ref.i31 is both an administrative instruction and an instruction #470

Activity

1. rossberg commented on Nov 6, 2023

   

   rossberg

   on Nov 6, 2023

   Member

   More actions

   Yeah, this is analogous to ref.func, where we go from ref.func $x to ref.func 𝑎𝑑𝑑𝑟, essentially overloading the mnemonic. We should probably have a better naming scheme for administrative instructions that are reducts of regular ones.
2. redianthus commented on Nov 8, 2023

   

   redianthus

   on Nov 8, 2023

   ContributorAuthor

   More actions

   We should probably have a better naming scheme for administrative instructions that are reducts of regular ones.

   Indeed, that would help in some cases. I'm writing a proof and I either have to explain that there are two instruction with the same name (which I'd like to avoid because it's disturbing the reader) or use another name (likely i31.new) otherwise it will be too much confusing for potential reviewers that are not familiar with the Wasm semantics.
3. rossberg commented on Nov 8, 2023

   

   rossberg

   on Nov 8, 2023

   Member

   More actions

   The only convention I could think of so far is to add a prime to the administrative versions, which we e.g. used in the OOPSLA'19 memory model paper. How about going with that for your paper as well?
4. redianthus commented on Nov 8, 2023

   

   redianthus

   on Nov 8, 2023

   ContributorAuthor

   More actions

   That could be a solution. I'm going to think about it and see if I can think of something else.

   I believe it would be good to also harmonize the script format. For instance, here:

   (assert_return (invoke "f") (ref.i31))

   The ref.i31 should probably match what we have as the administrative instruction, what do you think ?
5. rossberg commented on Nov 8, 2023

   

   rossberg

   on Nov 8, 2023

   Member

   More actions

   Hm, yeah, but then all ref patterns should still follow a consistent naming scheme as well.
6. redianthus commented on Nov 8, 2023

   

   redianthus

   on Nov 8, 2023

   Last edited by redianthus

   ContributorAuthor

   More actions

   Here are a few possibilities:

   Name	i31 instr	i31 admin	i31 result	struct instr	struct admin	struct result	func instr	func admin	func result
   Current	ref.i31	ref.i31 u31	ref.i31	struct.new	ref.struct structaddr	ref.struct	ref.func	ref.func funcaddr	ref.func
   Old	i31.new	ref.i31 u31	ref.i31	struct.new	ref.struct structaddr	ref.struct	ref.func	ref.func funcaddr	ref.func
   Prime	ref.i31	ref.i31' u31	ref.i31'	struct.new	ref.struct' structaddr	ref.struct'	ref.func	ref.func' funcaddr	ref.func'
   Ptr	ref.i31	ptr.i31 u31	ptr.i31	struct.new	ptr.struct structaddr	ptr.struct	ref.func	ptr.func funcaddr	ptr.func
   New	i31.new	ref.i31 u31	ref.i31	struct.new	ref.struct structaddr	ref.struct	func.new	ref.func funcaddr	ref.func
   Ptr+Const	ref.i31	i31.const u31	i31.const u31?	struct.new	ptr.struct structaddr	ptr.struct	ref.func	ptr.func funcaddr	ptr.func
   New+Const	i31.new	i31.const u31	i31.const u31?	struct.new	ref.struct structaddr	ref.struct	func.new	ref.func funcaddr	ref.func

   The one I like the most is probably the New one.
7. rossberg commented on Nov 10, 2023

   

   rossberg

   on Nov 10, 2023

   Member

   More actions

   Well, we should find a convention that works uniformly for more than just references. For example, we have invoke which is just call', threads have wait', and possibly others. And of course I dropped i31.new for a reason: to distinguish allocations from just referencing.
8. redianthus commented on Nov 10, 2023

   

   redianthus

   on Nov 10, 2023

   ContributorAuthor

   More actions

   I can't find the call' you're referring to, where is it defined ?
9. rossberg commented on Nov 10, 2023

   

   rossberg

   on Nov 10, 2023

   Last edited by rossberg

   Member

   More actions

   It's currently called invoke (on main repo), but relates to call in exactly the same way as ref.func relates to ref.func with addr. Ideally, they should use the same convention.
10. redianthus commented on Nov 10, 2023

    

    redianthus

    on Nov 10, 2023

    Last edited by redianthus

    ContributorAuthor

    More actions

    Ideally, they should use the same convention.

    Ideally yes, but I'm not sure we can find a convention that is not confusing. To me, adding a prime is confusing because these are not variables but instruction literals. I believe it would make everything simpler if we were to find a proper name for each of them - of course, that may not always be easy.

    If we want to distinguish allocating instructions, we could have i31.of_i32 instead of i31.new. For func.new, I don't have a good name right now, I'll try to think about it.

    I knew the invoke one (I thought you were saying it was reducing to call' and I was getting worried that I missed something in my formalization). I believe this one is a perfectly fine name.

    For wait', I don't know. I'm not familiar yet with the threads proposal semantics.

    EDIT:

    Quoting your answer in #478:

    See my reply on the other issue. Also, ref.null is already standardised as part of Wasm 2.0 (as is ref.func, btw), so we cannot change it anymore.

    OK, so we probably won't be able to fix the issue in the way I propose by having proper name for each instructions. Still, I believe we could fix those that are not yet fully part of a released standard such as i31.ref being renamed to i31.of_i32 or whatever.