Stream: git-wasmtime

Topic: wasmtime / issue #11024 Cranelift: Removing irrelevant in...

Wasmtime GitHub notifications bot (Jun 12 2025 at 13:52):

akldc edited issue #11024:

.clif Test Case

test optimize
    set opt_level=none
    set preserve_frame_pointers=true
    set enable_multi_ret_implicit_sret=true

function u1:1() -> f64 fast {
    ss0 = explicit_slot 32

block0():
    v1 = f64const 0x1.d707d3c2dd690p-1

    v7 = stack_addr.i64 ss0
    v8 = load.f32x4 little v7+8
    return v1
}
function %main() -> i64x2, f64,f64x2 fast {
    ss0 = explicit_slot 32
    ss1 = explicit_slot 32
    sig0 = () -> f64 fast
    fn0 = u1:1 sig0
    const0 = 0x655a67ef3826115b29aeb0d774a11922

block0:

    v1 = iconst.i64 -8142416529737083292
    v2 = vconst.i64x2 const0

    v3 = stack_addr.i64 ss0
    v4 = stack_addr.i64 ss1

    store little v1, v3
    store little v1, v3+8
    store little v1, v3+16
    store little v1, v3+24

    store little v1, v4
    store little v1, v4+8
    store little v1, v4+16
    store little v1, v4+24

    v5 = stack_addr.i64 ss1
    atomic_store little v1, v5

    v6 = fcvt_from_uint.f64 v1
    v7 = scalar_to_vector.f64x2 v6

    v8 = call fn0()
    return v2,v6,v7
}


; print: %main()

Result

[x86    ] %main() -> [0x655a67ef3826115b29aeb0d774a11922, 0x1.1e00a69cf7813p63, 0x655a67ef3826115b43e1e00a69cf7813]
[aarch64] %main() -> [0x655a67ef3826115b29aeb0d774a11922, 0x1.1e00a69cf7813p63, 0x000000000000000043e1e00a69cf7813]

When I try removing the call instruction from block0, or removing the atomic_store instruction, or removing the store little instruction, any of these changes makes the x86 result consistent with the aarch64 result. But in theory, these actions should have no effect on v7.

[x86    ] %main() -> [0x655a67ef3826115b29aeb0d774a11922, 0x1.1e00a69cf7813p63, 0x000000000000000043e1e00a69cf7813]
[aarch64] %main() -> [0x655a67ef3826115b29aeb0d774a11922, 0x1.1e00a69cf7813p63, 0x000000000000000043e1e00a69cf7813]

Wasmtime GitHub notifications bot (Jun 12 2025 at 20:34):

alexcrichton closed issue #11024:

.clif Test Case

test optimize
    set opt_level=none
    set preserve_frame_pointers=true
    set enable_multi_ret_implicit_sret=true

function u1:1() -> f64 fast {
    ss0 = explicit_slot 32

block0():
    v1 = f64const 0x1.d707d3c2dd690p-1

    v7 = stack_addr.i64 ss0
    v8 = load.f32x4 little v7+8
    return v1
}
function %main() -> i64x2, f64,f64x2 fast {
    ss0 = explicit_slot 32
    ss1 = explicit_slot 32
    sig0 = () -> f64 fast
    fn0 = u1:1 sig0
    const0 = 0x655a67ef3826115b29aeb0d774a11922

block0:

    v1 = iconst.i64 -8142416529737083292
    v2 = vconst.i64x2 const0

    v3 = stack_addr.i64 ss0
    v4 = stack_addr.i64 ss1

    store little v1, v3
    store little v1, v3+8
    store little v1, v3+16
    store little v1, v3+24

    store little v1, v4
    store little v1, v4+8
    store little v1, v4+16
    store little v1, v4+24

    v5 = stack_addr.i64 ss1
    atomic_store little v1, v5

    v6 = fcvt_from_uint.f64 v1
    v7 = scalar_to_vector.f64x2 v6

    v8 = call fn0()
    return v2,v6,v7
}


; print: %main()

Result

[x86    ] %main() -> [0x655a67ef3826115b29aeb0d774a11922, 0x1.1e00a69cf7813p63, 0x655a67ef3826115b43e1e00a69cf7813]
[aarch64] %main() -> [0x655a67ef3826115b29aeb0d774a11922, 0x1.1e00a69cf7813p63, 0x000000000000000043e1e00a69cf7813]

When I try removing the call instruction from block0, or removing the atomic_store instruction, or removing the store little instruction, any of these changes makes the x86 result consistent with the aarch64 result. But in theory, these actions should have no effect on v7.

[x86    ] %main() -> [0x655a67ef3826115b29aeb0d774a11922, 0x1.1e00a69cf7813p63, 0x000000000000000043e1e00a69cf7813]
[aarch64] %main() -> [0x655a67ef3826115b29aeb0d774a11922, 0x1.1e00a69cf7813p63, 0x000000000000000043e1e00a69cf7813]

Last updated: Dec 06 2025 at 06:05 UTC