| 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/tanf.c |
| 5 | //! https://git.musl-libc.org/cgit/musl/tree/src/math/tan.c |
| 6 | //! https://git.musl-libc.org/cgit/musl/tree/src/math/tanl.c |
| 7 | //! https://golang.org/src/math/tan.go |
| 8 | |
| 9 | const std = @import("std"); |
| 10 | const builtin = @import("builtin"); |
| 11 | const math = std.math; |
| 12 | const ld = math.long_double; |
| 13 | const mem = std.mem; |
| 14 | const expect = std.testing.expect; |
| 15 | const expectApproxEqAbs = std.testing.expectApproxEqAbs; |
| 16 | |
| 17 | const kernel = @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 | const arch = builtin.cpu.arch; |
| 23 | const compiler_rt = @import("../compiler_rt.zig"); |
| 24 | const symbol = compiler_rt.symbol; |
| 25 | |
| 26 | comptime { |
| 27 | symbol(&__tanh, "__tanh"); |
| 28 | symbol(&tanf, "tanf"); |
| 29 | symbol(&tan, "tan"); |
| 30 | symbol(&__tanx, "__tanx"); |
| 31 | symbol(&tanq, "tanf128"); |
| 32 | symbol(&tanl, "tanl"); |
| 33 | } |
| 34 | |
| 35 | fn __tanh(x: compiler_rt.f16.Abi) callconv(.c) compiler_rt.f16.Abi { |
| 36 | return compiler_rt.f16.toAbi(tan_f16(compiler_rt.f16.fromAbi(x))); |
| 37 | } |
| 38 | pub fn tan_f16(x: f16) f16 { |
| 39 | // TODO: more efficient implementation |
| 40 | return @floatCast(tan_f32(x)); |
| 41 | } |
| 42 | |
| 43 | fn tanf(x: compiler_rt.f32.Abi) callconv(.c) compiler_rt.f32.Abi { |
| 44 | return compiler_rt.f32.toAbi(tan_f32(compiler_rt.f32.fromAbi(x))); |
| 45 | } |
| 46 | pub fn tan_f32(x: f32) f32 { |
| 47 | // Small multiples of pi/2 rounded to double precision. |
| 48 | const t1pio2: f64 = 1.0 * math.pi / 2.0; // 0x3FF921FB, 0x54442D18 |
| 49 | const t2pio2: f64 = 2.0 * math.pi / 2.0; // 0x400921FB, 0x54442D18 |
| 50 | const t3pio2: f64 = 3.0 * math.pi / 2.0; // 0x4012D97C, 0x7F3321D2 |
| 51 | const t4pio2: f64 = 4.0 * math.pi / 2.0; // 0x401921FB, 0x54442D18 |
| 52 | |
| 53 | var ix: u32 = @bitCast(x); |
| 54 | const sign = ix >> 31 != 0; |
| 55 | ix &= 0x7fffffff; |
| 56 | |
| 57 | if (ix <= 0x3f490fda) { // |x| ~<= pi/4 |
| 58 | if (ix < 0x39800000) { // |x| < 2**-12 |
| 59 | // raise inexact if x!=0 and underflow if subnormal |
| 60 | if (compiler_rt.want_float_exceptions) { |
| 61 | if (ix < 0x00800000) { |
| 62 | mem.doNotOptimizeAway(x / 0x1p120); |
| 63 | } else { |
| 64 | mem.doNotOptimizeAway(x + 0x1p120); |
| 65 | } |
| 66 | } |
| 67 | return x; |
| 68 | } |
| 69 | return kernel.tandf(x, false); |
| 70 | } |
| 71 | if (ix <= 0x407b53d1) { // |x| ~<= 5*pi/4 |
| 72 | if (ix <= 0x4016cbe3) { // |x| ~<= 3pi/4 |
| 73 | return kernel.tandf((if (sign) x + t1pio2 else x - t1pio2), true); |
| 74 | } else { |
| 75 | return kernel.tandf((if (sign) x + t2pio2 else x - t2pio2), false); |
| 76 | } |
| 77 | } |
| 78 | if (ix <= 0x40e231d5) { // |x| ~<= 9*pi/4 |
| 79 | if (ix <= 0x40afeddf) { // |x| ~<= 7*pi/4 |
| 80 | return kernel.tandf((if (sign) x + t3pio2 else x - t3pio2), true); |
| 81 | } else { |
| 82 | return kernel.tandf((if (sign) x + t4pio2 else x - t4pio2), false); |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | // tan(Inf or NaN) is NaN |
| 87 | if (ix >= 0x7f800000) { |
| 88 | return x - x; |
| 89 | } |
