On Fri, May 27, 2016 at 07:58:06PM +0300, Yury Norov wrote:

So, we have 3 options for now:
1. Clear top halves in entry.S which means we pass off_t as a pair.
   The cost is performance (didn't measure it yet and doubt about it
   makes serious impact). The advantage is simplicity and unification with
   aarch32, as I mentioned above. And David likes it. And it mininizes
   the amount of changes on glibc side.
2. Clear top halves in in separated file hosted wrappers.
3. Clear top halves in I-cache and tail optimization friendly in-site wrappers.

2 and 3 are the same from ABI point of view.

2 is the worst for me as it is the most complex in implementation and 
I-cache and tail optimization non-friendly. But Heiko likes it.
 
3 is what Catalin is talking about, and it was my initial approach.
Though I didn't made compiler to do tail optimization, I think we can
do it.

I don't fully understand the difference between 2 and 3. My comment was
more around annotating the wrappers in (2) with __naked to no longer
generate function prologue/epilogue. They would still be in a separate
kernel/compat_wrapper.c file.

I can't figure out how you would have in-place wrappers for all
syscalls. You can indeed handle the current COMPAT_SYSCALL_DEFINE via
__SC_DELOUSE (and penalising the AArch32/compat support slightly) but
there is no solution for native SYSCALL_DEFINE functions to do it
in-place.

But 2 is what we have now. And I'd choose it. We'll never get ilp32 done
if will roll back previously agreed decisions again and again.

I would rather roll back a decision than going ahead with a wrong one.
Note that this is *ABI*, not a driver that you can fix upstream later.

Since yesterday, I realised that (2) requires further annotations and
wrapping for the native and compat syscalls used by ILP32 just to cope
with pointers. Also given davem's comments, (1) starts to look a bit
more appealing (I don't like reverting such decisions either, I'd have
to review the code again and again).

-- 
Catalin