| 1 | const std = @import("std"); |
| 2 | const builtin = @import("builtin"); |
| 3 | const arch = builtin.cpu.arch; |
| 4 | |
| 5 | const compiler_rt = @import("../compiler_rt.zig"); |
| 6 | const symbol = compiler_rt.symbol; |
| 7 | const normalize = compiler_rt.normalize; |
| 8 | const wideMultiply = compiler_rt.wideMultiply; |
| 9 | |
| 10 | comptime { |
| 11 | symbol(&__divxf3, "__divxf3"); |
| 12 | } |
| 13 | |
| 14 | fn __divxf3(a: compiler_rt.f80.Abi, b: compiler_rt.f80.Abi) callconv(.c) compiler_rt.f80.Abi { |
| 15 | return compiler_rt.f80.toAbi(div_f80(compiler_rt.f80.fromAbi(a), compiler_rt.f80.fromAbi(b))); |
| 16 | } |
| 17 | pub fn div_f80(a: f80, b: f80) f80 { |
| 18 | const T = f80; |
| 19 | const Z = @Int(.unsigned, @bitSizeOf(T)); |
| 20 | |
| 21 | const significandBits = std.math.floatMantissaBits(T); |
| 22 | const fractionalBits = std.math.floatFractionalBits(T); |
| 23 | const exponentBits = std.math.floatExponentBits(T); |
| 24 | |
| 25 | const signBit = (@as(Z, 1) << (significandBits + exponentBits)); |
| 26 | const maxExponent = ((1 << exponentBits) - 1); |
| 27 | const exponentBias = (maxExponent >> 1); |
| 28 | |
| 29 | const integerBit = (@as(Z, 1) << fractionalBits); |
| 30 | const quietBit = integerBit >> 1; |
| 31 | const significandMask = (@as(Z, 1) << significandBits) - 1; |
| 32 | |
| 33 | const absMask = signBit - 1; |
| 34 | const qnanRep = @as(Z, @bitCast(std.math.nan(T))) | quietBit; |
| 35 | const infRep: Z = @bitCast(std.math.inf(T)); |
| 36 | |
| 37 | const aExponent: u32 = @truncate((@as(Z, @bitCast(a)) >> significandBits) & maxExponent); |
| 38 | const bExponent: u32 = @truncate((@as(Z, @bitCast(b)) >> significandBits) & maxExponent); |
| 39 | const quotientSign: Z = (@as(Z, @bitCast(a)) ^ @as(Z, @bitCast(b))) & signBit; |
| 40 | |
| 41 | var aSignificand: Z = @as(Z, @bitCast(a)) & significandMask; |
| 42 | var bSignificand: Z = @as(Z, @bitCast(b)) & significandMask; |
| 43 | var scale: i32 = 0; |
| 44 | |
| 45 | // Detect if a or b is zero, denormal, infinity, or NaN. |
| 46 | if (aExponent -% 1 >= maxExponent - 1 or bExponent -% 1 >= maxExponent - 1) { |
| 47 | const aAbs: Z = @as(Z, @bitCast(a)) & absMask; |
| 48 | const bAbs: Z = @as(Z, @bitCast(b)) & absMask; |
| 49 | |
| 50 | // NaN / anything = qNaN |
| 51 | if (aAbs > infRep) return @bitCast(@as(Z, @bitCast(a)) | quietBit); |
| 52 | // anything / NaN = qNaN |
| 53 | if (bAbs > infRep) return @bitCast(@as(Z, @bitCast(b)) | quietBit); |
| 54 | |
| 55 | if (aAbs == infRep) { |
| 56 | // infinity / infinity = NaN |
| 57 | if (bAbs == infRep) { |
| 58 | return @bitCast(qnanRep); |
| 59 | } |
| 60 | // infinity / anything else = +/- infinity |
| 61 | else { |
| 62 | return @bitCast(aAbs | quotientSign); |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | // anything else / infinity = +/- 0 |
| 67 | if (bAbs == infRep) return @bitCast(quotientSign); |
| 68 | |
| 69 | if (aAbs == 0) { |
| 70 | // zero / zero = NaN |
| 71 | if (bAbs == 0) { |
| 72 | return @bitCast(qnanRep); |
| 73 | } |
| 74 | // zero / anything else = +/- zero |
| 75 | else { |
| 76 | return @bitCast(quotientSign); |
| 77 | } |
| 78 | } |
| 79 | // anything else / zero = +/- infinity |
| 80 | if (bAbs == 0) return @bitCast(infRep | quotientSign); |
