1 /* Double-precision x^y function.
2    Copyright (c) 2018 Arm Ltd.  All rights reserved.
3 
4    SPDX-License-Identifier: BSD-3-Clause
5 
6    Redistribution and use in source and binary forms, with or without
7    modification, are permitted provided that the following conditions
8    are met:
9    1. Redistributions of source code must retain the above copyright
10       notice, this list of conditions and the following disclaimer.
11    2. Redistributions in binary form must reproduce the above copyright
12       notice, this list of conditions and the following disclaimer in the
13       documentation and/or other materials provided with the distribution.
14    3. The name of the company may not be used to endorse or promote
15       products derived from this software without specific prior written
16       permission.
17 
18    THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
19    WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
20    MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21    IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
22    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
23    TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
24    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
25    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
26    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
27    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */
28 
29 #include "fdlibm.h"
30 #if !__OBSOLETE_MATH_DOUBLE
31 
32 #include <math.h>
33 #include <stdint.h>
34 #include "math_config.h"
35 
36 /*
37 Worst-case error: 0.54 ULP (~= ulperr_exp + 1024*Ln2*relerr_log*2^53)
38 relerr_log: 1.3 * 2^-68 (Relative error of log, 1.5 * 2^-68 without fma)
39 ulperr_exp: 0.509 ULP (ULP error of exp, 0.511 ULP without fma)
40 */
41 
42 #define T __pow_log_data.tab
43 #define A __pow_log_data.poly
44 #define Ln2hi __pow_log_data.ln2hi
45 #define Ln2lo __pow_log_data.ln2lo
46 #define N (1 << POW_LOG_TABLE_BITS)
47 #define OFF 0x3fe6955500000000
48 
49 /* Top 12 bits of a double (sign and exponent bits).  */
50 static inline uint32_t
top12(double x)51 top12 (double x)
52 {
53   return asuint64 (x) >> 52;
54 }
55 
56 /* Compute y+TAIL = log(x) where the rounded result is y and TAIL has about
57    additional 15 bits precision.  IX is the bit representation of x, but
58    normalized in the subnormal range using the sign bit for the exponent.  */
59 static inline double_t
log_inline(uint64_t ix,double_t * tail)60 log_inline (uint64_t ix, double_t *tail)
61 {
62   /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
63   double_t z, r, y, invc, logc, logctail, kd, hi, t1, t2, lo, lo1, lo2, p;
64   uint64_t iz, tmp;
65   int k, i;
66 
67   /* x = 2^k z; where z is in range [OFF,2*OFF) and exact.
68      The range is split into N subintervals.
69      The ith subinterval contains z and c is near its center.  */
70   tmp = ix - OFF;
71   i = (tmp >> (52 - POW_LOG_TABLE_BITS)) % N;
72   k = (int64_t) tmp >> 52; /* arithmetic shift */
73   iz = ix - (tmp & 0xfffULL << 52);
74   z = asdouble (iz);
75   kd = (double_t) k;
76 
77   /* log(x) = k*Ln2 + log(c) + log1p(z/c-1).  */
78   invc = T[i].invc;
79   logc = T[i].logc;
80   logctail = T[i].logctail;
81 
82   /* Note: 1/c is j/N or j/N/2 where j is an integer in [N,2N) and
83      |z/c - 1| < 1/N, so r = z/c - 1 is exactly representible.  */
84 #if _HAVE_FAST_FMA
85   r = fma (z, invc, -1.0);
86 #else
87   /* Split z such that rhi, rlo and rhi*rhi are exact and |rlo| <= |r|.  */
88   double_t zhi = asdouble ((iz + (1ULL << 31)) & (-1ULL << 32));
89   double_t zlo = z - zhi;
90   double_t rhi = zhi * invc - 1.0;
91   double_t rlo = zlo * invc;
92   r = rhi + rlo;
93 #endif
94 
95   /* k*Ln2 + log(c) + r.  */
