On Tue, 21 Nov 2023, Richard Biener wrote:
> and this, too, btw. - the DSE actually happens, the latter transform not. > We specifically "opt out" of doing that for QOI to not make undefined > behavior worse. The more correct transform would be to replace the > load with a __builtin_trap () during path isolation (or wire in path isolation > to value-numbering where we actually figure out there's no valid definition > to reach for the load). > > So yes, if you want to avoid these kind of transforms earlier instrumentation > is better. And then attempting to schedule it immediately after pass_ccp in the early-opts pipeline is already too late, right? Thanks! Alexander
