Compiler options

1

I have tweeted about the -o and -c compiler options and will link to Compiler options comparison. What compiler options do you think are worth mentioning, especially those that people are unaware of?

I plan to write about

limiting the number of error messages
gfortran -fmax-errors=n
ifort -error-limit:n

redirecting compiler errors with 2>

standard-conformance options

dusty-deck options
such as gfortran -std=legacy

optimization options
-O1 -O2 -O3, -march=native, and for gfortran --ffast-math, but with caveats

debugging options
such as gfortran -fbounds-check -g

options to promote single to double precision
primarily to demonstrate that problems that can arise when relying on such options

Options related to coarrays or OpenMP will be discussed when topics are.

5 Likes
2

One thing I think many is unaware of is that it’s perfectly fine to compile with optimization (like - O2) and debug information (-g) at the same time. This can be useful to debug issues in long and running computations. The compiler may have reordered some code etc. so the debugging experience might not be as intuitive as with -Od/-debug though.

4 Likes
4

Standard conformance options are of two kinds, both of which useful. One gives diagnostics when non-standard features are used - this is useful in the development process as sometimes it highlights a usage you had not intended. The second adjusts how the compiler behaves, such as ifort’s -standard-semantics. Compilers with a long history often have default behaviors that don’t match what the current standard says. I have seen many users complain that something doesn’t work right when in fact they were getting the compiler’s traditional, but now non-standard, behavior.

6 Likes
5

I may mention that

gfortran -c -fdump-fortran-original -fdump-tree-original a.f90
prints the internal representation of the Fortran abstract syntax tree
to standard output and writes a C-like representation that is passed
to the middle end.

For example, the

Fortran code to compute Fibonacci numbers 
module fib_mod
implicit none
contains
elemental function fib(i) result(ifib)
integer, intent(in) :: i
integer             :: ifib,j,fib_old(2)
ifib = merge(0, 1, i<1)
if (i < 3) return
fib_old = [1,1]
do j=3,i
   ifib = fib_old(1) + fib_old(2)
   if (j == i) exit
   fib_old = [ifib,fib_old(1)]
end do
end function fib
end module fib_mod
!
program test_fib
use fib_mod, only: fib
implicit none
print*,"fib(6) =",fib(6) ! 8
end program test_fib
gives an AST 
Namespace: A-Z: (UNKNOWN 0)
procedure name = fib_mod
  symtree: 'fib'         || symbol: 'fib'          
    type spec : (INTEGER 4)
    attributes: (PROCEDURE MODULE-PROC  FUNCTION ELEMENTAL PURE CONTAINED ALWAYS-EXPLICIT)
    result: ifib
    Formal arglist: i
  symtree: 'fib_mod'     || symbol: 'fib_mod'      
    type spec : (UNKNOWN 0)
    attributes: (MODULE )

  code:
CONTAINS

  Namespace: A-Z: (UNKNOWN 0)
  procedure name = fib
    symtree: 'fib'         || symbol: 'fib' from namespace 'fib_mod'
    symtree: 'fib_old'     || symbol: 'fib_old'      
      type spec : (INTEGER 4)
      attributes: (VARIABLE  DIMENSION)
      Array spec:(1 [0] AS_EXPLICIT 1 2 )
    symtree: 'i'           || symbol: 'i'            
      type spec : (INTEGER 4)
      attributes: (VARIABLE  DUMMY(IN))
    symtree: 'ifib'        || symbol: 'ifib'         
      type spec : (INTEGER 4)
      attributes: (VARIABLE  RESULT ALWAYS-EXPLICIT)
    symtree: 'j'           || symbol: 'j'            
      type spec : (INTEGER 4)
      attributes: (VARIABLE )
    symtree: 'merge'       || symbol: 'merge'        
      type spec : (UNKNOWN 0)
      attributes: (PROCEDURE INTRINSIC-PROC  FUNCTION ARRAY-OUTER-DEPENDENCY)
      result: merge

    code:
    ASSIGN fib:ifib __merge_i4[[((0) (1) ((< fib:i 1)))]]
    IF (< fib:i 3)
      RETURN 
    ENDIF
    ASSIGN fib:fib_old(FULL) (/ 1 , 1 /)
    DO fib:j=3 fib:i 1
      ASSIGN fib:ifib (+ fib:fib_old(1) fib:fib_old(2))
      IF (== fib:j fib:i)
        EXIT
      ENDIF
      ASSIGN fib:fib_old(FULL) (/ fib:ifib , fib:fib_old(1) /)
    END DO
    


