Stream: git-wasmtime

Topic: wasmtime / issue #10460 Winch x64 simd: disagreement with...

Wasmtime GitHub notifications bot (Mar 22 2025 at 17:44):

alexcrichton added the fuzz-bug label to Issue #10460.

Wasmtime GitHub notifications bot (Mar 22 2025 at 17:44):

alexcrichton added the wasm-proposal:simd label to Issue #10460.

Wasmtime GitHub notifications bot (Mar 22 2025 at 17:44):

alexcrichton added the winch label to Issue #10460.

Wasmtime GitHub notifications bot (Mar 22 2025 at 17:44):

alexcrichton opened issue #10460:

This input:

(module
  (func (export "x")
    (param i32 v128 i32)
    (param v128 i32 v128 i32 v128)
    (param i32 v128 i32 v128 i32)
    (param v128 i32 v128 i32)
    (param $last v128)
    (result v128)
    local.get $last
  )
)

(assert_return
  (invoke "x"
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
  )
  (v128.const i64x2 0 0))

yields:

$ cargo run -q wast -Ccompiler=winch repro.wat
Error: failed to run script file 'repro.wat'

Caused by:
    0: failed directive on repro.wat:13:1
    1: result 0 didn't match
    2: expected [                   0,                    0]
       actual   [      94750456038144,                    0]

       expected (hex) [0000000000000000, 0000000000000000]
       actual (hex)   [0000562ccf464300, 0000000000000000]

cc @jeffcharles @saulecabrera

Wasmtime GitHub notifications bot (Mar 24 2025 at 10:13):

saulecabrera commented on issue #10460:

Thanks for the report! We'll take a look into it.

Wasmtime GitHub notifications bot (Mar 24 2025 at 22:05):

jeffcharles commented on issue #10460:

I wasn't able to resolve this today but I narrowed down the repro case to:

(module
  (func (export "x")
    (param f32 f32 f32 f32 f32 f32 f32 f32 f32) (param $last v128)
    (result v128)
    local.get $last
  )
)

(assert_return
  (invoke "x"
    (f32.const 1)
    (f32.const 2)
    (f32.const 3)
    (f32.const 4)
    (f32.const 5)
    (f32.const 6)
    (f32.const 7)
    (f32.const 8)
    (f32.const 9)
    (v128.const i64x2 10 11)
  )
  (v128.const i64x2 10 11))

The key detail is there needs to be enough float/vector parameters to require using the stack for the last two and the last parameter must be a vector. Cranelift seems to be using offsets of 0 for the f32 and 16 for the v128 for the last two parameters whereas it'll use offsets of 0 and 8 for two f32 parameters. Winch uses offsets of 16 for the f32 and 24 for the v128 from the base pointer in both cases. I'm not very familiar with exactly how Cranelift uses those offsets so I can't speak to whether there's an issue with the offsets being different for the different argument types.

Wasmtime GitHub notifications bot (Mar 31 2025 at 13:32):

saulecabrera assigned saulecabrera to issue #10460.

Wasmtime GitHub notifications bot (Apr 02 2025 at 20:05):

fitzgen closed issue #10460:

This input:

(module
  (func (export "x")
    (param i32 v128 i32)
    (param v128 i32 v128 i32 v128)
    (param i32 v128 i32 v128 i32)
    (param v128 i32 v128 i32)
    (param $last v128)
    (result v128)
    local.get $last
  )
)

(assert_return
  (invoke "x"
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
    (i32.const 0)
    (v128.const i64x2 0 0)
  )
  (v128.const i64x2 0 0))

yields:

$ cargo run -q wast -Ccompiler=winch repro.wat
Error: failed to run script file 'repro.wat'

Caused by:
    0: failed directive on repro.wat:13:1
    1: result 0 didn't match
    2: expected [                   0,                    0]
       actual   [      94750456038144,                    0]

       expected (hex) [0000000000000000, 0000000000000000]
       actual (hex)   [0000562ccf464300, 0000000000000000]

cc @jeffcharles @saulecabrera

Last updated: Apr 18 2025 at 22:03 UTC