I found a problem because the FP emulation of long doubles in the m68k port
doesn't treat -0.0 correctly. A cast from a long double -0.0 to a double results
in a double with the value -2.0.
The relevant function is __truncxfdf2 in gcc/config/m68k/fpgnulib.c. There are
two issues here.
First of all the short-cut test for zero only tested for +0.0, not -0.0.
However that by itself shouldn't have shown up a problem as the more lengthy
code underneath should have still come up with the right answer. The problem is
that it doesn't treat a long double exponent of 0 as special. Without this
fixed, subnormal numbers continue to be broken when converted to doubles.
I believe this is the correct fix, but I'm not sure:
--- fpgnulib.c~ Mon Jan 12 03:55:42 2004
+++ fpgnulib.c Sun Feb 27 04:05:13 2005
@@ -315,8 +315,14 @@
ldl.ld = ld;
/*printf ("xfdf in: %s\n", dumpxf (ld));*/
- if (!ldl.l.upper && !ldl.l.middle && !ldl.l.lower)
- return 0;
+ /* Match +/- 0 and subnormals, which get flushed to 0 */
+ if ( EXPX(ldl) == 0 )
+ {
+ /* Preserve sign */
+ dl.l.upper = SIGNX (ldl);
+ dl.l.lower = 0;
+ return dl.d;
+ }
exp = EXPX (ldl) - EXCESSX + EXCESSD;
/* ??? quick and dirty: keep `exp' sane */
In the above case, both +/- 0 and subnormals cause a correctly signed double 0.0
to be returned. Is this the right fix?
In case it is:
2005-02-27 Jonathan Larmour <[EMAIL PROTECTED]>
* config/m68k/fpgnulib.c (__truncxfdf2): Match subnormals as well as
-0.0 in special case test, and preserve sign.
--
Summary: [m68k] long double -> double cast fails with -0.0
Product: gcc
Version: 3.4.3
Status: UNCONFIRMED
Severity: minor
Priority: P3
Component: target
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: jifl-bugzilla at jifvik dot org
CC: gcc-bugs at gcc dot gnu dot org
GCC build triplet: i686-pc-linux-gnu
GCC host triplet: i686-pc-linux-gnu
GCC target triplet: m68k-elf
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=20227
