Hacker News Reader

Comment by klodolph | Hacker News Reader

If you were designing a bytecode as a compilation target, you would provide an easy correspondence in the bytecode to basic blocks.

See: https://en.wikipedia.org/wiki/Basic_block

WASM instead provides traditional control structures. So the compiler either has to preserve control structures through to the IR, or has to work backwards from basic blocks to control structures. Both options are undesirable, from the perspective of compiler writers, and would be unnecessary if the VM were a greenfield project.

View on HN

Nested, structured control can be broken into basic blocks without flattening it into if/goto first. It sounds easier, in fact.

In fact, the construction of a basic block from flat code is a form of recovery of control structure. If the original code had nested loops, nesting will emerge in the basic block graph.

A recursive traversal of the original structure could produce that graph more directly; it doesn't have to walk a flat list of instructions asking, "does this have a label on it which is the target of a branch". without having to walk a flat list of instructions asking questions like "is this the target of a branch".

For instance, if we walk an if/then/else AST node, we can recursively get the basic block graph for the test, the then and else part, and then integrate that into a larger basic block graph according to a rigid pattern.

The latter is very achievable using the Relooper algorithm developed by Alon Zakai and described in the Emscripten paper: https://dl.acm.org/doi/10.1145/2048147.2048224

Achievable, but extra front-end work that belongs in the backend.

I get the impression WASM is really a very clunky representation, far from what any greenfield project would ever have arrived at. That is to say, its decisions aren't just tradeoffs that people disagree about, they're simply inappropriate for what it's trying to do. More than once I've encountered comments and blog-posts lamenting its fundamentals, e.g. [0].

[0] http://troubles.md/wasm-is-not-a-stack-machine/

I don't think that's true. It was designed for its purpose by people that knew what they were doing based on experience gathered from implementing previous incremental steps including emscripten's original JavaScript output and asm.js (a formalization of some techniques emscripten employed). The design rationale is right there in the spec repository: https://github.com/WebAssembly/design/blob/master/Rationale....

It was built to be a better target than asm.js for compiled languages to run in JavaScript VMs and it seems to have succeeded on that front. That it's not a perfect fully-general bytecode format doesn't seem like a real knock against it, that's a much harder problem to solve. In any event it seems to be getting a lot more traction than any previous entries in the space, which is exciting!

The criticisms in the linked article are simply not grounded in fact, and the article was obviously written by someone without any expertise in how modern compilers work. There are also a number of basic errors in the article.

Liveness information simply doesn’t belong in the bytecode. SSA is trivial to recreate from local mutable variables (it would make a good homework assignment for someone in an undergrad “intro to compilers” class). WebAssembly is obviously a register machine.

> With this, it becomes possible to get rid of locals entirely.

Both factually incorrect and pointless. There is no tangible benefit in getting rid of locals entirely. You are merely changing out one representation (register machine) for a different one (stack machine).

There are valid criticisms of WASM, but the linked article doesn’t have any.

The weird part of WASM is the control structures. The rest of it is a fairly sensible, actually rather nice register machine. You can see that older bytecode systems like the JVM are stack machines, but newer ones tend to be register machines. This isn’t because people are getting stupid, it’s because there are legitimate reasons to prefer register machines, and on the balance of things, my observations are that people with experience in the field tend to prefer register machines.

Interesting, thanks for the analysis.

> You can see that older bytecode systems like the JVM are stack machines, but newer ones tend to be register machines.

Like LLVM. Is there a good reason for WebAssembly not being SSA?

There’s no good reason for WebAssembly to be SSA. That is a good enough reason for it not to be.

The purpose of SSA is to make it easier to write optimizations. However, you wouldn’t be optimizing WASM anyway, you would necessarily translate it first into some kind of IR suitable for optimization passes. Might as well convert to SSA at that point, rather than bloat the bytecode by exposing implementation details of the target.

Remember that WASM’s purpose is to be portable and safe. Making it more complicated just in order to make sophisticated back-ends slightly faster is a net loss. Using SSA would also make naïve/simple backends slower.

> The weird part of WASM is the control structures.

Out of curiosity, what do you find weird about the control structure portion?

I just wrote a basic language that compiles to wasm and found the built in control structures made my life easier.

Open any compilers book, WASM will look alien. Control flow analysis is normally done through basic blocks. If you make a toy language and don’t do control flow analysis you won’t notice what’s so weird about WASM. If you take an existing compiler and port it to WASM, you’ll cry, then gather yourself together and read how the relooper algorithm works.

The reason wasm is like this is that with "normal" control flow chrome would have had to implement itself the relooper algorithm internally before the MVP release.

Overall the main negative point seems to be that compilers toward wasm are more complex, but under the constraints of the web moving complexity from the client looks like a valuable positive point.

Two other negative points would possibly be performance and code size (which is the reason for a if-then-else specific instruction for example) about which I know very little. Do you think it would have made a difference here?

