On Thu, Jan 30, 2020 at 04:44:09PM -0800, John Fastabend wrote: > > > > But verifier seems handling <<= and >>= correctly, right? > > Even we have it, the verifier should reach the same conclusion > > compared to not having it, right? > > > > No, verifier marks it unknown and doesn't track it fully. We > can perhaps improve the verifier but the above is a nice > fix/improvement for the backend imo regardless. I don't see how the verifier can be taught to carry smax information after <<32 shift. The verifier has to do dst_reg->smax_value = S64_MAX. The only way to carry smax is to recognize a sequence of <<32 s>>32. While at it we can recognize both <<32 >>32 and <<32 s>32 as pseudo insns.
