Re: [PATCH] tree level if-conversion for vectorizer

[Devang Patel <dpatel at apple dot com>]

On Aug 30, 2004, at 3:51 PM, Richard Henderson wrote:

> On Fri, Aug 27, 2004 at 06:41:53PM -0700, Devang Patel wrote:
> > *************** expand_expr_real_1 (tree exp, rtx target
> > *** 7851,7859 ****
> >     	 for side effect only.  If type is void, we must do likewise.  */
> >
> >           if (TREE_ADDRESSABLE (type)
> > !  	  || ignore
> > !  	  || TREE_TYPE (TREE_OPERAND (exp, 1)) == void_type_node
> > !  	  || TREE_TYPE (TREE_OPERAND (exp, 2)) == void_type_node)
> >    	abort ();
> >
> >          /* If we are not to produce a result, we have no target.  
> > Otherwise,
> > --- 7851,7857 ----
> >     	 for side effect only.  If type is void, we must do likewise.  */
> >
> >           if (TREE_ADDRESSABLE (type)
> > ! 	    || ignore)
> >    	abort ();
>
> If you need this, then you havn't addressed my concerns wrt
>
> 	T = cond ? (void)B : A

First of all, I never meant to say "(void)B". I always wanted to say 
"(void)0". Sorry.

Do you agree that when sibling statement is not available, I can not 
convert
 T = cond ? A : (void)0;
in to
 T = A;

Or are you suggesting that if-conversion is not suitable for such loops?

> > *************** is_gimple_stmt (tree t)
> > *** 363,369 ****
> >     switch (code)
> >       {
> >       case BIND_EXPR:
> > -     case COND_EXPR:
> >         /* These are only valid if they're void.  */
> >         return TREE_TYPE (t) == NULL || VOID_TYPE_P (TREE_TYPE (t));
> >
> > --- 363,368 ----
> > *************** is_gimple_stmt (tree t)
> > *** 380,385 ****
> > --- 379,385 ----
> >       case RESX_EXPR:
> >       case PHI_NODE:
> >       case STATEMENT_LIST:
> > +     case COND_EXPR:
> >         /* These are always void.  */
>
> Still in the wrong place.

ok

> > +   /* Create temporary for C in < A = COND_EXPR < C, B>>.
> > +      Vectorizier prefers to have gimple condexpr as part of 
> > COND_EXPR
> > +      cond.  */
> > +   if (is_gimple_condexpr (cond))
> > +     new_cond = unshare_expr (cond);
> > +   else
> > +     new_cond = ifc_temp_var (boolean_type_node, unshare_expr 
> > (cond), bsi, true);
>
> You should be creating one instane of C for the entire basic block.
> Not one per statement.

ok

>   And your comment wrt the vectorizer is
> inaccurate: the vectorizer prefers to have a gimple *value*.  And
> you should say why, not just that it does.

ok

-
Devang