diff options
Diffstat (limited to 'src/tools/test-float-parse')
18 files changed, 2395 insertions, 0 deletions
diff --git a/src/tools/test-float-parse/Cargo.lock b/src/tools/test-float-parse/Cargo.lock new file mode 100644 index 00000000000..3f60423fed3 --- /dev/null +++ b/src/tools/test-float-parse/Cargo.lock @@ -0,0 +1,75 @@ +# This file is automatically @generated by Cargo. +# It is not intended for manual editing. +version = 3 + +[[package]] +name = "cfg-if" +version = "1.0.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "baf1de4339761588bc0619e3cbc0120ee582ebb74b53b4efbf79117bd2da40fd" + +[[package]] +name = "getrandom" +version = "0.2.10" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "be4136b2a15dd319360be1c07d9933517ccf0be8f16bf62a3bee4f0d618df427" +dependencies = [ + "cfg-if", + "libc", + "wasi", +] + +[[package]] +name = "libc" +version = "0.2.147" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "b4668fb0ea861c1df094127ac5f1da3409a82116a4ba74fca2e58ef927159bb3" + +[[package]] +name = "ppv-lite86" +version = "0.2.17" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "5b40af805b3121feab8a3c29f04d8ad262fa8e0561883e7653e024ae4479e6de" + +[[package]] +name = "rand" +version = "0.8.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "34af8d1a0e25924bc5b7c43c079c942339d8f0a8b57c39049bef581b46327404" +dependencies = [ + "libc", + "rand_chacha", + "rand_core", +] + +[[package]] +name = "rand_chacha" +version = "0.3.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e6c10a63a0fa32252be49d21e7709d4d4baf8d231c2dbce1eaa8141b9b127d88" +dependencies = [ + "ppv-lite86", + "rand_core", +] + +[[package]] +name = "rand_core" +version = "0.6.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ec0be4795e2f6a28069bec0b5ff3e2ac9bafc99e6a9a7dc3547996c5c816922c" +dependencies = [ + "getrandom", +] + +[[package]] +name = "test-float-parse" +version = "0.1.0" +dependencies = [ + "rand", +] + +[[package]] +name = "wasi" +version = "0.11.0+wasi-snapshot-preview1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "9c8d87e72b64a3b4db28d11ce29237c246188f4f51057d65a7eab63b7987e423" diff --git a/src/tools/test-float-parse/Cargo.toml b/src/tools/test-float-parse/Cargo.toml new file mode 100644 index 00000000000..e407e322f9e --- /dev/null +++ b/src/tools/test-float-parse/Cargo.toml @@ -0,0 +1,22 @@ +[package] +name = "test-float-parse" +version = "0.1.0" +edition = "2024" +publish = false + +[dependencies] +indicatif = { version = "0.17.8", default-features = false } +num = "0.4.3" +rand = "0.9.0" +rand_chacha = "0.9.0" +rayon = "1" + +[lib] +name = "test_float_parse" + +[lints.rust.unexpected_cfgs] +level = "warn" +check-cfg = [ + # Internal features aren't marked known config by default + 'cfg(target_has_reliable_f16)', +] diff --git a/src/tools/test-float-parse/README.md b/src/tools/test-float-parse/README.md new file mode 100644 index 00000000000..5e2c43d1cad --- /dev/null +++ b/src/tools/test-float-parse/README.md @@ -0,0 +1,55 @@ +# Float Parsing Tests + +These are tests designed to test decimal to float conversions (`dec2flt`) used +by the standard library. + +It consists of a collection of test generators that each generate a set of +patterns intended to test a specific property. In addition, there are exhaustive +tests (for <= `f32`) and fuzzers (for anything that can't be run exhaustively). + +The generators work as follows: + +- Each generator is a struct that lives somewhere in the `gen` module. Usually + it is generic over a float type. +- These generators must implement `Iterator`, which should return a context type + that can be used to construct a test string (but usually not the string + itself). +- They must also implement the `Generator` trait, which provides a method to + write test context to a string as a test case, as well as some extra metadata. + + The split between context generation and string construction is so that we can + reuse string allocations. +- Each generator gets registered once for each float type. Each of these + generators then get their iterator called, and each test case checked against + the float type's parse implementation. + +Some generators produce decimal strings, others create bit patterns that need to +be bitcasted to the float type, which then uses its `Display` implementation to +write to a string. For these, float to decimal (`flt2dec`) conversions also get +tested, if unintentionally. + +For each test case, the following is done: + +- The test string is parsed to the float type using the standard library's + implementation. +- The test string is parsed separately to a `BigRational`, which acts as a + representation with infinite precision. +- The rational value then gets checked that it is within the float's + representable values (absolute value greater than the smallest number to round + to zero, but less less than the first value to round to infinity). If these + limits are exceeded, check that the parsed float reflects that. +- For real nonzero numbers, the parsed float is converted into a rational using + `significand * 2^exponent`. It is then checked against the actual rational + value, and verified to be within half a bit's precision of the parsed value. + Also it is checked that ties round to even. + +This is all highly parallelized with `rayon`; test generators can run in +parallel, and their tests get chunked and run in parallel. + +There is a simple command line that allows filtering which tests are run, +setting the number of iterations for fuzzing tests, limiting failures, setting +timeouts, etc. See `main.rs` or run with `--help` for options. + +Note that when running via `./x`, only tests that take less than a few minutes +are run by default. Navigate to the crate (or pass `-C` to Cargo) and run it +directly to run all tests or pass specific arguments. diff --git a/src/tools/test-float-parse/src/gen_/exhaustive.rs b/src/tools/test-float-parse/src/gen_/exhaustive.rs new file mode 100644 index 00000000000..01458fb0b60 --- /dev/null +++ b/src/tools/test-float-parse/src/gen_/exhaustive.rs @@ -0,0 +1,42 @@ +use std::fmt::Write; +use std::ops::RangeInclusive; + +use crate::{Float, Generator, Int}; + +/// Test every possible bit pattern. This is infeasible to run on any float types larger than +/// `f32` (which takes about an hour). +pub struct Exhaustive<F: Float> { + iter: RangeInclusive<F::Int>, +} + +impl<F: Float> Generator<F> for Exhaustive<F> +where + RangeInclusive<F::Int>: Iterator<Item = F::Int>, +{ + const SHORT_NAME: &'static str = "exhaustive"; + + type WriteCtx = F; + + fn total_tests() -> u64 { + 1u64.checked_shl(F::Int::BITS).expect("More than u64::MAX tests") + } + + fn new() -> Self { + Self { iter: F::Int::ZERO..=F::Int::MAX } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + write!(s, "{ctx:e}").unwrap(); + } +} + +impl<F: Float> Iterator for Exhaustive<F> +where + RangeInclusive<F::Int>: Iterator<Item = F::Int>, +{ + type Item = F; + + fn next(&mut self) -> Option<Self::Item> { + Some(F::from_bits(self.iter.next()?)) + } +} diff --git a/src/tools/test-float-parse/src/gen_/exponents.rs b/src/tools/test-float-parse/src/gen_/exponents.rs new file mode 100644 index 00000000000..3748e9d380c --- /dev/null +++ b/src/tools/test-float-parse/src/gen_/exponents.rs @@ -0,0 +1,95 @@ +use std::fmt::Write; +use std::ops::RangeInclusive; + +use crate::traits::BoxGenIter; +use crate::{Float, Generator}; + +const SMALL_COEFF_MAX: i32 = 10_000; +const SMALL_EXP_MAX: i32 = 300; + +const SMALL_COEFF_RANGE: RangeInclusive<i32> = (-SMALL_COEFF_MAX)..=SMALL_COEFF_MAX; +const SMALL_EXP_RANGE: RangeInclusive<i32> = (-SMALL_EXP_MAX)..=SMALL_EXP_MAX; + +const LARGE_COEFF_RANGE: RangeInclusive<u32> = 0..=100_000; +const LARGE_EXP_RANGE: RangeInclusive<u32> = 300..=350; + +/// Check exponential values around zero. +pub struct SmallExponents<F: Float> { + iter: BoxGenIter<Self, F>, +} + +impl<F: Float> Generator<F> for SmallExponents<F> { + const NAME: &'static str = "small exponents"; + const SHORT_NAME: &'static str = "small exp"; + + /// `(coefficient, exponent)` + type WriteCtx = (i32, i32); + + fn total_tests() -> u64 { + ((1 + SMALL_COEFF_RANGE.end() - SMALL_COEFF_RANGE.start()) + * (1 + SMALL_EXP_RANGE.end() - SMALL_EXP_RANGE.start())) + .try_into() + .unwrap() + } + + fn new() -> Self { + let iter = SMALL_EXP_RANGE.flat_map(|exp| SMALL_COEFF_RANGE.map(move |coeff| (coeff, exp))); + + Self { iter: Box::new(iter) } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + let (coeff, exp) = ctx; + write!(s, "{coeff}e{exp}").unwrap(); + } +} + +impl<F: Float> Iterator for SmallExponents<F> { + type Item = (i32, i32); + + fn next(&mut self) -> Option<Self::Item> { + self.iter.next() + } +} + +/// Check exponential values further from zero. +pub struct LargeExponents<F: Float> { + iter: BoxGenIter<Self, F>, +} + +impl<F: Float> Generator<F> for LargeExponents<F> { + const NAME: &'static str = "large positive exponents"; + const SHORT_NAME: &'static str = "large exp"; + + /// `(coefficient, exponent, is_positive)` + type WriteCtx = (u32, u32, bool); + + fn total_tests() -> u64 { + ((1 + LARGE_EXP_RANGE.end() - LARGE_EXP_RANGE.start()) + * (1 + LARGE_COEFF_RANGE.end() - LARGE_COEFF_RANGE.start()) + * 2) + .into() + } + + fn new() -> Self { + let iter = LARGE_EXP_RANGE + .flat_map(|exp| LARGE_COEFF_RANGE.map(move |coeff| (coeff, exp))) + .flat_map(|(coeff, exp)| [(coeff, exp, false), (coeff, exp, true)]); + + Self { iter: Box::new(iter) } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + let (coeff, exp, is_positive) = ctx; + let sign = if is_positive { "" } else { "-" }; + write!