Commit ad6f8e3a59

castholm <carl@astholm.se>
2023-10-06 20:44:47
std.math: add nextAfter (#16894)
`nextAfter()` returns the next representable value after `x` in the direction of `y` and is a standard math library function ([C++](https://en.cppreference.com/w/cpp/numeric/math/nextafter), [Java](https://docs.oracle.com/javase/8/docs/api/java/lang/Math.html#nextAfter-double-double-)). It is primarily useful for bitwise incrementing/decrementing floats. This implementation supports runtime integers, runtime floats and `comptime_int`. `comptime_float` is not supported because NaNs/infinities are intentionally difficult to obtain and because I'm not sure if the fact that it's backed by `f128` is supposed to be an implementation detail. Either way, the user could just call the function with the floating-point type whose behavior they want at comptime and then cast the result to `comptime_float`. The float implementation was ported from mingw-w64 with some slight changes made possible because the Zig standard library doesn't care about raising FP exceptions. The number of test cases may seem excessive but they should cover every normal and edge case for every float type and are especially important for verifying that `f80` works.
master
1 parent e6590fe
Changed files (2)
lib
std
math
nextafter.zig
math.zig
lib/std/math/nextafter.zig
@@ -0,0 +1,327 @@
+const builtin = @import("builtin");
+const std = @import("../std.zig");
+const math = std.math;
+const assert = std.debug.assert;
+const expect = std.testing.expect;
+
+/// Returns the next representable value after `x` in the direction of `y`.
+///
+/// Special cases:
+///
+/// - If `x == y`, `y` is returned.
+/// - For floats, if either `x` or `y` is a NaN, a NaN is returned.
+/// - For floats, if `x == 0.0` and `@abs(y) > 0.0`, the smallest subnormal number with the sign of
+///   `y` is returned.
+///
+pub fn nextAfter(comptime T: type, x: T, y: T) T {
+    return switch (@typeInfo(T)) {
+        .Int, .ComptimeInt => nextAfterInt(T, x, y),
+        .Float => nextAfterFloat(T, x, y),
+        else => @compileError("expected int or non-comptime float, found '" ++ @typeName(T) ++ "'"),
+    };
+}
+
+fn nextAfterInt(comptime T: type, x: T, y: T) T {
+    comptime assert(@typeInfo(T) == .Int or @typeInfo(T) == .ComptimeInt);
+    return if (@typeInfo(T) == .Int and @bitSizeOf(T) < 2)
+        // Special case for `i0`, `u0`, `i1`, and `u1`.
+        y
+    else if (y > x)
+        x + 1
+    else if (y < x)
+        x - 1
+    else
+        y;
+}
+
+// Based on nextafterf/nextafterl from mingw-w64 which are both public domain.
+// <https://github.com/mingw-w64/mingw-w64/blob/e89de847dd3e05bb8e46344378ce3e124f4e7d1c/mingw-w64-crt/math/nextafterf.c>
+// <https://github.com/mingw-w64/mingw-w64/blob/e89de847dd3e05bb8e46344378ce3e124f4e7d1c/mingw-w64-crt/math/nextafterl.c>
+
+fn nextAfterFloat(comptime T: type, x: T, y: T) T {
+    comptime assert(@typeInfo(T) == .Float);
+    if (x == y) {
+        // Returning `y` ensures that (0.0, -0.0) returns -0.0 and that (-0.0, 0.0) returns 0.0.
+        return y;
+    }
+    if (math.isNan(x) or math.isNan(y)) {
+        return math.nan(T);
+    }
+    if (x == 0.0) {
+        return if (y > 0.0)
+            math.floatTrueMin(T)
+        else
+            -math.floatTrueMin(T);
+    }
+    if (@bitSizeOf(T) == 80) {
+        // Unlike other floats, `f80` has an explicitly stored integer bit between the fractional
+        // part and the exponent and thus requires special handling. This integer bit *must* be set
+        // when the value is normal, an infinity or a NaN and *should* be cleared otherwise.
