| 1 | //===-- Implementation header for acosf -------------------------*- C++ -*-===// |
| 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 | #ifndef LLVM_LIBC_SRC___SUPPORT_MATH_ACOSF_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_ACOSF_H |
| 11 | |
| 12 | #include "inv_trigf_utils.h" |
| 13 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 14 | #include "src/__support/FPUtil/FPBits.h" |
| 15 | #include "src/__support/FPUtil/except_value_utils.h" |
| 16 | #include "src/__support/FPUtil/multiply_add.h" |
| 17 | #include "src/__support/FPUtil/sqrt.h" |
| 18 | #include "src/__support/macros/config.h" |
| 19 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
| 20 | |
| 21 | namespace LIBC_NAMESPACE_DECL { |
| 22 | |
| 23 | namespace math { |
| 24 | |
| 25 | namespace acosf_internal { |
| 26 | |
| 27 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 28 | |
| 29 | LIBC_INLINE_VAR constexpr size_t N_EXCEPTS = 4; |
| 30 | |
| 31 | // Exceptional values when |x| <= 0.5 |
| 32 | LIBC_INLINE_VAR constexpr fputil::ExceptValues<float, N_EXCEPTS> ACOSF_EXCEPTS = |
| 33 | {.values: { |
| 34 | // (inputs, RZ output, RU offset, RD offset, RN offset) |
| 35 | // x = 0x1.110b46p-26, acosf(x) = 0x1.921fb4p0 (RZ) |
| 36 | {.input: 0x328885a3, .rnd_towardzero_result: 0x3fc90fda, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
| 37 | // x = -0x1.110b46p-26, acosf(x) = 0x1.921fb4p0 (RZ) |
| 38 | {.input: 0xb28885a3, .rnd_towardzero_result: 0x3fc90fda, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
| 39 | // x = 0x1.04c444p-12, acosf(x) = 0x1.920f68p0 (RZ) |
| 40 | {.input: 0x39826222, .rnd_towardzero_result: 0x3fc907b4, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
| 41 | // x = -0x1.04c444p-12, acosf(x) = 0x1.923p0 (RZ) |
| 42 | {.input: 0xb9826222, .rnd_towardzero_result: 0x3fc91800, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
| 43 | }}; |
| 44 | |
| 45 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 46 | |
| 47 | } // namespace acosf_internal |
| 48 | |
| 49 | LIBC_INLINE float acosf(float x) { |
| 50 | using namespace acosf_internal; |
| 51 | using namespace inv_trigf_utils_internal; |
| 52 | using FPBits = typename fputil::FPBits<float>; |
| 53 | |
| 54 | FPBits xbits(x); |
| 55 | uint32_t x_uint = xbits.uintval(); |
| 56 | uint32_t x_abs = xbits.uintval() & 0x7fff'ffffU; |
| 57 | uint32_t x_sign = x_uint >> 31; |
| 58 | |
| 59 | // |x| <= 0.5 |
| 60 | if (LIBC_UNLIKELY(x_abs <= 0x3f00'0000U)) { |
| 61 | // |x| < 0x1p-10 |
| 62 | if (LIBC_UNLIKELY(x_abs < 0x3a80'0000U)) { |
| 63 | // When |x| < 2^-10, we use the following approximation: |
| 64 | // acos(x) = pi/2 - asin(x) |
| 65 | // ~ pi/2 - x - x^3 / 6 |
| 66 | |
| 67 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 68 | // Check for exceptional values |
| 69 | if (auto r = ACOSF_EXCEPTS.lookup(x_bits: x_uint); LIBC_UNLIKELY(r.has_value())) |
| 70 | return r.value(); |
| 71 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 72 | |
