Rewrite the random test generator

Currently, all inputs are generated and then cached. This works
reasonably well but it isn't very configurable or extensible (adding
`f16` and `f128` is awkward).

Replace this with a trait for generating random sequences of tuples.
This also removes possible storage limitations of caching all inputs.

Author: tgross35

compiler-builtins/libm/crates/libm-test/benches/random.rs

@@ -2,8 +2,9 @@ use std::hint::black_box;
 use std::time::Duration;
 
 use criterion::{Criterion, criterion_main};
-use libm_test::gen::{CachedInput, random};
-use libm_test::{CheckBasis, CheckCtx, GenerateInput, MathOp, TupleCall};
+use libm_test::gen::random;
+use libm_test::gen::random::RandomInput;
+use libm_test::{CheckBasis, CheckCtx, MathOp, TupleCall};
 
 /// Benchmark with this many items to get a variety
 const BENCH_ITER_ITEMS: usize = if cfg!(feature = "short-benchmarks") { 50 } else { 500 };
@@ -47,7 +48,7 @@ macro_rules! musl_rand_benches {
 fn bench_one<Op>(c: &mut Criterion, musl_extra: MuslExtra<Op::CFn>)
 where
     Op: MathOp,
-    CachedInput: GenerateInput<Op::RustArgs>,
+    Op::RustArgs: RandomInput,
 {
     let name = Op::NAME;
 

compiler-builtins/libm/crates/libm-test/src/gen.rs

@@ -1,6 +1,5 @@
 //! Different generators that can create random or systematic bit patterns.
 
-use crate::GenerateInput;
 pub mod domain_logspace;
 pub mod edge_cases;
 pub mod random;
@@ -41,71 +40,3 @@ impl<I: Iterator> Iterator for KnownSize<I> {
 }
 
 impl<I: Iterator> ExactSizeIterator for KnownSize<I> {}
-
-/// Helper type to turn any reusable input into a generator.
-#[derive(Clone, Debug, Default)]
-pub struct CachedInput {
-    pub inputs_f32: Vec<(f32, f32, f32)>,
-    pub inputs_f64: Vec<(f64, f64, f64)>,
-    pub inputs_i32: Vec<(i32, i32, i32)>,
-}
-
-impl GenerateInput<(f32,)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (f32,)> {
-        self.inputs_f32.iter().map(|f| (f.0,))
-    }
-}
-
-impl GenerateInput<(f32, f32)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (f32, f32)> {
-        self.inputs_f32.iter().map(|f| (f.0, f.1))
-    }
-}
-
-impl GenerateInput<(i32, f32)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (i32, f32)> {
-        self.inputs_i32.iter().zip(self.inputs_f32.iter()).map(|(i, f)| (i.0, f.0))
-    }
-}
-
-impl GenerateInput<(f32, i32)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (f32, i32)> {
-        GenerateInput::<(i32, f32)>::get_cases(self).map(|(i, f)| (f, i))
-    }
-}
-
-impl GenerateInput<(f32, f32, f32)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (f32, f32, f32)> {
-        self.inputs_f32.iter().copied()
-    }
-}
-
-impl GenerateInput<(f64,)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (f64,)> {
-        self.inputs_f64.iter().map(|f| (f.0,))
-    }
-}
-
-impl GenerateInput<(f64, f64)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (f64, f64)> {
-        self.inputs_f64.iter().map(|f| (f.0, f.1))
-    }
-}
-
-impl GenerateInput<(i32, f64)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (i32, f64)> {
-        self.inputs_i32.iter().zip(self.inputs_f64.iter()).map(|(i, f)| (i.0, f.0))
-    }
-}
-
-impl GenerateInput<(f64, i32)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (f64, i32)> {
-        GenerateInput::<(i32, f64)>::get_cases(self).map(|(i, f)| (f, i))
-    }
-}
-
-impl GenerateInput<(f64, f64, f64)> for CachedInput {
-    fn get_cases(&self) -> impl Iterator<Item = (f64, f64, f64)> {
-        self.inputs_f64.iter().copied()
-    }
-}

compiler-builtins/libm/crates/libm-test/src/gen/random.rs

@@ -1,120 +1,118 @@
-//! A simple generator that produces deterministic random input, caching to use the same
-//! inputs for all functions.
-
+use std::env;
+use std::ops::RangeInclusive;
 use std::sync::LazyLock;
 
