1 /*
2 * Copyright (c) 2015, Intel Corporation.
3 * Intel Math Library (LIBM) Source Code
4 *
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 *
7 * This code is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 only, as
9 * published by the Free Software Foundation.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 *
25 */
26
27 #include "precompiled.hpp"
28 #include "asm/assembler.hpp"
29 #include "asm/assembler.inline.hpp"
30 #include "macroAssembler_x86.hpp"
31
32 #ifdef _MSC_VER
33 #define ALIGNED_(x) __declspec(align(x))
34 #else
35 #define ALIGNED_(x) __attribute__ ((aligned(x)))
36 #endif
37
38 /******************************************************************************/
39 // ALGORITHM DESCRIPTION - EXP()
40 // ---------------------
41 //
42 // Description:
43 // Let K = 64 (table size).
44 // x x/log(2) n
45 // e = 2 = 2 * T[j] * (1 + P(y))
46 // where
47 // x = m*log(2)/K + y, y in [-log(2)/K..log(2)/K]
48 // m = n*K + j, m,n,j - signed integer, j in [-K/2..K/2]
49 // j/K
50 // values of 2 are tabulated as T[j] = T_hi[j] ( 1 + T_lo[j]).
51 //
52 // P(y) is a minimax polynomial approximation of exp(x)-1
53 // on small interval [-log(2)/K..log(2)/K] (were calculated by Maple V).
54 //
55 // To avoid problems with arithmetic overflow and underflow,
56 // n n1 n2
57 // value of 2 is safely computed as 2 * 2 where n1 in [-BIAS/2..BIAS/2]
58 // where BIAS is a value of exponent bias.
59 //
60 // Special cases:
61 // exp(NaN) = NaN
62 // exp(+INF) = +INF
63 // exp(-INF) = 0
64 // exp(x) = 1 for subnormals
65 // for finite argument, only exp(0)=1 is exact
66 // For IEEE double
67 // if x > 709.782712893383973096 then exp(x) overflow
68 // if x < -745.133219101941108420 then exp(x) underflow
69 //
70 /******************************************************************************/
71
72 #ifdef _LP64
73
74 ALIGNED_(16) juint _cv[] =
75 {
76 0x652b82feUL, 0x40571547UL, 0x652b82feUL, 0x40571547UL, 0xfefa0000UL,
77 0x3f862e42UL, 0xfefa0000UL, 0x3f862e42UL, 0xbc9e3b3aUL, 0x3d1cf79aUL,
78 0xbc9e3b3aUL, 0x3d1cf79aUL, 0xfffffffeUL, 0x3fdfffffUL, 0xfffffffeUL,
79 0x3fdfffffUL, 0xe3289860UL, 0x3f56c15cUL, 0x555b9e25UL, 0x3fa55555UL,
80 0xc090cf0fUL, 0x3f811115UL, 0x55548ba1UL, 0x3fc55555UL
81 };
82
83 ALIGNED_(16) juint _shifter[] =
84 {
85 0x00000000UL, 0x43380000UL, 0x00000000UL, 0x43380000UL
86 };
87
88 ALIGNED_(16) juint _mmask[] =
89 {
90 0xffffffc0UL, 0x00000000UL, 0xffffffc0UL, 0x00000000UL
91 };
92
93 ALIGNED_(16) juint _bias[] =
94 {
95 0x0000ffc0UL, 0x00000000UL, 0x0000ffc0UL, 0x00000000UL
96 };
97
98 ALIGNED_(16) juint _Tbl_addr[] =
99 {
100 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x0e03754dUL,
101 0x3cad7bbfUL, 0x3e778060UL, 0x00002c9aUL, 0x3567f613UL, 0x3c8cd252UL,
102 0xd3158574UL, 0x000059b0UL, 0x61e6c861UL, 0x3c60f74eUL, 0x18759bc8UL,
103 0x00008745UL, 0x5d837b6cUL, 0x3c979aa6UL, 0x6cf9890fUL, 0x0000b558UL,
104 0x702f9cd1UL, 0x3c3ebe3dUL, 0x32d3d1a2UL, 0x0000e3ecUL, 0x1e63bcd8UL,
105 0x3ca3516eUL, 0xd0125b50UL, 0x00011301UL, 0x26f0387bUL, 0x3ca4c554UL,
106 0xaea92ddfUL, 0x0001429aUL, 0x62523fb6UL, 0x3ca95153UL, 0x3c7d517aUL,
107 0x000172b8UL, 0x3f1353bfUL, 0x3c8b898cUL, 0xeb6fcb75UL, 0x0001a35bUL,
108 0x3e3a2f5fUL, 0x3c9aecf7UL, 0x3168b9aaUL, 0x0001d487UL, 0x44a6c38dUL,
109 0x3c8a6f41UL, 0x88628cd6UL, 0x0002063bUL, 0xe3a8a894UL, 0x3c968efdUL,
110 0x6e756238UL, 0x0002387aUL, 0x981fe7f2UL, 0x3c80472bUL, 0x65e27cddUL,
111 0x00026b45UL, 0x6d09ab31UL, 0x3c82f7e1UL, 0xf51fdee1UL, 0x00029e9dUL,
112 0x720c0ab3UL, 0x3c8b3782UL, 0xa6e4030bUL, 0x0002d285UL, 0x4db0abb6UL,
113 0x3c834d75UL, 0x0a31b715UL, 0x000306feUL, 0x5dd3f84aUL, 0x3c8fdd39UL,
114 0xb26416ffUL, 0x00033c08UL, 0xcc187d29UL, 0x3ca12f8cUL, 0x373aa9caUL,
115 0x000371a7UL, 0x738b5e8bUL, 0x3ca7d229UL, 0x34e59ff6UL, 0x0003a7dbUL,
116 0xa72a4c6dUL, 0x3c859f48UL, 0x4c123422UL, 0x0003dea6UL, 0x259d9205UL,
117 0x3ca8b846UL, 0x21f72e29UL, 0x0004160aUL, 0x60c2ac12UL, 0x3c4363edUL,
118 0x6061892dUL, 0x00044e08UL, 0xdaa10379UL, 0x3c6ecce1UL, 0xb5c13cd0UL,
119 0x000486a2UL, 0xbb7aafb0UL, 0x3c7690ceUL, 0xd5362a27UL, 0x0004bfdaUL,
120 0x9b282a09UL, 0x3ca083ccUL, 0x769d2ca6UL, 0x0004f9b2UL, 0xc1aae707UL,
121 0x3ca509b0UL, 0x569d4f81UL, 0x0005342bUL, 0x18fdd78eUL, 0x3c933505UL,
122 0x36b527daUL, 0x00056f47UL, 0xe21c5409UL, 0x3c9063e1UL, 0xdd485429UL,
123 0x0005ab07UL, 0x2b64c035UL, 0x3c9432e6UL, 0x15ad2148UL, 0x0005e76fUL,
124 0x99f08c0aUL, 0x3ca01284UL, 0xb03a5584UL, 0x0006247eUL, 0x0073dc06UL,
125 0x3c99f087UL, 0x82552224UL, 0x00066238UL, 0x0da05571UL, 0x3c998d4dUL,
126 0x667f3bccUL, 0x0006a09eUL, 0x86ce4786UL, 0x3ca52bb9UL, 0x3c651a2eUL,
127 0x0006dfb2UL, 0x206f0dabUL, 0x3ca32092UL, 0xe8ec5f73UL, 0x00071f75UL,
128 0x8e17a7a6UL, 0x3ca06122UL, 0x564267c8UL, 0x00075febUL, 0x461e9f86UL,
129 0x3ca244acUL, 0x73eb0186UL, 0x0007a114UL, 0xabd66c55UL, 0x3c65ebe1UL,
130 0x36cf4e62UL, 0x0007e2f3UL, 0xbbff67d0UL, 0x3c96fe9fUL, 0x994cce12UL,
131 0x00082589UL, 0x14c801dfUL, 0x3c951f14UL, 0x9b4492ecUL, 0x000868d9UL,
132 0xc1f0eab4UL, 0x3c8db72fUL, 0x422aa0dbUL, 0x0008ace5UL, 0x59f35f44UL,
133 0x3c7bf683UL, 0x99157736UL, 0x0008f1aeUL, 0x9c06283cUL, 0x3ca360baUL,
134 0xb0cdc5e4UL, 0x00093737UL, 0x20f962aaUL, 0x3c95e8d1UL, 0x9fde4e4fUL,
135 0x00097d82UL, 0x2b91ce27UL, 0x3c71affcUL, 0x82a3f090UL, 0x0009c491UL,
136 0x589a2ebdUL, 0x3c9b6d34UL, 0x7b5de564UL, 0x000a0c66UL, 0x9ab89880UL,
137 0x3c95277cUL, 0xb23e255cUL, 0x000a5503UL, 0x6e735ab3UL, 0x3c846984UL,
138 0x5579fdbfUL, 0x000a9e6bUL, 0x92cb3387UL, 0x3c8c1a77UL, 0x995ad3adUL,
139 0x000ae89fUL, 0xdc2d1d96UL, 0x3ca22466UL, 0xb84f15faUL, 0x000b33a2UL,
140 0xb19505aeUL, 0x3ca1112eUL, 0xf2fb5e46UL, 0x000b7f76UL, 0x0a5fddcdUL,
141 0x3c74ffd7UL, 0x904bc1d2UL, 0x000bcc1eUL, 0x30af0cb3UL, 0x3c736eaeUL,
142 0xdd85529cUL, 0x000c199bUL, 0xd10959acUL, 0x3c84e08fUL, 0x2e57d14bUL,
143 0x000c67f1UL, 0x6c921968UL, 0x3c676b2cUL, 0xdcef9069UL, 0x000cb720UL,
144 0x36df99b3UL, 0x3c937009UL, 0x4a07897bUL, 0x000d072dUL, 0xa63d07a7UL,
145 0x3c74a385UL, 0xdcfba487UL, 0x000d5818UL, 0xd5c192acUL, 0x3c8e5a50UL,
146 0x03db3285UL, 0x000da9e6UL, 0x1c4a9792UL, 0x3c98bb73UL, 0x337b9b5eUL,
147 0x000dfc97UL, 0x603a88d3UL, 0x3c74b604UL, 0xe78b3ff6UL, 0x000e502eUL,
148 0x92094926UL, 0x3c916f27UL, 0xa2a490d9UL, 0x000ea4afUL, 0x41aa2008UL,
149 0x3c8ec3bcUL, 0xee615a27UL, 0x000efa1bUL, 0x31d185eeUL, 0x3c8a64a9UL,
150 0x5b6e4540UL, 0x000f5076UL, 0x4d91cd9dUL, 0x3c77893bUL, 0x819e90d8UL,
151 0x000fa7c1UL
152 };
153
154 ALIGNED_(16) juint _ALLONES[] =
155 {
156 0xffffffffUL, 0xffffffffUL, 0xffffffffUL, 0xffffffffUL
157 };
158
159 ALIGNED_(16) juint _ebias[] =
160 {
161 0x00000000UL, 0x3ff00000UL, 0x00000000UL, 0x3ff00000UL
162 };
163
164 ALIGNED_(4) juint _XMAX[] =
165 {
166 0xffffffffUL, 0x7fefffffUL
167 };
168
169 ALIGNED_(4) juint _XMIN[] =
170 {
171 0x00000000UL, 0x00100000UL
172 };
173
174 ALIGNED_(4) juint _INF[] =
175 {
176 0x00000000UL, 0x7ff00000UL
177 };
178
179 ALIGNED_(4) juint _ZERO[] =
180 {
181 0x00000000UL, 0x00000000UL
182 };
183
184 ALIGNED_(4) juint _ONE_val[] =
185 {
186 0x00000000UL, 0x3ff00000UL
187 };
188
189
190 // Registers:
191 // input: xmm0
192 // scratch: xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7
193 // rax, rdx, rcx, tmp - r11
194
195 // Code generated by Intel C compiler for LIBM library
196
197 void MacroAssembler::fast_exp(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp) {
198 Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
199 Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
200 Label L_2TAG_PACKET_8_0_2, L_2TAG_PACKET_9_0_2, L_2TAG_PACKET_10_0_2, L_2TAG_PACKET_11_0_2;
201 Label L_2TAG_PACKET_12_0_2, B1_3, B1_5, start;
202
203 assert_different_registers(tmp, eax, ecx, edx);
204 jmp(start);
205 address cv = (address)_cv;
206 address Shifter = (address)_shifter;
207 address mmask = (address)_mmask;
208 address bias = (address)_bias;
209 address Tbl_addr = (address)_Tbl_addr;
210 address ALLONES = (address)_ALLONES;
211 address ebias = (address)_ebias;
212 address XMAX = (address)_XMAX;
213 address XMIN = (address)_XMIN;
214 address INF = (address)_INF;
215 address ZERO = (address)_ZERO;
216 address ONE_val = (address)_ONE_val;
217
218 bind(start);
219 subq(rsp, 24);
220 movsd(Address(rsp, 8), xmm0);
221 unpcklpd(xmm0, xmm0);
222 movdqu(xmm1, ExternalAddress(cv)); // 0x652b82feUL, 0x40571547UL, 0x652b82feUL, 0x40571547UL
223 movdqu(xmm6, ExternalAddress(Shifter)); // 0x00000000UL, 0x43380000UL, 0x00000000UL, 0x43380000UL
224 movdqu(xmm2, ExternalAddress(16+cv)); // 0xfefa0000UL, 0x3f862e42UL, 0xfefa0000UL, 0x3f862e42UL
225 movdqu(xmm3, ExternalAddress(32+cv)); // 0xbc9e3b3aUL, 0x3d1cf79aUL, 0xbc9e3b3aUL, 0x3d1cf79aUL
226 pextrw(eax, xmm0, 3);
227 andl(eax, 32767);
228 movl(edx, 16527);
229 subl(edx, eax);
230 subl(eax, 15504);
231 orl(edx, eax);
232 cmpl(edx, INT_MIN);
233 jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);
234 mulpd(xmm1, xmm0);
235 addpd(xmm1, xmm6);
236 movapd(xmm7, xmm1);
237 subpd(xmm1, xmm6);
238 mulpd(xmm2, xmm1);
239 movdqu(xmm4, ExternalAddress(64+cv)); // 0xe3289860UL, 0x3f56c15cUL, 0x555b9e25UL, 0x3fa55555UL
240 mulpd(xmm3, xmm1);
241 movdqu(xmm5, ExternalAddress(80+cv)); // 0xc090cf0fUL, 0x3f811115UL, 0x55548ba1UL, 0x3fc55555UL
242 subpd(xmm0, xmm2);
243 movdl(eax, xmm7);
244 movl(ecx, eax);
245 andl(ecx, 63);
246 shll(ecx, 4);
247 sarl(eax, 6);
248 movl(edx, eax);
249 movdqu(xmm6, ExternalAddress(mmask)); // 0xffffffc0UL, 0x00000000UL, 0xffffffc0UL, 0x00000000UL
250 pand(xmm7, xmm6);
251 movdqu(xmm6, ExternalAddress(bias)); // 0x0000ffc0UL, 0x00000000UL, 0x0000ffc0UL, 0x00000000UL
252 paddq(xmm7, xmm6);
253 psllq(xmm7, 46);
254 subpd(xmm0, xmm3);
255 lea(tmp, ExternalAddress(Tbl_addr));
256 movdqu(xmm2, Address(ecx,tmp));
257 mulpd(xmm4, xmm0);
258 movapd(xmm6, xmm0);
259 movapd(xmm1, xmm0);
260 mulpd(xmm6, xmm6);
261 mulpd(xmm0, xmm6);
262 addpd(xmm5, xmm4);
263 mulsd(xmm0, xmm6);
264 mulpd(xmm6, ExternalAddress(48+cv)); // 0xfffffffeUL, 0x3fdfffffUL, 0xfffffffeUL, 0x3fdfffffUL
265 addsd(xmm1, xmm2);
266 unpckhpd(xmm2, xmm2);
267 mulpd(xmm0, xmm5);
268 addsd(xmm1, xmm0);
269 por(xmm2, xmm7);
270 unpckhpd(xmm0, xmm0);
271 addsd(xmm0, xmm1);
272 addsd(xmm0, xmm6);
273 addl(edx, 894);
274 cmpl(edx, 1916);
275 jcc (Assembler::above, L_2TAG_PACKET_1_0_2);
276 mulsd(xmm0, xmm2);
277 addsd(xmm0, xmm2);
278 jmp (B1_5);
279
280 bind(L_2TAG_PACKET_1_0_2);
281 xorpd(xmm3, xmm3);
282 movdqu(xmm4, ExternalAddress(ALLONES)); // 0xffffffffUL, 0xffffffffUL, 0xffffffffUL, 0xffffffffUL
283 movl(edx, -1022);
284 subl(edx, eax);
285 movdl(xmm5, edx);
286 psllq(xmm4, xmm5);
287 movl(ecx, eax);
288 sarl(eax, 1);
289 pinsrw(xmm3, eax, 3);
290 movdqu(xmm6, ExternalAddress(ebias)); // 0x00000000UL, 0x3ff00000UL, 0x00000000UL, 0x3ff00000UL
291 psllq(xmm3, 4);
292 psubd(xmm2, xmm3);
293 mulsd(xmm0, xmm2);
294 cmpl(edx, 52);
295 jcc(Assembler::greater, L_2TAG_PACKET_2_0_2);
296 pand(xmm4, xmm2);
297 paddd(xmm3, xmm6);
298 subsd(xmm2, xmm4);
299 addsd(xmm0, xmm2);
300 cmpl(ecx, 1023);
301 jcc(Assembler::greaterEqual, L_2TAG_PACKET_3_0_2);
302 pextrw(ecx, xmm0, 3);
303 andl(ecx, 32768);
304 orl(edx, ecx);
305 cmpl(edx, 0);
306 jcc(Assembler::equal, L_2TAG_PACKET_4_0_2);
307 movapd(xmm6, xmm0);
308 addsd(xmm0, xmm4);
309 mulsd(xmm0, xmm3);
310 pextrw(ecx, xmm0, 3);
311 andl(ecx, 32752);
312 cmpl(ecx, 0);
313 jcc(Assembler::equal, L_2TAG_PACKET_5_0_2);
314 jmp(B1_5);
315
316 bind(L_2TAG_PACKET_5_0_2);
317 mulsd(xmm6, xmm3);
318 mulsd(xmm4, xmm3);
319 movdqu(xmm0, xmm6);
320 pxor(xmm6, xmm4);
321 psrad(xmm6, 31);
322 pshufd(xmm6, xmm6, 85);
323 psllq(xmm0, 1);
324 psrlq(xmm0, 1);
325 pxor(xmm0, xmm6);
326 psrlq(xmm6, 63);
327 paddq(xmm0, xmm6);
328 paddq(xmm0, xmm4);
329 movl(Address(rsp,0), 15);
330 jmp(L_2TAG_PACKET_6_0_2);
331
332 bind(L_2TAG_PACKET_4_0_2);
333 addsd(xmm0, xmm4);
334 mulsd(xmm0, xmm3);
335 jmp(B1_5);
336
337 bind(L_2TAG_PACKET_3_0_2);
338 addsd(xmm0, xmm4);
339 mulsd(xmm0, xmm3);
340 pextrw(ecx, xmm0, 3);
341 andl(ecx, 32752);
342 cmpl(ecx, 32752);
343 jcc(Assembler::aboveEqual, L_2TAG_PACKET_7_0_2);
344 jmp(B1_5);
345
346 bind(L_2TAG_PACKET_2_0_2);
347 paddd(xmm3, xmm6);
348 addpd(xmm0, xmm2);
349 mulsd(xmm0, xmm3);
350 movl(Address(rsp,0), 15);
351 jmp(L_2TAG_PACKET_6_0_2);
352
353 bind(L_2TAG_PACKET_8_0_2);
354 cmpl(eax, 2146435072);
355 jcc(Assembler::aboveEqual, L_2TAG_PACKET_9_0_2);
356 movl(eax, Address(rsp,12));
357 cmpl(eax, INT_MIN);
358 jcc(Assembler::aboveEqual, L_2TAG_PACKET_10_0_2);
359 movsd(xmm0, ExternalAddress(XMAX)); // 0xffffffffUL, 0x7fefffffUL
360 mulsd(xmm0, xmm0);
361
362 bind(L_2TAG_PACKET_7_0_2);
363 movl(Address(rsp,0), 14);
364 jmp(L_2TAG_PACKET_6_0_2);
365
366 bind(L_2TAG_PACKET_10_0_2);
367 movsd(xmm0, ExternalAddress(XMIN)); // 0x00000000UL, 0x00100000UL
368 mulsd(xmm0, xmm0);
369 movl(Address(rsp,0), 15);
370 jmp(L_2TAG_PACKET_6_0_2);
371
372 bind(L_2TAG_PACKET_9_0_2);
373 movl(edx, Address(rsp,8));
374 cmpl(eax, 2146435072);
375 jcc(Assembler::above, L_2TAG_PACKET_11_0_2);
376 cmpl(edx, 0);
377 jcc(Assembler::notEqual, L_2TAG_PACKET_11_0_2);
378 movl(eax, Address(rsp,12));
379 cmpl(eax, 2146435072);
380 jcc(Assembler::notEqual, L_2TAG_PACKET_12_0_2);
381 movsd(xmm0, ExternalAddress(INF)); // 0x00000000UL, 0x7ff00000UL
382 jmp(B1_5);
383
384 bind(L_2TAG_PACKET_12_0_2);
385 movsd(xmm0, ExternalAddress(ZERO)); // 0x00000000UL, 0x00000000UL
386 jmp(B1_5);
387
388 bind(L_2TAG_PACKET_11_0_2);
389 movsd(xmm0, Address(rsp, 8));
390 addsd(xmm0, xmm0);
391 jmp(B1_5);
392
393 bind(L_2TAG_PACKET_0_0_2);
394 movl(eax, Address(rsp, 12));
395 andl(eax, 2147483647);
396 cmpl(eax, 1083179008);
397 jcc(Assembler::aboveEqual, L_2TAG_PACKET_8_0_2);
398 movsd(Address(rsp, 8), xmm0);
399 addsd(xmm0, ExternalAddress(ONE_val)); // 0x00000000UL, 0x3ff00000UL
400 jmp(B1_5);
401
402 bind(L_2TAG_PACKET_6_0_2);
403 movq(Address(rsp, 16), xmm0);
404
405 bind(B1_3);
406 movq(xmm0, Address(rsp, 16));
407
408 bind(B1_5);
409 addq(rsp, 24);
410 }
411
412 #endif
413
414 #ifndef _LP64
415
416 ALIGNED_(16) juint _static_const_table[] =
417 {
