alexcrichton commented on issue #5709:
Alternatively instead of modifying clif one possibility would be to split apart band_not during either legalization or during egraph optimizations and then have the backend pattern match both band_not and (band (bnot ..))
cfallin commented on issue #5709:
The general tack we've taken is to decompose ops, so yes, removing the
_not
variants and making backends match when able to useandn
instructions seems like the better approach here. In general the more "normalized" our IR is, the easier it'll be to write rules that apply appropriately. cc @elliottt -- something we can think about next week...
Last updated: Nov 22 2024 at 16:03 UTC