fcm make: build: compiler aggressive inlining breaks incremental build assumption

[matthewrmshin]

Consider a simple case where main.f90 depends on mod1.f90.

! In main.f90
program main
use mod1, only: sub1
integer :: c
call sub1(c, 1, 2)
write(*, '(i0)') c
end program main

! In mod1.f90
module mod1
contains
subroutine sub1(c, a, b)
integer, intent(out) :: c
integer, intent(in) :: a
integer, intent(in) :: b
c = a + b + 1
end subroutine sub1
end module mod1

Some compiler (e.g. Cray), on -O3, may inline logic of mod1.f90 into main.o. If we subsequently change the line c = a + b + 1 to say c = a + b + 20, the module will be recompiled, updating mod1.o but mod1.mod will remain unchanged - as there is no change to the module interface. This causes fcm make to think that it does not need to recompile main.f90, although it will re-link the executable.

The result of the executable will then be wrong in incremental mode.

[steoxley]

Is the fix in this case to simply add a dependency on the .o as well as the .mod or is it more complicated than this?

Developers of VAR occasionally have issues with incremental builds and I would like to rule this out as being the cause before investigating further.

[matthewrmshin]

Is the fix in this case to simply add a dependency on the .o as well as the .mod or is it more complicated than this?

Probably.