1 /*
2  * Double-precision log(x) function.
3  *
4  * Copyright (c) 2018, Arm Limited.
5  * SPDX-License-Identifier: MIT
6  */
7 
8 #include <math.h>
9 #include <stdint.h>
10 #include "libm.h"
11 #include "log_data.h"
12 
13 #define T __log_data.tab
14 #define T2 __log_data.tab2
15 #define B __log_data.poly1
16 #define A __log_data.poly
17 #define Ln2hi __log_data.ln2hi
18 #define Ln2lo __log_data.ln2lo
19 #define N (1 << LOG_TABLE_BITS)
20 #define OFF 0x3fe6000000000000
21 
22 /* Top 16 bits of a double.  */
top16(double x)23 static inline uint32_t top16(double x)
24 {
25 	return asuint64(x) >> 48;
26 }
27 
log(double x)28 double log(double x)
29 {
30 	double_t w, z, r, r2, r3, y, invc, logc, kd, hi, lo;
31 	uint64_t ix, iz, tmp;
32 	uint32_t top;
33 	int k, i;
34 
35 	ix = asuint64(x);
36 	top = top16(x);
37 #define LO asuint64(1.0 - 0x1p-4)
38 #define HI asuint64(1.0 + 0x1.09p-4)
39 	if (predict_false(ix - LO < HI - LO)) {
40 		/* Handle close to 1.0 inputs separately.  */
41 		/* Fix sign of zero with downward rounding when x==1.  */
42 		if (WANT_ROUNDING && predict_false(ix == asuint64(1.0)))
43 			return 0;
44 		r = x - 1.0;
45 		r2 = r * r;
46 		r3 = r * r2;
47 		y = r3 *
48 		    (B[1] + r * B[2] + r2 * B[3] +
49 		     r3 * (B[4] + r * B[5] + r2 * B[6] +
50 			   r3 * (B[7] + r * B[8] + r2 * B[9] + r3 * B[10])));
51 		/* Worst-case error is around 0.507 ULP.  */
52 		w = r * 0x1p27;
53 		double_t rhi = r + w - w;
54 		double_t rlo = r - rhi;
55 		w = rhi * rhi * B[0]; /* B[0] == -0.5.  */
56 		hi = r + w;
57 		lo = r - hi + w;
58 		lo += B[0] * rlo * (rhi + r);
59 		y += lo;
60 		y += hi;
61 		return eval_as_double(y);
62 	}
63 	if (predict_false(top - 0x0010 >= 0x7ff0 - 0x0010)) {
64 		/* x < 0x1p-1022 or inf or nan.  */
65 		if (ix * 2 == 0)
66 			return __math_divzero(1);
67 		if (ix == asuint64(INFINITY)) /* log(inf) == inf.  */
68 			return x;
69 		if ((top & 0x8000) || (top & 0x7ff0) == 0x7ff0)
70 			return __math_invalid(x);
71 		/* x is subnormal, normalize it.  */
72 		ix = asuint64(x * 0x1p52);
73 		ix -= 52ULL << 52;
74 	}
75 
76 	/* x = 2^k z; where z is in range [OFF,2*OFF) and exact.
77 	   The range is split into N subintervals.
78 	   The ith subinterval contains z and c is near its center.  */
79 	tmp = ix - OFF;
80 	i = (tmp >> (52 - LOG_TABLE_BITS)) % N;
81 	k = (int64_t)tmp >> 52; /* arithmetic shift */
82 	iz = ix - (tmp & 0xfffULL << 52);
83 	invc = T[i].invc;
84 	logc = T[i].logc;
85 	z = asdouble(iz);
86 
87 	/* log(x) = log1p(z/c-1) + log(c) + k*Ln2.  */
88 	/* r ~= z/c - 1, |r| < 1/(2*N).  */
89 #if __FP_FAST_FMA
90 	/* rounding error: 0x1p-55/N.  */
91 	r = __builtin_fma(z, invc, -1.0);
92 #else
93 	/* rounding error: 0x1p-55/N + 0x1p-66.  */
94 	r = (z - T2[i].chi - T2[i].clo) * invc;
95 #endif
96 	kd = (double_t)k;
97 
98 	/* hi + lo = r + log(c) + k*Ln2.  */
99 	w = kd * Ln2hi + logc;
100 	hi = w + r;
101 	lo = w - hi + r + kd * Ln2lo;
102 
103 	/* log(x) = lo + (log1p(r) - r) + hi.  */
104 	r2 = r * r; /* rounding error: 0x1p-54/N^2.  */
105 	/* Worst case error if |y| > 0x1p-5:
106 	   0.5 + 4.13/N + abs-poly-error*2^57 ULP (+ 0.002 ULP without fma)
107 	   Worst case error if |y| > 0x1p-4:
108 	   0.5 + 2.06/N + abs-poly-error*2^56 ULP (+ 0.001 ULP without fma).  */
109 	y = lo + r2 * A[0] +
110 	    r * r2 * (A[1] + r * A[2] + r2 * (A[3] + r * A[4])) + hi;
111 	return eval_as_double(y);
112 }
113