+use libm::support::Float;
+use rand::distributions::{Alphanumeric, Standard};
+use rand::prelude::Distribution;
 use rand::{Rng, SeedableRng};
 use rand_chacha::ChaCha8Rng;
 
-use super::CachedInput;
-use crate::{BaseName, CheckCtx, GenerateInput};
-
-const SEED: [u8; 32] = *b"3.141592653589793238462643383279";
-
-/// Number of tests to run.
-// FIXME(ntests): clean this up when possible
-const NTESTS: usize = {
-    if cfg!(optimizations_enabled) {
-        if crate::emulated()
-            || !cfg!(target_pointer_width = "64")
-            || cfg!(all(target_arch = "x86_64", target_vendor = "apple"))
-        {
-            // Tests are pretty slow on non-64-bit targets, x86 MacOS, and targets that run
-            // in QEMU.
-            100_000
-        } else {
-            5_000_000
-        }
-    } else {
-        // Without optimizations just run a quick check
-        800
-    }
-};
-
-/// Tested inputs.
-static TEST_CASES: LazyLock<CachedInput> = LazyLock::new(|| make_test_cases(NTESTS));
-
-/// The first argument to `jn` and `jnf` is the number of iterations. Make this a reasonable
-/// value so tests don't run forever.
-static TEST_CASES_JN: LazyLock<CachedInput> = LazyLock::new(|| {
-    // Start with regular test cases
-    let mut cases = (*TEST_CASES).clone();
-
-    // These functions are extremely slow, limit them
-    let ntests_jn = (NTESTS / 1000).max(80);
-    cases.inputs_i32.truncate(ntests_jn);
-    cases.inputs_f32.truncate(ntests_jn);
-    cases.inputs_f64.truncate(ntests_jn);
-
-    // It is easy to overflow the stack with these in debug mode
-    let max_iterations = if cfg!(optimizations_enabled) && cfg!(target_pointer_width = "64") {
-        0xffff
-    } else if cfg!(windows) {
-        0x00ff
-    } else {
-        0x0fff
-    };
+use super::KnownSize;
+use crate::run_cfg::{int_range, iteration_count};
+use crate::{CheckCtx, GeneratorKind};
 
-    let mut rng = ChaCha8Rng::from_seed(SEED);
+pub(crate) const SEED_ENV: &str = "LIBM_SEED";
 
-    for case in cases.inputs_i32.iter_mut() {
-        case.0 = rng.gen_range(3..=max_iterations);
-    }
+pub(crate) static SEED: LazyLock<[u8; 32]> = LazyLock::new(|| {
+    let s = env::var(SEED_ENV).unwrap_or_else(|_| {
+        let mut rng = rand::thread_rng();
+        (0..32).map(|_| rng.sample(Alphanumeric) as char).collect()
+    });
 
-    cases
+    s.as_bytes().try_into().unwrap_or_else(|_| {
+        panic!("Seed must be 32 characters, got `{s}`");
+    })
 });
 
-fn make_test_cases(ntests: usize) -> CachedInput {
-    let mut rng = ChaCha8Rng::from_seed(SEED);
-
-    // make sure we include some basic cases
-    let mut inputs_i32 = vec![(0, 0, 0), (1, 1, 1), (-1, -1, -1)];
-    let mut inputs_f32 = vec![
-        (0.0, 0.0, 0.0),
-        (f32::EPSILON, f32::EPSILON, f32::EPSILON),
-        (f32::INFINITY, f32::INFINITY, f32::INFINITY),
-        (f32::NEG_INFINITY, f32::NEG_INFINITY, f32::NEG_INFINITY),
-        (f32::MAX, f32::MAX, f32::MAX),
-        (f32::MIN, f32::MIN, f32::MIN),
-        (f32::MIN_POSITIVE, f32::MIN_POSITIVE, f32::MIN_POSITIVE),
-        (f32::NAN, f32::NAN, f32::NAN),
-    ];
-    let mut inputs_f64 = vec![
-        (0.0, 0.0, 0.0),
-        (f64::EPSILON, f64::EPSILON, f64::EPSILON),
-        (f64::INFINITY, f64::INFINITY, f64::INFINITY),
-        (f64::NEG_INFINITY, f64::NEG_INFINITY, f64::NEG_INFINITY),
-        (f64::MAX, f64::MAX, f64::MAX),
-        (f64::MIN, f64::MIN, f64::MIN),
-        (f64::MIN_POSITIVE, f64::MIN_POSITIVE, f64::MIN_POSITIVE),
-        (f64::NAN, f64::NAN, f64::NAN),
-    ];
-
-    inputs_i32.extend((0..(ntests - inputs_i32.len())).map(|_| rng.gen::<(i32, i32, i32)>()));
-
-    // Generate integers to get a full range of bitpatterns, then convert back to
-    // floats.
-    inputs_f32.extend((0..(ntests - inputs_f32.len())).map(|_| {
-        let ints = rng.gen::<(u32, u32, u32)>();
-        (f32::from_bits(ints.0), f32::from_bits(ints.1), f32::from_bits(ints.2))
-    }));
-    inputs_f64.extend((0..(ntests - inputs_f64.len())).map(|_| {
-        let ints = rng.gen::<(u64, u64, u64)>();
-        (f64::from_bits(ints.0), f64::from_bits(ints.1), f64::from_bits(ints.2))
-    }));
-
-    CachedInput { inputs_f32, inputs_f64, inputs_i32 }
+/// Generate a sequence of random values of this type.
+pub trait RandomInput {
+    fn get_cases(ctx: &CheckCtx) -> impl ExactSizeIterator<Item = Self>;
 }
 
