Completely overhaul fuzz testing

adds testing for almost every numerical intrinsic

src/int/mod.rs

@@ -72,6 +72,9 @@ pub trait Int:
     /// Prevents the need for excessive conversions between signed and unsigned
     fn logical_shr(self, other: u32) -> Self;
 
+    /// Absolute difference between two integers.
+    fn abs_diff(self, other: Self) -> Self::UnsignedInt;
+
     // copied from primitive integers, but put in a trait
     fn is_zero(self) -> bool;
     fn max_value() -> Self;
@@ -251,6 +254,10 @@ macro_rules! int_impl {
                 me
             }
 
+            fn abs_diff(self, other: Self) -> Self {
+                (self.wrapping_sub(other) as $ity).wrapping_abs() as $uty
+            }
+
             int_impl_common!($uty, $bits);
         }
 
@@ -274,6 +281,10 @@ macro_rules! int_impl {
                 me as $ity
             }
 
+            fn abs_diff(self, other: Self) -> $uty {
+                self.wrapping_sub(other).wrapping_abs() as $uty
+            }
+
             int_impl_common!($ity, $bits);
         }
     };

testcrate/Cargo.toml

@@ -11,7 +11,7 @@ doctest = false
 [build-dependencies]
 rand = "0.7"
 
-[dev-dependencies]
+[dependencies]
 # For fuzzing tests we want a deterministic seedable RNG. We also eliminate potential
 # problems with system RNGs on the variety of platforms this crate is tested on.
 # `xoshiro128**` is used for its quality, size, and speed at generating `u32` shift amounts.