+
+        const fractional_bits_mask = (1 << math.floatFractionalBits(f80)) - 1;
+        const integer_bit_mask = 1 << math.floatFractionalBits(f80);
+        const exponent_bits_mask = (1 << math.floatExponentBits(f80)) - 1;
+
+        var x_parts = math.break_f80(x);
+
+        // Bitwise increment/decrement the fractional part while also taking care to update the
+        // exponent if we overflow the fractional part. This might flip the integer bit; this is
+        // intentional.
+        if ((x > 0.0) == (y > x)) {
+            x_parts.fraction +%= 1;
+            if (x_parts.fraction & fractional_bits_mask == 0) {
+                x_parts.exp += 1;
+            }
+        } else {
+            if (x_parts.fraction & fractional_bits_mask == 0) {
+                x_parts.exp -= 1;
+            }
+            x_parts.fraction -%= 1;
+        }
+
+        // If the new value is normal or an infinity (indicated by at least one bit in the exponent
+        // being set), the integer bit might have been cleared from an overflow, so we must ensure
+        // that it remains set.
+        if (x_parts.exp & exponent_bits_mask != 0) {
+            x_parts.fraction |= integer_bit_mask;
+        }
+        // Otherwise, the new value is subnormal and the integer bit will have either flipped from
+        // set to cleared (if the old value was normal) or remained cleared (if the old value was
+        // subnormal), both of which are the outcomes we want.
+
+        return math.make_f80(x_parts);
+    } else {
+        const Bits = std.meta.Int(.unsigned, @bitSizeOf(T));
+        var x_bits: Bits = @bitCast(x);
+        if ((x > 0.0) == (y > x)) {
+            x_bits += 1;
+        } else {
+            x_bits -= 1;
+        }
+        return @bitCast(x_bits);
+    }
+}
+
+test "math.nextAfter.int" {
+    try expect(nextAfter(i0, 0, 0) == 0);
+    try expect(nextAfter(u0, 0, 0) == 0);
+    try expect(nextAfter(i1, 0, 0) == 0);
+    try expect(nextAfter(i1, 0, -1) == -1);
+    try expect(nextAfter(i1, -1, -1) == -1);
+    try expect(nextAfter(i1, -1, 0) == 0);
+    try expect(nextAfter(u1, 0, 0) == 0);
+    try expect(nextAfter(u1, 0, 1) == 1);
+    try expect(nextAfter(u1, 1, 1) == 1);
+    try expect(nextAfter(u1, 1, 0) == 0);
+    inline for (.{ i8, i16, i32, i64, i128, i333 }) |T| {
+        try expect(nextAfter(T, 3, 7) == 4);
+        try expect(nextAfter(T, 3, -7) == 2);
+        try expect(nextAfter(T, -3, -7) == -4);
+        try expect(nextAfter(T, -3, 7) == -2);
+        try expect(nextAfter(T, 5, 5) == 5);
+        try expect(nextAfter(T, -5, -5) == -5);
+        try expect(nextAfter(T, 0, 0) == 0);
+        try expect(nextAfter(T, math.minInt(T), math.minInt(T)) == math.minInt(T));
+        try expect(nextAfter(T, math.maxInt(T), math.maxInt(T)) == math.maxInt(T));
+    }
+    inline for (.{ u8, u16, u32, u64, u128, u333 }) |T| {
+        try expect(nextAfter(T, 3, 7) == 4);
+        try expect(nextAfter(T, 7, 3) == 6);
