| 1 | //! Ported from musl, which is licensed under the MIT license: |
| 2 | //! https://git.musl-libc.org/cgit/musl/tree/COPYRIGHT |
| 3 | //! |
| 4 | //! https://git.musl-libc.org/cgit/musl/tree/src/math/cosf.c |
| 5 | //! https://git.musl-libc.org/cgit/musl/tree/src/math/cos.c |
| 6 | //! https://git.musl-libc.org/cgit/musl/tree/src/math/cosl.c |
| 7 | |
| 8 | const std = @import("std"); |
| 9 | const math = std.math; |
| 10 | const ld = math.long_double; |
| 11 | const mem = std.mem; |
| 12 | const expect = std.testing.expect; |
| 13 | const expectApproxEqAbs = std.testing.expectApproxEqAbs; |
| 14 | |
| 15 | const compiler_rt = @import("../compiler_rt.zig"); |
| 16 | const symbol = compiler_rt.symbol; |
| 17 | const trig = @import("trig.zig"); |
| 18 | const rem_pio2 = @import("rem_pio2.zig").rem_pio2; |
| 19 | const rem_pio2f = @import("rem_pio2f.zig").rem_pio2f; |
| 20 | const rem_pio2l = @import("rem_pio2l.zig").rem_pio2l; |
| 21 | |
| 22 | comptime { |
| 23 | symbol(&__cosh, "__cosh"); |
| 24 | symbol(&cosf, "cosf"); |
| 25 | symbol(&cos, "cos"); |
| 26 | symbol(&__cosx, "__cosx"); |
| 27 | symbol(&cosq, "cosf128"); |
| 28 | symbol(&cosl, "cosl"); |
| 29 | symbol(&cosl, "__cosl"); // required by musl |
| 30 | } |
| 31 | |
| 32 | fn __cosh(x: compiler_rt.f16.Abi) callconv(.c) compiler_rt.f16.Abi { |
| 33 | return compiler_rt.f16.toAbi(cos_f16(compiler_rt.f16.fromAbi(x))); |
| 34 | } |
| 35 | pub fn cos_f16(x: f16) f16 { |
| 36 | // TODO: more efficient implementation |
| 37 | return @floatCast(cos_f32(x)); |
| 38 | } |
| 39 | |
| 40 | fn cosf(x: compiler_rt.f32.Abi) callconv(.c) compiler_rt.f32.Abi { |
| 41 | return compiler_rt.f32.toAbi(cos_f32(compiler_rt.f32.fromAbi(x))); |
| 42 | } |
| 43 | pub fn cos_f32(x: f32) f32 { |
| 44 | // Small multiples of pi/2 rounded to double precision. |
| 45 | const c1pio2: f64 = 1.0 * math.pi / 2.0; // 0x3FF921FB, 0x54442D18 |
| 46 | const c2pio2: f64 = 2.0 * math.pi / 2.0; // 0x400921FB, 0x54442D18 |
| 47 | const c3pio2: f64 = 3.0 * math.pi / 2.0; // 0x4012D97C, 0x7F3321D2 |
| 48 | const c4pio2: f64 = 4.0 * math.pi / 2.0; // 0x401921FB, 0x54442D18 |
| 49 | |
| 50 | var ix: u32 = @bitCast(x); |
| 51 | const sign = ix >> 31 != 0; |
| 52 | ix &= 0x7fffffff; |
| 53 | |
| 54 | if (ix <= 0x3f490fda) { // |x| ~<= pi/4 |
| 55 | if (ix < 0x39800000) { // |x| < 2**-12 |
| 56 | // raise inexact if x != 0 |
| 57 | if (compiler_rt.want_float_exceptions) mem.doNotOptimizeAway(x + 0x1p120); |
| 58 | return 1.0; |
| 59 | } |
| 60 | return trig.cosdf(x); |
| 61 | } |
| 62 | if (ix <= 0x407b53d1) { // |x| ~<= 5*pi/4 |
| 63 | if (ix > 0x4016cbe3) { // |x| ~> 3*pi/4 |