418 0x00000000UL, 0xfff00000UL, 0x00000000UL, 0xfff00000UL, 0xffffffc0UL,
419 0x00000000UL, 0xffffffc0UL, 0x00000000UL, 0x0000ffc0UL, 0x00000000UL,
420 0x0000ffc0UL, 0x00000000UL, 0x00000000UL, 0x43380000UL, 0x00000000UL,
421 0x43380000UL, 0x652b82feUL, 0x40571547UL, 0x652b82feUL, 0x40571547UL,
422 0xfefa0000UL, 0x3f862e42UL, 0xfefa0000UL, 0x3f862e42UL, 0xbc9e3b3aUL,
423 0x3d1cf79aUL, 0xbc9e3b3aUL, 0x3d1cf79aUL, 0xfffffffeUL, 0x3fdfffffUL,
424 0xfffffffeUL, 0x3fdfffffUL, 0xe3289860UL, 0x3f56c15cUL, 0x555b9e25UL,
425 0x3fa55555UL, 0xc090cf0fUL, 0x3f811115UL, 0x55548ba1UL, 0x3fc55555UL,
426 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x0e03754dUL,
427 0x3cad7bbfUL, 0x3e778060UL, 0x00002c9aUL, 0x3567f613UL, 0x3c8cd252UL,
428 0xd3158574UL, 0x000059b0UL, 0x61e6c861UL, 0x3c60f74eUL, 0x18759bc8UL,
429 0x00008745UL, 0x5d837b6cUL, 0x3c979aa6UL, 0x6cf9890fUL, 0x0000b558UL,
430 0x702f9cd1UL, 0x3c3ebe3dUL, 0x32d3d1a2UL, 0x0000e3ecUL, 0x1e63bcd8UL,
431 0x3ca3516eUL, 0xd0125b50UL, 0x00011301UL, 0x26f0387bUL, 0x3ca4c554UL,
432 0xaea92ddfUL, 0x0001429aUL, 0x62523fb6UL, 0x3ca95153UL, 0x3c7d517aUL,
433 0x000172b8UL, 0x3f1353bfUL, 0x3c8b898cUL, 0xeb6fcb75UL, 0x0001a35bUL,
434 0x3e3a2f5fUL, 0x3c9aecf7UL, 0x3168b9aaUL, 0x0001d487UL, 0x44a6c38dUL,
435 0x3c8a6f41UL, 0x88628cd6UL, 0x0002063bUL, 0xe3a8a894UL, 0x3c968efdUL,
436 0x6e756238UL, 0x0002387aUL, 0x981fe7f2UL, 0x3c80472bUL, 0x65e27cddUL,
437 0x00026b45UL, 0x6d09ab31UL, 0x3c82f7e1UL, 0xf51fdee1UL, 0x00029e9dUL,
438 0x720c0ab3UL, 0x3c8b3782UL, 0xa6e4030bUL, 0x0002d285UL, 0x4db0abb6UL,
439 0x3c834d75UL, 0x0a31b715UL, 0x000306feUL, 0x5dd3f84aUL, 0x3c8fdd39UL,
440 0xb26416ffUL, 0x00033c08UL, 0xcc187d29UL, 0x3ca12f8cUL, 0x373aa9caUL,
441 0x000371a7UL, 0x738b5e8bUL, 0x3ca7d229UL, 0x34e59ff6UL, 0x0003a7dbUL,
442 0xa72a4c6dUL, 0x3c859f48UL, 0x4c123422UL, 0x0003dea6UL, 0x259d9205UL,
443 0x3ca8b846UL, 0x21f72e29UL, 0x0004160aUL, 0x60c2ac12UL, 0x3c4363edUL,
444 0x6061892dUL, 0x00044e08UL, 0xdaa10379UL, 0x3c6ecce1UL, 0xb5c13cd0UL,
445 0x000486a2UL, 0xbb7aafb0UL, 0x3c7690ceUL, 0xd5362a27UL, 0x0004bfdaUL,
446 0x9b282a09UL, 0x3ca083ccUL, 0x769d2ca6UL, 0x0004f9b2UL, 0xc1aae707UL,
447 0x3ca509b0UL, 0x569d4f81UL, 0x0005342bUL, 0x18fdd78eUL, 0x3c933505UL,
448 0x36b527daUL, 0x00056f47UL, 0xe21c5409UL, 0x3c9063e1UL, 0xdd485429UL,
449 0x0005ab07UL, 0x2b64c035UL, 0x3c9432e6UL, 0x15ad2148UL, 0x0005e76fUL,
450 0x99f08c0aUL, 0x3ca01284UL, 0xb03a5584UL, 0x0006247eUL, 0x0073dc06UL,
451 0x3c99f087UL, 0x82552224UL, 0x00066238UL, 0x0da05571UL, 0x3c998d4dUL,
452 0x667f3bccUL, 0x0006a09eUL, 0x86ce4786UL, 0x3ca52bb9UL, 0x3c651a2eUL,
453 0x0006dfb2UL, 0x206f0dabUL, 0x3ca32092UL, 0xe8ec5f73UL, 0x00071f75UL,
454 0x8e17a7a6UL, 0x3ca06122UL, 0x564267c8UL, 0x00075febUL, 0x461e9f86UL,
455 0x3ca244acUL, 0x73eb0186UL, 0x0007a114UL, 0xabd66c55UL, 0x3c65ebe1UL,
456 0x36cf4e62UL, 0x0007e2f3UL, 0xbbff67d0UL, 0x3c96fe9fUL, 0x994cce12UL,
457 0x00082589UL, 0x14c801dfUL, 0x3c951f14UL, 0x9b4492ecUL, 0x000868d9UL,
458 0xc1f0eab4UL, 0x3c8db72fUL, 0x422aa0dbUL, 0x0008ace5UL, 0x59f35f44UL,
459 0x3c7bf683UL, 0x99157736UL, 0x0008f1aeUL, 0x9c06283cUL, 0x3ca360baUL,
460 0xb0cdc5e4UL, 0x00093737UL, 0x20f962aaUL, 0x3c95e8d1UL, 0x9fde4e4fUL,
461 0x00097d82UL, 0x2b91ce27UL, 0x3c71affcUL, 0x82a3f090UL, 0x0009c491UL,
462 0x589a2ebdUL, 0x3c9b6d34UL, 0x7b5de564UL, 0x000a0c66UL, 0x9ab89880UL,
463 0x3c95277cUL, 0xb23e255cUL, 0x000a5503UL, 0x6e735ab3UL, 0x3c846984UL,
464 0x5579fdbfUL, 0x000a9e6bUL, 0x92cb3387UL, 0x3c8c1a77UL, 0x995ad3adUL,
465 0x000ae89fUL, 0xdc2d1d96UL, 0x3ca22466UL, 0xb84f15faUL, 0x000b33a2UL,
466 0xb19505aeUL, 0x3ca1112eUL, 0xf2fb5e46UL, 0x000b7f76UL, 0x0a5fddcdUL,
467 0x3c74ffd7UL, 0x904bc1d2UL, 0x000bcc1eUL, 0x30af0cb3UL, 0x3c736eaeUL,
468 0xdd85529cUL, 0x000c199bUL, 0xd10959acUL, 0x3c84e08fUL, 0x2e57d14bUL,
469 0x000c67f1UL, 0x6c921968UL, 0x3c676b2cUL, 0xdcef9069UL, 0x000cb720UL,
470 0x36df99b3UL, 0x3c937009UL, 0x4a07897bUL, 0x000d072dUL, 0xa63d07a7UL,
471 0x3c74a385UL, 0xdcfba487UL, 0x000d5818UL, 0xd5c192acUL, 0x3c8e5a50UL,
472 0x03db3285UL, 0x000da9e6UL, 0x1c4a9792UL, 0x3c98bb73UL, 0x337b9b5eUL,
473 0x000dfc97UL, 0x603a88d3UL, 0x3c74b604UL, 0xe78b3ff6UL, 0x000e502eUL,
474 0x92094926UL, 0x3c916f27UL, 0xa2a490d9UL, 0x000ea4afUL, 0x41aa2008UL,
475 0x3c8ec3bcUL, 0xee615a27UL, 0x000efa1bUL, 0x31d185eeUL, 0x3c8a64a9UL,
476 0x5b6e4540UL, 0x000f5076UL, 0x4d91cd9dUL, 0x3c77893bUL, 0x819e90d8UL,
477 0x000fa7c1UL, 0x00000000UL, 0x3ff00000UL, 0x00000000UL, 0x7ff00000UL,
478 0x00000000UL, 0x00000000UL, 0xffffffffUL, 0x7fefffffUL, 0x00000000UL,
479 0x00100000UL
480 };
481
482 //registers,
483 // input: (rbp + 8)
484 // scratch: xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7
485 // rax, rdx, rcx, rbx (tmp)
486
487 // Code generated by Intel C compiler for LIBM library
488
489 void MacroAssembler::fast_exp(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp) {
490 Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
491 Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
492 Label L_2TAG_PACKET_8_0_2, L_2TAG_PACKET_9_0_2, L_2TAG_PACKET_10_0_2, L_2TAG_PACKET_11_0_2;
493 Label L_2TAG_PACKET_12_0_2, L_2TAG_PACKET_13_0_2, B1_3, B1_5, start;
494
495 assert_different_registers(tmp, eax, ecx, edx);
496 jmp(start);
497 address static_const_table = (address)_static_const_table;
498
499 bind(start);
500 subl(rsp, 120);
501 movl(Address(rsp, 64), tmp);
502 lea(tmp, ExternalAddress(static_const_table));
503 movdqu(xmm0, Address(rsp, 128));
504 unpcklpd(xmm0, xmm0);
505 movdqu(xmm1, Address(tmp, 64)); // 0x652b82feUL, 0x40571547UL, 0x652b82feUL, 0x40571547UL
506 movdqu(xmm6, Address(tmp, 48)); // 0x00000000UL, 0x43380000UL, 0x00000000UL, 0x43380000UL
507 movdqu(xmm2, Address(tmp, 80)); // 0xfefa0000UL, 0x3f862e42UL, 0xfefa0000UL, 0x3f862e42UL
508 movdqu(xmm3, Address(tmp, 96)); // 0xbc9e3b3aUL, 0x3d1cf79aUL, 0xbc9e3b3aUL, 0x3d1cf79aUL
509 pextrw(eax, xmm0, 3);
510 andl(eax, 32767);
511 movl(edx, 16527);
512 subl(edx, eax);
513 subl(eax, 15504);
514 orl(edx, eax);
515 cmpl(edx, INT_MIN);
516 jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);
517 mulpd(xmm1, xmm0);
518 addpd(xmm1, xmm6);
519 movapd(xmm7, xmm1);
520 subpd(xmm1, xmm6);
521 mulpd(xmm2, xmm1);
522 movdqu(xmm4, Address(tmp, 128)); // 0xe3289860UL, 0x3f56c15cUL, 0x555b9e25UL, 0x3fa55555UL
523 mulpd(xmm3, xmm1);
524 movdqu(xmm5, Address(tmp, 144)); // 0xc090cf0fUL, 0x3f811115UL, 0x55548ba1UL, 0x3fc55555UL
525 subpd(xmm0, xmm2);
526 movdl(eax, xmm7);
527 movl(ecx, eax);
528 andl(ecx, 63);
529 shll(ecx, 4);
530 sarl(eax, 6);
531 movl(edx, eax);
532 movdqu(xmm6, Address(tmp, 16)); // 0xffffffc0UL, 0x00000000UL, 0xffffffc0UL, 0x00000000UL
533 pand(xmm7, xmm6);
534 movdqu(xmm6, Address(tmp, 32)); // 0x0000ffc0UL, 0x00000000UL, 0x0000ffc0UL, 0x00000000UL
535 paddq(xmm7, xmm6);
536 psllq(xmm7, 46);
537 subpd(xmm0, xmm3);
538 movdqu(xmm2, Address(tmp, ecx, Address::times_1, 160));
539 mulpd(xmm4, xmm0);
540 movapd(xmm6, xmm0);
541 movapd(xmm1, xmm0);
542 mulpd(xmm6, xmm6);
543 mulpd(xmm0, xmm6);
544 addpd(xmm5, xmm4);
545 mulsd(xmm0, xmm6);
546 mulpd(xmm6, Address(tmp, 112)); // 0xfffffffeUL, 0x3fdfffffUL, 0xfffffffeUL, 0x3fdfffffUL
547 addsd(xmm1, xmm2);
548 unpckhpd(xmm2, xmm2);
549 mulpd(xmm0, xmm5);
550 addsd(xmm1, xmm0);
551 por(xmm2, xmm7);
552 unpckhpd(xmm0, xmm0);
553 addsd(xmm0, xmm1);
554 addsd(xmm0, xmm6);
555 addl(edx, 894);
556 cmpl(edx, 1916);
557 jcc (Assembler::above, L_2TAG_PACKET_1_0_2);
558 mulsd(xmm0, xmm2);
559 addsd(xmm0, xmm2);
560 jmp(L_2TAG_PACKET_2_0_2);
561
562 bind(L_2TAG_PACKET_1_0_2);
563 fnstcw(Address(rsp, 24));
564 movzwl(edx, Address(rsp, 24));
565 orl(edx, 768);
566 movw(Address(rsp, 28), edx);
567 fldcw(Address(rsp, 28));
568 movl(edx, eax);
569 sarl(eax, 1);
570 subl(edx, eax);
571 movdqu(xmm6, Address(tmp, 0)); // 0x00000000UL, 0xfff00000UL, 0x00000000UL, 0xfff00000UL
572 pandn(xmm6, xmm2);
573 addl(eax, 1023);
574 movdl(xmm3, eax);
575 psllq(xmm3, 52);
576 por(xmm6, xmm3);
577 addl(edx, 1023);
578 movdl(xmm4, edx);
579 psllq(xmm4, 52);
580 movsd(Address(rsp, 8), xmm0);
581 fld_d(Address(rsp, 8));
582 movsd(Address(rsp, 16), xmm6);
583 fld_d(Address(rsp, 16));
584 fmula(1);
585 faddp(1);
586 movsd(Address(rsp, 8), xmm4);
587 fld_d(Address(rsp, 8));
588 fmulp(1);
589 fstp_d(Address(rsp, 8));
590 movsd(xmm0,Address(rsp, 8));
591 fldcw(Address(rsp, 24));
592 pextrw(ecx, xmm0, 3);
593 andl(ecx, 32752);
594 cmpl(ecx, 32752);
595 jcc(Assembler::greaterEqual, L_2TAG_PACKET_3_0_2);
596 cmpl(ecx, 0);
597 jcc(Assembler::equal, L_2TAG_PACKET_4_0_2);
598 jmp(L_2TAG_PACKET_2_0_2);
599 cmpl(ecx, INT_MIN);
600 jcc(Assembler::less, L_2TAG_PACKET_3_0_2);
601 cmpl(ecx, -1064950997);
602 jcc(Assembler::less, L_2TAG_PACKET_2_0_2);
603 jcc(Assembler::greater, L_2TAG_PACKET_4_0_2);
604 movl(edx, Address(rsp, 128));
605 cmpl(edx ,-17155601);
606 jcc(Assembler::less, L_2TAG_PACKET_2_0_2);
607 jmp(L_2TAG_PACKET_4_0_2);
608
609 bind(L_2TAG_PACKET_3_0_2);
610 movl(edx, 14);
611 jmp(L_2TAG_PACKET_5_0_2);
612
613 bind(L_2TAG_PACKET_4_0_2);
614 movl(edx, 15);
615
616 bind(L_2TAG_PACKET_5_0_2);
617 movsd(Address(rsp, 0), xmm0);
618 movsd(xmm0, Address(rsp, 128));
619 fld_d(Address(rsp, 0));
620 jmp(L_2TAG_PACKET_6_0_2);
621
622 bind(L_2TAG_PACKET_7_0_2);
623 cmpl(eax, 2146435072);
624 jcc(Assembler::greaterEqual, L_2TAG_PACKET_8_0_2);
625 movl(eax, Address(rsp, 132));
626 cmpl(eax, INT_MIN);
627 jcc(Assembler::greaterEqual, L_2TAG_PACKET_9_0_2);
628 movsd(xmm0, Address(tmp, 1208)); // 0xffffffffUL, 0x7fefffffUL
629 mulsd(xmm0, xmm0);
630 movl(edx, 14);
631 jmp(L_2TAG_PACKET_5_0_2);
632
633 bind(L_2TAG_PACKET_9_0_2);
634 movsd(xmm0, Address(tmp, 1216));
635 mulsd(xmm0, xmm0);
636 movl(edx, 15);
637 jmp(L_2TAG_PACKET_5_0_2);
638
639 bind(L_2TAG_PACKET_8_0_2);
640 movl(edx, Address(rsp, 128));
641 cmpl(eax, 2146435072);
642 jcc(Assembler::above, L_2TAG_PACKET_10_0_2);
643 cmpl(edx, 0);
644 jcc(Assembler::notEqual, L_2TAG_PACKET_10_0_2);
645 movl(eax, Address(rsp, 132));
646 cmpl(eax, 2146435072);
647 jcc(Assembler::notEqual, L_2TAG_PACKET_11_0_2);
648 movsd(xmm0, Address(tmp, 1192)); // 0x00000000UL, 0x7ff00000UL
649 jmp(L_2TAG_PACKET_2_0_2);
650
651 bind(L_2TAG_PACKET_11_0_2);
652 movsd(xmm0, Address(tmp, 1200)); // 0x00000000UL, 0x00000000UL
653 jmp(L_2TAG_PACKET_2_0_2);
654
655 bind(L_2TAG_PACKET_10_0_2);
656 movsd(xmm0, Address(rsp, 128));
657 addsd(xmm0, xmm0);
658 jmp(L_2TAG_PACKET_2_0_2);
659
660 bind(L_2TAG_PACKET_0_0_2);
661 movl(eax, Address(rsp, 132));
662 andl(eax, 2147483647);
663 cmpl(eax, 1083179008);
664 jcc(Assembler::aboveEqual, L_2TAG_PACKET_7_0_2);
665 movsd(xmm0, Address(rsp, 128));
666 addsd(xmm0, Address(tmp, 1184)); // 0x00000000UL, 0x3ff00000UL
667 jmp(L_2TAG_PACKET_2_0_2);
668
669 bind(L_2TAG_PACKET_2_0_2);
670 movsd(Address(rsp, 48), xmm0);
671 fld_d(Address(rsp, 48));
672
673 bind(L_2TAG_PACKET_6_0_2);
674 movl(tmp, Address(rsp, 64));
675 }
676
677 #endif
678
679 /******************************************************************************/
680 // ALGORITHM DESCRIPTION - LOG()
681 // ---------------------
682 //
683 // x=2^k * mx, mx in [1,2)
684 //
685 // Get B~1/mx based on the output of rcpss instruction (B0)
686 // B = int((B0*2^7+0.5))/2^7
687 //
688 // Reduced argument: r=B*mx-1.0 (computed accurately in high and low parts)
689 //
690 // Result: k*log(2) - log(B) + p(r) if |x-1| >= small value (2^-6) and
691 // p(r) is a degree 7 polynomial
692 // -log(B) read from data table (high, low parts)
693 // Result is formed from high and low parts
694 //
695 // Special cases:
696 // log(NaN) = quiet NaN, and raise invalid exception
697 // log(+INF) = that INF
698 // log(0) = -INF with divide-by-zero exception raised
699 // log(1) = +0
700 // log(x) = NaN with invalid exception raised if x < -0, including -INF
701 //
702 /******************************************************************************/
703
704 #ifdef _LP64
705
706 ALIGNED_(16) juint _L_tbl[] =
707 {
708 0xfefa3800UL, 0x3fe62e42UL, 0x93c76730UL, 0x3d2ef357UL, 0xaa241800UL,
709 0x3fe5ee82UL, 0x0cda46beUL, 0x3d220238UL, 0x5c364800UL, 0x3fe5af40UL,
710 0xac10c9fbUL, 0x3d2dfa63UL, 0x26bb8c00UL, 0x3fe5707aUL, 0xff3303ddUL,
711 0x3d09980bUL, 0x26867800UL, 0x3fe5322eUL, 0x5d257531UL, 0x3d05ccc4UL,
712 0x835a5000UL, 0x3fe4f45aUL, 0x6d93b8fbUL, 0xbd2e6c51UL, 0x6f970c00UL,
713 0x3fe4b6fdUL, 0xed4c541cUL, 0x3cef7115UL, 0x27e8a400UL, 0x3fe47a15UL,
714 0xf94d60aaUL, 0xbd22cb6aUL, 0xf2f92400UL, 0x3fe43d9fUL, 0x481051f7UL,
715 0xbcfd984fUL, 0x2125cc00UL, 0x3fe4019cUL, 0x30f0c74cUL, 0xbd26ce79UL,
716 0x0c36c000UL, 0x3fe3c608UL, 0x7cfe13c2UL, 0xbd02b736UL, 0x17197800UL,
717 0x3fe38ae2UL, 0xbb5569a4UL, 0xbd218b7aUL, 0xad9d8c00UL, 0x3fe35028UL,
718 0x9527e6acUL, 0x3d10b83fUL, 0x44340800UL, 0x3fe315daUL, 0xc5a0ed9cUL,
719 0xbd274e93UL, 0x57b0e000UL, 0x3fe2dbf5UL, 0x07b9dc11UL, 0xbd17a6e5UL,
720 0x6d0ec000UL, 0x3fe2a278UL, 0xe797882dUL, 0x3d206d2bUL, 0x1134dc00UL,
721 0x3fe26962UL, 0x05226250UL, 0xbd0b61f1UL, 0xd8bebc00UL, 0x3fe230b0UL,
722 0x6e48667bUL, 0x3d12fc06UL, 0x5fc61800UL, 0x3fe1f863UL, 0xc9fe81d3UL,
723 0xbd2a7242UL, 0x49ae6000UL, 0x3fe1c078UL, 0xed70e667UL, 0x3cccacdeUL,
724 0x40f23c00UL, 0x3fe188eeUL, 0xf8ab4650UL, 0x3d14cc4eUL, 0xf6f29800UL,
725 0x3fe151c3UL, 0xa293ae49UL, 0xbd2edd97UL, 0x23c75c00UL, 0x3fe11af8UL,
726 0xbb9ddcb2UL, 0xbd258647UL, 0x8611cc00UL, 0x3fe0e489UL, 0x07801742UL,
727 0x3d1c2998UL, 0xe2d05400UL, 0x3fe0ae76UL, 0x887e7e27UL, 0x3d1f486bUL,
728 0x0533c400UL, 0x3fe078bfUL, 0x41edf5fdUL, 0x3d268122UL, 0xbe760400UL,
729 0x3fe04360UL, 0xe79539e0UL, 0xbd04c45fUL, 0xe5b20800UL, 0x3fe00e5aUL,
730 0xb1727b1cUL, 0xbd053ba3UL, 0xaf7a4800UL, 0x3fdfb358UL, 0x3c164935UL,
731 0x3d0085faUL, 0xee031800UL, 0x3fdf4aa7UL, 0x6f014a8bUL, 0x3d12cde5UL,
732 0x56b41000UL, 0x3fdee2a1UL, 0x5a470251UL, 0x3d2f27f4UL, 0xc3ddb000UL,
733 0x3fde7b42UL, 0x5372bd08UL, 0xbd246550UL, 0x1a272800UL, 0x3fde148aUL,
734 0x07322938UL, 0xbd1326b2UL, 0x484c9800UL, 0x3fddae75UL, 0x60dc616aUL,
735 0xbd1ea42dUL, 0x46def800UL, 0x3fdd4902UL, 0xe9a767a8UL, 0x3d235bafUL,
736 0x18064800UL, 0x3fdce42fUL, 0x3ec7a6b0UL, 0xbd0797c3UL, 0xc7455800UL,
