Re: Bug in ra-colorize.c:merge_moves?

[Daniel Berlin <dberlin at dberlin dot org>]

> > Um, look again.
> > You've *decreased* the possibility of graph colorization in this
> > example.
> > The colorability of a graph is *not* related to the number of webs,
> > it's related to the number of conflicts.
> > web1 is now live *longer*, and thus, conflicts with more webs than it
> > did before.
>
> Please point in which place.

Sure.

> > It now conflicts with everything in the ...'s, *decreasing*
> > colorability.
>
> May be you misunderstand me ?
> web1-real-stack-slot is a memory slot.
> Please point in which place I have *decreased* the possibility of
> graph colorization.

Yes, but unless you've made symbol_refs of some new symbol to represent 
the stack space you are using, they are still based off a register (in 
this case, the stack register).
This register now conflicts with the webin *all* of the  ...'s
Before, each new web would conflict with a single ...


To put it in symbolic terms, let's number the ...'s.

> > > i2  use web1-real-stack-slot
> > > ...[1]
> > > i3  use web1-real-stack-slot
> > > ...[2]
> > > i4  use web1-real-stack-slot
> > > ...[3]
> > >  you now have:

web1-real-stack-slot's base register conflicts with all webs in 1, 2, 
and 3.
Let's say each ... has 3 unique webs in it.
Your way now has 6-9 webs (depending on when the base register dies).

Before, we had:
> > > We generate spill insns:
> > >
> > > i21 def web11 web1-pseudo-spill-slot
> > > i2  use web11
> > > ...
> > > i31 def web12 web1-pseudo-spill-slot
> > > i3  use web12
> > > ...
> > > i41 def web13 web1-pseudo-spill-slot
> > > i4  use web13

Here, given the same parameters, we have:
0-3 webs conflicting with web11
0-3 webs conflicting with web12
0-3 webs conflicting with web13

The colorability of each of these is quite different.

In the first case, you can't color any of the webs with 
web1-real-stack-slot's base register. That means 6 -9 webs can't use 
that color
In the second case, web11, web12, and web13 can each be assigned a 
different color, and 0-3 webs conflicting with it can't use that color.

IE in the first case, if our stack register is hard reg 3, 6-9 webs 
can't use hard reg 3.
In the second case, if we choose different colors for each pseudo-spill 
slot, say hard regs 4, 5, and 6, 0-3 webs can't use reg 4, 0-3 webs 
can't use reg 5, and 0-3 webs can't use reg 6.

This is usually much more colorable, and in the worst case, as 
colorable.
In the case the registers all only had 3 open, 6-9 webs now can't be 
colored.
before, if all the registers only had 3 open, only 0-3 webs can't be 
colored.

--Dan