CMake Fortran Issues: Difference between revisions

From KitwarePublic
Jump to navigationJump to search
Line 146: Line 146:
This is now serialized to a file at the build directory (at the moment I'm using  
This is now serialized to a file at the build directory (at the moment I'm using  
boost.serialization. Are there serialization capabilities in CMake?)
boost.serialization. Are there serialization capabilities in CMake?)
: '''Brad''': Instead of serialization we usually encode the information in a .cmake file that is later loaded with ReadListFile.


== -E cmake_depends ==
== -E cmake_depends ==

Revision as of 16:46, 18 December 2007

Introduction to this wikipage

CMake has a number of Fortran issues that have been discussed many different times on list and duplicated a fair number of times in the bug tracker as well.

Maik Beckmann is trying to make sense of all the confusion by collecting information on all Fortran issues at http://www.cmake.org/Bug/view.php?id=5809

Please join the work there by

  • Contributing patches.
  • Testing the patches that already exist there.
  • Reporting things that don't work.
  • Sending simplified examples of things which don't work.
  • Sharing your expert knowledge of CMake.

Introduction to the problem

Hello CMake developers,

I pushed myself during the last weekends to get more familiar with CMakes codebase. Not for fun only ;), but make me smart enough to sketch an approach for handling fortrans module dependencies.

I will try to write down what I'm thinking about

To give you (and me) some orientation here are the steps which are done by cmake so far and what I'm going to plan.

CMakes atom of dependency is a sourcefile. It knows the language of the sourcefile and handles it by the related class

cmDepends**

When CMake processes foo/.c/.cxx/.java/.f/.f90 it doesn't know anything about the CONTENT! of any other sourcefile. This is ok for include dependencies, since the included file is almost somewhere at the disk.

Anyway, this is not sufficient for fortran. If a fortran source foo.f90 contains something like

 module bar
   ...
 end module

the compiler will generate a file called foo.mod. A .mod file can be considered as a __dynamically generated header file__. Every fortran source can potentially create a .mod file! So if cmake processes a fortran sourcefile which needs bar.mod it know that foo.f90 provides this module.

Abstract solution

The Plan is to scan all fortran source in the sourcetree before the per file dependency generation is done. The extracted information are serialized. When cmake processes a fortran source it know if a required module is provided by

- a source of the current target
- a source of another target in the source tree
- if none of both it must be part of an prebuild library

Concrete solution

Current CMake does: After cmake ran once of the source tree (i.e cmake -G"Unix Makefiles" /path/to/sourcetree)

1. check build system
...
k    : $(MAKE) -f  ../foo.dir/build.make ../foo.dir/depend
k+1: $(MAKE) -f ../foo.dir/build.make  ../foo.dir/requires
k+2: $(MAKE) -f ../foo.dir/build.make  ../foo.dir/build
...
l    : $(MAKE) -f  ../bar.dir/build.make ../bar.dir/depend
l+1: $(MAKE) -f ../bar.dir/build.make  ../bar.dir/requires
l+1: $(MAKE) -f ../bar.dir/build.make  ../bar.dir/build

It would be cool if it will do:

1  : check build system
...
# maybe only for targets which have fortran sources
j    : $(MAKE) -f    ../foo.dir/build.make ../foo.dir/fortran_module_scan
j+1: $(MAKE) -f    ../bar.dir/build.make ../bar.dir/fortran_module_scan
...
k    : $(MAKE) -f  ../foo.dir/build.make ../foo.dir/depend
k+1: $(MAKE) -f ../foo.dir/build.make  ../foo.dir/build
...
l    : $(MAKE) -f  ../bar.dir/build.make ../bar.dir/depend
l+1: $(MAKE) -f ../bar.dir/build.make  ../bar.dir/build

.dir/fortran_module_scan will trigger a new cmake command

$(CMAKE_COMMAND) -E fortran_module_scan ...

while .dir/depend triggers still tiggers

$(CMAKE_COMMAND) -E cmake_depends ...

-E fortran_module_scan

This is a new cmake command.

