Use of global for data offset causes lucetc error

[kanaka]

Consider the following code:

(module

  (memory 1)
  (global $memoryBase i32 (i32.const 0))
  (data
    (global.get $memoryBase)
    "abcd")

  (func $main (nop))

  (export "_start" (func $main))
)

Attempting to compile with lucetc-wasi has the following results (shows from .wat followed by from .wasm):

# lucetc-wasi wasitests/data_global_offset.wat -o wasitests/data_global_offset.so
Error(Validate("test.wast:6:6: error: initializer expression can only reference an imported global\n    (global.get $memoryBase)\n     ^^^^^^^^^^\n"))

# wasm-as wasitests/data_global_offset.wat -o wasitests/data_global_offset.wasm
# lucetc-wasi wasitests/data_global_offset.wasm -o wasitests/data_global_offset.so
LucetcError { inner: ErrorMessage { msg: "cannot create sparse data: data initializer uses global as base" }

Unsupported }

wasm-as has no problem compiling the module and wasmtime has no problem running the resulting module.

[pchickey]

Hi @kanaka, thanks for the report. This is a known bug - we don't support import globals, or the use of globals in init expressions yet. We didn't prioritize implementing these because none of the toolchains we use emits them, but we plan to implement them soon as part of spec compliance.

[awortman-fastly]

Hi! Following up on this: I've been working on import/export and globals support lately, and there's a few angles to this one:

First, I think wasm-as might be in the wrong in accepting your example program. As the first quoted error indicates, validation failed, but wasm-as happily accepts. That lead me to The Spec:

• Data Segments makes quite clear that an i32 const expression can be the offset of a data segment, which is fine
• But the footnote for Constant Expressions specifically points out that global.get is only allowed to refer to imported globals.

Indeed, the following program:

(module

  (global $memoryBase (import "env" "mem_base") i32)
  (memory 1)
  (data
    (global.get $memoryBase)
    "abcd")

  (func $main (nop))

  (export "_start" (func $main))
)

validates and is rejected by lucetc later. There's a small digression about what wasm-as actually produces, but it looks like the right program assuming the global.get constraint is loosened.

With regard to the second program, that's definitely on us. While imported globals are supported lucetc (as of #204), the runtime provides no mechanism to satisfy imported globals (à la bindings.json for import functions/hostcalls), so it will fall over there instead.

I'm hoping to spend some time this week adjusting the data representation to be amenable to a runtime-defined base, whichever way globals pans out.