On 23/05/17 22:27, Daniel Borkmann wrote: > On 05/23/2017 09:45 PM, Alexei Starovoitov wrote: >> On 5/23/17 7:41 AM, Edward Cree wrote: >>> Hmm, that means that we can't do arithmetic on a >>> PTR_TO_MAP_VALUE_OR_NULL, we have to convert it to a PTR_TO_MAP_VALUE >>> first by NULL-checking it. That's probably fine, but I can just about >>> imagine some compiler optimisation reordering them. Any reason not to >>> split this out into a different reg->field, rather than overloading id? >> >> 'id' is sort of like 'version' of a pointer and has the same meaning in >> both cases. How exactly do you see this split? I was thinking there would be reg->id and reg->map_id. Both could share the env->id_gen, since that's not likely to run out, but they'd be separate fields so that a PTR_TO_MAP_VALUE_OR_NULL could say "this is either map_value plus a 4-byte-aligned offset less than 24, or NULL plus that same offset", and then if another pointer with the same map_id and no variable-offset part was NULL-checked, we could convert both pointers to PTR_TO_MAP_VALUE. (I'm getting rid of PTR_TO_MAP_VALUE_ADJ in my patch, along with several other types, by taking the 'we have an offset' part out of the bpf_reg_type.) > So far we haven't run into this kind of optimization > from llvm side yet[...] Out of curiosity, did you run into it with llvm? No, purely theoretical. I haven't even built/installed llvm yet, I'm just working with the bytecode in test_verifier.c for now. I'm merely trying to not have restrictions that are unnecessary; but since allowing this kind of construct would take a non-zero amount of work, I'll file it for later.
-Ed
