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/config.h" |
17 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
18 | |
19 | #include "inv_trigf_utils.h" |
20 | |
21 | namespace LIBC_NAMESPACE_DECL { |
22 | |
23 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
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 = {{ |
28 | // (inputs, RZ output, RU offset, RD offset, RN offset) |
29 | // x = 0x1.110b46p-26, acosf(x) = 0x1.921fb4p0 (RZ) |
30 | {0x328885a3, 0x3fc90fda, 1, 0, 1}, |
31 | // x = -0x1.110b46p-26, acosf(x) = 0x1.921fb4p0 (RZ) |
32 | {0xb28885a3, 0x3fc90fda, 1, 0, 1}, |
33 | // x = 0x1.04c444p-12, acosf(x) = 0x1.920f68p0 (RZ) |
34 | {0x39826222, 0x3fc907b4, 1, 0, 1}, |
35 | // x = -0x1.04c444p-12, acosf(x) = 0x1.923p0 (RZ) |
36 | {0xb9826222, 0x3fc91800, 1, 0, 1}, |
37 | }}; |
38 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
39 | |
40 | LLVM_LIBC_FUNCTION(float, acosf, (float x)) { |
41 | using FPBits = typename fputil::FPBits<float>; |
42 | |
43 | FPBits xbits(x); |
44 | uint32_t x_uint = xbits.uintval(); |
45 | uint32_t x_abs = xbits.uintval() & 0x7fff'ffffU; |
46 | uint32_t x_sign = x_uint >> 31; |
47 | |
48 | // |x| <= 0.5 |
49 | if (LIBC_UNLIKELY(x_abs <= 0x3f00'0000U)) { |
50 | // |x| < 0x1p-10 |
51 | if (LIBC_UNLIKELY(x_abs < 0x3a80'0000U)) { |
52 | // When |x| < 2^-10, we use the following approximation: |
53 | // acos(x) = pi/2 - asin(x) |
54 | // ~ pi/2 - x - x^3 / 6 |
55 | |
56 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
57 | // Check for exceptional values |
58 | if (auto r = ACOSF_EXCEPTS.lookup(x_uint); LIBC_UNLIKELY(r.has_value())) |
59 | return r.value(); |
60 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
61 | |
62 | double xd = static_cast<double>(x); |
63 | return static_cast<float>(fputil::multiply_add( |
64 | -0x1.5555555555555p-3 * xd, xd * xd, M_MATH_PI_2 - xd)); |
65 | } |
66 | |
67 | // For |x| <= 0.5, we approximate acosf(x) by: |
68 | // acos(x) = pi/2 - asin(x) = pi/2 - x * P(x^2) |
69 | // Where P(X^2) = Q(X) is a degree-20 minimax even polynomial approximating |
70 | // asin(x)/x on [0, 0.5] generated by Sollya with: |
71 | // > Q = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20|], |
72 | // [|1, D...|], [0, 0.5]); |
73 | double xd = static_cast<double>(x); |
74 | double xsq = xd * xd; |
75 | double x3 = xd * xsq; |
76 | double r = asin_eval(xsq); |
77 | return static_cast<float>(fputil::multiply_add(-x3, r, M_MATH_PI_2 - xd)); |
78 | } |
79 | |
80 | // |x| >= 1, return 0, 2pi, or NaNs. |
81 | if (LIBC_UNLIKELY(x_abs >= 0x3f80'0000U)) { |
82 | if (x_abs == 0x3f80'0000U) |
83 | return x_sign ? /* x == -1.0f */ fputil::round_result_slightly_down( |
84 | 0x1.921fb6p+1f) |
85 | : /* x == 1.0f */ 0.0f; |
86 | |
87 | if (xbits.is_signaling_nan()) { |
88 | fputil::raise_except_if_required(FE_INVALID); |
89 | return FPBits::quiet_nan().get_val(); |
90 | } |
91 | |
92 | // |x| <= +/-inf |
93 | if (x_abs <= 0x7f80'0000U) { |
94 | fputil::set_errno_if_required(EDOM); |
95 | fputil::raise_except_if_required(FE_INVALID); |
96 | } |
97 | |
98 | return x + FPBits::quiet_nan().get_val(); |
99 | } |
100 | |
101 | // When 0.5 < |x| < 1, we perform range reduction as follow: |
102 | // |
103 | // Assume further that 0.5 < x <= 1, and let: |
104 | // y = acos(x) |
105 | // We use the double angle formula: |
106 | // x = cos(y) = 1 - 2 sin^2(y/2) |
107 | // So: |
108 | // sin(y/2) = sqrt( (1 - x)/2 ) |
109 | // And hence: |
110 | // y = 2 * asin( sqrt( (1 - x)/2 ) ) |
111 | // Let u = (1 - x)/2, then |
112 | // acos(x) = 2 * asin( sqrt(u) ) |
113 | // Moreover, since 0.5 < x <= 1, |
114 | // 0 <= u < 1/4, and 0 <= sqrt(u) < 0.5, |
115 | // And hence we can reuse the same polynomial approximation of asin(x) when |
116 | // |x| <= 0.5: |
117 | // acos(x) ~ 2 * sqrt(u) * P(u). |
118 | // |
119 | // When -1 < x <= -0.5, we use the identity: |
120 | // acos(x) = pi - acos(-x) |
121 | // which is reduced to the postive case. |
122 | |
123 | xbits.set_sign(Sign::POS); |
124 | double xd = static_cast<double>(xbits.get_val()); |
125 | double u = fputil::multiply_add(-0.5, xd, 0.5); |
126 | double cv = 2 * fputil::sqrt<double>(u); |
127 | |
128 | double r3 = asin_eval(u); |
129 | double r = fputil::multiply_add(cv * u, r3, cv); |
130 | return static_cast<float>(x_sign ? M_MATH_PI - r : r); |
131 | } |
132 | |
133 | } // namespace LIBC_NAMESPACE_DECL |
134 | |