Front page | perl.perl5.porters |
Postings from December 2003
I'm a bit confused because I'm not quite sure if this could be a bug in GCC.
I tried building bleadperl with a recent gcc-3.4 snapshot (20031217) using
./Configure -des -Dusedevel -Duse64bitint -Dcc=gcc-3.4 -Doptimize='-g -O3'
Running 'make test', test 67 of op/sprintf failed:
not ok 67 >%hd< >1< >1< >104< # More extensive testing of
It's easily reproducible from the command line:
mhx@r2d2 perl-current-gcc-3.4 $ ./perl -e 'printf"%hd\n",1'
104
What turns the 1 into a 104 is the (short) cast
case 'h': iv = (short)iv; break;
in this part of sv.c (at line 9062):
else {
iv = SvIVx(argsv);
switch (intsize) {
case 'h': iv = (short)iv; break;
default: break;
case 'l': iv = (long)iv; break;
case 'V': break;
#ifdef HAS_QUAD
case 'q': iv = (Quad_t)iv; break;
#endif
}
}
Adding some debug output
else {
iv = SvIVx(argsv);
DEBUG_c(PerlIO_printf(Perl_debug_log, "+++ %"IVdf" (before)\n", iv));
switch (intsize) {
case 'h': iv = (short)iv; break;
default: break;
case 'l': iv = (long)iv; break;
case 'V': break;
#ifdef HAS_QUAD
case 'q': iv = (Quad_t)iv; break;
#endif
}
DEBUG_c(PerlIO_printf(Perl_debug_log, "+++ %"IVdf" (after)\n", iv));
}
gave me:
mhx@r2d2 perl-current-gcc-3.4 $ ./perl -Dc -e 'printf"%hd\n",1'
EXECUTING...
+++ 1 (before)
+++ 104 (after)
104
The following plain C program should be equivalent to the above:
int main(void)
{
long long a = 1;
printf("%lld\n", a);
a = (short) a;
printf("%lld\n", a);
return 0;
}
And indeed, when compiled with 'gcc-3.4 -O3' gives me:
mhx@r2d2 ~/src/c/test $ gcc-3.4 -O3 -o test test.c && ./test
1
2
Not really 104, but at least 2. But interestingly it doesn't matter which of
my various gcc's I'm using. The behaviour is the same for gcc-3.2.3, gcc-3.3.2
and gcc-3.4. It also doesn't matter if I choose -O1, -O2 or -O3. Only if I
choose -O0 or -g, I get the correct, expected result.
This is interesting, because I don't seem to be able to get perl to fail with
e.g. gcc-3.4 -O2 or gcc-3.3.2 -O3.
However, the above seems to me like a bug in GCC when trying to cast long longs
to shorts (or chars, which I've also checked).
BTW, it works when assigning to an int variable before casting to a short:
int main(void)
{
long long a = 1;
int b;
printf("%lld\n", a);
b = a;
a = (short) b;
printf("%lld\n", a);
return 0;
}
If this is a bug in GCC, should there be a workaround for it in perl?
With the attached (very experimental and most probably incomplete) patch,
I'm at least getting "All tests successful" with my gcc-3.4 configuration.
Is this problem real or am I missing something here?
If anyone can confirm that this is a bug, I'll enter a bug report in the
GCC bugzilla.
-- Marcus
Thread Next