737 0x3fdc7ff9UL, 0xc15249aeUL, 0xbd29b6ddUL, 0x693fa000UL, 0x3fdc1c60UL,
738 0x7fe8e180UL, 0x3d2cec80UL, 0x1b80e000UL, 0x3fdbb961UL, 0xf40a666dUL,
739 0x3d27d85bUL, 0x04462800UL, 0x3fdb56faUL, 0x2d841995UL, 0x3d109525UL,
740 0x5248d000UL, 0x3fdaf529UL, 0x52774458UL, 0xbd217cc5UL, 0x3c8ad800UL,
741 0x3fda93edUL, 0xbea77a5dUL, 0x3d1e36f2UL, 0x0224f800UL, 0x3fda3344UL,
742 0x7f9d79f5UL, 0x3d23c645UL, 0xea15f000UL, 0x3fd9d32bUL, 0x10d0c0b0UL,
743 0xbd26279eUL, 0x43135800UL, 0x3fd973a3UL, 0xa502d9f0UL, 0xbd152313UL,
744 0x635bf800UL, 0x3fd914a8UL, 0x2ee6307dUL, 0xbd1766b5UL, 0xa88b3000UL,
745 0x3fd8b639UL, 0xe5e70470UL, 0xbd205ae1UL, 0x776dc800UL, 0x3fd85855UL,
746 0x3333778aUL, 0x3d2fd56fUL, 0x3bd81800UL, 0x3fd7fafaUL, 0xc812566aUL,
747 0xbd272090UL, 0x687cf800UL, 0x3fd79e26UL, 0x2efd1778UL, 0x3d29ec7dUL,
748 0x76c67800UL, 0x3fd741d8UL, 0x49dc60b3UL, 0x3d2d8b09UL, 0xe6af1800UL,
749 0x3fd6e60eUL, 0x7c222d87UL, 0x3d172165UL, 0x3e9c6800UL, 0x3fd68ac8UL,
750 0x2756eba0UL, 0x3d20a0d3UL, 0x0b3ab000UL, 0x3fd63003UL, 0xe731ae00UL,
751 0xbd2db623UL, 0xdf596000UL, 0x3fd5d5bdUL, 0x08a465dcUL, 0xbd0a0b2aUL,
752 0x53c8d000UL, 0x3fd57bf7UL, 0xee5d40efUL, 0x3d1fadedUL, 0x0738a000UL,
753 0x3fd522aeUL, 0x8164c759UL, 0x3d2ebe70UL, 0x9e173000UL, 0x3fd4c9e0UL,
754 0x1b0ad8a4UL, 0xbd2e2089UL, 0xc271c800UL, 0x3fd4718dUL, 0x0967d675UL,
755 0xbd2f27ceUL, 0x23d5e800UL, 0x3fd419b4UL, 0xec90e09dUL, 0x3d08e436UL,
756 0x77333000UL, 0x3fd3c252UL, 0xb606bd5cUL, 0x3d183b54UL, 0x76be1000UL,
757 0x3fd36b67UL, 0xb0f177c8UL, 0x3d116ecdUL, 0xe1d36000UL, 0x3fd314f1UL,
758 0xd3213cb8UL, 0xbd28e27aUL, 0x7cdc9000UL, 0x3fd2bef0UL, 0x4a5004f4UL,
759 0x3d2a9cfaUL, 0x1134d800UL, 0x3fd26962UL, 0xdf5bb3b6UL, 0x3d2c93c1UL,
760 0x6d0eb800UL, 0x3fd21445UL, 0xba46baeaUL, 0x3d0a87deUL, 0x635a6800UL,
761 0x3fd1bf99UL, 0x5147bdb7UL, 0x3d2ca6edUL, 0xcbacf800UL, 0x3fd16b5cUL,
762 0xf7a51681UL, 0x3d2b9acdUL, 0x8227e800UL, 0x3fd1178eUL, 0x63a5f01cUL,
763 0xbd2c210eUL, 0x67616000UL, 0x3fd0c42dUL, 0x163ceae9UL, 0x3d27188bUL,
764 0x604d5800UL, 0x3fd07138UL, 0x16ed4e91UL, 0x3cf89cdbUL, 0x5626c800UL,
765 0x3fd01eaeUL, 0x1485e94aUL, 0xbd16f08cUL, 0x6cb3b000UL, 0x3fcf991cUL,
766 0xca0cdf30UL, 0x3d1bcbecUL, 0xe4dd0000UL, 0x3fcef5adUL, 0x65bb8e11UL,
767 0xbcca2115UL, 0xffe71000UL, 0x3fce530eUL, 0x6041f430UL, 0x3cc21227UL,
768 0xb0d49000UL, 0x3fcdb13dUL, 0xf715b035UL, 0xbd2aff2aUL, 0xf2656000UL,
769 0x3fcd1037UL, 0x75b6f6e4UL, 0xbd084a7eUL, 0xc6f01000UL, 0x3fcc6ffbUL,
770 0xc5962bd2UL, 0xbcf1ec72UL, 0x383be000UL, 0x3fcbd087UL, 0x595412b6UL,
771 0xbd2d4bc4UL, 0x575bd000UL, 0x3fcb31d8UL, 0x4eace1aaUL, 0xbd0c358dUL,
772 0x3c8ae000UL, 0x3fca93edUL, 0x50562169UL, 0xbd287243UL, 0x07089000UL,
773 0x3fc9f6c4UL, 0x6865817aUL, 0x3d29904dUL, 0xdcf70000UL, 0x3fc95a5aUL,
774 0x58a0ff6fUL, 0x3d07f228UL, 0xeb390000UL, 0x3fc8beafUL, 0xaae92cd1UL,
775 0xbd073d54UL, 0x6551a000UL, 0x3fc823c1UL, 0x9a631e83UL, 0x3d1e0ddbUL,
776 0x85445000UL, 0x3fc7898dUL, 0x70914305UL, 0xbd1c6610UL, 0x8b757000UL,
777 0x3fc6f012UL, 0xe59c21e1UL, 0xbd25118dUL, 0xbe8c1000UL, 0x3fc6574eUL,
778 0x2c3c2e78UL, 0x3d19cf8bUL, 0x6b544000UL, 0x3fc5bf40UL, 0xeb68981cUL,
779 0xbd127023UL, 0xe4a1b000UL, 0x3fc527e5UL, 0xe5697dc7UL, 0x3d2633e8UL,
780 0x8333b000UL, 0x3fc4913dUL, 0x54fdb678UL, 0x3d258379UL, 0xa5993000UL,
781 0x3fc3fb45UL, 0x7e6a354dUL, 0xbd2cd1d8UL, 0xb0159000UL, 0x3fc365fcUL,
782 0x234b7289UL, 0x3cc62fa8UL, 0x0c868000UL, 0x3fc2d161UL, 0xcb81b4a1UL,
783 0x3d039d6cUL, 0x2a49c000UL, 0x3fc23d71UL, 0x8fd3df5cUL, 0x3d100d23UL,
784 0x7e23f000UL, 0x3fc1aa2bUL, 0x44389934UL, 0x3d2ca78eUL, 0x8227e000UL,
785 0x3fc1178eUL, 0xce2d07f2UL, 0x3d21ef78UL, 0xb59e4000UL, 0x3fc08598UL,
786 0x7009902cUL, 0xbd27e5ddUL, 0x39dbe000UL, 0x3fbfe891UL, 0x4fa10afdUL,
787 0xbd2534d6UL, 0x830a2000UL, 0x3fbec739UL, 0xafe645e0UL, 0xbd2dc068UL,
788 0x63844000UL, 0x3fbda727UL, 0x1fa71733UL, 0x3d1a8940UL, 0x01bc4000UL,
789 0x3fbc8858UL, 0xc65aacd3UL, 0x3d2646d1UL, 0x8dad6000UL, 0x3fbb6ac8UL,
790 0x2bf768e5UL, 0xbd139080UL, 0x40b1c000UL, 0x3fba4e76UL, 0xb94407c8UL,
791 0xbd0e42b6UL, 0x5d594000UL, 0x3fb9335eUL, 0x3abd47daUL, 0x3d23115cUL,
792 0x2f40e000UL, 0x3fb8197eUL, 0xf96ffdf7UL, 0x3d0f80dcUL, 0x0aeac000UL,
793 0x3fb700d3UL, 0xa99ded32UL, 0x3cec1e8dUL, 0x4d97a000UL, 0x3fb5e95aUL,
794 0x3c5d1d1eUL, 0xbd2c6906UL, 0x5d208000UL, 0x3fb4d311UL, 0x82f4e1efUL,
795 0xbcf53a25UL, 0xa7d1e000UL, 0x3fb3bdf5UL, 0xa5db4ed7UL, 0x3d2cc85eUL,
796 0xa4472000UL, 0x3fb2aa04UL, 0xae9c697dUL, 0xbd20b6e8UL, 0xd1466000UL,
797 0x3fb1973bUL, 0x560d9e9bUL, 0xbd25325dUL, 0xb59e4000UL, 0x3fb08598UL,
798 0x7009902cUL, 0xbd17e5ddUL, 0xc006c000UL, 0x3faeea31UL, 0x4fc93b7bUL,
799 0xbd0e113eUL, 0xcdddc000UL, 0x3faccb73UL, 0x47d82807UL, 0xbd1a68f2UL,
800 0xd0fb0000UL, 0x3faaaef2UL, 0x353bb42eUL, 0x3d20fc1aUL, 0x149fc000UL,
801 0x3fa894aaUL, 0xd05a267dUL, 0xbd197995UL, 0xf2d4c000UL, 0x3fa67c94UL,
802 0xec19afa2UL, 0xbd029efbUL, 0xd42e0000UL, 0x3fa466aeUL, 0x75bdfd28UL,
803 0xbd2c1673UL, 0x2f8d0000UL, 0x3fa252f3UL, 0xe021b67bUL, 0x3d283e9aUL,
804 0x89e74000UL, 0x3fa0415dUL, 0x5cf1d753UL, 0x3d0111c0UL, 0xec148000UL,
805 0x3f9c63d2UL, 0x3f9eb2f3UL, 0x3d2578c6UL, 0x28c90000UL, 0x3f984925UL,
806 0x325a0c34UL, 0xbd2aa0baUL, 0x25980000UL, 0x3f9432a9UL, 0x928637feUL,
807 0x3d098139UL, 0x58938000UL, 0x3f902056UL, 0x06e2f7d2UL, 0xbd23dc5bUL,
808 0xa3890000UL, 0x3f882448UL, 0xda74f640UL, 0xbd275577UL, 0x75890000UL,
809 0x3f801015UL, 0x999d2be8UL, 0xbd10c76bUL, 0x59580000UL, 0x3f700805UL,
810 0xcb31c67bUL, 0x3d2166afUL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
811 0x80000000UL
812 };
813
814 ALIGNED_(16) juint _log2[] =
815 {
816 0xfefa3800UL, 0x3fa62e42UL, 0x93c76730UL, 0x3ceef357UL
817 };
818
819 ALIGNED_(16) juint _coeff[] =
820 {
821 0x92492492UL, 0x3fc24924UL, 0x00000000UL, 0xbfd00000UL, 0x3d6fb175UL,
822 0xbfc5555eUL, 0x55555555UL, 0x3fd55555UL, 0x9999999aUL, 0x3fc99999UL,
823 0x00000000UL, 0xbfe00000UL
824 };
825
826 //registers,
827 // input: xmm0
828 // scratch: xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7
829 // rax, rdx, rcx, r8, r11
830
831 void MacroAssembler::fast_log(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp1, Register tmp2) {
832 Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
833 Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
834 Label L_2TAG_PACKET_8_0_2;
835 Label L_2TAG_PACKET_12_0_2, L_2TAG_PACKET_13_0_2, B1_3, B1_5, start;
836
837 assert_different_registers(tmp1, tmp2, eax, ecx, edx);
838 jmp(start);
839 address L_tbl = (address)_L_tbl;
840 address log2 = (address)_log2;
841 address coeff = (address)_coeff;
842
843 bind(start);
844 subq(rsp, 24);
845 movsd(Address(rsp, 0), xmm0);
846 mov64(rax, 0x3ff0000000000000);
847 movdq(xmm2, rax);
848 mov64(rdx, 0x77f0000000000000);
849 movdq(xmm3, rdx);
850 movl(ecx, 32768);
851 movdl(xmm4, rcx);
852 mov64(tmp1, 0xffffe00000000000);
853 movdq(xmm5, tmp1);
854 movdqu(xmm1, xmm0);
855 pextrw(eax, xmm0, 3);
856 por(xmm0, xmm2);
857 movl(ecx, 16352);
858 psrlq(xmm0, 27);
859 lea(tmp2, ExternalAddress(L_tbl));
860 psrld(xmm0, 2);
861 rcpps(xmm0, xmm0);
862 psllq(xmm1, 12);
863 pshufd(xmm6, xmm5, 228);
864 psrlq(xmm1, 12);
865 subl(eax, 16);
866 cmpl(eax, 32736);
867 jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);
868
869 bind(L_2TAG_PACKET_1_0_2);
870 paddd(xmm0, xmm4);
871 por(xmm1, xmm3);
872 movdl(edx, xmm0);
873 psllq(xmm0, 29);
874 pand(xmm5, xmm1);
875 pand(xmm0, xmm6);
876 subsd(xmm1, xmm5);
877 mulpd(xmm5, xmm0);
878 andl(eax, 32752);
879 subl(eax, ecx);
880 cvtsi2sdl(xmm7, eax);
881 mulsd(xmm1, xmm0);
882 movq(xmm6, ExternalAddress(log2)); // 0xfefa3800UL, 0x3fa62e42UL
883 movdqu(xmm3, ExternalAddress(coeff)); // 0x92492492UL, 0x3fc24924UL, 0x00000000UL, 0xbfd00000UL
884 subsd(xmm5, xmm2);
885 andl(edx, 16711680);
886 shrl(edx, 12);
887 movdqu(xmm0, Address(tmp2, edx));
888 movdqu(xmm4, ExternalAddress(16 + coeff)); // 0x3d6fb175UL, 0xbfc5555eUL, 0x55555555UL, 0x3fd55555UL
889 addsd(xmm1, xmm5);
890 movdqu(xmm2, ExternalAddress(32 + coeff)); // 0x9999999aUL, 0x3fc99999UL, 0x00000000UL, 0xbfe00000UL
891 mulsd(xmm6, xmm7);
892 movddup(xmm5, xmm1);
893 mulsd(xmm7, ExternalAddress(8 + log2)); // 0x93c76730UL, 0x3ceef357UL
894 mulsd(xmm3, xmm1);
895 addsd(xmm0, xmm6);
896 mulpd(xmm4, xmm5);
897 mulpd(xmm5, xmm5);
898 movddup(xmm6, xmm0);
899 addsd(xmm0, xmm1);
900 addpd(xmm4, xmm2);
901 mulpd(xmm3, xmm5);
902 subsd(xmm6, xmm0);
903 mulsd(xmm4, xmm1);
904 pshufd(xmm2, xmm0, 238);
905 addsd(xmm1, xmm6);
906 mulsd(xmm5, xmm5);
907 addsd(xmm7, xmm2);
908 addpd(xmm4, xmm3);
909 addsd(xmm1, xmm7);
910 mulpd(xmm4, xmm5);
911 addsd(xmm1, xmm4);
912 pshufd(xmm5, xmm4, 238);
913 addsd(xmm1, xmm5);
914 addsd(xmm0, xmm1);
915 jmp(B1_5);
916
917 bind(L_2TAG_PACKET_0_0_2);
918 movq(xmm0, Address(rsp, 0));
919 movq(xmm1, Address(rsp, 0));
920 addl(eax, 16);
921 cmpl(eax, 32768);
922 jcc(Assembler::aboveEqual, L_2TAG_PACKET_2_0_2);
923 cmpl(eax, 16);
924 jcc(Assembler::below, L_2TAG_PACKET_3_0_2);
925
926 bind(L_2TAG_PACKET_4_0_2);
927 addsd(xmm0, xmm0);
928 jmp(B1_5);
929
930 bind(L_2TAG_PACKET_5_0_2);
931 jcc(Assembler::above, L_2TAG_PACKET_4_0_2);
932 cmpl(edx, 0);
933 jcc(Assembler::above, L_2TAG_PACKET_4_0_2);
934 jmp(L_2TAG_PACKET_6_0_2);
935
936 bind(L_2TAG_PACKET_3_0_2);
937 xorpd(xmm1, xmm1);
938 addsd(xmm1, xmm0);
939 movdl(edx, xmm1);
940 psrlq(xmm1, 32);
941 movdl(ecx, xmm1);
942 orl(edx, ecx);
943 cmpl(edx, 0);
944 jcc(Assembler::equal, L_2TAG_PACKET_7_0_2);
945 xorpd(xmm1, xmm1);
946 movl(eax, 18416);
947 pinsrw(xmm1, eax, 3);
948 mulsd(xmm0, xmm1);
949 movdqu(xmm1, xmm0);
950 pextrw(eax, xmm0, 3);
951 por(xmm0, xmm2);
952 psrlq(xmm0, 27);
953 movl(ecx, 18416);
954 psrld(xmm0, 2);
955 rcpps(xmm0, xmm0);
956 psllq(xmm1, 12);
957 pshufd(xmm6, xmm5, 228);
958 psrlq(xmm1, 12);
959 jmp(L_2TAG_PACKET_1_0_2);
960
961 bind(L_2TAG_PACKET_2_0_2);
962 movdl(edx, xmm1);
963 psrlq(xmm1, 32);
964 movdl(ecx, xmm1);
965 addl(ecx, ecx);
966 cmpl(ecx, -2097152);
967 jcc(Assembler::aboveEqual, L_2TAG_PACKET_5_0_2);
968 orl(edx, ecx);
969 cmpl(edx, 0);
970 jcc(Assembler::equal, L_2TAG_PACKET_7_0_2);
971
972 bind(L_2TAG_PACKET_6_0_2);
973 xorpd(xmm1, xmm1);
974 xorpd(xmm0, xmm0);
975 movl(eax, 32752);
976 pinsrw(xmm1, eax, 3);
977 mulsd(xmm0, xmm1);
978 movl(Address(rsp, 16), 3);
979 jmp(L_2TAG_PACKET_8_0_2);
980 bind(L_2TAG_PACKET_7_0_2);
981 xorpd(xmm1, xmm1);
982 xorpd(xmm0, xmm0);
983 movl(eax, 49136);
984 pinsrw(xmm0, eax, 3);
985 divsd(xmm0, xmm1);
986 movl(Address(rsp, 16), 2);
987
988 bind(L_2TAG_PACKET_8_0_2);
989 movq(Address(rsp, 8), xmm0);
990
991 bind(B1_3);
992 movq(xmm0, Address(rsp, 8));
993
994 bind(B1_5);
995 addq(rsp, 24);
996 }
997
998 #endif
999
1000 #ifndef _LP64
1001
1002 ALIGNED_(16) juint _static_const_table_log[] =
1003 {
1004 0xfefa3800UL, 0x3fe62e42UL, 0x93c76730UL, 0x3d2ef357UL, 0xaa241800UL,
1005 0x3fe5ee82UL, 0x0cda46beUL, 0x3d220238UL, 0x5c364800UL, 0x3fe5af40UL,
1006 0xac10c9fbUL, 0x3d2dfa63UL, 0x26bb8c00UL, 0x3fe5707aUL, 0xff3303ddUL,
1007 0x3d09980bUL, 0x26867800UL, 0x3fe5322eUL, 0x5d257531UL, 0x3d05ccc4UL,
1008 0x835a5000UL, 0x3fe4f45aUL, 0x6d93b8fbUL, 0xbd2e6c51UL, 0x6f970c00UL,
1009 0x3fe4b6fdUL, 0xed4c541cUL, 0x3cef7115UL, 0x27e8a400UL, 0x3fe47a15UL,
1010 0xf94d60aaUL, 0xbd22cb6aUL, 0xf2f92400UL, 0x3fe43d9fUL, 0x481051f7UL,
1011 0xbcfd984fUL, 0x2125cc00UL, 0x3fe4019cUL, 0x30f0c74cUL, 0xbd26ce79UL,
1012 0x0c36c000UL, 0x3fe3c608UL, 0x7cfe13c2UL, 0xbd02b736UL, 0x17197800UL,
1013 0x3fe38ae2UL, 0xbb5569a4UL, 0xbd218b7aUL, 0xad9d8c00UL, 0x3fe35028UL,
1014 0x9527e6acUL, 0x3d10b83fUL, 0x44340800UL, 0x3fe315daUL, 0xc5a0ed9cUL,
1015 0xbd274e93UL, 0x57b0e000UL, 0x3fe2dbf5UL, 0x07b9dc11UL, 0xbd17a6e5UL,
1016 0x6d0ec000UL, 0x3fe2a278UL, 0xe797882dUL, 0x3d206d2bUL, 0x1134dc00UL,
1017 0x3fe26962UL, 0x05226250UL, 0xbd0b61f1UL, 0xd8bebc00UL, 0x3fe230b0UL,
1018 0x6e48667bUL, 0x3d12fc06UL, 0x5fc61800UL, 0x3fe1f863UL, 0xc9fe81d3UL,
1019 0xbd2a7242UL, 0x49ae6000UL, 0x3fe1c078UL, 0xed70e667UL, 0x3cccacdeUL,
1020 0x40f23c00UL, 0x3fe188eeUL, 0xf8ab4650UL, 0x3d14cc4eUL, 0xf6f29800UL,
1021 0x3fe151c3UL, 0xa293ae49UL, 0xbd2edd97UL, 0x23c75c00UL, 0x3fe11af8UL,
1022 0xbb9ddcb2UL, 0xbd258647UL, 0x8611cc00UL, 0x3fe0e489UL, 0x07801742UL,
1023 0x3d1c2998UL, 0xe2d05400UL, 0x3fe0ae76UL, 0x887e7e27UL, 0x3d1f486bUL,
1024 0x0533c400UL, 0x3fe078bfUL, 0x41edf5fdUL, 0x3d268122UL, 0xbe760400UL,
1025 0x3fe04360UL, 0xe79539e0UL, 0xbd04c45fUL, 0xe5b20800UL, 0x3fe00e5aUL,
1026 0xb1727b1cUL, 0xbd053ba3UL, 0xaf7a4800UL, 0x3fdfb358UL, 0x3c164935UL,
1027 0x3d0085faUL, 0xee031800UL, 0x3fdf4aa7UL, 0x6f014a8bUL, 0x3d12cde5UL,
1028 0x56b41000UL, 0x3fdee2a1UL, 0x5a470251UL, 0x3d2f27f4UL, 0xc3ddb000UL,
1029 0x3fde7b42UL, 0x5372bd08UL, 0xbd246550UL, 0x1a272800UL, 0x3fde148aUL,
1030 0x07322938UL, 0xbd1326b2UL, 0x484c9800UL, 0x3fddae75UL, 0x60dc616aUL,
1031 0xbd1ea42dUL, 0x46def800UL, 0x3fdd4902UL, 0xe9a767a8UL, 0x3d235bafUL,
1032 0x18064800UL, 0x3fdce42fUL, 0x3ec7a6b0UL, 0xbd0797c3UL, 0xc7455800UL,
1033 0x3fdc7ff9UL, 0xc15249aeUL, 0xbd29b6ddUL, 0x693fa000UL, 0x3fdc1c60UL,
1034 0x7fe8e180UL, 0x3d2cec80UL, 0x1b80e000UL, 0x3fdbb961UL, 0xf40a666dUL,
1035 0x3d27d85bUL, 0x04462800UL, 0x3fdb56faUL, 0x2d841995UL, 0x3d109525UL,
1036 0x5248d000UL, 0x3fdaf529UL, 0x52774458UL, 0xbd217cc5UL, 0x3c8ad800UL,
1037 0x3fda93edUL, 0xbea77a5dUL, 0x3d1e36f2UL, 0x0224f800UL, 0x3fda3344UL,
1038 0x7f9d79f5UL, 0x3d23c645UL, 0xea15f000UL, 0x3fd9d32bUL, 0x10d0c0b0UL,
1039 0xbd26279eUL, 0x43135800UL, 0x3fd973a3UL, 0xa502d9f0UL, 0xbd152313UL,
1040 0x635bf800UL, 0x3fd914a8UL, 0x2ee6307dUL, 0xbd1766b5UL, 0xa88b3000UL,
1041 0x3fd8b639UL, 0xe5e70470UL, 0xbd205ae1UL, 0x776dc800UL, 0x3fd85855UL,
1042 0x3333778aUL, 0x3d2fd56fUL, 0x3bd81800UL, 0x3fd7fafaUL, 0xc812566aUL,
1043 0xbd272090UL, 0x687cf800UL, 0x3fd79e26UL, 0x2efd1778UL, 0x3d29ec7dUL,
1044 0x76c67800UL, 0x3fd741d8UL, 0x49dc60b3UL, 0x3d2d8b09UL, 0xe6af1800UL,
1045 0x3fd6e60eUL, 0x7c222d87UL, 0x3d172165UL, 0x3e9c6800UL, 0x3fd68ac8UL,
1046 0x2756eba0UL, 0x3d20a0d3UL, 0x0b3ab000UL, 0x3fd63003UL, 0xe731ae00UL,
1047 0xbd2db623UL, 0xdf596000UL, 0x3fd5d5bdUL, 0x08a465dcUL, 0xbd0a0b2aUL,