| 64 | return -trig.cosdf(if (sign) x + c2pio2 else x - c2pio2); |
| 65 | } else { |
| 66 | if (sign) { |
| 67 | return trig.sindf(x + c1pio2); |
| 68 | } else { |
| 69 | return trig.sindf(c1pio2 - x); |
| 70 | } |
| 71 | } |
| 72 | } |
| 73 | if (ix <= 0x40e231d5) { // |x| ~<= 9*pi/4 |
| 74 | if (ix > 0x40afeddf) { // |x| ~> 7*pi/4 |
| 75 | return trig.cosdf(if (sign) x + c4pio2 else x - c4pio2); |
| 76 | } else { |
| 77 | if (sign) { |
| 78 | return trig.sindf(-x - c3pio2); |
| 79 | } else { |
| 80 | return trig.sindf(x - c3pio2); |
| 81 | } |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | // cos(Inf or NaN) is NaN |
| 86 | if (ix >= 0x7f800000) { |
| 87 | return x - x; |
| 88 | } |
| 89 | |
| 90 | var y: f64 = undefined; |
| 91 | const n = rem_pio2f(x, &y); |
| 92 | return switch (n & 3) { |
| 93 | 0 => trig.cosdf(y), |
| 94 | 1 => trig.sindf(-y), |
| 95 | 2 => -trig.cosdf(y), |
| 96 | else => trig.sindf(y), |
| 97 | }; |
| 98 | } |
| 99 | |
| 100 | fn cos(x: compiler_rt.f64.Abi) callconv(.c) compiler_rt.f64.Abi { |
| 101 | return compiler_rt.f64.toAbi(cos_f64(compiler_rt.f64.fromAbi(x))); |
| 102 | } |
| 103 | pub fn cos_f64(x: f64) f64 { |
| 104 | var ix = @as(u64, @bitCast(x)) >> 32; |
| 105 | ix &= 0x7fffffff; |
| 106 | |
| 107 | // |x| ~< pi/4 |
| 108 | if (ix <= 0x3fe921fb) { |
| 109 | if (ix < 0x3e46a09e) { // |x| < 2**-27 * sqrt(2) |
| 110 | // raise inexact if x!=0 |
| 111 | if (compiler_rt.want_float_exceptions) mem.doNotOptimizeAway(x + 0x1p120); |
| 112 | return 1.0; |
| 113 | } |
| 114 | return trig.cos(x, 0); |
| 115 | } |
| 116 | |
| 117 | // cos(Inf or NaN) is NaN |
| 118 | if (ix >= 0x7ff00000) { |
| 119 | return x - x; |
| 120 | } |
| 121 | |
| 122 | var y: [2]f64 = undefined; |
| 123 | const n = rem_pio2(x, &y); |
| 124 | return switch (n & 3) { |
| 125 | 0 => trig.cos(y[0], y[1]), |
| 126 | 1 => -trig.sin(y[0], y[1], 1), |
| 127 | 2 => -trig.cos(y[0], y[1]), |
| 128 | else => trig.sin(y[0], y[1], 1), |
| 129 | }; |
| 130 | } |
| 131 | |
| 132 | fn __cosx(x: compiler_rt.f80.Abi) callconv(.c) compiler_rt.f80.Abi { |
| 133 | return compiler_rt.f80.toAbi(cos_f80(compiler_rt.f80.fromAbi(x))); |
| 134 | } |
| 135 | pub fn cos_f80(x: f80) f80 { |
| 136 | const se = ld.signExponent(x) & 0x7fff; |
| 137 | if (se == 0x7fff) { |
| 138 | return x - x; |
| 139 | } |
| 140 | |
| 141 | if (@abs(x) < trig.pi_4) { |
| 142 | if (se < 0x3fff - math.floatMantissaBits(f80)) { |
| 143 | // raise inexact if x!=0 |
| 144 | return 1.0 + x; |
| 145 | } |
| 146 | return trig.cosx(x, 0.0); |
| 147 | } |
| 148 | |
| 149 | var y: [2]f80 = undefined; |
| 150 | const n = rem_pio2l(f80, x, &y); |
