runtime timers are unimplemented

[ryskiwt]

Using time.NewTicker causes a runtime error:

package main

import (
	"time"
)

func main() {
	ticker := time.NewTicker(time.Second)
	defer ticker.Stop()

	<-ticker.C

	neverDone := make(chan struct{})
	<-neverDone
}

Error:

Uncaught (in promise) LinkError: WebAssembly.instantiate(): Import #12 module="env" function="time.startTimer" error: function import requires a callable

Compiled using the 0.12.0 docker image:

docker run ... \
    -e GOOS=js -e GOARCH=wasm \
    tinygo/tinygo:0.12.0 tinygo build -target wasm -o hoge.wasm .

[niaow]

TinyGo does not implement the runtimeTimer system which is used to implement time.NewTicker and time.AfterFunc. We either need to somehow implement it (not currently possible due to how //go:linkname is implemented), or replace parts of the time package.

[ryskiwt]

Thanks for reply. I got it.
Is there any timeline or roadmap for implementing go:linkname ?

[niaow]

So //go:linkname has been implemented, but in TinyGo it preserves type information. Since time.runtimeTimer is an unexported type, another package cannot linkname time.startTimer, because the types would be different and cause a compiler error.

[niaow]

For now, if possible, use time.Sleep or loops. I do not know when we are going to be able to efficiently implement this.

[niaow]

As a workaround, I wonder whether we could do something like //go:linkname but with types.

[aykevl]

Maybe we need to support //go:linkname with slightly-mismatched types after all. Yes, we could allow custom type names for named types, but doing something like that would still be quite invasive. I still don't like the //go:extern trick for globals, it requires parsing the AST. I would prefer to avoid one more dependency on having the AST available while building the LLVM IR.

Slightly-mismatched types would be much less of a problem if opaque pointer types were completed in LLVM. Unfortunately, it looks like it'll still take quite a while (if it is ever finished). Until then, maybe the simplest solution is the following: if a function to call has a changed linkname (with //go:linkname), then check whether the types are the same and if not bitcast the function pointer. I think that would be the least invasive change to the compiler. Perhaps it would be easier after #1008.

[deadprogram]

I was just doing some experimenting with trying to run https://github.com/mbertschler/dragon-iss-docking-autopilot using TinyGo and ran into this exact problem pretty quickly.

Was reading the proposed work-arounds. Any more ideas on this currently, or do we just need to get #1008 merged in and proceed from there?

[aykevl]

#1008 is not really related, it would just make one implementation a bit easier.

I had another idea how //go:linkname could be fixed: every pointer to a named struct could be cast to *i8 or similar during the call. I don't expect that to affect optimizations, but it should be checked to be sure. It does add a bit of noise to the IR, however. And it might make some TinyGo optimization passes more complicated.

[deadprogram]

I still really would like to be able to write code like

case <-time.NewTimer(3 * time.Second).C:

[aykevl]

Yes, and we will need this to support the stdlib testing package (which makes use of these timers).

[aykevl]

I have an implementation for timers and tickers, will clean up the patch and send a PR soon.