96   t1 = kd * Ln2hi + logc;
97   t2 = t1 + r;
98   lo1 = kd * Ln2lo + logctail;
99   lo2 = t1 - t2 + r;
100 
101   /* Evaluation is optimized assuming superscalar pipelined execution.  */
102   double_t ar, ar2, ar3, lo3, lo4;
103   ar = A[0] * r; /* A[0] = -0.5.  */
104   ar2 = r * ar;
105   ar3 = r * ar2;
106   /* k*Ln2 + log(c) + r + A[0]*r*r.  */
107 #if _HAVE_FAST_FMA
108   hi = t2 + ar2;
109   lo3 = fma (ar, r, -ar2);
110   lo4 = t2 - hi + ar2;
111 #else
112   double_t arhi = A[0] * rhi;
113   double_t arhi2 = rhi * arhi;
114   hi = t2 + arhi2;
115   lo3 = rlo * (ar + arhi);
116   lo4 = t2 - hi + arhi2;
117 #endif
118   /* p = log1p(r) - r - A[0]*r*r.  */
119 #if POW_LOG_POLY_ORDER == 8
120   p = (ar3
121        * (A[1] + r * A[2] + ar2 * (A[3] + r * A[4] + ar2 * (A[5] + r * A[6]))));
122 #endif
123   lo = lo1 + lo2 + lo3 + lo4 + p;
124   y = hi + lo;
125   *tail = hi - y + lo;
126   return y;
127 }
128 
129 #undef N
130 #undef T
131 #define N (1 << EXP_TABLE_BITS)
132 #define InvLn2N __exp_data.invln2N
133 #define NegLn2hiN __exp_data.negln2hiN
134 #define NegLn2loN __exp_data.negln2loN
135 #define Shift __exp_data.shift
136 #define T __exp_data.tab
137 #define C2 __exp_data.poly[5 - EXP_POLY_ORDER]
138 #define C3 __exp_data.poly[6 - EXP_POLY_ORDER]
139 #define C4 __exp_data.poly[7 - EXP_POLY_ORDER]
140 #define C5 __exp_data.poly[8 - EXP_POLY_ORDER]
141 #define C6 __exp_data.poly[9 - EXP_POLY_ORDER]
142 
143 /* Handle cases that may overflow or underflow when computing the result that
144    is scale*(1+TMP) without intermediate rounding.  The bit representation of
145    scale is in SBITS, however it has a computed exponent that may have
146    overflown into the sign bit so that needs to be adjusted before using it as
147    a double.  (int32_t)KI is the k used in the argument reduction and exponent
148    adjustment of scale, positive k here means the result may overflow and
149    negative k means the result may underflow.  */
150 static inline double
specialcase(double_t tmp,uint64_t sbits,uint64_t ki)151 specialcase (double_t tmp, uint64_t sbits, uint64_t ki)
152 {
153   double_t scale, y;
154 
155   if ((ki & 0x80000000) == 0)
156     {
157       /* k > 0, the exponent of scale might have overflowed by <= 460.  */
158       sbits -= 1009ull << 52;
159       scale = asdouble (sbits);
160       y = 0x1p1009 * (scale + scale * tmp);
161       return check_oflow (y);
162     }
163   /* k < 0, need special care in the subnormal range.  */
164   sbits += 1022ull << 52;
165   /* Note: sbits is signed scale.  */
166   scale = asdouble (sbits);
167   y = scale + scale * tmp;
168 #if FLT_EVAL_METHOD == 2
169 #define fabs(x) fabsl(x)
170 #endif
171   if (fabs (y) < 1.0)
172 #undef fabs
173     {
174       /* Round y to the right precision before scaling it into the subnormal
175 	 range to avoid double rounding that can cause 0.5+E/2 ulp error where
176 	 E is the worst-case ulp error outside the subnormal range.  So this
177 	 is only useful if the goal is better than 1 ulp worst-case error.  */
178       double_t hi, lo, one = 1.0;
179       if (y < 0.0)
180 	one = -1.0;
181       lo = scale - y + scale * tmp;
182       hi = one + y;
183       lo = one - hi + y + lo;
184       y = eval_as_double (hi + lo) - one;
185       /* Fix the sign of 0.  */
186       if (y == 0.0)
187 	y = asdouble (sbits & 0x8000000000000000);
188       /* The underflow exception needs to be signaled explicitly.  */
189       force_eval_double (opt_barrier_double (0x1p-1022) * 0x1p-1022);
190     }
191   y = 0x1p-1022 * y;
192   return check_uflow (y);
193 }
194 
195 #define SIGN_BIAS ((int32_t) 0x800 << EXP_TABLE_BITS)
196 
197 /* Computes sign*exp(x+xtail) where |xtail| < 2^-8/N and |xtail| <= |x|.
