num_put::put(..., 1.0e+4932L, ...) formats "inf"

• Key: STDCXX-880
• Type: Bug
• Status: Closed
• Resolution: Fixed
• Priority: Blocker
• Assignee: Farid Zaripov
• Reporter: Martin Sebor
• Component/s: 22. Localization
• Fix Version/s: 4.2.1

Description

The 22.locale.num.put.cpp fails 4 assertions with the head of trunk (see below). The test passes at 100% with 4.2.0. I reduced the failures to the following test case:

$ cat t.cpp && gmake t && ./t 
#include <cassert>
#include <cstdio>
#include <math.h>
#include <sstream>

int main ()
{
    const long double x = 1.0e+4932L;
    assert (isnormal (x));

    std::ostringstream strm;
    strm << x;

    char buf [40];
    std::sprintf (buf, "%Lg", x);

    if (strm.str () == buf)
        return 0;

    std::printf ("%s != %s\n", strm.str ().c_str (), buf);

    assert (0);
}
aCC -c  -D_RWSTDDEBUG   -mt -I/amd/devco/sebor/stdcxx-4.2.x/include -I/build/seb
or/stdcxx-4.2.x-aCC-6.16-15D//include -I/amd/devco/sebor/stdcxx-4.2.x/tests/incl
ude  -AA  -g +d  +DD64 +w +W392,655,684,818,819,849 +W2193,2236,2261,2340,2401,2
487 +W4227,4229,4231,4235,4237,4249 +W4255,4272,4284,4285,4286,4296,4297 +W3348 
  t.cpp
"/amd/devco/sebor/stdcxx-4.2.x/include/loc/_ctype.h", line 545: warning #2186-D:
 
          pointless comparison of unsigned integer with zero
      if (   0 <= __c
               ^

"/amd/devco/sebor/stdcxx-4.2.x/include/loc/_ctype.h", line 555: warning #2186-D:
 
          pointless comparison of unsigned integer with zero
      if (   0 <= __c
               ^

aCC t.o -o t -L/build/sebor/stdcxx-4.2.x-aCC-6.16-15D//rwtest -lrwtest15D -AA +n
ostl -Wl,+s   -mt +DD64 -L/build/sebor/stdcxx-4.2.x-aCC-6.16-15D//lib  -Wl,+b/bu
ild/sebor/stdcxx-4.2.x-aCC-6.16-15D//lib:/build/sebor/stdcxx-4.2.x-aCC-6.16-15D/
/rwtest -lstd15D  -lm 
inf != 1e+4932
Assertion failed: 0, file t.cpp, line 22
ABORT instruction (core dumped)

The output of 22.locale.num.put follows:

$ ./22.locale.num.put -q
# ASSERTION (S7) (4 lines):
# TEXT: line 377: num_put<char>::put (..., long double = 1.18973e+4932) wrote "inf", expected "1.18973149535723176508575932662800702e+4932", flags = fmtflags(0), precision = 36
# CLAUSE: lib.locale.num.put
# LINE: 1854

# ASSERTION (S7) (4 lines):
# TEXT: line 377: num_put<char>::put (..., long double = 1e+4932) wrote "inf", expected { "1", '0' <repeats 34 times>, "1", '0' <repeats 4897 times> }, flags = fixed, precision = 0
# CLAUSE: lib.locale.num.put
# LINE: 1958

# ASSERTION (S7) (4 lines):
# TEXT: line 377: num_put<wchar_t>::put (..., long double = 1.18973e+4932) wrote L"inf", expected "1.18973149535723176508575932662800702e+4932", flags = fmtflags(0), precision = 36
# CLAUSE: lib.locale.num.put
# LINE: 1854

# ASSERTION (S7) (4 lines):
# TEXT: line 377: num_put<wchar_t>::put (..., long double = 1e+4932) wrote L"inf", expected { "1", '0' <repeats 34 times>, "1", '0' <repeats 4897 times> }, flags = fixed, precision = 0
# CLAUSE: lib.locale.num.put
# LINE: 1958

# +-----------------------+----------+----------+----------+
# | DIAGNOSTIC            |  ACTIVE  |   TOTAL  | INACTIVE |
# +-----------------------+----------+----------+----------+
# | (S1) INFO             |       96 |       96 |       0% |
# | (S2) NOTE             |        1 |        1 |       0% |
# | (S5) WARNING          |        2 |        2 |       0% |
# | (S7) ASSERTION        |        4 |     1671 |      99% |
# +-----------------------+----------+----------+----------+

Comments

[Martin Sebor]

Reverting r646396 fixes the problem.
Farid please take a look at this (FWIW, it looks like we're missing long double overloads for the new functions added in the change).

[Farid Zaripov]

The patch is attached.

ChangeLog:

	STDCXX-880
	* src/num_put.cpp (max_double_t): New typedef for defining the
	floating-point type with max precision. Implement generic functions
	that not using the math library functions, instead of just "return false".
	(__rw_isfinite): Define function for max_double_t type instead of double.
	(__rw_signbit): Ditto.
	(__rw_isinf): Ditto.
	(__rw_isnan): Ditto.
	(__rw_isqnan): Ditto.
	(__rw_issnan): Ditto.
	(__rw_fmat_infinite): Ditto.

[Farid Zaripov]

I'm only not sure about Solaris OS. I don't see in the documentation at docs.sun.com that there are present fpclass(), finite(), ... functions for long double type.
But now we need the only finite() implemented for long double. Other functions can be implemented for double type (I believe that inf and nan values are correctly converting from long double into double).

[Martin Sebor]

I'm not sure about converting double to max_double_t. Is it safe in terms of traps and floating point exceptions? Is it efficient? Why can't we just have three overloads of these functions for each of the three floating point types?

In any event, it's far too late to be changing the implementation of the double overloads to long double, or to be adding non-trivial implementations the generic functions for 4.2.1. Besides, the new generic functions are unsafe. They must avoid traps, so any kind of floating point arithmetic is out of the question. On platforms with limited support for these kinds of things (e.g., HP-UX on PA-RISC where the arguments must be float or double – see for example the fpclassify(3m) HP-UX man page) we'll have to implement our own versions by examining the bit patterns ourselves.

For 4.2.1, we need the simplest, safest possible fix, or we need to revert the patch.

[Farid Zaripov]

Fixed in trunk thus http://svn.apache.org/viewvc?rev=650902&view=rev and http://svn.apache.org/viewvc?rev=650933&view=rev

Will be closed after merging the chanbges in 4.2.x branch.

[Farid Zaripov]

Merged in 4.2.x branch thus: http://svn.apache.org/viewvc?rev=651034&view=rev