The XFAILed test in test/CodeGen/AMDGPU/reg-coalescer-sched-crash.ll is a verifier error introduced by the register coalescer. I think it has something to do with rematerializing an instruction into a sub register when the super register isn't fully defined.
This is a reduced testcase from an OpenCV regression
Is this new? I.e., is this a regression and if yes, what introduced it?
(In reply to comment #2) > Is this new? > I.e., is this a regression and if yes, what introduced it? The regression showed up in r251860, so I think this was uncovering a existing problem.
I think this bug always existed, it only affects targets with acticated subregister liveness. We already had a longer IRC session analyzing it (just have to grep my logs when I start working on it), so I'll look into it.
(In reply to comment #4) > I think this bug always existed, it only affects targets with acticated > subregister liveness. We already had a longer IRC session analyzing it (just > have to grep my logs when I start working on it), so I'll look into it. Sounds good. Thanks both. For the sake of future archeology ;), would you mind summarizing the content of the IRC discussion?
> For the sake of future archeology ;), would you mind summarizing the content of the IRC discussion? Sorry I can't remember the details (intricate stuff in the rematerialization logic of the register coalescer). Once there is a fix it is hopefully clear that the problem was :)
(In reply to comment #6) > > For the sake of future archeology ;), would you mind summarizing the content of the IRC discussion? > > Sorry I can't remember the details (intricate stuff in the rematerialization > logic of the register coalescer). Once there is a fix it is hopefully clear > that the problem was :) Fair enough :).
Fixed in r262768