stack height reduction in tree-SSA

[Richard Henderson]

On Wed, Sep 01, 2004 at 10:57:41AM -0700, Mike Stump wrote:
> float f[20];
> double d[10];
> 
> If we used a new alias set that was defined to conflict with the two 
> original sets involved for each of the large variables, I think we 
> could fix the original problem... not that others would not exist.

It turns out that we *can't* fix this without entirely turning off
alias sets.  You don't just have to worry about f[3] and d[6], but
also any random pointer dereference.

There's a test in the testsuite that triggers this.

> For small variables, I think not reusing may be best, though, that's
> really up to the scheduler in general.

Recall that only *addressable* small variables live on the stack.
I disbelieve that these matter that much.  I'm not going to do
anything to distinguish this case.


r~