The information which is extracted from a source "foo.f90" are

- required modules 
- provided modules
- includes

These three vectors and the name "foo.f90" are stored in a data structure called "FortranSourceDependInfo".

struct cmFortranSourceDependInfo
{
  std::string Name;
  std::vector<std::string> Requires;
  std::vector<std::string> Provides;
  std::vector<std::string> Includes;
  // std::set needs this operator
  bool operator<(cmFortranSourceDependInfo const& rhs) const
   {  this->Name < rhs.Name; }
};
This happens for every source of a Target i.e. "mylib".  Another data 

structure "FortranTargetDependInfo" holds the name "mylib" and a set of "FortranSourceDependInfo" instances.

struct cmFortranTargetDependInfo
{
  std::string Name;
  std::set<cmFortranSourceDependInfo> Sources;
  void GetModuleInfoOfSource(char const* src,  cmFortranSourceDependInfo&)
    const;
  // does this target source provide module mod?
  bool Provides(char const* mod) const;
};

This "cmFortranTargetDependInfo" instance is finally stored into

class cmFortranDependInfos
{
public:
  typedef std::map<std::string, cmFortranTargetDependInfo > targets_type;

  cmFortranDependInfos(char const* homeDir);
  virtual ~cmFortranDependInfos();

  void InsertSourceDependInfo(char const* targetName,
    cmFortranSourceDependInfo const&);

  void GetModuleInfoOfTarget(char const* targetName,
    cmFortranTargetDependInfo&) const;

  void Serialize();
  void Load(char const* homeDir);

private:
  static char const* ArchiveName;
  targets_type Targets;
  std::string Archive;
};

This is now serialized to a file at the build directory (at the moment I'm using boost.serialization. Are there serialization capabilities in CMake?)

Brad: Instead of serialization we usually encode the information in a .cmake file that is later loaded with ReadListFile.

-E cmake_depends

One of the very first things which happens when -E cmake_depends is executed, is the construction of an "class cmake" instance, which itself constructs a cmGlobalGenerator. cmake#CreateGlobalGenerator(..) looks for an the serialized cmFortranDependInfos instance and loads it. Now the per sourcefile dependency generation scheme of cmake knows enough to do the right thing for fortran modules.

OUTDATED: Concepts expressed using Makefiles

This section is intended to discuss the Makefile rules which CMake has to generate. All examples are fully working. You can download them as tarball examples_using_Makefiles.tar.gz at http://www.cmake.org/Bug/view.php?id=5809. To build an example, change into the corresponding build directory and run the

$ make

command. After this initial build, check dependencies by touching source files of your choice and running the

$ make

command again.

Note: For examples which show how an external library providing modules is handled, the external library which resides at directory extLib for each of these examples has to be built and installed by changing into the corresponding extLib directory and running the

$ make install && make clean

command.

A simple program

A f9x program which is build by compiling in linking two source files a.f90 and main.f90. The tree structure is:

  • example_simpleProgram
    • build
      • Makefile
      • prog.dir
        • build.make
    • main.f90
    • a.f90

a.f90:

SUBROUTINE printHello
    WRITE(*,*) "Hello f9x world"	
END SUBROUTINE

main.f90:

PROGRAM hello 
    CALL printHello
END PROGRAM

Makefile:

all: prog.dir/all

prog.dir/all:
	$(MAKE) -f prog.dir/build.make prog.dir/all

clean:
	$(MAKE) -f prog.dir/build.make prog.dir/clean

build.make:

prog.dir/all: prog.dir/prog

prog.dir/prog:  prog.dir/a.o prog.dir/main.o
	gfortran -o prog.dir/prog  prog.dir/a.o prog.dir/main.o
	
prog.dir/a.o: ../a.f90
	gfortran -o prog.dir/a.o  -c ../a.f90
	
prog.dir/main.o: ../main.f90 
	gfortran -o prog.dir/main.o  -c ../main.f90 
	
prog.dir/clean:
	rm prog.dir/a.o prog.dir/main.o prog.dir/prog

The rules generated by the current CMake covers all dependencies which can occur as long as no modules are used.