testcrate/src/lib.rs

@@ -1 +1,259 @@
+//! This crate is for integration testing and fuzz testing of functions in `compiler-builtins`. This
+//! includes publicly documented intrinsics and some internal alternative implementation functions
+//! such as `usize_leading_zeros_riscv` (which are tested because they are configured for
+//! architectures not tested by the CI).
+//!
+//! The general idea is to use a combination of edge case testing and randomized fuzz testing. The
+//! edge case testing is crucial for checking cases like where both inputs are equal or equal to
+//! special values such as `i128::MIN`, which is unlikely for the random fuzzer by itself to
+//! encounter. The randomized fuzz testing is specially designed to cover wide swaths of search
+//! space in as few iterations as possible. See `fuzz_values` in `testcrate/tests/misc.rs` for an
+//! example.
+//!
+//! Some floating point tests are disabled for specific architectures, because they do not have
+//! correct rounding.
 #![no_std]
+
+use compiler_builtins::float::Float;
+use compiler_builtins::int::Int;
+
+use rand_xoshiro::rand_core::{RngCore, SeedableRng};
+use rand_xoshiro::Xoshiro128StarStar;
+
+/// Sets the number of fuzz iterations run for most tests. In practice, the vast majority of bugs
+/// are caught by the edge case testers. Most of the remaining bugs triggered by more complex
+/// sequences are caught well within 10_000 fuzz iterations. For classes of algorithms like division
+/// that are vulnerable to rare edge cases, we want 1_000_000 iterations to be more confident. In
+/// practical CI, however, we only want to run the more strenuous test once to catch algorithmic
+/// level bugs, and run the 10_000 iteration test on most targets. Target-dependent bugs are likely
+/// to involve miscompilation and misconfiguration that is likely to break algorithms in quickly
+/// caught ways. We choose to configure `N = 1_000_000` iterations for `x86_64` targets (and if
+/// debug assertions are disabled. Tests without `--release` would take too long) which are likely
+/// to have fast hardware, and run `N = 10_000` for all other targets.
+pub const N: u32 = if cfg!(target_arch = "x86_64") && !cfg!(debug_assertions) {
+    1_000_000
+} else {
+    10_000
+};
+
+/// Random fuzzing step. When run several times, it results in excellent fuzzing entropy such as:
+/// 11110101010101011110111110011111
+/// 10110101010100001011101011001010
+/// 1000000000000000
+/// 10000000000000110111110000001010
+/// 1111011111111101010101111110101
+/// 101111111110100000000101000000
+/// 10000000110100000000100010101
+/// 1010101010101000
+fn fuzz_step<I: Int>(rng: &mut Xoshiro128StarStar, x: &mut I) {
+    let ones = !I::ZERO;
+    let bit_indexing_mask: u32 = I::BITS - 1;
+    // It happens that all the RNG we need can come from one call. 7 bits are needed to index a
+    // worst case 128 bit integer, and there are 4 indexes that need to be made plus 4 bits for
+    // selecting operations
+    let rng32 = rng.next_u32();
+
+    // Randomly OR, AND, and XOR randomly sized and shifted continuous strings of
+    // ones with `lhs` and `rhs`.
+    let r0 = bit_indexing_mask & rng32;
+    let r1 = bit_indexing_mask & (rng32 >> 7);
+    let mask = ones.wrapping_shl(r0).rotate_left(r1);
+    match (rng32 >> 14) % 4 {
+        0 => *x |= mask,
+        1 => *x &= mask,
+        // both 2 and 3 to make XORs as common as ORs and ANDs combined
+        _ => *x ^= mask,
+    }
+
+    // Alternating ones and zeros (e.x. 0b1010101010101010). This catches second-order
+    // problems that might occur for algorithms with two modes of operation (potentially
+    // there is some invariant that can be broken and maintained via alternating between modes,
+    // breaking the algorithm when it reaches the end).
+    let mut alt_ones = I::ONE;
+    for _ in 0..(I::BITS / 2) {
+        alt_ones <<= 2;
+        alt_ones |= I::ONE;
+    }
+    let r0 = bit_indexing_mask & (rng32 >> 16);
+    let r1 = bit_indexing_mask & (rng32 >> 23);
+    let mask = alt_ones.wrapping_shl(r0).rotate_left(r1);
+    match rng32 >> 30 {
+        0 => *x |= mask,
+        1 => *x &= mask,
+        _ => *x ^= mask,
+    }
+}
+
+// We need macros like this, because `#![no_std]` prevents us from using iterators
+macro_rules! edge_cases {
+    ($I:ident, $case:ident, $inner:block) => {
+        for i0 in 0..$I::FUZZ_NUM {
+            let mask_lo = (!$I::UnsignedInt::ZERO).wrapping_shr($I::FUZZ_LENGTHS[i0] as u32);
+            for i1 in i0..I::FUZZ_NUM {
+                let mask_hi =
+                    (!$I::UnsignedInt::ZERO).wrapping_shl($I::FUZZ_LENGTHS[i1 - i0] as u32);
+                let $case = I::from_unsigned(mask_lo & mask_hi);
+                $inner
+            }
+        }
+    };
+}
+
+/// Feeds a series of fuzzing inputs to `f`. The fuzzer first uses an algorithm designed to find
+/// edge cases, followed by a more random fuzzer that runs `n` times.