198    The sign_bias argument is SIGN_BIAS or 0 and sets the sign to -1 or 1.  */
199 static inline double
exp_inline(double x,double xtail,uint32_t sign_bias)200 exp_inline (double x, double xtail, uint32_t sign_bias)
201 {
202   uint32_t abstop;
203   uint64_t ki, idx, top, sbits;
204   /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
205   double_t kd, z, r, r2, scale, tail, tmp;
206 
207   abstop = top12 (x) & 0x7ff;
208   if (unlikely (abstop - top12 (0x1p-54) >= top12 (512.0) - top12 (0x1p-54)))
209     {
210       if (abstop - top12 (0x1p-54) >= 0x80000000)
211 	{
212 	  /* Avoid spurious underflow for tiny x.  */
213 	  /* Note: 0 is common input.  */
214 	  double_t one = WANT_ROUNDING ? 1.0 + x : 1.0;
215 	  return sign_bias ? -one : one;
216 	}
217       if (abstop >= top12 (1024.0))
218 	{
219 	  /* Note: inf and nan are already handled.  */
220 	  if (asuint64 (x) >> 63)
221 	    return __math_uflow (sign_bias);
222 	  else
223 	    return __math_oflow (sign_bias);
224 	}
225       /* Large x is special cased below.  */
226       abstop = 0;
227     }
228 
229   /* exp(x) = 2^(k/N) * exp(r), with exp(r) in [2^(-1/2N),2^(1/2N)].  */
230   /* x = ln2/N*k + r, with int k and r in [-ln2/2N, ln2/2N].  */
231   z = InvLn2N * x;
232 #if TOINT_INTRINSICS
233   kd = roundtoint (z);
234   ki = converttoint (z);
235 #elif EXP_USE_TOINT_NARROW
236   /* z - kd is in [-0.5-2^-16, 0.5] in all rounding modes.  */
237   kd = eval_as_double (z + Shift);
238   ki = asuint64 (kd) >> 16;
239   kd = (double_t) (int32_t) ki;
240 #else
241   /* z - kd is in [-1, 1] in non-nearest rounding modes.  */
242   kd = eval_as_double (z + Shift);
243   ki = asuint64 (kd);
244   kd -= Shift;
245 #endif
246   r = x + kd * NegLn2hiN + kd * NegLn2loN;
247   /* The code assumes 2^-200 < |xtail| < 2^-8/N.  */
248   r += xtail;
249   /* 2^(k/N) ~= scale * (1 + tail).  */
250   idx = 2 * (ki % N);
251   top = (ki + sign_bias) << (52 - EXP_TABLE_BITS);
252   tail = asdouble (T[idx]);
253   /* This is only a valid scale when -1023*N < k < 1024*N.  */
254   sbits = T[idx + 1] + top;
255   /* exp(x) = 2^(k/N) * exp(r) ~= scale + scale * (tail + exp(r) - 1).  */
256   /* Evaluation is optimized assuming superscalar pipelined execution.  */
257   r2 = r * r;
258   /* Without fma the worst case error is 0.25/N ulp larger.  */
259   /* Worst case error is less than 0.5+1.11/N+(abs poly error * 2^53) ulp.  */
260 #if EXP_POLY_ORDER == 4
261   tmp = tail + r + r2 * C2 + r * r2 * (C3 + r * C4);
262 #elif EXP_POLY_ORDER == 5
263   tmp = tail + r + r2 * (C2 + r * C3) + r2 * r2 * (C4 + r * C5);
264 #elif EXP_POLY_ORDER == 6
265   tmp = tail + r + r2 * (0.5 + r * C3) + r2 * r2 * (C4 + r * C5 + r2 * C6);
266 #endif
267   if (unlikely (abstop == 0))
268     return specialcase (tmp, sbits, ki);
269   scale = asdouble (sbits);
270   /* Note: tmp == 0 or |tmp| > 2^-200 and scale > 2^-739, so there
271      is no spurious underflow here even without fma.  */
272   return scale + scale * tmp;
273 }
274 
275 /* Returns 0 if not int, 1 if odd int, 2 if even int.  The argument is
276    the bit representation of a non-zero finite floating-point value.  */
277 static inline int
checkint(uint64_t iy)278 checkint (uint64_t iy)
279 {
280   int e = iy >> 52 & 0x7ff;
281   if (e < 0x3ff)
282     return 0;
283   if (e > 0x3ff + 52)
284     return 2;
285   if (iy & ((1ULL << (0x3ff + 52 - e)) - 1))
286     return 0;
287   if (iy & (1ULL << (0x3ff + 52 - e)))
288     return 1;
289   return 2;
290 }
291 
292 /* Returns 1 if input is the bit representation of 0, infinity or nan.  */
293 static inline int
