1 | /* |
2 | * Copyright (c) 2005, 2017, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "jni.h" |
27 | #include "runtime/interfaceSupport.inline.hpp" |
28 | #include "runtime/sharedRuntime.hpp" |
29 | |
30 | // This file contains copies of the fdlibm routines used by |
31 | // StrictMath. It turns out that it is almost always required to use |
32 | // these runtime routines; the Intel CPU doesn't meet the Java |
33 | // specification for sin/cos outside a certain limited argument range, |
34 | // and the SPARC CPU doesn't appear to have sin/cos instructions. It |
35 | // also turns out that avoiding the indirect call through function |
36 | // pointer out to libjava.so in SharedRuntime speeds these routines up |
37 | // by roughly 15% on both Win32/x86 and Solaris/SPARC. |
38 | |
39 | // Enabling optimizations in this file causes incorrect code to be |
40 | // generated; can not figure out how to turn down optimization for one |
41 | // file in the IDE on Windows |
42 | #ifdef WIN32 |
43 | # pragma warning( disable: 4748 ) // /GS can not protect parameters and local variables from local buffer overrun because optimizations are disabled in function |
44 | # pragma optimize ( "", off ) |
45 | #endif |
46 | |
47 | #include "runtime/sharedRuntimeMath.hpp" |
48 | |
49 | /* __ieee754_log(x) |
50 | * Return the logarithm of x |
51 | * |
52 | * Method : |
53 | * 1. Argument Reduction: find k and f such that |
54 | * x = 2^k * (1+f), |
55 | * where sqrt(2)/2 < 1+f < sqrt(2) . |
56 | * |
57 | * 2. Approximation of log(1+f). |
58 | * Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s) |
59 | * = 2s + 2/3 s**3 + 2/5 s**5 + ....., |
60 | * = 2s + s*R |
61 | * We use a special Reme algorithm on [0,0.1716] to generate |
62 | * a polynomial of degree 14 to approximate R The maximum error |
63 | * of this polynomial approximation is bounded by 2**-58.45. In |
64 | * other words, |
65 | * 2 4 6 8 10 12 14 |
66 | * R(z) ~ Lg1*s +Lg2*s +Lg3*s +Lg4*s +Lg5*s +Lg6*s +Lg7*s |
67 | * (the values of Lg1 to Lg7 are listed in the program) |
68 | * and |
69 | * | 2 14 | -58.45 |
70 | * | Lg1*s +...+Lg7*s - R(z) | <= 2 |
71 | * | | |
72 | * Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2. |
73 | * In order to guarantee error in log below 1ulp, we compute log |
74 | * by |
75 | * log(1+f) = f - s*(f - R) (if f is not too large) |
76 | * log(1+f) = f - (hfsq - s*(hfsq+R)). (better accuracy) |
77 | * |
78 | * 3. Finally, log(x) = k*ln2 + log(1+f). |
79 | * = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo))) |
80 | * Here ln2 is split into two floating point number: |
81 | * ln2_hi + ln2_lo, |
82 | * where n*ln2_hi is always exact for |n| < 2000. |
83 | * |
84 | * Special cases: |
85 | * log(x) is NaN with signal if x < 0 (including -INF) ; |
86 | * log(+INF) is +INF; log(0) is -INF with signal; |
87 | * log(NaN) is that NaN with no signal. |
88 | * |
89 | * Accuracy: |
90 | * according to an error analysis, the error is always less than |
