🚀 go-pugleaf

	[Possible gcc backend bug for RS/6000 architecture.  libg++-2.3
does not pass its self-tests (make check) and calls abort because of assert
in tFile.C:t11().]

	I have tracked down the bug in formatted I/O further by comparing
the steps that the Sun version takes versus the RS/6000 version.  For
t11(), cout.form(...), the Sun version calls __outfloat() from
streambuf::vform() with precision=2 and fill=' ' while the RS/6000 version
uses precision=6 and fill=GARBAGE.  fill is set correctly before the call
but is corrupted in the call itself (it is wrong in the called subroutine
itself so this presumably is not a gdb-4.7 bug).  precision is set
differently because the Sun and RS/6000 take slightly different paths
through the code for the '.' case of the format.  The Sun executes:


                        while (is_digit(ch)) {
                                n = 10 * n + to_digit(ch);
                                ch = *fmt++;
                        }

but the RS/6000 skips that fragment and proceeds directly to:

                        prec = n < 0 ? -1 : n;
                        goto reswitch;

so that it ignores the precision specification.  This appears to be an
error in the generated code.  I have not tried to check the assembly output
itself.  As I stated in the original report, I am using gcc-2.3.2 for both
the Sun and RS/6000 tests.  gdb call and traceback information follows.

					David
(Disclaimer: I do not speak for IBM and it does not speak for me.)

	Sun call and traceback:

__outfloat (value=5.5, sb=0xf7fff150, type=102, width=0, precision=2, flags=0,
    sign_mode=0 '\000', fill=32 ' ') at outfloat.C:31
31          int count = 0;
(gdb) where
#0  __outfloat (value=5.5, sb=0xf7fff150, type=102, width=0, precision=2,
    flags=0, sign_mode=0 '\000', fill=32 ' ') at outfloat.C:31
#1  0xb8d8 in streambuf::vform (this=0xf7fff150, fmt0=0x0,
    ap=0xf7fff90c "@\031") at sbufvform.C:339
#2  0xb40c in help_vform (sb=0x1a9a4, fmt0=0x4fb0 "%.2f+%.2f = %4.3e\n%n",
    ap=0xf7fff904 "@\026") at sbufvform.C:86
#3  0xb4bc in streambuf::vform (this=0x1a9a4,
    fmt0=0x4fb0 "%.2f+%.2f = %4.3e\n%n", ap=0xf7fff904 "@\026")
    at sbufvform.C:200
#4  0xa6dc in ostream::form (this=0x1a3bc,
    format=0x4fb0 "%.2f+%.2f = %4.3e\n%n") at iostream.C:638
#5  0x5060 in t11 () at tFile.C:423
#6  0x52e0 in main (argc=0, argv=0xf7fffa34) at tFile.C:460
(gdb)

	RS/6000 call and traceback:

__outfloat (value=5.5, sb=0x2ff7f778, type=102 'f', width=0, precision=6,
    flags=0, sign_mode=0 '\000', fill=536870912) at outfloat.C:31
31          int count = 0;
(gdb) where
#0  __outfloat (value=5.5, sb=0x2ff7f778, type=102 'f', width=0, precision=6,
    flags=0, sign_mode=0 '\000', fill=536870912) at outfloat.C:31
#1  0x10013270 in streambuf::vform (this=<unknown struct>,
    fmt0=0x10018a58 "%.2f+%.2f = %4.3e\n%n", ap=0x2ff7fa28 "@\031")
    at sbufvform.C:339
#2  0x10012e1c in help_vform (sb=<unknown struct>,
    fmt0=0x10018a58 "%.2f+%.2f = %4.3e\n%n", ap=0x2ff7fa20 "@\026")
    at sbufvform.C:86
#3  0x10012eac in streambuf::vform (this=<unknown struct>,
    fmt0=0x10018a58 "%.2f+%.2f = %4.3e\n%n", ap=0x2ff7fa20 "@\026")
    at sbufvform.C:200
#4  0x10006ed4 in ostream::form (this=0x20051fa4,
    format=0x10018a58 "%.2f+%.2f = %4.3e\n%n") at streambuf.h:5
#5  0x10002a6c in t11 () at tFile.C:423
#6  0x10002c80 in main (argc=1, argv=0x2ff7faf0) at tFile.C:460
#7  0x1000042c in __start ()
(gdb)
===============================================================================
David Edelsohn                                      T.J. Watson Research Center
c$dje@watson.ibm.com                                P.O. Box 218
+1 914 945 3011 (Tieline 862)                       Yorktown Heights, NY 10598
"The Church doesn't have problems with sex; the world does" -- Vatican official
"A good theory should fit on a T-shirt" -- Astronomer at Jan 1992 AAS meeting