Re: [PATCH] Updating IPCP to SSA

[Razya Ladelsky <RAZYA at il dot ibm dot com>]

Jan Hubicka <jh@suse.cz> wrote on 01/15/2007 04:27:41 PM:

> > 
> > You're right about recognizing PHI nodes exactly how we do it for
> > MODFIY_STMT.
> > The part which I don't get is what you're saying about ADDRESS_EXPR.
> > If the parameter's address is taken, then it is not in ssa form.
> > (I check it by TREE_ADDRESSABLE(param) in ipa_method_compute_modify()) 

> > Should I be checking ADDR_EXPR explicitly?
> 
> What I was referring to is that you can have
> 
>   ssa_var = PHI (&non_ssa_var, ...)
> 
> since ADDR_EXPR of variable is considered to be function wide constant.
> So you simple need to look into operands of PHIs for possible ADDR_EXPRs
> same way as you look into RHS of MODIFY_STMT.
> ADDR_EXPR of SSA_NAME is wrong, so you don't need to worry about it.
> 

Still not sure why I need to explicitly look for ADDR_EXPR.
If the parameter is TREE_ADDRESSABLE, I automatically mark it modified.
If it's not, then its address is not taken anywhere.

> Thank you,
> Honza