+        try expect(nextAfter(T, 5, 5) == 5);
+        try expect(nextAfter(T, 0, 0) == 0);
+        try expect(nextAfter(T, math.minInt(T), math.minInt(T)) == math.minInt(T));
+        try expect(nextAfter(T, math.maxInt(T), math.maxInt(T)) == math.maxInt(T));
+    }
+    comptime {
+        try expect(nextAfter(comptime_int, 3, 7) == 4);
+        try expect(nextAfter(comptime_int, 3, -7) == 2);
+        try expect(nextAfter(comptime_int, -3, -7) == -4);
+        try expect(nextAfter(comptime_int, -3, 7) == -2);
+        try expect(nextAfter(comptime_int, 5, 5) == 5);
+        try expect(nextAfter(comptime_int, -5, -5) == -5);
+        try expect(nextAfter(comptime_int, 0, 0) == 0);
+        try expect(nextAfter(comptime_int, math.maxInt(u512), math.maxInt(u512)) == math.maxInt(u512));
+    }
+}
+
+test "math.nextAfter.float" {
+    @setEvalBranchQuota(2000);
+
+    // normal -> normal
+    try expect(nextAfter(f16, 0x1.234p0, 2.0) == 0x1.238p0);
+    try expect(nextAfter(f16, 0x1.234p0, -2.0) == 0x1.230p0);
+    try expect(nextAfter(f16, 0x1.234p0, 0x1.234p0) == 0x1.234p0);
+    try expect(nextAfter(f16, -0x1.234p0, -2.0) == -0x1.238p0);
+    try expect(nextAfter(f16, -0x1.234p0, 2.0) == -0x1.230p0);
+    try expect(nextAfter(f16, -0x1.234p0, -0x1.234p0) == -0x1.234p0);
+    try expect(nextAfter(f32, 0x1.001234p0, 2.0) == 0x1.001236p0);
+    try expect(nextAfter(f32, 0x1.001234p0, -2.0) == 0x1.001232p0);
+    try expect(nextAfter(f32, 0x1.001234p0, 0x1.001234p0) == 0x1.001234p0);
+    try expect(nextAfter(f32, -0x1.001234p0, -2.0) == -0x1.001236p0);
+    try expect(nextAfter(f32, -0x1.001234p0, 2.0) == -0x1.001232p0);
+    try expect(nextAfter(f32, -0x1.001234p0, -0x1.001234p0) == -0x1.001234p0);
+    inline for (.{f64} ++ if (@bitSizeOf(c_longdouble) == 64) .{c_longdouble} else .{}) |T64| {
+        try expect(nextAfter(T64, 0x1.0000000001234p0, 2.0) == 0x1.0000000001235p0);
+        try expect(nextAfter(T64, 0x1.0000000001234p0, -2.0) == 0x1.0000000001233p0);
+        try expect(nextAfter(T64, 0x1.0000000001234p0, 0x1.0000000001234p0) == 0x1.0000000001234p0);
+        try expect(nextAfter(T64, -0x1.0000000001234p0, -2.0) == -0x1.0000000001235p0);
+        try expect(nextAfter(T64, -0x1.0000000001234p0, 2.0) == -0x1.0000000001233p0);
+        try expect(nextAfter(T64, -0x1.0000000001234p0, -0x1.0000000001234p0) == -0x1.0000000001234p0);
+    }
+    inline for (.{f80} ++ if (@bitSizeOf(c_longdouble) == 80) .{c_longdouble} else .{}) |T80| {
+        try expect(nextAfter(T80, 0x1.0000000000001234p0, 2.0) == 0x1.0000000000001236p0);
+        try expect(nextAfter(T80, 0x1.0000000000001234p0, -2.0) == 0x1.0000000000001232p0);
+        try expect(nextAfter(T80, 0x1.0000000000001234p0, 0x1.0000000000001234p0) == 0x1.0000000000001234p0);
+        try expect(nextAfter(T80, -0x1.0000000000001234p0, -2.0) == -0x1.0000000000001236p0);
+        try expect(nextAfter(T80, -0x1.0000000000001234p0, 2.0) == -0x1.0000000000001232p0);
