If you're compiling anyway, do you actually need the verify_program() call?

I'm thinking the try should just be around this one line:

        try:
            nvvm.compile_program(prog, len(options), options)
        except nvvm.nvvmError as e:
            # can we add something here to ensure we're not masking errors other than invalid options?

I believe it's really important to take great care that we're not masking actual errors; e.g. the hard-wired _PRECHECK_NVVM_IR might need tweaks for future GPU generations. If we're simply reporting any error as "invalid compiler option", it'll potentially take someone downstream a long time to drill down all the way back here.

While I do agree verify_program() is not needed, the location of proposed try could eventually throw as then the destroy_program() call will never get triggered.

What I was thinking is that :

program = nvvm.create_program()
try:
  <..rest of code>
  nvvm.compile_program()
except nvvm.nvvmError as e:
 <exception raising>
finally:
  destroy_program()

Could we use a non-numba example here?