Simplify HUGE_VAL definitions [committed]
Commit Message
On 09/01/2017 09:24 AM, Joseph Myers wrote:
> On Fri, 1 Sep 2017, Zack Weinberg wrote:
>
>>> Annex F requires conversions of constants (with at most DECIMAL_DIG
>>> digits) to be to-nearest. (TS 18661-1 adds the FENV_ROUND pragma to
>>> choose a different rounding mode for them and for certain operations.)
>>
>> Yes, but for any constant greater than <type>_MAX, <type>_MAX is
>> arguably nearer to that constant than +Infinity is.
>
> The IEEE rounding modes (all of them) define overflow on the basis of
> whether the result with normal precision but infinite exponent range would
> have an exponent that's too big (and in the case of overflow, the rounded
> result is determined by the rounding mode, so +Inf for to-nearest).
OK, I think I get it.
One more problem, which I should have thought to check earlier: 1e10000L
isn't big enough if long double is 128 bits wide (which it actually
appears to be on this computer):
$ gcc -E -dM -xc - < /dev/null | grep -E '(DBL|LDBL|FLT[0-9]*)_MAX_EXP'
#define __LDBL_MAX_EXP__ 16384
#define __DBL_MAX_EXP__ 1024
#define __FLT32_MAX_EXP__ 128
#define __FLT128_MAX_EXP__ 16384
#define __FLT_MAX_EXP__ 128
#define __FLT64_MAX_EXP__ 1024
So I propose this patch (not yet tested):
Comments
On Fri, 1 Sep 2017, Zack Weinberg wrote:
> One more problem, which I should have thought to check earlier: 1e10000L
> isn't big enough if long double is 128 bits wide (which it actually
> appears to be on this computer):
1e10000L is big enough. For both x86 extended and binary128, the largest
finite value is just below 2^16384 (about 1.2e4932).
It's true that it wouldn't be enough for _Float128x (which must have a
precision of at least 128 bits and maximum exponent at least 65535; e.g.
binary192 would meet those requirements, binary160 wouldn't; GCC does not
support _Float128x, or any such format, on any target). But if we
supported such types I don't expect we'd have a HUGE_VAL fallback for them
any more than we do for HUGE_VAL_F128.
@@ -37,20 +37,25 @@ __BEGIN_DECLS
/* Gather machine dependent type support. */
#include <bits/floatn.h>
-/* Value returned on overflow. On all IEEE754 machines, this is
- +Infinity. */
+/* Value returned on overflow. With IEEE 754 floating point, this is
+ +Infinity, otherwise the largest representable positive value. */
#if __GNUC_PREREQ (3, 3)
# define HUGE_VAL (__builtin_huge_val ())
#else
-# define HUGE_VAL 1e10000
+/* This may provoke compiler warnings, and may not be rounded to
+ +Infinity in all IEEE 754 rounding modes, but is the best that
+ can be done in ISO C while remaining a constant expression.
+ 100,000 is greater than the maximum exponent for all supported
+ floating-point formats and widths. */
+# define HUGE_VAL 1e100000
#endif
#ifdef __USE_ISOC99
# if __GNUC_PREREQ (3, 3)
# define HUGE_VALF (__builtin_huge_valf ())
# define HUGE_VALL (__builtin_huge_vall ())
# else
-# define HUGE_VALF 1e10000f
-# define HUGE_VALL 1e10000L
+# define HUGE_VALF 1e100000f
+# define HUGE_VALL 1e100000L
# endif
#endif
#if __HAVE_FLOAT128 && __GLIBC_USE (IEC_60559_TYPES_EXT)