Inlining changes behavior of memory load

[alexcrichton]

This input:

(component
  (core module $M
    (memory (export "mem") 1)
    (func (export "writeit") (param i32 i32)
      (i32.store (local.get 0) (local.get 1)))
  )
  (core instance $m (instantiate $M))
  (core module $N
    (import "" "mem" (memory 1))
    (import "" "writeit" (func $writeit (param i32 i32)))
    (func (export "g") (result i32)
      (i32.load (i32.const 0))
      drop
      (call $writeit (i32.const 0) (i32.const 123))
      (i32.load (i32.const 0)))
  )
  (core instance $n (instantiate $N
      (with "" (instance
        (export "mem" (memory $m "mem"))
        (export "writeit" (func $m "writeit"))))))
  (func (export "g") (result u32) (canon lift (core func $n "g"))))

changes behavior with inlining:

$ cargo run -- -C inlining=n --invoke 'g()' repro.wat
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 0.15s
     Running `/home/alex/code/wasmtime2/target/debug/wasmtime -C inlining=n --invoke 'g()' repro.wat`
123
$ cargo run -- -C inlining=y --invoke 'g()' repro.wat
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 0.11s
     Running `/home/alex/code/wasmtime2/target/debug/wasmtime -C inlining=y --invoke 'g()' repro.wat`
0

An LLM-produced report is here and this looks like a regression after #13525 and #13354 -- cc @fitzgen

[fitzgen]

Ooo this is a good one.

We can sometimes resolve this by turning the imported memory alias region into a precise defined memory alias region when the importing module is always instantiated with the same memory (e.g. it is instantiated exactly once). However, if it is instantiated multiple times with multiple different memories, then we can't do that because sometimes the imported memory might alias and other times it might not. We would need monomorphization to keep applying precise alias regions.

I guess the conservative fix, when we don't have a single precise memory to reference, would be to make accesses to exported memories use the imported memory alias region (potentially only when inlining is also enabled).

[fitzgen]

I guess the conservative fix, when we don't have a single precise memory to reference, would be to make accesses to exported memories use the imported memory alias region (potentially only when inlining is also enabled).

And we could potentially only do this for the memories that can flow into an ambiguous import, as well, fwiw.