1048 0x53c8d000UL, 0x3fd57bf7UL, 0xee5d40efUL, 0x3d1fadedUL, 0x0738a000UL,
1049 0x3fd522aeUL, 0x8164c759UL, 0x3d2ebe70UL, 0x9e173000UL, 0x3fd4c9e0UL,
1050 0x1b0ad8a4UL, 0xbd2e2089UL, 0xc271c800UL, 0x3fd4718dUL, 0x0967d675UL,
1051 0xbd2f27ceUL, 0x23d5e800UL, 0x3fd419b4UL, 0xec90e09dUL, 0x3d08e436UL,
1052 0x77333000UL, 0x3fd3c252UL, 0xb606bd5cUL, 0x3d183b54UL, 0x76be1000UL,
1053 0x3fd36b67UL, 0xb0f177c8UL, 0x3d116ecdUL, 0xe1d36000UL, 0x3fd314f1UL,
1054 0xd3213cb8UL, 0xbd28e27aUL, 0x7cdc9000UL, 0x3fd2bef0UL, 0x4a5004f4UL,
1055 0x3d2a9cfaUL, 0x1134d800UL, 0x3fd26962UL, 0xdf5bb3b6UL, 0x3d2c93c1UL,
1056 0x6d0eb800UL, 0x3fd21445UL, 0xba46baeaUL, 0x3d0a87deUL, 0x635a6800UL,
1057 0x3fd1bf99UL, 0x5147bdb7UL, 0x3d2ca6edUL, 0xcbacf800UL, 0x3fd16b5cUL,
1058 0xf7a51681UL, 0x3d2b9acdUL, 0x8227e800UL, 0x3fd1178eUL, 0x63a5f01cUL,
1059 0xbd2c210eUL, 0x67616000UL, 0x3fd0c42dUL, 0x163ceae9UL, 0x3d27188bUL,
1060 0x604d5800UL, 0x3fd07138UL, 0x16ed4e91UL, 0x3cf89cdbUL, 0x5626c800UL,
1061 0x3fd01eaeUL, 0x1485e94aUL, 0xbd16f08cUL, 0x6cb3b000UL, 0x3fcf991cUL,
1062 0xca0cdf30UL, 0x3d1bcbecUL, 0xe4dd0000UL, 0x3fcef5adUL, 0x65bb8e11UL,
1063 0xbcca2115UL, 0xffe71000UL, 0x3fce530eUL, 0x6041f430UL, 0x3cc21227UL,
1064 0xb0d49000UL, 0x3fcdb13dUL, 0xf715b035UL, 0xbd2aff2aUL, 0xf2656000UL,
1065 0x3fcd1037UL, 0x75b6f6e4UL, 0xbd084a7eUL, 0xc6f01000UL, 0x3fcc6ffbUL,
1066 0xc5962bd2UL, 0xbcf1ec72UL, 0x383be000UL, 0x3fcbd087UL, 0x595412b6UL,
1067 0xbd2d4bc4UL, 0x575bd000UL, 0x3fcb31d8UL, 0x4eace1aaUL, 0xbd0c358dUL,
1068 0x3c8ae000UL, 0x3fca93edUL, 0x50562169UL, 0xbd287243UL, 0x07089000UL,
1069 0x3fc9f6c4UL, 0x6865817aUL, 0x3d29904dUL, 0xdcf70000UL, 0x3fc95a5aUL,
1070 0x58a0ff6fUL, 0x3d07f228UL, 0xeb390000UL, 0x3fc8beafUL, 0xaae92cd1UL,
1071 0xbd073d54UL, 0x6551a000UL, 0x3fc823c1UL, 0x9a631e83UL, 0x3d1e0ddbUL,
1072 0x85445000UL, 0x3fc7898dUL, 0x70914305UL, 0xbd1c6610UL, 0x8b757000UL,
1073 0x3fc6f012UL, 0xe59c21e1UL, 0xbd25118dUL, 0xbe8c1000UL, 0x3fc6574eUL,
1074 0x2c3c2e78UL, 0x3d19cf8bUL, 0x6b544000UL, 0x3fc5bf40UL, 0xeb68981cUL,
1075 0xbd127023UL, 0xe4a1b000UL, 0x3fc527e5UL, 0xe5697dc7UL, 0x3d2633e8UL,
1076 0x8333b000UL, 0x3fc4913dUL, 0x54fdb678UL, 0x3d258379UL, 0xa5993000UL,
1077 0x3fc3fb45UL, 0x7e6a354dUL, 0xbd2cd1d8UL, 0xb0159000UL, 0x3fc365fcUL,
1078 0x234b7289UL, 0x3cc62fa8UL, 0x0c868000UL, 0x3fc2d161UL, 0xcb81b4a1UL,
1079 0x3d039d6cUL, 0x2a49c000UL, 0x3fc23d71UL, 0x8fd3df5cUL, 0x3d100d23UL,
1080 0x7e23f000UL, 0x3fc1aa2bUL, 0x44389934UL, 0x3d2ca78eUL, 0x8227e000UL,
1081 0x3fc1178eUL, 0xce2d07f2UL, 0x3d21ef78UL, 0xb59e4000UL, 0x3fc08598UL,
1082 0x7009902cUL, 0xbd27e5ddUL, 0x39dbe000UL, 0x3fbfe891UL, 0x4fa10afdUL,
1083 0xbd2534d6UL, 0x830a2000UL, 0x3fbec739UL, 0xafe645e0UL, 0xbd2dc068UL,
1084 0x63844000UL, 0x3fbda727UL, 0x1fa71733UL, 0x3d1a8940UL, 0x01bc4000UL,
1085 0x3fbc8858UL, 0xc65aacd3UL, 0x3d2646d1UL, 0x8dad6000UL, 0x3fbb6ac8UL,
1086 0x2bf768e5UL, 0xbd139080UL, 0x40b1c000UL, 0x3fba4e76UL, 0xb94407c8UL,
1087 0xbd0e42b6UL, 0x5d594000UL, 0x3fb9335eUL, 0x3abd47daUL, 0x3d23115cUL,
1088 0x2f40e000UL, 0x3fb8197eUL, 0xf96ffdf7UL, 0x3d0f80dcUL, 0x0aeac000UL,
1089 0x3fb700d3UL, 0xa99ded32UL, 0x3cec1e8dUL, 0x4d97a000UL, 0x3fb5e95aUL,
1090 0x3c5d1d1eUL, 0xbd2c6906UL, 0x5d208000UL, 0x3fb4d311UL, 0x82f4e1efUL,
1091 0xbcf53a25UL, 0xa7d1e000UL, 0x3fb3bdf5UL, 0xa5db4ed7UL, 0x3d2cc85eUL,
1092 0xa4472000UL, 0x3fb2aa04UL, 0xae9c697dUL, 0xbd20b6e8UL, 0xd1466000UL,
1093 0x3fb1973bUL, 0x560d9e9bUL, 0xbd25325dUL, 0xb59e4000UL, 0x3fb08598UL,
1094 0x7009902cUL, 0xbd17e5ddUL, 0xc006c000UL, 0x3faeea31UL, 0x4fc93b7bUL,
1095 0xbd0e113eUL, 0xcdddc000UL, 0x3faccb73UL, 0x47d82807UL, 0xbd1a68f2UL,
1096 0xd0fb0000UL, 0x3faaaef2UL, 0x353bb42eUL, 0x3d20fc1aUL, 0x149fc000UL,
1097 0x3fa894aaUL, 0xd05a267dUL, 0xbd197995UL, 0xf2d4c000UL, 0x3fa67c94UL,
1098 0xec19afa2UL, 0xbd029efbUL, 0xd42e0000UL, 0x3fa466aeUL, 0x75bdfd28UL,
1099 0xbd2c1673UL, 0x2f8d0000UL, 0x3fa252f3UL, 0xe021b67bUL, 0x3d283e9aUL,
1100 0x89e74000UL, 0x3fa0415dUL, 0x5cf1d753UL, 0x3d0111c0UL, 0xec148000UL,
1101 0x3f9c63d2UL, 0x3f9eb2f3UL, 0x3d2578c6UL, 0x28c90000UL, 0x3f984925UL,
1102 0x325a0c34UL, 0xbd2aa0baUL, 0x25980000UL, 0x3f9432a9UL, 0x928637feUL,
1103 0x3d098139UL, 0x58938000UL, 0x3f902056UL, 0x06e2f7d2UL, 0xbd23dc5bUL,
1104 0xa3890000UL, 0x3f882448UL, 0xda74f640UL, 0xbd275577UL, 0x75890000UL,
1105 0x3f801015UL, 0x999d2be8UL, 0xbd10c76bUL, 0x59580000UL, 0x3f700805UL,
1106 0xcb31c67bUL, 0x3d2166afUL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
1107 0x80000000UL, 0xfefa3800UL, 0x3fa62e42UL, 0x93c76730UL, 0x3ceef357UL,
1108 0x92492492UL, 0x3fc24924UL, 0x00000000UL, 0xbfd00000UL, 0x3d6fb175UL,
1109 0xbfc5555eUL, 0x55555555UL, 0x3fd55555UL, 0x9999999aUL, 0x3fc99999UL,
1110 0x00000000UL, 0xbfe00000UL, 0x00000000UL, 0xffffe000UL, 0x00000000UL,
1111 0xffffe000UL
1112 };
1113 //registers,
1114 // input: xmm0
1115 // scratch: xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7
1116 // rax, rdx, rcx, rbx (tmp)
1117
1118 void MacroAssembler::fast_log(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp) {
1119 Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
1120 Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
1121 Label L_2TAG_PACKET_8_0_2, L_2TAG_PACKET_9_0_2;
1122 Label L_2TAG_PACKET_10_0_2, start;
1123
1124 assert_different_registers(tmp, eax, ecx, edx);
1125 jmp(start);
1126 address static_const_table = (address)_static_const_table_log;
1127
1128 bind(start);
1129 subl(rsp, 104);
1130 movl(Address(rsp, 40), tmp);
1131 lea(tmp, ExternalAddress(static_const_table));
1132 xorpd(xmm2, xmm2);
1133 movl(eax, 16368);
1134 pinsrw(xmm2, eax, 3);
1135 xorpd(xmm3, xmm3);
1136 movl(edx, 30704);
1137 pinsrw(xmm3, edx, 3);
1138 movsd(xmm0, Address(rsp, 112));
1139 movapd(xmm1, xmm0);
1140 movl(ecx, 32768);
1141 movdl(xmm4, ecx);
1142 movsd(xmm5, Address(tmp, 2128)); // 0x00000000UL, 0xffffe000UL
1143 pextrw(eax, xmm0, 3);
1144 por(xmm0, xmm2);
1145 psllq(xmm0, 5);
1146 movl(ecx, 16352);
1147 psrlq(xmm0, 34);
1148 rcpss(xmm0, xmm0);
1149 psllq(xmm1, 12);
1150 pshufd(xmm6, xmm5, 228);
1151 psrlq(xmm1, 12);
1152 subl(eax, 16);
1153 cmpl(eax, 32736);
1154 jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);
1155
1156 bind(L_2TAG_PACKET_1_0_2);
1157 paddd(xmm0, xmm4);
1158 por(xmm1, xmm3);
1159 movdl(edx, xmm0);
1160 psllq(xmm0, 29);
1161 pand(xmm5, xmm1);
1162 pand(xmm0, xmm6);
1163 subsd(xmm1, xmm5);
1164 mulpd(xmm5, xmm0);
1165 andl(eax, 32752);
1166 subl(eax, ecx);
1167 cvtsi2sdl(xmm7, eax);
1168 mulsd(xmm1, xmm0);
1169 movsd(xmm6, Address(tmp, 2064)); // 0xfefa3800UL, 0x3fa62e42UL
1170 movdqu(xmm3, Address(tmp, 2080)); // 0x92492492UL, 0x3fc24924UL, 0x00000000UL, 0xbfd00000UL
1171 subsd(xmm5, xmm2);
1172 andl(edx, 16711680);
1173 shrl(edx, 12);
1174 movdqu(xmm0, Address(tmp, edx));
1175 movdqu(xmm4, Address(tmp, 2096)); // 0x3d6fb175UL, 0xbfc5555eUL, 0x55555555UL, 0x3fd55555UL
1176 addsd(xmm1, xmm5);
1177 movdqu(xmm2, Address(tmp, 2112)); // 0x9999999aUL, 0x3fc99999UL, 0x00000000UL, 0xbfe00000UL
1178 mulsd(xmm6, xmm7);
1179 pshufd(xmm5, xmm1, 68);
1180 mulsd(xmm7, Address(tmp, 2072)); // 0x93c76730UL, 0x3ceef357UL, 0x92492492UL, 0x3fc24924UL
1181 mulsd(xmm3, xmm1);
1182 addsd(xmm0, xmm6);
1183 mulpd(xmm4, xmm5);
1184 mulpd(xmm5, xmm5);
1185 pshufd(xmm6, xmm0, 228);
1186 addsd(xmm0, xmm1);
1187 addpd(xmm4, xmm2);
1188 mulpd(xmm3, xmm5);
1189 subsd(xmm6, xmm0);
1190 mulsd(xmm4, xmm1);
1191 pshufd(xmm2, xmm0, 238);
1192 addsd(xmm1, xmm6);
1193 mulsd(xmm5, xmm5);
1194 addsd(xmm7, xmm2);
1195 addpd(xmm4, xmm3);
1196 addsd(xmm1, xmm7);
1197 mulpd(xmm4, xmm5);
1198 addsd(xmm1, xmm4);
1199 pshufd(xmm5, xmm4, 238);
1200 addsd(xmm1, xmm5);
1201 addsd(xmm0, xmm1);
1202 jmp(L_2TAG_PACKET_2_0_2);
1203
1204 bind(L_2TAG_PACKET_0_0_2);
1205 movsd(xmm0, Address(rsp, 112));
1206 movdqu(xmm1, xmm0);
1207 addl(eax, 16);
1208 cmpl(eax, 32768);
1209 jcc(Assembler::aboveEqual, L_2TAG_PACKET_3_0_2);
1210 cmpl(eax, 16);
1211 jcc(Assembler::below, L_2TAG_PACKET_4_0_2);
1212
1213 bind(L_2TAG_PACKET_5_0_2);
1214 addsd(xmm0, xmm0);
1215 jmp(L_2TAG_PACKET_2_0_2);
1216
1217 bind(L_2TAG_PACKET_6_0_2);
1218 jcc(Assembler::above, L_2TAG_PACKET_5_0_2);
1219 cmpl(edx, 0);
1220 jcc(Assembler::above, L_2TAG_PACKET_5_0_2);
1221 jmp(L_2TAG_PACKET_7_0_2);
1222
1223 bind(L_2TAG_PACKET_3_0_2);
1224 movdl(edx, xmm1);
1225 psrlq(xmm1, 32);
1226 movdl(ecx, xmm1);
1227 addl(ecx, ecx);
1228 cmpl(ecx, -2097152);
1229 jcc(Assembler::aboveEqual, L_2TAG_PACKET_6_0_2);
1230 orl(edx, ecx);
1231 cmpl(edx, 0);
1232 jcc(Assembler::equal, L_2TAG_PACKET_8_0_2);
1233
1234 bind(L_2TAG_PACKET_7_0_2);
1235 xorpd(xmm1, xmm1);
1236 xorpd(xmm0, xmm0);
1237 movl(eax, 32752);
1238 pinsrw(xmm1, eax, 3);
1239 movl(edx, 3);
1240 mulsd(xmm0, xmm1);
1241
1242 bind(L_2TAG_PACKET_9_0_2);
1243 movsd(Address(rsp, 0), xmm0);
1244 movsd(xmm0, Address(rsp, 112));
1245 fld_d(Address(rsp, 0));
1246 jmp(L_2TAG_PACKET_10_0_2);
1247
1248 bind(L_2TAG_PACKET_8_0_2);
1249 xorpd(xmm1, xmm1);
1250 xorpd(xmm0, xmm0);
1251 movl(eax, 49136);
1252 pinsrw(xmm0, eax, 3);
1253 divsd(xmm0, xmm1);
1254 movl(edx, 2);
1255 jmp(L_2TAG_PACKET_9_0_2);
1256
1257 bind(L_2TAG_PACKET_4_0_2);
1258 movdl(edx, xmm1);
1259 psrlq(xmm1, 32);
1260 movdl(ecx, xmm1);
1261 orl(edx, ecx);
1262 cmpl(edx, 0);
1263 jcc(Assembler::equal, L_2TAG_PACKET_8_0_2);
1264 xorpd(xmm1, xmm1);
1265 movl(eax, 18416);
1266 pinsrw(xmm1, eax, 3);
1267 mulsd(xmm0, xmm1);
1268 movapd(xmm1, xmm0);
1269 pextrw(eax, xmm0, 3);
1270 por(xmm0, xmm2);
1271 psllq(xmm0, 5);
1272 movl(ecx, 18416);
1273 psrlq(xmm0, 34);
1274 rcpss(xmm0, xmm0);
1275 psllq(xmm1, 12);
1276 pshufd(xmm6, xmm5, 228);
1277 psrlq(xmm1, 12);
1278 jmp(L_2TAG_PACKET_1_0_2);
1279
1280 bind(L_2TAG_PACKET_2_0_2);
1281 movsd(Address(rsp, 24), xmm0);
1282 fld_d(Address(rsp, 24));
1283
1284 bind(L_2TAG_PACKET_10_0_2);
1285 movl(tmp, Address(rsp, 40));
1286 }
1287
1288 #endif
1289
1290 /******************************************************************************/
1291 // ALGORITHM DESCRIPTION - SIN()
1292 // ---------------------
1293 //
1294 // 1. RANGE REDUCTION
1295 //
1296 // We perform an initial range reduction from X to r with
1297 //
1298 // X =~= N * pi/32 + r
1299 //
1300 // so that |r| <= pi/64 + epsilon. We restrict inputs to those
1301 // where |N| <= 932560. Beyond this, the range reduction is
1302 // insufficiently accurate. For extremely small inputs,
1303 // denormalization can occur internally, impacting performance.
1304 // This means that the main path is actually only taken for
1305 // 2^-252 <= |X| < 90112.
1306 //
1307 // To avoid branches, we perform the range reduction to full
1308 // accuracy each time.
1309 //
1310 // X - N * (P_1 + P_2 + P_3)
1311 //
1312 // where P_1 and P_2 are 32-bit numbers (so multiplication by N
1313 // is exact) and P_3 is a 53-bit number. Together, these
1314 // approximate pi well enough for all cases in the restricted
1315 // range.
1316 //
1317 // The main reduction sequence is:
1318 //
1319 // y = 32/pi * x
1320 // N = integer(y)
1321 // (computed by adding and subtracting off SHIFTER)
1322 //
1323 // m_1 = N * P_1
1324 // m_2 = N * P_2
1325 // r_1 = x - m_1
1326 // r = r_1 - m_2
1327 // (this r can be used for most of the calculation)
1328 //
1329 // c_1 = r_1 - r
1330 // m_3 = N * P_3
1331 // c_2 = c_1 - m_2
1332 // c = c_2 - m_3
1333 //
1334 // 2. MAIN ALGORITHM
1335 //
1336 // The algorithm uses a table lookup based on B = M * pi / 32
1337 // where M = N mod 64. The stored values are:
1338 // sigma closest power of 2 to cos(B)
1339 // C_hl 53-bit cos(B) - sigma
1340 // S_hi + S_lo 2 * 53-bit sin(B)
1341 //
1342 // The computation is organized as follows:
1343 //
1344 // sin(B + r + c) = [sin(B) + sigma * r] +
1345 // r * (cos(B) - sigma) +
1346 // sin(B) * [cos(r + c) - 1] +
1347 // cos(B) * [sin(r + c) - r]
1348 //
1349 // which is approximately:
1350 //
1351 // [S_hi + sigma * r] +
1352 // C_hl * r +
1353 // S_lo + S_hi * [(cos(r) - 1) - r * c] +
1354 // (C_hl + sigma) * [(sin(r) - r) + c]
1355 //
1356 // and this is what is actually computed. We separate this sum
1357 // into four parts:
1358 //
1359 // hi + med + pols + corr
1360 //
1361 // where
1362 //
1363 // hi = S_hi + sigma r
1364 // med = C_hl * r
1365 // pols = S_hi * (cos(r) - 1) + (C_hl + sigma) * (sin(r) - r)
1366 // corr = S_lo + c * ((C_hl + sigma) - S_hi * r)
1367 //
1368 // 3. POLYNOMIAL
1369 //
1370 // The polynomial S_hi * (cos(r) - 1) + (C_hl + sigma) *
1371 // (sin(r) - r) can be rearranged freely, since it is quite
1372 // small, so we exploit parallelism to the fullest.
1373 //
1374 // psc4 = SC_4 * r_1
1375 // msc4 = psc4 * r
1376 // r2 = r * r
1377 // msc2 = SC_2 * r2
1378 // r4 = r2 * r2
1379 // psc3 = SC_3 + msc4
1380 // psc1 = SC_1 + msc2
1381 // msc3 = r4 * psc3
1382 // sincospols = psc1 + msc3
1383 // pols = sincospols *
1384 // <S_hi * r^2 | (C_hl + sigma) * r^3>
1385 //
1386 // 4. CORRECTION TERM
1387 //
1388 // This is where the "c" component of the range reduction is
1389 // taken into account; recall that just "r" is used for most of
1390 // the calculation.
1391 //
1392 // -c = m_3 - c_2
1393 // -d = S_hi * r - (C_hl + sigma)
1394 // corr = -c * -d + S_lo
1395 //
1396 // 5. COMPENSATED SUMMATIONS
1397 //
1398 // The two successive compensated summations add up the high
1399 // and medium parts, leaving just the low parts to add up at
1400 // the end.
1401 //
1402 // rs = sigma * r
1403 // res_int = S_hi + rs
1404 // k_0 = S_hi - res_int
1405 // k_2 = k_0 + rs
1406 // med = C_hl * r
1407 // res_hi = res_int + med
1408 // k_1 = res_int - res_hi
1409 // k_3 = k_1 + med
1410 //
1411 // 6. FINAL SUMMATION
1412 //
1413 // We now add up all the small parts:
1414 //
1415 // res_lo = pols(hi) + pols(lo) + corr + k_1 + k_3
1416 //
1417 // Now the overall result is just:
1418 //
1419 // res_hi + res_lo
1420 //
1421 // 7. SMALL ARGUMENTS
1422 //
1423 // If |x| < SNN (SNN meaning the smallest normal number), we
1424 // simply perform 0.1111111 cdots 1111 * x. For SNN <= |x|, we
1425 // do 2^-55 * (2^55 * x - x).