You can download this example as tarball example_simpleProgram.tar.gz at http://www.cmake.org/Bug/view.php?id=5809

A simple program with module

The same as before, but now a.f90 provides a module which main.f90 uses. The tree structure is:

  • example_simpleProgram_withModule
    • build
      • Makefile
      • prog.dir
        • build.make
    • main.f90
    • a.f90

a.f90:

MODULE localMod 
!
CONTAINS
    SUBROUTINE printHello
        WRITE(*,*) "Hello f9x world"
    END SUBROUTINE
END MODULE

main.f90:

PROGRAM hello 
    USE localMod
    CALL printHello
END PROGRAM

Rules like those generated by current CMake

Makefile:

all: prog.dir/all

prog.dir/all:
	$(MAKE) -f prog.dir/build.make prog.dir/requires
	$(MAKE) -f prog.dir/build.make prog.dir/all

clean:
	$(MAKE) -f prog.dir/build.make prog.dir/clean

build.make:

prog.dir/all: prog.dir/prog


prog.dir/prog:  prog.dir/a.o prog.dir/main.o
	gfortran -o prog.dir/prog  prog.dir/a.o prog.dir/main.o 

prog.dir/a.o: ../a.f90
	gfortran -o prog.dir/a.o  -c ../a.f90 -M prog.dir
	
prog.dir/localmod.mod: prog.dir/a.o	
	
prog.dir/main.o: ../main.f90 
	gfortran -o prog.dir/main.o  -c ../main.f90 -I prog.dir 
	
prog.dir/clean:
	rm prog.dir/localmod.mod prog.dir/a.o prog.dir/main.o prog.dir/prog
	
	
localmod.mod.proxy: prog.dir/a.o

prog.dir/main.o.requires: localmod.mod.proxy
	
prog.dir/requires: prog.dir/main.o.requires

After you build prog using this set of Makefiles do (you're at the build directory)

 $ touch ../a.f90

and enter

 $ make

You'll see that a.f90 is recompiled and prog.dir/prog is linked again. But main.f90 has to recompiled too, since a module dependency is a compile time dependency like an include.

Rules like those that should be generated by CMake

Makefile:

all: prog.dir/all

prog.dir/all:
	$(MAKE) -f prog.dir/build.make prog.dir/all

clean:
	$(MAKE) -f prog.dir/build.make prog.dir/clean

build.make:

prog.dir/all: prog.dir/prog


prog.dir/prog:  prog.dir/a.o prog.dir/main.o
	gfortran -o prog.dir/prog  prog.dir/a.o prog.dir/main.o 

prog.dir/a.o: ../a.f90
	gfortran -o prog.dir/a.o  -c ../a.f90 -M prog.dir
	
prog.dir/localmod.mod: prog.dir/a.o	
	
prog.dir/main.o: ../main.f90 prog.dir/localmod.mod
	gfortran -o prog.dir/main.o  -c ../main.f90 -I prog.dir 
	
prog.dir/clean:
	rm prog.dir/localmod.mod prog.dir/a.o prog.dir/main.o prog.dir/prog

After you build prog using this set of Makefiles do (you're at the build directory)

 $ touch ../a.f90

and enter

 $ make

You'll see that a.f90 is recompiled, like the current CMake does, but main.f90 is recompiled too, as it should be.

Executable depending on external lib

This example build a executable target which

  1. provides a module
  2. uses the provided module
  3. uses a module of a external library

structure:

  • example_dependingOn_externalLib
    • extLib
      • include
        • externalmod.mod
      • lib
        • libmyextlib.a
    • myproject
      • build
        • Makefile
        • prog.dir
          • build.make
      • a.f90
      • main.f90

Contents of myproject...

a.f90:

MODULE localMod 
!
CONTAINS
    SUBROUTINE printLocalModGreeting
        WRITE(*,*) "Greetings from Module localMod"
    END SUBROUTINE
END MODULE

main.f90:

PROGRAM hello 
    USE localMod
    USE externalMod
    CALL printLocalModGreeting
    CALL printExtModGreeting
END PROGRAM

Rules like those generated by current CMake

Makefile:

all: prog.dir/all

prog.dir/all:
	$(MAKE) -f prog.dir/build.make prog.dir/requires
	$(MAKE) -f prog.dir/build.make prog.dir/all

clean:
	$(MAKE) -f prog.dir/build.make prog.dir/clean

build.make:

prog.dir/all: prog.dir/prog


prog.dir/prog: ../../extLib/lib/libmyextlib.a
prog.dir/prog: prog.dir/a.o prog.dir/main.o
	gfortran -o prog.dir/prog  prog.dir/a.o prog.dir/main.o ../../extLib/lib/libmyextlib.a

prog.dir/a.o: ../a.f90
	gfortran -o prog.dir/a.o  -c ../a.f90 -M prog.dir
	
prog.dir/main.o: ../main.f90 
	gfortran -o prog.dir/main.o	 -c ../main.f90 -I prog.dir -I ../../extLib/include
	
prog.dir/clean:
	rm prog.dir/localmod.mod prog.dir/a.o prog.dir/main.o prog.dir/prog
	
	
externalmod.mod.proxy: # dummy 

localmod.mod.proxy: prog.dir/a.o

prog.dir/main.o.requires: localmod.mod.proxy externalmod.mod.proxy
	
prog.dir/requires: prog.dir/main.o.requires

This rules got the same problem as the example above (simple Program with module) plus it doesn't recognizes if the external modules got updated.

