jyknight added a comment. The idea of moving the copies to a new MachineBasicBlock seems a reasonable solution. That said, it does mean there will be allocatable physical registers which are live-in to the following block, which is generally not allowed. As far as I can tell, I _think_ that should be fine in this particular circumstance, but I'm a little uneasy that I might be missing some reason why it'll be incorrect.
I'm more uncomfortable with the scanning/splitting after-the-fact in ScheduleDAGSDNodes.cpp, and then the successor lists subsequently being incorrect. However, I wasn't immediately able to say what I'd suggest doing instead, so I spent some time yesterday poking around with this patch to see if I could find something that seemed better. I now believe it will be cleaner to have the inline-asm tell SelectionDAGISel::FinishBasicBlock to just emit the copies in another block, which we do in some other situations, already. But, I'm still poking at that to see how it'll end up -- I can't say I'm sure that's the right answer at the moment. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D69868/new/ https://reviews.llvm.org/D69868 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
