Re: [avr-gcc-list] >4.5.1 better than this at register-structure (xmega) access?

[David Brown]

On 10/10/2012 23:13, Erik Walthinsen wrote:
> On 10/10/2012 01:43 PM, Georg-Johann Lay wrote:
>  > Such a bug has never been reported to the GCC bug tracker.
> Well then, I guess I need to file one.  I just have no idea how to
> actually fix it, and no time to do so at this point ;-(
>
> > If nobody found it important enough to file a problem report for over 3
> > years (4.5 was released early 2009) I'd guess this is simply not an
> > important issue?
> It's not a bug per se but a missing optimization, so it's entirely
> possible nobody's cared enough to check.  I'm a) doing some extremely
> time-critical ISR routines, and b) trying to actually make use of the
> struct convenience.

You should file a "missed optimisation" bug on the issue - then it will 
not be forgotten or repeated.  Missed optimisation bugs don't get as 
high priority as wrong code bugs, or adding support for new devices, but 
they /do/ get worked on whenever someone has the time (and can figure 
out a way to improve the issue).  In particular, when there are other 
changes in the code in related areas, developers look through the list 
of outstanding bugs to see if there are other issues that could be fixed 
at the same time.

In this case, the structure Z+offset access is the most space-efficient 
if you are making three or more accesses to the same structure, and for 
two accesses it takes the same space and as using the direct access.  So 
using the Z+offset is not faster than direct access.

The exception here is if it is possible to use Z or Z+ modes, which are 
single-cycle on the XMega.

So ideally, the compiler should use direct access when there are two or 
less accesses to the same structure, and Z+q for three or more accesses 
(in -Os) or stick with direct access (in -O2).  If it first uses Z+q, 
then Z should be loaded with the lowest accessed address to allow Z mode 
for a slight speed gain.  And multiple adjacent access can be done using 
Z+ for faster access.


That's the theory, anyway, as far as I understand it - but I don't know 
how it could be implemented in practice...