(s, "{sign}{coeff}e{exp}").unwrap(); + } +} + +impl<F: Float> Iterator for LargeExponents<F> { + type Item = (u32, u32, bool); + + fn next(&mut self) -> Option<Self::Item> { + self.iter.next() + } +} diff --git a/src/tools/test-float-parse/src/gen_/fuzz.rs b/src/tools/test-float-parse/src/gen_/fuzz.rs new file mode 100644 index 00000000000..1d6c5562a14 --- /dev/null +++ b/src/tools/test-float-parse/src/gen_/fuzz.rs @@ -0,0 +1,87 @@ +use std::any::{TypeId, type_name}; +use std::collections::BTreeMap; +use std::fmt::Write; +use std::marker::PhantomData; +use std::ops::Range; +use std::sync::Mutex; + +use rand::Rng; +use rand::distr::{Distribution, StandardUniform}; +use rand_chacha::ChaCha8Rng; +use rand_chacha::rand_core::SeedableRng; + +use crate::{Float, Generator, Int, SEED}; + +/// Mapping of float types to the number of iterations that should be run. +/// +/// We could probably make `Generator::new` take an argument instead of the global state, +/// but we only load this once so it works. +static FUZZ_COUNTS: Mutex<BTreeMap<TypeId, u64>> = Mutex::new(BTreeMap::new()); + +/// Generic fuzzer; just tests deterministic random bit patterns N times. +pub struct Fuzz<F> { + iter: Range<u64>, + rng: ChaCha8Rng, + /// Allow us to use generics in `Iterator`. + marker: PhantomData<F>, +} + +impl<F: Float> Fuzz<F> { + /// Register how many iterations the fuzzer should run for a type. Uses some logic by + /// default, but if `from_cfg` is `Some`, that will be used instead. + pub fn set_iterations(from_cfg: Option<u64>) { + let count = if let Some(cfg_count) = from_cfg { + cfg_count + } else if F::BITS <= crate::MAX_BITS_FOR_EXHAUUSTIVE { + // If we run exhaustively, still fuzz but only do half as many bits. The only goal here is + // to catch failures from e.g. high bit patterns before exhaustive tests would get to them. + (F::Int::MAX >> (F::BITS / 2)).try_into().unwrap() + } else { + // Eveything bigger gets a fuzz test with as many iterations as `f32` exhaustive. + u32::MAX.into() + }; + + let _ = FUZZ_COUNTS.lock().unwrap().insert(TypeId::of::<F>(), count); + } +} + +impl<F: Float> Generator<F> for Fuzz<F> +where + StandardUniform: Distribution<<F as Float>::Int>, +{ + const SHORT_NAME: &'static str = "fuzz"; + + type WriteCtx = F; + + fn total_tests() -> u64 { + *FUZZ_COUNTS + .lock() + .unwrap() + .get(&TypeId::of::<F>()) + .unwrap_or_else(|| panic!("missing fuzz count for {}", type_name::<F>())) + } + + fn new() -> Self { + let rng = ChaCha8Rng::from_seed(SEED); + + Self { iter: 0..Self::total_tests(), rng, marker: PhantomData } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + write!(s, "{ctx:e}").unwrap(); + } +} + +impl<F: Float> Iterator for Fuzz<F> +where + StandardUniform: Distribution<<F as Float>::Int>, +{ + type Item = <Self as Generator<F>>::WriteCtx; + + fn next(&mut self) -> Option<Self::Item> { + let _ = self.iter.next()?; + let i: F::Int = self.rng.random(); + + Some(F::from_bits(i)) + } +} diff --git a/src/tools/test-float-parse/src/gen_/integers.rs b/src/tools/test-float-parse/src/gen_/integers.rs new file mode 100644 index 00000000000..070d188e88c --- /dev/null +++ b/src/tools/test-float-parse/src/gen_/integers.rs @@ -0,0 +1,104 @@ +use std::fmt::Write; +use std::ops::{Range, RangeInclusive}; + +use crate::traits::BoxGenIter; +use crate::{Float, Generator}; + +const SMALL_MAX_POW2: u32 = 19; + +/// All values up to the max power of two +const SMALL_VALUES: RangeInclusive<i32> = { + let max = 1i32 << SMALL_MAX_POW2; + (-max)..=max +}; + +/// Large values only get tested around powers of two +const LARGE_POWERS: Range<u32> = SMALL_MAX_POW2..128; + +/// We perturbe each large value around these ranges +const LARGE_PERTURBATIONS: RangeInclusive<i128> = -256..=256; + +/// Test all integers up to `2 ^ MAX_POW2` +pub struct SmallInt { + iter: RangeInclusive<i32>, +} + +impl<F: Float> Generator<F> for SmallInt { + const NAME: &'static str = "small integer values"; + const SHORT_NAME: &'static str = "int small"; + + type WriteCtx = i32; + + fn total_tests() -> u64 { + (SMALL_VALUES.end() + 1 - SMALL_VALUES.start()).try_into().unwrap() + } + + fn new() -> Self { + Self { iter: SMALL_VALUES } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + write!(s, "{ctx}").unwrap(); + } +} + +impl Iterator for SmallInt { + type Item = i32; + + fn next(&mut self) -> Option<Self::Item> { + self.iter.next() + } +} + +/// Test much bigger integers than [`SmallInt`]. +pub struct LargeInt<F: Float> { + iter: BoxGenIter<Self, F>, +} + +impl<F: Float> LargeInt<F> { + const EDGE_CASES: [i128; 7] = [ + i32::MIN as i128, + i32::MAX as i128, + i64::MIN as i128, + i64::MAX as i128, + u64::MAX as i128, + i128::MIN, + i128::MAX, + ]; +} + +impl<F: Float> Generator<F> for LargeInt<F> { + const NAME: &'static str = "large integer values"; + const SHORT_NAME: &'static str = "int large"; + + type WriteCtx = i128; + + fn total_tests() -> u64 { + u64::try_from( + (i128::from(LARGE_POWERS.end - LARGE_POWERS.start) + + i128::try_from(Self::EDGE_CASES.len()).unwrap()) + * (LARGE_PERTURBATIONS.end() + 1 - LARGE_PERTURBATIONS.start()), + ) + .unwrap() + } + + fn new() -> Self { + let iter = LARGE_POWERS + .map(|pow| 1i128 << pow) + .chain(Self::EDGE_CASES) + .flat_map(|base| LARGE_PERTURBATIONS.map(move |perturb| base.saturating_add(perturb))); + + Self { iter: Box::new(iter) } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + write!(s, "{ctx}").unwrap(); + } +} +impl<F: Float> Iterator for LargeInt<F> { + type Item = i128; + + fn next(&mut self) -> Option<Self::Item> { + self.iter.next() + } +} diff --git a/src/tools/test-float-parse/src/gen_/long_fractions.rs b/src/tools/test-float-parse/src/gen_/long_fractions.rs new file mode 100644 index 00000000000..b75148b779c --- /dev/null +++ b/src/tools/test-float-parse/src/gen_/long_fractions.rs @@ -0,0 +1,58 @@ +use std::char; +use std::fmt::Write; + +use crate::{Float, Generator}; + +/// Number of decimal digits to check (all of them). +const MAX_DIGIT: u32 = 9; +/// Test with this many decimals in the string. +const MAX_DECIMALS: usize = 410; +const PREFIX: &str = "0."; + +/// Test e.g. `0.1`, `0.11`, `0.111`, `0.1111`, ..., `0.2`, `0.22`, ... +pub struct RepeatingDecimal { + digit: u32, + buf: String, +} + +impl<F: Float> Generator<F> for RepeatingDecimal { + const NAME: &'static str = "repeating decimal"; + const SHORT_NAME: &'static str = "dec rep"; + + type WriteCtx = String; + + fn total_tests() -> u64 { + u64::from(MAX_DIGIT + 1) * u64::try_from(MAX_DECIMALS + 1).unwrap() + 1 + } + + fn new() -> Self { + Self { digit: 0, buf: PREFIX.to_owned() } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + *s = ctx; + } +} + +impl Iterator for RepeatingDecimal { + type Item = String; + + fn next(&mut self) -> Option<Self::Item> { + if self.digit > MAX_DIGIT { + return None; + } + + let digit = self.digit; + let inc_digit = self.buf.len() - PREFIX.len() > MAX_DECIMALS; + + if inc_digit { + // Reset the string + self.buf.clear(); + self.digit += 1; + self.buf.write_str(PREFIX).unwrap(); + } + + self.buf.push(char::from_digit(digit, 10).unwrap()); + Some(self.buf.clone()) + } +} diff --git a/src/tools/test-float-parse/src/gen_/many_digits.rs b/src/tools/test-float-parse/src/gen_/many_digits.rs new file mode 100644 index 00000000000..741e11437fe --- /dev/null +++ b/src/tools/test-float-parse/src/gen_/many_digits.rs @@ -0,0 +1,84 @@ +use std::char; +use std::fmt::Write; +use std::marker::PhantomData; +use std::ops::{Range, RangeInclusive}; + +use rand::distr::{Distribution, Uniform}; +use rand::{Rng, SeedableRng}; +use rand_chacha::ChaCha8Rng; + +use crate::{Float, Generator, SEED}; + +/// Total iterations +const ITERATIONS: u64 = 5_000_000; + +/// Possible lengths of the string, excluding decimals and exponents +const POSSIBLE_NUM_DIGITS: RangeInclusive<usize> = 100..=400; + +/// Range of possible exponents +const EXP_RANGE: Range<i32> = -4500..4500; + +/// Try strings of random digits. +pub struct RandDigits<F> { + rng: ChaCha8Rng, + iter: Range<u64>, + uniform: Uniform<u32>, + /// Allow us to use generics in `Iterator`. + marker: PhantomData<F>, +} + +impl<F: Float> Generator<F> for RandDigits<F> { + const NAME: &'static str = "random digits"; + + const SHORT_NAME: &'static str = "rand digits"; + + type WriteCtx = String; + + fn total_tests() -> u64 { + ITERATIONS + } + + fn new() -> Self { + let rng = ChaCha8Rng::from_seed(SEED); + let range = Uniform::try_from(0..10).unwrap(); + + Self { rng, iter: 0..ITERATIONS, uniform: range, marker: PhantomData } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + *s = ctx; + } +} + +impl<F: Float> Iterator for RandDigits<F> { + type Item = String; + + fn next(&mut self) -> Option<Self::Item> { + let _ = self.iter.next()?; + let num_digits = self.rng.random_range(POSSIBLE_NUM_DIGITS); + let has_decimal = self.rng.random_bool(0.2); + let has_exp = self.rng.random_bool(0.2); + + let dec_pos = if has_decimal { Some(self.rng.random_range(0..num_digits)) } else { None }; + + let mut s = String::with_capacity(num_digits); + + for pos in 0..num_digits { + let digit = char::from_digit(self.uniform.sample(&mut self.rng), 10).unwrap(); + s.push(digit); + + if let Some(dec_pos) = dec_pos { + if pos == dec_pos { + s.push('.'); + } + } + } + + if has_exp { + let exp = self.rng.random_range(EXP_RANGE); + write!(s, "e{exp}").unwrap(); + } + + Some(s) + } +} diff --git a/src/tools/test-float-parse/src/gen_/sparse.rs b/src/tools/test-float-parse/src/gen_/sparse.rs new file mode 100644 index 00000000000..72b65d4ce7f --- /dev/null +++ b/src/tools/test-float-parse/src/gen_/sparse.rs @@ -0,0 +1,99 @@ +use std::fmt::Write; + +use crate::traits::BoxGenIter; +use crate::{Float, Generator}; + +const POWERS_OF_TWO: [u128; 128] = make_powers_of_two(); + +const fn make_powers_of_two() -> [u128; 128] { + let mut ret = [0; 128]; + let mut i = 0; + while i < 128 { + ret[i] = 1 << i; + i += 1; + } + + ret +} + +/// Can't clone this result because of lifetime errors, just use a macro. +macro_rules! pow_iter { + () => { + (0..F::BITS).map(|i| F::Int::try_from(POWERS_OF_TWO[i as usize]).unwrap()) + }; +} + +/// Test all numbers that include three 1s in the binary representation as integers. +pub struct FewOnesInt<F: Float> +where + FewOnesInt<F>: Generator<F>, +{ + iter: BoxGenIter<Self, F>, +} + +impl<F: Float> Generator<F> for FewOnesInt<F> +where + <F::Int as TryFrom<u128>>::Error: std::fmt::Debug, +{ + const SHORT_NAME: &'static str = "few ones int"; + + type WriteCtx = F::Int; + + fn total_tests() -> u64 { + u64::from(F::BITS).pow(3) + } + + fn new() -> Self { + let iter = pow_iter!() + .flat_map(move |a| pow_iter!().map(move |b| (a, b))) + .flat_map(move |(a, b)| pow_iter!().map(move |c| (a, b, c))) + .map(|(a, b, c)| a | b | c); + + Self { iter: Box::new(iter) } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + write!(s, "{ctx}").unwrap(); + } +} + +impl<F: Float> Iterator for FewOnesInt<F> { + type Item = F::Int; + + fn next(&mut self) -> Option<Self::Item> { + self.iter.next() + } +} + +/// Similar to `FewOnesInt` except test those bit patterns as a float. +pub struct FewOnesFloat<F: Float>(FewOnesInt<F>); + +impl<F: Float> Generator<F> for FewOnesFloat<F> +where + <F::Int as TryFrom<u128>>::Error: std::fmt::Debug, +{ + const NAME: &'static str = "few ones float"; + const SHORT_NAME: &'static str = "few ones float"; + + type WriteCtx = F; + + fn total_tests() -> u64 { + FewOnesInt::<F>::total_tests() + } + + fn new() -> Self { + Self(FewOnesInt::new()) + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + write!