| 81 | |
| 82 | // one or both of a or b is denormal, the other (if applicable) is a |
| 83 | // normal number. Renormalize one or both of a and b, and set scale to |
| 84 | // include the necessary exponent adjustment. |
| 85 | if (aAbs < integerBit) scale +%= normalize(T, &aSignificand); |
| 86 | if (bAbs < integerBit) scale -%= normalize(T, &bSignificand); |
| 87 | } |
| 88 | var quotientExponent: i32 = @as(i32, @bitCast(aExponent -% bExponent)) +% scale; |
| 89 | |
| 90 | // Align the significand of b as a Q63 fixed-point number in the range |
| 91 | // [1, 2.0) and get a Q64 approximate reciprocal using a small minimax |
| 92 | // polynomial approximation: reciprocal = 3/4 + 1/sqrt(2) - b/2. This |
| 93 | // is accurate to about 3.5 binary digits. |
| 94 | const q63b: u64 = @intCast(bSignificand); |
| 95 | var recip64 = @as(u64, 0x7504f333F9DE6484) -% q63b; |
| 96 | // 0x7504f333F9DE6484 / 2^64 + 1 = 3/4 + 1/sqrt(2) |
| 97 | |
| 98 | // Now refine the reciprocal estimate using a Newton-Raphson iteration: |
| 99 | // |
| 100 | // x1 = x0 * (2 - x0 * b) |
| 101 | // |
| 102 | // This doubles the number of correct binary digits in the approximation |
| 103 | // with each iteration. |
| 104 | var correction64: u64 = undefined; |
| 105 | correction64 = @truncate(~(@as(u128, recip64) *% q63b >> 64) +% 1); |
| 106 | recip64 = @truncate(@as(u128, recip64) *% correction64 >> 63); |
| 107 | correction64 = @truncate(~(@as(u128, recip64) *% q63b >> 64) +% 1); |
| 108 | recip64 = @truncate(@as(u128, recip64) *% correction64 >> 63); |
| 109 | correction64 = @truncate(~(@as(u128, recip64) *% q63b >> 64) +% 1); |
| 110 | recip64 = @truncate(@as(u128, recip64) *% correction64 >> 63); |
| 111 | correction64 = @truncate(~(@as(u128, recip64) *% q63b >> 64) +% 1); |
| 112 | recip64 = @truncate(@as(u128, recip64) *% correction64 >> 63); |
| 113 | correction64 = @truncate(~(@as(u128, recip64) *% q63b >> 64) +% 1); |
| 114 | recip64 = @truncate(@as(u128, recip64) *% correction64 >> 63); |
| 115 | |
| 116 | // The reciprocal may have overflowed to zero if the upper half of b is |
| 117 | // exactly 1.0. This would sabatoge the full-width final stage of the |
| 118 | // computation that follows, so we adjust the reciprocal down by one bit. |
| 119 | recip64 -%= 1; |
| 120 | |
| 121 | // We need to perform one more iteration to get us to 112 binary digits; |
| 122 | // The last iteration needs to happen with extra precision. |
| 123 | |
| 124 | // NOTE: This operation is equivalent to __multi3, which is not implemented |
| 125 | // in some architechures |
| 126 | var reciprocal: u128 = undefined; |
| 127 | var correction: u128 = undefined; |
| 128 | var dummy: u128 = undefined; |
| 129 | wideMultiply(u128, recip64, q63b, &dummy, &correction); |
| 130 | |
| 131 | correction = -%correction; |
| 132 | |
| 133 | const cHi: u64 = @truncate(correction >> 64); |
| 134 | const cLo: u64 = @truncate(correction); |
| 135 | |
| 136 | var r64cH: u128 = undefined; |
| 137 | var r64cL: u128 = undefined; |
| 138 | wideMultiply(u128, recip64, cHi, &dummy, &r64cH); |
| 139 | wideMultiply(u128, recip64, cLo, &dummy, &r64cL); |
| 140 | |
| 141 | reciprocal = r64cH + (r64cL >> 64); |
| 142 | |
