Hey Ryan,
Sorry for the slow response. The bug here is that DTrace isn't appropriately
sign extending the variable. Here's what we generate:
DIFO 0x831740 returns D type (integer) (size 4)
OFF OPCODE INSTRUCTION
00: 2c050001 ldts DT_VAR(1280), %r1 ! DT_VAR(1280) = "x"
01: 25000002 setx DT_INTEGER[0], %r2 ! 0x0
02: 0f010200 cmp %r1, %r2
03: 18000006 bl 6
04: 0e000001 mov %r0, %r1
05: 11000007 ba 7
06: 25000101 setx DT_INTEGER[1], %r1 ! 0x1
07: 23000001 ret %r1
If we don't specify the type for self->x, but instead cast it to an int we get
the right thing:
DIFO 0x8317d0 returns D type (integer) (size 4)
OFF OPCODE INSTRUCTION
00: 2c050001 ldts DT_VAR(1280), %r1 ! DT_VAR(1280) = "x"
01: 25000002 setx DT_INTEGER[0], %r2 ! 0x20
02: 04010201 sll %r1, %r2, %r1
03: 2e010201 sra %r1, %r2, %r1
04: 25000102 setx DT_INTEGER[1], %r2 ! 0x0
05: 0f010200 cmp %r1, %r2
06: 18000009 bl 9
07: 0e000001 mov %r0, %r1
08: 1100000a ba 10
09: 25000201 setx DT_INTEGER[2], %r1 ! 0x1
10: 23000001 ret %r1
Note that after we do the ldts we sign extend the result (%r1).
It looks like we need an implicit call to dt_cg_typecast() in the case where
the type of the variable is not a uintptr_t. I've filed this bug:
6978322 libdtrace compiler fails to sign extend certain variables
Adam
On Jul 20, 2010, at 7:31 AM, Ryan Johnson wrote:
> Hi all,
>
> I would have expected that x < 0, below, but it isn't (even though it
> theoretically has only 32 bits, as confirmed by the second probe)
>
> self signed int x;
> self signed long y;
> self signed int z;
> BEGIN {self->x = 0xffffffff; self->y = 0xffffffffffffffffl; self->z =
> 0xffffffffffffffffl; }
> BEGIN/self->x & 0xffffffff00000000/ { printf("x has 33+ bits"); }
> BEGIN/self->x < 0/ {printf("x < 0"); }
> BEGIN/self->y < 0/ {printf("y < 0"); }
> BEGIN/self->z < 0/ {printf("z < 0"); }
> BEGIN {exit(0); }
>
> I'm using Solaris 10, so maybe I've got an old-n-buggy version of dtrace, but
> google didn't turn up anything obvious when I searched.
>
> Ideas?
> Ryan
> _______________________________________________
> dtrace-discuss mailing list
> dtrace-discuss@opensolaris.org
--
Adam Leventhal, Fishworks http://blogs.sun.com/ahl
_______________________________________________
dtrace-discuss mailing list
dtrace-discuss@opensolaris.org