91 | * 1 ulp (unit in the last place). |
92 | * |
93 | * Constants: |
94 | * The hexadecimal values are the intended ones for the following |
95 | * constants. The decimal values may be used, provided that the |
96 | * compiler will convert from decimal to binary accurately enough |
97 | * to produce the hexadecimal values shown. |
98 | */ |
99 | |
100 | static const double |
101 | ln2_hi = 6.93147180369123816490e-01, /* 3fe62e42 fee00000 */ |
102 | ln2_lo = 1.90821492927058770002e-10, /* 3dea39ef 35793c76 */ |
103 | Lg1 = 6.666666666666735130e-01, /* 3FE55555 55555593 */ |
104 | Lg2 = 3.999999999940941908e-01, /* 3FD99999 9997FA04 */ |
105 | Lg3 = 2.857142874366239149e-01, /* 3FD24924 94229359 */ |
106 | Lg4 = 2.222219843214978396e-01, /* 3FCC71C5 1D8E78AF */ |
107 | Lg5 = 1.818357216161805012e-01, /* 3FC74664 96CB03DE */ |
108 | Lg6 = 1.531383769920937332e-01, /* 3FC39A09 D078C69F */ |
109 | Lg7 = 1.479819860511658591e-01; /* 3FC2F112 DF3E5244 */ |
110 | |
111 | static double zero = 0.0; |
112 | |
113 | static double __ieee754_log(double x) { |
114 | double hfsq,f,s,z,R,w,t1,t2,dk; |
115 | int k,hx,i,j; |
116 | unsigned lx; |
117 | |
118 | hx = high(x); /* high word of x */ |
119 | lx = low(x); /* low word of x */ |
120 | |
121 | k=0; |
122 | if (hx < 0x00100000) { /* x < 2**-1022 */ |
123 | if (((hx&0x7fffffff)|lx)==0) |
124 | return -two54/zero; /* log(+-0)=-inf */ |
125 | if (hx<0) return (x-x)/zero; /* log(-#) = NaN */ |
126 | k -= 54; x *= two54; /* subnormal number, scale up x */ |
127 | hx = high(x); /* high word of x */ |
128 | } |
129 | if (hx >= 0x7ff00000) return x+x; |
130 | k += (hx>>20)-1023; |
131 | hx &= 0x000fffff; |
132 | i = (hx+0x95f64)&0x100000; |
133 | set_high(&x, hx|(i^0x3ff00000)); /* normalize x or x/2 */ |
134 | k += (i>>20); |
135 | f = x-1.0; |
136 | if((0x000fffff&(2+hx))<3) { /* |f| < 2**-20 */ |
137 | if(f==zero) { |
138 | if (k==0) return zero; |
139 | else {dk=(double)k; return dk*ln2_hi+dk*ln2_lo;} |
140 | } |
141 | R = f*f*(0.5-0.33333333333333333*f); |
142 | if(k==0) return f-R; else {dk=(double)k; |
143 | return dk*ln2_hi-((R-dk*ln2_lo)-f);} |
144 | } |
145 | s = f/(2.0+f); |
146 | dk = (double)k; |
147 | z = s*s; |
148 | i = hx-0x6147a; |
149 | w = z*z; |
150 | j = 0x6b851-hx; |
151 | t1= w*(Lg2+w*(Lg4+w*Lg6)); |
152 | t2= z*(Lg1+w*(Lg3+w*(Lg5+w*Lg7))); |
153 | i |= j; |
154 | R = t2+t1; |
155 | if(i>0) { |
156 | hfsq=0.5*f*f; |
157 | if(k==0) return f-(hfsq-s*(hfsq+R)); else |
158 | return dk*ln2_hi-((hfsq-(s*(hfsq+R)+dk*ln2_lo))-f); |
159 | } else { |
160 | if(k==0) return f-s*(f-R); else |
161 | return dk*ln2_hi-((s*(f-R)-dk*ln2_lo)-f); |
162 | } |
163 | } |
164 | |
165 | JRT_LEAF(jdouble, SharedRuntime::dlog(jdouble x)) |
166 | return __ieee754_log(x); |
167 | JRT_END |
168 | |
169 | /* __ieee754_log10(x) |
170 | * Return the base 10 logarithm of x |
171 | * |
172 | * Method : |
173 | * Let log10_2hi = leading 40 bits of log10(2) and |
174 | * log10_2lo = log10(2) - log10_2hi, |
175 | * ivln10 = 1/log(10) rounded. |