+pub fn fuzz<I: Int, F: FnMut(I)>(n: u32, mut f: F) {
+    // edge case tester. Calls `f` 210 times for u128.
+    // zero gets skipped by the loop
+    f(I::ZERO);
+    edge_cases!(I, case, {
+        f(case);
+    });
+
+    // random fuzzer
+    let mut rng = Xoshiro128StarStar::seed_from_u64(0);
+    let mut x: I = Int::ZERO;
+    for _ in 0..n {
+        fuzz_step(&mut rng, &mut x);
+        f(x)
+    }
+}
+
+/// The same as `fuzz`, except `f` has two inputs.
+pub fn fuzz_2<I: Int, F: Fn(I, I)>(n: u32, f: F) {
+    // Check cases where the first and second inputs are zero. Both call `f` 210 times for `u128`.
+    edge_cases!(I, case, {
+        f(I::ZERO, case);
+    });
+    edge_cases!(I, case, {
+        f(case, I::ZERO);
+    });
+    // Nested edge tester. Calls `f` 44100 times for `u128`.
+    edge_cases!(I, case0, {
+        edge_cases!(I, case1, {
+            f(case0, case1);
+        })
+    });
+
+    // random fuzzer
+    let mut rng = Xoshiro128StarStar::seed_from_u64(0);
+    let mut x: I = I::ZERO;
+    let mut y: I = I::ZERO;
+    for _ in 0..n {
+        fuzz_step(&mut rng, &mut x);
+        fuzz_step(&mut rng, &mut y);
+        f(x, y)
+    }
+}
+
+/// Tester for shift functions
+pub fn fuzz_shift<I: Int, F: Fn(I, u32)>(f: F) {
+    // Shift functions are very simple and do not need anything other than shifting a small
+    // set of random patterns for every fuzz length.
+    let mut rng = Xoshiro128StarStar::seed_from_u64(0);
+    let mut x: I = Int::ZERO;
+    for i in 0..I::FUZZ_NUM {
+        fuzz_step(&mut rng, &mut x);
+        f(x, Int::ZERO);
+        f(x, I::FUZZ_LENGTHS[i] as u32);
+    }
+}
+
+fn fuzz_float_step<F: Float>(rng: &mut Xoshiro128StarStar, f: &mut F) {
+    let rng32 = rng.next_u32();
+    // we need to fuzz the different parts of the float separately, because the masking on larger
+    // significands will tend to set the exponent to all ones or all zeros frequently
+
+    // sign bit fuzzing
+    let sign = (rng32 & 1) != 0;
+
+    // exponent fuzzing. Only 4 bits for the selector needed.
+    let ones = (F::Int::ONE << F::EXPONENT_BITS) - F::Int::ONE;
+    let r0 = (rng32 >> 1) % F::EXPONENT_BITS;
+    let r1 = (rng32 >> 5) % F::EXPONENT_BITS;
+    // custom rotate shift. Note that `F::Int` is unsigned, so we can shift right without smearing
+    // the sign bit.
+    let mask = if r1 == 0 {
+        ones.wrapping_shr(r0)
+    } else {
+        let tmp = ones.wrapping_shr(r0);
+        (tmp.wrapping_shl(r1) | tmp.wrapping_shr(F::EXPONENT_BITS - r1)) & ones
+    };
+    let mut exp = (f.repr() & F::EXPONENT_MASK) >> F::SIGNIFICAND_BITS;
+    match (rng32 >> 9) % 4 {
+        0 => exp |= mask,
+        1 => exp &= mask,
+        _ => exp ^= mask,
+    }
+
+    // significand fuzzing
+    let mut sig = f.repr() & F::SIGNIFICAND_MASK;
+    fuzz_step(rng, &mut sig);
+    sig &= F::SIGNIFICAND_MASK;
+
+    *f = F::from_parts(sign, exp, sig);
+}
+
+macro_rules! float_edge_cases {
+    ($F:ident, $case:ident, $inner:block) => {
+        for exponent in [
+            F::Int::ZERO,
+            F::Int::ONE,
+            F::Int::ONE << (F::EXPONENT_BITS / 2),
+            (F::Int::ONE << (F::EXPONENT_BITS - 1)) - F::Int::ONE,
+            F::Int::ONE << (F::EXPONENT_BITS - 1),
+            (F::Int::ONE << (F::EXPONENT_BITS - 1)) + F::Int::ONE,
+            (F::Int::ONE << F::EXPONENT_BITS) - F::Int::ONE,
+        ]
+        .iter()
+        {
+            for significand in [
+                F::Int::ZERO,
+                F::Int::ONE,
+                F::Int::ONE << (F::SIGNIFICAND_BITS / 2),
+                (F::Int::ONE << (F::SIGNIFICAND_BITS - 1)) - F::Int::ONE,
+                F::Int::ONE << (F::SIGNIFICAND_BITS - 1),
+                (F::Int::ONE << (F::SIGNIFICAND_BITS - 1)) + F::Int::ONE,
+                (F::Int::ONE << F::SIGNIFICAND_BITS) - F::Int::ONE,
+            ]
+            .iter()
+            {
+                for sign in [false, true].iter() {
+                    let $case = F::from_parts(*sign, *exponent, *significand);
+                    $inner
+                }
+            }
+        }
+    };
+}
+
+pub fn fuzz_float<F: Float, E: Fn(F)>(n: u32, f: E) {
+    float_edge_cases!(F, case, {
+        f(case);
+    });
+
+    // random fuzzer
+    let mut rng = Xoshiro128StarStar::seed_from_u64(0);
+    let mut x = F::ZERO;
+    for _ in 0..n {
+        fuzz_float_step(&mut rng, &mut x);
+        f(x);
+    }
+}
+
+pub fn fuzz_float_2<F: Float, E: Fn(F, F)>(n: u32, f: E) {
+    float_edge_cases!(F, case0, {
+        float_edge_cases!(F, case1, {
+            f(case0, case1);
+        });
+    });
+
+    // random fuzzer
+    let mut rng = Xoshiro128StarStar::seed_from_u64(0);
+    let mut x = F::ZERO;
+    let mut y = F::ZERO;
+    for _ in 0..n {
+        fuzz_float_step(&mut rng, &mut x);
+        fuzz_float_step(&mut rng, &mut y);
+        f(x, y)
+    }
+}

