xref: /aosp_15_r20/external/llvm-libc/src/math/generic/asinf.cpp (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
1 //===-- Single-precision asin 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/asinf.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 #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA
19 
20 #include "inv_trigf_utils.h"
21 
22 namespace LIBC_NAMESPACE_DECL {
23 
24 static constexpr size_t N_EXCEPTS = 2;
25 
26 // Exceptional values when |x| <= 0.5
27 static constexpr fputil::ExceptValues<float, N_EXCEPTS> ASINF_EXCEPTS_LO = {{
28     // (inputs, RZ output, RU offset, RD offset, RN offset)
29     // x = 0x1.137f0cp-5, asinf(x) = 0x1.138c58p-5 (RZ)
30     {0x3d09bf86, 0x3d09c62c, 1, 0, 1},
31     // x = 0x1.cbf43cp-4, asinf(x) = 0x1.cced1cp-4 (RZ)
32     {0x3de5fa1e, 0x3de6768e, 1, 0, 0},
33 }};
34 
35 // Exceptional values when 0.5 < |x| <= 1
36 static constexpr fputil::ExceptValues<float, N_EXCEPTS> ASINF_EXCEPTS_HI = {{
37     // (inputs, RZ output, RU offset, RD offset, RN offset)
38     // x = 0x1.107434p-1, asinf(x) = 0x1.1f4b64p-1 (RZ)
39     {0x3f083a1a, 0x3f0fa5b2, 1, 0, 0},
40     // x = 0x1.ee836cp-1, asinf(x) = 0x1.4f0654p0 (RZ)
41     {0x3f7741b6, 0x3fa7832a, 1, 0, 0},
42 }};
43 
44 LLVM_LIBC_FUNCTION(float, asinf, (float x)) {
45   using FPBits = typename fputil::FPBits<float>;
46 
47   FPBits xbits(x);
48   uint32_t x_uint = xbits.uintval();
49   uint32_t x_abs = xbits.uintval() & 0x7fff'ffffU;
50   constexpr double SIGN[2] = {1.0, -1.0};
51   uint32_t x_sign = x_uint >> 31;
52 
53   // |x| <= 0.5-ish
54   if (x_abs < 0x3f04'471dU) {
55     // |x| < 0x1.d12edp-12
56     if (LIBC_UNLIKELY(x_abs < 0x39e8'9768U)) {
57       // When |x| < 2^-12, the relative error of the approximation asin(x) ~ x
58       // is:
59       //   |asin(x) - x| / |asin(x)| < |x^3| / (6|x|)
60       //                             = x^2 / 6
61       //                             < 2^-25
62       //                             < epsilon(1)/2.
63       // So the correctly rounded values of asin(x) are:
64       //   = x + sign(x)*eps(x) if rounding mode = FE_TOWARDZERO,
65       //                        or (rounding mode = FE_UPWARD and x is
66       //                        negative),
67       //   = x otherwise.
68       // To simplify the rounding decision and make it more efficient, we use
69       //   fma(x, 2^-25, x) instead.
70       // An exhaustive test shows that this formula work correctly for all
71       // rounding modes up to |x| < 0x1.d12edp-12.
72       // Note: to use the formula x + 2^-25*x to decide the correct rounding, we
73       // do need fma(x, 2^-25, x) to prevent underflow caused by 2^-25*x when
74       // |x| < 2^-125. For targets without FMA instructions, we simply use
75       // double for intermediate results as it is more efficient than using an
76       // emulated version of FMA.
77 #if defined(LIBC_TARGET_CPU_HAS_FMA)
78       return fputil::multiply_add(x, 0x1.0p-25f, x);
79 #else
80       double xd = static_cast<double>(x);
81       return static_cast<float>(fputil::multiply_add(xd, 0x1.0p-25, xd));
82 #endif // LIBC_TARGET_CPU_HAS_FMA
83     }
84 
85     // Check for exceptional values
86     if (auto r = ASINF_EXCEPTS_LO.lookup_odd(x_abs, x_sign);
87         LIBC_UNLIKELY(r.has_value()))
88       return r.value();
89 
90     // For |x| <= 0.5, we approximate asinf(x) by:
91     //   asin(x) = x * P(x^2)
92     // Where P(X^2) = Q(X) is a degree-20 minimax even polynomial approximating
93     // asin(x)/x on [0, 0.5] generated by Sollya with:
94     // > Q = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20|],
95     //                 [|1, D...|], [0, 0.5]);
96     // An exhaustive test shows that this approximation works well up to a
97     // little more than 0.5.
98     double xd = static_cast<double>(x);
99     double xsq = xd * xd;
100     double x3 = xd * xsq;
101     double r = asin_eval(xsq);
102     return static_cast<float>(fputil::multiply_add(x3, r, xd));
103   }
104 
105   // |x| > 1, return NaNs.
106   if (LIBC_UNLIKELY(x_abs > 0x3f80'0000U)) {
107     if (x_abs <= 0x7f80'0000U) {
108       fputil::set_errno_if_required(EDOM);
109       fputil::raise_except_if_required(FE_INVALID);
110     }
111     return FPBits::quiet_nan().get_val();
112   }
113 
114   // Check for exceptional values
115   if (auto r = ASINF_EXCEPTS_HI.lookup_odd(x_abs, x_sign);
116       LIBC_UNLIKELY(r.has_value()))
117     return r.value();
118 
119   // When |x| > 0.5, we perform range reduction as follow:
120   //
121   // Assume further that 0.5 < x <= 1, and let:
122   //   y = asin(x)
123   // We will use the double angle formula:
124   //   cos(2y) = 1 - 2 sin^2(y)
125   // and the complement angle identity:
126   //   x = sin(y) = cos(pi/2 - y)
127   //              = 1 - 2 sin^2 (pi/4 - y/2)
128   // So:
129   //   sin(pi/4 - y/2) = sqrt( (1 - x)/2 )
130   // And hence:
131   //   pi/4 - y/2 = asin( sqrt( (1 - x)/2 ) )
132   // Equivalently:
133   //   asin(x) = y = pi/2 - 2 * asin( sqrt( (1 - x)/2 ) )
134   // Let u = (1 - x)/2, then:
135   //   asin(x) = pi/2 - 2 * asin( sqrt(u) )
136   // Moreover, since 0.5 < x <= 1:
137   //   0 <= u < 1/4, and 0 <= sqrt(u) < 0.5,
138   // And hence we can reuse the same polynomial approximation of asin(x) when
139   // |x| <= 0.5:
140   //   asin(x) ~ pi/2 - 2 * sqrt(u) * P(u),
141 
142   xbits.set_sign(Sign::POS);
143   double sign = SIGN[x_sign];
144   double xd = static_cast<double>(xbits.get_val());
145   double u = fputil::multiply_add(-0.5, xd, 0.5);
146   double c1 = sign * (-2 * fputil::sqrt<double>(u));
147   double c2 = fputil::multiply_add(sign, M_MATH_PI_2, c1);
148   double c3 = c1 * u;
149 
150   double r = asin_eval(u);
151   return static_cast<float>(fputil::multiply_add(c3, r, c2));
152 }
153 
154 } // namespace LIBC_NAMESPACE_DECL
155