Re: [gomp4] rewrite simd clone argument adjustment to avoid regimplification

[Aldy Hernandez <aldyh at redhat dot com>]

On 11/07/13 00:38, Jakub Jelinek wrote:
> On Wed, Nov 06, 2013 at 05:37:03PM -0700, Aldy Hernandez wrote:
> > Hmmm, good point.  I've moved update_stmt and company to the caller,
> > and modified the caller to call regimplify_operands only for
> > GIMPLE_RETURN which is the special case.
>
> Can't you (later) handle that without regimplification too?

Sure!  See attached patch.

But as discussed on IRC, I wonder whether we can do without the 
following in the attached patch:

+		    tree repl = make_ssa_name (TREE_TYPE (retval), NULL);
+		    stmt = gimple_build_assign (repl, retval);
+		    gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
+		    stmt = gimple_build_assign (ref, repl);

...and unconditionally do:

+		  stmt = gimple_build_assign (ref, retval);

...since it seems all the GIMPLE_RETURNs I see can be replaced by 
``retval_array[iter] = whatever'' without creating something non-gimple 
(thus avoiding an SSA variable).

Either way, I'm ok.  Let me know.

Aldy

Attachment: curr
Description: Text document