Le 03/11/2016 à 20:47, Richard Henderson a écrit :
> On 11/02/2016 03:15 PM, Laurent Vivier wrote:
>> + if ((insn & 7) + 8 == i &&
>> + m68k_feature(s->env, M68K_FEATURE_EXT_FULL)) {
>> + /* M68020+: if the addressing register is the
>> + * register moved to memory, the value written
>> + * is the initial value decremented by the
>> size of
>> + * the operation
>> + * M68000/M68010: the value is the initial value
>> + */
>> + TCGv tmp = tcg_temp_new();
>> + tcg_gen_sub_i32(tmp, mreg(i), incr);
>> + gen_store(s, opsize, addr, tmp);
>> + tcg_temp_free(tmp);
>
> This doesn't look right. Is the value stored the intermediate value of
> the decremented register, or the final value? What you're storing is
> reg-4, which is neither of these things.
>
> I could see, maybe, that reg-4 might well turn out to be the right value
> for
>
> movem {a0-a7}, (sp)-
>
> since sp == a7, and therefore stored first. But I question that's the
> correct result for
>
> movem {a0-a7}, (a1)-
>
> If it's the incremental value, then you can just store "addr" and you
> don't need a temp. If it's the final value, then you can compute
>
> tcg_gen_subi_i32(tmp, AREG(insn, 0), ctpop32(mask) * 4);
>
As it was not clear for me, I have written a test to see what was the
good value.
my test program is:
top:
.space 64,0
stack:
.text
.globl _start
_start:
lea stack,%a4
lea 1,%a0
lea 2,%a1
lea 3,%a2
lea 4,%a3
lea 5,%a5
lea 6,%a6
moveq.l #8, %d0
moveq.l #9, %d1
moveq.l #10, %d2
moveq.l #11, %d3
moveq.l #12, %d4
moveq.l #13, %d5
moveq.l #14, %d6
moveq.l #15, %d7
movem.l %a0-%a7/%d0-%d7,-(%a4)
on a real 68040:
initial value of A4 is 0x800020ec
final value of A4 is 0x800020ac
(gdb) x/15x 0x800020ac
0x800020ac: 0x00000008 0x00000009 0x0000000a 0x0000000b
0x800020bc: 0x0000000c 0x0000000d 0x0000000e 0x0000000f
0x800020cc: 0x00000001 0x00000002 0x00000003 0x00000004
0x800020dc: 0x800020e8 0x00000005 0x00000006
Stored value is thus 0x800020e8 so this is initial value - 4.
[I have tried the same test with a1, for the same result]
Laurent