-/// Create a test case iterator.
-pub fn get_test_cases<RustArgs>(ctx: &CheckCtx) -> impl Iterator<Item = RustArgs>
+/// Generate a sequence of deterministically random floats.
+fn random_floats<F: Float>(count: u64) -> impl Iterator<Item = F>
 where
-    CachedInput: GenerateInput<RustArgs>,
+    Standard: Distribution<F::Int>,
 {
-    let inputs = if ctx.base_name == BaseName::Jn || ctx.base_name == BaseName::Yn {
-        &TEST_CASES_JN
-    } else {
-        &TEST_CASES
+    let mut rng = ChaCha8Rng::from_seed(*SEED);
+
+    // Generate integers to get a full range of bitpatterns (including NaNs), then convert back
+    // to the float type.
+    (0..count).map(move |_| F::from_bits(rng.gen::<F::Int>()))
+}
+
+/// Generate a sequence of deterministically random `i32`s within a specified range.
+fn random_ints(count: u64, range: RangeInclusive<i32>) -> impl Iterator<Item = i32> {
+    let mut rng = ChaCha8Rng::from_seed(*SEED);
+    (0..count).map(move |_| rng.gen_range::<i32, _>(range.clone()))
+}
+
+macro_rules! impl_random_input {
+    ($fty:ty) => {
+        impl RandomInput for ($fty,) {
+            fn get_cases(ctx: &CheckCtx) -> impl ExactSizeIterator<Item = Self> {
+                let count = iteration_count(ctx, GeneratorKind::Random, 0);
+                let iter = random_floats(count).map(|f: $fty| (f,));
+                KnownSize::new(iter, count)
+            }
+        }
+
+        impl RandomInput for ($fty, $fty) {
+            fn get_cases(ctx: &CheckCtx) -> impl ExactSizeIterator<Item = Self> {
+                let count0 = iteration_count(ctx, GeneratorKind::Random, 0);
+                let count1 = iteration_count(ctx, GeneratorKind::Random, 1);
+                let iter = random_floats(count0)
+                    .flat_map(move |f1: $fty| random_floats(count1).map(move |f2: $fty| (f1, f2)));
+                KnownSize::new(iter, count0 * count1)
+            }
+        }
+
+        impl RandomInput for ($fty, $fty, $fty) {
+            fn get_cases(ctx: &CheckCtx) -> impl ExactSizeIterator<Item = Self> {
+                let count0 = iteration_count(ctx, GeneratorKind::Random, 0);
+                let count1 = iteration_count(ctx, GeneratorKind::Random, 1);
+                let count2 = iteration_count(ctx, GeneratorKind::Random, 2);
+                let iter = random_floats(count0).flat_map(move |f1: $fty| {
+                    random_floats(count1).flat_map(move |f2: $fty| {
+                        random_floats(count2).map(move |f3: $fty| (f1, f2, f3))
+                    })
+                });
+                KnownSize::new(iter, count0 * count1 * count2)
+            }
+        }
+
+        impl RandomInput for (i32, $fty) {
+            fn get_cases(ctx: &CheckCtx) -> impl ExactSizeIterator<Item = Self> {
+                let count0 = iteration_count(ctx, GeneratorKind::Random, 0);
+                let count1 = iteration_count(ctx, GeneratorKind::Random, 1);
+                let range0 = int_range(ctx, 0);
+                let iter = random_ints(count0, range0)
+                    .flat_map(move |f1: i32| random_floats(count1).map(move |f2: $fty| (f1, f2)));
+                KnownSize::new(iter, count0 * count1)
+            }
+        }
+
+        impl RandomInput for ($fty, i32) {
+            fn get_cases(ctx: &CheckCtx) -> impl ExactSizeIterator<Item = Self> {
+                let count0 = iteration_count(ctx, GeneratorKind::Random, 0);
+                let count1 = iteration_count(ctx, GeneratorKind::Random, 1);
+                let range1 = int_range(ctx, 1);
+                let iter = random_floats(count0).flat_map(move |f1: $fty| {
+                    random_ints(count1, range1.clone()).map(move |f2: i32| (f1, f2))
+                });
+                KnownSize::new(iter, count0 * count1)
+            }
+        }
     };
-    inputs.get_cases()
+}
+
+impl_random_input!(f32);
+impl_random_input!(f64);
+
+/// Create a test case iterator.
+pub fn get_test_cases<RustArgs: RandomInput>(
+    ctx: &CheckCtx,
+) -> impl Iterator<Item = RustArgs> + use<'_, RustArgs> {
+    RustArgs::get_cases(ctx)
 }

