Re: [tree-ssa] Supefluous phi removal

[Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>]

Hello,

> + 	      if (TREE_CODE (stmt) != PHI_NODE
> + 		  /* Prevent copy propagation into arguments of phi nodes
> + 		     corresponding to abnormal edges, by setting their
> + 		     value to top.  */
> + 		  || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (t))
> + 		{
> + 		  eq_to[aver] = t;
> + 		  raise_value (phi, t, eq_to);
> 
> 
> So what's not terribly obvious is that the assignment to eq_to here is
> preventing copy propagation of the PHI argument, but that the PHI result
> is still eligible for copy propagation.
> 
> Instead I would write something like this
> 
>   /* Prevent copy propagation of SSA_NAMES which occur in abnormal PHI
>      nodes.  */
>   if (SSA_NAME_OCCURS_IN_ABORMAL_PHI (t))
>     eq_to[aver] = t;
> 
>   /* Raise the result of PHI by joining it to T.  */
>   raise_value (phi, t, eq_to);
> 
> It seems to me that you always want to call raise_value, regardless of
> whether or not T is set from a normal statement or a PHI node.

no -- if it is a normal phi node, it either has assigned value of
bottom, or raise_value was called for it already recursively, so there
is no point in calling raise_value for it again.

We must only call raise_value if we need to change value for the ssa
name from some reason, which may be because

-- the ssa name is defined in the real statement
-- we cannot replace its uses (in our case this only happens when it is an
   argument on abnormal edge)
-- value of some argument of the phi node it is defined in has changed

Zdenek