On Wed, Mar 2, 2011 at 11:40 AM, David Miller <davem@xxxxxxxxxxxxx> wrote:
>
> We purposely don't do that "page table entry typedef'd to aggregate" stuff
> on sparc32 because otherwise such values get passed on the stack.
>
> Architectures can currently avoid this bad code generation for the
> page table case, but with this new code they won't be able to avoid
> pass-by-value.
Well, the thing is, on architectures that _can_ pass by value, it
avoids one indirection.
And if you do pass it on stack, then the code generated will be the
same as if we passed a pointer. So sparc may not be able to take
advantage of the optimization, but I don't think the code generation
would be worse.
For the page table case, we don't have that kind of trade-off: the
trade-off there is literally just between "pass in registers, or pass
on stack". Here the trade-off is "pass as an aggregate value or pass
as a pointer to an aggregate value".
That said, since I suspect that the main user will always just get
inlined (ie the helper function that actually fetches the pointers), I
suspect even sparc will see the advantage of the pass-by-value model.
But you might want to actually test the difference and look at the
code generation.
Linus
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>