(s, "{ctx:e}").unwrap(); + } +} + +impl<F: Float> Iterator for FewOnesFloat<F> { + type Item = F; + + fn next(&mut self) -> Option<Self::Item> { + self.0.next().map(|i| F::from_bits(i)) + } +} diff --git a/src/tools/test-float-parse/src/gen_/spot_checks.rs b/src/tools/test-float-parse/src/gen_/spot_checks.rs new file mode 100644 index 00000000000..18691f9d6cf --- /dev/null +++ b/src/tools/test-float-parse/src/gen_/spot_checks.rs @@ -0,0 +1,101 @@ +use std::fmt::Write; + +use crate::traits::{Float, Generator}; + +const SPECIAL: &[&str] = &[ + "inf", "Inf", "iNf", "INF", "-inf", "-Inf", "-iNf", "-INF", "+inf", "+Inf", "+iNf", "+INF", + "nan", "NaN", "NAN", "nAn", "-nan", "-NaN", "-NAN", "-nAn", "+nan", "+NaN", "+NAN", "+nAn", + "1", "-1", "+1", "1e1", "-1e1", "+1e1", "1e-1", "-1e-1", "+1e-1", "1e+1", "-1e+1", "+1e+1", + "1E1", "-1E1", "+1E1", "1E-1", "-1E-1", "+1E-1", "1E+1", "-1E+1", "+1E+1", "0", "-0", "+0", +]; + +/// Check various non-numeric special strings. +pub struct Special { + iter: std::slice::Iter<'static, &'static str>, +} + +impl<F: Float> Generator<F> for Special { + const NAME: &'static str = "special values"; + + const SHORT_NAME: &'static str = "special"; + + type WriteCtx = &'static str; + + fn total_tests() -> u64 { + SPECIAL.len().try_into().unwrap() + } + + fn new() -> Self { + Self { iter: SPECIAL.iter() } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + s.write_str(ctx).unwrap(); + } +} + +impl Iterator for Special { + type Item = &'static str; + + fn next(&mut self) -> Option<Self::Item> { + self.iter.next().copied() + } +} + +/// Strings that we know have failed in the past +const REGRESSIONS: &[&str] = &[ + // From <https://github.com/rust-lang/rust/issues/31407> + "1234567890123456789012345678901234567890e-340", + "2.225073858507201136057409796709131975934819546351645648023426109724822222021076945516529523908135087914149158913039621106870086438694594645527657207407820621743379988141063267329253552286881372149012981122451451889849057222307285255133155755015914397476397983411801999323962548289017107081850690630666655994938275772572015763062690663332647565300009245888316433037779791869612049497390377829704905051080609940730262937128958950003583799967207254304360284078895771796150945516748243471030702609144621572289880258182545180325707018860872113128079512233426288368622321503775666622503982534335974568884423900265498198385487948292206894721689831099698365846814022854243330660339850886445804001034933970427567186443383770486037861622771738545623065874679014086723327636718749999999999999999999999999999999999999e-308", + "2.22507385850720113605740979670913197593481954635164564802342610972482222202107694551652952390813508791414915891303962110687008643869459464552765720740782062174337998814106326732925355228688137214901298112245145188984905722230728525513315575501591439747639798341180199932396254828901710708185069063066665599493827577257201576306269066333264756530000924588831643303777979186961204949739037782970490505108060994073026293712895895000358379996720725430436028407889577179615094551674824347103070260914462157228988025818254518032570701886087211312807951223342628836862232150377566662250398253433597456888442390026549819838548794829220689472168983109969836584681402285424333066033985088644580400103493397042756718644338377048603786162277173854562306587467901408672332763671875e-308", + "0.0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000222507385850720138309023271733240406421921598046233183055332741688720443481391819585428315901251102056406733973103581100515243416155346010885601238537771882113077799353200233047961014744258363607192156504694250373420837525080665061665815894872049117996859163964850063590877011830487479978088775374994945158045160505091539985658247081864511353793580499211598108576605199243335211435239014879569960959128889160299264151106346631339366347758651302937176204732563178148566435087212282863764204484681140761391147706280168985324411002416144742161856716615054015428508471675290190316132277889672970737312333408698898317506783884692609277397797285865965494109136909540613646756870239867831529068098461721092462539672851562500000000000000001", + "179769313486231580793728971405303415079934132710037826936173778980444968292764750946649017977587207096330286416692887910946555547851940402630657488671505820681908902000708383676273854845817711531764475730270069855571366959622842914819860834936475292719074168444365510704342711559699508093042880177904174497791.9999999999999999999999999999999999999999999999999999999999999999999999", + "2.47032822920623272e-324", + "6.631236871469758276785396630275967243399099947355303144249971758736286630139265439618068200788048744105960420552601852889715006376325666595539603330361800519107591783233358492337208057849499360899425128640718856616503093444922854759159988160304439909868291973931426625698663157749836252274523485312442358651207051292453083278116143932569727918709786004497872322193856150225415211997283078496319412124640111777216148110752815101775295719811974338451936095907419622417538473679495148632480391435931767981122396703443803335529756003353209830071832230689201383015598792184172909927924176339315507402234836120730914783168400715462440053817592702766213559042115986763819482654128770595766806872783349146967171293949598850675682115696218943412532098591327667236328125E-316", + "3.237883913302901289588352412501532174863037669423108059901297049552301970670676565786835742587799557860615776559838283435514391084153169252689190564396459577394618038928365305143463955100356696665629202017331344031730044369360205258345803431471660032699580731300954848363975548690010751530018881758184174569652173110473696022749934638425380623369774736560008997404060967498028389191878963968575439222206416981462690113342524002724385941651051293552601421155333430225237291523843322331326138431477823591142408800030775170625915670728657003151953664260769822494937951845801530895238439819708403389937873241463484205608000027270531106827387907791444918534771598750162812548862768493201518991668028251730299953143924168545708663913273994694463908672332763671875E-319", + "6.953355807847677105972805215521891690222119817145950754416205607980030131549636688806115726399441880065386399864028691275539539414652831584795668560082999889551357784961446896042113198284213107935110217162654939802416034676213829409720583759540476786936413816541621287843248433202369209916612249676005573022703244799714622116542188837770376022371172079559125853382801396219552418839469770514904192657627060319372847562301074140442660237844114174497210955449896389180395827191602886654488182452409583981389442783377001505462015745017848754574668342161759496661766020028752888783387074850773192997102997936619876226688096314989645766000479009083731736585750335262099860150896718774401964796827166283225641992040747894382698751809812609536720628966577351093292236328125E-310", + "3.339068557571188581835713701280943911923401916998521771655656997328440314559615318168849149074662609099998113009465566426808170378434065722991659642619467706034884424989741080790766778456332168200464651593995817371782125010668346652995912233993254584461125868481633343674905074271064409763090708017856584019776878812425312008812326260363035474811532236853359905334625575404216060622858633280744301892470300555678734689978476870369853549413277156622170245846166991655321535529623870646888786637528995592800436177901746286272273374471701452991433047257863864601424252024791567368195056077320885329384322332391564645264143400798619665040608077549162173963649264049738362290606875883456826586710961041737908872035803481241600376705491726170293986797332763671875E-319", + "2.4703282292062327208828439643411068618252990130716238221279284125033775363510437593264991818081799618989828234772285886546332835517796989819938739800539093906315035659515570226392290858392449105184435931802849936536152500319370457678249219365623669863658480757001585769269903706311928279558551332927834338409351978015531246597263579574622766465272827220056374006485499977096599470454020828166226237857393450736339007967761930577506740176324673600968951340535537458516661134223766678604162159680461914467291840300530057530849048765391711386591646239524912623653881879636239373280423891018672348497668235089863388587925628302755995657524455507255189313690836254779186948667994968324049705821028513185451396213837722826145437693412532098591327667236328124999e-324", + "2.4703282292062327208828439643411068618252990130716238221279284125033775363510437593264991818081799618989828234772285886546332835517796989819938739800539093906315035659515570226392290858392449105184435931802849936536152500319370457678249219365623669863658480757001585769269903706311928279558551332927834338409351978015531246597263579574622766465272827220056374006485499977096599470454020828166226237857393450736339007967761930577506740176324673600968951340535537458516661134223766678604162159680461914467291840300530057530849048765391711386591646239524912623653881879636239373280423891018672348497668235089863388587925628302755995657524455507255189313690836254779186948667994968324049705821028513185451396213837722826145437693412532098591327667236328125e-324", + "2.4703282292062327208828439643411068618252990130716238221279284125033775363510437593264991818081799618989828234772285886546332835517796989819938739800539093906315035659515570226392290858392449105184435931802849936536152500319370457678249219365623669863658480757001585769269903706311928279558551332927834338409351978015531246597263579574622766465272827220056374006485499977096599470454020828166226237857393450736339007967761930577506740176324673600968951340535537458516661134223766678604162159680461914467291840300530057530849048765391711386591646239524912623653881879636239373280423891018672348497668235089863388587925628302755995657524455507255189313690836254779186948667994968324049705821028513185451396213837722826145437693412532098591327667236328125001e-324", + "7.4109846876186981626485318930233205854758970392148714663837852375101326090531312779794975454245398856969484704316857659638998506553390969459816219401617281718945106978546710679176872575177347315553307795408549809608457500958111373034747658096871009590975442271004757307809711118935784838675653998783503015228055934046593739791790738723868299395818481660169122019456499931289798411362062484498678713572180352209017023903285791732520220528974020802906854021606612375549983402671300035812486479041385743401875520901590172592547146296175134159774938718574737870961645638908718119841271673056017045493004705269590165763776884908267986972573366521765567941072508764337560846003984904972149117463085539556354188641513168478436313080237596295773983001708984374999e-324", + "7.4109846876186981626485318930233205854758970392148714663837852375101326090531312779794975454245398856969484704316857659638998506553390969459816219401617281718945106978546710679176872575177347315553307795408549809608457500958111373