Rules like those that should be generated by CMake

Makefile:

all: prog.dir/all

prog.dir/all:
	$(MAKE) -f prog.dir/build.make prog.dir/all

clean:
	$(MAKE) -f prog.dir/build.make prog.dir/clean

build.make:

prog.dir/all: prog.dir/prog


prog.dir/prog: ../../extLib/lib/libmyextlib.a
prog.dir/prog: prog.dir/a.o prog.dir/main.o
	gfortran -o prog.dir/prog  prog.dir/a.o prog.dir/main.o ../../extLib/lib/libmyextlib.a

prog.dir/a.o: ../a.f90
	gfortran -o prog.dir/a.o  -c ../a.f90 -M prog.dir
	
prog.dir/localmod.mod: prog.dir/a.o	
	
prog.dir/main.o: ../main.f90 prog.dir/localmod.mod
	gfortran -o prog.dir/main.o	 -c ../main.f90 -I prog.dir -I ../../extLib/include
	
prog.dir/clean:
	rm prog.dir/localmod.mod prog.dir/a.o prog.dir/main.o prog.dir/prog

These rules build everything in proper order and consider the timestamp of externalmod.mod.

Executable target depending on lib target

structure:

  • example_depending_libTarget
    • build
      • Makefile
      • lib.dir
        • build.make
        • libmodx.mod.stamp
        • libmody.mod.stamp
      • prog.dir
        • build.make
    • lib
      • a.f90
      • b.f90
    • prog
      • a.f90
      • main.f90

contents...

lib/a.f90:

MODULE libModX 
    USE libModY
END MODULE

lib/b.f90:

MODULE libModY
END MODULE

prog/a.f90:

MODULE localMod
END MODULE

prog/main.f90:

PROGRAM hello
    USE localMod
    USE libModX

    WRITE(*,*) 'Hello, F90 world.'
END PROGRAM

Rules like those generated by current CMake

build/Makefile:

all: lib.dir/all prog.dir/all
	

lib.dir/all:
	$(MAKE) -f lib.dir/build.make lib.dir/requires
	$(MAKE) -f lib.dir/build.make lib.dir/all

prog.dir/all: lib.dir/all
	$(MAKE) -f prog.dir/build.make  prog.dir/requires
	$(MAKE) -f prog.dir/build.make  prog.dir/all

	
clean:
	$(MAKE) -f prog.dir/build.make prog.dir/clean
	$(MAKE) -f lib.dir/build.make lib.dir/clean

build/lib.dir/build.make:

lib.dir/all: lib.dir/mylib

lib.dir/mylib: lib.dir/libmylib.a

lib.dir/libmylib.a: lib.dir/a.o lib.dir/b.o 
	ar rc lib.dir/libmylib.a lib.dir/a.o lib.dir/b.o  
	ranlib lib.dir/libmylib.a

lib.dir/a.o: ../lib/a.f90 
	gfortran -o lib.dir/a.o -c ../lib/a.f90 -M lib.dir

	
lib.dir/b.o: ../lib/b.f90 
	gfortran -o lib.dir/b.o -c ../lib/b.f90  -M lib.dir

 
libmody.mod.proxy: lib.dir/b.o

lib.dir/a.o.requires: libmody.mod.proxy
	
lib.dir/requires:  lib.dir/a.o.requires  


lib.dir/clean:
	rm lib.dir/a.o lib.dir/b.o lib.dir/libmylib.a
	rm lib.dir/libmodx.mod lib.dir/libmody.mod

build/prog.dir/build.make:

prog.dir/all: prog.dir/prog


prog.dir/prog: prog.dir/main.o prog.dir/a.o
	gfortran -o prog.dir/prog  prog.dir/main.o prog.dir/a.o  lib.dir/libmylib.a 

