1 | //===-- Single-precision acos 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/acosf.h" |
10 | #include "src/__support/FPUtil/FEnvImpl.h" |
11 | #include "src/__support/FPUtil/FPBits.h" |
12 | #include "src/__support/FPUtil/PolyEval.h" |
13 | #include "src/__support/FPUtil/except_value_utils.h" |
14 | #include "src/__support/FPUtil/multiply_add.h" |
15 | #include "src/__support/FPUtil/sqrt.h" |
16 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
17 | |
18 | #include <errno.h> |
19 | |
20 | #include "inv_trigf_utils.h" |
21 | |
22 | namespace LIBC_NAMESPACE { |
23 | |
24 | static constexpr size_t N_EXCEPTS = 4; |
25 | |
26 | // Exceptional values when |x| <= 0.5 |
27 | static constexpr fputil::ExceptValues<float, N_EXCEPTS> ACOSF_EXCEPTS = {.values: { |
28 | // (inputs, RZ output, RU offset, RD offset, RN offset) |
29 | // x = 0x1.110b46p-26, acosf(x) = 0x1.921fb4p0 (RZ) |
30 | {.input: 0x328885a3, .rnd_towardzero_result: 0x3fc90fda, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
31 | // x = -0x1.110b46p-26, acosf(x) = 0x1.921fb4p0 (RZ) |
32 | {.input: 0xb28885a3, .rnd_towardzero_result: 0x3fc90fda, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
33 | // x = 0x1.04c444p-12, acosf(x) = 0x1.920f68p0 (RZ) |
34 | {.input: 0x39826222, .rnd_towardzero_result: 0x3fc907b4, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
35 | // x = -0x1.04c444p-12, acosf(x) = 0x1.923p0 (RZ) |
36 | {.input: 0xb9826222, .rnd_towardzero_result: 0x3fc91800, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
37 | }}; |
38 | |
39 | LLVM_LIBC_FUNCTION(float, acosf, (float x)) { |
40 | using FPBits = typename fputil::FPBits<float>; |
41 | |
42 | FPBits xbits(x); |
43 | uint32_t x_uint = xbits.uintval(); |
44 | uint32_t x_abs = xbits.uintval() & 0x7fff'ffffU; |
45 | uint32_t x_sign = x_uint >> 31; |
46 | |
47 | // |x| <= 0.5 |
48 | if (LIBC_UNLIKELY(x_abs <= 0x3f00'0000U)) { |
49 | // |x| < 0x1p-10 |
50 | if (LIBC_UNLIKELY(x_abs < 0x3a80'0000U)) { |
51 | // When |x| < 2^-10, we use the following approximation: |
52 | // acos(x) = pi/2 - asin(x) |
53 | // ~ pi/2 - x - x^3 / 6 |
54 | |
55 | // Check for exceptional values |
56 | if (auto r = ACOSF_EXCEPTS.lookup(x_bits: x_uint); LIBC_UNLIKELY(r.has_value())) |
57 | return r.value(); |
58 | |
59 | double xd = static_cast<double>(x); |
60 | return static_cast<float>(fputil::multiply_add( |
61 | x: -0x1.5555555555555p-3 * xd, y: xd * xd, z: M_MATH_PI_2 - xd)); |
62 | } |
63 | |
64 | // For |x| <= 0.5, we approximate acosf(x) by: |
65 | // acos(x) = pi/2 - asin(x) = pi/2 - x * P(x^2) |
66 | // Where P(X^2) = Q(X) is a degree-20 minimax even polynomial approximating |
67 | // asin(x)/x on [0, 0.5] generated by Sollya with: |
68 | // > Q = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20|], |
69 | // [|1, D...|], [0, 0.5]); |
70 | double xd = static_cast<double>(x); |
71 | double xsq = xd * xd; |
72 | double x3 = xd * xsq; |
73 | double r = asin_eval(xsq); |
74 | return static_cast<float>(fputil::multiply_add(x: -x3, y: r, z: M_MATH_PI_2 - xd)); |
75 | } |
76 | |
77 | // |x| >= 1, return 0, 2pi, or NaNs. |
78 | if (LIBC_UNLIKELY(x_abs >= 0x3f80'0000U)) { |
79 | if (x_abs == 0x3f80'0000U) |
80 | return x_sign ? /* x == -1.0f */ fputil::round_result_slightly_down( |
81 | value_rn: 0x1.921fb6p+1f) |
82 | : /* x == 1.0f */ 0.0f; |
83 | |
84 | if (x_abs <= 0x7f80'0000U) { |
85 | fputil::set_errno_if_required(EDOM); |
86 | fputil::raise_except_if_required(FE_INVALID); |
87 | } |
88 | return x + FPBits::quiet_nan().get_val(); |
89 | } |
90 | |
91 | // When 0.5 < |x| < 1, we perform range reduction as follow: |
92 | // |
93 | // Assume further that 0.5 < x <= 1, and let: |
94 | // y = acos(x) |
95 | // We use the double angle formula: |
96 | // x = cos(y) = 1 - 2 sin^2(y/2) |
97 | // So: |
98 | // sin(y/2) = sqrt( (1 - x)/2 ) |
99 | // And hence: |
100 | // y = 2 * asin( sqrt( (1 - x)/2 ) ) |
101 | // Let u = (1 - x)/2, then |
102 | // acos(x) = 2 * asin( sqrt(u) ) |
103 | // Moreover, since 0.5 < x <= 1, |
104 | // 0 <= u < 1/4, and 0 <= sqrt(u) < 0.5, |
105 | // And hence we can reuse the same polynomial approximation of asin(x) when |
106 | // |x| <= 0.5: |
107 | // acos(x) ~ 2 * sqrt(u) * P(u). |
108 | // |
109 | // When -1 < x <= -0.5, we use the identity: |
110 | // acos(x) = pi - acos(-x) |
111 | // which is reduced to the postive case. |
112 | |
113 | xbits.set_sign(Sign::POS); |
114 | double xd = static_cast<double>(xbits.get_val()); |
115 | double u = fputil::multiply_add(x: -0.5, y: xd, z: 0.5); |
116 | double cv = 2 * fputil::sqrt(x: u); |
117 | |
118 | double r3 = asin_eval(xsq: u); |
119 | double r = fputil::multiply_add(x: cv * u, y: r3, z: cv); |
120 | return static_cast<float>(x_sign ? M_MATH_PI - r : r); |
121 | } |
122 | |
123 | } // namespace LIBC_NAMESPACE |
124 | |