034747658096871009590975442271004757307809711118935784838675653998783503015228055934046593739791790738723868299395818481660169122019456499931289798411362062484498678713572180352209017023903285791732520220528974020802906854021606612375549983402671300035812486479041385743401875520901590172592547146296175134159774938718574737870961645638908718119841271673056017045493004705269590165763776884908267986972573366521765567941072508764337560846003984904972149117463085539556354188641513168478436313080237596295773983001708984375e-324", + "7.4109846876186981626485318930233205854758970392148714663837852375101326090531312779794975454245398856969484704316857659638998506553390969459816219401617281718945106978546710679176872575177347315553307795408549809608457500958111373034747658096871009590975442271004757307809711118935784838675653998783503015228055934046593739791790738723868299395818481660169122019456499931289798411362062484498678713572180352209017023903285791732520220528974020802906854021606612375549983402671300035812486479041385743401875520901590172592547146296175134159774938718574737870961645638908718119841271673056017045493004705269590165763776884908267986972573366521765567941072508764337560846003984904972149117463085539556354188641513168478436313080237596295773983001708984375001e-324", + "94393431193180696942841837085033647913224148539854e-358", + "104308485241983990666713401708072175773165034278685682646111762292409330928739751702404658197872319129036519947435319418387839758990478549477777586673075945844895981012024387992135617064532141489278815239849108105951619997829153633535314849999674266169258928940692239684771590065027025835804863585454872499320500023126142553932654370362024104462255244034053203998964360882487378334860197725139151265590832887433736189468858614521708567646743455601905935595381852723723645799866672558576993978025033590728687206296379801363024094048327273913079612469982585674824156000783167963081616214710691759864332339239688734656548790656486646106983450809073750535624894296242072010195710276073042036425579852459556183541199012652571123898996574563824424330960027873516082763671875e-1075", + "0.00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000247032822920623272088284396434110686182529901307162382212792841250337753635104375932649918180817996189898282347722858865463328355177969898199387398005390939063150356595155702263922908583924491051844359318028499365361525003193704576782492193656236698636584807570015857692699037063119282795585513329278343384093519780155312465972635795746227664652728272200563740064854999770965994704540208281662262378573934507363390079677619305775067401763246736009689513405355374585166611342237666786041621596804619144672918403005300575308490487653917113865916462395249126236538818796362393732804238910186723484976682350898633885879256283027559956575244555072551893136908362547791869486679949683240497058210285131854513962138377228261454376934125320985913276672363281249", + "0.00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000247032822920623272088284396434110686182529901307162382212792841250337753635104375932649918180817996189898282347722858865463328355177969898199387398005390939063150356595155702263922908583924491051844359318028499365361525003193704576782492193656236698636584807570015857692699037063119282795585513329278343384093519780155312465972635795746227664652728272200563740064854999770965994704540208281662262378573934507363390079677619305775067401763246736009689513405355374585166611342237666786041621596804619144672918403005300575308490487653917113865916462395249126236538818796362393732804238910186723484976682350898633885879256283027559956575244555072551893136908362547791869486679949683240497058210285131854513962138377228261454376934125320985913276672363281251", +]; + +/// Check items that failed in the past. +pub struct RegressionCheck { + iter: std::slice::Iter<'static, &'static str>, +} + +impl<F: Float> Generator<F> for RegressionCheck { + const NAME: &'static str = "regression check"; + + const SHORT_NAME: &'static str = "regression"; + + type WriteCtx = &'static str; + + fn total_tests() -> u64 { + REGRESSIONS.len().try_into().unwrap() + } + + fn new() -> Self { + Self { iter: REGRESSIONS.iter() } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + s.write_str(ctx).unwrap(); + } +} + +impl Iterator for RegressionCheck { + type Item = &'static str; + + fn next(&mut self) -> Option<Self::Item> { + self.iter.next().copied() + } +} diff --git a/src/tools/test-float-parse/src/gen_/subnorm.rs b/src/tools/test-float-parse/src/gen_/subnorm.rs new file mode 100644 index 00000000000..654f324b9b0 --- /dev/null +++ b/src/tools/test-float-parse/src/gen_/subnorm.rs @@ -0,0 +1,108 @@ +use std::fmt::Write; +use std::ops::RangeInclusive; + +use crate::{Float, Generator, Int}; + +/// Spot check some edge cases for subnormals. +pub struct SubnormEdgeCases<F: Float> { + cases: [F::Int; 6], + index: usize, +} + +impl<F: Float> SubnormEdgeCases<F> { + /// Shorthand + const I1: F::Int = F::Int::ONE; + + fn edge_cases() -> [F::Int; 6] { + // Comments use an 8-bit mantissa as a demo + [ + // 0b00000001 + Self::I1, + // 0b10000000 + Self::I1 << (F::MAN_BITS - 1), + // 0b00001000 + Self::I1 << ((F::MAN_BITS / 2) - 1), + // 0b00001111 + Self::I1 << ((F::MAN_BITS / 2) - 1), + // 0b00001111 + Self::I1 << ((F::MAN_BITS / 2) - 1), + // 0b11111111 + F::MAN_MASK, + ] + } +} + +impl<F: Float> Generator<F> for SubnormEdgeCases<F> { + const NAME: &'static str = "subnormal edge cases"; + const SHORT_NAME: &'static str = "subnorm edge"; + + type WriteCtx = F; + + fn new() -> Self { + Self { cases: Self::edge_cases(), index: 0 } + } + + fn total_tests() -> u64 { + Self::edge_cases().len().try_into().unwrap() + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + write!(s, "{ctx:e}").unwrap(); + } +} + +impl<F: Float> Iterator for SubnormEdgeCases<F> { + type Item = F; + + fn next(&mut self) -> Option<Self::Item> { + let i = self.cases.get(self.index)?; + self.index += 1; + + Some(F::from_bits(*i)) + } +} + +/// Test all subnormals up to `1 << 22`. +pub struct SubnormComplete<F: Float> { + iter: RangeInclusive<F::Int>, +} + +impl<F: Float> Generator<F> for SubnormComplete<F> +where + RangeInclusive<F::Int>: Iterator<Item = F::Int>, +{ + const NAME: &'static str = "subnormal"; + const SHORT_NAME: &'static str = "subnorm "; + + type WriteCtx = F; + + fn total_tests() -> u64 { + let iter = Self::new().iter; + (F::Int::ONE + *iter.end() - *iter.start()).try_into().unwrap() + } + + fn new() -> Self { + let upper_lim = if F::MAN_BITS >= 22 { + F::Int::ONE << 22 + } else { + (F::Int::ONE << F::MAN_BITS) - F::Int::ONE + }; + + Self { iter: F::Int::ZERO..=upper_lim } + } + + fn write_string(s: &mut String, ctx: Self::WriteCtx) { + write!(s, "{ctx:e}").unwrap(); + } +} + +impl<F: Float> Iterator for SubnormComplete<F> +where + RangeInclusive<F::Int>: Iterator<Item = F::Int>, +{ + type Item = F; + + fn next(&mut self) -> Option<Self::Item> { + Some(F::from_bits(self.iter.next()?)) + } +} diff --git a/src/tools/test-float-parse/src/lib.rs b/src/tools/test-float-parse/src/lib.rs new file mode 100644 index 00000000000..f590149523b --- /dev/null +++ b/src/tools/test-float-parse/src/lib.rs @@ -0,0 +1,419 @@ +#![feature(f16)] +#![feature(cfg_target_has_reliable_f16_f128)] +#![expect(internal_features)] // reliable_f16_f128 + +mod traits; +mod ui; +mod validate; + +use std::any::type_name; +use std::cmp::min; +use std::ops::RangeInclusive; +use std::process::ExitCode; +use std::sync::OnceLock; +use std::sync::atomic::{AtomicU64, Ordering}; +use std::{fmt, time}; + +use rand::distr::{Distribution, StandardUniform}; +use rayon::prelude::*; +use time::{Duration, Instant}; +use traits::{Float, Generator, Int}; +use validate::CheckError; + +/// Test generators. +mod gen_ { + pub mod exhaustive; + pub mod exponents; + pub mod fuzz; + pub mod integers; + pub mod long_fractions; + pub mod many_digits; + pub mod sparse; + pub mod spot_checks; + pub mod subnorm; +} + +/// How many failures to exit after if unspecified. +const DEFAULT_MAX_FAILURES: u64 = 20; + +/// Register exhaustive tests only for <= 32 bits. No more because it would take years. +const MAX_BITS_FOR_EXHAUUSTIVE: u32 = 32; + +/// If there are more tests than this threshold, the test will be deferred until after all +/// others run (so as to avoid thread pool starvation). They also can be excluded with +/// `--skip-huge`. +const HUGE_TEST_CUTOFF: u64 = 5_000_000; + +/// Seed for tests that use a deterministic RNG. +const SEED: [u8; 32] = *b"3.141592653589793238462643383279"; + +/// Global configuration. +#[derive(Debug)] +pub struct Config { + pub timeout: Duration, + /// Failures per test + pub max_failures: u64, + pub disable_max_failures: bool, + /// If `None`, the default will be used + pub fuzz_count: Option<u64>, + pub skip_huge: bool, +} + +impl Default for Config { + fn default() -> Self { + Self { + timeout: Duration::from_secs(60 * 60 * 3), + max_failures: DEFAULT_MAX_FAILURES, + disable_max_failures: false, + fuzz_count: None, + skip_huge: false, + } + } +} + +/// Collect, filter, and launch all tests. +pub fn run(cfg: Config, include: &[String], exclude: &[String]) -> ExitCode { + // With default parallelism, the CPU doesn't saturate. We don't need to be nice to + // other processes, so do 1.5x to make sure we use all available resources. + let threads = std::thread::available_parallelism().map(Into::into).unwrap_or(0) * 3 / 2; + rayon::ThreadPoolBuilder::new().num_threads(threads).build_global().unwrap(); + + let mut tests = register_tests(&cfg); + println!("registered"); + let initial_tests: Vec<_> = tests.iter().map(|t| t.name.clone()).collect(); + + let unmatched: Vec<_> = include + .iter() + .chain(exclude.iter()) + .filter(|filt| !tests.iter().any(|t| t.matches(filt))) + .collect(); + + assert!( + unmatched.is_empty(), + "filters were provided that have no matching tests: {unmatched:#?}" + ); + + tests.retain(|test| !exclude.iter().any(|exc| test.matches(exc))); + + if cfg.skip_huge { + tests.retain(|test| !test.is_huge_test()); + } + + if !include.is_empty() { + tests.retain(|test| include.iter().any(|inc| test.matches(inc))); + } + + for exc in initial_tests.iter().filter(|orig_name| !tests.iter().any(|t| t.name == **orig_name)) + { + println!("Skipping test '{exc}'"); + } + + println!("Launching all"); + let elapsed = launch_tests(&mut tests, &cfg); + ui::finish_all(&tests, elapsed, &cfg) +} + +/// Enumerate tests to run but don't actually run them. +pub fn register_tests(cfg: &Config) -> Vec<TestInfo> { + let mut tests = Vec::new(); + + // Register normal generators for all floats. + + #[cfg(not(bootstrap))] + #[cfg(target_has_reliable_f16)] + register_float::<f16>(&mut tests, cfg); + register_float::<f32>(&mut tests, cfg); + register_float::<f64>(&mut tests, cfg); + + tests.sort_unstable_by_key(|t| (t.float_name, t.gen_name)); + for i in 0..