+        try expect(nextAfter(T80, -0x1.0000000000001234p0, -0x1.0000000000001234p0) == -0x1.0000000000001234p0);
+    }
+    inline for (.{f128} ++ if (@bitSizeOf(c_longdouble) == 128) .{c_longdouble} else .{}) |T128| {
+        try expect(nextAfter(T128, 0x1.0000000000000000000000001234p0, 2.0) == 0x1.0000000000000000000000001235p0);
+        try expect(nextAfter(T128, 0x1.0000000000000000000000001234p0, -2.0) == 0x1.0000000000000000000000001233p0);
+        try expect(nextAfter(T128, 0x1.0000000000000000000000001234p0, 0x1.0000000000000000000000001234p0) == 0x1.0000000000000000000000001234p0);
+        try expect(nextAfter(T128, -0x1.0000000000000000000000001234p0, -2.0) == -0x1.0000000000000000000000001235p0);
+        try expect(nextAfter(T128, -0x1.0000000000000000000000001234p0, 2.0) == -0x1.0000000000000000000000001233p0);
+        try expect(nextAfter(T128, -0x1.0000000000000000000000001234p0, -0x1.0000000000000000000000001234p0) == -0x1.0000000000000000000000001234p0);
+    }
+
+    // subnormal -> subnormal
+    try expect(nextAfter(f16, 0x0.234p-14, 1.0) == 0x0.238p-14);
+    try expect(nextAfter(f16, 0x0.234p-14, -1.0) == 0x0.230p-14);
+    try expect(nextAfter(f16, 0x0.234p-14, 0x0.234p-14) == 0x0.234p-14);
+    try expect(nextAfter(f16, -0x0.234p-14, -1.0) == -0x0.238p-14);
+    try expect(nextAfter(f16, -0x0.234p-14, 1.0) == -0x0.230p-14);
+    try expect(nextAfter(f16, -0x0.234p-14, -0x0.234p-14) == -0x0.234p-14);
+    try expect(nextAfter(f32, 0x0.001234p-126, 1.0) == 0x0.001236p-126);
+    try expect(nextAfter(f32, 0x0.001234p-126, -1.0) == 0x0.001232p-126);
+    try expect(nextAfter(f32, 0x0.001234p-126, 0x0.001234p-126) == 0x0.001234p-126);
+    try expect(nextAfter(f32, -0x0.001234p-126, -1.0) == -0x0.001236p-126);
+    try expect(nextAfter(f32, -0x0.001234p-126, 1.0) == -0x0.001232p-126);
+    try expect(nextAfter(f32, -0x0.001234p-126, -0x0.001234p-126) == -0x0.001234p-126);
+    inline for (.{f64} ++ if (@bitSizeOf(c_longdouble) == 64) .{c_longdouble} else .{}) |T64| {
+        try expect(nextAfter(T64, 0x0.0000000001234p-1022, 1.0) == 0x0.0000000001235p-1022);
+        try expect(nextAfter(T64, 0x0.0000000001234p-1022, -1.0) == 0x0.0000000001233p-1022);
+        try expect(nextAfter(T64, 0x0.0000000001234p-1022, 0x0.0000000001234p-1022) == 0x0.0000000001234p-1022);
+        try expect(nextAfter(T64, -0x0.0000000001234p-1022, -1.0) == -0x0.0000000001235p-1022);
+        try expect(nextAfter(T64, -0x0.0000000001234p-1022, 1.0) == -0x0.0000000001233p-1022);
+        try expect(nextAfter(T64, -0x0.0000000001234p-1022, -0x0.0000000001234p-1022) == -0x0.0000000001234p-1022);
+    }
+    inline for (.{f80} ++ if (@bitSizeOf(c_longdouble) == 80) .{c_longdouble} else .{}) |T80| {
+        try expect(nextAfter(T80, 0x0.0000000000001234p-16382, 1.0) == 0x0.0000000000001236p-16382);
+        try expect(nextAfter(T80, 0x0.0000000000001234p-16382, -1.0) == 0x0.0000000000001232p-16382);
+        try expect(nextAfter(T80, 0x0.0000000000001234p-16382, 0x0.0000000000001234p-16382) == 0x0.0000000000001234p-16382);
