| 1 | const builtin = @import("builtin"); |
| 2 | const std = @import("std"); |
| 3 | const math = std.math; |
| 4 | const Log2Int = std.math.Log2Int; |
| 5 | const assert = std.debug.assert; |
| 6 | const expect = std.testing.expect; |
| 7 | |
| 8 | /// Returns x * 2^n. |
| 9 | pub fn ldexp(x: anytype, n: i32) @TypeOf(x) { |
| 10 | const T = @TypeOf(x); |
| 11 | const TBits = @Int(.unsigned, @typeInfo(T).float.bits); |
| 12 | |
| 13 | const exponent_bits = math.floatExponentBits(T); |
| 14 | const mantissa_bits = math.floatMantissaBits(T); |
| 15 | const fractional_bits = math.floatFractionalBits(T); |
| 16 | |
| 17 | const max_biased_exponent = 2 * math.floatExponentMax(T); |
| 18 | const mantissa_mask = @as(TBits, (1 << mantissa_bits) - 1); |
| 19 | |
| 20 | const repr = @as(TBits, @bitCast(x)); |
| 21 | const sign_bit = repr & (1 << (exponent_bits + mantissa_bits)); |
| 22 | |
| 23 | if (math.isNan(x) or !math.isFinite(x)) |
| 24 | return x; |
| 25 | |
| 26 | var exponent: i32 = @as(i32, @intCast((repr << 1) >> (mantissa_bits + 1))); |
| 27 | if (exponent == 0) |
| 28 | exponent += (@as(i32, exponent_bits) + @intFromBool(T == f80)) - @clz(repr << 1); |
| 29 | |
| 30 | if (n >= 0) { |
| 31 | if (n > max_biased_exponent - exponent) { |
| 32 | // Overflow. Return +/- inf |
| 33 | return @as(T, @bitCast(@as(TBits, @bitCast(math.inf(T))) | sign_bit)); |
| 34 | } else if (exponent + n <= 0) { |
| 35 | // Result is subnormal |
| 36 | return @as(T, @bitCast((repr << @as(Log2Int(TBits), @intCast(n))) | sign_bit)); |
| 37 | } else if (exponent <= 0) { |
| 38 | // Result is normal, but needs shifting |
| 39 | var result = @as(TBits, @intCast(n + exponent)) << mantissa_bits; |
| 40 | result |= (repr << @as(Log2Int(TBits), @intCast(1 - exponent))) & mantissa_mask; |
| 41 | return @as(T, @bitCast(result | sign_bit)); |
| 42 | } |
| 43 | |
| 44 | // Result needs no shifting |
| 45 | return @as(T, @bitCast(repr + (@as(TBits, @intCast(n)) << mantissa_bits))); |
| 46 | } else { |
| 47 | if (n <= -exponent) { |
| 48 | if (n < -(mantissa_bits + exponent)) |
| 49 | return @as(T, @bitCast(sign_bit)); // Severe underflow. Return +/- 0 |
| 50 | |
| 51 | // Result underflowed, we need to shift and round |
| 52 | const shift = @as(Log2Int(TBits), @intCast(@min(-n, -(exponent + n) + 1))); |
| 53 | const exact_tie: bool = @ctz(repr) == shift - 1; |
| 54 | var result = repr & mantissa_mask; |
| 55 | |
| 56 | if (T != f80) // Include integer bit |
| 57 | result |= @as(TBits, @intFromBool(exponent > 0)) << fractional_bits; |
| 58 | result = @as(TBits, @intCast((result >> (shift - 1)))); |
| 59 | |
| 60 | // Round result, including round-to-even for exact ties |
| 61 | result = ((result + 1) >> 1) & ~@as(TBits, @intFromBool(exact_tie)); |
| 62 | return @as(T, @bitCast(result | sign_bit)); |
| 63 | } |
| 64 | |
| 65 | // Result is exact, and needs no shifting |
| 66 | return @as(T, @bitCast(repr - (@as(TBits, @intCast(-n)) << mantissa_bits))); |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | test ldexp { |
| 71 | if (builtin.zig_backend == .stage2_c and builtin.cpu.arch.isAarch64() and builtin.os.tag == .netbsd) return error.SkipZigTest; // https://codeberg.org/ziglang/zig/issues/36765 |
| 72 | if (builtin.zig_backend == .stage2_llvm and builtin.cpu.arch.isAarch64() and builtin.os.tag == .netbsd) return error.SkipZigTest; // https://codeberg.org/ziglang/zig/issues/36765 |
| 73 | |
| 74 | // subnormals |
| 75 | try expect(ldexp(@as(f16, 0x1.1FFp14), -14 - 9 - 15) == math.floatTrueMin(f16)); |
| 76 | try expect(ldexp(@as(f32, 0x1.3FFFFFp-1), -126 - 22) == math.floatTrueMin(f32)); |
| 77 | try expect(ldexp(@as(f64, 0x1.7FFFFFFFFFFFFp-1), -1022 - 51) == math.floatTrueMin(f64)); |
| 78 | try expect(ldexp(@as(f80, 0x1.7FFFFFFFFFFFFFFEp-1), -16382 - 62) == math.floatTrueMin(f80)); |
