Stream: git-wasmtime

Topic: wasmtime / issue #6659 `cranelift-filetest`: Allow `run` ...

Wasmtime GitHub notifications bot (Jun 28 2023 at 15:59):

fitzgen opened issue #6659:

And allow passing in arbitrary values as vmctx as well, basically treating it like any other argument.

While this wouldn't allow complex global values based on the vmctx, it would allow simple ones like gv0 = vmctx.

We would also want the CLIF interpreter to support this as well, since basically all of our runtests are also interpreter tests.

Wasmtime GitHub notifications bot (Jun 28 2023 at 15:59):

fitzgen labeled issue #6659:

And allow passing in arbitrary values as vmctx as well, basically treating it like any other argument.

While this wouldn't allow complex global values based on the vmctx, it would allow simple ones like gv0 = vmctx.

We would also want the CLIF interpreter to support this as well, since basically all of our runtests are also interpreter tests.

Wasmtime GitHub notifications bot (Jun 28 2023 at 16:30):

afonso360 commented on issue #6659:

We used to have some limited support for vmctx and global values in the interpreter (#4396). But that was essentially just used to support custom heaps, and was removed in (#5386).

We can still resolve global_value's in the interpreter, so we should still be able to use it, but we might have to manually call the function and pass the address to a stack slot or something along those lines.

Is that what you were looking to do?

We don't support something like ; run: %test(vmctx, 1, 2, 3) == 10, but I'm not sure what we would pass in instead of vmctx in the runtests.

<details>
<summary>Here's an example testcase using vmctx that works today</summary>

test interpret
test run
target x86_64
target s390x
target aarch64
target riscv64


function %vm_state() -> i64 {
    fn0 = %load_at_0_and_add(i64 vmctx) -> i64

    ;; This is our vmctx struct
    ss1 = explicit_slot 8

block0:
    ;; Store a 1 in the vmctx struct
    v1 = iconst.i64 1
    stack_store.i64 v1, ss1

    ;; Call %load_at_0_and_add with vmctx
    v2 = stack_addr.i64 ss1
    v3 = call fn0(v2)

    return v3
}
; run: %vm_state() == 2


function %load_at_0_and_add(i64 vmctx) -> i64 {
    gv0 = vmctx
    gv1 = load.i64 notrap aligned gv0+0

block0(v0: i64):
    v1 = global_value.i64 gv1
    v2 = iadd_imm.i64 v1, 1
    return v2
}

</details>

Wasmtime GitHub notifications bot (Jun 28 2023 at 16:34):

fitzgen commented on issue #6659:

I was essentially trying to do something like this:

function %stack_limit(i64 vmctx, i64) -> i64 {
    gv0 = vmctx
    stack_limit = gv0

block0(v0: i64, v1: i64):
    return v1
}

; run(0, 42) == 42

Wasmtime GitHub notifications bot (Jun 28 2023 at 16:35):

fitzgen commented on issue #6659:

I guess I could have got that working with another trampoline function to init the vmctx, that's a good trick. Still a bit of a workaround though compared to just passing in a value.

Wasmtime GitHub notifications bot (Jun 28 2023 at 16:36):

afonso360 commented on issue #6659:

Oh, weird that it doesn't work when we call the vmctx function directly. That should be a somewhat easy fix, I don't mind working on it.

I think we don't support stack_limit in the interpreter yet though.

Wasmtime GitHub notifications bot (Jun 28 2023 at 16:36):

afonso360 assigned issue #6659 to afonso360.

Wasmtime GitHub notifications bot (Jun 28 2023 at 18:15):

fitzgen closed issue #6659:

And allow passing in arbitrary values as vmctx as well, basically treating it like any other argument.

While this wouldn't allow complex global values based on the vmctx, it would allow simple ones like gv0 = vmctx.

We would also want the CLIF interpreter to support this as well, since basically all of our runtests are also interpreter tests.

Last updated: Oct 23 2024 at 20:03 UTC