Stream: git-wasmtime

Topic: wasmtime / issue #9143 Cranelift: Lowering fails for `fma...

Wasmtime GitHub notifications bot (Aug 19 2024 at 19:55):

kcvlex added the bug label to Issue #9143.

Wasmtime GitHub notifications bot (Aug 19 2024 at 19:55):

kcvlex added the cranelift label to Issue #9143.

Wasmtime GitHub notifications bot (Aug 19 2024 at 19:55):

kcvlex opened issue #9143:

.clif Test Case

test compile
target x86_64

function %fmax(i64, f32x4) -> f32x4 {
block0(v0: i64, v1: f32x4):
    v2 = load.f32x4 v0
    v3 = fmax v1, v2
    return v3
}

function %fmin(i64, f32x4) -> f32x4 {
block0(v0: i64, v1: f32x4):
    v2 = load.f32x4 v0
    v3 = fmin v1, v2
    return v3
}

Steps to Reproduce

clif-util test ./the-above.clif

Expected Results

Successful

Actual Results

Assertion failure

thread 'worker #0' panicked at cranelift/codegen/src/machinst/lower.rs:1449:13:
assertion failed: !self.inst_sunk.contains(&inst)
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
 ERROR cranelift_filetests::concurrent > FAIL: panicked in worker #0: assertion failed: !self.inst_sunk.contains(&inst)
FAIL ./fmax.clif: panicked in worker #0: assertion failed: !self.inst_sunk.contains(&inst)
1 tests
Error: 1 failure

Versions and Environment

Cranelift version or commit: #2c3e64bc5de0a67963e0e34eabbb6de95ecda0ba

Operating system: Linux

Architecture: X86_64

Extra Info

I tried other similar cases and found that the test passes if there are no load instructions. Also, the problem doesn't occur when inserting a jump like the one below.

test compile
target x86_64

function %fmax(i64, f32x4) -> f32x4 {
block0(v0: i64, v1: f32x4):
    v2 = load.f32x4 v0
    jump block1(v1, v2)

block1(v3: f32x4, v4: f32x4):
    v5 = fmax v3, v4
    return v5
}

Wasmtime GitHub notifications bot (Aug 19 2024 at 20:00):

kcvlex edited issue #9143:

.clif Test Case

test compile
target x86_64

function %fmax(i64, f32x4) -> f32x4 {
block0(v0: i64, v1: f32x4):
    v2 = load.f32x4 v0
    v3 = fmax v1, v2
    return v3
}

function %fmin(i64, f32x4) -> f32x4 {
block0(v0: i64, v1: f32x4):
    v2 = load.f32x4 v0
    v3 = fmin v1, v2
    return v3
}

Steps to Reproduce

clif-util test ./the-above.clif

Expected Results

Successful

Actual Results

Assertion failure

thread 'worker #0' panicked at cranelift/codegen/src/machinst/lower.rs:1449:13:
assertion failed: !self.inst_sunk.contains(&inst)
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
 ERROR cranelift_filetests::concurrent > FAIL: panicked in worker #0: assertion failed: !self.inst_sunk.contains(&inst)
FAIL ./fmax.clif: panicked in worker #0: assertion failed: !self.inst_sunk.contains(&inst)
1 tests
Error: 1 failure

Versions and Environment

Cranelift version or commit: #2c3e64bc5de0a67963e0e34eabbb6de95ecda0ba

Operating system: Linux

Architecture: X86_64

Extra Info

I tried other similar cases and found that the test passes if there are no load instructions. Also, the problem doesn't occur when inserting a jump like the one below.

test compile
target x86_64

function %fmax(i64, f32x4) -> f32x4 {
block0(v0: i64, v1: f32x4):
    v2 = load.f32x4 v0
    jump block1(v1, v2)

block1(v3: f32x4, v4: f32x4):
    v5 = fmax v3, v4
    return v5
}

Wasmtime GitHub notifications bot (Aug 19 2024 at 20:02):

fitzgen commented on issue #9143:

Maybe a regression from https://github.com/bytecodealliance/wasmtime/pull/9137 ? cc @alexcrichton

Wasmtime GitHub notifications bot (Aug 19 2024 at 21:27):

alexcrichton commented on issue #9143:

Bisection points to #5841, "very old" issue, continuing to investigate.

Wasmtime GitHub notifications bot (Aug 19 2024 at 21:33):

alexcrichton commented on issue #9143:

This appears to not affect wasm because wasm always loads/stores as i8x16 and the bitcast afterwards to go to f32x4 thwarts the load sinking here. Given that I think this is just a fix for clif/lowerings which makes it so this doesn't have quite as large of an impact. I'll work on fixing.

Wasmtime GitHub notifications bot (Aug 19 2024 at 21:42):

alexcrichton commented on issue #9143:

Thank you for the clear and concise bug report!

Wasmtime GitHub notifications bot (Aug 19 2024 at 22:06):

fitzgen closed issue #9143:

.clif Test Case

test compile
target x86_64

function %fmax(i64, f32x4) -> f32x4 {
block0(v0: i64, v1: f32x4):
    v2 = load.f32x4 v0
    v3 = fmax v1, v2
    return v3
}

function %fmin(i64, f32x4) -> f32x4 {
block0(v0: i64, v1: f32x4):
    v2 = load.f32x4 v0
    v3 = fmin v1, v2
    return v3
}

Steps to Reproduce

clif-util test ./the-above.clif

Expected Results

Successful

Actual Results

Assertion failure

thread 'worker #0' panicked at cranelift/codegen/src/machinst/lower.rs:1449:13:
assertion failed: !self.inst_sunk.contains(&inst)
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
 ERROR cranelift_filetests::concurrent > FAIL: panicked in worker #0: assertion failed: !self.inst_sunk.contains(&inst)
FAIL ./fmax.clif: panicked in worker #0: assertion failed: !self.inst_sunk.contains(&inst)
1 tests
Error: 1 failure

Versions and Environment

Cranelift version or commit: #2c3e64bc5de0a67963e0e34eabbb6de95ecda0ba

Operating system: Linux

Architecture: X86_64

Extra Info

I tried other similar cases and found that the test passes if there are no load instructions. Also, the problem doesn't occur when inserting a jump like the one below.

test compile
target x86_64

function %fmax(i64, f32x4) -> f32x4 {
block0(v0: i64, v1: f32x4):
    v2 = load.f32x4 v0
    jump block1(v1, v2)

block1(v3: f32x4, v4: f32x4):
    v5 = fmax v3, v4
    return v5
}

Wasmtime GitHub notifications bot (Aug 20 2024 at 10:38):

kcvlex commented on issue #9143:

Thanks for your quick fix!

Last updated: Dec 23 2024 at 13:07 UTC