Files
wasmtime/crates/fuzzing/src/generators/value.rs
Andrew Brown cd982c5a3f [fuzz] Add SIMD to single-instruction generator (#4778)
* [fuzz] Add SIMD to single-instruction generator

This change extends the single-instruction generator with most of the
SIMD instructions. Examples of instructions that were excluded are: all
memory-related instructions, any instruction with an immediate.

* [fuzz] Generate V128s with known values from each type

To better cover the fuzzing search space, `DiffValue` will generate
better known values for the `V128` type. First, it uses arbitrary data
to select a sub-type (e.g., `I8x16`, `F32x4`, etc.) and then it fills in
the bytes by generating biased values for each of the lanes.

* [fuzz] Canonicalize NaN values in SIMD lanes

This change ports the NaN canonicalization logic from `wasm-smith`
([here]) to the single-instruction generator.

[here]: https://github.com/bytecodealliance/wasm-tools/blob/6c127a6/crates/wasm-smith/src/core/code_builder.rs#L927
2022-09-06 14:54:39 -07:00

310 lines
12 KiB
Rust

//! Generate Wasm values, primarily for differential execution.
use arbitrary::{Arbitrary, Unstructured};
use std::hash::Hash;
/// A value passed to and from evaluation. Note that reference types are not
/// (yet) supported.
#[derive(Clone, Debug)]
#[allow(missing_docs)]
pub enum DiffValue {
I32(i32),
I64(i64),
F32(u32),
F64(u64),
V128(u128),
FuncRef { null: bool },
ExternRef { null: bool },
}
impl DiffValue {
fn ty(&self) -> DiffValueType {
match self {
DiffValue::I32(_) => DiffValueType::I32,
DiffValue::I64(_) => DiffValueType::I64,
DiffValue::F32(_) => DiffValueType::F32,
DiffValue::F64(_) => DiffValueType::F64,
DiffValue::V128(_) => DiffValueType::V128,
DiffValue::FuncRef { .. } => DiffValueType::FuncRef,
DiffValue::ExternRef { .. } => DiffValueType::ExternRef,
}
}
/// Generate a [`DiffValue`] of the given `ty` type.
///
/// This function will bias the returned value 50% of the time towards one
/// of a set of known values (e.g., NaN, -1, 0, infinity, etc.).
pub fn arbitrary_of_type(
u: &mut Unstructured<'_>,
ty: DiffValueType,
) -> arbitrary::Result<Self> {
use DiffValueType::*;
let val = match ty {
I32 => DiffValue::I32(biased_arbitrary_value(u, KNOWN_I32_VALUES)?),
I64 => DiffValue::I64(biased_arbitrary_value(u, KNOWN_I64_VALUES)?),
F32 => {
// TODO once `to_bits` is stable as a `const` function, move
// this to a `const` definition.
let known_f32_values = &[
f32::NAN.to_bits(),
f32::INFINITY.to_bits(),
f32::NEG_INFINITY.to_bits(),
f32::MIN.to_bits(),
(-1.0f32).to_bits(),
(0.0f32).to_bits(),
(1.0f32).to_bits(),
f32::MAX.to_bits(),
];
let bits = biased_arbitrary_value(u, known_f32_values)?;
// If the chosen bits are NAN then always use the canonical bit
// pattern of nan to enable better compatibility with engines
// where arbitrary nan patterns can't make their way into wasm
// (e.g. v8 through JS can't do that).
let bits = if f32::from_bits(bits).is_nan() {
f32::NAN.to_bits()
} else {
bits
};
DiffValue::F32(bits)
}
F64 => {