176 | * Then |
177 | * n = ilogb(x), |
178 | * if(n<0) n = n+1; |
179 | * x = scalbn(x,-n); |
180 | * log10(x) := n*log10_2hi + (n*log10_2lo + ivln10*log(x)) |
181 | * |
182 | * Note 1: |
183 | * To guarantee log10(10**n)=n, where 10**n is normal, the rounding |
184 | * mode must set to Round-to-Nearest. |
185 | * Note 2: |
186 | * [1/log(10)] rounded to 53 bits has error .198 ulps; |
187 | * log10 is monotonic at all binary break points. |
188 | * |
189 | * Special cases: |
190 | * log10(x) is NaN with signal if x < 0; |
191 | * log10(+INF) is +INF with no signal; log10(0) is -INF with signal; |
192 | * log10(NaN) is that NaN with no signal; |
193 | * log10(10**N) = N for N=0,1,...,22. |
194 | * |
195 | * Constants: |
196 | * The hexadecimal values are the intended ones for the following constants. |
197 | * The decimal values may be used, provided that the compiler will convert |
198 | * from decimal to binary accurately enough to produce the hexadecimal values |
199 | * shown. |
200 | */ |
201 | |
202 | static const double |
203 | ivln10 = 4.34294481903251816668e-01, /* 0x3FDBCB7B, 0x1526E50E */ |
204 | log10_2hi = 3.01029995663611771306e-01, /* 0x3FD34413, 0x509F6000 */ |
205 | log10_2lo = 3.69423907715893078616e-13; /* 0x3D59FEF3, 0x11F12B36 */ |
206 | |
207 | static double __ieee754_log10(double x) { |
208 | double y,z; |
209 | int i,k,hx; |
210 | unsigned lx; |
211 | |
212 | hx = high(x); /* high word of x */ |
213 | lx = low(x); /* low word of x */ |
214 | |
215 | k=0; |
216 | if (hx < 0x00100000) { /* x < 2**-1022 */ |
217 | if (((hx&0x7fffffff)|lx)==0) |
218 | return -two54/zero; /* log(+-0)=-inf */ |
219 | if (hx<0) return (x-x)/zero; /* log(-#) = NaN */ |
220 | k -= 54; x *= two54; /* subnormal number, scale up x */ |
221 | hx = high(x); /* high word of x */ |
222 | } |
223 | if (hx >= 0x7ff00000) return x+x; |
224 | k += (hx>>20)-1023; |
225 | i = ((unsigned)k&0x80000000)>>31; |
226 | hx = (hx&0x000fffff)|((0x3ff-i)<<20); |
227 | y = (double)(k+i); |
228 | set_high(&x, hx); |
229 | z = y*log10_2lo + ivln10*__ieee754_log(x); |
230 | return z+y*log10_2hi; |
231 | } |
232 | |
233 | JRT_LEAF(jdouble, SharedRuntime::dlog10(jdouble x)) |
234 | return __ieee754_log10(x); |
235 | JRT_END |
236 | |
237 | |
238 | /* __ieee754_exp(x) |
239 | * Returns the exponential of x. |
240 | * |
241 | * Method |
242 | * 1. Argument reduction: |
243 | * Reduce x to an r so that |r| <= 0.5*ln2 ~ 0.34658. |
244 | * Given x, find r and integer k such that |
245 | * |
246 | * x = k*ln2 + r, |r| <= 0.5*ln2. |
247 | * |
248 | * Here r will be represented as r = hi-lo for better |
249 | * accuracy. |
250 | * |
251 | * 2. Approximation of exp(r) by a special rational function on |
252 | * the interval [0,0.34658]: |
253 | * Write |
254 | * R(r**2) = r*(exp(r)+1)/(exp(r)-1) = 2 + r*r/6 - r**4/360 + ... |
255 | * We use a special Reme algorithm on [0,0.34658] to generate |
256 | * a polynomial of degree 5 to approximate R. The maximum error |