1426 //
1427 // Special cases:
1428 // sin(NaN) = quiet NaN, and raise invalid exception
1429 // sin(INF) = NaN and raise invalid exception
1430 // sin(+/-0) = +/-0
1431 //
1432 /******************************************************************************/
1433
1434 #ifdef _LP64
1435
1436 ALIGNED_(16) juint _ONEHALF[] =
1437 {
1438 0x00000000UL, 0x3fe00000UL, 0x00000000UL, 0x3fe00000UL
1439 };
1440
1441 ALIGNED_(16) juint _P_2[] =
1442 {
1443 0x1a600000UL, 0x3d90b461UL, 0x1a600000UL, 0x3d90b461UL
1444 };
1445
1446 ALIGNED_(16) juint _SC_4[] =
1447 {
1448 0xa556c734UL, 0x3ec71de3UL, 0x1a01a01aUL, 0x3efa01a0UL
1449 };
1450
1451 ALIGNED_(16) juint _Ctable[] =
1452 {
1453 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
1454 0x00000000UL, 0x00000000UL, 0x3ff00000UL, 0x176d6d31UL, 0xbf73b92eUL,
1455 0xbc29b42cUL, 0x3fb917a6UL, 0xe0000000UL, 0xbc3e2718UL, 0x00000000UL,
1456 0x3ff00000UL, 0x011469fbUL, 0xbf93ad06UL, 0x3c69a60bUL, 0x3fc8f8b8UL,
1457 0xc0000000UL, 0xbc626d19UL, 0x00000000UL, 0x3ff00000UL, 0x939d225aUL,
1458 0xbfa60beaUL, 0x2ed59f06UL, 0x3fd29406UL, 0xa0000000UL, 0xbc75d28dUL,
1459 0x00000000UL, 0x3ff00000UL, 0x866b95cfUL, 0xbfb37ca1UL, 0xa6aea963UL,
1460 0x3fd87de2UL, 0xe0000000UL, 0xbc672cedUL, 0x00000000UL, 0x3ff00000UL,
1461 0x73fa1279UL, 0xbfbe3a68UL, 0x3806f63bUL, 0x3fde2b5dUL, 0x20000000UL,
1462 0x3c5e0d89UL, 0x00000000UL, 0x3ff00000UL, 0x5bc57974UL, 0xbfc59267UL,
1463 0x39ae68c8UL, 0x3fe1c73bUL, 0x20000000UL, 0x3c8b25ddUL, 0x00000000UL,
1464 0x3ff00000UL, 0x53aba2fdUL, 0xbfcd0dfeUL, 0x25091dd6UL, 0x3fe44cf3UL,
1465 0x20000000UL, 0x3c68076aUL, 0x00000000UL, 0x3ff00000UL, 0x99fcef32UL,
1466 0x3fca8279UL, 0x667f3bcdUL, 0x3fe6a09eUL, 0x20000000UL, 0xbc8bdd34UL,
1467 0x00000000UL, 0x3fe00000UL, 0x94247758UL, 0x3fc133ccUL, 0x6b151741UL,
1468 0x3fe8bc80UL, 0x20000000UL, 0xbc82c5e1UL, 0x00000000UL, 0x3fe00000UL,
1469 0x9ae68c87UL, 0x3fac73b3UL, 0x290ea1a3UL, 0x3fea9b66UL, 0xe0000000UL,
1470 0x3c39f630UL, 0x00000000UL, 0x3fe00000UL, 0x7f909c4eUL, 0xbf9d4a2cUL,
1471 0xf180bdb1UL, 0x3fec38b2UL, 0x80000000UL, 0xbc76e0b1UL, 0x00000000UL,
1472 0x3fe00000UL, 0x65455a75UL, 0xbfbe0875UL, 0xcf328d46UL, 0x3fed906bUL,
1473 0x20000000UL, 0x3c7457e6UL, 0x00000000UL, 0x3fe00000UL, 0x76acf82dUL,
1474 0x3fa4a031UL, 0x56c62ddaUL, 0x3fee9f41UL, 0xe0000000UL, 0x3c8760b1UL,
1475 0x00000000UL, 0x3fd00000UL, 0x0e5967d5UL, 0xbfac1d1fUL, 0xcff75cb0UL,
1476 0x3fef6297UL, 0x20000000UL, 0x3c756217UL, 0x00000000UL, 0x3fd00000UL,
1477 0x0f592f50UL, 0xbf9ba165UL, 0xa3d12526UL, 0x3fefd88dUL, 0x40000000UL,
1478 0xbc887df6UL, 0x00000000UL, 0x3fc00000UL, 0x00000000UL, 0x00000000UL,
1479 0x00000000UL, 0x3ff00000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
1480 0x00000000UL, 0x0f592f50UL, 0x3f9ba165UL, 0xa3d12526UL, 0x3fefd88dUL,
1481 0x40000000UL, 0xbc887df6UL, 0x00000000UL, 0xbfc00000UL, 0x0e5967d5UL,
1482 0x3fac1d1fUL, 0xcff75cb0UL, 0x3fef6297UL, 0x20000000UL, 0x3c756217UL,
1483 0x00000000UL, 0xbfd00000UL, 0x76acf82dUL, 0xbfa4a031UL, 0x56c62ddaUL,
1484 0x3fee9f41UL, 0xe0000000UL, 0x3c8760b1UL, 0x00000000UL, 0xbfd00000UL,
1485 0x65455a75UL, 0x3fbe0875UL, 0xcf328d46UL, 0x3fed906bUL, 0x20000000UL,
1486 0x3c7457e6UL, 0x00000000UL, 0xbfe00000UL, 0x7f909c4eUL, 0x3f9d4a2cUL,
1487 0xf180bdb1UL, 0x3fec38b2UL, 0x80000000UL, 0xbc76e0b1UL, 0x00000000UL,
1488 0xbfe00000UL, 0x9ae68c87UL, 0xbfac73b3UL, 0x290ea1a3UL, 0x3fea9b66UL,
1489 0xe0000000UL, 0x3c39f630UL, 0x00000000UL, 0xbfe00000UL, 0x94247758UL,
1490 0xbfc133ccUL, 0x6b151741UL, 0x3fe8bc80UL, 0x20000000UL, 0xbc82c5e1UL,
1491 0x00000000UL, 0xbfe00000UL, 0x99fcef32UL, 0xbfca8279UL, 0x667f3bcdUL,
1492 0x3fe6a09eUL, 0x20000000UL, 0xbc8bdd34UL, 0x00000000UL, 0xbfe00000UL,
1493 0x53aba2fdUL, 0x3fcd0dfeUL, 0x25091dd6UL, 0x3fe44cf3UL, 0x20000000UL,
1494 0x3c68076aUL, 0x00000000UL, 0xbff00000UL, 0x5bc57974UL, 0x3fc59267UL,
1495 0x39ae68c8UL, 0x3fe1c73bUL, 0x20000000UL, 0x3c8b25ddUL, 0x00000000UL,
1496 0xbff00000UL, 0x73fa1279UL, 0x3fbe3a68UL, 0x3806f63bUL, 0x3fde2b5dUL,
1497 0x20000000UL, 0x3c5e0d89UL, 0x00000000UL, 0xbff00000UL, 0x866b95cfUL,
1498 0x3fb37ca1UL, 0xa6aea963UL, 0x3fd87de2UL, 0xe0000000UL, 0xbc672cedUL,
1499 0x00000000UL, 0xbff00000UL, 0x939d225aUL, 0x3fa60beaUL, 0x2ed59f06UL,
1500 0x3fd29406UL, 0xa0000000UL, 0xbc75d28dUL, 0x00000000UL, 0xbff00000UL,
1501 0x011469fbUL, 0x3f93ad06UL, 0x3c69a60bUL, 0x3fc8f8b8UL, 0xc0000000UL,
1502 0xbc626d19UL, 0x00000000UL, 0xbff00000UL, 0x176d6d31UL, 0x3f73b92eUL,
1503 0xbc29b42cUL, 0x3fb917a6UL, 0xe0000000UL, 0xbc3e2718UL, 0x00000000UL,
1504 0xbff00000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
1505 0x00000000UL, 0x00000000UL, 0x00000000UL, 0xbff00000UL, 0x176d6d31UL,
1506 0x3f73b92eUL, 0xbc29b42cUL, 0xbfb917a6UL, 0xe0000000UL, 0x3c3e2718UL,
1507 0x00000000UL, 0xbff00000UL, 0x011469fbUL, 0x3f93ad06UL, 0x3c69a60bUL,
1508 0xbfc8f8b8UL, 0xc0000000UL, 0x3c626d19UL, 0x00000000UL, 0xbff00000UL,
1509 0x939d225aUL, 0x3fa60beaUL, 0x2ed59f06UL, 0xbfd29406UL, 0xa0000000UL,
1510 0x3c75d28dUL, 0x00000000UL, 0xbff00000UL, 0x866b95cfUL, 0x3fb37ca1UL,
1511 0xa6aea963UL, 0xbfd87de2UL, 0xe0000000UL, 0x3c672cedUL, 0x00000000UL,
1512 0xbff00000UL, 0x73fa1279UL, 0x3fbe3a68UL, 0x3806f63bUL, 0xbfde2b5dUL,
1513 0x20000000UL, 0xbc5e0d89UL, 0x00000000UL, 0xbff00000UL, 0x5bc57974UL,
1514 0x3fc59267UL, 0x39ae68c8UL, 0xbfe1c73bUL, 0x20000000UL, 0xbc8b25ddUL,
1515 0x00000000UL, 0xbff00000UL, 0x53aba2fdUL, 0x3fcd0dfeUL, 0x25091dd6UL,
1516 0xbfe44cf3UL, 0x20000000UL, 0xbc68076aUL, 0x00000000UL, 0xbff00000UL,
1517 0x99fcef32UL, 0xbfca8279UL, 0x667f3bcdUL, 0xbfe6a09eUL, 0x20000000UL,
1518 0x3c8bdd34UL, 0x00000000UL, 0xbfe00000UL, 0x94247758UL, 0xbfc133ccUL,
1519 0x6b151741UL, 0xbfe8bc80UL, 0x20000000UL, 0x3c82c5e1UL, 0x00000000UL,
1520 0xbfe00000UL, 0x9ae68c87UL, 0xbfac73b3UL, 0x290ea1a3UL, 0xbfea9b66UL,
1521 0xe0000000UL, 0xbc39f630UL, 0x00000000UL, 0xbfe00000UL, 0x7f909c4eUL,
1522 0x3f9d4a2cUL, 0xf180bdb1UL, 0xbfec38b2UL, 0x80000000UL, 0x3c76e0b1UL,
1523 0x00000000UL, 0xbfe00000UL, 0x65455a75UL, 0x3fbe0875UL, 0xcf328d46UL,
1524 0xbfed906bUL, 0x20000000UL, 0xbc7457e6UL, 0x00000000UL, 0xbfe00000UL,
1525 0x76acf82dUL, 0xbfa4a031UL, 0x56c62ddaUL, 0xbfee9f41UL, 0xe0000000UL,
1526 0xbc8760b1UL, 0x00000000UL, 0xbfd00000UL, 0x0e5967d5UL, 0x3fac1d1fUL,
1527 0xcff75cb0UL, 0xbfef6297UL, 0x20000000UL, 0xbc756217UL, 0x00000000UL,
1528 0xbfd00000UL, 0x0f592f50UL, 0x3f9ba165UL, 0xa3d12526UL, 0xbfefd88dUL,
1529 0x40000000UL, 0x3c887df6UL, 0x00000000UL, 0xbfc00000UL, 0x00000000UL,
1530 0x00000000UL, 0x00000000UL, 0xbff00000UL, 0x00000000UL, 0x00000000UL,
1531 0x00000000UL, 0x00000000UL, 0x0f592f50UL, 0xbf9ba165UL, 0xa3d12526UL,
1532 0xbfefd88dUL, 0x40000000UL, 0x3c887df6UL, 0x00000000UL, 0x3fc00000UL,
1533 0x0e5967d5UL, 0xbfac1d1fUL, 0xcff75cb0UL, 0xbfef6297UL, 0x20000000UL,
1534 0xbc756217UL, 0x00000000UL, 0x3fd00000UL, 0x76acf82dUL, 0x3fa4a031UL,
1535 0x56c62ddaUL, 0xbfee9f41UL, 0xe0000000UL, 0xbc8760b1UL, 0x00000000UL,
1536 0x3fd00000UL, 0x65455a75UL, 0xbfbe0875UL, 0xcf328d46UL, 0xbfed906bUL,
1537 0x20000000UL, 0xbc7457e6UL, 0x00000000UL, 0x3fe00000UL, 0x7f909c4eUL,
1538 0xbf9d4a2cUL, 0xf180bdb1UL, 0xbfec38b2UL, 0x80000000UL, 0x3c76e0b1UL,
1539 0x00000000UL, 0x3fe00000UL, 0x9ae68c87UL, 0x3fac73b3UL, 0x290ea1a3UL,
1540 0xbfea9b66UL, 0xe0000000UL, 0xbc39f630UL, 0x00000000UL, 0x3fe00000UL,
1541 0x94247758UL, 0x3fc133ccUL, 0x6b151741UL, 0xbfe8bc80UL, 0x20000000UL,
1542 0x3c82c5e1UL, 0x00000000UL, 0x3fe00000UL, 0x99fcef32UL, 0x3fca8279UL,
1543 0x667f3bcdUL, 0xbfe6a09eUL, 0x20000000UL, 0x3c8bdd34UL, 0x00000000UL,
1544 0x3fe00000UL, 0x53aba2fdUL, 0xbfcd0dfeUL, 0x25091dd6UL, 0xbfe44cf3UL,
1545 0x20000000UL, 0xbc68076aUL, 0x00000000UL, 0x3ff00000UL, 0x5bc57974UL,
1546 0xbfc59267UL, 0x39ae68c8UL, 0xbfe1c73bUL, 0x20000000UL, 0xbc8b25ddUL,
1547 0x00000000UL, 0x3ff00000UL, 0x73fa1279UL, 0xbfbe3a68UL, 0x3806f63bUL,
1548 0xbfde2b5dUL, 0x20000000UL, 0xbc5e0d89UL, 0x00000000UL, 0x3ff00000UL,
1549 0x866b95cfUL, 0xbfb37ca1UL, 0xa6aea963UL, 0xbfd87de2UL, 0xe0000000UL,
1550 0x3c672cedUL, 0x00000000UL, 0x3ff00000UL, 0x939d225aUL, 0xbfa60beaUL,
1551 0x2ed59f06UL, 0xbfd29406UL, 0xa0000000UL, 0x3c75d28dUL, 0x00000000UL,
1552 0x3ff00000UL, 0x011469fbUL, 0xbf93ad06UL, 0x3c69a60bUL, 0xbfc8f8b8UL,
1553 0xc0000000UL, 0x3c626d19UL, 0x00000000UL, 0x3ff00000UL, 0x176d6d31UL,
1554 0xbf73b92eUL, 0xbc29b42cUL, 0xbfb917a6UL, 0xe0000000UL, 0x3c3e2718UL,
1555 0x00000000UL, 0x3ff00000UL
1556 };
1557
1558 ALIGNED_(16) juint _SC_2[] =
1559 {
1560 0x11111111UL, 0x3f811111UL, 0x55555555UL, 0x3fa55555UL
1561 };
1562
1563 ALIGNED_(16) juint _SC_3[] =
1564 {
1565 0x1a01a01aUL, 0xbf2a01a0UL, 0x16c16c17UL, 0xbf56c16cUL
1566 };
1567
1568 ALIGNED_(16) juint _SC_1[] =
1569 {
1570 0x55555555UL, 0xbfc55555UL, 0x00000000UL, 0xbfe00000UL
1571 };
1572
1573 ALIGNED_(16) juint _PI_INV_TABLE[] =
1574 {
1575 0x00000000UL, 0x00000000UL, 0xa2f9836eUL, 0x4e441529UL, 0xfc2757d1UL,
1576 0xf534ddc0UL, 0xdb629599UL, 0x3c439041UL, 0xfe5163abUL, 0xdebbc561UL,
1577 0xb7246e3aUL, 0x424dd2e0UL, 0x06492eeaUL, 0x09d1921cUL, 0xfe1deb1cUL,
1578 0xb129a73eUL, 0xe88235f5UL, 0x2ebb4484UL, 0xe99c7026UL, 0xb45f7e41UL,
1579 0x3991d639UL, 0x835339f4UL, 0x9c845f8bUL, 0xbdf9283bUL, 0x1ff897ffUL,
1580 0xde05980fUL, 0xef2f118bUL, 0x5a0a6d1fUL, 0x6d367ecfUL, 0x27cb09b7UL,
1581 0x4f463f66UL, 0x9e5fea2dUL, 0x7527bac7UL, 0xebe5f17bUL, 0x3d0739f7UL,
1582 0x8a5292eaUL, 0x6bfb5fb1UL, 0x1f8d5d08UL, 0x56033046UL, 0xfc7b6babUL,
1583 0xf0cfbc21UL
1584 };
1585
1586 ALIGNED_(8) juint _PI_4[] =
1587 {
1588 0x40000000UL, 0x3fe921fbUL, 0x18469899UL, 0x3e64442dUL
1589 };
1590
1591 ALIGNED_(8) juint _PI32INV[] =
1592 {
1593 0x6dc9c883UL, 0x40245f30UL
1594 };
1595
1596 ALIGNED_(8) juint _SHIFTER[] =
1597 {
1598 0x00000000UL, 0x43380000UL
1599 };
1600
1601 ALIGNED_(8) juint _SIGN_MASK[] =
1602 {
1603 0x00000000UL, 0x80000000UL
1604 };
1605
1606 ALIGNED_(8) juint _P_3[] =
1607 {
1608 0x2e037073UL, 0x3b63198aUL
1609 };
1610
1611 ALIGNED_(8) juint _ALL_ONES[] =
1612 {
1613 0xffffffffUL, 0x3fefffffUL
1614 };
1615
1616 ALIGNED_(8) juint _TWO_POW_55[] =
1617 {
1618 0x00000000UL, 0x43600000UL
1619 };
1620
1621 ALIGNED_(8) juint _TWO_POW_M55[] =
1622 {
1623 0x00000000UL, 0x3c800000ULL
1624 };
1625
1626 ALIGNED_(8) juint _P_1[] =
1627 {
1628 0x54400000UL, 0x3fb921fbUL
1629 };
1630
1631 ALIGNED_(8) juint _NEG_ZERO[] =
1632 {
1633 0x00000000UL, 0x80000000UL
1634 };
1635
1636 void MacroAssembler::fast_sin(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ebx, Register ecx, Register edx, Register tmp1, Register tmp2, Register tmp3, Register tmp4) {
1637 Label L_2TAG_PACKET_0_0_1, L_2TAG_PACKET_1_0_1, L_2TAG_PACKET_2_0_1, L_2TAG_PACKET_3_0_1;
1638 Label L_2TAG_PACKET_4_0_1, L_2TAG_PACKET_5_0_1, L_2TAG_PACKET_6_0_1, L_2TAG_PACKET_7_0_1;
1639 Label L_2TAG_PACKET_8_0_1, L_2TAG_PACKET_9_0_1, L_2TAG_PACKET_10_0_1, L_2TAG_PACKET_11_0_1;
1640 Label L_2TAG_PACKET_13_0_1, L_2TAG_PACKET_14_0_1;
1641 Label L_2TAG_PACKET_12_0_1, B1_1, B1_2, B1_4, start;
1642
1643 assert_different_registers(tmp1, tmp2, tmp3, tmp4, eax, ebx, ecx, edx);
1644 address ONEHALF = (address)_ONEHALF;
1645 address P_2 = (address)_P_2;
1646 address SC_4 = (address)_SC_4;
1647 address Ctable = (address)_Ctable;
1648 address SC_2 = (address)_SC_2;
1649 address SC_3 = (address)_SC_3;
1650 address SC_1 = (address)_SC_1;
1651 address PI_INV_TABLE = (address)_PI_INV_TABLE;
1652 address PI_4 = (address)_PI_4;
1653 address PI32INV = (address)_PI32INV;
1654 address SHIFTER = (address)_SHIFTER;
1655 address SIGN_MASK = (address)_SIGN_MASK;
1656 address P_3 = (address)_P_3;
1657 address ALL_ONES = (address)_ALL_ONES;
1658 address TWO_POW_55 = (address)_TWO_POW_55;
1659 address TWO_POW_M55 = (address)_TWO_POW_M55;
1660 address P_1 = (address)_P_1;
1661 address NEG_ZERO = (address)_NEG_ZERO;
1662
1663 bind(start);
1664 push(rbx);
1665 subq(rsp, 16);
1666 movsd(Address(rsp, 8), xmm0);
1667 movl(eax, Address(rsp, 12));
1668 movq(xmm1, ExternalAddress(PI32INV)); //0x6dc9c883UL, 0x40245f30UL
1669 movq(xmm2, ExternalAddress(SHIFTER)); //0x00000000UL, 0x43380000UL
1670 andl(eax, 2147418112);
1671 subl(eax, 808452096);
1672 cmpl(eax, 281346048);
1673 jcc(Assembler::above, L_2TAG_PACKET_0_0_1);
1674 mulsd(xmm1, xmm0);
1675 movdqu(xmm5, ExternalAddress(ONEHALF)); //0x00000000UL, 0x3fe00000UL, 0x00000000UL, 0x3fe00000UL
1676 movq(xmm4, ExternalAddress(SIGN_MASK)); //0x00000000UL, 0x80000000UL
1677 pand(xmm4, xmm0);
1678 por(xmm5, xmm4);
1679 addpd(xmm1, xmm5);
1680 cvttsd2sil(edx, xmm1);
1681 cvtsi2sdl(xmm1, edx);
1682 movdqu(xmm6, ExternalAddress(P_2)); //0x1a600000UL, 0x3d90b461UL, 0x1a600000UL, 0x3d90b461UL
1683 mov64(r8, 0x3fb921fb54400000);
1684 movdq(xmm3, r8);
1685 movdqu(xmm5, ExternalAddress(SC_4)); //0xa556c734UL, 0x3ec71de3UL, 0x1a01a01aUL, 0x3efa01a0UL
1686 pshufd(xmm4, xmm0, 68);
1687 mulsd(xmm3, xmm1);
1688 movddup(xmm1, xmm1);
1689 andl(edx, 63);
1690 shll(edx, 5);
1691 lea(rax, ExternalAddress(Ctable));
1692 addq(rax, rdx);
1693 mulpd(xmm6, xmm1);