| 151 | return switch (n & 3) { |
| 152 | 0 => trig.cosx(y[0], y[1]), |
| 153 | 1 => -trig.sinx(y[0], y[1], 1), |
| 154 | 2 => -trig.cosx(y[0], y[1]), |
| 155 | else => trig.sinx(y[0], y[1], 1), |
| 156 | }; |
| 157 | } |
| 158 | |
| 159 | fn cosq(x: compiler_rt.f128.Abi) callconv(.c) compiler_rt.f128.Abi { |
| 160 | return compiler_rt.f128.toAbi(cos_f128(compiler_rt.f128.fromAbi(x))); |
| 161 | } |
| 162 | pub fn cos_f128(x: f128) f128 { |
| 163 | const se = ld.signExponent(x) & 0x7fff; |
| 164 | if (se == 0x7fff) { |
| 165 | return x - x; |
| 166 | } |
| 167 | |
| 168 | if (@abs(x) < trig.pi_4) { |
| 169 | if (se < 0x3fff - math.floatMantissaBits(f128)) { |
| 170 | // raise inexact if x!=0 |
| 171 | return 1.0 + x; |
| 172 | } |
| 173 | return trig.cosq(x, 0.0); |
| 174 | } |
| 175 | |
| 176 | var y: [2]f128 = undefined; |
| 177 | const n = rem_pio2l(f128, x, &y); |
| 178 | return switch (n & 3) { |
| 179 | 0 => trig.cosq(y[0], y[1]), |
| 180 | 1 => -trig.sinq(y[0], y[1], 1), |
| 181 | 2 => -trig.cosq(y[0], y[1]), |
| 182 | else => trig.sinq(y[0], y[1], 1), |
| 183 | }; |
| 184 | } |
| 185 | |
| 186 | pub fn cosl(x: c_longdouble) callconv(.c) c_longdouble { |
| 187 | switch (@typeInfo(c_longdouble).float.bits) { |
| 188 | 64 => return cos_f64(x), |
| 189 | 80 => return cos_f80(x), |
| 190 | 128 => return cos_f128(x), |
| 191 | else => comptime unreachable, |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | fn testCosSpecial(comptime T: type) !void { |
| 196 | const f = switch (T) { |
| 197 | f16 => cos_f16, |
| 198 | f32 => cos_f32, |
| 199 | f64 => cos_f64, |
| 200 | f80 => cos_f80, |
| 201 | f128 => cos_f128, |
| 202 | else => comptime unreachable, |
| 203 | }; |
| 204 | |
| 205 | try expect(f(0.0) == 1.0); |
| 206 | try expect(f(-0.0) == 1.0); |
| 207 | try expect(math.isNan(f(math.inf(T)))); |
| 208 | try expect(math.isNan(f(-math.inf(T)))); |
| 209 | try expect(math.isNan(f(math.nan(T)))); |
| 210 | } |
| 211 | |
| 212 | test "cos32.normal" { |
| 213 | const epsilon = math.floatEps(f32); |
| 214 | try expectApproxEqAbs(@as(f32, 1.0), cos_f32(0.0), epsilon); |
| 215 | try expectApproxEqAbs(@as(f32, 0.9800666), cos_f32(0.2), epsilon); |
| 216 | try expectApproxEqAbs(@as(f32, 0.6276231), cos_f32(0.8923), epsilon); |
| 217 | try expectApproxEqAbs(@as(f32, 0.0707372), cos_f32(1.5), epsilon); |
| 218 | try expectApproxEqAbs(@as(f32, 0.0707372), cos_f32(-1.5), epsilon); |
| 219 | try expectApproxEqAbs(@as(f32, 0.96913195), cos_f32(37.45), epsilon); |
| 220 | try expectApproxEqAbs(@as(f32, 0.40079966), cos_f32(89.123), epsilon); |
| 221 | } |
| 222 | |
| 223 | test "cos32.special" { |
