http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60128
--- Comment #4 from Dominique d'Humieres <dominiq at lps dot ens.fr> ---
The regression reported in comment 3 disappears with the following patch
--- ../_clean/libgfortran/io/write_float.def 2014-01-21 08:30:57.000000000
+0100
+++ libgfortran/io/write_float.def 2014-02-10 11:35:55.000000000 +0100
@@ -326,7 +322,7 @@ output_float (st_parameter_dt *dtp, cons
for IEEE 754, this ought to be round to nearest, ties to
even, corresponding to the Fortran ROUND='NEAREST'. */
case ROUND_PROCDEFINED:
- case ROUND_UNSPECIFIED:
+ /* case ROUND_UNSPECIFIED: */
case ROUND_ZERO: /* Do nothing and truncation occurs. */
goto skip;
case ROUND_UP:
@@ -338,6 +334,7 @@ output_float (st_parameter_dt *dtp, cons
goto skip;
goto updown;
case ROUND_NEAREST:
+ case ROUND_UNSPECIFIED:
/* Round compatible unless there is a tie. A tie is a 5 with
all trailing zero's. */
i = nafter + nbefore;
@@ -1215,8 +1212,8 @@ determine_en_precision (st_parameter_dt
nbefore = 3 - nbefore;
}
int prec = f->u.real.d + nbefore;
- if (dtp->u.p.current_unit->round_status != ROUND_UNSPECIFIED
- && dtp->u.p.current_unit->round_status != ROUND_PROCDEFINED)
+ /* if (dtp->u.p.current_unit->round_status != ROUND_UNSPECIFIED
+ && dtp->u.p.current_unit->round_status != ROUND_PROCDEFINED) */
prec += 2 * len + 4;
return prec;
}
>From the comment
/* For processor defined and unspecified rounding we use
snprintf to print the exact number of digits needed, and thus
let snprintf handle the rounding. On system claiming support
for IEEE 754, this ought to be round to nearest, ties to
even, corresponding to the Fortran ROUND='NEAREST'. */
case ROUND_PROCDEFINED:
case ROUND_UNSPECIFIED:
case ROUND_ZERO: /* Do nothing and truncation occurs. */
I think it is wrong to handle ROUND_PROCDEFINED and ROUND_UNSPECIFIED as
ROUND_ZERO.
