Currently the accessor implementation stores references to functions that encode/decode normalized vertex values. When cloning an accessor, those functions held references to their previous scope, and looked up the previous accessor's normalization state.
This PR changes and simplifies the implementation to avoid storing any unnecessary state. The previous implementation was a premature optimization that (as shown by the benchmarks below) did little or nothing. There might be a small regression in join() performance, but that's ongoing work for #1253.
Currently the accessor implementation stores references to functions that encode/decode normalized vertex values. When cloning an accessor, those functions held references to their previous scope, and looked up the previous accessor's normalization state.
This PR changes and simplifies the implementation to avoid storing any unnecessary state. The previous implementation was a premature optimization that (as shown by the benchmarks below) did little or nothing. There might be a small regression in
join()
performance, but that's ongoing work for #1253.Changes
Benchmark - before
Benchmark - after