// TODO once `to_bits` is stable as a `const` function, move
// this to a `const` definition.
let known_f64_values = &[
f64::NAN.to_bits(),
f64::INFINITY.to_bits(),
f64::NEG_INFINITY.to_bits(),
f64::MIN.to_bits(),
(-1.0f64).to_bits(),
(0.0f64).to_bits(),
(1.0f64).to_bits(),
f64::MAX.to_bits(),
];
let bits = biased_arbitrary_value(u, known_f64_values)?;
// See `f32` above for why canonical nan patterns are always
// used.
let bits = if f64::from_bits(bits).is_nan() {
f64::NAN.to_bits()
} else {
bits
};
DiffValue::F64(bits)
}
V128 => {
// Generate known values for each sub-type of V128.
let ty: DiffSimdTy = u.arbitrary()?;
match ty {
DiffSimdTy::I8x16 => {
let mut i8 = || biased_arbitrary_value(u, KNOWN_I8_VALUES).map(|b| b as u8);
let vector = u128::from_le_bytes([
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
i8()?,
]);
DiffValue::V128(vector)
}
DiffSimdTy::I16x8 => {
let mut i16 =
|| biased_arbitrary_value(u, KNOWN_I16_VALUES).map(i16::to_le_bytes);
let vector: Vec<u8> = i16()?
.into_iter()
.chain(i16()?)
.chain(i16()?)
.chain(i16()?)
.chain(i16()?)
.chain(i16()?)
.chain(i16()?)
.chain(i16()?)
.collect();
DiffValue::V128(u128::from_le_bytes(vector.try_into().unwrap()))
}
DiffSimdTy::I32x4 => {
let mut i32 =
|| biased_arbitrary_value(u, KNOWN_I32_VALUES).map(i32::to_le_bytes);
let vector: Vec<u8> = i32()?
.into_iter()
.chain(i32()?)
.chain(i32()?)
.chain(i32()?)
.collect();
DiffValue::V128(u128::from_le_bytes(vector.try_into().unwrap()))
}
DiffSimdTy::I64x2 => {
let mut i64 =
|| biased_arbitrary_value(u, KNOWN_I64_VALUES).map(i64::to_le_bytes);
let vector: Vec<u8> = i64()?.into_iter().chain(i64()?).collect();
DiffValue::V128(u128::from_le_bytes(vector.try_into().unwrap()))
}
DiffSimdTy::F32x4 => {
let mut f32 = || {
Self::arbitrary_of_type(u, DiffValueType::F32).map(|v| match v {
DiffValue::F32(v) => v.to_le_bytes(),
_ => unreachable!(),
})
};
let vector: Vec<u8> = f32()?
.into_iter()
.chain(f32()?)
.chain(f32()?)
.chain(f32()?)
.collect();
DiffValue::V128(u128::from_le_bytes(vector.try_into().unwrap()))
}
DiffSimdTy::F64x2 => {
let mut f64 = || {
Self::arbitrary_of_type(u, DiffValueType::F64).map(|v| match v {
DiffValue::F64(v) => v.to_le_bytes(),
_ => unreachable!(),
})
};
let vector: Vec<u8> = f64()?.into_iter().chain(f64()?).collect();
DiffValue::V128(u128::from_le_bytes(vector.try_into().unwrap()))
}
}
}
// TODO: this isn't working in most engines so just always pass a
// null in which if an engine supports this is should at least
// support doing that.
FuncRef => DiffValue::FuncRef { null: true },
ExternRef => DiffValue::ExternRef { null: true },
};
arbitrary::Result::Ok(val)
}
}
const KNOWN_I8_VALUES: &[i8] = &[i8::MIN, -1, 0, 1, i8::MAX];
const KNOWN_I16_VALUES: &[i16] = &[i16::MIN, -1, 0, 1, i16::MAX];
const KNOWN_I32_VALUES: &[i32] = &[i32::MIN, -1, 0, 1, i32::MAX];
const KNOWN_I64_VALUES: &[i64] = &[i64::MIN, -1, 0, 1, i64::MAX];
/// Helper function to pick a known value from the list of `known_values` half
/// the time.
fn biased_arbitrary_value<'a, T>(
u: &mut Unstructured<'a>,
known_values: &[T],