257 | * of this polynomial approximation is bounded by 2**-59. In |
258 | * other words, |
259 | * R(z) ~ 2.0 + P1*z + P2*z**2 + P3*z**3 + P4*z**4 + P5*z**5 |
260 | * (where z=r*r, and the values of P1 to P5 are listed below) |
261 | * and |
262 | * | 5 | -59 |
263 | * | 2.0+P1*z+...+P5*z - R(z) | <= 2 |
264 | * | | |
265 | * The computation of exp(r) thus becomes |
266 | * 2*r |
267 | * exp(r) = 1 + ------- |
268 | * R - r |
269 | * r*R1(r) |
270 | * = 1 + r + ----------- (for better accuracy) |
271 | * 2 - R1(r) |
272 | * where |
273 | * 2 4 10 |
274 | * R1(r) = r - (P1*r + P2*r + ... + P5*r ). |
275 | * |
276 | * 3. Scale back to obtain exp(x): |
277 | * From step 1, we have |
278 | * exp(x) = 2^k * exp(r) |
279 | * |
280 | * Special cases: |
281 | * exp(INF) is INF, exp(NaN) is NaN; |
282 | * exp(-INF) is 0, and |
283 | * for finite argument, only exp(0)=1 is exact. |
284 | * |
285 | * Accuracy: |
286 | * according to an error analysis, the error is always less than |
287 | * 1 ulp (unit in the last place). |
288 | * |
289 | * Misc. info. |
290 | * For IEEE double |
291 | * if x > 7.09782712893383973096e+02 then exp(x) overflow |
292 | * if x < -7.45133219101941108420e+02 then exp(x) underflow |
293 | * |
294 | * Constants: |
295 | * The hexadecimal values are the intended ones for the following |
296 | * constants. The decimal values may be used, provided that the |
297 | * compiler will convert from decimal to binary accurately enough |
298 | * to produce the hexadecimal values shown. |
299 | */ |
300 | |
301 | static const double |
302 | one = 1.0, |
303 | halF[2] = {0.5,-0.5,}, |
304 | twom1000= 9.33263618503218878990e-302, /* 2**-1000=0x01700000,0*/ |
305 | o_threshold= 7.09782712893383973096e+02, /* 0x40862E42, 0xFEFA39EF */ |
306 | u_threshold= -7.45133219101941108420e+02, /* 0xc0874910, 0xD52D3051 */ |
307 | ln2HI[2] ={ 6.93147180369123816490e-01, /* 0x3fe62e42, 0xfee00000 */ |
308 | -6.93147180369123816490e-01,},/* 0xbfe62e42, 0xfee00000 */ |
309 | ln2LO[2] ={ 1.90821492927058770002e-10, /* 0x3dea39ef, 0x35793c76 */ |
310 | -1.90821492927058770002e-10,},/* 0xbdea39ef, 0x35793c76 */ |
311 | invln2 = 1.44269504088896338700e+00, /* 0x3ff71547, 0x652b82fe */ |
312 | P1 = 1.66666666666666019037e-01, /* 0x3FC55555, 0x5555553E */ |
313 | P2 = -2.77777777770155933842e-03, /* 0xBF66C16C, 0x16BEBD93 */ |
314 | P3 = 6.61375632143793436117e-05, /* 0x3F11566A, 0xAF25DE2C */ |
315 | P4 = -1.65339022054652515390e-06, /* 0xBEBBBD41, 0xC5D26BF1 */ |
316 | P5 = 4.13813679705723846039e-08; /* 0x3E663769, 0x72BEA4D0 */ |
317 | |
318 | static double __ieee754_exp(double x) { |
319 | double y,hi=0,lo=0,c,t; |
320 | int k=0,xsb; |
321 | unsigned hx; |
322 | |
323 | hx = high(x); /* high word of x */ |
324 | xsb = (hx>>31)&1; /* sign bit of x */ |
325 | hx &= 0x7fffffff; /* high word of |x| */ |
326 | |
327 | /* filter out non-finite argument */ |
328 | if(hx >= 0x40862E42) { /* if |x|>=709.78... */ |
329 | if(hx>=0x7ff00000) { |