1694 mulsd(xmm1, ExternalAddress(P_3)); //0x2e037073UL, 0x3b63198aUL
1695 subsd(xmm4, xmm3);
1696 movq(xmm7, Address(rax, 8));
1697 subsd(xmm0, xmm3);
1698 movddup(xmm3, xmm4);
1699 subsd(xmm4, xmm6);
1700 pshufd(xmm0, xmm0, 68);
1701 movdqu(xmm2, Address(rax, 0));
1702 mulpd(xmm5, xmm0);
1703 subpd(xmm0, xmm6);
1704 mulsd(xmm7, xmm4);
1705 subsd(xmm3, xmm4);
1706 mulpd(xmm5, xmm0);
1707 mulpd(xmm0, xmm0);
1708 subsd(xmm3, xmm6);
1709 movdqu(xmm6, ExternalAddress(SC_2)); //0x11111111UL, 0x3f811111UL, 0x55555555UL, 0x3fa55555UL
1710 subsd(xmm1, xmm3);
1711 movq(xmm3, Address(rax, 24));
1712 addsd(xmm2, xmm3);
1713 subsd(xmm7, xmm2);
1714 mulsd(xmm2, xmm4);
1715 mulpd(xmm6, xmm0);
1716 mulsd(xmm3, xmm4);
1717 mulpd(xmm2, xmm0);
1718 mulpd(xmm0, xmm0);
1719 addpd(xmm5, ExternalAddress(SC_3)); //0x1a01a01aUL, 0xbf2a01a0UL, 0x16c16c17UL, 0xbf56c16cUL
1720 mulsd(xmm4, Address(rax, 0));
1721 addpd(xmm6, ExternalAddress(SC_1)); //0x55555555UL, 0xbfc55555UL, 0x00000000UL, 0xbfe00000UL
1722 mulpd(xmm5, xmm0);
1723 movdqu(xmm0, xmm3);
1724 addsd(xmm3, Address(rax, 8));
1725 mulpd(xmm1, xmm7);
1726 movdqu(xmm7, xmm4);
1727 addsd(xmm4, xmm3);
1728 addpd(xmm6, xmm5);
1729 movq(xmm5, Address(rax, 8));
1730 subsd(xmm5, xmm3);
1731 subsd(xmm3, xmm4);
1732 addsd(xmm1, Address(rax, 16));
1733 mulpd(xmm6, xmm2);
1734 addsd(xmm5, xmm0);
1735 addsd(xmm3, xmm7);
1736 addsd(xmm1, xmm5);
1737 addsd(xmm1, xmm3);
1738 addsd(xmm1, xmm6);
1739 unpckhpd(xmm6, xmm6);
1740 movdqu(xmm0, xmm4);
1741 addsd(xmm1, xmm6);
1742 addsd(xmm0, xmm1);
1743 jmp(B1_4);
1744
1745 bind(L_2TAG_PACKET_0_0_1);
1746 jcc(Assembler::greater, L_2TAG_PACKET_1_0_1);
1747 shrl(eax, 20);
1748 cmpl(eax, 3325);
1749 jcc(Assembler::notEqual, L_2TAG_PACKET_2_0_1);
1750 mulsd(xmm0, ExternalAddress(ALL_ONES)); //0xffffffffUL, 0x3fefffffUL
1751 jmp(B1_4);
1752
1753 bind(L_2TAG_PACKET_2_0_1);
1754 movq(xmm3, ExternalAddress(TWO_POW_55)); //0x00000000UL, 0x43600000UL
1755 mulsd(xmm3, xmm0);
1756 subsd(xmm3, xmm0);
1757 mulsd(xmm3, ExternalAddress(TWO_POW_M55)); //0x00000000UL, 0x3c800000UL
1758 jmp(B1_4);
1759
1760 bind(L_2TAG_PACKET_1_0_1);
1761 pextrw(eax, xmm0, 3);
1762 andl(eax, 32752);
1763 cmpl(eax, 32752);
1764 jcc(Assembler::equal, L_2TAG_PACKET_3_0_1);
1765 pextrw(ecx, xmm0, 3);
1766 andl(ecx, 32752);
1767 subl(ecx, 16224);
1768 shrl(ecx, 7);
1769 andl(ecx, 65532);
1770 lea(r11, ExternalAddress(PI_INV_TABLE));
1771 addq(rcx, r11);
1772 movdq(rax, xmm0);
1773 movl(r10, Address(rcx, 20));
1774 movl(r8, Address(rcx, 24));
1775 movl(edx, eax);
1776 shrq(rax, 21);
1777 orl(eax, INT_MIN);
1778 shrl(eax, 11);
1779 movl(r9, r10);
1780 imulq(r10, rdx);
1781 imulq(r9, rax);
1782 imulq(r8, rax);
1783 movl(rsi, Address(rcx, 16));
1784 movl(rdi, Address(rcx, 12));
1785 movl(r11, r10);
1786 shrq(r10, 32);
1787 addq(r9, r10);
1788 addq(r11, r8);
1789 movl(r8, r11);
1790 shrq(r11, 32);
1791 addq(r9, r11);
1792 movl(r10, rsi);
1793 imulq(rsi, rdx);
1794 imulq(r10, rax);
1795 movl(r11, rdi);
1796 imulq(rdi, rdx);
1797 movl(ebx, rsi);
1798 shrq(rsi, 32);
1799 addq(r9, rbx);
1800 movl(ebx, r9);
1801 shrq(r9, 32);
1802 addq(r10, rsi);
1803 addq(r10, r9);
1804 shlq(rbx, 32);
1805 orq(r8, rbx);
1806 imulq(r11, rax);
1807 movl(r9, Address(rcx, 8));
1808 movl(rsi, Address(rcx, 4));
1809 movl(ebx, rdi);
1810 shrq(rdi, 32);
1811 addq(r10, rbx);
1812 movl(ebx, r10);
1813 shrq(r10, 32);
1814 addq(r11, rdi);
1815 addq(r11, r10);
1816 movq(rdi, r9);
1817 imulq(r9, rdx);
1818 imulq(rdi, rax);
1819 movl(r10, r9);
1820 shrq(r9, 32);
1821 addq(r11, r10);
1822 movl(r10, r11);
1823 shrq(r11, 32);
1824 addq(rdi, r9);
1825 addq(rdi, r11);
1826 movq(r9, rsi);
1827 imulq(rsi, rdx);
1828 imulq(r9, rax);
1829 shlq(r10, 32);
1830 orq(r10, rbx);
1831 movl(eax, Address(rcx, 0));
1832 movl(r11, rsi);
1833 shrq(rsi, 32);
1834 addq(rdi, r11);
1835 movl(r11, rdi);
1836 shrq(rdi, 32);
1837 addq(r9, rsi);
1838 addq(r9, rdi);
1839 imulq(rdx, rax);
1840 pextrw(ebx, xmm0, 3);
1841 lea(rdi, ExternalAddress(PI_INV_TABLE));
1842 subq(rcx, rdi);
1843 addl(ecx, ecx);
1844 addl(ecx, ecx);
1845 addl(ecx, ecx);
1846 addl(ecx, 19);
1847 movl(rsi, 32768);
1848 andl(rsi, ebx);
1849 shrl(ebx, 4);
1850 andl(ebx, 2047);
1851 subl(ebx, 1023);
1852 subl(ecx, ebx);
1853 addq(r9, rdx);
1854 movl(edx, ecx);
1855 addl(edx, 32);
1856 cmpl(ecx, 1);
1857 jcc(Assembler::less, L_2TAG_PACKET_4_0_1);
1858 negl(ecx);
1859 addl(ecx, 29);
1860 shll(r9);
1861 movl(rdi, r9);
1862 andl(r9, 536870911);
1863 testl(r9, 268435456);
1864 jcc(Assembler::notEqual, L_2TAG_PACKET_5_0_1);
1865 shrl(r9);
1866 movl(ebx, 0);
1867 shlq(r9, 32);
1868 orq(r9, r11);
1869
1870 bind(L_2TAG_PACKET_6_0_1);
1871
1872 bind(L_2TAG_PACKET_7_0_1);
1873
1874 cmpq(r9, 0);
1875 jcc(Assembler::equal, L_2TAG_PACKET_8_0_1);
1876
1877 bind(L_2TAG_PACKET_9_0_1);
1878 bsrq(r11, r9);
1879 movl(ecx, 29);
1880 subl(ecx, r11);
1881 jcc(Assembler::lessEqual, L_2TAG_PACKET_10_0_1);
1882 shlq(r9);
1883 movq(rax, r10);
1884 shlq(r10);
1885 addl(edx, ecx);
1886 negl(ecx);
1887 addl(ecx, 64);
1888 shrq(rax);
1889 shrq(r8);
1890 orq(r9, rax);
1891 orq(r10, r8);
1892
1893 bind(L_2TAG_PACKET_11_0_1);
1894 cvtsi2sdq(xmm0, r9);
1895 shrq(r10, 1);
1896 cvtsi2sdq(xmm3, r10);
1897 xorpd(xmm4, xmm4);
1898 shll(edx, 4);
1899 negl(edx);
1900 addl(edx, 16368);
1901 orl(edx, rsi);
1902 xorl(edx, ebx);
1903 pinsrw(xmm4, edx, 3);
1904 movq(xmm2, ExternalAddress(PI_4)); //0x40000000UL, 0x3fe921fbUL, 0x18469899UL, 0x3e64442dUL
1905 movq(xmm6, ExternalAddress(8 + PI_4)); //0x3fe921fbUL, 0x18469899UL, 0x3e64442dUL
1906 xorpd(xmm5, xmm5);
1907 subl(edx, 1008);
1908 pinsrw(xmm5, edx, 3);
1909 mulsd(xmm0, xmm4);
1910 shll(rsi, 16);
1911 sarl(rsi, 31);
1912 mulsd(xmm3, xmm5);
1913 movdqu(xmm1, xmm0);
1914 mulsd(xmm0, xmm2);
1915 shrl(rdi, 29);
1916 addsd(xmm1, xmm3);
1917 mulsd(xmm3, xmm2);
1918 addl(rdi, rsi);
1919 xorl(rdi, rsi);
1920 mulsd(xmm6, xmm1);
1921 movl(eax, rdi);
1922 addsd(xmm6, xmm3);
1923 movdqu(xmm2, xmm0);
1924 addsd(xmm0, xmm6);
1925 subsd(xmm2, xmm0);
1926 addsd(xmm6, xmm2);
1927
1928 bind(L_2TAG_PACKET_12_0_1);
1929 movq(xmm1, ExternalAddress(PI32INV)); //0x6dc9c883UL, 0x40245f30UL
1930 mulsd(xmm1, xmm0);
1931 movq(xmm5, ExternalAddress(ONEHALF)); //0x00000000UL, 0x3fe00000UL, 0x00000000UL, 0x3fe00000UL
1932 movq(xmm4, ExternalAddress(SIGN_MASK)); //0x00000000UL, 0x80000000UL
1933 pand(xmm4, xmm0);
1934 por(xmm5, xmm4);
1935 addpd(xmm1, xmm5);
1936 cvttsd2sil(edx, xmm1);
1937 cvtsi2sdl(xmm1, edx);
1938 movq(xmm3, ExternalAddress(P_1)); //0x54400000UL, 0x3fb921fbUL
1939 movdqu(xmm2, ExternalAddress(P_2)); //0x1a600000UL, 0x3d90b461UL, 0x1a600000UL, 0x3d90b461UL
1940 mulsd(xmm3, xmm1);
1941 unpcklpd(xmm1, xmm1);
1942 shll(eax, 3);
1943 addl(edx, 1865216);
1944 movdqu(xmm4, xmm0);
1945 addl(edx, eax);
1946 andl(edx, 63);
1947 movdqu(xmm5, ExternalAddress(SC_4)); //0x54400000UL, 0x3fb921fbUL
1948 lea(rax, ExternalAddress(Ctable));
1949 shll(edx, 5);
1950 addq(rax, rdx);
1951 mulpd(xmm2, xmm1);
1952 subsd(xmm0, xmm3);
1953 mulsd(xmm1, ExternalAddress(P_3)); //0x2e037073UL, 0x3b63198aUL
1954 subsd(xmm4, xmm3);
1955 movq(xmm7, Address(rax, 8));
1956 unpcklpd(xmm0, xmm0);
1957 movdqu(xmm3, xmm4);
1958 subsd(xmm4, xmm2);
1959 mulpd(xmm5, xmm0);
1960 subpd(xmm0, xmm2);
1961 mulsd(xmm7, xmm4);
1962 subsd(xmm3, xmm4);
1963 mulpd(xmm5, xmm0);
1964 mulpd(xmm0, xmm0);
1965 subsd(xmm3, xmm2);
1966 movdqu(xmm2, Address(rax, 0));
1967 subsd(xmm1, xmm3);
1968 movq(xmm3, Address(rax, 24));
1969 addsd(xmm2, xmm3);
1970 subsd(xmm7, xmm2);
1971 subsd(xmm1, xmm6);
1972 movdqu(xmm6, ExternalAddress(SC_2)); //0x11111111UL, 0x3f811111UL, 0x55555555UL, 0x3fa55555UL
1973 mulsd(xmm2, xmm4);
1974 mulpd(xmm6, xmm0);
1975 mulsd(xmm3, xmm4);
1976 mulpd(xmm2, xmm0);
1977 mulpd(xmm0, xmm0);
1978 addpd(xmm5, ExternalAddress(SC_3)); //0x1a01a01aUL, 0xbf2a01a0UL, 0x16c16c17UL, 0xbf56c16cUL
1979 mulsd(xmm4, Address(rax, 0));
1980 addpd(xmm6, ExternalAddress(SC_1)); //0x55555555UL, 0xbfc55555UL, 0x00000000UL, 0xbfe00000UL
1981 mulpd(xmm5, xmm0);
1982 movdqu(xmm0, xmm3);
1983 addsd(xmm3, Address(rax, 8));
1984 mulpd(xmm1, xmm7);
1985 movdqu(xmm7, xmm4);
1986 addsd(xmm4, xmm3);
1987 addpd(xmm6, xmm5);
1988 movq(xmm5, Address(rax, 8));
1989 subsd(xmm5, xmm3);
1990 subsd(xmm3, xmm4);
1991 addsd(xmm1, Address(rax, 16));
1992 mulpd(xmm6, xmm2);
1993 addsd(xmm5, xmm0);
1994 addsd(xmm3, xmm7);
1995 addsd(xmm1, xmm5);
1996 addsd(xmm1, xmm3);
1997 addsd(xmm1, xmm6);
1998 unpckhpd(xmm6, xmm6);
1999 movdqu(xmm0, xmm4);
2000 addsd(xmm1, xmm6);
2001 addsd(xmm0, xmm1);
2002 jmp(B1_4);
2003
2004 bind(L_2TAG_PACKET_8_0_1);
2005 addl(edx, 64);
2006 movq(r9, r10);
2007 movq(r10, r8);
2008 movl(r8, 0);
2009 cmpq(r9, 0);
2010 jcc(Assembler::notEqual, L_2TAG_PACKET_9_0_1);
2011 addl(edx, 64);
2012 movq(r9, r10);
2013 movq(r10, r8);
2014 cmpq(r9, 0);
2015 jcc(Assembler::notEqual, L_2TAG_PACKET_9_0_1);
2016 xorpd(xmm0, xmm0);
2017 xorpd(xmm6, xmm6);
2018 jmp(L_2TAG_PACKET_12_0_1);
2019
2020 bind(L_2TAG_PACKET_10_0_1);
2021 jcc(Assembler::equal, L_2TAG_PACKET_11_0_1);
2022 negl(ecx);
2023 shrq(r10);
2024 movq(rax, r9);
2025 shrq(r9);
2026 subl(edx, ecx);
2027 negl(ecx);
2028 addl(ecx, 64);
2029 shlq(rax);
2030 orq(r10, rax);
2031 jmp(L_2TAG_PACKET_11_0_1);
2032
2033 bind(L_2TAG_PACKET_4_0_1);
2034 negl(ecx);
2035 shlq(r9, 32);
2036 orq(r9, r11);
2037 shlq(r9);
2038 movq(rdi, r9);
2039 testl(r9, INT_MIN);
2040 jcc(Assembler::notEqual, L_2TAG_PACKET_13_0_1);
2041 shrl(r9);
2042 movl(ebx, 0);
2043 shrq(rdi, 3);
2044 jmp(L_2TAG_PACKET_7_0_1);
2045
2046 bind(L_2TAG_PACKET_5_0_1);
2047 shrl(r9);
2048 movl(ebx, 536870912);
2049 shrl(ebx);
2050 shlq(r9, 32);
2051 orq(r9, r11);
2052 shlq(rbx, 32);
2053 addl(rdi, 536870912);
2054 movl(rcx, 0);
2055 movl(r11, 0);
2056 subq(rcx, r8);
2057 sbbq(r11, r10);
2058 sbbq(rbx, r9);
2059 movq(r8, rcx);
2060 movq(r10, r11);
2061 movq(r9, rbx);
2062 movl(ebx, 32768);
2063 jmp(L_2TAG_PACKET_6_0_1);
2064
2065 bind(L_2TAG_PACKET_13_0_1);
2066 shrl(r9);
2067 mov64(rbx, 0x100000000);
2068 shrq(rbx);
2069 movl(rcx, 0);
2070 movl(r11, 0);
2071 subq(rcx, r8);
2072 sbbq(r11, r10);
2073 sbbq(rbx, r9);
2074 movq(r8, rcx);
2075 movq(r10, r11);
2076 movq(r9, rbx);
2077 movl(ebx, 32768);
2078 shrq(rdi, 3);
2079 addl(rdi, 536870912);
2080 jmp(L_2TAG_PACKET_7_0_1);
2081
2082 bind(L_2TAG_PACKET_3_0_1);
2083 movq(xmm0, Address(rsp, 8));
2084 mulsd(xmm0, ExternalAddress(NEG_ZERO)); //0x00000000UL, 0x80000000UL
2085 movq(Address(rsp, 0), xmm0);
2086
2087 bind(L_2TAG_PACKET_14_0_1);
2088
2089 bind(B1_4);
2090 addq(rsp, 16);
2091 pop(rbx);
2092 }
2093
2094 #endif
2095
2096 #ifndef _LP64
2097
2098 ALIGNED_(8) juint _zero_none[] =
2099 {
2100 0x00000000UL, 0x00000000UL, 0x00000000UL, 0xbff00000UL
2101 };
2102
2103 ALIGNED_(4) juint __4onpi_d[] =
2104 {
2105 0x6dc9c883UL, 0x3ff45f30UL
2106 };
2107
2108 ALIGNED_(4) juint _TWO_32H[] =
2109 {
2110 0x00000000UL, 0x41f80000UL
2111 };
2112
2113 ALIGNED_(4) juint _pi04_3d[] =
2114 {
2115 0x54442d00UL, 0x3fe921fbUL, 0x98cc5180UL, 0x3ce84698UL, 0xcbb5bf6cUL,
2116 0xb9dfc8f8UL
2117 };
2118
2119 ALIGNED_(4) juint _pi04_5d[] =
2120 {
2121 0x54400000UL, 0x3fe921fbUL, 0x1a600000UL, 0x3dc0b461UL, 0x2e000000UL,
2122 0x3b93198aUL, 0x25200000UL, 0x396b839aUL, 0x533e63a0UL, 0x37027044UL
2123 };
2124
2125 ALIGNED_(4) juint _SCALE[] =
2126 {
2127 0x00000000UL, 0x32600000UL
2128 };
2129
2130 ALIGNED_(4) juint _zeros[] =
2131 {
2132 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x80000000UL
2133 };
2134
2135 ALIGNED_(4) juint _pi04_2d[] =
2136 {
2137 0x54400000UL, 0x3fe921fbUL, 0x1a626331UL, 0x3dc0b461UL
2138 };
2139
2140 ALIGNED_(4) juint _TWO_12H[] =
2141 {
2142 0x00000000UL, 0x40b80000UL
2143 };
2144
2145 ALIGNED_(2) jushort __4onpi_31l[] =
2146 {
2147 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x836e, 0xa2f9,
2148 0x40d8, 0x0000, 0x0000, 0x0000, 0x2a50, 0x9c88, 0x40b7, 0x0000, 0x0000, 0x0000,
2149 0xabe8, 0xfe13, 0x4099, 0x0000, 0x0000, 0x0000, 0x6ee0, 0xfa9a, 0x4079, 0x0000,
2150 0x0000, 0x0000, 0x9580, 0xdb62, 0x4058, 0x0000, 0x0000, 0x0000, 0x1c82, 0xc9e2,
2151 0x403d, 0x0000, 0x0000, 0x0000, 0xb1c0, 0xff28, 0x4019, 0x0000, 0x0000, 0x0000,
2152 0xef14, 0xaf7a, 0x3ffe, 0x0000, 0x0000, 0x0000, 0x48dc, 0xc36e, 0x3fdf, 0x0000,
2153 0x0000, 0x0000, 0x3740, 0xe909, 0x3fbe, 0x0000, 0x0000, 0x0000, 0x924a, 0xb801,
2154 0x3fa2, 0x0000, 0x0000, 0x0000, 0x3a32, 0xdd41, 0x3f83, 0x0000, 0x0000, 0x0000,
2155 0x8778, 0x873f, 0x3f62, 0x0000, 0x0000, 0x0000, 0x1298, 0xb1cb, 0x3f44, 0x0000,
2156 0x0000, 0x0000, 0xa208, 0x9cfb, 0x3f26, 0x0000, 0x0000, 0x0000, 0xbaec, 0xd7d4,
2157 0x3f06, 0x0000, 0x0000, 0x0000, 0xd338, 0x8909, 0x3ee7, 0x0000, 0x0000, 0x0000,
2158 0x68b8, 0xe04d, 0x3ec7, 0x0000, 0x0000, 0x0000, 0x4e64, 0xdf90, 0x3eaa, 0x0000,
2159 0x0000, 0x0000, 0xc1a8, 0xeb1c, 0x3e89, 0x0000, 0x0000, 0x0000, 0x2720, 0xce7d,
2160 0x3e6a, 0x0000, 0x0000, 0x0000, 0x77b8, 0x8bf1, 0x3e4b, 0x0000, 0x0000, 0x0000,
2161 0xec7e, 0xe4a0, 0x3e2e, 0x0000, 0x0000, 0x0000, 0xffbc, 0xf12f, 0x3e0f, 0x0000,
