Re: Best way to simulate C-assert using gfortran

[Tobias Burnus <burnus at net-b dot de>]

Am 04.02.2013 09:50, schrieb Arjen Markus:
>      pure subroutine assertion(cond)
>          implicit none
>          logical, intent(in) :: cond
>          real, volatile :: r
>          if ( noassert ) return   !<===
>          r = 1.0
>          if (.not. cond) r = r / 0.0
>      end subroutine assertion
>
> The parameter noassert should - in principle - enable the compiler to delete
> the actual call to the subroutine. (I have been unable to verify this from the
> call tree, as the option -fdump-fortran-optimized did not produce any output.)

Try compiling with "-O1". "if (.true.)" should already at -O0 generate 
unconditional code (although, the - unreachable - code after the 
"return" is generated).

Instead of "r = r/0.0" consider "call abort()". "abort()" is a (common) 
vendor extension. The call to abort has two advantages: First, it really 
force-stops the code while "1.0 / 0.0" requires trapping.

Secondly, "abort" is marked as noreturn. If you have code like:
   if (cond) then
      call abort ()
   else ! or "end if"
   ...
the compiler assumes that "cond" evaluates as false. This branch 
prediction can speed up the code. The downside is that if the compiler 
predicted it wrongly, some cycles are wasted to recover from the wrong 
branch. Well, as "abort()" exits the code, the wasted cycles shouldn't 
matter. (Hence, this kind of optimization is made if one branch contains 
a noreturn procedure call.)

The downside of "abort" is that it is not part of the Fortran standard, 
i.e. not all compilers support it.

Tobias