**Document**: WG14/N1361

**Date**: 2009/02/27

**References**: WG14/N1321,
WG14/N1346

**Authors**: Jim
Thomas

**Reply
to**:

**Subject**:
FLT_EVAL_METHOD and constants

**Problem:** The
September 2008 meeting minutes (WG14/N1346), item 4.21, record:

Straw
Poll:

1.
Accept N1321 to the WP, removing const from FLT_EVAL_METHOD 8‐1‐8
YES

Removing
constants from the FLT_EVAL_METHOD would cause silent changes in the behavior
of C99 strictly conforming programs. For example, 5.2.4.2.2 #8 says that if
FLT_EVAL_METHOD equals 1, then float constants are evaluated to the range and
precision of the double type. Annex F says that if the implementation defines
__STDC_IEC_559__, then the double type and its basic arithmetic is fully
specified by the floating-point standard and the conversion from decimal
character sequences (of length up to DECIMAL_DIG) to double is correctly
rounded. Therefore, the following program should never print “non-conforming
implementation”:

#include
<float.h>

#include
<stdio.h>

Int
main(void)

{

#if
defined(__STDC_IEC_559__) && (FLT_EVAL_METHOD == 1)

If (0.1F != (1.F / 10.F))
printf(“non-conforming implementation\n”);

#endif

}

However,
it would print “non-conforming implementation” if the #if conditions were
satisfied but the constant 0.1F were not widened to double like the divide
operation.

The
reason C99 included constants in FLT_EVAL_METHOD is the following. Commonly, programs
use a given floating type because the type was assumed (without benefit of
error analysis) to be wide enough, not because wider precision would be
incorrect. That assumption might be incorrect for unanticipated data sets. Wide
evaluation can make these programs more robust. Constants, which amount to
translation-time decimal-to-binary conversion operations, can be widened like
arithmetic operations and contribute to the overall positive effects of wide
evaluation.

On the
other hand, there are arguments against including constants in FLT_EVAL_METHOD.
Constants wider than their suffix type might surprise some users. Also, some
special programs require constants rounded to a specific width (and not wider).
The hexadecimal form can be used for floating constants whose value is
independent of the evaluation method, though this option may be unfamiliar to
some users.

**Recommendation**: Don’t remove constants from the current
standard evaluation methods, namely those for FLT_EVAL_METHOD values 0, 1, or 2.

If standard-supported evaluation methods that widen
operations but not constants are deemed desirable then these could be defined
by the standard and specified by positive FLT_EVAL_METHOD values other than 1
or 2.