Re: [tree-ssa]: out-of-ssa breaks if something adds a pred

[Diego Novillo <dnovillo at redhat dot com>]

On Fri, 2003-06-06 at 15:52, Daniel Berlin wrote:
> When something ends up adding a predecessor block to a block with phi's 
> in it, out-of-ssa breaks.
> This is because it expects all the phi's to have the same number of 
> args/edges as preds, but phi's created *before* the pred was added have 
> one less argument/edge.
> 
Then whoever created the new edge should have added a
> 
> IE into-ssa makes the following:
> BLOCK 13, preds 2, 1
> a_8 = PHI<a_9 (1), a_10 (2)>
> 
> Later, an edge insertion causes the following (due to creation of a new 
> block):
> BLOCK 13, preds 2, 1, 15
> a_8 = PHI<a_9 (1), a_10 (2)>
> 
Wait, why does it matter that a new block is being inserted?  When we
create a new block to break critical edges, we should not be altering
the number of incoming edges to the target block.

	B1	  B2
	| \	  /
	|  \E1   /
	|   \   /
	|    B13
	|
	B3

Presumably, you are inserting in edge E1?  If so, the new block B15
should not create a *third* incoming block to B13:

	B1	   B2
	| \	   /
	|  \      /
	|  B15   /
	|    \  /
	|     \/
	|    B13
	|
	B3

The PHI nodes at B13 should just be rewritten to change all the
arguments that were coming from B1, to come from B15.  I think this is
something the edge insertion commit routine can do automatically. 
Andrew?


Diego.