xref: /aosp_15_r20/external/llvm-libc/src/math/generic/acosf.cpp (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
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 static constexpr size_t N_EXCEPTS = 4;
24 
25 // Exceptional values when |x| <= 0.5
26 static constexpr fputil::ExceptValues<float, N_EXCEPTS> ACOSF_EXCEPTS = {{
27     // (inputs, RZ output, RU offset, RD offset, RN offset)
28     // x = 0x1.110b46p-26, acosf(x) = 0x1.921fb4p0 (RZ)
29     {0x328885a3, 0x3fc90fda, 1, 0, 1},
30     // x = -0x1.110b46p-26, acosf(x) = 0x1.921fb4p0 (RZ)
31     {0xb28885a3, 0x3fc90fda, 1, 0, 1},
32     // x = 0x1.04c444p-12, acosf(x) = 0x1.920f68p0 (RZ)
33     {0x39826222, 0x3fc907b4, 1, 0, 1},
34     // x = -0x1.04c444p-12, acosf(x) = 0x1.923p0 (RZ)
35     {0xb9826222, 0x3fc91800, 1, 0, 1},
36 }};
37 
38 LLVM_LIBC_FUNCTION(float, acosf, (float x)) {
39   using FPBits = typename fputil::FPBits<float>;
40 
41   FPBits xbits(x);
42   uint32_t x_uint = xbits.uintval();
43   uint32_t x_abs = xbits.uintval() & 0x7fff'ffffU;
44   uint32_t x_sign = x_uint >> 31;
45 
46   // |x| <= 0.5
47   if (LIBC_UNLIKELY(x_abs <= 0x3f00'0000U)) {
48     // |x| < 0x1p-10
49     if (LIBC_UNLIKELY(x_abs < 0x3a80'0000U)) {
50       // When |x| < 2^-10, we use the following approximation:
51       //   acos(x) = pi/2 - asin(x)
52       //           ~ pi/2 - x - x^3 / 6
53 
54       // Check for exceptional values
55       if (auto r = ACOSF_EXCEPTS.lookup(x_uint); LIBC_UNLIKELY(r.has_value()))
56         return r.value();
57 
58       double xd = static_cast<double>(x);
59       return static_cast<float>(fputil::multiply_add(
60           -0x1.5555555555555p-3 * xd, xd * xd, M_MATH_PI_2 - xd));
61     }
62 
63     // For |x| <= 0.5, we approximate acosf(x) by:
64     //   acos(x) = pi/2 - asin(x) = pi/2 - x * P(x^2)
65     // Where P(X^2) = Q(X) is a degree-20 minimax even polynomial approximating
66     // asin(x)/x on [0, 0.5] generated by Sollya with:
67     // > Q = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20|],
68     //                 [|1, D...|], [0, 0.5]);
69     double xd = static_cast<double>(x);
70     double xsq = xd * xd;
71     double x3 = xd * xsq;
72     double r = asin_eval(xsq);
73     return static_cast<float>(fputil::multiply_add(-x3, r, M_MATH_PI_2 - xd));
74   }
75 
76   // |x| >= 1, return 0, 2pi, or NaNs.
77   if (LIBC_UNLIKELY(x_abs >= 0x3f80'0000U)) {
78     if (x_abs == 0x3f80'0000U)
79       return x_sign ? /* x == -1.0f */ fputil::round_result_slightly_down(
80                           0x1.921fb6p+1f)
81                     : /* x == 1.0f */ 0.0f;
82 
83     if (x_abs <= 0x7f80'0000U) {
84       fputil::set_errno_if_required(EDOM);
85       fputil::raise_except_if_required(FE_INVALID);
86     }
87     return x + FPBits::quiet_nan().get_val();
88   }
89 
90   // When 0.5 < |x| < 1, we perform range reduction as follow:
91   //
92   // Assume further that 0.5 < x <= 1, and let:
93   //   y = acos(x)
94   // We use the double angle formula:
95   //   x = cos(y) = 1 - 2 sin^2(y/2)
96   // So:
97   //   sin(y/2) = sqrt( (1 - x)/2 )
98   // And hence:
99   //   y = 2 * asin( sqrt( (1 - x)/2 ) )
100   // Let u = (1 - x)/2, then
101   //   acos(x) = 2 * asin( sqrt(u) )
102   // Moreover, since 0.5 < x <= 1,
103   //   0 <= u < 1/4, and 0 <= sqrt(u) < 0.5,
104   // And hence we can reuse the same polynomial approximation of asin(x) when
105   // |x| <= 0.5:
106   //   acos(x) ~ 2 * sqrt(u) * P(u).
107   //
108   // When -1 < x <= -0.5, we use the identity:
109   //   acos(x) = pi - acos(-x)
110   // which is reduced to the postive case.
111 
112   xbits.set_sign(Sign::POS);
113   double xd = static_cast<double>(xbits.get_val());
114   double u = fputil::multiply_add(-0.5, xd, 0.5);
115   double cv = 2 * fputil::sqrt<double>(u);
116 
117   double r3 = asin_eval(u);
118   double r = fputil::multiply_add(cv * u, r3, cv);
119   return static_cast<float>(x_sign ? M_MATH_PI - r : r);
120 }
121 
122 } // namespace LIBC_NAMESPACE_DECL
123