RFR: 8315361: C2: Create a superclass of SuperWord [v2]

[Fei Gao]

On Mon, 30 Oct 2023 11:01:29 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

> My question: should we not also move the `velt` (vector element type) and maybe even the dependence graph parts? If I remember right, then the post-loop-vectorizer also needs some way to determine element type (`_elem_bt` [here](https://github.com/openjdk/jdk/pull/14581)), but not sure about dependence graph. And also the memory slices may be helpful for all vectorizers. We could also have more classes that have different parts of these features (eg. `Vectorizer` with only basic functionality, and then a `VectorizerWithDependencyGraph`).

Hi @eme64, thanks for your review!

Yes, and both SuperWord and post-loop-vectorizer need to determine vector element type. There are still some deficiencies in current [`compute_vector_element_type()`](https://github.com/openjdk/jdk/blob/b3fec6b5f32c338ae1a84dd20bdcbd3d9b7186f3/src/hotspot/share/opto/superword.cpp#L3350) of SuperWord. For example, some loop cases of subword types, including byte and short, can't be recognized very well. See the description in https://github.com/openjdk/jdk/pull/7954. Maybe the main idea of [`find_vector_element_types()`](https://github.com/openjdk/jdk/blob/bd1b939b21a23675fd072b91d4eab538ff6d2a7d/src/hotspot/share/opto/vmaskloop.cpp#L318) proposed in Pengfei’s post-loop-vectorizer can be improved to solve it?

About the dependence graph part, SuperWord works better than post-loop-vectorizer, because it can support more cases like reading-forward, after your great refactoring 👍 . But now, post-loop-vectorizer can’t support any cases with index offset.

So, maybe both two parts need more complex refactoring and improving, before they can be shared well by two vectorizers, not like code parts in this patch.

Thanks!

-------------

PR Comment: https://git.openjdk.org/jdk/pull/16353#issuecomment-1788340550