1 | //===-- Single-precision tan function -------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #include "src/math/tanf.h" |
10 | #include "sincosf_utils.h" |
11 | #include "src/__support/FPUtil/FEnvImpl.h" |
12 | #include "src/__support/FPUtil/FPBits.h" |
13 | #include "src/__support/FPUtil/PolyEval.h" |
14 | #include "src/__support/FPUtil/except_value_utils.h" |
15 | #include "src/__support/FPUtil/multiply_add.h" |
16 | #include "src/__support/FPUtil/nearest_integer.h" |
17 | #include "src/__support/common.h" |
18 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
19 | #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA |
20 | |
21 | #include <errno.h> |
22 | |
23 | namespace LIBC_NAMESPACE { |
24 | |
25 | // Exceptional cases for tanf. |
26 | constexpr size_t N_EXCEPTS = 6; |
27 | |
28 | constexpr fputil::ExceptValues<float, N_EXCEPTS> TANF_EXCEPTS{.values: { |
29 | // (inputs, RZ output, RU offset, RD offset, RN offset) |
30 | // x = 0x1.ada6aap27, tan(x) = 0x1.e80304p-3 (RZ) |
31 | {.input: 0x4d56d355, .rnd_towardzero_result: 0x3e740182, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 0}, |
32 | // x = 0x1.862064p33, tan(x) = -0x1.8dee56p-3 (RZ) |
33 | {.input: 0x50431032, .rnd_towardzero_result: 0xbe46f72b, .rnd_upward_offset: 0, .rnd_downward_offset: 1, .rnd_tonearest_offset: 1}, |
34 | // x = 0x1.af61dap48, tan(x) = 0x1.60d1c6p-2 (RZ) |
35 | {.input: 0x57d7b0ed, .rnd_towardzero_result: 0x3eb068e3, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
36 | // x = 0x1.0088bcp52, tan(x) = 0x1.ca1edp0 (RZ) |
37 | {.input: 0x5980445e, .rnd_towardzero_result: 0x3fe50f68, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 0}, |
38 | // x = 0x1.f90dfcp72, tan(x) = 0x1.597f9cp-1 (RZ) |
39 | {.input: 0x63fc86fe, .rnd_towardzero_result: 0x3f2cbfce, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 0}, |
40 | // x = 0x1.a6ce12p86, tan(x) = -0x1.c5612ep-1 (RZ) |
41 | {.input: 0x6ad36709, .rnd_towardzero_result: 0xbf62b097, .rnd_upward_offset: 0, .rnd_downward_offset: 1, .rnd_tonearest_offset: 0}, |
42 | }}; |
43 | |
44 | LLVM_LIBC_FUNCTION(float, tanf, (float x)) { |
45 | using FPBits = typename fputil::FPBits<float>; |
46 | FPBits xbits(x); |
47 | bool x_sign = xbits.uintval() >> 31; |
48 | uint32_t x_abs = xbits.uintval() & 0x7fff'ffffU; |
49 | |
50 | // |x| < pi/32 |
51 | if (LIBC_UNLIKELY(x_abs <= 0x3dc9'0fdbU)) { |
52 | double xd = static_cast<double>(x); |
53 | |
54 | // |x| < 0x1.0p-12f |
55 | if (LIBC_UNLIKELY(x_abs < 0x3980'0000U)) { |
56 | if (LIBC_UNLIKELY(x_abs == 0U)) { |
57 | // For signed zeros. |
58 | return x; |
59 | } |
60 | // When |x| < 2^-12, the relative error of the approximation tan(x) ~ x |
61 | // is: |
62 | // |tan(x) - x| / |tan(x)| < |x^3| / (3|x|) |
63 | // = x^2 / 3 |
64 | // < 2^-25 |
65 | // < epsilon(1)/2. |
66 | // So the correctly rounded values of tan(x) are: |
67 | // = x + sign(x)*eps(x) if rounding mode = FE_UPWARD and x is positive, |
68 | // or (rounding mode = FE_DOWNWARD and x is |
69 | // negative), |
70 | // = x otherwise. |
