requests for pkg-dirs with version included

Avoid loops when installing; instead unroll them in automake

* for new autoconf:
  * completely handle multi-":" mode for AC_CONFIG_HEADER
  * Scan multiple input files when Makefile is generated?
    This would provide flexibility for large projects; subsumes
    the "Makefile.tmpl" idea

   [ can't do this.  must explain why in manual.
     basically, solving all the problems is too hard
     like: how to remove redundancies between generated .in files
     instead should implement `include' directive for Makefile.am ]
* for multi-":" mode and AC_OUTPUT, it might be good to pick the
  first input file that has a corresponding .am file.

Some long-term projects:
* if $(FOO) is used somewhere, ensure FOO is defined, either by
  user or by automake if possible

[ include, += support ]
* even better would be allowing targets in different included
  fragments to be merged.  e.g., `install-local'.

consider putting all check-* targets onto @check?
To support --help/--version checking?

take diff-n-query code from libit

Per Bothner says:
Per> 1) Being able to build a set of non-source programs
Per> from source programs, without necessarily linking them together.
Per> I.e. one should be able to say something like:
Per>    dummy_SOURCES=foo.c bar.c
Per> and automake should realize that it needs to build foo.o and bar.o.
Per> 2) Being intelligent about new kinds of suffixes.
Per> If it sees:
Per>    SUFFIXES = .class .java
Per> and a suffix rule of the form:
Per>    .java.class:
Per> then it should be able to realize it can build .class files from
Per> .java files, and thus be able to generate a list of
Per> .class files from a list of .java source files.

!! Must fix require_file stuff.  It is really gross, and I don't
   understand it any more.

Jim's idea: should look for @setfilename and warn if filenames too long
* guess split size

** many requests for a way to omit a file from the distribution.
   Should be done like `!foo' or `~foo' in _SOURCES, etc.
   Such files should be removed explicitly after the copy step!
   Doing this requires rewriting macros before generating Makefile.in.

from joerg-martin schwarz:
 -- If Makefile.am contains $(CC), $(COMPILE), $(YLWRAP), ....
    in an explicitly written rule,  you should emit the corresponding
    Makefile variables automatically.

Configuring in the large:
* allow hierarchy of dirs to share one aclocal.m4
  How?

consider printing full file name of Makefile.am or configure.in when
giving error.  This would help for very large trees with many
configure.in scripts

From the GNU Standards.  These things could be checked, and probably
should be if --gnu.
*    Make sure that the directory into which the distribution unpacks (as
well as any subdirectories) are all world-writable (octal mode 777).
*   Make sure that no file name in the distribution is more than 14
characters long.
*    Don't include any symbolic links in the distribution itself.
     (ditto hard links)
*    Make sure that all the files in the distribution are world-readable.
** also, check --help output and --version output.  Idea from François
* standards no longer prohibit ANSI C.  What does this imply
  for the de-ansi-fication feature? [ must keep it -- some users rely on it ]

should be able to determine what is built by looking at rules (and
configure.in).  Then built man pages (eg) could automatically be
omitted from the distribution.

Henrik Frystyk Nielsen says:
Henrik> 4) Flags like --include-deps are lost when you make changes to
Henrik> Makefile.am files and automake is run automatically. It would
Henrik> be nice to keep these flags as I now have to redo everything
Henrik> manually.
... what about other options here too?

Think about: maybe "make check" should just bomb if error occurs?
Then user must use "make -k check".  This is probably more natural.

Consider: "cvs" option adds some cvs-specific rules?

Right now, targets generated internally (eg "install") are not
overridable by user code.  This should probably be possible, even
though it isn't very important.  This could be done by generating all
internal rules via a function call instead of just appending to
$output_rules.
 [ this will be harder to implement when scanning a rule like all-recursive
   from subdirs.am ]

* Should be a way to have "nobuild_PROGRAMS" which aren't even built,
  but which could be by running the magic make command.

Other priorities:
* Must rewrite am_install_var.  Should break into multiple functions.
  This will allow the callers to be a little smarter.
* Rewrite clean targets.
* Must rewrite error handling code.  Right now it is a real mess
  Should fix up require_file junk at the same time

djm wants ``LINKS'' variable; list of things to link together after
install.  In BSD environment, use:
        LINKS = from1 to1 from2 to2 ...

