1 //===-- Single-precision e^x - 1 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/expm1f.h"
10 #include "common_constants.h" // Lookup tables EXP_M1 and EXP_M2.
11 #include "src/__support/FPUtil/BasicOperations.h"
12 #include "src/__support/FPUtil/FEnvImpl.h"
13 #include "src/__support/FPUtil/FMA.h"
14 #include "src/__support/FPUtil/FPBits.h"
15 #include "src/__support/FPUtil/PolyEval.h"
16 #include "src/__support/common.h"
17 
18 #include <errno.h>
19 
20 namespace __llvm_libc {
21 
22 LLVM_LIBC_FUNCTION(float, expm1f, (float x)) {
23   using FPBits = typename fputil::FPBits<float>;
24   FPBits xbits(x);
25 
26   uint32_t x_u = xbits.uintval();
27   uint32_t x_abs = x_u & 0x7fff'ffffU;
28 
29   // Exceptional value
30   if (unlikely(x_u == 0x3e35'bec5U)) { // x = 0x1.6b7d8ap-3f
31     int round_mode = fputil::get_round();
32     if (round_mode == FE_TONEAREST || round_mode == FE_UPWARD)
33       return 0x1.8dbe64p-3f;
34     return 0x1.8dbe62p-3f;
35   }
36 
37 #if !defined(LIBC_TARGET_HAS_FMA)
38   if (unlikely(x_u == 0xbdc1'c6cbU)) { // x = -0x1.838d96p-4f
39     int round_mode = fputil::get_round();
40     if (round_mode == FE_TONEAREST || round_mode == FE_DOWNWARD)
41       return -0x1.71c884p-4f;
42     return -0x1.71c882p-4f;
43   }
44 #endif // LIBC_TARGET_HAS_FMA
45 
46   // When |x| > 25*log(2), or nan
47   if (unlikely(x_abs >= 0x418a'a123U)) {
48     // x < log(2^-25)
49     if (xbits.get_sign()) {
50       // exp(-Inf) = 0
51       if (xbits.is_inf())
52         return -1.0f;
53       // exp(nan) = nan
54       if (xbits.is_nan())
55         return x;
56       int round_mode = fputil::get_round();
57       if (round_mode == FE_UPWARD || round_mode == FE_TOWARDZERO)
58         return -0x1.ffff'fep-1f; // -1.0f + 0x1.0p-24f
59       return -1.0f;
60     } else {
61       // x >= 89 or nan
62       if (xbits.uintval() >= 0x42b2'0000) {
63         if (xbits.uintval() < 0x7f80'0000U) {
64           int rounding = fputil::get_round();
65           if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO)
66             return static_cast<float>(FPBits(FPBits::MAX_NORMAL));
67 
68           errno = ERANGE;
69         }
70         return x + static_cast<float>(FPBits::inf());
71       }
72     }
73   }
74 
75   // |x| < 2^-4
76   if (x_abs < 0x3d80'0000U) {
77     // |x| < 2^-25
78     if (x_abs < 0x3300'0000U) {
79       // x = -0.0f
80       if (unlikely(xbits.uintval() == 0x8000'0000U))
81         return x;
82         // When |x| < 2^-25, the relative error of the approximation e^x - 1 ~ x
83         // is:
84         //   |(e^x - 1) - x| / |e^x - 1| < |x^2| / |x|
85         //                               = |x|
86         //                               < 2^-25
87         //                               < epsilon(1)/2.
88         // So the correctly rounded values of expm1(x) are:
89         //   = x + eps(x) if rounding mode = FE_UPWARD,
90         //                   or (rounding mode = FE_TOWARDZERO and x is
91         //                   negative),
92         //   = x otherwise.
93         // To simplify the rounding decision and make it more efficient, we use
94         //   fma(x, x, x) ~ x + x^2 instead.
95         // Note: to use the formula x + x^2 to decide the correct rounding, we
96         // do need fma(x, x, x) to prevent underflow caused by x*x when |x| <
97         // 2^-76. For targets without FMA instructions, we simply use double for
98         // intermediate results as it is more efficient than using an emulated
99         // version of FMA.
100 #if defined(LIBC_TARGET_HAS_FMA)
101       return fputil::fma(x, x, x);
102 #else
103       double xd = x;
104       return static_cast<float>(fputil::multiply_add(xd, xd, xd));
105 #endif // LIBC_TARGET_HAS_FMA
106     }
107 
108     // 2^-25 <= |x| < 2^-4
109     double xd = static_cast<double>(x);
110     double xsq = xd * xd;
111     // Degree-8 minimax polynomial generated by Sollya with:
112     // > display = hexadecimal;
113     // > P = fpminimax((expm1(x) - x)/x^2, 6, [|D...|], [-2^-4, 2^-4]);
114     double r =
115         fputil::polyeval(xd, 0x1p-1, 0x1.55555555557ddp-3, 0x1.55555555552fap-5,
116                          0x1.111110fcd58b7p-7, 0x1.6c16c1717660bp-10,
117                          0x1.a0241f0006d62p-13, 0x1.a01e3f8d3c06p-16);
118     return static_cast<float>(fputil::multiply_add(r, xsq, xd));
119   }
120 
121   // For -18 < x < 89, to compute expm1(x), we perform the following range
122   // reduction: find hi, mid, lo such that:
123   //   x = hi + mid + lo, in which
124   //     hi is an integer,
125   //     mid * 2^7 is an integer
126   //     -2^(-8) <= lo < 2^-8.
127   // In particular,
128   //   hi + mid = round(x * 2^7) * 2^(-7).
129   // Then,
130   //   expm1(x) = exp(hi + mid + lo) - 1 = exp(hi) * exp(mid) * exp(lo) - 1.
131   // We store exp(hi) and exp(mid) in the lookup tables EXP_M1 and EXP_M2
132   // respectively.  exp(lo) is computed using a degree-4 minimax polynomial
133   // generated by Sollya.
134 
135   // x_hi = hi + mid.
136   int x_hi = static_cast<int>(x * 0x1.0p7f + (xbits.get_sign() ? -0.5f : 0.5f));
137   // Subtract (hi + mid) from x to get lo.
138   x -= static_cast<float>(x_hi) * 0x1.0p-7f;
139   double xd = static_cast<double>(x);
140   x_hi += 104 << 7;
141   // hi = x_hi >> 7
142   double exp_hi = EXP_M1[x_hi >> 7];
143   // lo = x_hi & 0x0000'007fU;
144   double exp_mid = EXP_M2[x_hi & 0x7f];
145   double exp_hi_mid = exp_hi * exp_mid;
146   // Degree-4 minimax polynomial generated by Sollya with the following
147   // commands:
148   //   > display = hexadecimal;
149   //   > Q = fpminimax(expm1(x)/x, 3, [|D...|], [-2^-8, 2^-8]);
150   //   > Q;
151   double exp_lo =
152       fputil::polyeval(xd, 0x1.0p0, 0x1.ffffffffff777p-1, 0x1.000000000071cp-1,
153                        0x1.555566668e5e7p-3, 0x1.55555555ef243p-5);
154   return static_cast<float>(fputil::multiply_add(exp_hi_mid, exp_lo, -1.0));
155 }
156 
157 } // namespace __llvm_libc
158