1 //===-- Single-precision 2^x function -------------------------------------===//
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
7 //===----------------------------------------------------------------------===//
9 #include "src/math/exp2f.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/multiply_add.h"
14 #include "src/__support/FPUtil/nearest_integer.h"
15 #include "src/__support/common.h"
16 #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY
22 namespace __llvm_libc
{
24 constexpr uint32_t EXVAL1
= 0x3b42'9d37U
;
25 constexpr uint32_t EXVAL2
= 0xbcf3'a937U
;
26 constexpr uint32_t EXVAL_MASK
= EXVAL1
& EXVAL2
;
28 LLVM_LIBC_FUNCTION(float, exp2f
, (float x
)) {
29 using FPBits
= typename
fputil::FPBits
<float>;
32 uint32_t x_u
= xbits
.uintval();
33 uint32_t x_abs
= x_u
& 0x7fff'ffffU
;
36 if (LIBC_UNLIKELY(x_abs
<= 0x3280'0000U
)) {
40 // // When |x| >= 128, or x is nan
41 if (LIBC_UNLIKELY(x_abs
>= 0x4300'0000U
)) {
44 if (!xbits
.get_sign()) {
46 if (x_u
< 0x7f80'0000U
) {
47 int rounding
= fputil::get_round();
48 if (rounding
== FE_DOWNWARD
|| rounding
== FE_TOWARDZERO
)
49 return static_cast<float>(FPBits(FPBits::MAX_NORMAL
));
51 fputil::set_errno_if_required(ERANGE
);
52 fputil::raise_except_if_required(FE_OVERFLOW
);
55 return x
+ FPBits::inf().get_val();
58 if (x_u
>= 0xc316'0000U
) {
65 if (fputil::get_round() == FE_UPWARD
)
66 return FPBits(FPBits::MIN_SUBNORMAL
).get_val();
68 fputil::set_errno_if_required(ERANGE
);
69 fputil::raise_except_if_required(FE_UNDERFLOW
);
75 // Check exceptional values.
76 if (LIBC_UNLIKELY((x_u
& EXVAL_MASK
) == EXVAL_MASK
)) {
77 if (LIBC_UNLIKELY(x_u
== EXVAL1
)) { // x = 0x1.853a6ep-9f
78 if (fputil::get_round() == FE_TONEAREST
)
79 return 0x1.00870ap
+0f
;
80 } else if (LIBC_UNLIKELY(x_u
== EXVAL2
)) { // x = -0x1.e7526ep-6f
81 if (fputil::get_round() == FE_TONEAREST
)
82 return 0x1.f58d62p
-1f
;
86 // For -150 < x < 128, to compute 2^x, we perform the following range
87 // reduction: find hi, mid, lo such that:
88 // x = hi + mid + lo, in which
90 // 0 <= mid * 2^5 < 32 is an integer
91 // -2^(-6) <= lo <= 2^-6.
93 // hi + mid = round(x * 2^5) * 2^(-5).
95 // 2^x = 2^(hi + mid + lo) = 2^hi * 2^mid * 2^lo.
96 // 2^mid is stored in the lookup table of 32 elements.
97 // 2^lo is computed using a degree-5 minimax polynomial
98 // generated by Sollya.
99 // We perform 2^hi * 2^mid by simply add hi to the exponent field
102 // kf = (hi + mid) * 2^5 = round(x * 2^5)
103 float kf
= fputil::nearest_integer(x
* 32.0f
);
104 // dx = lo = x - (hi + mid) = x - kf * 2^(-5)
105 double dx
= fputil::multiply_add(-0x1.0p
-5f
, kf
, x
);
107 int k
= static_cast<int>(kf
);
108 // hi = floor(kf * 2^(-4))
109 // exp_hi = shift hi to the exponent field of double precision.
110 int64_t exp_hi
= static_cast<int64_t>(k
>> ExpBase::MID_BITS
)
111 << fputil::FloatProperties
<double>::MANTISSA_WIDTH
;
113 // mh_bits = bit field of mh
114 int64_t mh_bits
= ExpBase::EXP_2_MID
[k
& ExpBase::MID_MASK
] + exp_hi
;
115 double mh
= fputil::FPBits
<double>(uint64_t(mh_bits
)).get_val();
117 // Degree-5 polynomial approximating (2^x - 1)/x generating by Sollya with:
118 // > P = fpminimax((2^x - 1)/x, 5, [|D...|], [-1/32. 1/32]);
119 constexpr double COEFFS
[5] = {0x1.62e42fefa39efp
-1, 0x1.ebfbdff8131c4p
-3,
120 0x1.c6b08d7061695p
-5, 0x1.3b2b1bee74b2ap
-7,
121 0x1.5d88091198529p
-10};
122 double dx_sq
= dx
* dx
;
123 double c1
= fputil::multiply_add(dx
, COEFFS
[0], 1.0);
124 double c2
= fputil::multiply_add(dx
, COEFFS
[2], COEFFS
[1]);
125 double c3
= fputil::multiply_add(dx
, COEFFS
[4], COEFFS
[3]);
126 double p
= fputil::multiply_add(dx_sq
, c3
, c2
);
127 // 2^x = 2^(hi + mid + lo)
128 // = 2^(hi + mid) * 2^lo
129 // ~ mh * (1 + lo * P(lo))
130 // = mh + (mh*lo) * P(lo)
131 return static_cast<float>(fputil::multiply_add(p
, dx_sq
* mh
, c1
* mh
));
134 } // namespace __llvm_libc