Michael Matz wrote:
Hi,
On Fri, 13 Feb 2009, Paolo Bonzini wrote:
We'd want to encode [early insn alternative selection] information in
the conflict graph so that IRA would allocate registers so as to fit
the constraints of the early insn alternative selection. Right? In
the case where the graph is uncolorable, do we allow IRA to override
the alternative selection, or do we insert copies to simplify the
conflict graph or some mixture of both?
If the initial alternative selection was done cleverly (like chose the
alternatives allowing the largest register sets which don't immediately
create conflicting demands for a pseudo register) the opportunities for
making an uncolorable graph colorable by chosing another alternative will
be very small. This can only happen if that new alternative somehow
allows for the uncolorable node a completely new set of register (like say
float instead of integer regs), which would mean also selecting other
alternatives for all instructions where this pseudo also is used.
So it's not impossible, but I think it would happen relatively seldom that
changing the alternatives improves the situation.
Of course. However, we might want to pick a narrower class if it has a
smaller cost. The mn103 targets come to mind. In general you're better
off with d0-d3/a0-a3 as they're the cheapest (cost & space). However,
you've got some extended registers which can be used just like
d0-d3/a0-a3, but which are more expensive (but still cheaper than memory).
Jeff