testcrate/tests/addsub.rs

@@ -0,0 +1,109 @@
+use testcrate::*;
+
+macro_rules! sum {
+    ($($i:ty, $fn_add:ident, $fn_sub:ident);*;) => {
+        $(
+            fuzz_2(N, |x: $i, y: $i| {
+                let add0 = x.wrapping_add(y);
+                let sub0 = x.wrapping_sub(y);
+                let add1: $i = $fn_add(x, y);
+                let sub1: $i = $fn_sub(x, y);
+                if add0 != add1 {
+                    panic!(
+                        "{}({}, {}): std: {}, builtins: {}",
+                        stringify!($fn_add), x, y, add0, add1
+                    );
+                }
+                if sub0 != sub1 {
+                    panic!(
+                        "{}({}, {}): std: {}, builtins: {}",
+                        stringify!($fn_sub), x, y, sub0, sub1
+                    );
+                }
+            });
+        )*
+    };
+}
+
+macro_rules! overflowing_sum {
+    ($($i:ty, $fn_add:ident, $fn_sub:ident);*;) => {
+        $(
+            fuzz_2(N, |x: $i, y: $i| {
+                let add0 = x.overflowing_add(y);
+                let sub0 = x.overflowing_sub(y);
+                let add1: ($i, bool) = $fn_add(x, y);
+                let sub1: ($i, bool) = $fn_sub(x, y);
+                if add0.0 != add1.0 || add0.1 != add1.1 {
+                    panic!(
+                        "{}({}, {}): std: {:?}, builtins: {:?}",
+                        stringify!($fn_add), x, y, add0, add1
+                    );
+                }
+                if sub0.0 != sub1.0 || sub0.1 != sub1.1 {
+                    panic!(
+                        "{}({}, {}): std: {:?}, builtins: {:?}",
+                        stringify!($fn_sub), x, y, sub0, sub1
+                    );
+                }
+            });
+        )*
+    };
+}
+
+#[test]
+fn addsub() {
+    use compiler_builtins::int::addsub::{
+        __rust_i128_add, __rust_i128_addo, __rust_i128_sub, __rust_i128_subo, __rust_u128_add,
+        __rust_u128_addo, __rust_u128_sub, __rust_u128_subo,
+    };
+
+    // Integer addition and subtraction is very simple, so 100 fuzzing passes should be plenty.
+    sum!(
+        u128, __rust_u128_add, __rust_u128_sub;
+        i128, __rust_i128_add, __rust_i128_sub;
+    );
+    overflowing_sum!(
+        u128, __rust_u128_addo, __rust_u128_subo;
+        i128, __rust_i128_addo, __rust_i128_subo;
+    );
+}
+
+macro_rules! float_sum {
+    ($($f:ty, $fn_add:ident, $fn_sub:ident);*;) => {
+        $(
+            fuzz_float_2(N, |x: $f, y: $f| {
+                let add0 = x + y;
+                let sub0 = x - y;
+                let add1: $f = $fn_add(x, y);
+                let sub1: $f = $fn_sub(x, y);
+                if !Float::eq_repr(add0, add1) {
+                    panic!(
+                        "{}({}, {}): std: {}, builtins: {}",
+                        stringify!($fn_add), x, y, add0, add1
+                    );
+                }
+                if !Float::eq_repr(sub0, sub1) {
+                    panic!(
+                        "{}({}, {}): std: {}, builtins: {}",
+                        stringify!($fn_sub), x, y, sub0, sub1
+                    );
+                }
+            });
+        )*
+    };
+}
+
+#[cfg(not(all(target_arch = "x86", not(target_feature = "sse"))))]
+#[test]
+fn float_addsub() {
+    use compiler_builtins::float::{
+        add::{__adddf3, __addsf3},
+        sub::{__subdf3, __subsf3},
+        Float,
+    };
+
+    float_sum!(
+        f32, __addsf3, __subsf3;
+        f64, __adddf3, __subdf3;
+    );
+}

