xref: /aosp_15_r20/external/llvm-libc/src/__support/FPUtil/generic/sqrt_80_bit_long_double.h (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
1 //===-- Square root of x86 long double numbers ------------------*- 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_FPUTIL_GENERIC_SQRT_80_BIT_LONG_DOUBLE_H
10 #define LLVM_LIBC_SRC___SUPPORT_FPUTIL_GENERIC_SQRT_80_BIT_LONG_DOUBLE_H
11 
12 #include "src/__support/CPP/bit.h"
13 #include "src/__support/FPUtil/FEnvImpl.h"
14 #include "src/__support/FPUtil/FPBits.h"
15 #include "src/__support/FPUtil/rounding_mode.h"
16 #include "src/__support/common.h"
17 #include "src/__support/macros/config.h"
18 #include "src/__support/uint128.h"
19 
20 namespace LIBC_NAMESPACE_DECL {
21 namespace fputil {
22 namespace x86 {
23 
normalize(int & exponent,FPBits<long double>::StorageType & mantissa)24 LIBC_INLINE void normalize(int &exponent,
25                            FPBits<long double>::StorageType &mantissa) {
26   const unsigned int shift = static_cast<unsigned int>(
27       cpp::countl_zero(static_cast<uint64_t>(mantissa)) -
28       (8 * sizeof(uint64_t) - 1 - FPBits<long double>::FRACTION_LEN));
29   exponent -= shift;
30   mantissa <<= shift;
31 }
32 
33 // if constexpr statement in sqrt.h still requires x86::sqrt to be declared
34 // even when it's not used.
35 LIBC_INLINE long double sqrt(long double x);
36 
37 // Correctly rounded SQRT for all rounding modes.
38 // Shift-and-add algorithm.
39 #if defined(LIBC_TYPES_LONG_DOUBLE_IS_X86_FLOAT80)
sqrt(long double x)40 LIBC_INLINE long double sqrt(long double x) {
41   using LDBits = FPBits<long double>;
42   using StorageType = typename LDBits::StorageType;
43   constexpr StorageType ONE = StorageType(1) << int(LDBits::FRACTION_LEN);
44   constexpr auto LDNAN = LDBits::quiet_nan().get_val();
45 
46   LDBits bits(x);
47 
48   if (bits == LDBits::inf(Sign::POS) || bits.is_zero() || bits.is_nan()) {
49     // sqrt(+Inf) = +Inf
50     // sqrt(+0) = +0
51     // sqrt(-0) = -0
52     // sqrt(NaN) = NaN
53     // sqrt(-NaN) = -NaN
54     return x;
55   } else if (bits.is_neg()) {
56     // sqrt(-Inf) = NaN
57     // sqrt(-x) = NaN
58     return LDNAN;
59   } else {
60     int x_exp = bits.get_explicit_exponent();
61     StorageType x_mant = bits.get_mantissa();
62 
63     // Step 1a: Normalize denormal input
64     if (bits.get_implicit_bit()) {
65       x_mant |= ONE;
66     } else if (bits.is_subnormal()) {
67       normalize(x_exp, x_mant);
68     }
69 
70     // Step 1b: Make sure the exponent is even.
71     if (x_exp & 1) {
72       --x_exp;
73       x_mant <<= 1;
74     }
75 
76     // After step 1b, x = 2^(x_exp) * x_mant, where x_exp is even, and
77     // 1 <= x_mant < 4.  So sqrt(x) = 2^(x_exp / 2) * y, with 1 <= y < 2.
78     // Notice that the output of sqrt is always in the normal range.
79     // To perform shift-and-add algorithm to find y, let denote:
80     //   y(n) = 1.y_1 y_2 ... y_n, we can define the nth residue to be:
81     //   r(n) = 2^n ( x_mant - y(n)^2 ).
82     // That leads to the following recurrence formula:
83     //   r(n) = 2*r(n-1) - y_n*[ 2*y(n-1) + 2^(-n-1) ]
84     // with the initial conditions: y(0) = 1, and r(0) = x - 1.
85     // So the nth digit y_n of the mantissa of sqrt(x) can be found by:
86     //   y_n = 1 if 2*r(n-1) >= 2*y(n - 1) + 2^(-n-1)
87     //         0 otherwise.
88     StorageType y = ONE;
89     StorageType r = x_mant - ONE;
90 
91     for (StorageType current_bit = ONE >> 1; current_bit; current_bit >>= 1) {
92       r <<= 1;
93       StorageType tmp = (y << 1) + current_bit; // 2*y(n - 1) + 2^(-n-1)
94       if (r >= tmp) {
95         r -= tmp;
96         y += current_bit;
97       }
98     }
99 
100     // We compute one more iteration in order to round correctly.
101     bool lsb = static_cast<bool>(y & 1); // Least significant bit
102     bool rb = false;                     // Round bit
103     r <<= 2;
104     StorageType tmp = (y << 2) + 1;
105     if (r >= tmp) {
106       r -= tmp;
107       rb = true;
108     }
109 
110     // Append the exponent field.
111     x_exp = ((x_exp >> 1) + LDBits::EXP_BIAS);
112     y |= (static_cast<StorageType>(x_exp) << (LDBits::FRACTION_LEN + 1));
113 
114     switch (quick_get_round()) {
115     case FE_TONEAREST:
116       // Round to nearest, ties to even
117       if (rb && (lsb || (r != 0)))
118         ++y;
119       break;
120     case FE_UPWARD:
121       if (rb || (r != 0))
122         ++y;
123       break;
124     }
125 
126     // Extract output
127     FPBits<long double> out(0.0L);
128     out.set_biased_exponent(x_exp);
129     out.set_implicit_bit(1);
130     out.set_mantissa((y & (ONE - 1)));
131 
132     return out.get_val();
133   }
134 }
135 #endif // LIBC_TYPES_LONG_DOUBLE_IS_X86_FLOAT80
136 
137 } // namespace x86
138 } // namespace fputil
139 } // namespace LIBC_NAMESPACE_DECL
140 
141 #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_GENERIC_SQRT_80_BIT_LONG_DOUBLE_H
142