WASM is not quite a stack machine

ufo•about 1 hour ago

The author seems to complain about a lack of stack manip expressions like dup and rot, but at least for me that's what I would expect from an average programming language stack machine. Even Java, which does have those instructions, doesn't use them --- reuse happens via local variables.

The way I see it, the difference between register and stack vms is all about the instruction encoding. Register VMs have fatter instructions in exchange for needing fewer LOAD and STORE operations. Despite the name, register VMs also have a stack.

pjjpo•3 minutes ago

> Despite the name, register VMs also have a stack.

Out of curiosity what do you think about this - in spite of the name, stack machines also have yet another stack. Ok I don't like that wording, but locals are basically the stack frames people know of from their computer arch class I think.

It doesn't change the fact that Wasm operations have to have the execution stack as one or more of the operands. Seems like a stack machines to me too, though I don't know more details on why the specific design of Wasm would make optimizing compilers harder to write than JVM as the article suggests (I think?).

U1F984•6 minutes ago

Java does use dup in some cases, e.g.

   public static void test() { 
      new Object();
   }

         0: new           #2                  // class java/lang/Object
         3: dup
         4: invokespecial #1                  // Method java/lang/Object."<init>":()V
         7: pop
         8: return

asibahi•10 minutes ago

I dont really disagree with the main premise of the article, which is that WASM is not really a stack language, but this part just gave me pause:

> In textual Wasm, for example, they are instead represented in a LISP-like notation – not any less or more efficient

The Text format, at least when it comes to instructions, it 1 to 1 with the binary format. The LISP-like syntax is mainly just syntax sugar[1].

    ‘(’ plaininstr  instrs ‘)’ ≡  instrs plaininstr

So (in theory, as far as I understand it) you can just do `(local.get 2 local.get 0 local.get 1)` to mean `local.get 0 local.get 1 local.get 2`, and it works for (almost) any instruction.

Unfortunately, in my limited testing, tools like `wat2wasm` and Binaryen's `wasm-as` don't seem to adhere to (my perhaps faulty understanding of) the spec, and demand all instructions in a folded block be folded and have the "correct" amount of arguments, which makes Binaryen do weird things like

    (return
      (tuple.make     ;; Binaryen only pseudoinstruction
        (local.get 0) ;; or w/e expression
        (local.get 1) ;; or w/e expression
      )
    )

when this is perfectly valid

    local.get 0
    local.get 1
    return

tl;dr: the LISP syntax is just syntax sugar. The textual format is as "stack-like" as the binary format.

[1]: https://webassembly.github.io/spec/core/text/instructions.ht...

Hendrikto•40 minutes ago

The series of articles linked at the end (troubles.md/posts/wasm-is-not-a-stack-machine/) is even more interesting, imo.

Very well articulated and concise critique by somebody who seems to have a great amount of knowledge and experience with the topics.

stevefan1999•about 3 hours ago

I'm trying to implement a WASM to C compiler, and because of that not-quite-so-stack behavior, I can actually guarantee that it will always build an expression and I don't have to discard or reset stack value! Everything stays within that function, which is very neat, and I think it is one of the reason WAT, the textual format is so neat, that you can represent it with a S-Expression.

jedisct1•about 2 hours ago

Compiling WASM to C is a really good option: https://00f.net/2023/12/11/webassembly-compilation-to-c/

ncruces•about 1 hour ago

Shameless plug… compiling it to Go is a great option too: https://github.com/ncruces/wasm2go

I've used it to translate SQLite (with a few extensions) and, that I know of, it's been used (to varying degrees of success) to translate the MARISA trie library (C++), libghostty (Zig), zlib, Perl, and QuickJS.

More on-topic, I use a mix of an unevaluated expression stack and a stack-to-locals approach to translate Wasm.

bsder•about 2 hours ago

But how do you handle arguments or loop index variables? Your liveness is the entire function? You have to compile all the WASM chunks together in order to do any optimization? That seems ... problematic.

Edit: Yep. In article referenced from the original: http://troubles.md/posts/wasm-is-not-a-stack-machine/

Double edit: Some of this has already been fixed in WASM: https://github.com/WebAssembly/multi-value

kg•about 1 hour ago

The lack of a dup opcode in Wasm as mentioned in the post is quite annoying when trying to generate compact code. I wish something like it had made it into the spec.

thomasmg•28 minutes ago

You could use "local.tee". I kind of is "store" + "duplicate".

asibahi•19 minutes ago

`local.tee` doesn't duplicate. it just doesn't remove the value from the stack. (so it is "just" `local.set` followed by `local.get`)

WASM is not quite a stack machine

⚡ Community Insights

Discussion (12 Comments)Read Original on HackerNews