The OGL issue occurs with both LLVM 8 and 9-git, so there might not be a
connection to your radv issue here. I suppose LLVM 9 would also be fine with
AMD_DEBUG=nodma & radeonsi, which indicates that the bug really is in the async
DMA code of radeonsi (or at least triggers it).

I also would suppose that RADV_PERFTEST= doesn't have to know the same values
like AMD_DEBUG=, which afaik applies only to radeonsi.