+        try expect(nextAfter(T80, -0x0.0000000000001234p-16382, -1.0) == -0x0.0000000000001236p-16382);
+        try expect(nextAfter(T80, -0x0.0000000000001234p-16382, 1.0) == -0x0.0000000000001232p-16382);
+        try expect(nextAfter(T80, -0x0.0000000000001234p-16382, -0x0.0000000000001234p-16382) == -0x0.0000000000001234p-16382);
+    }
+    inline for (.{f128} ++ if (@bitSizeOf(c_longdouble) == 128) .{c_longdouble} else .{}) |T128| {
+        try expect(nextAfter(T128, 0x0.0000000000000000000000001234p-16382, 1.0) == 0x0.0000000000000000000000001235p-16382);
+        try expect(nextAfter(T128, 0x0.0000000000000000000000001234p-16382, -1.0) == 0x0.0000000000000000000000001233p-16382);
+        try expect(nextAfter(T128, 0x0.0000000000000000000000001234p-16382, 0x0.0000000000000000000000001234p-16382) == 0x0.0000000000000000000000001234p-16382);
+        try expect(nextAfter(T128, -0x0.0000000000000000000000001234p-16382, -1.0) == -0x0.0000000000000000000000001235p-16382);
+        try expect(nextAfter(T128, -0x0.0000000000000000000000001234p-16382, 1.0) == -0x0.0000000000000000000000001233p-16382);
+        try expect(nextAfter(T128, -0x0.0000000000000000000000001234p-16382, -0x0.0000000000000000000000001234p-16382) == -0x0.0000000000000000000000001234p-16382);
+    }
+
+    // normal -> normal (change in exponent)
+    try expect(nextAfter(f16, 0x1.FFCp3, math.inf(f16)) == 0x1p4);
+    try expect(nextAfter(f16, 0x1p4, -math.inf(f16)) == 0x1.FFCp3);
+    try expect(nextAfter(f16, -0x1.FFCp3, -math.inf(f16)) == -0x1p4);
+    try expect(nextAfter(f16, -0x1p4, math.inf(f16)) == -0x1.FFCp3);
+    try expect(nextAfter(f32, 0x1.FFFFFEp3, math.inf(f32)) == 0x1p4);
+    try expect(nextAfter(f32, 0x1p4, -math.inf(f32)) == 0x1.FFFFFEp3);
+    try expect(nextAfter(f32, -0x1.FFFFFEp3, -math.inf(f32)) == -0x1p4);
+    try expect(nextAfter(f32, -0x1p4, math.inf(f32)) == -0x1.FFFFFEp3);
+    inline for (.{f64} ++ if (@bitSizeOf(c_longdouble) == 64) .{c_longdouble} else .{}) |T64| {
+        try expect(nextAfter(T64, 0x1.FFFFFFFFFFFFFp3, math.inf(T64)) == 0x1p4);
+        try expect(nextAfter(T64, 0x1p4, -math.inf(T64)) == 0x1.FFFFFFFFFFFFFp3);
+        try expect(nextAfter(T64, -0x1.FFFFFFFFFFFFFp3, -math.inf(T64)) == -0x1p4);
+        try expect(nextAfter(T64, -0x1p4, math.inf(T64)) == -0x1.FFFFFFFFFFFFFp3);
+    }
+    inline for (.{f80} ++ if (@bitSizeOf(c_longdouble) == 80) .{c_longdouble} else .{}) |T80| {
+        try expect(nextAfter(T80, 0x1.FFFFFFFFFFFFFFFEp3, math.inf(T80)) == 0x1p4);
+        try expect(nextAfter(T80, 0x1p4, -math.inf(T80)) == 0x1.FFFFFFFFFFFFFFFEp3);
+        try expect(nextAfter(T80, -0x1.FFFFFFFFFFFFFFFEp3, -math.inf(T80)) == -0x1p4);
+        try expect(nextAfter(T80, -0x1p4, math.inf(T80)) == -0x1.FFFFFFFFFFFFFFFEp3);
+    }
+    inline for (.{f128} ++ if (@bitSizeOf(c_longdouble) == 128) .{c_longdouble} else .{}) |T128| {
+        try expect(nextAfter(T128, 0x1.FFFFFFFFFFFFFFFFFFFFFFFFFFFFp3, math.inf(T128)) == 0x1p4);