330 | if(((hx&0xfffff)|low(x))!=0) |
331 | return x+x; /* NaN */ |
332 | else return (xsb==0)? x:0.0; /* exp(+-inf)={inf,0} */ |
333 | } |
334 | if(x > o_threshold) return hugeX*hugeX; /* overflow */ |
335 | if(x < u_threshold) return twom1000*twom1000; /* underflow */ |
336 | } |
337 | |
338 | /* argument reduction */ |
339 | if(hx > 0x3fd62e42) { /* if |x| > 0.5 ln2 */ |
340 | if(hx < 0x3FF0A2B2) { /* and |x| < 1.5 ln2 */ |
341 | hi = x-ln2HI[xsb]; lo=ln2LO[xsb]; k = 1-xsb-xsb; |
342 | } else { |
343 | k = (int)(invln2*x+halF[xsb]); |
344 | t = k; |
345 | hi = x - t*ln2HI[0]; /* t*ln2HI is exact here */ |
346 | lo = t*ln2LO[0]; |
347 | } |
348 | x = hi - lo; |
349 | } |
350 | else if(hx < 0x3e300000) { /* when |x|<2**-28 */ |
351 | if(hugeX+x>one) return one+x;/* trigger inexact */ |
352 | } |
353 | else k = 0; |
354 | |
355 | /* x is now in primary range */ |
356 | t = x*x; |
357 | c = x - t*(P1+t*(P2+t*(P3+t*(P4+t*P5)))); |
358 | if(k==0) return one-((x*c)/(c-2.0)-x); |
359 | else y = one-((lo-(x*c)/(2.0-c))-hi); |
360 | if(k >= -1021) { |
361 | set_high(&y, high(y) + (k<<20)); /* add k to y's exponent */ |
362 | return y; |
363 | } else { |
364 | set_high(&y, high(y) + ((k+1000)<<20)); /* add k to y's exponent */ |
365 | return y*twom1000; |
366 | } |
367 | } |
368 | |
369 | JRT_LEAF(jdouble, SharedRuntime::dexp(jdouble x)) |
370 | return __ieee754_exp(x); |
371 | JRT_END |
372 | |
373 | /* __ieee754_pow(x,y) return x**y |
374 | * |
375 | * n |
376 | * Method: Let x = 2 * (1+f) |
377 | * 1. Compute and return log2(x) in two pieces: |
378 | * log2(x) = w1 + w2, |
379 | * where w1 has 53-24 = 29 bit trailing zeros. |
380 | * 2. Perform y*log2(x) = n+y' by simulating muti-precision |
381 | * arithmetic, where |y'|<=0.5. |
382 | * 3. Return x**y = 2**n*exp(y'*log2) |
383 | * |
384 | * Special cases: |
385 | * 1. (anything) ** 0 is 1 |
386 | * 2. (anything) ** 1 is itself |
387 | * 3. (anything) ** NAN is NAN |
388 | * 4. NAN ** (anything except 0) is NAN |
389 | * 5. +-(|x| > 1) ** +INF is +INF |
390 | * 6. +-(|x| > 1) ** -INF is +0 |
391 | * 7. +-(|x| < 1) ** +INF is +0 |
392 | * 8. +-(|x| < 1) ** -INF is +INF |
393 | * 9. +-1 ** +-INF is NAN |
394 | * 10. +0 ** (+anything except 0, NAN) is +0 |
395 | * 11. -0 ** (+anything except 0, NAN, odd integer) is +0 |
396 | * 12. +0 ** (-anything except 0, NAN) is +INF |
397 | * 13. -0 ** (-anything except 0, NAN, odd integer) is +INF |
398 | * 14. -0 ** (odd integer) = -( +0 ** (odd integer) ) |
399 | * 15. +INF ** (+anything except 0,NAN) is +INF |
400 | * 16. +INF ** (-anything except 0,NAN) is +0 |
401 | * 17. -INF ** (anything) = -0 ** (-anything) |
402 | * 18. (-anything) ** (integer) is (-1)**(integer)*(+anything**integer) |
403 | * 19. (-anything except 0 and inf) ** (non-integer) is NAN |
404 | * |
405 | * Accuracy: |
406 | * pow(x,y) returns x**y nearly rounded. In particular |
407 | * pow(integer,integer) |
408 | * always returns the correct integer provided it is |
409 | * representable. |
410 | * |