| 73 | double xd = static_cast<double>(x); |
| 74 | return static_cast<float>(fputil::multiply_add( |
| 75 | x: -0x1.5555555555555p-3 * xd, y: xd * xd, z: M_MATH_PI_2 - xd)); |
| 76 | } |
| 77 | |
| 78 | // For |x| <= 0.5, we approximate acosf(x) by: |
| 79 | // acos(x) = pi/2 - asin(x) = pi/2 - x * P(x^2) |
| 80 | // Where P(X^2) = Q(X) is a degree-24 minimax even polynomial approximating |
| 81 | // asin(x)/x on [0, 0.5] generated by Sollya with: |
| 82 | // > Q = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, |
| 83 | // 22, 24|], [|1, D...|], [0, 0.5]); |
| 84 | double xd = static_cast<double>(x); |
| 85 | double xsq = xd * xd; |
| 86 | double x3 = xd * xsq; |
| 87 | double r = asin_eval(xsq); |
| 88 | return static_cast<float>(fputil::multiply_add(x: -x3, y: r, z: M_MATH_PI_2 - xd)); |
| 89 | } |
| 90 | |
| 91 | // |x| >= 1, return 0, 2pi, or NaNs. |
| 92 | if (LIBC_UNLIKELY(x_abs >= 0x3f80'0000U)) { |
| 93 | if (x_abs == 0x3f80'0000U) |
| 94 | return x_sign ? /* x == -1.0f */ fputil::round_result_slightly_down( |
| 95 | value_rn: 0x1.921fb6p+1f) |
| 96 | : /* x == 1.0f */ 0.0f; |
| 97 | |
| 98 | if (xbits.is_signaling_nan()) { |
| 99 | fputil::raise_except_if_required(FE_INVALID); |
| 100 | return FPBits::quiet_nan().get_val(); |
| 101 | } |
| 102 | |
| 103 | // |x| <= +/-inf |
| 104 | if (x_abs <= 0x7f80'0000U) { |
| 105 | fputil::set_errno_if_required(EDOM); |
| 106 | fputil::raise_except_if_required(FE_INVALID); |
| 107 | } |
| 108 | |
| 109 | return x + FPBits::quiet_nan().get_val(); |
| 110 | } |
| 111 | |
| 112 | // When 0.5 < |x| < 1, we perform range reduction as follow: |
| 113 | // |
| 114 | // Assume further that 0.5 < x <= 1, and let: |
| 115 | // y = acos(x) |
| 116 | // We use the double angle formula: |
| 117 | // x = cos(y) = 1 - 2 sin^2(y/2) |
| 118 | // So: |
| 119 | // sin(y/2) = sqrt( (1 - x)/2 ) |
| 120 | // And hence: |
| 121 | // y = 2 * asin( sqrt( (1 - x)/2 ) ) |
| 122 | // Let u = (1 - x)/2, then |
| 123 | // acos(x) = 2 * asin( sqrt(u) ) |
| 124 | // Moreover, since 0.5 < x <= 1, |
| 125 | // 0 <= u < 1/4, and 0 <= sqrt(u) < 0.5, |
| 126 | // And hence we can reuse the same polynomial approximation of asin(x) when |
| 127 | // |x| <= 0.5: |
| 128 | // acos(x) ~ 2 * sqrt(u) * P(u). |
| 129 | // |
| 130 | // When -1 < x <= -0.5, we use the identity: |
| 131 | // acos(x) = pi - acos(-x) |
| 132 | // which is reduced to the postive case. |
| 133 | |
| 134 | xbits.set_sign(Sign::POS); |
| 135 | double xd = static_cast<double>(xbits.get_val()); |
| 136 | double u = fputil::multiply_add(x: -0.5, y: xd, z: 0.5); |
| 137 | double cv = 2 * fputil::sqrt<double>(x: u); |
| 138 | |
| 139 | double r3 = asin_eval(xsq: u); |
| 140 | double r = fputil::multiply_add(x: cv * u, y: r3, z: cv); |
| 141 | return static_cast<float>(x_sign ? M_MATH_PI - r : r); |
| 142 | } |
| 143 | |
| 144 | } // namespace math |
| 145 | |
| 146 | } // namespace LIBC_NAMESPACE_DECL |
| 147 | |
| 148 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_ACOSF_H |
| 149 | |