(tests.len() - 1) { + if tests[i].gen_name == tests[i + 1].gen_name { + panic!("duplicate test name {}", tests[i].gen_name); + } + } + + tests +} + +/// Register all generators for a single float. +fn register_float<F: Float>(tests: &mut Vec<TestInfo>, cfg: &Config) +where + RangeInclusive<F::Int>: Iterator<Item = F::Int>, + <F::Int as TryFrom<u128>>::Error: std::fmt::Debug, + StandardUniform: Distribution<<F as traits::Float>::Int>, +{ + if F::BITS <= MAX_BITS_FOR_EXHAUUSTIVE { + // Only run exhaustive tests if there is a chance of completion. + TestInfo::register::<F, gen_::exhaustive::Exhaustive<F>>(tests); + } + + gen_::fuzz::Fuzz::<F>::set_iterations(cfg.fuzz_count); + + TestInfo::register::<F, gen_::exponents::LargeExponents<F>>(tests); + TestInfo::register::<F, gen_::exponents::SmallExponents<F>>(tests); + TestInfo::register::<F, gen_::fuzz::Fuzz<F>>(tests); + TestInfo::register::<F, gen_::integers::LargeInt<F>>(tests); + TestInfo::register::<F, gen_::integers::SmallInt>(tests); + TestInfo::register::<F, gen_::long_fractions::RepeatingDecimal>(tests); + TestInfo::register::<F, gen_::many_digits::RandDigits<F>>(tests); + TestInfo::register::<F, gen_::sparse::FewOnesFloat<F>>(tests); + TestInfo::register::<F, gen_::sparse::FewOnesInt<F>>(tests); + TestInfo::register::<F, gen_::spot_checks::RegressionCheck>(tests); + TestInfo::register::<F, gen_::spot_checks::Special>(tests); + TestInfo::register::<F, gen_::subnorm::SubnormComplete<F>>(tests); + TestInfo::register::<F, gen_::subnorm::SubnormEdgeCases<F>>(tests); +} + +/// Configuration for a single test. +#[derive(Debug)] +pub struct TestInfo { + pub name: String, + float_name: &'static str, + float_bits: u32, + gen_name: &'static str, + /// Name for display in the progress bar. + short_name: String, + /// Pad the short name to a common width for progress bar use. + short_name_padded: String, + total_tests: u64, + /// Function to launch this test. + launch: fn(&TestInfo, &Config), + /// Progress bar to be updated. + progress: Option<ui::Progress>, + /// Once completed, this will be set. + completed: OnceLock<Completed>, +} + +impl TestInfo { + /// Check if either the name or short name is a match, for filtering. + fn matches(&self, pat: &str) -> bool { + self.short_name.contains(pat) || self.name.contains(pat) + } + + /// Create a `TestInfo` for a given float and generator, then add it to a list. + fn register<F: Float, G: Generator<F>>(v: &mut Vec<Self>) { + let f_name = type_name::<F>(); + let gen_name = G::NAME; + let gen_short_name = G::SHORT_NAME; + let name = format!("{f_name} {gen_name}"); + let short_name = format!("{f_name} {gen_short_name}"); + let short_name_padded = format!("{short_name:18}"); + + let info = TestInfo { + float_name: f_name, + float_bits: F::BITS, + gen_name, + progress: None, + name, + short_name_padded, + short_name, + launch: test_runner::<F, G>, + total_tests: G::total_tests(), + completed: OnceLock::new(), + }; + v.push(info); + } + + /// True if this should be run after all others. + fn is_huge_test(&self) -> bool { + self.total_tests >= HUGE_TEST_CUTOFF + } + + /// When the test is finished, update progress bar messages and finalize. + fn complete(&self, c: Completed) { + self.progress.as_ref().unwrap().complete(&c, 0); + self.completed.set(c).unwrap(); + } +} + +/// Result of an input did not parsing successfully. +#[derive(Clone, Debug)] +enum CheckFailure { + /// Above the zero cutoff but got rounded to zero. + UnexpectedZero, + /// Below the infinity cutoff but got rounded to infinity. + UnexpectedInf, + /// Above the negative infinity cutoff but got rounded to negative infinity. + UnexpectedNegInf, + /// Got a `NaN` when none was expected. + UnexpectedNan, + /// Expected `NaN`, got none. + ExpectedNan, + /// Expected infinity, got finite. + ExpectedInf, + /// Expected negative infinity, got finite. + ExpectedNegInf, + /// The value exceeded its error tolerance. + InvalidReal { + /// Error from the expected value, as a float. + error_float: Option<f64>, + /// Error as a rational string (since it can't always be represented as a float). + error_str: Box<str>, + /// True if the error was caused by not rounding to even at the midpoint between + /// two representable values. + incorrect_midpoint_rounding: bool, + }, + /// String did not parse successfully. + ParsingFailed(Box<str>), + /// A panic was caught. + Panic(Box<str>), +} + +impl fmt::Display for CheckFailure { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + match self { + CheckFailure::UnexpectedZero => { + write!(f, "incorrectly rounded to 0 (expected nonzero)") + } + CheckFailure::UnexpectedInf => { + write!(f, "incorrectly rounded to +inf (expected finite)") + } + CheckFailure::UnexpectedNegInf => { + write!(f, "incorrectly rounded to -inf (expected finite)") + } + CheckFailure::UnexpectedNan => write!(f, "got a NaN where none was expected"), + CheckFailure::ExpectedNan => write!(f, "expected a NaN but did not get it"), + CheckFailure::ExpectedInf => write!(f, "expected +inf but did not get it"), + CheckFailure::ExpectedNegInf => write!(f, "expected -inf but did not get it"), + CheckFailure::InvalidReal { error_float, error_str, incorrect_midpoint_rounding } => { + if *incorrect_midpoint_rounding { + write!( + f, + "midpoint between two representable values did not correctly \ + round to even; error: {error_str}" + )?; + } else { + write!(f, "real number did not parse correctly; error: {error_str}")?; + } + + if let Some(float) = error_float { + write!(f, " ({float})")?; + } + Ok(()) + } + CheckFailure::ParsingFailed(e) => write!(f, "parsing failed: {e}"), + CheckFailure::Panic(e) => write!(f, "function panicked: {e}"), + } + } +} + +/// Information about a completed test generator. +#[derive(Clone, Debug)] +struct Completed { + /// Finished tests (both successful and failed). + executed: u64, + /// Failed tests. + failures: u64, + /// Extra exit information if unsuccessful. + result: Result<FinishedAll, EarlyExit>, + /// If there is something to warn about (e.g bad estimate), leave it here. + warning: Option<Box<str>>, + /// Total time to run the test. + elapsed: Duration, +} + +/// Marker for completing all tests (used in `Result` types). +#[derive(Clone, Debug)] +struct FinishedAll; + +/// Reasons for exiting early. +#[derive(Clone, Debug)] +enum EarlyExit { + Timeout, + MaxFailures, +} + +/// Run all tests in `tests`. +/// +/// This launches a main thread that receives messages and handlees UI updates, and uses the +/// rest of the thread pool to execute the tests. +fn launch_tests(tests: &mut [TestInfo], cfg: &Config) -> Duration { + // Run shorter tests and smaller float types first. + tests.sort_unstable_by_key(|test| (test.total_tests, test.float_bits)); + + for test in tests.iter() { + println!("Launching test '{}'", test.name); + } + + let mut all_progress_bars = Vec::new(); + let start = Instant::now(); + + for test in tests.iter_mut() { + test.progress = Some(ui::Progress::new(test, &mut all_progress_bars)); + ui::set_panic_hook(&all_progress_bars); + ((test.launch)(test, cfg)); + } + + start.elapsed() +} + +/// Test runer for a single generator. +/// +/// This calls the generator's iterator multiple times (in parallel) and validates each output. +fn test_runner<F: Float, G: Generator<F>>(test: &TestInfo, cfg: &Config) { + let gen_ = G::new(); + let executed = AtomicU64::new(0); + let failures = AtomicU64::new(0); + + let checks_per_update = min(test.total_tests, 1000); + let started = Instant::now(); + + // Function to execute for a single test iteration. + let check_one = |buf: &mut String, ctx: G::WriteCtx| { + let executed = executed.fetch_add(1, Ordering::Relaxed); + buf.clear(); + G::write_string(buf, ctx); + + match validate::validate::<F>(buf) { + Ok(()) => (), + Err(e) => { + let CheckError { fail, input, float_res } = e; + test.progress.as_ref().unwrap().println(&format!( + "Failure in '{}': {fail}. parsing '{input}'. Parsed as: {float_res}", + test.name + )); + + let f = failures.fetch_add(1, Ordering::Relaxed); + // End early if the limit is exceeded. + if f >= cfg.max_failures { + return Err(EarlyExit::MaxFailures); + } + } + }; + + // Send periodic updates + if executed % checks_per_update == 0 { + let failures = failures.load(Ordering::Relaxed); + test.progress.as_ref().unwrap().update(executed, failures); + if started.elapsed() > cfg.timeout { + return Err(EarlyExit::Timeout); + } + } + + Ok(()) + }; + + // Run the test iterations in parallel. Each thread gets a string buffer to write + // its check values to. + let res = gen_.par_bridge().try_for_each_init(String::new, check_one); + + let elapsed = started.elapsed(); + let executed = executed.into_inner(); + let failures = failures.into_inner(); + + // Warn about bad estimates if relevant. + let warning = if executed != test.total_tests && res.is_ok() { + let msg = format!( + "executed tests != estimated ({executed} != {}) for {}", + test.total_tests, + G::NAME + ); + + Some(msg.into()) + } else { + None + }; + + let result = res.map(|()| FinishedAll); + test.complete(Completed { executed, failures, result, warning, elapsed }); +} diff --git a/src/tools/test-float-parse/src/main.rs b/src/tools/test-float-parse/src/main.rs new file mode 100644 index 00000000000..9c6cad7324f --- /dev/null +++ b/src/tools/test-float-parse/src/main.rs @@ -0,0 +1,129 @@ +use std::process::ExitCode; +use std::time::Duration; + +use test_float_parse as tfp; + +static HELP: &str = r#"Usage: + + ./test-float-parse [--timeout x] [--exclude x] [--max-failures x] [INCLUDE ...] + ./test-float-parse [--fuzz-count x] [INCLUDE ...] + ./test-float-parse [--skip-huge] [INCLUDE ...] + ./test-float-parse --list + +Args: + + INCLUDE Include only tests with names containing these + strings. If this argument is not specified, all tests + are run. + --timeout N Exit after this amount of time (in seconds). + --exclude FILTER Skip tests containing this string. May be specified + more than once. + --list List available tests. + --max-failures N Limit to N failures per test. Defaults to 20. Pass + "--max-failures none" to remove this limit. + --fuzz-count N Run the fuzzer with N iterations. Only has an effect + if fuzz tests are enabled. Pass `--fuzz-count none` + to remove this limit. + --skip-huge Skip tests that run for a long time. + --all Reset previous `--exclude`, `--skip-huge`, and + `INCLUDE` arguments (useful for running all tests + via `./x`). +"#; + +enum ArgMode { + Any, + Timeout, + Exclude, + FuzzCount, + MaxFailures, +} + +fn main() -> ExitCode { + if cfg!