411 | * Constants : |
412 | * The hexadecimal values are the intended ones for the following |
413 | * constants. The decimal values may be used, provided that the |
414 | * compiler will convert from decimal to binary accurately enough |
415 | * to produce the hexadecimal values shown. |
416 | */ |
417 | |
418 | static const double |
419 | bp[] = {1.0, 1.5,}, |
420 | dp_h[] = { 0.0, 5.84962487220764160156e-01,}, /* 0x3FE2B803, 0x40000000 */ |
421 | dp_l[] = { 0.0, 1.35003920212974897128e-08,}, /* 0x3E4CFDEB, 0x43CFD006 */ |
422 | zeroX = 0.0, |
423 | two = 2.0, |
424 | two53 = 9007199254740992.0, /* 0x43400000, 0x00000000 */ |
425 | /* poly coefs for (3/2)*(log(x)-2s-2/3*s**3 */ |
426 | L1X = 5.99999999999994648725e-01, /* 0x3FE33333, 0x33333303 */ |
427 | L2X = 4.28571428578550184252e-01, /* 0x3FDB6DB6, 0xDB6FABFF */ |
428 | L3X = 3.33333329818377432918e-01, /* 0x3FD55555, 0x518F264D */ |
429 | L4X = 2.72728123808534006489e-01, /* 0x3FD17460, 0xA91D4101 */ |
430 | L5X = 2.30660745775561754067e-01, /* 0x3FCD864A, 0x93C9DB65 */ |
431 | L6X = 2.06975017800338417784e-01, /* 0x3FCA7E28, 0x4A454EEF */ |
432 | lg2 = 6.93147180559945286227e-01, /* 0x3FE62E42, 0xFEFA39EF */ |
433 | lg2_h = 6.93147182464599609375e-01, /* 0x3FE62E43, 0x00000000 */ |
434 | lg2_l = -1.90465429995776804525e-09, /* 0xBE205C61, 0x0CA86C39 */ |
435 | ovt = 8.0085662595372944372e-0017, /* -(1024-log2(ovfl+.5ulp)) */ |
436 | cp = 9.61796693925975554329e-01, /* 0x3FEEC709, 0xDC3A03FD =2/(3ln2) */ |
437 | cp_h = 9.61796700954437255859e-01, /* 0x3FEEC709, 0xE0000000 =(float)cp */ |
438 | cp_l = -7.02846165095275826516e-09, /* 0xBE3E2FE0, 0x145B01F5 =tail of cp_h*/ |
439 | ivln2 = 1.44269504088896338700e+00, /* 0x3FF71547, 0x652B82FE =1/ln2 */ |
440 | ivln2_h = 1.44269502162933349609e+00, /* 0x3FF71547, 0x60000000 =24b 1/ln2*/ |
441 | ivln2_l = 1.92596299112661746887e-08; /* 0x3E54AE0B, 0xF85DDF44 =1/ln2 tail*/ |
442 | |
443 | double __ieee754_pow(double x, double y) { |
444 | double z,ax,z_h,z_l,p_h,p_l; |
445 | double y1,t1,t2,r,s,t,u,v,w; |
446 | int i0,i1,i,j,k,yisint,n; |
447 | int hx,hy,ix,iy; |
448 | unsigned lx,ly; |
449 | |
450 | i0 = ((*(int*)&one)>>29)^1; i1=1-i0; |
451 | hx = high(x); lx = low(x); |
452 | hy = high(y); ly = low(y); |
453 | ix = hx&0x7fffffff; iy = hy&0x7fffffff; |
454 | |
455 | /* y==zero: x**0 = 1 */ |
456 | if((iy|ly)==0) return one; |
457 | |
458 | /* +-NaN return x+y */ |
459 | if(ix > 0x7ff00000 || ((ix==0x7ff00000)&&(lx!=0)) || |
460 | iy > 0x7ff00000 || ((iy==0x7ff00000)&&(ly!=0))) |
461 | return x+y; |
462 | |
463 | /* determine if y is an odd int when x < 0 |
464 | * yisint = 0 ... y is not an integer |
465 | * yisint = 1 ... y is an odd int |
466 | * yisint = 2 ... y is an even int |
467 | */ |
468 | yisint = 0; |
469 | if(hx<0) { |
470 | if(iy>=0x43400000) yisint = 2; /* even integer y */ |
471 | else if(iy>=0x3ff00000) { |
472 | k = (iy>>20)-0x3ff; /* exponent */ |
473 | if(k>20) { |
474 | j = ly>>(52-k); |