71 | // To simplify the rounding decision and make it more efficient, we use |
72 | // fma(x, 2^-25, x) instead. |
73 | // Note: to use the formula x + 2^-25*x to decide the correct rounding, we |
74 | // do need fma(x, 2^-25, x) to prevent underflow caused by 2^-25*x when |
75 | // |x| < 2^-125. For targets without FMA instructions, we simply use |
76 | // double for intermediate results as it is more efficient than using an |
77 | // emulated version of FMA. |
78 | #if defined(LIBC_TARGET_CPU_HAS_FMA) |
79 | return fputil::multiply_add(x, y: 0x1.0p-25f, z: x); |
80 | #else |
81 | return static_cast<float>(fputil::multiply_add(xd, 0x1.0p-25, xd)); |
82 | #endif // LIBC_TARGET_CPU_HAS_FMA |
83 | } |
84 | |
85 | // |x| < pi/32 |
86 | double xsq = xd * xd; |
87 | |
88 | // Degree-9 minimax odd polynomial of tan(x) generated by Sollya with: |
89 | // > P = fpminimax(tan(x)/x, [|0, 2, 4, 6, 8|], [|1, D...|], [0, pi/32]); |
90 | double result = |
91 | fputil::polyeval(x: xsq, a0: 1.0, a: 0x1.555555553d022p-2, a: 0x1.111111ce442c1p-3, |
92 | a: 0x1.ba180a6bbdecdp-5, a: 0x1.69c0a88a0b71fp-6); |
93 | return static_cast<float>(xd * result); |
94 | } |
95 | |
96 | // Check for exceptional values |
97 | if (LIBC_UNLIKELY(x_abs == 0x3f8a1f62U)) { |
98 | // |x| = 0x1.143ec4p0 |
99 | float sign = x_sign ? -1.0f : 1.0f; |
100 | |
101 | // volatile is used to prevent compiler (gcc) from optimizing the |
102 | // computation, making the results incorrect in different rounding modes. |
103 | volatile float tmp = 0x1.ddf9f4p0f; |
104 | tmp = fputil::multiply_add(x: sign, y: tmp, z: sign * 0x1.1p-24f); |
105 | |
106 | return tmp; |
107 | } |
108 | |
109 | // |x| > 0x1.ada6a8p+27f |
110 | if (LIBC_UNLIKELY(x_abs > 0x4d56'd354U)) { |
111 | // Inf or NaN |
112 | if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) { |
113 | if (x_abs == 0x7f80'0000U) { |
114 | fputil::set_errno_if_required(EDOM); |
115 | fputil::raise_except_if_required(FE_INVALID); |
116 | } |
117 | return x + FPBits::quiet_nan().get_val(); |
118 | } |
119 | // Other large exceptional values |
120 | if (auto r = TANF_EXCEPTS.lookup_odd(x_abs, sign: x_sign); |
121 | LIBC_UNLIKELY(r.has_value())) |
122 | return r.value(); |
123 | } |
124 | |
125 | // For |x| >= pi/32, we use the definition of tan(x) function: |
126 | // tan(x) = sin(x) / cos(x) |
127 | // The we follow the same computations of sin(x) and cos(x) as sinf, cosf, |
128 | // and sincosf. |
129 | |
130 | double xd = static_cast<double>(x); |
131 | double sin_k, cos_k, sin_y, cosm1_y; |
132 | |
133 | sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y); |
134 | // tan(x) = sin(x) / cos(x) |
135 | // = (sin_y * cos_k + cos_y * sin_k) / (cos_y * cos_k - sin_y * sin_k) |
136 | using fputil::multiply_add; |
137 | return static_cast<float>( |
138 | multiply_add(x: sin_y, y: cos_k, z: multiply_add(x: cosm1_y, y: sin_k, z: sin_k)) / |
139 | multiply_add(x: sin_y, y: -sin_k, z: multiply_add(x: cosm1_y, y: cos_k, z: cos_k))); |
140 | } |
141 | |
142 | } // namespace LIBC_NAMESPACE |
143 | |