Fix PR tree-optimization/20940

[Daniel Berlin]

On Wed, 2005-03-16 at 11:37 -0500, Diego Novillo wrote:
> On Wed, Mar 16, 2005 at 11:31:53AM -0500, Daniel Berlin wrote:
> 
> > > It is certainly annoying to have to deal with non-GIMPLE code
> > > coming out of fold.  We currently do that when folding some
> > > builtins, but if we allow fold to return whatever expression it
> > > feels like, there will be trouble down the road.
> > 
> > Such as where?
> > 
> In SRA and the builtin folder.  We have to add code to rescan the
> newly inserted statements, add the new symbols to the symbol
> table and put them in SSA form.  And if fold will not guarantee
> GIMPLE output, then we potentially need to do this everywhere
> fold_stmt is called.
> 
> Not impossible.  Annoying.

This was my argument :)
He wasn't too impressed, since we can, of course, abstract it away.

> 
> 
> Diego.