Ok.

Please check in file #10478 (patch-gnumach-seg-1) and then mark this bug as `fixed' and `closed'.

I'd suggest the following for the ChangeLog entry:

#v+
[current date]  Samuel Thibault  <samuel.thibault@ens-lyon.org>

        * i386/i386/seg.h (struct pseudo_descriptor): Pack structure and move the
        padding field to the end.
        (lgdt): Pass the whole structure to the lgdt assembly command.
        (lidt): Likewise.
#v-

Quoting from file #10478 (patch-gnumach-seg-1):

#v+
[...]
--- i386/i386/seg.h     5 Apr 2001 06:39:20 -0000       1.2
+++ i386/i386/seg.h     7 Aug 2006 03:59:04 -0000
@@ -121,20 +121,20 @@ struct real_gate {
 /* Format of a "pseudo-descriptor", used for loading the IDT and GDT.  */
 struct pseudo_descriptor
 {
-       short pad;
        unsigned short limit;
        unsigned long linear_base;
-};
+       short pad;
+} _attribute_((packed));
[...]
#v-

Why do we need that padding field at all?

GCC gets smarter and smarter...  When seeing this:

struct pseudo_descriptor
{
short pad;
unsigned short limit;
unsigned long linear_base;
}
MACH_INLINE void lgdt(struct pseudo_descriptor *pdesc)
{
        __asm volatile("lgdt %0" : : "m" (pdesc->limit));
}
gdt_init() {
...
{
                struct pseudo_descriptor pdesc;

                pdesc.limit = sizeof(gdt)-1;
                pdesc.linear_base = kvtolin(&gdt);
                lgdt(&pdesc);
        }
}

, it understands that filling the linear_base field is not needed, since we only give the limit field to the asm statement... The result it this:

gcc 4.0:
<gdt_init+102>:      movw   $0x57,0xfffffffa(%ebp)
<gdt_init+108>:      movl   $0xc0000000,0xfffffffc(%ebp)
<gdt_init+115>:      lgdtl  0xfffffffa(%ebp)

gcc 4.1:
<gdt_init+102>:      movw   $0x57,0xfffffffa(%ebp)
<gdt_init+108>:      lgdtl  0xfffffffa(%ebp)

it doesn't fill the linear_base field!

There are two possible fixes, that I'll attach:

- reorder the structure, pack it, and give it as a whole to the asm statement.
- clobber memory in the asm statement.

My preference goes to the first, since it explains to gcc what we exactly want. That supposes the support of the "packed" attribute, however.