Namespace: A-Z: (UNKNOWN 0)
procedure name = test_fib
  symtree: 'fib'         || symbol: 'fib'          
    type spec : (INTEGER 4)
    attributes: (PROCEDURE MODULE-PROC  USE-ASSOC(fib_mod) FUNCTION ELEMENTAL PURE ALWAYS-EXPLICIT)
    result: ifib
    Formal arglist: i
  symtree: 'fib_mod'     || symbol: 'fib_mod'      
    type spec : (UNKNOWN 0)
    attributes: (MODULE  USE-ASSOC(fib_mod))
  symtree: 'test_fib'    || symbol: 'test_fib'     
    type spec : (UNKNOWN 0)
    attributes: (PROGRAM PUBLIC  SUBROUTINE IS-MAIN-PROGRAM)

  code:
  WRITE UNIT=6 FMT=-1
  TRANSFER 'fib(6) ='
  TRANSFER fib[[((6))]]
  DT_END
and C-like representation 
__attribute__((fn spec (". r ")))
integer(kind=4) fib (integer(kind=4) & restrict i)
{
  integer(kind=4) fib_old[2];
  integer(kind=4) ifib;
  integer(kind=4) j;

  {
    static integer(kind=4) A.0[2] = {1, 1};

    ifib = (integer(kind=4)) (*i > 0);
    if (*i <= 2)
      {
        return ifib;
      }
    L.1:;
    (void) __builtin_memcpy ((void *) &fib_old, (void *) &A.0, 8);
    {
      integer(kind=4) D.4223;

      D.4223 = *i;
      j = 3;
      while (1)
        {
          {
            logical(kind=4) D.4226;

            D.4226 = j > D.4223;
            if (D.4226) goto L.3;
            ifib = fib_old[0] + fib_old[1];
            if (*i == j) goto L.3;
            L.4:;
            {
              struct array01_integer(kind=4) atmp.1;
              integer(kind=4) A.2[2];

                            typedef integer(kind=4) [2];
              atmp.1.dtype = {.elem_len=4, .rank=1, .type=1};
              atmp.1.dim[0].stride = 1;
              atmp.1.dim[0].lbound = 0;
              atmp.1.dim[0].ubound = 1;
              atmp.1.span = 4;
              atmp.1.data = (void * restrict) &A.2;
              atmp.1.offset = 0;
              (*(integer(kind=4)[2] * restrict) atmp.1.data)[0] = ifib;
              (*(integer(kind=4)[2] * restrict) atmp.1.data)[1] = fib_old[0];
              {
                integer(kind=8) S.4;

                S.4 = 0;
                while (1)
                  {
                    if (S.4 > 1) goto L.5;
                    fib_old[S.4] = (*(integer(kind=4)[2] * restrict) atmp.1.data)[S.4];
                    S.4 = S.4 + 1;
                  }
                L.5:;
              }
            }
            L.2:;
            j = j + 1;
          }
        }
      L.3:;
    }
  }
  return ifib;
}


__attribute__((fn spec (". ")))
void test_fib ()
{
  {
    struct __st_parameter_dt dt_parm.5;

    dt_parm.5.common.filename = &"fib.f90"[1]{lb: 1 sz: 1};
    dt_parm.5.common.line = 21;
    dt_parm.5.common.flags = 128;
    dt_parm.5.common.unit = 6;
    _gfortran_st_write (&dt_parm.5);
    _gfortran_transfer_character_write (&dt_parm.5, &"fib(6) ="[1]{lb: 1 sz: 1}, 8);
    {
      static integer(kind=4) C.4258 = 6;
      integer(kind=4) D.4259;

      D.4259 = fib (&C.4258);
      _gfortran_transfer_integer_write (&dt_parm.5, &D.4259, 4);
    }
    _gfortran_st_write_done (&dt_parm.5);
  }
}


__attribute__((externally_visible))
integer(kind=4) main (integer(kind=4) argc, character(kind=1) * * argv)
{
  static integer(kind=4) options.6[7] = {2116, 4095, 0, 1, 1, 0, 31};

  _gfortran_set_args (argc, argv);
  _gfortran_set_options (7, &options.6[0]);
  test_fib ();
  return 0;
}

Looking at these stages will be useful to a compiler developer when there is a problem. When would it be helpful to an end user? I noticed that when

integer :: ifib,j,fib_old(2)

is replaced by the equivalent

integer             :: ifib,j
integer, dimension(2) :: fib_old

that the AST’s of the two codes are the same. Maybe looking at intermediate stages could be used to verify that cosmetic code changes have not changed the output of a code (although comparing the results of executables is the first thing to do.)

6

A project I work on actually did this in early stages. We used a tool to convert from fixed format to free format and used the AST output of a compiler to verify that the converted code was equivalent, since the produced AST was identical.

2 Likes