On 9/26/26 00:27, Mark Cave-Ayland wrote:
Always add the segment base when an override prefix is specified in segmented
mode. Otherwise the generated address is calculated incorrectly when an override
prefix is specified and HF_ADDSEG_MASK is not set.
Fixes: 55a15eb3504 ("target/i386/tcg: simplify effective address calculation")
Signed-off-by: Mark Cave-Ayland <[email protected]>
---
target/i386/tcg/translate.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
Paolo/Richard: I'm not sure if this patch is exactly correct, however it fixes
a regression when launching the OpenWatcom installer under MS-DOS. Current
git master aborts the installer with a GPF error in the guest caused by this
instruction:
0x0041e0b5: 66 2e 8e 1d 0c 00 00 00 movw %cs:0xc, %ds
where def_seg == 3, ovr_seg == 1 and ADDSEG(s) is 0. Without this fix the target
address is calculated as 0xc instead of 0x41e00c.
It isn't the root cause, no. Here we go:
x86_update_hflags:
hflags |= ((env->segs[R_DS].base | env->segs[R_ES].base |
env->segs[R_SS].base) != 0) << HF_ADDSEG_SHIFT;
cpu_x86_load_seg_cache:
new_hflags |= ((env->segs[R_DS].base |
env->segs[R_ES].base |
env->segs[R_SS].base) != 0) <<
HF_ADDSEG_SHIFT;
So we've been incorrectly including CS in ADDSEG since 55a15eb3504.
Thus
- has_base = ovr_seg >= R_FS || (ovr_seg >= 0 && ADDSEG(s));
switch (ovr_seg) {
case R_DS:
case R_ES:
case R_SS:
has_base = ADDSEG(s);
break;
case R_CS:
case R_FS:
case R_GS:
has_base = true;
break;
default:
assert(ovr_seg < 0);
has_base = false;
break;
}
might be the most legible way to fix the above expression.
r~