2162 0x0000, 0x0000, 0xfdc0, 0xb301, 0x3deb, 0x0000, 0x0000, 0x0000, 0xc5ac, 0x9788,
2163 0x3dd1, 0x0000, 0x0000, 0x0000, 0x47da, 0x829b, 0x3db2, 0x0000, 0x0000, 0x0000,
2164 0xd9e4, 0xa6cf, 0x3d93, 0x0000, 0x0000, 0x0000, 0x36e8, 0xf961, 0x3d73, 0x0000,
2165 0x0000, 0x0000, 0xf668, 0xf463, 0x3d54, 0x0000, 0x0000, 0x0000, 0x5168, 0xf2ff,
2166 0x3d35, 0x0000, 0x0000, 0x0000, 0x758e, 0xea4f, 0x3d17, 0x0000, 0x0000, 0x0000,
2167 0xf17a, 0xebe5, 0x3cf8, 0x0000, 0x0000, 0x0000, 0x9cfa, 0x9e83, 0x3cd9, 0x0000,
2168 0x0000, 0x0000, 0xa4ba, 0xe294, 0x3cba, 0x0000, 0x0000, 0x0000, 0xd7ec, 0x9afe,
2169 0x3c9a, 0x0000, 0x0000, 0x0000, 0xae80, 0x8fc6, 0x3c79, 0x0000, 0x0000, 0x0000,
2170 0x3304, 0x8560, 0x3c5c, 0x0000, 0x0000, 0x0000, 0x6d70, 0xdf8f, 0x3c3b, 0x0000,
2171 0x0000, 0x0000, 0x3ef0, 0xafc3, 0x3c1e, 0x0000, 0x0000, 0x0000, 0xd0d8, 0x826b,
2172 0x3bfe, 0x0000, 0x0000, 0x0000, 0x1c80, 0xed4f, 0x3bdd, 0x0000, 0x0000, 0x0000,
2173 0x730c, 0xb0af, 0x3bc1, 0x0000, 0x0000, 0x0000, 0x6660, 0xc219, 0x3ba2, 0x0000,
2174 0x0000, 0x0000, 0x940c, 0xabe2, 0x3b83, 0x0000, 0x0000, 0x0000, 0xdffc, 0x8408,
2175 0x3b64, 0x0000, 0x0000, 0x0000, 0x6b98, 0xc402, 0x3b45, 0x0000, 0x0000, 0x0000,
2176 0x1818, 0x9cc4, 0x3b26, 0x0000, 0x0000, 0x0000, 0x5390, 0xaab6, 0x3b05, 0x0000,
2177 0x0000, 0x0000, 0xb070, 0xd464, 0x3ae9, 0x0000, 0x0000, 0x0000, 0x231a, 0x9ef0,
2178 0x3aca, 0x0000, 0x0000, 0x0000, 0x0670, 0xd1f1, 0x3aaa, 0x0000, 0x0000, 0x0000,
2179 0x7738, 0xd9f3, 0x3a8a, 0x0000, 0x0000, 0x0000, 0xa834, 0x8092, 0x3a6c, 0x0000,
2180 0x0000, 0x0000, 0xb45c, 0xce23, 0x3a4d, 0x0000, 0x0000, 0x0000, 0x36e8, 0xb0e5,
2181 0x3a2d, 0x0000, 0x0000, 0x0000, 0xd156, 0xaf44, 0x3a10, 0x0000, 0x0000, 0x0000,
2182 0x9f52, 0x8c82, 0x39f1, 0x0000, 0x0000, 0x0000, 0x829c, 0xff83, 0x39d1, 0x0000,
2183 0x0000, 0x0000, 0x7d06, 0xefc6, 0x39b3, 0x0000, 0x0000, 0x0000, 0x93e0, 0xb0b7,
2184 0x3992, 0x0000, 0x0000, 0x0000, 0xedde, 0xc193, 0x3975, 0x0000, 0x0000, 0x0000,
2185 0xbbc0, 0xcf49, 0x3952, 0x0000, 0x0000, 0x0000, 0xbdf0, 0xd63c, 0x3937, 0x0000,
2186 0x0000, 0x0000, 0x1f34, 0x9f3a, 0x3918, 0x0000, 0x0000, 0x0000, 0x3f8e, 0xe579,
2187 0x38f9, 0x0000, 0x0000, 0x0000, 0x90c8, 0xc3f8, 0x38d9, 0x0000, 0x0000, 0x0000,
2188 0x48c0, 0xf8f8, 0x38b7, 0x0000, 0x0000, 0x0000, 0xed56, 0xafa6, 0x389c, 0x0000,
2189 0x0000, 0x0000, 0x8218, 0xb969, 0x387d, 0x0000, 0x0000, 0x0000, 0x1852, 0xec57,
2190 0x385e, 0x0000, 0x0000, 0x0000, 0x670c, 0xd674, 0x383e, 0x0000, 0x0000, 0x0000,
2191 0xad40, 0xc2c4, 0x3820, 0x0000, 0x0000, 0x0000, 0x2e80, 0xa696, 0x3801, 0x0000,
2192 0x0000, 0x0000, 0xd800, 0xc467, 0x37dc, 0x0000, 0x0000, 0x0000, 0x3c72, 0xc5ae,
2193 0x37c3, 0x0000, 0x0000, 0x0000, 0xb006, 0xac69, 0x37a4, 0x0000, 0x0000, 0x0000,
2194 0x34a0, 0x8cdf, 0x3782, 0x0000, 0x0000, 0x0000, 0x9ed2, 0xd25e, 0x3766, 0x0000,
2195 0x0000, 0x0000, 0x6fec, 0xaaaa, 0x3747, 0x0000, 0x0000, 0x0000, 0x6040, 0xfb5c,
2196 0x3726, 0x0000, 0x0000, 0x0000, 0x764c, 0xa3fc, 0x3708, 0x0000, 0x0000, 0x0000,
2197 0xb254, 0x954e, 0x36e9, 0x0000, 0x0000, 0x0000, 0x3e1c, 0xf5dc, 0x36ca, 0x0000,
2198 0x0000, 0x0000, 0x7b06, 0xc635, 0x36ac, 0x0000, 0x0000, 0x0000, 0xa8ba, 0xd738,
2199 0x368d, 0x0000, 0x0000, 0x0000, 0x06cc, 0xb24e, 0x366d, 0x0000, 0x0000, 0x0000,
2200 0x7108, 0xac76, 0x364f, 0x0000, 0x0000, 0x0000, 0x2324, 0xa7cb, 0x3630, 0x0000,
2201 0x0000, 0x0000, 0xac40, 0xef15, 0x360f, 0x0000, 0x0000, 0x0000, 0xae46, 0xd516,
2202 0x35f2, 0x0000, 0x0000, 0x0000, 0x615e, 0xe003, 0x35d3, 0x0000, 0x0000, 0x0000,
2203 0x0cf0, 0xefe7, 0x35b1, 0x0000, 0x0000, 0x0000, 0xfb50, 0xf98c, 0x3595, 0x0000,
2204 0x0000, 0x0000, 0x0abc, 0xf333, 0x3575, 0x0000, 0x0000, 0x0000, 0xdd60, 0xca3f,
2205 0x3555, 0x0000, 0x0000, 0x0000, 0x7eb6, 0xd87f, 0x3538, 0x0000, 0x0000, 0x0000,
2206 0x44f4, 0xb291, 0x3519, 0x0000, 0x0000, 0x0000, 0xff80, 0xc982, 0x34f6, 0x0000,
2207 0x0000, 0x0000, 0x9de0, 0xd9b8, 0x34db, 0x0000, 0x0000, 0x0000, 0xcd42, 0x9366,
2208 0x34bc, 0x0000, 0x0000, 0x0000, 0xbef0, 0xfaee, 0x349d, 0x0000, 0x0000, 0x0000,
2209 0xdac4, 0xb6f1, 0x347d, 0x0000, 0x0000, 0x0000, 0xf140, 0x94de, 0x345d, 0x0000,
2210 0x0000, 0x0000, 0xa218, 0x8b4b, 0x343e, 0x0000, 0x0000, 0x0000, 0x6380, 0xa135,
2211 0x341e, 0x0000, 0x0000, 0x0000, 0xb184, 0x8cb2, 0x3402, 0x0000, 0x0000, 0x0000,
2212 0x196e, 0xdc61, 0x33e3, 0x0000, 0x0000, 0x0000, 0x0c00, 0xde05, 0x33c4, 0x0000,
2213 0x0000, 0x0000, 0xef9a, 0xbd38, 0x33a5, 0x0000, 0x0000, 0x0000, 0xc1a0, 0xdf00,
2214 0x3385, 0x0000, 0x0000, 0x0000, 0x1090, 0x9973, 0x3365, 0x0000, 0x0000, 0x0000,
2215 0x4882, 0x8301, 0x3348, 0x0000, 0x0000, 0x0000, 0x7abe, 0xadc7, 0x3329, 0x0000,
2216 0x0000, 0x0000, 0x7cba, 0xec2b, 0x330a, 0x0000, 0x0000, 0x0000, 0xa520, 0x8f21,
2217 0x32e9, 0x0000, 0x0000, 0x0000, 0x710c, 0x8d36, 0x32cc, 0x0000, 0x0000, 0x0000,
2218 0x5212, 0xc6ed, 0x32ad, 0x0000, 0x0000, 0x0000, 0x7308, 0xfd76, 0x328d, 0x0000,
2219 0x0000, 0x0000, 0x5014, 0xd548, 0x326f, 0x0000, 0x0000, 0x0000, 0xd3f2, 0xb499,
2220 0x3250, 0x0000, 0x0000, 0x0000, 0x7f74, 0xa606, 0x3230, 0x0000, 0x0000, 0x0000,
2221 0xf0a8, 0xd720, 0x3212, 0x0000, 0x0000, 0x0000, 0x185c, 0xe20f, 0x31f2, 0x0000,
2222 0x0000, 0x0000, 0xa5a8, 0x8738, 0x31d4, 0x0000, 0x0000, 0x0000, 0xdd74, 0xcafb,
2223 0x31b4, 0x0000, 0x0000, 0x0000, 0x98b6, 0xbd8e, 0x3196, 0x0000, 0x0000, 0x0000,
2224 0xe9de, 0x977f, 0x3177, 0x0000, 0x0000, 0x0000, 0x67c0, 0x818d, 0x3158, 0x0000,
2225 0x0000, 0x0000, 0xe52a, 0x9322, 0x3139, 0x0000, 0x0000, 0x0000, 0xe568, 0x9b6c,
2226 0x3119, 0x0000, 0x0000, 0x0000, 0x2358, 0xaa0a, 0x30fa, 0x0000, 0x0000, 0x0000,
2227 0xe480, 0xe13b, 0x30d9, 0x0000, 0x0000, 0x0000, 0x3024, 0x90a1, 0x30bd, 0x0000,
2228 0x0000, 0x0000, 0x9620, 0xda30, 0x309d, 0x0000, 0x0000, 0x0000, 0x898a, 0xb388,
2229 0x307f, 0x0000, 0x0000, 0x0000, 0xb24c, 0xc891, 0x3060, 0x0000, 0x0000, 0x0000,
2230 0x8056, 0xf98b, 0x3041, 0x0000, 0x0000, 0x0000, 0x72a4, 0xa1ea, 0x3021, 0x0000,
2231 0x0000, 0x0000, 0x6af8, 0x9488, 0x3001, 0x0000, 0x0000, 0x0000, 0xe00c, 0xdfcb,
2232 0x2fe4, 0x0000, 0x0000, 0x0000, 0xeeec, 0xc941, 0x2fc4, 0x0000, 0x0000, 0x0000,
2233 0x53e0, 0xe70f, 0x2fa4, 0x0000, 0x0000, 0x0000, 0x8f60, 0x9c07, 0x2f85, 0x0000,
2234 0x0000, 0x0000, 0xb328, 0xc3e7, 0x2f68, 0x0000, 0x0000, 0x0000, 0x9404, 0xf8c7,
2235 0x2f48, 0x0000, 0x0000, 0x0000, 0x38e0, 0xc99f, 0x2f29, 0x0000, 0x0000, 0x0000,
2236 0x9778, 0xd984, 0x2f09, 0x0000, 0x0000, 0x0000, 0xe700, 0xd142, 0x2eea, 0x0000,
2237 0x0000, 0x0000, 0xd904, 0x9443, 0x2ecd, 0x0000, 0x0000, 0x0000, 0xd4ba, 0xae7e,
2238 0x2eae, 0x0000, 0x0000, 0x0000, 0x8e5e, 0x8524, 0x2e8f, 0x0000, 0x0000, 0x0000,
2239 0xb550, 0xc9ed, 0x2e6e, 0x0000, 0x0000, 0x0000, 0x53b8, 0x8648, 0x2e51, 0x0000,
2240 0x0000, 0x0000, 0xdae4, 0x87f9, 0x2e32, 0x0000, 0x0000, 0x0000, 0x2942, 0xd966,
2241 0x2e13, 0x0000, 0x0000, 0x0000, 0x4f28, 0xcf3c, 0x2df3, 0x0000, 0x0000, 0x0000,
2242 0xfa40, 0xc4ef, 0x2dd1, 0x0000, 0x0000, 0x0000, 0x4424, 0xbca7, 0x2db5, 0x0000,
2243 0x0000, 0x0000, 0x2e62, 0xcdc5, 0x2d97, 0x0000, 0x0000, 0x0000, 0xed88, 0x996b,
2244 0x2d78, 0x0000, 0x0000, 0x0000, 0x7c30, 0xd97d, 0x2d56, 0x0000, 0x0000, 0x0000,
2245 0xed26, 0xbf6e, 0x2d3a, 0x0000, 0x0000, 0x0000, 0x2918, 0x921b, 0x2d1a, 0x0000,
2246 0x0000, 0x0000, 0x4e24, 0xe84e, 0x2cfb, 0x0000, 0x0000, 0x0000, 0x6dc0, 0x92ec,
2247 0x2cdd, 0x0000, 0x0000, 0x0000, 0x4f2c, 0xacf8, 0x2cbd, 0x0000, 0x0000, 0x0000,
2248 0xc634, 0xf094, 0x2c9e, 0x0000, 0x0000, 0x0000, 0xdc70, 0xe5d3, 0x2c7e, 0x0000,
2249 0x0000, 0x0000, 0x2180, 0xa600, 0x2c5b, 0x0000, 0x0000, 0x0000, 0x8480, 0xd680,
2250 0x2c3c, 0x0000, 0x0000, 0x0000, 0x8b24, 0xd63b, 0x2c22, 0x0000, 0x0000, 0x0000,
2251 0x02e0, 0xaa47, 0x2c00, 0x0000, 0x0000, 0x0000, 0x9ad0, 0xee84, 0x2be3, 0x0000,
2252 0x0000, 0x0000, 0xf7dc, 0xf699, 0x2bc6, 0x0000, 0x0000, 0x0000, 0xddde, 0xe490,
2253 0x2ba7, 0x0000, 0x0000, 0x0000, 0x34a0, 0xb4fd, 0x2b85, 0x0000, 0x0000, 0x0000,
2254 0x91b4, 0x8ef6, 0x2b68, 0x0000, 0x0000, 0x0000, 0xa3e0, 0xa2a7, 0x2b47, 0x0000,
2255 0x0000, 0x0000, 0xcce4, 0x82b3, 0x2b2a, 0x0000, 0x0000, 0x0000, 0xe4be, 0x8207,
2256 0x2b0c, 0x0000, 0x0000, 0x0000, 0x1d92, 0xab43, 0x2aed, 0x0000, 0x0000, 0x0000,
2257 0xe818, 0xf9f6, 0x2acd, 0x0000, 0x0000, 0x0000, 0xff12, 0xba80, 0x2aaf, 0x0000,
2258 0x0000, 0x0000, 0x5254, 0x8529, 0x2a90, 0x0000, 0x0000, 0x0000, 0x1b88, 0xe032,
2259 0x2a71, 0x0000, 0x0000, 0x0000, 0x3248, 0xd86d, 0x2a50, 0x0000, 0x0000, 0x0000,
2260 0x3140, 0xc9d5, 0x2a2e, 0x0000, 0x0000, 0x0000, 0x14e6, 0xbd47, 0x2a14, 0x0000,
2261 0x0000, 0x0000, 0x5c10, 0xe544, 0x29f4, 0x0000, 0x0000, 0x0000, 0x9f50, 0x90b6,
2262 0x29d4, 0x0000, 0x0000, 0x0000, 0x9850, 0xab55, 0x29b6, 0x0000, 0x0000, 0x0000,
2263 0x2750, 0x9d07, 0x2998, 0x0000, 0x0000, 0x0000, 0x6700, 0x8bbb, 0x2973, 0x0000,
2264 0x0000, 0x0000, 0x5dba, 0xed31, 0x295a, 0x0000, 0x0000, 0x0000, 0x61dc, 0x85fe,
2265 0x293a, 0x0000, 0x0000, 0x0000, 0x9ba2, 0xd6b4, 0x291c, 0x0000, 0x0000, 0x0000,
2266 0x2d30, 0xe3a5, 0x28fb, 0x0000, 0x0000, 0x0000, 0x6630, 0xb566, 0x28dd, 0x0000,
2267 0x0000, 0x0000, 0x5ad4, 0xa829, 0x28bf, 0x0000, 0x0000, 0x0000, 0x89d8, 0xe290,
2268 0x28a0, 0x0000, 0x0000, 0x0000, 0x3916, 0xc428, 0x2881, 0x0000, 0x0000, 0x0000,
2269 0x0490, 0xbea4, 0x2860, 0x0000, 0x0000, 0x0000, 0xee06, 0x80ee, 0x2843, 0x0000,
2270 0x0000, 0x0000, 0xfc00, 0xf327, 0x2820, 0x0000, 0x0000, 0x0000, 0xea40, 0xa871,
2271 0x2800, 0x0000, 0x0000, 0x0000, 0x63d8, 0x9c26, 0x27e4, 0x0000, 0x0000, 0x0000,
2272 0x07ba, 0xc0c9, 0x27c7, 0x0000, 0x0000, 0x0000, 0x3fa2, 0x9797, 0x27a8, 0x0000,
2273 0x0000, 0x0000, 0x21c6, 0xfeca, 0x2789, 0x0000, 0x0000, 0x0000, 0xde40, 0x860d,
2274 0x2768, 0x0000, 0x0000, 0x0000, 0x9cc8, 0x98ce, 0x2749, 0x0000, 0x0000, 0x0000,
2275 0x3778, 0xa31c, 0x272a, 0x0000, 0x0000, 0x0000, 0xe778, 0xf6e2, 0x270b, 0x0000,
2276 0x0000, 0x0000, 0x59b8, 0xf841, 0x26ed, 0x0000, 0x0000, 0x0000, 0x02e0, 0xad04,
2277 0x26cd, 0x0000, 0x0000, 0x0000, 0x5a92, 0x9380, 0x26b0, 0x0000, 0x0000, 0x0000,
2278 0xc740, 0x8886, 0x268d, 0x0000, 0x0000, 0x0000, 0x0680, 0xfaf8, 0x266c, 0x0000,
2279 0x0000, 0x0000, 0xfb60, 0x897f, 0x2653, 0x0000, 0x0000, 0x0000, 0x8760, 0xf903,
2280 0x2634, 0x0000, 0x0000, 0x0000, 0xad2a, 0xc2c8, 0x2615, 0x0000, 0x0000, 0x0000,
2281 0x2d86, 0x8aef, 0x25f6, 0x0000, 0x0000, 0x0000, 0x1ef4, 0xe627, 0x25d6, 0x0000,
2282 0x0000, 0x0000, 0x09e4, 0x8020, 0x25b7, 0x0000, 0x0000, 0x0000, 0x7548, 0xd227,
2283 0x2598, 0x0000, 0x0000, 0x0000, 0x75dc, 0xfb5b, 0x2579, 0x0000, 0x0000, 0x0000,
2284 0xea84, 0xc8b6, 0x255a, 0x0000, 0x0000, 0x0000, 0xe4d0, 0x8145, 0x253b, 0x0000,
2285 0x0000, 0x0000, 0x3640, 0x9768, 0x251c, 0x0000, 0x0000, 0x0000, 0x246a, 0xccec,
2286 0x24fe, 0x0000, 0x0000, 0x0000, 0x51d0, 0xa075, 0x24dd, 0x0000, 0x0000, 0x0000,
2287 0x4638, 0xa385, 0x24bf, 0x0000, 0x0000, 0x0000, 0xd788, 0xd776, 0x24a1, 0x0000,
2288 0x0000, 0x0000, 0x1370, 0x8997, 0x2482, 0x0000, 0x0000, 0x0000, 0x1e88, 0x9b67,
2289 0x2462, 0x0000, 0x0000, 0x0000, 0x6c08, 0xd975, 0x2444, 0x0000, 0x0000, 0x0000,
2290 0xfdb0, 0xcfc0, 0x2422, 0x0000, 0x0000, 0x0000, 0x3100, 0xc026, 0x2406, 0x0000,
2291 0x0000, 0x0000, 0xc5b4, 0xae64, 0x23e6, 0x0000, 0x0000, 0x0000, 0x2280, 0xf687,
2292 0x23c3, 0x0000, 0x0000, 0x0000, 0x2de0, 0x9006, 0x23a9, 0x0000, 0x0000, 0x0000,
2293 0x24bc, 0xf631, 0x238a, 0x0000, 0x0000, 0x0000, 0xb8d4, 0xa975, 0x236b, 0x0000,
2294 0x0000, 0x0000, 0xd9a4, 0xb949, 0x234b, 0x0000, 0x0000, 0x0000, 0xb54e, 0xbd39,
2295 0x232d, 0x0000, 0x0000, 0x0000, 0x4aac, 0x9a52, 0x230e, 0x0000, 0x0000, 0x0000,
2296 0xbbbc, 0xd085, 0x22ef, 0x0000, 0x0000, 0x0000, 0xdf18, 0xc633, 0x22cf, 0x0000,
2297 0x0000, 0x0000, 0x16d0, 0xeca5, 0x22af, 0x0000, 0x0000, 0x0000, 0xf2a0, 0xdf6f,
2298 0x228e, 0x0000, 0x0000, 0x0000, 0x8c44, 0xe86b, 0x2272, 0x0000, 0x0000, 0x0000,
2299 0x35c0, 0xbbf4, 0x2253, 0x0000, 0x0000, 0x0000, 0x0c40, 0xdafb, 0x2230, 0x0000,
2300 0x0000, 0x0000, 0x92dc, 0x9935, 0x2216, 0x0000, 0x0000, 0x0000, 0x0ca0, 0xbda6,
2301 0x21f3, 0x0000, 0x0000, 0x0000, 0x5958, 0xa6fd, 0x21d6, 0x0000, 0x0000, 0x0000,
2302 0xa3dc, 0x9d7f, 0x21b9, 0x0000, 0x0000, 0x0000, 0x79dc, 0xfcb5, 0x2199, 0x0000,
2303 0x0000, 0x0000, 0xf264, 0xcebb, 0x217b, 0x0000, 0x0000, 0x0000, 0x0abe, 0x8308,
2304 0x215c, 0x0000, 0x0000, 0x0000, 0x30ae, 0xb463, 0x213d, 0x0000, 0x0000, 0x0000,
2305 0x6228, 0xb040, 0x211c, 0x0000, 0x0000, 0x0000, 0xc9b2, 0xf43b, 0x20ff, 0x0000,
2306 0x0000, 0x0000, 0x3d8e, 0xa4b3, 0x20e0, 0x0000, 0x0000, 0x0000, 0x84e6, 0x8dab,
2307 0x20c1, 0x0000, 0x0000, 0x0000, 0xa124, 0x9b74, 0x20a1, 0x0000, 0x0000, 0x0000,
2308 0xc276, 0xd497, 0x2083, 0x0000, 0x0000, 0x0000, 0x6354, 0xa466, 0x2063, 0x0000,
2309 0x0000, 0x0000, 0x8654, 0xaf0a, 0x2044, 0x0000, 0x0000, 0x0000, 0x1d20, 0xfa5c,
2310 0x2024, 0x0000, 0x0000, 0x0000, 0xbcd0, 0xf3f0, 0x2004, 0x0000, 0x0000, 0x0000,
2311 0xedf0, 0xf0b6, 0x1fe7, 0x0000, 0x0000, 0x0000, 0x45bc, 0x9182, 0x1fc9, 0x0000,
2312 0x0000, 0x0000, 0xe254, 0xdc85, 0x1faa, 0x0000, 0x0000, 0x0000, 0xb898, 0xe9b1,
