Added restat = 1 in the fortran dyndep example

[Volker-Weissmann]

Let's say you have a Fortran Project consisting of two files:

$ cat foo.f90
module foo
  implicit none
  integer, parameter :: rkind=SELECTED_REAL_KIND(15,307)

  contains
     function myProcessing(x) RESULT(z)
	    real(kind=rkind):: x,z

	 end function myProcessing
end module foo
$ cat bar.f90
program main
  use foo
  implicit none
  real(kind=rkind) :: val, res

  res = myProcessing(val)
end program main

If you do it like the manual suggests, you run into a problem:

$ cat build.ninja 
rule f95
  command = f95 -o $out -c $in
rule fscan
  command = echo -e "ninja_dyndep_version = 1\nbuild foo.o | foo.mod: dyndep\nbuild bar.o: dyndep | foo.mod" > $out
build foobar.dd: fscan foo.f90 bar.f90
build foo.o: f95 foo.f90 || foobar.dd
  dyndep = foobar.dd
build bar.o: f95 bar.f90 || foobar.dd
  dyndep = foobar.dd
$ ninja
[3/3] f95 -o bar.o -c bar.f90
$ ninja
ninja: no work to do.
$ touch foo.f90 
$ ninja
[3/3] f95 -o bar.o -c bar.f90
$ ninja -v -d explain
ninja explain: loading dyndep file 'foobar.dd'
ninja explain: output foo.mod older than most recent input foo.f90 (1679308368472041273 vs 1679308373408688040)
ninja explain: foo.mod is dirty
[1/2] f95 -o foo.o -c foo.f90
[2/2] f95 -o bar.o -c bar.f90
$ ninja -v -d explain
ninja explain: loading dyndep file 'foobar.dd'
ninja explain: output foo.mod older than most recent input foo.f90 (1679308368472041273 vs 1679308373408688040)
ninja explain: foo.mod is dirty
[1/2] f95 -o foo.o -c foo.f90
[2/2] f95 -o bar.o -c bar.f90

As you can see ninja rebuilds stuff even though nothing should get rebuild. The last two ninja calls should have printed ninja: no work to do..

I did some debugging and found that the following build.ninja file has the same problem:

rule f95
  command = f95 -o $out -c $in
build foo.o | foo.mod: f95 foo.f90
build bar.o: f95 bar.f90 | foo.mod

But why? The build.ninja file looks so correct!
The problem lies in the assumption that f95 -o foo.o -c foo.f90 will generate foo.mod. This is only half-true:

$ rm parent.mod
$ f95 -o foo.o -c foo.f90
$ stat foo.mod 
Access: 2023-03-20 11:36:14.018192964 +0100
Modify: 2023-03-20 11:36:14.018192964 +0100
Change: 2023-03-20 11:36:14.018192964 +0100
 Birth: 2023-03-20 11:36:14.018192964 +0100
$ f95 -o foo.o -c foo.f90
$ stat foo.mod 
Access: 2023-03-20 11:36:18.184855010 +0100
Modify: 2023-03-20 11:36:14.018192964 +0100
Change: 2023-03-20 11:36:14.018192964 +0100
 Birth: 2023-03-20 11:36:14.018192964 +0100
$ f95 -o foo.o -c foo.f90
$ stat foo.mod 
Access: 2023-03-20 11:36:18.184855010 +0100
Modify: 2023-03-20 11:36:14.018192964 +0100
Change: 2023-03-20 11:36:14.018192964 +0100
 Birth: 2023-03-20 11:36:14.018192964 +0100

As you can see, the fortran compiler indeed created foo.mod. But it does not update the modification time of foo.mod if it already does exist. If I insert a restat

build foo.o | foo.mod: f95 foo.f90
    restat = 1

everything seems to work.

I'm using gfortran 12.2.1