(debug_assertions) { + println!( + "WARNING: running in debug mode. Release mode is recommended to reduce test duration." + ); + std::thread::sleep(Duration::from_secs(2)); + } + + let args: Vec<_> = std::env::args().skip(1).collect(); + if args.iter().any(|arg| arg == "--help" || arg == "-h") { + println!("{HELP}"); + return ExitCode::SUCCESS; + } + + if args.iter().any(|arg| arg == "--list") { + let tests = tfp::register_tests(&tfp::Config::default()); + println!("Available tests:"); + for t in tests { + println!("{}", t.name); + } + + return ExitCode::SUCCESS; + } + + let (cfg, include, exclude) = parse_args(args); + + tfp::run(cfg, &include, &exclude) +} + +/// Simple command argument parser +fn parse_args(args: Vec<String>) -> (tfp::Config, Vec<String>, Vec<String>) { + let mut cfg = tfp::Config::default(); + + let mut mode = ArgMode::Any; + let mut include = Vec::new(); + let mut exclude = Vec::new(); + + for arg in args { + mode = match mode { + ArgMode::Any if arg == "--timeout" => ArgMode::Timeout, + ArgMode::Any if arg == "--exclude" => ArgMode::Exclude, + ArgMode::Any if arg == "--max-failures" => ArgMode::MaxFailures, + ArgMode::Any if arg == "--fuzz-count" => ArgMode::FuzzCount, + ArgMode::Any if arg == "--skip-huge" => { + cfg.skip_huge = true; + ArgMode::Any + } + ArgMode::Any if arg == "--all" => { + cfg.skip_huge = false; + include.clear(); + exclude.clear(); + ArgMode::Any + } + ArgMode::Any if arg.starts_with('-') => { + panic!("Unknown argument {arg}. Usage:\n{HELP}") + } + ArgMode::Any => { + include.push(arg); + ArgMode::Any + } + ArgMode::Timeout => { + cfg.timeout = Duration::from_secs(arg.parse().unwrap()); + ArgMode::Any + } + ArgMode::MaxFailures => { + if arg == "none" { + cfg.disable_max_failures = true; + } else { + cfg.max_failures = arg.parse().unwrap(); + } + ArgMode::Any + } + ArgMode::FuzzCount => { + if arg == "none" { + cfg.fuzz_count = Some(u64::MAX); + } else { + cfg.fuzz_count = Some(arg.parse().unwrap()); + } + ArgMode::Any + } + ArgMode::Exclude => { + exclude.push(arg); + ArgMode::Any + } + } + } + + (cfg, include, exclude) +} diff --git a/src/tools/test-float-parse/src/traits.rs b/src/tools/test-float-parse/src/traits.rs new file mode 100644 index 00000000000..16484f8fe2c --- /dev/null +++ b/src/tools/test-float-parse/src/traits.rs @@ -0,0 +1,206 @@ +//! Interfaces used throughout this crate. + +use std::str::FromStr; +use std::{fmt, ops}; + +use num::Integer; +use num::bigint::ToBigInt; + +use crate::validate::Constants; + +/// Integer types. +#[allow(dead_code)] // Some functions only used for testing +pub trait Int: + Clone + + Copy + + fmt::Debug + + fmt::Display + + fmt::LowerHex + + ops::Add<Output = Self> + + ops::Sub<Output = Self> + + ops::Shl<u32, Output = Self> + + ops::Shr<u32, Output = Self> + + ops::BitAnd<Output = Self> + + ops::BitOr<Output = Self> + + ops::Not<Output = Self> + + ops::AddAssign + + ops::BitAndAssign + + ops::BitOrAssign + + From<u8> + + TryFrom<i8> + + TryFrom<u32, Error: fmt::Debug> + + TryFrom<u64, Error: fmt::Debug> + + TryFrom<u128, Error: fmt::Debug> + + TryInto<u64, Error: fmt::Debug> + + TryInto<u32, Error: fmt::Debug> + + ToBigInt + + PartialOrd + + Integer + + Send + + 'static +{ + type Signed: Int; + type Bytes: Default + AsMut<[u8]>; + + const BITS: u32; + const ZERO: Self; + const ONE: Self; + const MAX: Self; + + fn to_signed(self) -> Self::Signed; + fn wrapping_neg(self) -> Self; + fn trailing_zeros(self) -> u32; + + fn hex(self) -> String { + format!("{self:x}") + } +} + +macro_rules! impl_int { + ($($uty:ty, $sty:ty);+) => { + $( + impl Int for $uty { + type Signed = $sty; + type Bytes = [u8; Self::BITS as usize / 8]; + const BITS: u32 = Self::BITS; + const ZERO: Self = 0; + const ONE: Self = 1; + const MAX: Self = Self::MAX; + fn to_signed(self) -> Self::Signed { + self.try_into().unwrap() + } + fn wrapping_neg(self) -> Self { + self.wrapping_neg() + } + fn trailing_zeros(self) -> u32 { + self.trailing_zeros() + } + } + + impl Int for $sty { + type Signed = Self; + type Bytes = [u8; Self::BITS as usize / 8]; + const BITS: u32 = Self::BITS; + const ZERO: Self = 0; + const ONE: Self = 1; + const MAX: Self = Self::MAX; + fn to_signed(self) -> Self::Signed { + self + } + fn wrapping_neg(self) -> Self { + self.wrapping_neg() + } + fn trailing_zeros(self) -> u32 { + self.trailing_zeros() + } + } + )+ + } +} + +impl_int!(u16, i16; u32, i32; u64, i64); + +/// Floating point types. +pub trait Float: + Copy + fmt::Debug + fmt::LowerExp + FromStr<Err: fmt::Display> + Sized + Send + 'static +{ + /// Unsigned integer of same width + type Int: Int<Signed = Self::SInt>; + type SInt: Int; + + /// Total bits + const BITS: u32; + + /// (Stored) bits in the mantissa) + const MAN_BITS: u32; + + /// Bits in the exponent + const EXP_BITS: u32 = Self::BITS - Self::MAN_BITS - 1; + + /// A saturated exponent (all ones) + const EXP_SAT: u32 = (1 << Self::EXP_BITS) - 1; + + /// The exponent bias, also its maximum value + const EXP_BIAS: u32 = Self::EXP_SAT >> 1; + + const MAN_MASK: Self::Int; + const SIGN_MASK: Self::Int; + + fn from_bits(i: Self::Int) -> Self; + fn to_bits(self) -> Self::Int; + + /// Rational constants associated with this float type. + fn constants() -> &'static Constants; + + fn is_sign_negative(self) -> bool { + (self.to_bits() & Self::SIGN_MASK) > Self::Int::ZERO + } + + /// Exponent without adjustment for bias. + fn exponent(self) -> u32 { + ((self.to_bits() >> Self::MAN_BITS) & Self::EXP_SAT.try_into().unwrap()).try_into().unwrap() + } + + fn mantissa(self) -> Self::Int { + self.to_bits() & Self::MAN_MASK + } +} + +macro_rules! impl_float { + ($($fty:ty, $ity:ty);+) => { + $( + impl Float for $fty { + type Int = $ity; + type SInt = <Self::Int as Int>::Signed; + const BITS: u32 = <$ity>::BITS; + const MAN_BITS: u32 = Self::MANTISSA_DIGITS - 1; + const MAN_MASK: Self::Int = (Self::Int::ONE << Self::MAN_BITS) - Self::Int::ONE; + const SIGN_MASK: Self::Int = Self::Int::ONE << (Self::BITS-1); + fn from_bits(i: Self::Int) -> Self { Self::from_bits(i) } + fn to_bits(self) -> Self::Int { self.to_bits() } + fn constants() -> &'static Constants { + use std::sync::LazyLock; + static CONSTANTS: LazyLock<Constants> = LazyLock::new(Constants::new::<$fty>); + &CONSTANTS + } + } + )+ + } +} + +impl_float!(f32, u32; f64, u64); + +#[cfg(not(bootstrap))] +#[cfg(target_has_reliable_f16)] +impl_float!(f16, u16); + +/// A test generator. Should provide an iterator that produces unique patterns to parse. +/// +/// The iterator needs to provide a `WriteCtx` (could be anything), which is then used to +/// write the string at a later step. This is done separately so that we can reuse string +/// allocations (which otherwise turn out to be a pretty expensive part of these tests). +pub trait Generator<F: Float>: Iterator<Item = Self::WriteCtx> + Send + 'static { + /// Full display and filtering name + const NAME: &'static str = Self::SHORT_NAME; + + /// Name for display with the progress bar + const SHORT_NAME: &'static str; + + /// The context needed to create a test string. + type WriteCtx: Send; + + /// Number of tests that will be run. + fn total_tests() -> u64; + + /// Constructor for this test generator. + fn new() -> Self; + + /// Create a test string given write context, which was produced as a step from the iterator. + /// + /// `s` will be provided empty. + fn write_string(s: &mut String, ctx: Self::WriteCtx); +} + +/// For tests that use iterator combinators, it is easier to just to box the iterator than trying +/// to specify its type. This is a shorthand for the usual type. +pub type BoxGenIter<This, F> = Box<dyn Iterator<Item = <This as Generator<F>>::WriteCtx> + Send>; diff --git a/src/tools/test-float-parse/src/ui.rs b/src/tools/test-float-parse/src/ui.rs new file mode 100644 index 00000000000..73473eef0bf --- /dev/null +++ b/src/tools/test-float-parse/src/ui.rs @@ -0,0 +1,168 @@ +//! Progress bars and such. + +use std::any::type_name; +use std::fmt; +use std::io::{self, Write}; +use std::process::ExitCode; +use std::time::Duration; + +use indicatif::{ProgressBar, ProgressStyle}; + +use crate::{Completed, Config, EarlyExit, FinishedAll, TestInfo}; + +/// Templates for progress bars. +const PB_TEMPLATE: &str = "[{elapsed:3} {percent:3}%] {bar:20.cyan/blue} NAME \ + {human_pos:>8}/{human_len:8} {msg} f {per_sec:14} eta {eta:8}"; +const PB_TEMPLATE_FINAL: &str = "[{elapsed:3} {percent:3}%] {bar:20.cyan/blue} NAME \ + {human_pos:>8}/{human_len:8} {msg:.COLOR} {per_sec:18} {elapsed_precise}"; + +/// Thin abstraction over our usage of a `ProgressBar`. +#[derive(Debug)] +pub struct Progress { + pb: ProgressBar, + make_final_style: NoDebug<Box<dyn Fn(&'static str) -> ProgressStyle + Sync>>, +} + +impl Progress { + /// Create a new progress bar within a multiprogress bar. + pub fn new(test: &TestInfo, all_bars: &mut Vec<ProgressBar>) -> Self { + let initial_template = PB_TEMPLATE.replace("NAME", &test.short_name_padded); + let final_template = PB_TEMPLATE_FINAL.replace("NAME", &test.short_name_padded); + let initial_style = + ProgressStyle::with_template(&initial_template).unwrap().progress_chars("##-"); + let make_final_style = move |color| { + ProgressStyle::with_template(&final_template.replace("COLOR", color)) + .unwrap() + .progress_chars("##-") + }; + + let pb = ProgressBar::new(test.total_tests); + pb.set_style(initial_style); + pb.set_length(test.total_tests); + pb.set_message("0"); + all_bars.push(pb.clone()); + + Progress { pb, make_final_style: NoDebug(Box::new(make_final_style)) } + } + + /// Completed a out of b tests. + pub fn update(&self, completed: u64, failures: u64) { + // Infrequently update the progress bar. + if completed % 5_000 == 0 || failures > 0 { + self.pb.set_position(completed); + } + + if failures > 0 { + self.pb.set_message(format! {"{failures}"}); + } + } + + /// Finalize the progress bar. + pub fn complete(&self, c: &Completed, real_total: u64) { + let f = c.failures; + let (color, msg, finish_fn): (&str, String, fn(&ProgressBar)) = match &c.result { + Ok(FinishedAll) if f > 0 => { + ("red", format!("{f} f (completed with errors)",), ProgressBar::finish) + } + Ok(FinishedAll) => { + ("green", format!("{f} f (completed successfully)",), ProgressBar::finish) + } + Err(EarlyExit::Timeout) => ("red", format!("{f} f (timed out)"), ProgressBar::abandon), + Err(EarlyExit::MaxFailures) => { + ("red", format!("{f} f (failure limit)"), ProgressBar::abandon) + } + }; + + self.pb.set_position(real_total); + self.pb.set_style(self.make_final_style.0(color)); + self.pb.set_message(msg); + finish_fn(&self.pb); + } + + /// Print a message to stdout above the current progress bar. + pub fn println(&self, msg: &str) { + self.pb.suspend(|| println!