Tejas Belagod <tbela...@arm.com> writes: > In the currect scheme, when multiple code/mode iterators are in an rtx > pattern, > they are expanded for each combination of iterator values in > apply_iterator_traverse () and a repeated traversal of the expanded rtx's for > each iterator achieves the 'cross-product' of rtx's for iterator value > combinations. This allows for having the place-holder codes/modes in new > rtx's > that are shallow-copied and a subsequent traversal will eventually replace > them > with actual code/mode iterator values. Doing away with the placeholders means > that we have to build a 'cross-product' of iterator values from a database of > iterator values as used in a pattern and make an rtx copy(deep) corresponding > to > each element of the cross-product with the iterator value combinations > substituted in. This is my understanding of the new implementation - am I on > the > right track?
That makes it sound a lot more complicated than it was supposed to be. :-) I'm testing a patch to convert the modes and codes now. Hopefully it'll be easy to add your stuff on top of that. Richard