testcrate/tests/cmp.rs

@@ -0,0 +1,52 @@
+use testcrate::*;
+
+macro_rules! cmp {
+    ($x:ident, $y:ident, $($unordered_val:expr, $fn:ident);*;) => {
+        $(
+            let cmp0 = if $x.is_nan() || $y.is_nan() {
+                $unordered_val
+            } else if $x < $y {
+                -1
+            } else if $x == $y {
+                0
+            } else {
+                1
+            };
+            let cmp1 = $fn($x, $y);
+            if cmp0 != cmp1 {
+                panic!("{}({}, {}): std: {}, builtins: {}", stringify!($fn_builtins), $x, $y, cmp0, cmp1);
+            }
+        )*
+    };
+}
+
+#[test]
+fn float_comparisons() {
+    use compiler_builtins::float::cmp::{
+        __eqdf2, __eqsf2, __gedf2, __gesf2, __gtdf2, __gtsf2, __ledf2, __lesf2, __ltdf2, __ltsf2,
+        __nedf2, __nesf2, __unorddf2, __unordsf2,
+    };
+
+    fuzz_float_2(N, |x: f32, y: f32| {
+        assert_eq!(__unordsf2(x, y) != 0, x.is_nan() || y.is_nan());
+        cmp!(x, y,
+            1, __ltsf2;
+            1, __lesf2;
+            1, __eqsf2;
+            -1, __gesf2;
+            -1, __gtsf2;
+            1, __nesf2;
+        );
+    });
+    fuzz_float_2(N, |x: f64, y: f64| {
+        assert_eq!(__unorddf2(x, y) != 0, x.is_nan() || y.is_nan());
+        cmp!(x, y,
+            1, __ltdf2;
+            1, __ledf2;
+            1, __eqdf2;
+            -1, __gedf2;
+            -1, __gtdf2;
+            1, __nedf2;
+        );
+    });
+}

testcrate/tests/conv.rs

@@ -0,0 +1,125 @@
+use testcrate::*;
+
+macro_rules! i_to_f {
+    ($($from:ty, $into:ty, $fn:ident);*;) => {
+        $(
+            fuzz(N, |x: $from| {
+                let f0 = x as $into;
+                let f1: $into = $fn(x);
+                // This makes sure that the conversion produced the best rounding possible, and does
+                // this independent of `x as $into` rounding correctly.
+                // This assumes that float to integer conversion is correct.
+                let y_minus_ulp = <$into>::from_bits(f1.to_bits().wrapping_sub(1)) as $from;
+                let y = f1 as $from;
+                let y_plus_ulp = <$into>::from_bits(f1.to_bits().wrapping_add(1)) as $from;
+                let error_minus = <$from as Int>::abs_diff(y_minus_ulp, x);
+                let error = <$from as Int>::abs_diff(y, x);
+                let error_plus = <$from as Int>::abs_diff(y_plus_ulp, x);
+                // The first two conditions check that none of the two closest float values are
+                // strictly closer in representation to `x`. The second makes sure that rounding is
+                // towards even significand if two float values are equally close to the integer.
+                if error_minus < error
+                    || error_plus < error
+                    || ((error_minus == error || error_plus == error)
+                        && ((f0.to_bits() & 1) != 0))
+                {
+                    panic!(
+                        "incorrect rounding by {}({}): {}, ({}, {}, {}), errors ({}, {}, {})",
+                        stringify!($fn),
+                        x,
+                        f1.to_bits(),
+                        y_minus_ulp,
+                        y,
+                        y_plus_ulp,
+                        error_minus,
+                        error,
+                        error_plus,
+                    );
+                }
+                // Test against native conversion. We disable testing on all `x86` because of
+                // rounding bugs with `i686`. `powerpc` also has the same rounding bug.
+                if f0 != f1 && !cfg!(any(
+                    target_arch = "x86",
+                    target_arch = "powerpc",
+                    target_arch = "powerpc64"
+                )) {
+                    panic!(
+                        "{}({}): std: {}, builtins: {}",
+                        stringify!($fn),
+                        x,
+                        f0,
+                        f1,
+                    );
+                }
+            });
+        )*
+    };
+}
+
+#[test]
+fn int_to_float() {
+    use compiler_builtins::float::conv::{
+        __floatdidf, __floatdisf, __floatsidf, __floatsisf, __floattidf, __floattisf,
+        __floatundidf, __floatundisf, __floatunsidf, __floatunsisf, __floatuntidf, __floatuntisf,
+    };
+    use compiler_builtins::int::Int;
+
+    i_to_f!(
+        u32, f32, __floatunsisf;
+        u32, f64, __floatunsidf;
+        i32, f32, __floatsisf;
+        i32, f64, __floatsidf;
+        u64, f32, __floatundisf;
+        u64, f64, __floatundidf;
+        i64, f32, __floatdisf;
+        i64, f64, __floatdidf;
+        u128, f32, __floatuntisf;
+        u128, f64, __floatuntidf;
+        i128, f32, __floattisf;
+        i128, f64, __floattidf;
+    );
+}
+
+macro_rules! f_to_i {
+    ($x:ident, $($f:ty, $fn:ident);*;) => {
+        $(
+            // it is undefined behavior in the first place to do conversions with NaNs
+            if !$x.is_nan() {
+                let conv0 = $x as $f;
+                let conv1: $f = $fn($x);
+                if conv0 != conv1 {
+                    panic!("{}({}): std: {}, builtins: {}", stringify!($fn), $x, conv0, conv1);
+                }
+            }
+        )*
+    };
+}
+
+#[test]
+fn float_to_int() {
+    use compiler_builtins::float::conv::{
+        __fixdfdi, __fixdfsi, __fixdfti, __fixsfdi, __fixsfsi, __fixsfti, __fixunsdfdi,
+        __fixunsdfsi, __fixunsdfti, __fixunssfdi, __fixunssfsi, __fixunssfti,
+    };
+
+    fuzz_float(N, |x: f32| {
+        f_to_i!(x,
+            u32, __fixunssfsi;
+            u64, __fixunssfdi;
+            u128, __fixunssfti;
+            i32, __fixsfsi;
+            i64, __fixsfdi;
+            i128, __fixsfti;
+        );
+    });
+    fuzz_float(N, |x: f64| {
+        f_to_i!(x,
+            u32, __fixunsdfsi;
+            u64, __fixunsdfdi;
+            u128, __fixunsdfti;
+            i32, __fixdfsi;
+            i64, __fixdfdi;
+            i128, __fixdfti;
+        );
+    });
+}

