Port-vax archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: Building current

>       size_t i;
>       uint64_t v, nv;

>       for (i = 0; i < 16; i++) {
>               v = ~(uint64_t)0 << i;
>               nv = ~v;
>               printf("%zu: mask %08llx not %08llx\n", i, v, nv);
>       }

> 0: mask ffffffff not ffffffff00000000
> 1: mask 1fffffffe not fffffffe00000001
> 2: mask 3fffffffc not fffffffc00000003

>         ashq %r6,$-1,%r2

Interestingly, NetBSD/vax 1.4T gives me the same behaviour (at least
once I s/%zu/%u/, because 1.4T doesn't have the z prefix), at least on
my simulated VAX.  The .s file says

        ashq -4(fp),$-1,-12(fp)

but the generated instruction stream says

 79 ad fc 8f ff ff ff ff 00 00 00 00 ad f4

which disassembles to

    ashq    -04(fp),$00000000ffffffff,-0c(fp)

(The disassembler uses hex, hence 00000000ffffffff not 4294967295 and
0c not 12.)

When I change it to

        uint64_t v, nv, o;

        o = ~(int64_t)0;
        for (i = 0; i < 16; i++) {
                v = o << i;

then it works, and the assignment to o turns into

        movq $0xffffffffffffffff,-28(fp)

This leads me to suspect that somewhere in a codegen path the original
code is losing track of the width of the literal and thus generating
"$-1" instead of "$0xffffffffffffffff", since it apparently knows how
to use the latter for that value.  If it's the same bug, it must be a
very durable one, though - but if the interfaces between gcc and the
per-arch codegen module have remained relatively stable and the VAX
codegen has thus remained stable, it could be a long-standing bug
there.  (I don't rule out a bug in my simulator masquerading as
something else, though the similarity of behaviour between what I see
and what's reported upthread leads me to suspect that's not it.)

All the above compiles were done with no optimization.  When I turn the
optimizer loose on it, each version has "ashq r6,$-1,r2" and exhibits
the misbehaviour.  But the modifed code, additionally modified by
making o volatile, with full optimization, says
"movq $0xffffffffffffffff,-8(fp)" and works.

/~\ The ASCII                             Mouse
\ / Ribbon Campaign
 X  Against HTML                mouse%rodents-montreal.org@localhost
/ \ Email!           7D C8 61 52 5D E7 2D 39  4E F1 31 3E E8 B3 27 4B

Home | Main Index | Thread Index | Old Index