prog.dir/a.o: ../prog/a.f90
	gfortran -o prog.dir/a.o -c ../prog/a.f90 -M prog.dir
	
prog.dir/localmod.mod: prog.dir/a.o


prog.dir/main.o: ../prog/main.f90 
	gfortran -o prog.dir/main.o -c ../prog/main.f90 -I lib.dir -I prog.dir


localmod.mod.proxy: prog.dir/a.o
libmodx.mod.proxy: # dummy 

prog.dir/main.o.requires: localmod.mod.proxy libmodx.mod.proxy
	
prog.dir/requires: prog.dir/main.o.requires	


prog.dir/clean:
	rm prog.dir/a.o prog.dir/main.o prog.dir/prog prog.dir/localmod.mod

Again everything is build, but isn't updated proper.

Rules like those that should be generated by CMake

build/Makefile:

all: lib.dir/all prog.dir/all
	

lib.dir/all:
	$(MAKE) -f lib.dir/build.make lib.dir/all

prog.dir/all: lib.dir/all
	$(MAKE) -f prog.dir/build.make prog.dir/all

build/lib.dir/build.make:

lib.dir/all: lib.dir/mylib

lib.dir/mylib: lib.dir/libmylib.a

lib.dir/libmylib.a: lib.dir/a.o lib.dir/b.o 
	ar rc lib.dir/libmylib.a lib.dir/a.o lib.dir/b.o  
	ranlib lib.dir/libmylib.a

lib.dir/a.o: ../lib/a.f90 lib.dir/libmody.mod 
	gfortran -o lib.dir/a.o -c ../lib/a.f90 -M lib.dir
	touch lib.dir/libmodx.mod.stamp
	
lib.dir/b.o: ../lib/b.f90 
	gfortran -o lib.dir/b.o -c ../lib/b.f90  -M lib.dir
	touch lib.dir/libmody.mod.stamp
	

lib.dir/libmodx.mod: lib.dir/a.o
lib.dir/libmody.mod: lib.dir/b.o


lib.dir/clean:
	rm lib.dir/a.o lib.dir/b.o lib.dir/libmylib.a
	rm lib.dir/libmodx.mod lib.dir/libmody.mod

build/prog.dir/build.make:

prog.dir/all: prog.dir/prog


prog.dir/prog: prog.dir/main.o prog.dir/a.o
	gfortran -o prog.dir/prog  prog.dir/main.o prog.dir/a.o  lib.dir/libmylib.a 

prog.dir/a.o: ../prog/a.f90
	gfortran -o prog.dir/a.o -c ../prog/a.f90 -M prog.dir
	
prog.dir/localmod.mod: prog.dir/a.o

prog.dir/main.o: lib.dir/libmodx.mod.stamp
prog.dir/main.o: ../prog/main.f90 prog.dir/localmod.mod 
	gfortran -o prog.dir/main.o -c ../prog/main.f90 -I lib.dir -I prog.dir



prog.dir/clean:
	rm prog.dir/a.o prog.dir/main.o prog.dir/prog prog.dir/localmod.mod

Finally: Executable target depending on lib target and external lib

structure:

  • example_final
    • extLib
      • include
        • externalmod.mod
      • lib
        • libmyextlib.a
    • myproject
      • build
        • Makefile
        • lib.dir
          • build.make
          • libmodx.mod.stamp
          • libmody.mod.stamp
        • prog.dir
          • build.make
      • lib
        • a.f90
        • b.f90
      • prog
        • a.f90
        • main.f90


Contents...

lib/a.f90:

MODULE libModX 
    USE libModY
END MODULE

lib/b.f90:

MODULE libModY
END MODULE

prog/a.f90:

MODULE localMod
END MODULE


prog/b.f90:

