Re: [Qemu-devel] [PATCH 06/11] target-i386: use floatx80 constants in helper_fld*_ST0()

["Andreas Färber" <andreas.faerber@web.de>]

Am 21.05.11 11:35, schrieb Andreas Färber:
> Am 20.05.2011 um 12:32 schrieb Peter Maydell:
>
>> On 15 May 2011 15:13, Aurelien Jarno <aurelien@aurel32.net> wrote:
>>> Instead of using a table which doesn't correspond to anything from
>>> physical in the CPU, use directly the constants in helper_fld*_ST0().
>>
>> Actually I rather suspect there is effectively a table in the CPU
>> indexed by the last 3 bits of the FLD* opcode... It would be
>> possible to implement this group of insns in QEMU with a single
>> helper function that took the index into the array, but since the
>> array seems to be causing weird compilation problems we might
>> as well stick with the lots-of-helpers approach, at which point
>> this is a sensible cleanup.
>
> In OpenBIOS we once ran into a similar error on ppc64 where a cast 
> would've resulted in the truncation of a static pointer ... could this 
> be an alignment issue here, that it's being truncated by the floatx80 
> cast? I tried using __attribute__((packed)) on the floatx80 type 
> without luck.
> Or maybe the constant width is being handled weird, with LL being 128 
> bits rather than the expected 64 bits? ;)

Some more info:

The issue only pops up with -std=c99 or -std=gnu99, with both gcc 3.4.3 
and 4.3.2. That's reproducible on Darwin gcc 4.0.1 and 4.2 as well.

The following trivial sample program triggers it, there's no magic 
Solaris headers involved:

#include <inttypes.h>
// on OpenSolaris:
// typedef unsigned short uint16_t;
// typedef unsigned long long uint64_t;

typedef struct {
     uint64_t low;
     uint16_t high;
} floatx80;
#define make_floatx80(exp, mant) ((floatx80) { mant, exp })

#define floatx80_l2t make_floatx80( 0x4000, 0xd49a784bcd1b8afeLL )

static const floatx80 mine[1] = {
     floatx80_l2t,
};

int main(void)
{
}

Any thoughts?

Andreas