| 90 | |
| 91 | var y: f64 = undefined; |
| 92 | const n = rem_pio2f(x, &y); |
| 93 | return kernel.tandf(y, n & 1 != 0); |
| 94 | } |
| 95 | |
| 96 | fn tan(x: compiler_rt.f64.Abi) callconv(.c) compiler_rt.f64.Abi { |
| 97 | return compiler_rt.f64.toAbi(tan_f64(compiler_rt.f64.fromAbi(x))); |
| 98 | } |
| 99 | pub fn tan_f64(x: f64) f64 { |
| 100 | var ix = @as(u64, @bitCast(x)) >> 32; |
| 101 | ix &= 0x7fffffff; |
| 102 | |
| 103 | // |x| ~< pi/4 |
| 104 | if (ix <= 0x3fe921fb) { |
| 105 | if (ix < 0x3e400000) { // |x| < 2**-27 |
| 106 | // raise inexact if x!=0 and underflow if subnormal |
| 107 | if (compiler_rt.want_float_exceptions) { |
| 108 | if (ix < 0x00100000) { |
| 109 | mem.doNotOptimizeAway(x / 0x1p120); |
| 110 | } else { |
| 111 | mem.doNotOptimizeAway(x + 0x1p120); |
| 112 | } |
| 113 | } |
| 114 | return x; |
| 115 | } |
| 116 | return kernel.tan(x, 0.0, false); |
| 117 | } |
| 118 | |
| 119 | // tan(Inf or NaN) is NaN |
| 120 | if (ix >= 0x7ff00000) { |
| 121 | return x - x; |
| 122 | } |
| 123 | |
| 124 | var y: [2]f64 = undefined; |
| 125 | const n = rem_pio2(x, &y); |
| 126 | return kernel.tan(y[0], y[1], n & 1 != 0); |
| 127 | } |
| 128 | |
| 129 | fn __tanx(x: compiler_rt.f80.Abi) callconv(.c) compiler_rt.f80.Abi { |
| 130 | return compiler_rt.f80.toAbi(tan_f80(compiler_rt.f80.fromAbi(x))); |
| 131 | } |
| 132 | pub fn tan_f80(x: f80) f80 { |
| 133 | const se = ld.signExponent(x) & 0x7fff; |
| 134 | if (se == 0x7fff) { |
| 135 | return x - x; |
| 136 | } |
| 137 | |
| 138 | if (@abs(x) < kernel.pi_4) { |
| 139 | if (se < 0x3fff - math.floatMantissaBits(f80) / 2) { |
| 140 | if (compiler_rt.want_float_exceptions) { |
| 141 | mem.doNotOptimizeAway(if (se == 0) x * 0x1p-120 else x + 0x1p120); |
| 142 | } |
| 143 | return x; |
| 144 | } |
| 145 | return kernel.tanx(x, 0.0, 0); |
| 146 | } |
| 147 | |
| 148 | var y: [2]f80 = undefined; |
| 149 | const n = rem_pio2l(f80, x, &y); |
| 150 | return kernel.tanx(y[0], y[1], n & 1); |
| 151 | } |
| 152 | |
| 153 | fn tanq(x: compiler_rt.f128.Abi) callconv(.c) compiler_rt.f128.Abi { |
| 154 | return compiler_rt.f128.toAbi(tan_f128(compiler_rt.f128.fromAbi(x))); |
| 155 | } |
| 156 | pub fn tan_f128(x: f128) f128 { |
| 157 | const se = ld.signExponent(x) & 0x7fff; |
| 158 | if (se == 0x7fff) { |
| 159 | return x - x; |
| 160 | } |
| 161 | |
| 162 | if (@abs(x) < kernel.pi_4) { |
| 163 | if (se < 0x3fff - math.floatMantissaBits(f128) / 2) { |
| 164 | if (compiler_rt.want_float_exceptions) { |
| 165 | mem.doNotOptimizeAway(if (se == 0) x * 0x1p-120 else x + 0x1p120); |
| 166 | } |
| 167 | return x; |
| 168 | } |
| 169 | return kernel.tanq(x, 0.0, 0); |
| 170 | } |
| 171 | |
| 172 | var y: [2]f128 = undefined; |
| 173 | const n = rem_pio2l(f128, x, &y); |
| 174 | return kernel.tanq(y[0], y[1], n & 1); |
| 175 | } |
| 176 | |
| 177 | pub fn tanl(x: c_longdouble) callconv(.c) c_longdouble { |
| 178 | switch (@typeInfo(c_longdouble).float.bits) { |
| 179 | 64 => return tan_f64(x), |
| 180 | 80 => return tan_f80(x), |
| 181 | 128 => return tan_f128(x), |
| 182 | else => comptime unreachable, |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | fn testTanNormal(comptime T: type) !void { |