I would still consider WASM a stack machine and not a register machine. Yes, there are mutable local variables in WASM but Java bytecode has them as well - which you consider a stack machine. BTW the designers of WASM explicitly call WASM a stack machine here: https://github.com/WebAssembly/design/blob/master/Rationale..... With WASM's MVP it was necessary to store e.g. loop state in local variables, thanks to recent changes this doesn't seem to be necessary anymore. I think this was the main argument that blog post considered WASM to be a register machine. javac also makes heavy use of variables in bytecode, but somehow no one considers the JVM a register machine.

> my observations are that people with experience in the field tend to prefer register machines

That's actually the opposite of my observation, they seem to prefer stack machines.

Java has instructions like dup, swap etc. To me, that is the critical difference here, and where I draw the line between “stack machine” and “register machine”.

I have to admit this line seems arbitrary to me. So WASM is a register machine to you but if they would simply add those 2 instruction would it suddenly become a stack machine then? Those instructions would actually be trivial to add. I think those terms are relatively well defined and when you argue that WASM is a register machine even though the inventors explicitly claim it's a stack machine you should have really good arguments for that. Personally I would be surprised if you could point me to any literature that supports your definition.

Turing completeness sounds like an arbitrary distinction to those outside of the field of CS, but it’s not.

To me, the distinction here is that the stack machine in WASM is restricted to the point that it corresponds 1:1 with an expression tree—not even a graph, just a tree. This means that every function in Web Assembly can be thought of as a collections of statements and expressions, and the stack machine abstraction is nothing more than a serialization format for the expressions.

Maybe dial it back a bit with the challenge to point at literature. The literature has not really caught up with the existence of WASM yet.

The multi-value proposal breaks the ability to turn Wasm into expressions easily, and thus makes it even more of a stack machine than it already is. Dup and swap may still be added in the future.

A defining feature of a register machine is that the actual instruction encoding has direct references to source and destination registers in it. Wasm doesn't have those, it has explicit get_local instructions instead.

That said, if you turn off LLVM's WebAssemblyRegStackify pass, all LLVM IR's values will end up in locals, with little to no stack usage. Still no register machine, but a bit more of a grey area :)

> Maybe dial it back a bit with the challenge to point at literature. The literature has not really caught up with the existence of WASM yet.

It wasn't me who claimed that WASM is "obviously" a register machine, despite the inventors saying otherwise. They even explicitly state that they decided against a register machine. I guess it's then reasonable for me to ask on what definition of stack vs register you are basing this opinion on. Let me be clear: I was not asking here for literature about WASM specifically but a definition of register/stack machines that supports your claim.

WASM's instruction encoding is very much based on a stack machine. Even with the initial limitations you mentioned I don't think it qualifies as "obviously a register machine". As already mentioned in multiple comments those restrictions were already lifted with the multi value proposal.

I understand that there is a grey area, but simply claiming "obviously a register machine" doesn't seem right to me. Implementation-wise WASM is a stack machine even if it needs/needed locals to be turing-complete.

dup and swap can be implemented with locals in wasm (on well typed stacks)

The clang disassembly given in the article sure makes it look like WASM is a nested expression tree, which leaves the choice of stack versus register to the implementation.

      (f32.add
        (f32.add
          (f32.mul
            (f32.load
              (local.get 0))
            (f32.load
              (local.get 1)))
          (f32.mul
            (f32.load offset=4
              (local.get 0))
            (f32.load offset=4
              (local.get 1))))
        (f32.mul
          (f32.load offset=8
            (local.get 0))
          (f32.load offset=8
            (local.get 1))))

The outer f32.add could translate into a byte code instruction that finds its two operands on a stack, or to one which gets them from registers.

The code only says that there is a f32.add call which has two operands that are the result of a f32.add and f32.mul and so on.

The implementations will agree in their treatment of locals: that there are two locals 0 and 1, which support loading at offsets and such.

Both stack and register machines can support locals.

Yes, that’s exactly true. The “stack machine” here can be seen as nothing more than a way of encoding the expression tree.

The stack machine is a model for the semantics of wasm, in the sense that the safety properties of wasm are defined in terms of stack types rather than SSA of register properties (for those that are unfamiliar with stack machines, this stack is a different type of concept from the call stack, that in wasm is used store the local variables). Whether during execution it is better implemented as a register machine or a stack machine is an implementation detail.

That's just the pprinter working its magic :-).

Just read the standard, it's quite clear from the semantics that it's a stack machine: https://webassembly.github.io/spec/core/exec/runtime.html#st... https://webassembly.github.io/spec/core/exec/instructions.ht...

The control structure as nested block are a must be for wasm, because chrome compilation model of Javascript is incompatible with an LLVM-like goto-based control flow.

Reply on news.ycombinator.com