testcrate/tests/div_rem.rs

@@ -1,8 +1,9 @@
-use rand_xoshiro::rand_core::{RngCore, SeedableRng};
-use rand_xoshiro::Xoshiro128StarStar;
-
 use compiler_builtins::int::sdiv::{__divmoddi4, __divmodsi4, __divmodti4};
-use compiler_builtins::int::udiv::{__udivmoddi4, __udivmodsi4, __udivmodti4};
+use compiler_builtins::int::udiv::{__udivmoddi4, __udivmodsi4, __udivmodti4, u128_divide_sparc};
+use testcrate::*;
+
+// Division algorithms have by far the nastiest and largest number of edge cases, and experience shows
+// that sometimes 100_000 iterations of the random fuzzer is needed.
 
 /// Creates intensive test functions for division functions of a certain size
 macro_rules! test {
@@ -16,14 +17,17 @@ macro_rules! test {
     ) => {
         #[test]
         fn $test_name() {
-            fn assert_invariants(lhs: $uX, rhs: $uX) {
-                let rem: &mut $uX = &mut 0;
-                let quo: $uX = $unsigned_name(lhs, rhs, Some(rem));
-                let rem = *rem;
+            fuzz_2(N, |lhs, rhs| {
+                if rhs == 0 {
+                    return;
+                }
+
+                let mut rem: $uX = 0;
+                let quo: $uX = $unsigned_name(lhs, rhs, Some(&mut rem));
                 if rhs <= rem || (lhs != rhs.wrapping_mul(quo).wrapping_add(rem)) {
                     panic!(
                         "unsigned division function failed with lhs:{} rhs:{} \
-                        expected:({}, {}) found:({}, {})",
+                        std:({}, {}) builtins:({}, {})",
                         lhs,
                         rhs,
                         lhs.wrapping_div(rhs),
@@ -55,7 +59,7 @@ macro_rules! test {
                 if incorrect_rem || lhs != rhs.wrapping_mul(quo).wrapping_add(rem) {
                     panic!(
                         "signed division function failed with lhs:{} rhs:{} \
-                        expected:({}, {}) found:({}, {})",
+                        std:({}, {}) builtins:({}, {})",
                         lhs,
                         rhs,
                         lhs.wrapping_div(rhs),
@@ -64,70 +68,7 @@ macro_rules! test {
                         rem
                     );
                 }
-            }
-
-            // Specially designed random fuzzer
-            let mut rng = Xoshiro128StarStar::seed_from_u64(0);
-            let mut lhs: $uX = 0;
-            let mut rhs: $uX = 0;
-            // all ones constant
-            let ones: $uX = !0;
-            // Alternating ones and zeros (e.x. 0b1010101010101010). This catches second-order
-            // problems that might occur for algorithms with two modes of operation (potentially
-            // there is some invariant that can be broken for large `duo` and maintained via
-            // alternating between modes, breaking the algorithm when it reaches the end).
-            let mut alt_ones: $uX = 1;
-            for _ in 0..($n / 2) {
-                alt_ones <<= 2;
-                alt_ones |= 1;
-            }
-            // creates a mask for indexing the bits of the type
-            let bit_indexing_mask = $n - 1;
-            for _ in 0..1_000_000 {
-                // Randomly OR, AND, and XOR randomly sized and shifted continuous strings of
-                // ones with `lhs` and `rhs`. This results in excellent fuzzing entropy such as:
-                // lhs:10101010111101000000000100101010 rhs: 1010101010000000000000001000001
-                // lhs:10101010111101000000000101001010 rhs: 1010101010101010101010100010100
-                // lhs:10101010111101000000000101001010 rhs:11101010110101010101010100001110
-                // lhs:10101010000000000000000001001010 rhs:10100010100000000000000000001010
-                // lhs:10101010000000000000000001001010 rhs:            10101010101010101000
-                // lhs:10101010000000000000000001100000 rhs:11111111111101010101010101001111
-                // lhs:10101010000000101010101011000000 rhs:11111111111101010101010100000111
-                // lhs:10101010101010101010101011101010 rhs:             1010100000000000000
-                // lhs:11111111110101101010101011010111 rhs:             1010100000000000000
-                // The msb is set half of the time by the fuzzer, but `assert_invariants` tests
-                // both the signed and unsigned functions.
-                let r0: u32 = bit_indexing_mask & rng.next_u32();
-                let r1: u32 = bit_indexing_mask & rng.next_u32();
-                let mask = ones.wrapping_shr(r0).rotate_left(r1);
-                match rng.next_u32() % 8 {
-                    0 => lhs |= mask,
-                    1 => lhs &= mask,
-                    // both 2 and 3 to make XORs as common as ORs and ANDs combined, otherwise
-                    // the entropy gets destroyed too often
-                    2 | 3 => lhs ^= mask,
-                    4 => rhs |= mask,
-                    5 => rhs &= mask,
-                    _ => rhs ^= mask,
-                }
-                // do the same for alternating ones and zeros
-                let r0: u32 = bit_indexing_mask & rng.next_u32();
-                let r1: u32 = bit_indexing_mask & rng.next_u32();
-                let mask = alt_ones.wrapping_shr(r0).rotate_left(r1);
-                match rng.next_u32() % 8 {
-                    0 => lhs |= mask,
-                    1 => lhs &= mask,
-                    // both 2 and 3 to make XORs as common as ORs and ANDs combined, otherwise
-                    // the entropy gets destroyed too often
-                    2 | 3 => lhs ^= mask,
-                    4 => rhs |= mask,
-                    5 => rhs &= mask,
-                    _ => rhs ^= mask,
-                }
-                if rhs != 0 {
-                    assert_invariants(lhs, rhs);
-                }
-            }
+            });
         }
     };
 }
@@ -135,3 +76,61 @@ macro_rules! test {
 test!(32, u32, i32, div_rem_si4, __udivmodsi4, __divmodsi4);
 test!(64, u64, i64, div_rem_di4, __udivmoddi4, __divmoddi4);
 test!(128, u128, i128, div_rem_ti4, __udivmodti4, __divmodti4);
+
+#[test]
+fn divide_sparc() {
+    fuzz_2(N, |lhs, rhs| {
+        if rhs == 0 {
+            return;
+        }
+
+        let mut rem: u128 = 0;
+        let quo: u128 = u128_divide_sparc(lhs, rhs, &mut rem);
+        if rhs <= rem || (lhs != rhs.wrapping_mul(quo).wrapping_add(rem)) {
+            panic!(
+                "u128_divide_sparc({}, {}): \
+                std:({}, {}), builtins:({}, {})",
+                lhs,
+                rhs,
+                lhs.wrapping_div(rhs),
+                lhs.wrapping_rem(rhs),
+                quo,
+                rem
+            );
+        }
+    });
+}
+
+macro_rules! float {
+    ($($i:ty, $fn:ident);*;) => {
+        $(
+            fuzz_float_2(N, |x: $i, y: $i| {
+                let quo0 = x / y;
+                let quo1: $i = $fn(x, y);
+                // division of subnormals is not currently handled
+                if !(Float::is_subnormal(&quo0) || Float::is_subnormal(&quo1)) {
+                    if !Float::eq_repr(quo0, quo1) {
+                        panic!(
+                            "{}({}, {}): std: {}, builtins: {}",
+                            stringify!($fn), x, y, quo0, quo1
+                        );
+                    }
+                }
+            });
+        )*
+    };
+}
+
+#[cfg(not(all(target_arch = "x86", not(target_feature = "sse"))))]
+#[test]
+fn float_div() {
+    use compiler_builtins::float::{
+        div::{__divdf3, __divsf3},
+        Float,
+    };
+
+    float!(
+        f32, __divsf3;
+        f64, __divdf3;
+    );
+}

testcrate/tests/leading_zeros.rs

@@ -1,54 +0,0 @@
-use rand_xoshiro::rand_core::{RngCore, SeedableRng};
-use rand_xoshiro::Xoshiro128StarStar;
-
-use compiler_builtins::int::__clzsi2;
-use compiler_builtins::int::leading_zeros::{
-    usize_leading_zeros_default, usize_leading_zeros_riscv,
-};
-
-#[test]
-fn __clzsi2_test() {
-    // Binary fuzzer. We cannot just send a random number directly to `__clzsi2()`, because we need
-    // large sequences of zeros to test. This XORs, ANDs, and ORs random length strings of 1s to
-    // `x`. ORs insure sequences of ones, ANDs insures sequences of zeros, and XORs are not often
-    // destructive but add entropy.
-    let mut rng = Xoshiro128StarStar::seed_from_u64(0);
-    let mut x = 0usize;
-    // creates a mask for indexing the bits of the type
-    let bit_indexing_mask = usize::MAX.count_ones() - 1;
-    // 10000 iterations is enough to make sure edge cases like single set bits are tested and to go
-    // through many paths.
-    for _ in 0..10_000 {
-        let r0 = bit_indexing_mask & rng.next_u32();
-        // random length of ones
-        let ones: usize = !0 >> r0;
-        let r1 = bit_indexing_mask & rng.next_u32();
-        // random circular shift
-        let mask = ones.rotate_left(r1);
-        match rng.next_u32() % 4 {
-            0 => x |= mask,
-            1 => x &= mask,
-            // both 2 and 3 to make XORs as common as ORs and ANDs combined
-            _ => x ^= mask,
-        }
-        let lz = x.leading_zeros() as usize;
-        let lz0 = __clzsi2(x);
-        let lz1 = usize_leading_zeros_default(x);
-        let lz2 = usize_leading_zeros_riscv(x);
-        if lz0 != lz {
-            panic!("__clzsi2({}): expected: {}, found: {}", x, lz, lz0);
-        }
-        if lz1 != lz {
-            panic!(
-                "usize_leading_zeros_default({}): expected: {}, found: {}",
-                x, lz, lz1
-            );
-        }
-        if lz2 != lz {
-            panic!(
-                "usize_leading_zeros_riscv({}): expected: {}, found: {}",
-                x, lz, lz2
-            );
-        }
-    }
-}

