On 8/10/24 8:31 AM, Robin Dapp wrote:
A bit of bikeshedding:
While it's obviously a bug, I'm not really sure it's useful to truncate before
emitting the widening shift. Do we save an instruction vs. the regular
non-widening shift by doing so?
At least for the test you added, there is no different before/after
Pan's patch. If we disable the pattern entirely, then yes it gets
notably worse -- instead of a single vwsll we end up two extensions, the
vsll, and a vset+vncvt for the final conversion. So yea, the pattern
still seems to be worth having :-0
Jeff