+        try expect(nextAfter(T128, 0x1p4, -math.inf(T128)) == 0x1.FFFFFFFFFFFFFFFFFFFFFFFFFFFFp3);
+        try expect(nextAfter(T128, -0x1.FFFFFFFFFFFFFFFFFFFFFFFFFFFFp3, -math.inf(T128)) == -0x1p4);
+        try expect(nextAfter(T128, -0x1p4, math.inf(T128)) == -0x1.FFFFFFFFFFFFFFFFFFFFFFFFFFFFp3);
+    }
+
+    // normal -> subnormal
+    try expect(nextAfter(f16, 0x1p-14, -math.inf(f16)) == 0x0.FFCp-14);
+    try expect(nextAfter(f16, -0x1p-14, math.inf(f16)) == -0x0.FFCp-14);
+    try expect(nextAfter(f32, 0x1p-126, -math.inf(f32)) == 0x0.FFFFFEp-126);
+    try expect(nextAfter(f32, -0x1p-126, math.inf(f32)) == -0x0.FFFFFEp-126);
+    inline for (.{f64} ++ if (@bitSizeOf(c_longdouble) == 64) .{c_longdouble} else .{}) |T64| {
+        try expect(nextAfter(T64, 0x1p-1022, -math.inf(T64)) == 0x0.FFFFFFFFFFFFFp-1022);
+        try expect(nextAfter(T64, -0x1p-1022, math.inf(T64)) == -0x0.FFFFFFFFFFFFFp-1022);
+    }
+    inline for (.{f80} ++ if (@bitSizeOf(c_longdouble) == 80) .{c_longdouble} else .{}) |T80| {
+        try expect(nextAfter(T80, 0x1p-16382, -math.inf(T80)) == 0x0.FFFFFFFFFFFFFFFEp-16382);
+        try expect(nextAfter(T80, -0x1p-16382, math.inf(T80)) == -0x0.FFFFFFFFFFFFFFFEp-16382);
+    }
+    inline for (.{f128} ++ if (@bitSizeOf(c_longdouble) == 128) .{c_longdouble} else .{}) |T128| {
+        try expect(nextAfter(T128, 0x1p-16382, -math.inf(T128)) == 0x0.FFFFFFFFFFFFFFFFFFFFFFFFFFFFp-16382);
+        try expect(nextAfter(T128, -0x1p-16382, math.inf(T128)) == -0x0.FFFFFFFFFFFFFFFFFFFFFFFFFFFFp-16382);
+    }
+
+    // subnormal -> normal
+    try expect(nextAfter(f16, 0x0.FFCp-14, math.inf(f16)) == 0x1p-14);
+    try expect(nextAfter(f16, -0x0.FFCp-14, -math.inf(f16)) == -0x1p-14);
+    try expect(nextAfter(f32, 0x0.FFFFFEp-126, math.inf(f32)) == 0x1p-126);
+    try expect(nextAfter(f32, -0x0.FFFFFEp-126, -math.inf(f32)) == -0x1p-126);
+    inline for (.{f64} ++ if (@bitSizeOf(c_longdouble) == 64) .{c_longdouble} else .{}) |T64| {
+        try expect(nextAfter(T64, 0x0.FFFFFFFFFFFFFp-1022, math.inf(T64)) == 0x1p-1022);
+        try expect(nextAfter(T64, -0x0.FFFFFFFFFFFFFp-1022, -math.inf(T64)) == -0x1p-1022);
+    }
+    inline for (.{f80} ++ if (@bitSizeOf(c_longdouble) == 80) .{c_longdouble} else .{}) |T80| {
+        try expect(nextAfter(T80, 0x0.FFFFFFFFFFFFFFFEp-16382, math.inf(T80)) == 0x1p-16382);
+        try expect(nextAfter(T80, -0x0.FFFFFFFFFFFFFFFEp-16382, -math.inf(T80)) == -0x1p-16382);
+    }
+    inline for (.{f128} ++ if (@bitSizeOf(c_longdouble) == 128) .{c_longdouble} else .{}) |T128| {
+        try expect(nextAfter(T128, 0x0.FFFFFFFFFFFFFFFFFFFFFFFFFFFFp-16382, math.inf(T128)) == 0x1p-16382);
+        try expect(nextAfter(T128, -0x0.FFFFFFFFFFFFFFFFFFFFFFFFFFFFp-16382, -math.inf(T128)) == -0x1p-16382);
+    }
+
+    // special values
+    inline for (.{ f16, f32, f64, f80, f128, c_longdouble }) |T| {
+        try expect(bitwiseEqual(T, nextAfter(T, 0.0, 0.0), 0.0));