475 | if((unsigned)(j<<(52-k))==ly) yisint = 2-(j&1); |
476 | } else if(ly==0) { |
477 | j = iy>>(20-k); |
478 | if((j<<(20-k))==iy) yisint = 2-(j&1); |
479 | } |
480 | } |
481 | } |
482 | |
483 | /* special value of y */ |
484 | if(ly==0) { |
485 | if (iy==0x7ff00000) { /* y is +-inf */ |
486 | if(((ix-0x3ff00000)|lx)==0) |
487 | return y - y; /* inf**+-1 is NaN */ |
488 | else if (ix >= 0x3ff00000)/* (|x|>1)**+-inf = inf,0 */ |
489 | return (hy>=0)? y: zeroX; |
490 | else /* (|x|<1)**-,+inf = inf,0 */ |
491 | return (hy<0)?-y: zeroX; |
492 | } |
493 | if(iy==0x3ff00000) { /* y is +-1 */ |
494 | if(hy<0) return one/x; else return x; |
495 | } |
496 | if(hy==0x40000000) return x*x; /* y is 2 */ |
497 | if(hy==0x3fe00000) { /* y is 0.5 */ |
498 | if(hx>=0) /* x >= +0 */ |
499 | return sqrt(x); |
500 | } |
501 | } |
502 | |
503 | ax = fabsd(x); |
504 | /* special value of x */ |
505 | if(lx==0) { |
506 | if(ix==0x7ff00000||ix==0||ix==0x3ff00000){ |
507 | z = ax; /*x is +-0,+-inf,+-1*/ |
508 | if(hy<0) z = one/z; /* z = (1/|x|) */ |
509 | if(hx<0) { |
510 | if(((ix-0x3ff00000)|yisint)==0) { |
511 | #ifdef CAN_USE_NAN_DEFINE |
512 | z = NAN; |
513 | #else |
514 | z = (z-z)/(z-z); /* (-1)**non-int is NaN */ |
515 | #endif |
516 | } else if(yisint==1) |
517 | z = -1.0*z; /* (x<0)**odd = -(|x|**odd) */ |
518 | } |
519 | return z; |
520 | } |
521 | } |
522 | |
523 | n = (hx>>31)+1; |
524 | |
525 | /* (x<0)**(non-int) is NaN */ |
526 | if((n|yisint)==0) |
527 | #ifdef CAN_USE_NAN_DEFINE |
528 | return NAN; |
529 | #else |
530 | return (x-x)/(x-x); |
531 | #endif |
532 | |
533 | s = one; /* s (sign of result -ve**odd) = -1 else = 1 */ |
534 | if((n|(yisint-1))==0) s = -one;/* (-ve)**(odd int) */ |
535 | |
536 | /* |y| is huge */ |
537 | if(iy>0x41e00000) { /* if |y| > 2**31 */ |
538 | if(iy>0x43f00000){ /* if |y| > 2**64, must o/uflow */ |
539 | if(ix<=0x3fefffff) return (hy<0)? hugeX*hugeX:tiny*tiny; |
540 | if(ix>=0x3ff00000) return (hy>0)? hugeX*hugeX:tiny*tiny; |
541 | } |
542 | /* over/underflow if x is not close to one */ |
543 | if(ix<0x3fefffff) return (hy<0)? s*hugeX*hugeX:s*tiny*tiny; |
544 | if(ix>0x3ff00000) return (hy>0)? s*hugeX*hugeX:s*tiny*tiny; |
545 | /* now |1-x| is tiny <= 2**-20, suffice to compute |
546 | log(x) by x-x^2/2+x^3/3-x^4/4 */ |
547 | t = ax-one; /* t has 20 trailing zeros */ |
548 | w = (t*t)*(0.5-t*(0.3333333333333333333333-t*0.25)); |
549 | u = ivln2_h*t; /* ivln2_h has 21 sig. bits */ |
550 | v = t*ivln2_l-w*ivln2; |
551 | t1 = u+v; |
552 | set_low(&t1, 0); |
553 | t2 = v-(t1-u); |
554 | } else { |
555 | double ss,s2,s_h,s_l,t_h,t_l; |
556 | n = 0; |
557 | /* take care subnormal number */ |
558 | if(ix<0x00100000) |
559 | {ax *= two53; n -= 53; ix = high(ax); } |
560 | n += ((ix)>>20)-0x3ff; |
561 | j = ix&0x000fffff; |
562 | /* determine interval */ |
563 | ix = j|0x3ff00000; /* normalize ix */ |
564 | if(j<=0x3988E) k=0; /* |x|<sqrt(3/2) */ |
565 | else if(j<0xBB67A) k=1; /* |x|<sqrt(3) */ |
566 | else {k=0;n+=1;ix -= 0x00100000;} |