| 224 | try testCosSpecial(f32); |
| 225 | } |
| 226 | |
| 227 | test "cos64.normal" { |
| 228 | const epsilon = math.floatEps(f64); |
| 229 | try expectApproxEqAbs(@as(f64, 1.0), cos_f64(0.0), epsilon); |
| 230 | try expectApproxEqAbs(@as(f64, 0.9800665778412416), cos_f64(0.2), epsilon); |
| 231 | try expectApproxEqAbs(@as(f64, 0.6276230983360804), cos_f64(0.8923), epsilon); |
| 232 | try expectApproxEqAbs(@as(f64, 0.0707372016677029), cos_f64(1.5), epsilon); |
| 233 | try expectApproxEqAbs(@as(f64, 0.0707372016677029), cos_f64(-1.5), epsilon); |
| 234 | try expectApproxEqAbs(@as(f64, 0.9691317730707778), cos_f64(37.45), epsilon); |
| 235 | try expectApproxEqAbs(@as(f64, 0.4008006809354791), cos_f64(89.123), epsilon); |
| 236 | } |
| 237 | |
| 238 | test "cos64.special" { |
| 239 | try testCosSpecial(f64); |
| 240 | } |
| 241 | |
| 242 | test "cos80.normal" { |
| 243 | const epsilon = math.floatEps(f80); |
| 244 | try expectApproxEqAbs(@as(f80, 1.0), cos_f80(0.0), epsilon); |
| 245 | try expectApproxEqAbs(@as(f80, 0.98006657784124163112419651674816888), cos_f80(0.2), epsilon); |
| 246 | try expectApproxEqAbs(@as(f80, 0.62762309833608037003563995939286067), cos_f80(0.8923), epsilon); |
| 247 | try expectApproxEqAbs(@as(f80, 0.070737201667702910088189851434268747), cos_f80(1.5), epsilon); |
| 248 | try expectApproxEqAbs(@as(f80, 0.070737201667702910088189851434268747), cos_f80(-1.5), epsilon); |
| 249 | try expectApproxEqAbs(@as(f80, 0.9691317730707771246), cos_f80(37.45), epsilon); |
| 250 | try expectApproxEqAbs(@as(f80, 0.4008006809354834001), cos_f80(89.123), epsilon); |
| 251 | } |
| 252 | |
| 253 | test "cos80.special" { |
| 254 | try testCosSpecial(f80); |
| 255 | } |
| 256 | |
| 257 | test "cos128.normal" { |
| 258 | const epsilon = math.floatEps(f128); |
| 259 | try expectApproxEqAbs(@as(f128, 1.0), cos_f128(0.0), epsilon); |
| 260 | try expectApproxEqAbs(@as(f128, 0.98006657784124163112419651674816888), cos_f128(0.2), epsilon); |
| 261 | try expectApproxEqAbs(@as(f128, 0.62762309833608037003563995939286067), cos_f128(0.8923), epsilon); |
| 262 | try expectApproxEqAbs(@as(f128, 0.070737201667702910088189851434268747), cos_f128(1.5), epsilon); |
| 263 | try expectApproxEqAbs(@as(f128, 0.070737201667702910088189851434268747), cos_f128(-1.5), epsilon); |
| 264 | try expectApproxEqAbs(@as(f128, 0.96913177307077712443149563847233230), cos_f128(37.45), epsilon); |
| 265 | try expectApproxEqAbs(@as(f128, 0.40080068093548339848199454493704702), cos_f128(89.123), epsilon); |
| 266 | } |
| 267 | |
| 268 | test "cos128.special" { |
| 269 | try testCosSpecial(f128); |
| 270 | } |