Hacker News Reader

Comment by titzer | Hacker News Reader

Structured control flow is more efficient to verify, more compact, and disallows irreducible loops by design.

Also note that "br" "br_if" and "br_table" are all just text format mnemonics. You can think of the "br" as "branch" instead of break. We didn't have different instructions for branching depending on the type of the label because it's not necessary, just a waste of opcode space.

View on HN

> You can think of the "br" as "branch" instead of break

Yes, but the branch target depends on the targeted block's kind: it's the block-exit label for ifs and blocks and the block-start for loops. Just a bit unusual.

> We didn't have different instructions for branching depending on the type of the label because it's not necessary, just a waste of opcode space.

Yeah, I understand that, but it still could be two different mnemonics: 0x0C/0x0D could disassemble as "continue/continue_if" if the label targets a loop, and as "break/break_if" if it targets a non-loop; similarly for assembling. I imagine the added overhead would be minuscule.

Anyway, thanks for answering. I suppose forcing the structured control flow in the IR beats e.g. simplicity of compiling a for-loop in one pass.

    block
        ;; code for INITIALIZATION
        loop
            ;; code for CONDITION
            i32.eqz
            br_if 1
        
            i32.const 0
            loop
                if
                    ;; code for INCREMENT
                    br 2
                end
        
                ;; code for BODY
                i32.const 1
                br 0
            end
        end
    end

Well, more efficient to verify, because it is more restrictive

Meaning some control flows are unrepresentable and certainly less compact; since they must be 'interpreted'

I think that's probably the best decision for quick webpage loading, but given that WASM is generally used for code heavy webpages -- not quick jquery replacements, I'm not convinced it was the best decision overall

-- I love a lot of wasm, I just wish it was a better target for open ended computation; which is something gotos and coroutines are better suited for

That said; WASM is a clever trade off; but the trade-off is hidden within compilers (ref Go's wasm for instance)

Nothing is unrepresentable, if I’m not mistaken.

https://en.m.wikipedia.org/wiki/Structured_program_theorem

Rather, I think the trade off was more in the realm of “aggressively optimize for what the web does well”.

With that being said, I do wish it was better suited for general purpose computation.

Irreducible loops (loops with more than one way to enter) are not representable by structured control flow. It's possible to model with with state variables (e.g. loop over switch). Incidentally if irreducible control flow is modeled this way (set the control variable to a constant and then branch back to the loop header which immediately then switches on the variable), and the compiler performs jump threading with constant propagation (i.e. duplicate the loop header up to the switch, and fold the switch), the resulting control graph will again have irreducible control flow. In fact it will be the same as the original control flow graph. So it's possible to undo the inefficiency of the state variable with some transformations in the Wasm consumer.

It's not just that it's more restrictive, it's that it forces the producer to do work for the verifier. Structured control flow follows a stack discipline and also pre-declares all labels, so the consumer (verifier) can use the minimum data structures necessary, immediately reusing metadata for each control structure when it is exited.

Since you're giving great answers ;)

What's the problem with multiple entry points anyway? If we're not using computed goto, or dynamically verifying it when we do, aren't all of our jump targets known, and so inherently safe?

What am i missing?

EDIT: is it just a question of speed of jit for register allocation / dominance frontiers?? But the verifier is providing safety right??

Sure, you can still have safe code with irreducible loops. Safety is orthogonal. There is a proposal to add "multi-loop" (https://gist.github.com/conrad-watt/6a620cb8b7d8f0191296e3eb...) to Wasm, but it hasn't been a high priority.

Of all the Wasm JITs I know of, none of the Web engine's optimizing compilers can handle irreducible control flow in their backends. Their register allocators and codegen algorithms would need to be modified. I worked on TurboFan, and several parts of it would choke on irreducible control flow. It's likely these engines would fall back to baseline compilation for such functions, which is a bigger perf impact than using state variables.

I was hoping we’d have someone more qualified to answer this question, but whatever. Holy shit!

https://news.ycombinator.com/user?id=titzer

I love this site

Reply on news.ycombinator.com