567 | set_high(&ax, ix); |
568 | |
569 | /* compute ss = s_h+s_l = (x-1)/(x+1) or (x-1.5)/(x+1.5) */ |
570 | u = ax-bp[k]; /* bp[0]=1.0, bp[1]=1.5 */ |
571 | v = one/(ax+bp[k]); |
572 | ss = u*v; |
573 | s_h = ss; |
574 | set_low(&s_h, 0); |
575 | /* t_h=ax+bp[k] High */ |
576 | t_h = zeroX; |
577 | set_high(&t_h, ((ix>>1)|0x20000000)+0x00080000+(k<<18)); |
578 | t_l = ax - (t_h-bp[k]); |
579 | s_l = v*((u-s_h*t_h)-s_h*t_l); |
580 | /* compute log(ax) */ |
581 | s2 = ss*ss; |
582 | r = s2*s2*(L1X+s2*(L2X+s2*(L3X+s2*(L4X+s2*(L5X+s2*L6X))))); |
583 | r += s_l*(s_h+ss); |
584 | s2 = s_h*s_h; |
585 | t_h = 3.0+s2+r; |
586 | set_low(&t_h, 0); |
587 | t_l = r-((t_h-3.0)-s2); |
588 | /* u+v = ss*(1+...) */ |
589 | u = s_h*t_h; |
590 | v = s_l*t_h+t_l*ss; |
591 | /* 2/(3log2)*(ss+...) */ |
592 | p_h = u+v; |
593 | set_low(&p_h, 0); |
594 | p_l = v-(p_h-u); |
595 | z_h = cp_h*p_h; /* cp_h+cp_l = 2/(3*log2) */ |
596 | z_l = cp_l*p_h+p_l*cp+dp_l[k]; |
597 | /* log2(ax) = (ss+..)*2/(3*log2) = n + dp_h + z_h + z_l */ |
598 | t = (double)n; |
599 | t1 = (((z_h+z_l)+dp_h[k])+t); |
600 | set_low(&t1, 0); |
601 | t2 = z_l-(((t1-t)-dp_h[k])-z_h); |
602 | } |
603 | |
604 | /* split up y into y1+y2 and compute (y1+y2)*(t1+t2) */ |
605 | y1 = y; |
606 | set_low(&y1, 0); |
607 | p_l = (y-y1)*t1+y*t2; |
608 | p_h = y1*t1; |
609 | z = p_l+p_h; |
610 | j = high(z); |
611 | i = low(z); |
612 | if (j>=0x40900000) { /* z >= 1024 */ |
613 | if(((j-0x40900000)|i)!=0) /* if z > 1024 */ |
614 | return s*hugeX*hugeX; /* overflow */ |
615 | else { |
616 | if(p_l+ovt>z-p_h) return s*hugeX*hugeX; /* overflow */ |
617 | } |
618 | } else if((j&0x7fffffff)>=0x4090cc00 ) { /* z <= -1075 */ |
619 | if(((j-0xc090cc00)|i)!=0) /* z < -1075 */ |
620 | return s*tiny*tiny; /* underflow */ |
621 | else { |
622 | if(p_l<=z-p_h) return s*tiny*tiny; /* underflow */ |
623 | } |
624 | } |
625 | /* |
626 | * compute 2**(p_h+p_l) |
627 | */ |
628 | i = j&0x7fffffff; |
629 | k = (i>>20)-0x3ff; |
630 | n = 0; |
631 | if(i>0x3fe00000) { /* if |z| > 0.5, set n = [z+0.5] */ |
632 | n = j+(0x00100000>>(k+1)); |
633 | k = ((n&0x7fffffff)>>20)-0x3ff; /* new k for n */ |
634 | t = zeroX; |
635 | set_high(&t, (n&~(0x000fffff>>k))); |
636 | n = ((n&0x000fffff)|0x00100000)>>(20-k); |
637 | if(j<0) n = -n; |
638 | p_h -= t; |
639 | } |
640 | t = p_l+p_h; |
641 | set_low(&t, 0); |
642 | u = t*lg2_h; |
643 | v = (p_l-(t-p_h))*lg2+t*lg2_l; |
644 | z = u+v; |
645 | w = v-(z-u); |
646 | t = z*z; |
647 | t1 = z - t*(P1+t*(P2+t*(P3+t*(P4+t*P5)))); |
648 | r = (z*t1)/(t1-two)-(w+z*w); |
649 | z = one-(r-z); |
650 | j = high(z); |
651 | j += (n<<20); |
652 | if((j>>20)<=0) z = scalbnA(z,n); /* subnormal output */ |
653 | else set_high(&z, high(z) + (n<<20)); |
654 | return s*z; |
655 | } |
656 | |
657 | |
658 | JRT_LEAF(jdouble, SharedRuntime::dpow(jdouble x, jdouble y)) |
659 | return __ieee754_pow(x, y); |
660 | JRT_END |
661 | |
662 | #ifdef WIN32 |
663 | # pragma optimize ( "", on ) |
664 | #endif |
665 | |