zeroinfnan(uint64_t i)294 zeroinfnan (uint64_t i)
295 {
296   return 2 * i - 1 >= 2 * asuint64 ((double) INFINITY) - 1;
297 }
298 
299 double
pow(double x,double y)300 pow (double x, double y)
301 {
302   uint32_t sign_bias = 0;
303   uint64_t ix, iy;
304   uint32_t topx, topy;
305 
306   ix = asuint64 (x);
307   iy = asuint64 (y);
308   topx = top12 (x);
309   topy = top12 (y);
310   if (unlikely (topx - 0x001 >= 0x7ff - 0x001
311 		|| (topy & 0x7ff) - 0x3be >= 0x43e - 0x3be))
312     {
313       /* Note: if |y| > 1075 * ln2 * 2^53 ~= 0x1.749p62 then pow(x,y) = inf/0
314 	 and if |y| < 2^-54 / 1075 ~= 0x1.e7b6p-65 then pow(x,y) = +-1.  */
315       /* Special cases: (x < 0x1p-126 or inf or nan) or
316 	 (|y| < 0x1p-65 or |y| >= 0x1p63 or nan).  */
317       if (unlikely (zeroinfnan (iy)))
318 	{
319 	  if (2 * iy == 0)
320 	    return issignaling_inline (x) ? x + y : 1.0;
321 	  if (ix == asuint64 (1.0))
322 	    return issignaling_inline (y) ? x + y : 1.0;
323 	  if (2 * ix > 2 * asuint64 ((double) INFINITY)
324 	      || 2 * iy > 2 * asuint64 ((double) INFINITY))
325 	    return x + y;
326 	  if (2 * ix == 2 * asuint64 (1.0))
327 	    return 1.0;
328 	  if ((2 * ix < 2 * asuint64 (1.0)) == !(iy >> 63))
329 	    return 0.0; /* |x|<1 && y==inf or |x|>1 && y==-inf.  */
330 	  return y * y;
331 	}
332       if (unlikely (zeroinfnan (ix)))
333 	{
334 	  double_t x2 = x * x;
335 	  if (ix >> 63 && checkint (iy) == 1)
336 	    {
337 	      x2 = -x2;
338 	      sign_bias = 1;
339 	    }
340 	  if (WANT_ERRNO && 2 * ix == 0 && iy >> 63)
341 	    return __math_divzero (sign_bias);
342 	  /* Without the barrier some versions of clang hoist the 1/x2 and
343 	     thus division by zero exception can be signaled spuriously.  */
344 	  return iy >> 63 ? opt_barrier_double (1 / x2) : x2;
345 	}
346       /* Here x and y are non-zero finite.  */
347       if (ix >> 63)
348 	{
349 	  /* Finite x < 0.  */
350 	  int yint = checkint (iy);
351 	  if (yint == 0)
352 	    return __math_invalid (x);
353 	  if (yint == 1)
354 	    sign_bias = SIGN_BIAS;
355 	  ix &= 0x7fffffffffffffff;
356 	  topx &= 0x7ff;
357 	}
358       if ((topy & 0x7ff) - 0x3be >= 0x43e - 0x3be)
359 	{
360 	  /* Note: sign_bias == 0 here because y is not odd.  */
361 	  if (ix == asuint64 (1.0))
362 	    return 1.0;
363 	  if ((topy & 0x7ff) < 0x3be)
364 	    {
365 	      /* |y| < 2^-65, x^y ~= 1 + y*log(x).  */
366 	      if (WANT_ROUNDING)
367 		return ix > asuint64 (1.0) ? 1.0 + y : 1.0 - y;
368 	      else
369 		return 1.0;
370 	    }
371 	  return (ix > asuint64 (1.0)) == (topy < 0x800) ? __math_oflow (0)
372 							 : __math_uflow (0);
373 	}
374       if (topx == 0)
375 	{
376 	  /* Normalize subnormal x so exponent becomes negative.  */
377 	  ix = asuint64 (x * 0x1p52);
378 	  ix &= 0x7fffffffffffffff;
379 	  ix -= 52ULL << 52;
380 	}
381     }
382 
383   double_t lo;
384   double_t hi = log_inline (ix, &lo);
385   double_t ehi, elo;
386 #if _HAVE_FAST_FMA
387   ehi = y * hi;
388   elo = y * lo + fma (y, hi, -ehi);
389 #else
390   double_t yhi = asdouble (iy & -1ULL << 27);
391   double_t ylo = y - yhi;
392   double_t lhi = asdouble (asuint64 (hi) & -1ULL << 27);
393   double_t llo = hi - lhi + lo;
394   ehi = yhi * lhi;
395   elo = ylo * lhi + y * llo; /* |elo| < |ehi| * 2^-25.  */
396 #endif
397   return exp_inline (ehi, elo, sign_bias);
398 }
399 
400 #if defined(_HAVE_ALIAS_ATTRIBUTE)
401 #ifndef __clang__
402 #pragma GCC diagnostic ignored "-Wmissing-attributes"
403 #endif
404 __strong_reference(pow, _pow);
405 #endif
406 
407 _MATH_ALIAS_d_dd(pow)
408 
409 #endif
410