C17/C23: "floating constant is out of range" for 1e10000, which is in range when double has infinities

[thinkoid]

Issue Summary

double h = 1e10000; is rejected with error #30 (bad_float_value) in --c17 and --c23 modes. C17 5.2.4.2.2p5 and C23 5.2.5.3.3p19 say that if positive infinity is representable in a floating type, the range of that type extends to all positive real numbers. On x86_64 double has an infinity, so 1e10000 is within the range of representable values, the constraint in C17 6.4.4p2 (C23 6.4.4.1p2) is met, and the constant is valid.

EDG already gives such constants their C17 meaning under GNU emulation: conv_host_fp_to_double (src/float_pt.c:1067-1068) reports the overflow unless gnu_mode is set, and in that mode "GNU C and C++ silently uses infinity for values that are too large". In C17 that is one of the results 6.4.4.2p3 allows (the nearest representable value or one adjacent to it), and the one 6.4.4.2p7 recommends: translation should match strtod, which returns infinity for this input.

This matters in practice: glibc's <math.h> defines HUGE_VAL as 1e10000 for compilers that do not define __GNUC__, so outside GNU emulation HUGE_VAL cannot be used.

Observed Behavior

cpfe built from the linux-gcc-release preset, run with the repository's own base:

$ export EDG_BASE=$PWD/bases/docker/dev-env/gcc
$ cat big.c
double h = 1e10000;
$ cpfe --c17 -A big.c
"big.c", line 1: error: floating constant is out of range
  double h = 1e10000;
             ^

1 error detected in the compilation of "big.c".
$ cpfe --c23 -A big.c
"big.c", line 1: error: floating constant is out of range
  double h = 1e10000;
             ^

1 error detected in the compilation of "big.c".

The same error appears without -A (cpfe --c17 big.c). The diagnostic is not discretionary: cpfe --c17 -A --diag_warning=30 big.c still reports the error. Under GNU emulation the constant is accepted: cpfe --gcc big.c reports nothing.

Expected Behavior

The constant is accepted in C17 and C23 modes, for instance converted to infinity as under GNU emulation. A warning would be fine: gcc -std=c17 -pedantic-errors -c big.c accepts it with only a -Woverflow warning.

C99 and C11 do not define the range of a floating type with infinities, so this report is limited to C17 and C23.

Additional Details

• EDG 7.0, edgcpp/compiler at ca5890bab8, x86_64 Linux.

Configuration: cpfe --dump_configuration output attached: dump_configuration.txt

[thinkoid]

New findings, which widen this report beyond C17 and C23.

A search of glibc's history shows where 1e10000 came from. Commit a60eca2e55 ("Simplify HUGE_VAL definitions", first in glibc 2.27) replaced the union-based fallbacks for non-GCC compilers with this constant. The commit message justifies the value by WG14's answer to DR #025: "C90 DR#025 said that if a floating-point format represents infinities, all real values lie within the range of representable values, so the constraints for constant expressions are not violated".

DR #25 (1992) asked what "representable floating-point value" means. The committee's response:

Implementations may elect to represent ``infinite'' values, in which case all real numbers would lie within the range of representable values.

and

This implies that a conforming implementation is allowed to accept a floating-point constant of any arbitrarily large or small value.

The constraint it reads is in every C dialect. C90 6.1.3: "The value of a constant shall be in the range of representable values for its type." C99 6.4.4p2 and C11 6.4.4p2 keep the same requirement. C17 5.2.4.2.2p5 later wrote the answer into the standard. So 1e10000 is in range in C89, C99 and C11 as well, and those strict modes give the same error:

$ cat big.c
double h = 1e10000;
$ cpfe --c89 -A big.c
"big.c", line 1: error: floating constant is out of range
  double h = 1e10000;
             ^

1 error detected in the compilation of "big.c".
$ cpfe --c99 -A big.c
"big.c", line 1: error: floating constant is out of range
  double h = 1e10000;
             ^

1 error detected in the compilation of "big.c".
$ cpfe --c11 -A big.c
"big.c", line 1: error: floating constant is out of range
  double h = 1e10000;
             ^

1 error detected in the compilation of "big.c".

gcc accepts the constant under -std=c89, -std=c99 and -std=c11, each with -pedantic-errors, and gives a -Woverflow warning.

So the fix asked for here should cover the C89, C99 and C11 modes as well as C17 and C23.