| 187 | const f = switch (T) { |
| 188 | f16 => tan_f16, |
| 189 | f32 => tan_f32, |
| 190 | f64 => tan_f64, |
| 191 | f80 => tan_f80, |
| 192 | f128 => tan_f128, |
| 193 | else => comptime unreachable, |
| 194 | }; |
| 195 | const epsilon = 0.00001; |
| 196 | |
| 197 | try expectApproxEqAbs(@as(T, 0.0), f(0.0), epsilon); |
| 198 | try expectApproxEqAbs(@as(T, 0.202710), f(0.2), epsilon); |
| 199 | try expectApproxEqAbs(@as(T, 1.240422), f(0.8923), epsilon); |
| 200 | try expectApproxEqAbs(@as(T, 14.101420), f(1.5), epsilon); |
| 201 | try expectApproxEqAbs(@as(T, -0.254397), f(37.45), epsilon); |
| 202 | try expectApproxEqAbs(@as(T, 2.285837), f(89.123), epsilon); |
| 203 | } |
| 204 | |
| 205 | fn testTanSpecial(comptime T: type) !void { |
| 206 | const f = switch (T) { |
| 207 | f16 => tan_f16, |
| 208 | f32 => tan_f32, |
| 209 | f64 => tan_f64, |
| 210 | f80 => tan_f80, |
| 211 | f128 => tan_f128, |
| 212 | else => comptime unreachable, |
| 213 | }; |
| 214 | |
| 215 | try expect(math.isPositiveZero(f(0.0))); |
| 216 | try expect(math.isNegativeZero(f(-0.0))); |
| 217 | try expect(math.isNan(f(math.inf(f32)))); |
| 218 | try expect(math.isNan(f(-math.inf(f32)))); |
| 219 | try expect(math.isNan(f(math.nan(f32)))); |
| 220 | } |
| 221 | |
| 222 | test "tan32.normal" { |
| 223 | try testTanNormal(f32); |
| 224 | } |
| 225 | |
| 226 | test "tan64.normal" { |
| 227 | try testTanNormal(f64); |
| 228 | } |
| 229 | |
| 230 | test "tan80.normal" { |
| 231 | const epsilon = math.floatEps(f80); |
| 232 | |
| 233 | try expectApproxEqAbs(@as(f80, 0.0), tan_f80(0.0), epsilon); |
| 234 | try expectApproxEqAbs(@as(f80, 0.2027100355086724833213582716475345), tan_f80(0.2), epsilon); |
| 235 | try expectApproxEqAbs(@as(f80, 1.2404217445497097995561220131857544), tan_f80(0.8923), epsilon); |
| 236 | try expectApproxEqAbs(@as(f80, 14.10141994717171938764), tan_f80(1.5), epsilon); |
| 237 | try expectApproxEqAbs(@as(f80, -0.25439607116885656232), tan_f80(37.45), epsilon); |
| 238 | try expectApproxEqAbs(@as(f80, 2.2858376251355320963), tan_f80(89.123), epsilon); |
| 239 | } |
| 240 | |
| 241 | test "tan128.normal" { |
| 242 | const epsilon = math.floatEps(f128); |
| 243 | |
| 244 | try expectApproxEqAbs(@as(f128, 0.0), tan_f128(0.0), epsilon); |
| 245 | try expectApproxEqAbs(@as(f128, 0.2027100355086724833213582716475345), tan_f128(0.2), epsilon); |
| 246 | try expectApproxEqAbs(@as(f128, 1.2404217445497097995561220131857544), tan_f128(0.8923), epsilon); |
| 247 | try expectApproxEqAbs(@as(f128, 14.101419947171719387646083651987755), tan_f128(1.5), epsilon); |
| 248 | try expectApproxEqAbs(@as(f128, -0.2543960711688565630469573224504774), tan_f128(37.45), epsilon); |
| 249 | try expectApproxEqAbs(@as(f128, 2.2858376251355321074066028114094292), tan_f128(89.123), epsilon); |
| 250 | } |
| 251 | |
| 252 | test "tan32.special" { |
| 253 | try testTanSpecial(f32); |
| 254 | } |
| 255 | |
| 256 | test "tan64.special" { |
| 257 | try testTanSpecial(f64); |
| 258 | } |
| 259 | |
| 260 | test "tan80.special" { |
| 261 | try testTanSpecial(f80); |
| 262 | } |
| 263 | |
| 264 | test "tan128.special" { |
| 265 | try testTanSpecial(f128); |
| 266 | } |