Need way to say there are no suffixes in a Makefile (Franc,ois'
"override" idea suffices here)

Check to make sure various scripts are executable (IE when looking for
them in a directory)

Handle dist-zoo.  Generally add more DOS support.  Maybe run "doschk"
(why isn't this merged with "pathchk"?) when doing a dist.  Do
whatever else François says here...

Add support for html via an option.  Use texi2html.  Use
"html_TEXINFOS", and htmldir = .../html.  Include html files in
distribution.  Also allow "html_DATA", for raw .html files.
  [ when will texinfo directly support html? ]

uninstall and pkg-dirs should rm -rf the dir.

a potential bug: configure puts "blah.o" into LIBOBJS, thus implying
these files can't be de-ansified.  Not a problem?
  [ fix by using ansi2knr wrapper program ]

In general most .am files should be merged into automake.  For
instance all the "clean" targets could be merged by keeping lists of
things to be removed.  This would be a lot nicer looking.  Note that
the install targets probably should not be merged; it is sometimes
useful to only install a small part.

* Clean up the output:
** Order rules sensibly
** Ensure every line has a purpose.  Omit unused stuff
** Eliminate extraneous rules when possible (eg 'install-am' stuff)
** Make sure vertical spacing is correct
Omit program transform vars from header if no program installed.  This
is currently pretty hard to do.  (But with beautification code it
would probably be easy)

* Lex, yacc support:
** It would be nice to automatically support using bison's better features
  to rename the output files.  This requires autoconf support
** Consider supporting syntax from autoconf "derived:source", eg:
        y.tab.c:perly.y
  for yacc and lex source
** what if you use flex and the option to avoid -lfl?
  should support this?

* Multi-language support:
** should have mapping of file extensions to languages
** should automatically handle the linking issue (special-case C++)
** must get compile rules for various languages; FORTRAN probably
  most important unimplemented language
This should be integrated in some way with Per's idea.
Eg .f.o rules should be recognized & auto-handled in _SOURCES
That way any random language can be treated with C/C++ on a first-class
basis (maybe)

It might be cool to generate .texi dependencies by grepping for
@include.  (If done, it should be done the same way C dependencies are
done)

It would be good to check some parts of GNU standards.  Already check
for install-sh and mkinstalldirs.  What else is required to be in
package by GNU standards or by automake?
Some things for --strictness=gnits:
* "cd $(foo); something" is an error in a rule.  Should be:
  "cd $(foo) && something"
* Look for 'ln -s' and warn about using $(LN) and AC_PROG_LN_S
* Look for $(LN) and require AC_PROG_LN_S

Auto-distribute "ChangeLog.[0-9]+"?  "ChangeLog.[a-z]+"?

Internationalize. [ gettext doesn't have the necessary machinery yet ]
am_error should use printf-style arguments (for eventual gettext scheme)

François says the ordering of files in a distribution should be as follows:
* README
* source files
* derived files
I agree, but I don't see how to implement this yet.
It might be easier if "derived files" is limited to those that
Automake itself knows about, eg output of yacc.

Check all source files to make sure that FSF address is up-to-date.
--gnits or --gnu only.

Merge each -vars.am file with corresponding ".am" file.  Can do this
because of changes to &file_contents.

Should libexec programs have the name transform done on them?

Order the output rules sensibly, so FOO_SOURCES and FOO_OBJECTS are
together and rules are in the usual order.

Make the output minimal: only output definitions for variables that
are used.

djm says:
David> To avoid comments like the one about subdirs getting buried in
David> the middle of a Makefile.in, how about pushing comments that
David> start with ### to the top of the Makefile.in (in order)?  Sort
David> of like how Autoconf uses diversions to force initialization
David> code to the top of configure.

Karl Berry says:
Karl> 2) Your Makefile variable names are generally uppercase, but GNU
Karl> generally uses lowercase. Not that it matters :-).