| 79 | try expect(ldexp(@as(f128, 0x1.7FFFFFFFFFFFFFFFFFFFFFFFFFFFp-1), -16382 - 111) == math.floatTrueMin(f128)); |
| 80 | |
| 81 | try expect(ldexp(math.floatMax(f32), -128 - 149) > 0.0); |
| 82 | try expect(ldexp(math.floatMax(f32), -128 - 149 - 1) == 0.0); |
| 83 | |
| 84 | @setEvalBranchQuota(10_000); |
| 85 | |
| 86 | inline for ([_]type{ f16, f32, f64, f80, f128 }) |T| { |
| 87 | const fractional_bits = math.floatFractionalBits(T); |
| 88 | |
| 89 | const min_exponent = math.floatExponentMin(T); |
| 90 | const max_exponent = math.floatExponentMax(T); |
| 91 | const exponent_bias = max_exponent; |
| 92 | |
| 93 | // basic usage |
| 94 | try expect(ldexp(@as(T, 1.5), 4) == 24.0); |
| 95 | |
| 96 | // normals -> subnormals |
| 97 | try expect(math.isNormal(ldexp(@as(T, 1.0), min_exponent))); |
| 98 | try expect(!math.isNormal(ldexp(@as(T, 1.0), min_exponent - 1))); |
| 99 | |
| 100 | // normals -> zero |
| 101 | try expect(ldexp(@as(T, 1.0), min_exponent - fractional_bits) > 0.0); |
| 102 | try expect(ldexp(@as(T, 1.0), min_exponent - fractional_bits - 1) == 0.0); |
| 103 | |
| 104 | // subnormals -> zero |
| 105 | try expect(ldexp(math.floatTrueMin(T), 0) > 0.0); |
| 106 | try expect(ldexp(math.floatTrueMin(T), -1) == 0.0); |
| 107 | |
| 108 | // Multiplications might flush the denormals to zero, esp. at |
| 109 | // runtime, so we manually construct the constants here instead. |
| 110 | const Z = @Int(.unsigned, @bitSizeOf(T)); |
| 111 | const EightTimesTrueMin = @as(T, @bitCast(@as(Z, 8))); |
| 112 | const TwoTimesTrueMin = @as(T, @bitCast(@as(Z, 2))); |
| 113 | |
| 114 | // subnormals -> subnormals |
| 115 | try expect(ldexp(math.floatTrueMin(T), 3) == EightTimesTrueMin); |
| 116 | try expect(ldexp(EightTimesTrueMin, -2) == TwoTimesTrueMin); |
| 117 | try expect(ldexp(EightTimesTrueMin, -3) == math.floatTrueMin(T)); |
| 118 | |
| 119 | // subnormals -> normals (+) |
| 120 | try expect(ldexp(math.floatTrueMin(T), fractional_bits) == math.floatMin(T)); |
| 121 | try expect(ldexp(math.floatTrueMin(T), fractional_bits - 1) == math.floatMin(T) * 0.5); |
| 122 | |
| 123 | // subnormals -> normals (-) |
| 124 | try expect(ldexp(-math.floatTrueMin(T), fractional_bits) == -math.floatMin(T)); |
| 125 | try expect(ldexp(-math.floatTrueMin(T), fractional_bits - 1) == -math.floatMin(T) * 0.5); |
| 126 | |
| 127 | // subnormals -> float limits (+inf) |
| 128 | try expect(math.isFinite(ldexp(math.floatTrueMin(T), max_exponent + exponent_bias + fractional_bits - 1))); |
| 129 | try expect(ldexp(math.floatTrueMin(T), max_exponent + exponent_bias + fractional_bits) == math.inf(T)); |
| 130 | |
| 131 | // subnormals -> float limits (-inf) |
| 132 | try expect(math.isFinite(ldexp(-math.floatTrueMin(T), max_exponent + exponent_bias + fractional_bits - 1))); |
| 133 | try expect(ldexp(-math.floatTrueMin(T), max_exponent + exponent_bias + fractional_bits) == -math.inf(T)); |
| 134 | |
| 135 | // infinity -> infinity |
| 136 | try expect(ldexp(math.inf(T), math.maxInt(i32)) == math.inf(T)); |
| 137 | try expect(ldexp(math.inf(T), math.minInt(i32)) == math.inf(T)); |
| 138 | try expect(ldexp(math.inf(T), max_exponent) == math.inf(T)); |
| 139 | try expect(ldexp(math.inf(T), min_exponent) == math.inf(T)); |
| 140 | try expect(ldexp(-math.inf(T), math.maxInt(i32)) == -math.inf(T)); |
| 141 | try expect(ldexp(-math.inf(T), math.minInt(i32)) == -math.inf(T)); |
| 142 | |
| 143 | // extremely large n |
| 144 | try expect(ldexp(math.floatMax(T), math.maxInt(i32)) == math.inf(T)); |
| 145 | try expect(ldexp(math.floatMax(T), -math.maxInt(i32)) == 0.0); |
| 146 | try expect(ldexp(math.floatMax(T), math.minInt(i32)) == 0.0); |
| 147 | try expect(ldexp(math.floatTrueMin(T), math.maxInt(i32)) == math.inf(T)); |
| 148 | try expect(ldexp(math.floatTrueMin(T), -math.maxInt(i32)) == 0.0); |
| 149 | try expect(ldexp(math.floatTrueMin(T), math.minInt(i32)) == 0.0); |
| 150 | } |
| 151 | } |