("{msg}")); + } +} + +/// Print final messages after all tests are complete. +pub fn finish_all(tests: &[TestInfo], total_elapsed: Duration, cfg: &Config) -> ExitCode { + println!("\n\nResults:"); + + let mut failed_generators = 0; + let mut stopped_generators = 0; + + for t in tests { + let Completed { executed, failures, elapsed, warning, result } = t.completed.get().unwrap(); + + let stat = if result.is_err() { + stopped_generators += 1; + "STOPPED" + } else if *failures > 0 { + failed_generators += 1; + "FAILURE" + } else { + "SUCCESS" + }; + + println!( + " {stat} for generator '{name}'. {passed}/{executed} passed in {elapsed:?}", + name = t.name, + passed = executed - failures, + ); + + if let Some(warning) = warning { + println!(" warning: {warning}"); + } + + match result { + Ok(FinishedAll) => (), + Err(EarlyExit::Timeout) => { + println!(" exited early; exceded {:?} timeout", cfg.timeout) + } + Err(EarlyExit::MaxFailures) => { + println!(" exited early; exceeded {:?} max failures", cfg.max_failures) + } + } + } + + println!( + "{passed}/{} tests succeeded in {total_elapsed:?} ({passed} passed, {} failed, {} stopped)", + tests.len(), + failed_generators, + stopped_generators, + passed = tests.len() - failed_generators - stopped_generators, + ); + + if failed_generators > 0 || stopped_generators > 0 { + ExitCode::FAILURE + } else { + ExitCode::SUCCESS + } +} + +/// indicatif likes to eat panic messages. This workaround isn't ideal, but it improves things. +/// <https://github.com/console-rs/indicatif/issues/121>. +pub fn set_panic_hook(drop_bars: &[ProgressBar]) { + let hook = std::panic::take_hook(); + let drop_bars = drop_bars.to_owned(); + std::panic::set_hook(Box::new(move |info| { + for bar in &drop_bars { + bar.abandon(); + println!(); + io::stdout().flush().unwrap(); + io::stderr().flush().unwrap(); + } + hook(info); + })); +} + +/// Allow non-Debug items in a `derive(Debug)` struct. +#[derive(Clone)] +struct NoDebug<T>(T); + +impl<T> fmt::Debug for NoDebug<T> { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + f.write_str(type_name::<Self>()) + } +} diff --git a/src/tools/test-float-parse/src/validate.rs b/src/tools/test-float-parse/src/validate.rs new file mode 100644 index 00000000000..40dda274e3b --- /dev/null +++ b/src/tools/test-float-parse/src/validate.rs @@ -0,0 +1,394 @@ +//! Everything related to verifying that parsed outputs are correct. + +use std::any::{Any, type_name}; +use std::collections::BTreeMap; +use std::ops::RangeInclusive; +use std::str::FromStr; +use std::sync::LazyLock; + +use num::bigint::ToBigInt; +use num::{BigInt, BigRational, FromPrimitive, Signed, ToPrimitive}; + +use crate::{CheckFailure, Float, Int}; + +/// Powers of two that we store for constants. Account for binary128 which has a 15-bit exponent. +const POWERS_OF_TWO_RANGE: RangeInclusive<i32> = (-(2 << 15))..=(2 << 15); + +/// Powers of ten that we cache. Account for binary128, which can fit +4932/-4931 +const POWERS_OF_TEN_RANGE: RangeInclusive<i32> = -5_000..=5_000; + +/// Cached powers of 10 so we can look them up rather than recreating. +static POWERS_OF_TEN: LazyLock<BTreeMap<i32, BigRational>> = LazyLock::new(|| { + POWERS_OF_TEN_RANGE.map(|exp| (exp, BigRational::from_u32(10).unwrap().pow(exp))).collect() +}); + +/// Rational property-related constants for a specific float type. +#[allow(dead_code)] +#[derive(Debug)] +pub struct Constants { + /// The minimum positive value (a subnormal). + min_subnormal: BigRational, + /// The maximum possible finite value. + max: BigRational, + /// Cutoff between rounding to zero and rounding to the minimum value (min subnormal). + zero_cutoff: BigRational, + /// Cutoff between rounding to the max value and rounding to infinity. + inf_cutoff: BigRational, + /// Opposite of `inf_cutoff` + neg_inf_cutoff: BigRational, + /// The powers of two for all relevant integers. + powers_of_two: BTreeMap<i32, BigRational>, + /// Half of each power of two. ULP = "unit in last position". + /// + /// This is a mapping from integers to half the precision available at that exponent. In other + /// words, `0.5 * 2^n` = `2^(n-1)`, which is half the distance between `m * 2^n` and + /// `(m + 1) * 2^n`, m ∈ ℤ. + /// + /// So, this is the maximum error from a real number to its floating point representation, + /// assuming the float type can represent the exponent. + half_ulp: BTreeMap<i32, BigRational>, + /// Handy to have around so we don't need to reallocate for it + two: BigInt, +} + +impl Constants { + pub fn new<F: Float>() -> Self { + let two_int = &BigInt::from_u32(2).unwrap(); + let two = &BigRational::from_integer(2.into()); + + // The minimum subnormal (aka minimum positive) value. Most negative power of two is the + // minimum exponent (bias - 1) plus the extra from shifting within the mantissa bits. + let min_subnormal = two.pow(-(F::EXP_BIAS + F::MAN_BITS - 1).to_signed()); + + // The maximum value is the maximum exponent with a fully saturated mantissa. This + // is easiest to calculate by evaluating what the next value up would be if representable + // (zeroed mantissa, exponent increments by one, i.e. `2^(bias + 1)`), and subtracting + // a single LSB (`2 ^ (-mantissa_bits)`). + let max = (two - two.pow(-F::MAN_BITS.to_signed())) * (two.pow(F::EXP_BIAS.to_signed())); + let zero_cutoff = &min_subnormal / two_int; + + let inf_cutoff = &max + two_int.pow(F::EXP_BIAS - F::MAN_BITS - 1); + let neg_inf_cutoff = -&inf_cutoff; + + let powers_of_two: BTreeMap<i32, _> = + (POWERS_OF_TWO_RANGE).map(|n| (n, two.pow(n))).collect(); + let mut half_ulp = powers_of_two.clone(); + half_ulp.iter_mut().for_each(|(_k, v)| *v = &*v / two_int); + + Self { + min_subnormal, + max, + zero_cutoff, + inf_cutoff, + neg_inf_cutoff, + powers_of_two, + half_ulp, + two: two_int.clone(), + } + } +} + +/// Validate that a string parses correctly +pub fn validate<F: Float>(input: &str) -> Result<(), CheckError> { + // Catch panics in case debug assertions within `std` fail. + let parsed = std::panic::catch_unwind(|| { + input.parse::<F>().map_err(|e| CheckError { + fail: CheckFailure::ParsingFailed(e.to_string().into()), + input: input.into(), + float_res: "none".into(), + }) + }) + .map_err(|e| convert_panic_error(&e, input))??; + + // Parsed float, decoded into significand and exponent + let decoded = decode(parsed); + + // Float parsed separately into a rational + let rational = Rational::parse(input); + + // Verify that the values match + decoded.check(rational, input) +} + +/// Turn panics into concrete error types. +fn convert_panic_error(e: &dyn Any, input: &str) -> CheckError { + let msg = e + .downcast_ref::<String>() + .map(|s| s.as_str()) + .or_else(|| e.downcast_ref::<&str>().copied()) + .unwrap_or("(no contents)"); + + CheckError { + fail: CheckFailure::Panic(msg.into()), + input: input.into(), + float_res: "none".into(), + } +} + +/// The result of parsing a string to a float type. +#[derive(Clone, Copy, Debug, PartialEq)] +pub enum FloatRes<F: Float> { + Inf, + NegInf, + Zero, + Nan, + /// A real number with significand and exponent. Value is `sig * 2 ^ exp`. + Real { + sig: F::SInt, + exp: i32, + }, +} + +#[derive(Clone, Debug)] +pub struct CheckError { + pub fail: CheckFailure, + /// String for which parsing was attempted. + pub input: Box<str>, + /// The parsed & decomposed `FloatRes`, already stringified so we don't need generics here. + pub float_res: Box<str>, +} + +impl<F: Float> FloatRes<F> { + /// Given a known exact rational, check that this representation is accurate within the + /// limits of the float representation. If not, construct a failure `Update` to send. + fn check(self, expected: Rational, input: &str) -> Result<(), CheckError> { + let consts = F::constants(); + // let bool_helper = |cond: bool, err| cond.then_some(()).ok_or(err); + + let res = match (expected, self) { + // Easy correct cases + (Rational::Inf, FloatRes::Inf) + | (Rational::NegInf, FloatRes::NegInf) + | (Rational::Nan, FloatRes::Nan) => Ok(()), + + // Easy incorrect cases + ( + Rational::Inf, + FloatRes::NegInf | FloatRes::Zero | FloatRes::Nan | FloatRes::Real { .. }, + ) => Err(CheckFailure::ExpectedInf), + ( + Rational::NegInf, + FloatRes::Inf | FloatRes::Zero | FloatRes::Nan | FloatRes::Real { .. }, + ) => Err(CheckFailure::ExpectedNegInf), + ( + Rational::Nan, + FloatRes::Inf | FloatRes::NegInf | FloatRes::Zero | FloatRes::Real { .. }, + ) => Err(CheckFailure::ExpectedNan), + (Rational::Finite(_), FloatRes::Nan) => Err(CheckFailure::UnexpectedNan), + + // Cases near limits + (Rational::Finite(r), FloatRes::Zero) => { + if r <= consts.zero_cutoff { + Ok(()) + } else { + Err(CheckFailure::UnexpectedZero) + } + } + (Rational::Finite(r), FloatRes::Inf) => { + if r >= consts.inf_cutoff { + Ok(()) + } else { + Err(CheckFailure::UnexpectedInf) + } + } + (Rational::Finite(r), FloatRes::NegInf) => { + if r <= consts.neg_inf_cutoff { + Ok(()) + } else { + Err(CheckFailure::UnexpectedNegInf) + } + } + + // Actual numbers + (Rational::Finite(r), FloatRes::Real { sig, exp }) => Self::validate_real(r, sig, exp), + }; + + res.map_err(|fail| CheckError { + fail, + input: input.into(), + float_res: format!("{self:?}").into(), + }) + } + + /// Check that `sig * 2^exp` is the same as `rational`, within the float's error margin. + fn validate_real(rational: BigRational, sig: F::SInt, exp: i32) -> Result<(), CheckFailure> { + let consts = F::constants(); + + // `2^exp`. Use cached powers of two to be faster. + let two_exp = consts + .powers_of_two + .get(&exp) + .unwrap_or_else(|| panic!("missing exponent {exp} for {}", type_name::<F>())); + + // Rational from the parsed value, `sig * 2^exp` + let parsed_rational = two_exp * sig.to_bigint().unwrap(); + let error = (parsed_rational - &rational).abs(); + + // Determine acceptable error at this exponent, which is halfway between this value + // (`sig * 2^exp`) and the next value up (`(sig+1) * 2^exp`). + let half_ulp = consts.half_ulp.get(&exp).unwrap(); + + // If we are within one error value (but not equal) then we rounded correctly. + if &error < half_ulp { + return Ok(()); + } + + // For values where we are exactly between two representable values, meaning that the error + // is exactly one half of the precision at that exponent, we need to round to an even + // binary value (i.e. mantissa ends in 0). + let incorrect_midpoint_rounding = if &error == half_ulp { + if sig & F::SInt::ONE == F::SInt::ZERO { + return Ok(()); + } + + // We rounded to odd rather than even; failing based on midpoint rounding. + true + } else { + // We are out of spec for some other reason. + false + }; + + let one_ulp = consts.half_ulp.get(&(exp + 1)).unwrap(); + assert_eq!