compiler-builtins/libm/crates/libm-test/src/lib.rs

@@ -26,7 +26,7 @@ pub use num::{FloatExt, logspace};
 pub use op::{BaseName, FloatTy, Identifier, MathOp, OpCFn, OpFTy, OpRustFn, OpRustRet, Ty};
 pub use precision::{MaybeOverride, SpecialCase, default_ulp};
 pub use run_cfg::{CheckBasis, CheckCtx, EXTENSIVE_ENV, GeneratorKind};
-pub use test_traits::{CheckOutput, GenerateInput, Hex, TupleCall};
+pub use test_traits::{CheckOutput, Hex, TupleCall};
 
 /// Result type for tests is usually from `anyhow`. Most times there is no success value to
 /// propagate.

compiler-builtins/libm/crates/libm-test/src/run_cfg.rs

@@ -1,8 +1,10 @@
 //! Configuration for how tests get run.
 
-use std::env;
+use std::ops::RangeInclusive;
 use std::sync::LazyLock;
+use std::{env, str};
 
+use crate::gen::random::{SEED, SEED_ENV};
 use crate::{BaseName, FloatTy, Identifier, test_log};
 
 /// The environment variable indicating which extensive tests should be run.
@@ -188,9 +190,16 @@ pub fn iteration_count(ctx: &CheckCtx, gen_kind: GeneratorKind, argnum: usize) -
     };
     let total = ntests.pow(t_env.input_count.try_into().unwrap());
 
+    let seed_msg = match gen_kind {
+        GeneratorKind::Domain => String::new(),
+        GeneratorKind::Random => {
+            format!(" using `{SEED_ENV}={}`", str::from_utf8(SEED.as_slice()).unwrap())
+        }
+    };
+
     test_log(&format!(
         "{gen_kind:?} {basis:?} {fn_ident} arg {arg}/{args}: {ntests} iterations \
-         ({total} total)",
+         ({total} total){seed_msg}",
         basis = ctx.basis,
         fn_ident = ctx.fn_ident,
         arg = argnum + 1,
@@ -200,6 +209,25 @@ pub fn iteration_count(ctx: &CheckCtx, gen_kind: GeneratorKind, argnum: usize) -
     ntests
 }
 
+/// Some tests require that an integer be kept within reasonable limits; generate that here.
+pub fn int_range(ctx: &CheckCtx, argnum: usize) -> RangeInclusive<i32> {
+    let t_env = TestEnv::from_env(ctx);
+
+    if !matches!(ctx.base_name, BaseName::Jn | BaseName::Yn) {
+        return i32::MIN..=i32::MAX;
+    }
+
+    assert_eq!(argnum, 0, "For `jn`/`yn`, only the first argument takes an integer");
+
+    // The integer argument to `jn` is an iteration count. Limit this to ensure tests can be
+    // completed in a reasonable amount of time.
+    if t_env.slow_platform || !cfg!(optimizations_enabled) {
+        (-0xf)..=0xff
+    } else {
+        (-0xff)..=0xffff
+    }
+}
+
 /// For domain tests, limit how many asymptotes or specified check points we test.
 pub fn check_point_count(ctx: &CheckCtx) -> usize {
     let t_env = TestEnv::from_env(ctx);