On Tue, 12 Aug 2025 05:58:47 GMT, Xiaohong Gong <xg...@openjdk.org> wrote:
>>> Q1: Is it possible that just passing `origin->get_con()` to >>> `VectorSliceNode` in case there are architectures that need it directly? >>> Or, maybe we'd better add comment telling that the origin passed to >>> `VectorSliceNode` is adjust to bytes. >>> >> >> Added comments. >> >>> Q2: If `origin` is not a constant, and there is an architecture that >>> support the index as a variable, will the code crash here? Can we just >>> limit the `origin` to a constant for this intrinsifaction in this PR? We >>> can consider to extend it to variable in case any architecture has such a >>> requirement. WDYT? >> >> Currently, inline expander only supports constant origin. I have added a >> check to fail intrinsification and inline fallback using the hybrid call >> generator. > > Thanks for your updating! So maybe the matcher function > `supports_vector_slice_with_non_constant_index()` could also be removed > totally? Yes, idea here is just to intrinsify a perticular scenario where slice index is a constant value and not burden the inline expander with full-blown intrinsification of all possible control paths without impacting the performance. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/24104#discussion_r2271958737