PROGRAM hello
    USE localMod
    USE libModX
    USE externalMod

    WRITE(*,*) 'Hello, F90 world.'
    CALL printExtModGreeting
END PROGRAM


Rules like those generated by current CMake

build/Makefile:

all: lib.dir/all prog.dir/all
	

lib.dir/all:
	$(MAKE) -f lib.dir/build.make lib.dir/requires
	$(MAKE) -f lib.dir/build.make lib.dir/all

prog.dir/all: lib.dir/all
	$(MAKE) -f prog.dir/build.make prog.dir/requires
	$(MAKE) -f prog.dir/build.make prog.dir/all

	
clean:
	$(MAKE) -f prog.dir/build.make prog.dir/clean
	$(MAKE) -f lib.dir/build.make lib.dir/clean

build/lib.dir/build.make:

lib.dir/all: lib.dir/mylib

lib.dir/mylib: lib.dir/libmylib.a

lib.dir/libmylib.a: lib.dir/a.o lib.dir/b.o 
	ar rc lib.dir/libmylib.a lib.dir/a.o lib.dir/b.o  
	ranlib lib.dir/libmylib.a

lib.dir/a.o: ../lib/a.f90 
	gfortran -o lib.dir/a.o -c ../lib/a.f90 -M lib.dir
	
lib.dir/b.o: ../lib/b.f90 
	gfortran -o lib.dir/b.o -c ../lib/b.f90  -M lib.dir
	

libmody.mod.proxy: lib.dir/b.o

lib.dir/b.o.requires: libmody.mod.proxy

lib.dir/requires: lib.dir/b.o.requires


lib.dir/clean:
	rm lib.dir/a.o lib.dir/b.o lib.dir/libmylib.a
	rm lib.dir/libmodx.mod lib.dir/libmody.mod

build/prog.dir/build.make:

prog.dir/all: prog.dir/prog

prog.dir/prog: ../../extLib/lib/libmyextlib.a
prog.dir/prog: prog.dir/main.o prog.dir/a.o ../../extLib/lib/libmyextlib.a
	gfortran -o prog.dir/prog  prog.dir/main.o prog.dir/a.o  lib.dir/libmylib.a ../../extLib/lib/libmyextlib.a

prog.dir/a.o: ../prog/a.f90
	gfortran -o prog.dir/a.o -c ../prog/a.f90 -M prog.dir
	
prog.dir/main.o: 
prog.dir/main.o: ../prog/main.f90 
	gfortran -o prog.dir/main.o -c ../prog/main.f90 -I lib.dir -I prog.dir -I ../../extLib/include


localmod.mod.proxy: prog.dir/a.o
libmodx.mod.proxy: # dummy
externalmod.mod.proxy: # dummy

prog.dir/main.o.requires: localmod.mod.proxy libmodx.mod.proxy externalmod.mod.proxy

prog.dir/requires: prog.dir/main.o.requires

prog.dir/clean:
	rm prog.dir/a.o prog.dir/main.o prog.dir/prog prog.dir/localmod.mod

Rules like those that should be generated by CMake

build/Makefile:

all: lib.dir/all prog.dir/all
	

lib.dir/all:
	$(MAKE) -f lib.dir/build.make lib.dir/all

prog.dir/all: lib.dir/all
	$(MAKE) -f prog.dir/build.make prog.dir/all

	
clean:
	$(MAKE) -f prog.dir/build.make prog.dir/clean
	$(MAKE) -f lib.dir/build.make lib.dir/clean

build/lib.dir/build.make:

lib.dir/all: lib.dir/mylib

lib.dir/mylib: lib.dir/libmylib.a

lib.dir/libmylib.a: lib.dir/a.o lib.dir/b.o 
	ar rc lib.dir/libmylib.a lib.dir/a.o lib.dir/b.o  
	ranlib lib.dir/libmylib.a

lib.dir/a.o: ../lib/a.f90 lib.dir/libmody.mod 
	gfortran -o lib.dir/a.o -c ../lib/a.f90 -M lib.dir
	touch lib.dir/libmodx.mod.stamp
	
lib.dir/b.o: ../lib/b.f90 
	gfortran -o lib.dir/b.o -c ../lib/b.f90  -M lib.dir
	touch lib.dir/libmody.mod.stamp
	

lib.dir/libmodx.mod: lib.dir/a.o
lib.dir/libmody.mod: lib.dir/b.o


lib.dir/clean:
	rm lib.dir/a.o lib.dir/b.o lib.dir/libmylib.a
	rm lib.dir/libmodx.mod lib.dir/libmody.mod

build/prog.dir/build.make:

prog.dir/all: prog.dir/prog

prog.dir/prog: ../../extLib/lib/libmyextlib.a
prog.dir/prog: prog.dir/main.o prog.dir/a.o ../../extLib/lib/libmyextlib.a
	gfortran -o prog.dir/prog  prog.dir/main.o prog.dir/a.o  lib.dir/libmylib.a ../../extLib/lib/libmyextlib.a