(one_ulp, &(half_ulp * &consts.two), "ULP values are incorrect"); + + let relative_error = error / one_ulp; + + Err(CheckFailure::InvalidReal { + error_float: relative_error.to_f64(), + error_str: relative_error.to_string().into(), + incorrect_midpoint_rounding, + }) + } + + /// Remove trailing zeros in the significand and adjust the exponent + #[cfg(test)] + fn normalize(self) -> Self { + use std::cmp::min; + + match self { + Self::Real { sig, exp } => { + // If there are trailing zeroes, remove them and increment the exponent instead + let shift = min(sig.trailing_zeros(), exp.wrapping_neg().try_into().unwrap()); + Self::Real { sig: sig >> shift, exp: exp + i32::try_from(shift).unwrap() } + } + _ => self, + } + } +} + +/// Decompose a float into its integral components. This includes the implicit bit. +/// +/// If `allow_nan` is `false`, panic if `NaN` values are reached. +fn decode<F: Float>(f: F) -> FloatRes<F> { + let ione = F::SInt::ONE; + let izero = F::SInt::ZERO; + + let mut exponent_biased = f.exponent(); + let mut mantissa = f.mantissa().to_signed(); + + if exponent_biased == 0 { + if mantissa == izero { + return FloatRes::Zero; + } + + exponent_biased += 1; + } else if exponent_biased == F::EXP_SAT { + if mantissa != izero { + return FloatRes::Nan; + } + + if f.is_sign_negative() { + return FloatRes::NegInf; + } + + return FloatRes::Inf; + } else { + // Set implicit bit + mantissa |= ione << F::MAN_BITS; + } + + let mut exponent = i32::try_from(exponent_biased).unwrap(); + + // Adjust for bias and the rnage of the mantissa + exponent -= i32::try_from(F::EXP_BIAS + F::MAN_BITS).unwrap(); + + if f.is_sign_negative() { + mantissa = mantissa.wrapping_neg(); + } + + FloatRes::Real { sig: mantissa, exp: exponent } +} + +/// A rational or its unrepresentable values. +#[derive(Clone, Debug, PartialEq)] +enum Rational { + Inf, + NegInf, + Nan, + Finite(BigRational), +} + +impl Rational { + /// Turn a string into a rational. `None` if `NaN`. + fn parse(s: &str) -> Rational { + let mut s = s; // lifetime rules + + if s.strip_prefix('+').unwrap_or(s).eq_ignore_ascii_case("nan") + || s.eq_ignore_ascii_case("-nan") + { + return Rational::Nan; + } + + if s.strip_prefix('+').unwrap_or(s).eq_ignore_ascii_case("inf") { + return Rational::Inf; + } + + if s.eq_ignore_ascii_case("-inf") { + return Rational::NegInf; + } + + // Fast path; no decimals or exponents ot parse + if s.bytes().all(|b| b.is_ascii_digit() || b == b'-') { + return Rational::Finite(BigRational::from_str(s).unwrap()); + } + + let mut ten_exp: i32 = 0; + + // Remove and handle e.g. `e-4`, `e+10`, `e5` suffixes + if let Some(pos) = s.bytes().position(|b| b == b'e' || b == b'E') { + let (dec, exp) = s.split_at(pos); + s = dec; + ten_exp = exp[1..].parse().unwrap(); + } + + // Remove the decimal and instead change our exponent + // E.g. "12.3456" becomes "123456 * 10^-4" + let mut s_owned; + if let Some(pos) = s.bytes().position(|b| b == b'.') { + ten_exp = ten_exp.checked_sub((s.len() - pos - 1).try_into().unwrap()).unwrap(); + s_owned = s.to_owned(); + s_owned.remove(pos); + s = &s_owned; + } + + // let pow = BigRational::from_u32(10).unwrap().pow(ten_exp); + let pow = + POWERS_OF_TEN.get(&ten_exp).unwrap_or_else(|| panic!("missing power of ten {ten_exp}")); + let r = pow + * BigInt::from_str(s) + .unwrap_or_else(|e| panic!("`BigInt::from_str(\"{s}\")` failed with {e}")); + Rational::Finite(r) + } + + #[cfg(test)] + fn expect_finite(self) -> BigRational { + let Self::Finite(r) = self else { + panic!("got non rational: {self:?}"); + }; + + r + } +} + +#[cfg(test)] +mod tests; diff --git a/src/tools/test-float-parse/src/validate/tests.rs b/src/tools/test-float-parse/src/validate/tests.rs new file mode 100644 index 00000000000..ab0e7d8a7ba --- /dev/null +++ b/src/tools/test-float-parse/src/validate/tests.rs @@ -0,0 +1,149 @@ +use num::ToPrimitive; + +use super::*; + +#[test] +fn test_parse_rational() { + assert_eq!(Rational::parse("1234").expect_finite(), BigRational::new(1234.into(), 1.into())); + assert_eq!( + Rational::parse("-1234").expect_finite(), + BigRational::new((-1234).into(), 1.into()) + ); + assert_eq!(Rational::parse("1e+6").expect_finite(), BigRational::new(1000000.into(), 1.into())); + assert_eq!(Rational::parse("1e-6").expect_finite(), BigRational::new(1.into(), 1000000.into())); + assert_eq!( + Rational::parse("10.4e6").expect_finite(), + BigRational::new(10400000.into(), 1.into()) + ); + assert_eq!( + Rational::parse("10.4e+6").expect_finite(), + BigRational::new(10400000.into(), 1.into()) + ); + assert_eq!( + Rational::parse("10.4e-6").expect_finite(), + BigRational::new(13.into(), 1250000.into()) + ); + assert_eq!( + Rational::parse("10.4243566462342456234124").expect_finite(), + BigRational::new(104243566462342456234124_i128.into(), 10000000000000000000000_i128.into()) + ); + assert_eq!(Rational::parse("inf"), Rational::Inf); + assert_eq!(Rational::parse("+inf"), Rational::Inf); + assert_eq!(Rational::parse("-inf"), Rational::NegInf); + assert_eq!(Rational::parse("NaN"), Rational::Nan); +} + +#[test] +fn test_decode() { + assert_eq!(decode(0f32), FloatRes::Zero); + assert_eq!(decode(f32::INFINITY), FloatRes::Inf); + assert_eq!(decode(f32::NEG_INFINITY), FloatRes::NegInf); + assert_eq!(decode(1.0f32).normalize(), FloatRes::Real { sig: 1, exp: 0 }); + assert_eq!(decode(-1.0f32).normalize(), FloatRes::Real { sig: -1, exp: 0 }); + assert_eq!(decode(100.0f32).normalize(), FloatRes::Real { sig: 100, exp: 0 }); + assert_eq!(decode(100.5f32).normalize(), FloatRes::Real { sig: 201, exp: -1 }); + assert_eq!(decode(-4.004f32).normalize(), FloatRes::Real { sig: -8396997, exp: -21 }); + assert_eq!(decode(0.0004f32).normalize(), FloatRes::Real { sig: 13743895, exp: -35 }); + assert_eq!(decode(f32::from_bits(0x1)).normalize(), FloatRes::Real { sig: 1, exp: -149 }); +} + +#[test] +fn test_validate() { + validate::<f32>("0").unwrap(); + validate::<f32>("-0").unwrap(); + validate::<f32>("1").unwrap(); + validate::<f32>("-1").unwrap(); + validate::<f32>("1.1").unwrap(); + validate::<f32>("-1.1").unwrap(); + validate::<f32>("1e10").unwrap(); + validate::<f32>("1e1000").unwrap(); + validate::<f32>("-1e1000").unwrap(); + validate::<f32>("1e-1000").unwrap(); + validate::<f32>("-1e-1000").unwrap(); +} + +#[test] +fn test_validate_real() { + // Most of the arbitrary values come from checking against <http://weitz.de/ieee/>. + let r = &BigRational::from_float(10.0).unwrap(); + FloatRes::<f32>::validate_real(r.clone(), 10, 0).unwrap(); + FloatRes::<f32>::validate_real(r.clone(), 10, -1).unwrap_err(); + FloatRes::<f32>::validate_real(r.clone(), 10, 1).unwrap_err(); + + let r = &BigRational::from_float(0.25).unwrap(); + FloatRes::<f32>::validate_real(r.clone(), 1, -2).unwrap(); + FloatRes::<f32>::validate_real(r.clone(), 2, -2).unwrap_err(); + + let r = &BigRational::from_float(1234.5678).unwrap(); + FloatRes::<f32>::validate_real(r.clone(), 0b100110100101001000101011, -13).unwrap(); + FloatRes::<f32>::validate_real(r.clone(), 0b100110100101001000101010, -13).unwrap_err(); + FloatRes::<f32>::validate_real(r.clone(), 0b100110100101001000101100, -13).unwrap_err(); + + let r = &BigRational::from_float(-1234.5678).unwrap(); + FloatRes::<f32>::validate_real(r.clone(), -0b100110100101001000101011, -13).unwrap(); + FloatRes::<f32>::validate_real(r.clone(), -0b100110100101001000101010, -13).unwrap_err(); + FloatRes::<f32>::validate_real(r.clone(), -0b100110100101001000101100, -13).unwrap_err(); +} + +#[test] +#[allow(unused)] +fn test_validate_real_rounding() { + // Check that we catch when values don't round to even. + + // For f32, the cutoff between 1.0 and the next value up (1.0000001) is + // 1.000000059604644775390625. Anything below it should round down, anything above it should + // round up, and the value itself should round _down_ because `1.0` has an even significand but + // 1.0000001 is odd. + let v1_low_down = Rational::parse("1.00000005960464477539062499999").expect_finite(); + let v1_mid_down = Rational::parse("1.000000059604644775390625").expect_finite(); + let v1_high_up = Rational::parse("1.00000005960464477539062500001").expect_finite(); + + let exp = -(f32::MAN_BITS as i32); + let v1_down_sig = 1 << f32::MAN_BITS; + let v1_up_sig = (1 << f32::MAN_BITS) | 0b1; + + FloatRes::<f32>::validate_real(v1_low_down.clone(), v1_down_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(v1_mid_down.clone(), v1_down_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(v1_high_up.clone(), v1_up_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(-v1_low_down.clone(), -v1_down_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(-v1_mid_down.clone(), -v1_down_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(-v1_high_up.clone(), -v1_up_sig, exp).unwrap(); + + // 1.000000178813934326171875 is between 1.0000001 and the next value up, 1.0000002. The middle + // value here should round _up_ since 1.0000002 has an even mantissa. + let v2_low_down = Rational::parse("1.00000017881393432617187499999").expect_finite(); + let v2_mid_up = Rational::parse("1.000000178813934326171875").expect_finite(); + let v2_high_up = Rational::parse("1.00000017881393432617187500001").expect_finite(); + + let v2_down_sig = v1_up_sig; + let v2_up_sig = (1 << f32::MAN_BITS) | 0b10; + + FloatRes::<f32>::validate_real(v2_low_down.clone(), v2_down_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(v2_mid_up.clone(), v2_up_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(v2_high_up.clone(), v2_up_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(-v2_low_down.clone(), -v2_down_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(-v2_mid_up.clone(), -v2_up_sig, exp).unwrap(); + FloatRes::<f32>::validate_real(-v2_high_up.clone(), -v2_up_sig, exp).unwrap(); + + // Rounding the wrong direction should error + for res in [ + FloatRes::<f32>::validate_real(v1_mid_down.clone(), v1_up_sig, exp), + FloatRes::<f32>::validate_real(v2_mid_up.clone(), v2_down_sig, exp), + FloatRes::<f32>::validate_real(-v1_mid_down.clone(), -v1_up_sig, exp), + FloatRes::<f32>::validate_real(-v2_mid_up.clone(), -v2_down_sig, exp), + ] { + let e = res.unwrap_err(); + let CheckFailure::InvalidReal { incorrect_midpoint_rounding: true, .. } = e else { + panic!("{e:?}"); + }; + } +} + +/// Just a quick check that the constants are what we expect. +#[test] +fn check_constants() { + assert_eq!(f32::constants().max.to_f32().unwrap(), f32::MAX); + assert_eq!(f32::constants().min_subnormal.to_f32().unwrap(), f32::from_bits(0x1)); + assert_eq!(f64::constants().max.to_f64().unwrap(), f64::MAX); + assert_eq!(f64::constants().min_subnormal.to_f64().unwrap(), f64::from_bits(0x1)); +} |