Re: additional refinements to the ra-conflict.

[Michael Matz <matz at suse dot de>]

Hi Kenny,

On Thu, 3 Apr 2008, Kenneth Zadeck wrote:

> This patch adds enough refinements to the existing conflict builder to 
> get rid of the REG_NO_CONFLICT notes from global.

I quickly skimmed over this to see the general approach.  One question: is 
it possible that you reach your goal of allocating two (partly 
conflicting) mw pseudos to the same hardreg by not creating the conflict 
edge at all?  I'm asking because this would generate wrong code sometimes.  
I don't see any changes in global.c which would allow it to assign the 
same hardreg to partly conflicting pseudos (see our mail exchange), except 
if the input conflicts are clear of half-conflicts.

If I got this wrong, you can ignore the following (and instead enlighten 
me why the necessary conflicts are there :) ).  Anyway, consider two mw 
pseudos A and B (with parts A.0, A.1, B.0 and B.1).
Suppose there's only one conflict: A.0 <-> B.1.  Hence a valid allocation 
could assign both to e.g. hardreg 1.

Further, consider how global.c:find_reg works: After having found a free 
reg for the allocno in question, it will mark all covered hardregs as 
taken away in all conflicts of that allocno.

Now, if that conflict is missing, you would also miss to mark some 
hardregs as taken.  For instance if A is allocated first and happens to 
get hardreg 1 (i.e. takes hardregs 1 and 2), and the conflict between A 
and B is invisible for find_reg, hardregs 1 and 2 are available (as far as 
find_reg knows) also for B.  Assuming that B only conflicts with A and 
nothing else, no hardreg will be taken away for it.  So find_reg would be 
free to assign B to hardreg 0.  And that would lead to wrong code as 
register 1 is now used by B.1 and A.0, which is exactly the assumed 
conflict.

Hmm, so either global.c doesn't yet handle partly conflicting mw pseudos 
or it generates wrong code.


Ciao,
Michael.