) -> arbitrary::Result<T>
where
T: Arbitrary<'a> + Copy,
{
let pick_from_known_values: bool = u.arbitrary()?;
if pick_from_known_values {
Ok(*u.choose(known_values)?)
} else {
u.arbitrary()
}
}
impl<'a> Arbitrary<'a> for DiffValue {
fn arbitrary(u: &mut Unstructured<'a>) -> arbitrary::Result<Self> {
let ty: DiffValueType = u.arbitrary()?;
DiffValue::arbitrary_of_type(u, ty)
}
}
impl Hash for DiffValue {
fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
self.ty().hash(state);
match self {
DiffValue::I32(n) => n.hash(state),
DiffValue::I64(n) => n.hash(state),
DiffValue::F32(n) => n.hash(state),
DiffValue::F64(n) => n.hash(state),
DiffValue::V128(n) => n.hash(state),
DiffValue::ExternRef { null } => null.hash(state),
DiffValue::FuncRef { null } => null.hash(state),
}
}
}
/// Implement equality checks. Note that floating-point values are not compared
/// bit-for-bit in the case of NaNs: because Wasm floating-point numbers may be
/// [arithmetic NaNs with arbitrary payloads] and Wasm operations are [not
/// required to propagate NaN payloads], we simply check that both sides are
/// NaNs here. We could be more strict, though: we could check that the NaN
/// signs are equal and that [canonical NaN payloads remain canonical].
///
/// [arithmetic NaNs with arbitrary payloads]:
/// https://webassembly.github.io/spec/core/bikeshed/index.html#floating-point%E2%91%A0
/// [not required to propagate NaN payloads]:
/// https://webassembly.github.io/spec/core/bikeshed/index.html#floating-point-operations%E2%91%A0
/// [canonical NaN payloads remain canonical]:
/// https://webassembly.github.io/spec/core/bikeshed/index.html#nan-propagation%E2%91%A0
impl PartialEq for DiffValue {
fn eq(&self, other: &Self) -> bool {
match (self, other) {
(Self::I32(l0), Self::I32(r0)) => l0 == r0,
(Self::I64(l0), Self::I64(r0)) => l0 == r0,
(Self::V128(l0), Self::V128(r0)) => l0 == r0,
(Self::F32(l0), Self::F32(r0)) => {
let l0 = f32::from_bits(*l0);
let r0 = f32::from_bits(*r0);
l0 == r0 || (l0.is_nan() && r0.is_nan())
}
(Self::F64(l0), Self::F64(r0)) => {
let l0 = f64::from_bits(*l0);
let r0 = f64::from_bits(*r0);
l0 == r0 || (l0.is_nan() && r0.is_nan())
}
(Self::FuncRef { null: a }, Self::FuncRef { null: b }) => a == b,
(Self::ExternRef { null: a }, Self::ExternRef { null: b }) => a == b,
_ => false,
}
}
}
/// Enumerate the supported value types.
#[derive(Copy, Clone, Debug, Arbitrary, Hash)]
#[allow(missing_docs)]
pub enum DiffValueType {
I32,
I64,
F32,
F64,
V128,
FuncRef,
ExternRef,
}
impl TryFrom<wasmtime::ValType> for DiffValueType {
type Error = &'static str;
fn try_from(ty: wasmtime::ValType) -> Result<Self, Self::Error> {
use wasmtime::ValType::*;
match ty {
I32 => Ok(Self::I32),
I64 => Ok(Self::I64),
F32 => Ok(Self::F32),
F64 => Ok(Self::F64),
V128 => Ok(Self::V128),
FuncRef => Ok(Self::FuncRef),
ExternRef => Ok(Self::ExternRef),
}
}
}
/// Enumerate the types of v128.
#[derive(Copy, Clone, Debug, Arbitrary, Hash)]
#[allow(missing_docs)]
pub enum DiffSimdTy {
I8x16,
I16x8,
I32x4,
I64x2,
F32x4,
F64x2,
}