Stream: wasm

Topic: ✔ typed funcrefs "undeclared" if not used in init exprs?

Chris Fallin (Feb 23 2024 at 17:17):

Hi all! I'm playing with typed funcrefs to see about using them for efficient indirection (ICs); specifically I want globals that are mutable typed non-nullable funcrefs.

I've hit some perplexing behavior that I can't square with anything I see in the proposal doc; specifically, a ref.func instruction yields an "undeclared function reference" error when validated by wasm-tools if the func it refers to is not also used in an init expression for a global or in a table element segment.

This validates:

(module
  (type $ic (func (param i32 i32 i32 i32) (result i32)))

  (global $ic1 (mut (ref $ic)) (ref.func $icfunc1))
  (global $ic2 (mut (ref $ic)) (ref.func $icfunc2))

  (func $icfunc1 (param i32 i32 i32 i32) (result i32)
        local.get 0)
  (func $icfunc2 (param i32 i32 i32 i32) (result i32)
        local.get 1)

  (func (export "setup")
        ref.func $icfunc2
        global.set $ic1))

but this does not:

(module
  (type $ic (func (param i32 i32 i32 i32) (result i32)))

  (global $ic1 (mut (ref $ic)) (ref.func $icfunc1))

  (func $icfunc1 (param i32 i32 i32 i32) (result i32)
        local.get 0)
  (func $icfunc2 (param i32 i32 i32 i32) (result i32)
        local.get 1)

  (func (export "setup")
        ref.func $icfunc2
        global.set $ic1))

which looks like weird action-at-a-distance to me (removing global $ic2 affects validation of a function that doesn't mention it at all). Any thoughts @fitzgen (he/him) / @Alex Crichton / anyone else?

(validating with latest main of wasm-tools, using wasm-tools validate -f all file.wat)

fitzgen (he/him) (Feb 23 2024 at 17:18):

Any function that isn’t exported needs to be in an (elem declare …) element for you to ref.func it

Alex Crichton (Feb 23 2024 at 17:18):

Ah yeah this is expected behavior (albeit with a bad error message), this came about with the reference-types proposal

Alex Crichton (Feb 23 2024 at 17:19):

IIRC the intention was to enable runtimes to, before the code section has arrived, be able to enumerate functions which could "escape" through globals/tables/etc

Chris Fallin (Feb 23 2024 at 17:19):

ah, great, I had missed that; thanks!

Notification Bot (Feb 23 2024 at 17:20):

Chris Fallin has marked this topic as resolved.

Joel Dice (Feb 23 2024 at 17:22):

Thanks for working on this @Chris Fallin. I naively tried to use typed funcrefs in componentize-py before realizing they weren't implemented yet. I look forward to revisiting that.

Chris Fallin (Feb 23 2024 at 17:22):

yup, I had a realization that they are exactly what I want for "just call a native function pointer with no checks plzkthx" performance

Chris Fallin (Feb 23 2024 at 17:22):

thanks folks (fitzgen I think?) for implementing them :-)

Joel Dice (Feb 23 2024 at 17:24):

Should I take that to mean typed funcrefs already work in recent versions of Wasmtime? If so, that's awesome.

Alex Crichton (Feb 23 2024 at 17:25):

I think that call_ref should work yeah, but I'll also note that the perf isn't quite there yet. Right now call_ref still does the fully general typecheck

Alex Crichton (Feb 23 2024 at 17:25):

I don't think we've implemented the optimization where the type information from validation flows directly into translation

Alex Crichton (Feb 23 2024 at 17:25):

that's what would enable completely skipping typechecks and such

Chris Fallin (Feb 23 2024 at 17:25):

my experiments last night showed a null check still (and there's a comment in visit_call_ref about removing it by threading type info, yeah), but no typecheck?

Alex Crichton (Feb 23 2024 at 17:26):

oh I must be misremembering the instruction then, I thought call_ref took a funcref as opposed to a (ref $ty)

Joel Dice (Feb 23 2024 at 17:26):

Does anyone happen to know if typed funcrefs work in Node and/or recent browsers (e.g. if I were to use them in componentize-py and run the result with Jco)?

fitzgen (he/him) (Feb 23 2024 at 17:27):

https://webassembly.org/features/

Chris Fallin (Feb 23 2024 at 17:27):

they're a prereq for GC, so according to that ^^ at least FF and Chrome, but not Safari I guess

fitzgen (he/him) (Feb 23 2024 at 17:28):

it pretty much folded into the gc proposal, and browsers don't generally differentiate at this point

Last updated: Nov 22 2024 at 16:03 UTC