================================================================

Stuff for aclocal:

probably should put each group of m4 files into a subdir owned by the
containing application.

================================================================

Document:

AM_MISSING_PROG

how to use the generated makefiles
 - standard targets
 - required targets
 - NORMAL_INSTALL junk

what goes in AC_CONFIG_AUX_DIR

multi-":" mode in AC_OUTPUT -- automake only looks at the first file
    also a note on how a .am file is found in this case

rationale for avoiding
        make CFLAGS="$CFLAGS" ...
in subdirs make rule

a package that installs its own aclocal macros

write example of using automake with dejagnu
follow calc example in dejagnu docs

document which variables are actually scanned and which are not.

Document customary ordering of Makefile.am.  From François.

Should include extended version of diagram from Autoconf (suggested by
Greg Woods)

Make a definition of the term "source"

document how to use Automake with CVS.  Idea from Mark Galassi.  Also
include Greg Woods' more sophisticated "cvs-dist" target.

document rebuilding configure.  CONFIGURE_DEPENDENCIES
CONFIG_STATUS_DEPENDENCIES

-- must document all variables that are supposed
   to be public knowledge

must document the targets required for integration with
non-automake-using subdirs

document the "make SHELL='/bin/sh -x'" trick for debugging

section on relationship to GNU make.  include notes on parallel makes

add a concept index

move discussion of cygwin32, etags, mkid under other gnu tools

CCLD, CXXLD, FLD

================================================================

Things to do for gcc:

Regularize dependency generation.  Add new flags:

-MH   Generate a dummy dependency for each header file mentioned.
-MT NAME
      Set name of target
-MF NAME
      Set name of output file

Then automake can use -MD -MH -MT 'foo.o foo.lo' -MF .deps/...

================================================================

Libraries:

* Should support standalone library along with subdir library in same
  Makefile.am.  Maybe: turn off "standalone" mode if library's Makefile.am
  is not only one specd? [ add an option for this ]

================================================================

Longer term:

Would it be useful to integrate in some way with the Debian package
building utility?  Must check.  maybe it would be possible to deal
with all the different package utilities somehow.  Lately I've been
hearing good things about the RedHat packaging utilities.  Why are
there so many of these?  Are they fun to write or something?
The RedHat package utility is called RPM; see
        ftp://ftp.redhat.com/pub/code/rpm
It actually has problems, like no configure script and no documentation.

For Cygnus it would probably be good to be able to handle the native
package utility on each platform.  There are probably 3 or 4 of these
(sysv, solaris?, aix?)

tcl/unix/Makefile.in has some code to generate a Solaris package.

Automake probably can't do all of this on its own.  A new tool might
be a better idea

I have some notes from a Debian developer on how the integration
should work

================================================================

A tool to guess what the local Makefile.am should look like:
(see Gord's Maint program!)

* Probably integrate with autoscan
* Use various simple rules to determine what to do:
  * get name of top directory, sans version info
  * search for .c files with 'main' in them
    * if in main.c, use directory name for program
    * if in more than one, generate multiple programs
    * if not found, generate a library named after directory
  * order subdir searches correctly: lib first, src last
  * assume 'testsuite' dir means we are using dejagnu
* maybe be smart about reading existing Makefile.am, so tool
  can be run for incremental changes?  You could imagine:

        Makefile.am:
                autoproject --incremental

================================================================

Stuff NOT to do, and why:

consider auto-including any file that matches "*.in".
  [ no: po/Makefile.in shouldn't be included ]

must look at mkid to see how it works (for subdir usage)
  [ right now, it doesn't.  i don't see a simple fix right now ]

if configure.in not found, move up a directory and try again?  This
could eliminate a common source of problems.
  [ this is just a bad idea ]

* scripts are installed in $exec_prefix/bin, not $prefix/bin
  Bug or feature?
  [ the consensus on Gnits is that this isn't required.
    doubters can work around it anyway ]

Scan source directories and warn about missing files, eg .c/.h files
that aren't mentioned?
  [ distcheck makes this less useful ]

* quoting bugs
  - how to install file with a space in its name?
  [ don't bother with this -- make is just too losing ]

Local Variables:
mode: outline
End: