This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [fortran PATCH] Implement a(:,:) = 0.0 using memset (take 2)

On Mon, Dec 18, 2006 at 06:55:22PM -0700, wrote:
> 2006-12-18  Roger Sayle  <>
>         * trans-expr.c (is_zero_initializer_p): Determine whether a given
>         constant expression is a zero initializer.
>         (gfc_trans_zero_assign): New function to attempt top optimize
>         "a(:) = 0.0" as a call to __builtin_memset (a, 0, sizeof(a));
>         (gfc_trans_assignment): Special case array assignments to zero
>         initializer constants, using gfc_trans_zero_assign.
>         * gfortran.dg/array_memset_1.f90: New test case.


I think this is OK for trunk; however, Andrew did bring up
the potential problem of a signed zero.  I'm not sure if
-0 occurs very often in Fortran code, but we can guard against
this by changing your code

> + 
> +     case BT_REAL:
> +       return mpfr_sgn (expr->value.real) == 0;

        case BT_REAL:
          return (mpfr_sgn (expr->value.real) == 0
                  && MPFR_SIGN  (expr->value.real) == 1);

> +     case BT_COMPLEX:
> +       return mpfr_sgn (expr->value.complex.r) == 0
> +              && mpfr_sgn (expr->value.complex.i) == 0;

A similar change is needed above.

MPFR_SIGN retrieves the sign bit from the mpfr_t structure.

As a side note, I found that your patch provides a 7.2% decrease
in computation time for the Bergen Ocean Model (ie., your patch
is beneficial!).  There are 176 __builtin_memset calls in BOM
due to your patch.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]