+        try expect(bitwiseEqual(T, nextAfter(T, 0.0, -0.0), -0.0));
+        try expect(bitwiseEqual(T, nextAfter(T, -0.0, -0.0), -0.0));
+        try expect(bitwiseEqual(T, nextAfter(T, -0.0, 0.0), 0.0));
+        try expect(nextAfter(T, 0.0, math.inf(T)) == math.floatTrueMin(T));
+        try expect(nextAfter(T, 0.0, -math.inf(T)) == -math.floatTrueMin(T));
+        try expect(nextAfter(T, -0.0, -math.inf(T)) == -math.floatTrueMin(T));
+        try expect(nextAfter(T, -0.0, math.inf(T)) == math.floatTrueMin(T));
+        try expect(bitwiseEqual(T, nextAfter(T, math.floatTrueMin(T), 0.0), 0.0));
+        try expect(bitwiseEqual(T, nextAfter(T, math.floatTrueMin(T), -0.0), 0.0));
+        try expect(bitwiseEqual(T, nextAfter(T, math.floatTrueMin(T), -math.inf(T)), 0.0));
+        try expect(bitwiseEqual(T, nextAfter(T, -math.floatTrueMin(T), -0.0), -0.0));
+        try expect(bitwiseEqual(T, nextAfter(T, -math.floatTrueMin(T), 0.0), -0.0));
+        try expect(bitwiseEqual(T, nextAfter(T, -math.floatTrueMin(T), math.inf(T)), -0.0));
+        try expect(nextAfter(T, math.inf(T), math.inf(T)) == math.inf(T));
+        try expect(nextAfter(T, math.inf(T), -math.inf(T)) == math.floatMax(T));
+        try expect(nextAfter(T, math.floatMax(T), math.inf(T)) == math.inf(T));
+        try expect(nextAfter(T, -math.inf(T), -math.inf(T)) == -math.inf(T));
+        try expect(nextAfter(T, -math.inf(T), math.inf(T)) == -math.floatMax(T));
+        try expect(nextAfter(T, -math.floatMax(T), -math.inf(T)) == -math.inf(T));
+        try expect(math.isNan(nextAfter(T, 1.0, math.nan(T))));
+        try expect(math.isNan(nextAfter(T, math.nan(T), 1.0)));
+        try expect(math.isNan(nextAfter(T, math.nan(T), math.nan(T))));
+        try expect(math.isNan(nextAfter(T, math.inf(T), math.nan(T))));
+        try expect(math.isNan(nextAfter(T, -math.inf(T), math.nan(T))));
+        try expect(math.isNan(nextAfter(T, math.nan(T), math.inf(T))));
+        try expect(math.isNan(nextAfter(T, math.nan(T), -math.inf(T))));
+    }
+}
+
+/// Helps ensure that 0.0 doesn't compare equal to -0.0.
+fn bitwiseEqual(comptime T: type, x: T, y: T) bool {
+    comptime assert(@typeInfo(T) == .Float);
+    const Bits = std.meta.Int(.unsigned, @bitSizeOf(T));
+    return @as(Bits, @bitCast(x)) == @as(Bits, @bitCast(y));
+}
lib/std/math.zig
@@ -223,6 +223,7 @@ pub const isInf = @import("math/isinf.zig").isInf;
 pub const isPositiveInf = @import("math/isinf.zig").isPositiveInf;
 pub const isNegativeInf = @import("math/isinf.zig").isNegativeInf;
 pub const isNormal = @import("math/isnormal.zig").isNormal;
+pub const nextAfter = @import("math/nextafter.zig").nextAfter;
 pub const signbit = @import("math/signbit.zig").signbit;
 pub const scalbn = @import("math/scalbn.zig").scalbn;
 pub const ldexp = @import("math/ldexp.zig").ldexp;
@@ -345,6 +346,7 @@ test {
     _ = isPositiveInf;
     _ = isNegativeInf;
     _ = isNormal;
+    _ = nextAfter;
     _ = signbit;
     _ = scalbn;
     _ = ldexp;