On Thu, Nov 8, 2018 at 4:32 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Thu, Nov 08, 2018 at 03:35:02PM -0800, Dave Hansen wrote:
> > On 11/8/18 2:00 PM, Matthew Wilcox wrote:
> > > struct a {
> > > char c;
> > > struct b b;
> > > };
> > >
> > > we want struct b to start at offset 8, but with __packed, it will start
> > > at offset 1.
> >
> > You're talking about how we want the struct laid out in memory if we
> > have control over the layout. I'm talking about what happens if
> > something *else* tells us the layout, like a hardware specification
> > which is what is in play with the XSAVE instruction dictated layout
> > that's in question here.
> >
> > What I'm concerned about is a structure like this:
> >
> > struct foo {
> > u32 i1;
> > u64 i2;
> > };
> >
> > If we leave that to natural alignment, we end up with a 16-byte
> > structure laid out like this:
> >
> > 0-3 i1
> > 3-8 alignment gap
> > 8-15 i2
>
> I know you actually meant:
>
> 0-3 i1
> 4-7 pad
> 8-15 i2
>
> > Which isn't what we want. We want a 12-byte structure, laid out like this:
> >
> > 0-3 i1
> > 4-11 i2
> >
> > Which we get with:
> >
> > struct foo {
> > u32 i1;
> > u64 i2;
> > } __packed;
>
> But we _also_ get pessimised accesses to i1 and i2. Because gcc can't
> rely on struct foo being aligned to a 4 or even 8 byte boundary (it
> might be embedded in "struct a" from above).
>
In the event we end up with a hardware structure that has
not-really-aligned elements, I suspect we can ask gcc for a new
extension to help. Or maybe some hack like:
struct foo {
u32 i1;
struct {
u64 i2;
} __attribute__((packed));
};
would do the trick.