testcrate/tests/misc.rs

@@ -0,0 +1,134 @@
+use testcrate::*;
+
+/// Make sure that the the edge case tester and randomized tester don't break, and list examples of
+/// fuzz values for documentation purposes.
+#[test]
+fn fuzz_values() {
+    const VALS: [u16; 47] = [
+        0b0, // edge cases
+        0b1111111111111111,
+        0b1111111111111110,
+        0b1111111111111100,
+        0b1111111110000000,
+        0b1111111100000000,
+        0b1110000000000000,
+        0b1100000000000000,
+        0b1000000000000000,
+        0b111111111111111,
+        0b111111111111110,
+        0b111111111111100,
+        0b111111110000000,
+        0b111111100000000,
+        0b110000000000000,
+        0b100000000000000,
+        0b11111111111111,
+        0b11111111111110,
+        0b11111111111100,
+        0b11111110000000,
+        0b11111100000000,
+        0b10000000000000,
+        0b111111111,
+        0b111111110,
+        0b111111100,
+        0b110000000,
+        0b100000000,
+        0b11111111,
+        0b11111110,
+        0b11111100,
+        0b10000000,
+        0b111,
+        0b110,
+        0b100,
+        0b11,
+        0b10,
+        0b1,
+        0b1010110100000, // beginning of random fuzzing
+        0b1100011001011010,
+        0b1001100101001111,
+        0b1101010100011010,
+        0b100010001,
+        0b1000000000000000,
+        0b1100000000000101,
+        0b1100111101010101,
+        0b1100010111111111,
+        0b1111110101111111,
+    ];
+    let mut i = 0;
+    fuzz(10, |x: u16| {
+        assert_eq!(x, VALS[i]);
+        i += 1;
+    });
+}
+
+#[test]
+fn leading_zeros() {
+    use compiler_builtins::int::__clzsi2;
+    use compiler_builtins::int::leading_zeros::{
+        usize_leading_zeros_default, usize_leading_zeros_riscv,
+    };
+    fuzz(N, |x: usize| {
+        let lz = x.leading_zeros() as usize;
+        let lz0 = __clzsi2(x);
+        let lz1 = usize_leading_zeros_default(x);
+        let lz2 = usize_leading_zeros_riscv(x);
+        if lz0 != lz {
+            panic!("__clzsi2({}): std: {}, builtins: {}", x, lz, lz0);
+        }
+        if lz1 != lz {
+            panic!(
+                "usize_leading_zeros_default({}): std: {}, builtins: {}",
+                x, lz, lz1
+            );
+        }
+        if lz2 != lz {
+            panic!(
+                "usize_leading_zeros_riscv({}): std: {}, builtins: {}",
+                x, lz, lz2
+            );
+        }
+    })
+}
+
+#[test]
+fn float_extend() {
+    fuzz_float(N, |x: f32| {
+        let tmp0 = x as f64;
+        let tmp1: f64 = compiler_builtins::float::extend::__extendsfdf2(x);
+        if !compiler_builtins::float::Float::eq_repr(tmp0, tmp1) {
+            panic!("__extendsfdf2({}): std: {}, builtins: {}", x, tmp0, tmp1);
+        }
+    });
+}
+
+// This doesn't quite work because of issues related to
+// https://github.com/rust-lang/rust/issues/73920.
+// TODO how do we resolve this?
+/*
+macro_rules! pow {
+    ($($f:ty, $fn:ident);*;) => {
+        $(
+            fuzz_float_2(N, |x: $f, y: $f| {
+                let n = y as i32;
+                let tmp0: $f = x.powi(n);
+                let tmp1: $f = $fn(x, n);
+                if tmp0 != tmp1 {
+                    panic!(
+                        "{}({}, {}): std: {}, builtins: {}",
+                        stringify!($fn), x, y, tmp0, tmp1
+                    );
+                }
+            });
+        )*
+    };
+}
+
+#[test]
+fn float_pow() {
+    use compiler_builtins::float::pow::{__powidf2, __powisf2};
+
+    pow!(
+        f32, __powisf2;
+        f64, __powidf2;
+    );
+}
+*/

