Debugging info on # spills?

[Mostafa Hagog]

gcc-owner@gcc.gnu.org wrote on 07/04/2004 19:10:07:

> Daniel Berlin wrote:
>
> >
> > On Apr 7, 2004, at 9:37 AM, Michael Matz wrote:
> >
> >> Hi,
> >>
> >> On Tue, 6 Apr 2004, Timothy J. Wood wrote:
> >>
> >>>   Is there some gcc option I can use that would allow me to determine
> >>> how many 'extra' registers would be necessary to allow my function to
> >>> run with zero spilling?
> >>
> >>
> >> Not really, no.  Solving it optimally is NP complete.
> >
> >
> > This actually doesn't appear to be true for all register interference
> > graphs generated by GCC (and other compilers for that matter).
> > See "Register Allocation by Optimal Graph Coloring" for some
statistics.
> > As I emailed Tim in private mail, I also tested this theory on all
> > register interference graphs generated by GCC during bootstrap, make
> > check, and compiling some random large programs i had.
> > They were all 1-perfect as well.
> >
>   Even if you can solve optimaly graph colouring problem, it is probably
> not what Tim asked.
>
>   The optimal graph colouring is just one of the register allocation
> problems (probably not the most complex).  Accoring to the solution a
> register is supposed to be allocated in one hard register or in memory.
>  Most register allocators can place register in different hard-registers
> (or in memory partially) making live range splitting and or web spliting
> .  Also optimal graph coloring problem is simplified in sense of
> register classes.  Gcc has more powerfull register model.  According to
> this model the same pseudo-register  sometimes has to be placed in
> different non-intersected classes, it will require register moving
> sometimes through memory (secondary reloads) if there is no direct
> moving one class register to another one.
>
>    Moreover gcc sometimes will always place pseudo-register in memory
> even if there are enough hard-regiseters because it will be more
> profitable (e.g. an x86 argument pseudo-register passed through stack
> and used only once).
>
>   Also with my point of view the problem is formulated inaccurately.
> What dose `extra' registers mean?   What classes?
>   So even if spills mean spills generated by reload and the problem will
> be formulated more accurately, in general writing code finding minimal
> number of hard registers enough for compiling programs by gcc without
> "spilling"  will be a difficult task.   And moreover the algorithm will
> be slow too.

How about a different but related question.
Is there a way to get the number of spilled registers in a given
compilation?
Since my interest is to check if a given optimization increases/decreases
register pressure, it doesn't matter if this includes the pseudo-registers
that will always be in memory.

Mostafa.