The following testcase is an example of code used in a glibc testcase. I'm
trying hard to shake out the bugs in the glibc testsuite for debian, and one
testsuite failure looks like a compiler issue.
The expected behaviour is for the testcase to print the raw IEEE754 value of
-NAN.
The observed behaviour, when -DALT is on the command line, is that the testcase
prints the incorrect raw value e.g. NAN.
GCC 4.4.3 in debian doesn't compile this code correctly.
cat >> tst-mul-nan.c <<EOF
#include <stdio.h>
#include <math.h>
#ifdef ALT
volatile double nanval;
#else
#define nanval NAN
#endif
int
main ()
{
#ifdef ALT
nanval = NAN;
#endif
printf ("0x%llx\n", -nanval);
return 0;
}
EOF
gcc -g3 -O0 -save-temps -o test-mul-nan-OK test-mul-nan.c; ./test-mul-nan-OK
0xfff7ffffffffffff
This is the correct result e.g. -NAN. In the correct case the compiler has
already computer -NAN and it's loaded directly from the local symbol e.g.
.LC1:
.word -524289
.word -1
This is the case that is not working correctly:
gcc -g3 -O0 -save-temps -DALT -o test-mul-nan-NG test-mul-nan.c;
./test-mul-nan-NG
0x7ff7ffffffffffff
That result is not -NAN, it is NAN. This is incorrect.
In the incorrect compilation the compiler loads NAN from a local constant:
.LC0:
.word 2146959359
.word -1
This is 0x7ff7ffffffffffff e.g. NAN.
Then it loads a 64-bit double -1.0.
.LC2:
.word -1074790400
.word 0
In the incorrect case the compiler tries to multiply -1.0 by NAN to get a
result of -NAN.
addil LR'nanval-$global$,%r27
copy %r1,%r19
ldo RR'nanval-$global$(%r19),%r19
fldds 0(%r19),%fr23
ldil LR'.LC2,%r19
ldo RR'.LC2(%r19),%r19
fldds 0(%r19),%fr22
fmpy,dbl %fr23,%fr22,%fr22
This is not going to work because -1.0 times NAN is NAN, and the sign of the
input will be ignored.
See: PA-RISC 2.0 Architecture, Floating Coprocessor 8-23 "Operations With
NaNs", and 8-24 "Sign Bit" can be referenced for information on NANs.
After the multiplication fr22 still contains NAN, and that is what is printed
instead of the expected result of -NAN.
John David Anglin indicates:
This is a bug.
The code should xor the sign bit when doing negation. The existing code
doesn't work for NANs. I'll try to fix negdf2 and negsf2.
The problem is PA1.1 doesn't have a fneg instruction, so the above takes a bit
of work to implement. You will get the correct result if you specify
-march=2.0.
Unfortunately I can't use -march=2.0 in this case, the C library has to work on
PA1.1 class systems.
--
Summary: Multiplying -1 by NaN is not valid.
Product: gcc
Version: 4.4.3
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: carlos at codesourcery dot com
GCC build triplet: hppa-linux-gnu
GCC host triplet: hppa-linux-gnu
GCC target triplet: hppa-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44261