testcrate/tests/mul.rs

@@ -0,0 +1,114 @@
+use testcrate::*;
+
+macro_rules! mul {
+    ($($i:ty, $fn:ident);*;) => {
+        $(
+            fuzz_2(N, |x: $i, y: $i| {
+                let mul0 = x.wrapping_mul(y);
+                let mul1: $i = $fn(x, y);
+                if mul0 != mul1 {
+                    panic!(
+                        "{}({}, {}): std: {}, builtins: {}",
+                        stringify!($fn), x, y, mul0, mul1
+                    );
+                }
+            });
+        )*
+    };
+}
+
+#[test]
+fn mul() {
+    use compiler_builtins::int::mul::{__muldi3, __multi3};
+
+    mul!(
+        u64, __muldi3;
+        i128, __multi3;
+    );
+}
+
+macro_rules! overflowing_mul {
+    ($($i:ty, $fn:ident);*;) => {
+        $(
+            fuzz_2(N, |x: $i, y: $i| {
+                let (mul0, o0) = x.overflowing_mul(y);
+                let mut o1 = 0i32;
+                let mul1: $i = $fn(x, y, &mut o1);
+                let o1 = o1 != 0;
+                if mul0 != mul1 || o0 != o1 {
+                    panic!(
+                        "{}({}, {}): std: ({}, {}), builtins: ({}, {})",
+                        stringify!($fn), x, y, mul0, o0, mul1, o1
+                    );
+                }
+            });
+        )*
+    };
+}
+
+#[test]
+fn overflowing_mul() {
+    use compiler_builtins::int::mul::{
+        __mulodi4, __mulosi4, __muloti4, __rust_i128_mulo, __rust_u128_mulo,
+    };
+
+    overflowing_mul!(
+        i32, __mulosi4;
+        i64, __mulodi4;
+        i128, __muloti4;
+    );
+    fuzz_2(N, |x: u128, y: u128| {
+        let (mul0, o0) = x.overflowing_mul(y);
+        let (mul1, o1) = __rust_u128_mulo(x, y);
+        if mul0 != mul1 || o0 != o1 {
+            panic!(
+                "__rust_u128_mulo({}, {}): std: ({}, {}), builtins: ({}, {})",
+                x, y, mul0, o0, mul1, o1
+            );
+        }
+        let x = x as i128;
+        let y = y as i128;
+        let (mul0, o0) = x.overflowing_mul(y);
+        let (mul1, o1) = __rust_i128_mulo(x, y);
+        if mul0 != mul1 || o0 != o1 {
+            panic!(
+                "__rust_i128_mulo({}, {}): std: ({}, {}), builtins: ({}, {})",
+                x, y, mul0, o0, mul1, o1
+            );
+        }
+    });
+}
+
+macro_rules! float_mul {
+    ($($f:ty, $fn:ident);*;) => {
+        $(
+            fuzz_float_2(N, |x: $f, y: $f| {
+                let mul0 = x * y;
+                let mul1: $f = $fn(x, y);
+                // multiplication of subnormals is not currently handled
+                if !(Float::is_subnormal(&mul0) || Float::is_subnormal(&mul1)) {
+                    if !Float::eq_repr(mul0, mul1) {
+                        panic!(
+                            "{}({}, {}): std: {}, builtins: {}",
+                            stringify!($fn), x, y, mul0, mul1
+                        );
+                    }
+                }
+            });
+        )*
+    };
+}
+
+#[cfg(not(all(target_arch = "x86", not(target_feature = "sse"))))]
+#[test]
+fn float_mul() {
+    use compiler_builtins::float::{
+        mul::{__muldf3, __mulsf3},
+        Float,
+    };
+
+    float_mul!(
+        f32, __mulsf3;
+        f64, __muldf3;
+    );
+}

testcrate/tests/shift.rs

@@ -0,0 +1,60 @@
+use testcrate::*;
+
+macro_rules! shift {
+    ($($i:ty, $fn_std:ident, $fn_builtins:ident);*;) => {
+        $(
+            fuzz_shift(|x: $i, s: u32| {
+                let tmp0: $i = x.$fn_std(s);
+                let tmp1: $i = $fn_builtins(x, s);
+                if tmp0 != tmp1 {
+                    panic!(
+                        "{}({}, {}): std: {}, builtins: {}",
+                        stringify!($fn_builtins), x, s, tmp0, tmp1
+                    );
+                }
+            });
+        )*
+    };
+}
+
+macro_rules! overflowing_shift {
+    ($($i:ty, $fn_std:ident, $fn_builtins:ident);*;) => {
+        $(
+            fuzz_shift(|x: $i, s: u32| {
+                let tmp0: $i = x.$fn_std(s);
+                let (tmp1, o1): ($i, bool) = $fn_builtins(x, s.into());
+                if tmp0 != tmp1 || o1 {
+                    panic!(
+                        "{}({}, {}): std: {}, builtins: {}",
+                        stringify!($fn_builtins), x, s, tmp0, tmp1
+                    );
+                }
+            });
+        )*
+    };
+}
+
+#[test]
+fn shift() {
+    use compiler_builtins::int::shift::{
+        __ashldi3, __ashlsi3, __ashlti3, __ashrdi3, __ashrsi3, __ashrti3, __lshrdi3, __lshrsi3,
+        __lshrti3, __rust_i128_shlo, __rust_i128_shro, __rust_u128_shlo, __rust_u128_shro,
+    };
+    shift!(
+        u32, wrapping_shl, __ashlsi3;
+        u64, wrapping_shl, __ashldi3;
+        u128, wrapping_shl, __ashlti3;
+        i32, wrapping_shr, __ashrsi3;
+        i64, wrapping_shr, __ashrdi3;
+        i128, wrapping_shr, __ashrti3;
+        u32, wrapping_shr, __lshrsi3;
+        u64, wrapping_shr, __lshrdi3;
+        u128, wrapping_shr, __lshrti3;
+    );
+    overflowing_shift!(
+        u128, wrapping_shl, __rust_u128_shlo;
+        i128, wrapping_shl, __rust_i128_shlo;
+        u128, wrapping_shr, __rust_u128_shro;
+        i128, wrapping_shr, __rust_i128_shro;
+    );
+}