[tree-ssa] Eliminating unnecessary variables during out-of-ssa

[Andrew MacLeod]

On Mon, 2003-05-12 at 14:43, Daniel Berlin wrote:
> > Currently, I dont think we need anything that has a derefernced 
> > variable
> > as a PHI_RESULT.
> 
> For SSAPRE store movement and load movement, I need it.
> 
> In particular, i need the vdefs for load movement, and the uses/phi 
> results for store movement.
> 
> Otherwise, i can't tell when a load is possibly killed without 
> calculating the exact same info (in the case of vdefs), or the possible 
> insertion points for the store (that's what the PHI's give us).
> 
> There are other optimizations i'm working on (loop related) that need 
> this info, too.
> 
> Just because our basic opts don't need it, doesn't mean you should 
> remove it.
> Can't you just ignore it in out-of-ssa, rather than remove it 
> altogether?
> 

Of course you need them for those optimizations, otherwise we wouldn't
generate them in the first place.

After we go out of SSA form, we no longer have an SSA form, so no one
needs these things. All PHI nodes are deleted as part of the process. 

I was asking if anyone could think of a reason why we might need to keep
them during the out of SSA translation phase. I'm trying to identify a
way of reducing the number of PHI nodes that I have to look at, thus
reducing the number of variables which have to be examined during the
entire phase. 

Andrew