nvcc/ptxas unnecessary lmem loads/stores Bug in nvcc alias analysis/PRE stages

[wwa]

Ah, thanks for the tip. I’ve written a quick script to parse the ELF data (using objdump) from new-style CUBINs and output something that looks like an old CUBIN for decuda

Be careful using that. ;) (that’s actually for other folks who find this thread, you guys probably know this)

There’s a crapload of stuff in 3.0 that decuda currently can’t handle.