| 143 | // Adjust the final 128-bit reciprocal estimate downward to ensure that it |
| 144 | // is strictly smaller than the infinitely precise exact reciprocal. Because |
| 145 | // the computation of the Newton-Raphson step is truncating at every step, |
| 146 | // this adjustment is small; most of the work is already done. |
| 147 | reciprocal -%= 2; |
| 148 | |
| 149 | // The numerical reciprocal is accurate to within 2^-112, lies in the |
| 150 | // interval [0.5, 1.0), and is strictly smaller than the true reciprocal |
| 151 | // of b. Multiplying a by this reciprocal thus gives a numerical q = a/b |
| 152 | // in Q127 with the following properties: |
| 153 | // |
| 154 | // 1. q < a/b |
| 155 | // 2. q is in the interval [0.5, 2.0) |
| 156 | // 3. The error in q is bounded away from 2^-63 (actually, we have |
| 157 | // many bits to spare, but this is all we need). |
| 158 | |
| 159 | // We need a 128 x 128 multiply high to compute q. |
| 160 | var quotient128: u128 = undefined; |
| 161 | var quotientLo: u128 = undefined; |
| 162 | wideMultiply(u128, aSignificand << 2, reciprocal, &quotient128, &quotientLo); |
| 163 | |
| 164 | // Two cases: quotient is in [0.5, 1.0) or quotient is in [1.0, 2.0). |
| 165 | // Right shift the quotient if it falls in the [1,2) range and adjust the |
| 166 | // exponent accordingly. |
| 167 | const quotient: u64 = if (quotient128 < (integerBit << 1)) b: { |
| 168 | quotientExponent -= 1; |
| 169 | break :b @intCast(quotient128); |
| 170 | } else @intCast(quotient128 >> 1); |
| 171 | |
| 172 | // We are going to compute a residual of the form |
| 173 | // |
| 174 | // r = a - q*b |
| 175 | // |
| 176 | // We know from the construction of q that r satisfies: |
| 177 | // |
| 178 | // 0 <= r < ulp(q)*b |
| 179 | // |
| 180 | // If r is greater than 1/2 ulp(q)*b, then q rounds up. Otherwise, we |
| 181 | // already have the correct result. The exact halfway case cannot occur. |
| 182 | const residual: u64 = -%(quotient *% q63b); |
| 183 | |
| 184 | const writtenExponent = quotientExponent + exponentBias; |
| 185 | if (writtenExponent >= maxExponent) { |
| 186 | // If we have overflowed the exponent, return infinity. |
| 187 | return @bitCast(infRep | quotientSign); |
| 188 | } else if (writtenExponent < 1) { |
| 189 | if (writtenExponent == 0) { |
| 190 | // Check whether the rounded result is normal. |
| 191 | if (residual > (bSignificand >> 1)) { // round |
| 192 | if (quotient == (integerBit - 1)) // If the rounded result is normal, return it |
| 193 | return @bitCast(@as(Z, @bitCast(std.math.floatMin(T))) | quotientSign); |
| 194 | } |
| 195 | } |
| 196 | // Flush denormals to zero. In the future, it would be nice to add |
| 197 | // code to round them correctly. |
| 198 | return @bitCast(quotientSign); |
| 199 | } else { |
| 200 | const round = @intFromBool(residual > (bSignificand >> 1)); |
| 201 | // Insert the exponent |
| 202 | var absResult = quotient | (@as(Z, @intCast(writtenExponent)) << significandBits); |
| 203 | // Round |
| 204 | absResult +%= round; |
| 205 | // Insert the sign and return |
| 206 | return @bitCast(absResult | quotientSign | integerBit); |
| 207 | } |
| 208 | } |
| 209 | |
| 210 | test { |
| 211 | _ = @import("divxf3_test.zig"); |
| 212 | } |