Re: conflict matrix vs reload

[Joern Rennecke <amylaar at cygnus dot co dot uk>]

> But those registers would be considered live on both paths.  Thus, on path
> A when we initialize reg44, we would create a conflict with reg46 because
> reg46 was live at the point where reg44 was set.
> 
> Similarly on the path where we initialized reg46 we would have created a
> conflict with reg44 because reg44 was live.

You have convinced me that your code is right.

But I believe your comment is wrong.

You said:
                3. Either X or Y is not evaluted on the path to P
                   (ie it is used uninitialized) and thus the
                   conflict can be ignored.

As I have pointed out, the conflict cannot be ignored in all cases where
either X or Y is not evaluated.  But your points 1. and 2. are already
general enough so that you can describe the remaining cases with:

                3. Neither X or Y are evaluted on the path to P
                   (ie it is used uninitialized) and thus the
                   conflict can be ignored.

As a corollary, we can examine when we'd run into problems with hard
reg conflicts if we removed record_conflicts altogether:
only iff both hard reg and pseudo are not set on the path to P, we won't
record a conflict for a pair of live hard / pseudo reg.
If both are indeed uninitialized, we won't mind any conflict.
A hard reg, however, can be initialized at the start of the function
without being set explicitly by rtl - it is live if it contains a function
argument or return value pointer.
We could inadvertantly clobber the hard register when we see the uninitalized
pseudo die.
So we need to call record_conflicts just once - for the entry block.