prog.dir/a.o: ../prog/a.f90
	gfortran -o prog.dir/a.o -c ../prog/a.f90 -M prog.dir
	
prog.dir/localmod.mod: prog.dir/a.o

prog.dir/main.o: lib.dir/libmodx.mod.stamp
prog.dir/main.o: ../../extLib/include/externalmod.mod
prog.dir/main.o: ../prog/main.f90 prog.dir/localmod.mod 
	gfortran -o prog.dir/main.o -c ../prog/main.f90 -I lib.dir -I prog.dir -I ../../extLib/include



prog.dir/clean:
	rm prog.dir/a.o prog.dir/main.o prog.dir/prog prog.dir/localmod.mod

Conclusion

What keeps CMake from doing it like could be done shown above?

For each target CMake parses the source files and writes the dependencies one by one. This is ok for includes. But doing it this way CMake cannot determine if a required module of source file a.f90 is provided by a source file b.f90 of the same target or not. This is IMHO the reason why the CMake developer droped a direct dependency of the source file to a required module. As shown in the rules like they are generated by current CMake sections an extra step called required was introduced. This works out if one want just to build, but isn't enough for developer needs (because recompilation isn't done as expected).

What kind of changes have to be done to make it happen?

  1. Before the actually dependency tracking starts, CMake has to parse all fortran sources and to create a corresponding file i.e. mymodule.mod.stamp for mymodule. This can be done source by source. Now CMake is able to search the build tree for a module.
  2. Rather than doing it one by one, all sources of a target have to be parsed before starting to write dependencies. This way CMake knows if a required modules is provided by itself or not.
  3. In case a required module mymod isn't provided by the same target
    1. search the build-tree for mymod.mod.stamp
    2. if not found search it at the include paths


Module dependency tracking superseeds include dependency tracking but

Note: IMHO the code responsible for C/C++/Java dependency generation shouldn't be touched, since speed is an imporant advantage of CMake for developers!

Notes from Brad

The "provides", "requires", and "proxy" stuff is necessary for correct minimal rebuilds even within a single target.

Maik: AFAI understand its necessary since there are no direct dependencies of target sources to the mod files, right?

However it has not been correctly implemented in the latest Makefile generator.

Maik: I know cmake since 2.4.2, but AFAIK former versions generated one big Makefile, right?

The idea is that after recompiling a source file that provides a module, there should be a copy-if-different step to update the mod.stamp file.

Maik: And this stamp is useless since the last Makefile generator was introduced.
Brad: Like I said it is not correctly implemented. The .o files need to depend on the .stamp files.

Then there must be a recursive call to make for each source that requires a module to evaluate the file-level dependencies with the updated stamp file.

I do not think we can search the whole build tree for mod.stamp files. The trees can be way to big and that would take forever.

Maik: ...(1)
Brad: There will have to be some kind of persistent scanning results for (1) to work. We cannot scan the whole tree every time CMake runs.

Also there could be leftover stamp files from modules that were once available and are not anymore or were moved.

Maik: good point! I didn't consider this.

Also, creating inter-target dependencies cannot be done during the dependency scanning stage, so we cannot create such dependencies based on information implicit in the source files. However, if one target is importing the modules from another target, it must link to that target in order to get the implementation. That linking should create the explicit inter-target dependency that is necessary.

Maik: Yes, that is I have in mind. We only have to search for implicit dependencies where explicit ones are given by the user. I think this solves your concern (1) too. Additionally we only have to search in fortran projects.

This just leaves creating the dependencies on the mod.stamp files for modules provided in other targets. I'm still thinking about this one.

Maik

I will delete my notes on your notes ;) as soon as I'm sure I didn't get you wrong.

The stamp file deletion problem is indeed tricky. I think the extra step which creates them should look for existing to delete them before writing new. The question is how can CMake determine if a fortran-target source got rid of a module to trigger this extra step. I doubt this is possible, so user interaction is required (rerun-CMake).

BTW: CMake only has to create a stamp file if the source file which contains the corresponding module belongs to a lib-target.

Brad

I've updated http://www.cmake.org/Bug/view.php?id=5809 with a note at "10-12-07 09:53". A partial fix for file-level dependencies with module timestamps has been implemented. Rebuilding should now work correctly for all object files whose required modules are provided in the same directory (even if in different targets).