2313 0x1f8a, 0x0000, 0x0000, 0x0000, 0x0ebe, 0xe6f0, 0x1f6c, 0x0000, 0x0000, 0x0000,
2314 0xa9b8, 0xf584, 0x1f4c, 0x0000, 0x0000, 0x0000, 0x12e8, 0xdf6b, 0x1f2e, 0x0000,
2315 0x0000, 0x0000, 0x9f9e, 0xcd55, 0x1f0f, 0x0000, 0x0000, 0x0000, 0x05a0, 0xec3a,
2316 0x1eef, 0x0000, 0x0000, 0x0000, 0xd8e0, 0x96f8, 0x1ed1, 0x0000, 0x0000, 0x0000,
2317 0x3bd4, 0xccc6, 0x1eb1, 0x0000, 0x0000, 0x0000, 0x4910, 0xb87b, 0x1e93, 0x0000,
2318 0x0000, 0x0000, 0xbefc, 0xd40b, 0x1e73, 0x0000, 0x0000, 0x0000, 0x317e, 0xa406,
2319 0x1e55, 0x0000, 0x0000, 0x0000, 0x6bb2, 0xc2b2, 0x1e36, 0x0000, 0x0000, 0x0000,
2320 0xb87e, 0xbb78, 0x1e17, 0x0000, 0x0000, 0x0000, 0xa03c, 0xdbbd, 0x1df7, 0x0000,
2321 0x0000, 0x0000, 0x5b6c, 0xe3c8, 0x1dd9, 0x0000, 0x0000, 0x0000, 0x8968, 0xca8e,
2322 0x1dba, 0x0000, 0x0000, 0x0000, 0xc024, 0xe6ab, 0x1d9a, 0x0000, 0x0000, 0x0000,
2323 0x4110, 0xd4eb, 0x1d7a, 0x0000, 0x0000, 0x0000, 0xa168, 0xbdb5, 0x1d5d, 0x0000,
2324 0x0000, 0x0000, 0x012e, 0xa5fa, 0x1d3e, 0x0000, 0x0000, 0x0000, 0x6838, 0x9c1f,
2325 0x1d1e, 0x0000, 0x0000, 0x0000, 0xa158, 0xaa76, 0x1d00, 0x0000, 0x0000, 0x0000,
2326 0x090a, 0xbd95, 0x1ce1, 0x0000, 0x0000, 0x0000, 0xf73e, 0x8b6d, 0x1cc2, 0x0000,
2327 0x0000, 0x0000, 0x5fda, 0xbcbf, 0x1ca3, 0x0000, 0x0000, 0x0000, 0xdbe8, 0xb89f,
2328 0x1c84, 0x0000, 0x0000, 0x0000, 0x6e4c, 0x96c7, 0x1c64, 0x0000, 0x0000, 0x0000,
2329 0x19c2, 0xf2a4, 0x1c46, 0x0000, 0x0000, 0x0000, 0xb800, 0xf855, 0x1c1e, 0x0000,
2330 0x0000, 0x0000, 0x87fc, 0x85ff, 0x1c08, 0x0000, 0x0000, 0x0000, 0x1418, 0x839f,
2331 0x1be9, 0x0000, 0x0000, 0x0000, 0x6186, 0xd9d8, 0x1bca, 0x0000, 0x0000, 0x0000,
2332 0xf500, 0xabaa, 0x1ba6, 0x0000, 0x0000, 0x0000, 0x7b36, 0xdafe, 0x1b8c, 0x0000,
2333 0x0000, 0x0000, 0xf394, 0xe6d8, 0x1b6c, 0x0000, 0x0000, 0x0000, 0x6efc, 0x9e55,
2334 0x1b4e, 0x0000, 0x0000, 0x0000, 0x5e10, 0xc523, 0x1b2e, 0x0000, 0x0000, 0x0000,
2335 0x8210, 0xb6f9, 0x1b0d, 0x0000, 0x0000, 0x0000, 0x9ab0, 0x96e3, 0x1af1, 0x0000,
2336 0x0000, 0x0000, 0x3864, 0x92e7, 0x1ad1, 0x0000, 0x0000, 0x0000, 0x9878, 0xdc65,
2337 0x1ab1, 0x0000, 0x0000, 0x0000, 0xfa20, 0xd6cb, 0x1a94, 0x0000, 0x0000, 0x0000,
2338 0x6c00, 0xa4e4, 0x1a70, 0x0000, 0x0000, 0x0000, 0xab40, 0xb41b, 0x1a53, 0x0000,
2339 0x0000, 0x0000, 0x43a4, 0x8ede, 0x1a37, 0x0000, 0x0000, 0x0000, 0x22e0, 0x9314,
2340 0x1a15, 0x0000, 0x0000, 0x0000, 0x6170, 0xb949, 0x19f8, 0x0000, 0x0000, 0x0000,
2341 0x6b00, 0xe056, 0x19d8, 0x0000, 0x0000, 0x0000, 0x9ba8, 0xa94c, 0x19b9, 0x0000,
2342 0x0000, 0x0000, 0xfaa0, 0xaa16, 0x199b, 0x0000, 0x0000, 0x0000, 0x899a, 0xf627,
2343 0x197d, 0x0000, 0x0000, 0x0000, 0x9f20, 0xfb70, 0x195d, 0x0000, 0x0000, 0x0000,
2344 0xa4b8, 0xc176, 0x193e, 0x0000, 0x0000, 0x0000, 0xb21c, 0x85c3, 0x1920, 0x0000,
2345 0x0000, 0x0000, 0x50d2, 0x9b19, 0x1901, 0x0000, 0x0000, 0x0000, 0xd4b0, 0xb708,
2346 0x18e0, 0x0000, 0x0000, 0x0000, 0xfb88, 0xf510, 0x18c1, 0x0000, 0x0000, 0x0000,
2347 0x31ec, 0xdc8d, 0x18a3, 0x0000, 0x0000, 0x0000, 0x3c00, 0xbff9, 0x1885, 0x0000,
2348 0x0000, 0x0000, 0x5020, 0xc30b, 0x1862, 0x0000, 0x0000, 0x0000, 0xd4f0, 0xda0c,
2349 0x1844, 0x0000, 0x0000, 0x0000, 0x20d2, 0x99a5, 0x1828, 0x0000, 0x0000, 0x0000,
2350 0x852e, 0xd159, 0x1809, 0x0000, 0x0000, 0x0000, 0x7cd8, 0x97a1, 0x17e9, 0x0000,
2351 0x0000, 0x0000, 0x423a, 0x997b, 0x17cb, 0x0000, 0x0000, 0x0000, 0xc1c0, 0xbe7d,
2352 0x17a8, 0x0000, 0x0000, 0x0000, 0xe8bc, 0xdcdd, 0x178d, 0x0000, 0x0000, 0x0000,
2353 0x8b28, 0xae06, 0x176e, 0x0000, 0x0000, 0x0000, 0x102e, 0xb8d4, 0x174f, 0x0000,
2354 0x0000, 0x0000, 0xaa00, 0xaa5c, 0x172f, 0x0000, 0x0000, 0x0000, 0x51f0, 0x9fc0,
2355 0x170e, 0x0000, 0x0000, 0x0000, 0xf858, 0xe181, 0x16f2, 0x0000, 0x0000, 0x0000,
2356 0x91a8, 0x8162, 0x16d3, 0x0000, 0x0000, 0x0000, 0x5f40, 0xcb6f, 0x16b1, 0x0000,
2357 0x0000, 0x0000, 0xbb50, 0xe55f, 0x1693, 0x0000, 0x0000, 0x0000, 0xacd2, 0xd895,
2358 0x1676, 0x0000, 0x0000, 0x0000, 0xef30, 0x97bf, 0x1654, 0x0000, 0x0000, 0x0000,
2359 0xf700, 0xb3d7, 0x1633, 0x0000, 0x0000, 0x0000, 0x3454, 0xa7b5, 0x1619, 0x0000,
2360 0x0000, 0x0000, 0x6b00, 0xa929, 0x15f6, 0x0000, 0x0000, 0x0000, 0x9f04, 0x89f7,
2361 0x15db, 0x0000, 0x0000, 0x0000, 0xad78, 0xd985, 0x15bc, 0x0000, 0x0000, 0x0000,
2362 0xa46a, 0xae3f, 0x159d, 0x0000, 0x0000, 0x0000, 0x63a0, 0xd0da, 0x157c, 0x0000,
2363 0x0000, 0x0000, 0x5e90, 0x817d, 0x155e, 0x0000, 0x0000, 0x0000, 0x1494, 0xb13f,
2364 0x1540, 0x0000, 0x0000, 0x0000, 0x0090, 0x9c40, 0x1521, 0x0000, 0x0000, 0x0000,
2365 0xdd70, 0xcc86, 0x1500, 0x0000, 0x0000, 0x0000, 0x64f8, 0xdb6f, 0x14e1, 0x0000,
2366 0x0000, 0x0000, 0xe22c, 0xac17, 0x14c3, 0x0000, 0x0000, 0x0000, 0x60e0, 0xa9ad,
2367 0x14a3, 0x0000, 0x0000, 0x0000, 0x4640, 0xd658, 0x1481, 0x0000, 0x0000, 0x0000,
2368 0x6490, 0xa181, 0x1467, 0x0000, 0x0000, 0x0000, 0x1df4, 0xaaa2, 0x1447, 0x0000,
2369 0x0000, 0x0000, 0xb94a, 0x8f61, 0x1429, 0x0000, 0x0000, 0x0000, 0x5198, 0x9d83,
2370 0x1409, 0x0000, 0x0000, 0x0000, 0x0f7a, 0xa818, 0x13eb, 0x0000, 0x0000, 0x0000,
2371 0xc45e, 0xc06c, 0x13cc, 0x0000, 0x0000, 0x0000, 0x4ec0, 0xfa29, 0x13a8, 0x0000,
2372 0x0000, 0x0000, 0x6418, 0x8cad, 0x138c, 0x0000, 0x0000, 0x0000, 0xbcc8, 0xe7d1,
2373 0x136f, 0x0000, 0x0000, 0x0000, 0xc934, 0xf9b0, 0x134f, 0x0000, 0x0000, 0x0000,
2374 0x6ce0, 0x98df, 0x1331, 0x0000, 0x0000, 0x0000, 0x3516, 0xe5e9, 0x1312, 0x0000,
2375 0x0000, 0x0000, 0xc6c0, 0xef8b, 0x12ef, 0x0000, 0x0000, 0x0000, 0xaf02, 0x913d,
2376 0x12d4, 0x0000, 0x0000, 0x0000, 0xd230, 0xe1d5, 0x12b5, 0x0000, 0x0000, 0x0000,
2377 0xfba8, 0xc232, 0x1295, 0x0000, 0x0000, 0x0000, 0x7ba4, 0xabeb, 0x1277, 0x0000,
2378 0x0000, 0x0000, 0x6e5c, 0xc692, 0x1258, 0x0000, 0x0000, 0x0000, 0x76a2, 0x9756,
2379 0x1239, 0x0000, 0x0000, 0x0000, 0xe180, 0xe423, 0x1214, 0x0000, 0x0000, 0x0000,
2380 0x8c3c, 0x90f8, 0x11fb, 0x0000, 0x0000, 0x0000, 0x9f3c, 0x9fd2, 0x11dc, 0x0000,
2381 0x0000, 0x0000, 0x53e0, 0xb73e, 0x11bd, 0x0000, 0x0000, 0x0000, 0x45be, 0x88d6,
2382 0x119e, 0x0000, 0x0000, 0x0000, 0x111a, 0x8bc0, 0x117f, 0x0000, 0x0000, 0x0000,
2383 0xe26a, 0xd7ff, 0x1160, 0x0000, 0x0000, 0x0000, 0xfb60, 0xdd8d, 0x113f, 0x0000,
2384 0x0000, 0x0000, 0x9370, 0xc108, 0x1120, 0x0000, 0x0000, 0x0000, 0x9654, 0x8baf,
2385 0x1103, 0x0000, 0x0000, 0x0000, 0xd6ec, 0xd6b9, 0x10e4, 0x0000, 0x0000, 0x0000,
2386 0x23e4, 0xd7b7, 0x10c4, 0x0000, 0x0000, 0x0000, 0x1aa6, 0xa847, 0x10a6, 0x0000,
2387 0x0000, 0x0000, 0xbee6, 0x9fef, 0x1087, 0x0000, 0x0000, 0x0000, 0x26d0, 0xa6eb,
2388 0x1066, 0x0000, 0x0000, 0x0000, 0x5b86, 0xa880, 0x1049, 0x0000, 0x0000, 0x0000,
2389 0x125c, 0xd971, 0x1029, 0x0000, 0x0000, 0x0000, 0x1f78, 0x9d18, 0x100a, 0x0000,
2390 0x0000, 0x0000, 0x0e84, 0xb15b, 0x0feb, 0x0000, 0x0000, 0x0000, 0xd0c0, 0xc150,
2391 0x0fcc, 0x0000, 0x0000, 0x0000, 0xa330, 0xc40c, 0x0fad, 0x0000, 0x0000, 0x0000,
2392 0x5202, 0xfc2c, 0x0f8f, 0x0000, 0x0000, 0x0000, 0x3f7c, 0xecf5, 0x0f6f, 0x0000,
2393 0x0000, 0x0000, 0xef44, 0xfdfd, 0x0f50, 0x0000, 0x0000, 0x0000, 0x3f6c, 0xab1b,
2394 0x0f31, 0x0000, 0x0000, 0x0000, 0xf658, 0x89ec, 0x0f11, 0x0000, 0x0000, 0x0000,
2395 0xbfc8, 0x9ba8, 0x0ef4, 0x0000, 0x0000, 0x0000, 0x3d40, 0xbe21, 0x0ed5, 0x0000,
2396 0x0000, 0x0000, 0xbbc4, 0xc70d, 0x0eb6, 0x0000, 0x0000, 0x0000, 0x5158, 0xdb16,
2397 0x0e96, 0x0000, 0x0000, 0x0000, 0xb5a8, 0xa8d8, 0x0e78, 0x0000, 0x0000, 0x0000,
2398 0xcccc, 0xb40e, 0x0e58, 0x0000, 0x0000, 0x0000, 0x448c, 0xcb62, 0x0e3a, 0x0000,
2399 0x0000, 0x0000, 0xf12a, 0x8aed, 0x0e1b, 0x0000, 0x0000, 0x0000, 0x79d0, 0xc59c,
2400 0x0dfb, 0x0000, 0x0000, 0x0000, 0x06b4, 0xcdc9, 0x0ddd, 0x0000, 0x0000, 0x0000,
2401 0xae70, 0xa979, 0x0dbe, 0x0000, 0x0000, 0x0000, 0x317c, 0xa8fb, 0x0d9e, 0x0000,
2402 0x0000, 0x0000, 0x5fe0, 0x8a50, 0x0d7d, 0x0000, 0x0000, 0x0000, 0x70b6, 0xfdfa,
2403 0x0d61, 0x0000, 0x0000, 0x0000, 0x1640, 0x9dc7, 0x0d41, 0x0000, 0x0000, 0x0000,
2404 0x9a9c, 0xdc50, 0x0d23, 0x0000, 0x0000, 0x0000, 0x4fcc, 0x9a9b, 0x0d04, 0x0000,
2405 0x0000, 0x0000, 0x7e48, 0x8f77, 0x0ce5, 0x0000, 0x0000, 0x0000, 0x84e4, 0xd4b9,
2406 0x0cc6, 0x0000, 0x0000, 0x0000, 0x84e0, 0xbd10, 0x0ca6, 0x0000, 0x0000, 0x0000,
2407 0x1b0a, 0xc8d9, 0x0c88, 0x0000, 0x0000, 0x0000, 0x6a48, 0xfc81, 0x0c68, 0x0000,
2408 0x0000, 0x0000, 0x070a, 0xbef6, 0x0c4a, 0x0000, 0x0000, 0x0000, 0x8a70, 0xf096,
2409 0x0c2b, 0x0000, 0x0000, 0x0000, 0xecc2, 0xc994, 0x0c0c, 0x0000, 0x0000, 0x0000,
2410 0x1540, 0x9537, 0x0bea, 0x0000, 0x0000, 0x0000, 0x1b02, 0xab5b, 0x0bce, 0x0000,
2411 0x0000, 0x0000, 0x5dc0, 0xb0c8, 0x0bad, 0x0000, 0x0000, 0x0000, 0xc928, 0xe034,
2412 0x0b8f, 0x0000, 0x0000, 0x0000, 0x2d12, 0xb4b0, 0x0b71, 0x0000, 0x0000, 0x0000,
2413 0x8fc2, 0xbb94, 0x0b52, 0x0000, 0x0000, 0x0000, 0xe236, 0xe22f, 0x0b33, 0x0000,
2414 0x0000, 0x0000, 0xb97c, 0xbe9e, 0x0b13, 0x0000, 0x0000, 0x0000, 0xe1a6, 0xe16d,
2415 0x0af5, 0x0000, 0x0000, 0x0000, 0xd330, 0xbaf0, 0x0ad6, 0x0000, 0x0000, 0x0000,
2416 0xc0bc, 0xbbd0, 0x0ab7, 0x0000, 0x0000, 0x0000, 0x8e66, 0xdd9b, 0x0a98, 0x0000,
2417 0x0000, 0x0000, 0xc95c, 0xf799, 0x0a79, 0x0000, 0x0000, 0x0000, 0xdac0, 0xbe4c,
2418 0x0a55, 0x0000, 0x0000, 0x0000, 0xafc0, 0xc378, 0x0a37, 0x0000, 0x0000, 0x0000,
2419 0xa880, 0xe341, 0x0a19, 0x0000, 0x0000, 0x0000, 0xc242, 0x81f6, 0x09fd, 0x0000,
2420 0x0000, 0x0000, 0x7470, 0xc777, 0x09de, 0x0000, 0x0000, 0x0000, 0x62bc, 0xb684,
2421 0x09be, 0x0000, 0x0000, 0x0000, 0x43ac, 0x8c58, 0x099f, 0x0000, 0x0000, 0x0000,
2422 0xcc3c, 0xf9ac, 0x0981, 0x0000, 0x0000, 0x0000, 0x1526, 0xb670, 0x0962, 0x0000,
2423 0x0000, 0x0000, 0xc9fe, 0xdf50, 0x0943, 0x0000, 0x0000, 0x0000, 0x6ae6, 0xc065,
2424 0x0924, 0x0000, 0x0000, 0x0000, 0xb114, 0xcf29, 0x0905, 0x0000, 0x0000, 0x0000,
2425 0xd388, 0x922a, 0x08e4, 0x0000, 0x0000, 0x0000, 0xcf54, 0xb926, 0x08c7, 0x0000,
2426 0x0000, 0x0000, 0x3826, 0xe855, 0x08a8, 0x0000, 0x0000, 0x0000, 0xe7c8, 0x829b,
2427 0x0888, 0x0000, 0x0000, 0x0000, 0x546c, 0xa903, 0x086a, 0x0000, 0x0000, 0x0000,
2428 0x8768, 0x99cc, 0x0849, 0x0000, 0x0000, 0x0000, 0x00ac, 0xf529, 0x082b, 0x0000,
2429 0x0000, 0x0000, 0x2658, 0x9f0b, 0x080c, 0x0000, 0x0000, 0x0000, 0xfe5c, 0x9e21,
2430 0x07ee, 0x0000, 0x0000, 0x0000, 0x6da2, 0x9910, 0x07cf, 0x0000, 0x0000, 0x0000,
2431 0x9220, 0xf9b3, 0x07b0, 0x0000, 0x0000, 0x0000, 0x3d90, 0xa541, 0x0791, 0x0000,
2432 0x0000, 0x0000, 0x6e4c, 0xe7cc, 0x0771, 0x0000, 0x0000, 0x0000, 0xa8fa, 0xe80a,
2433 0x0753, 0x0000, 0x0000, 0x0000, 0x4e14, 0xc3a7, 0x0734, 0x0000, 0x0000, 0x0000,
2434 0xf7e0, 0xbad9, 0x0712, 0x0000, 0x0000, 0x0000, 0xfea0, 0xeff2, 0x06f5, 0x0000,
2435 0x0000, 0x0000, 0xcef6, 0xbd48, 0x06d7, 0x0000, 0x0000, 0x0000, 0x7544, 0xf559,
2436 0x06b7, 0x0000, 0x0000, 0x0000, 0x2388, 0xf655, 0x0698, 0x0000, 0x0000, 0x0000,
2437 0xe900, 0xad56, 0x0676, 0x0000, 0x0000, 0x0000, 0x2cc0, 0x8437, 0x0659, 0x0000,
2438 0x0000, 0x0000, 0x3068, 0xc544, 0x063b, 0x0000, 0x0000, 0x0000, 0xdc70, 0xe73c,
2439 0x061b, 0x0000, 0x0000, 0x0000, 0xee50, 0x9d49, 0x05fc, 0x0000, 0x0000, 0x0000,
2440 0x93d2, 0x81f6, 0x05df, 0x0000, 0x0000, 0x0000, 0x941c, 0xadff, 0x05bf, 0x0000,
2441 0x0000, 0x0000, 0x2ce2, 0x8e45, 0x05a1, 0x0000, 0x0000, 0x0000, 0x4a60, 0x95fd,
2442 0x0581, 0x0000, 0x0000, 0x0000, 0x79f8, 0xb83a, 0x0563, 0x0000, 0x0000, 0x0000,
2443 0xcb58, 0xa1f5, 0x0543, 0x0000, 0x0000, 0x0000, 0x2a3a, 0xdc36, 0x0525, 0x0000,
2444 0x0000, 0x0000, 0x14ee, 0x890e, 0x0506, 0x0000, 0x0000, 0x0000, 0x8f20, 0xc432,
2445 0x04e3, 0x0000, 0x0000, 0x0000, 0x8440, 0xb21d, 0x04c6, 0x0000, 0x0000, 0x0000,
2446 0x5430, 0xf698, 0x04a7, 0x0000, 0x0000, 0x0000, 0x04ae, 0x8b20, 0x048a, 0x0000,
2447 0x0000, 0x0000, 0x04d0, 0xe872, 0x046b, 0x0000, 0x0000, 0x0000, 0xc78e, 0x8893,
2448 0x044c, 0x0000, 0x0000, 0x0000, 0x0f78, 0x9895, 0x042b, 0x0000, 0x0000, 0x0000,
2449 0x11d4, 0xdf2e, 0x040d, 0x0000, 0x0000, 0x0000, 0xe84c, 0x89d5, 0x03ef, 0x0000,
2450 0x0000, 0x0000, 0xf7be, 0x8a67, 0x03d0, 0x0000, 0x0000, 0x0000, 0x95d0, 0xc906,
2451 0x03b1, 0x0000, 0x0000, 0x0000, 0x64ce, 0xd96c, 0x0392, 0x0000, 0x0000, 0x0000,
2452 0x97ba, 0xa16f, 0x0373, 0x0000, 0x0000, 0x0000, 0x463c, 0xc51a, 0x0354, 0x0000,
2453 0x0000, 0x0000, 0xef0a, 0xe93e, 0x0335, 0x0000, 0x0000, 0x0000, 0x526a, 0xa466,
2454 0x0316, 0x0000, 0x0000, 0x0000, 0x4140, 0xa94d, 0x02f5, 0x0000, 0x0000, 0x0000,
2455 0xb4ec, 0xce68, 0x02d8, 0x0000, 0x0000, 0x0000, 0x4fa2, 0x8490, 0x02b9, 0x0000,
2456 0x0000, 0x0000, 0x4e60, 0xca98, 0x0298, 0x0000, 0x0000, 0x0000, 0x08dc, 0xe09c,
2457 0x027a, 0x0000, 0x0000, 0x0000, 0x2b90, 0xc7e3, 0x025c, 0x0000, 0x0000, 0x0000,
2458 0x5a7c, 0xf8ef, 0x023c, 0x0000, 0x0000, 0x0000, 0x5022, 0x9d58, 0x021e, 0x0000,
2459 0x0000, 0x0000, 0x553a, 0xe242, 0x01ff, 0x0000, 0x0000, 0x0000, 0x7e6e, 0xb54d,
2460 0x01e0, 0x0000, 0x0000, 0x0000, 0xd2d4, 0xa88c, 0x01c1, 0x0000, 0x0000, 0x0000,
2461 0x75b6, 0xfe6d, 0x01a2, 0x0000, 0x0000, 0x0000, 0x3bb2, 0xf04c, 0x0183, 0x0000,
2462 0x0000, 0x0000, 0xc2d0, 0xc046, 0x0163, 0x0000, 0x0000, 0x0000, 0x250c, 0xf9d6,
2463 0x0145, 0x0000, 0x0000, 0x0000, 0xb7b4, 0x8a0d, 0x0126, 0x0000, 0x0000, 0x0000,
2464 0x1a72, 0xe4f5, 0x0107, 0x0000, 0x0000, 0x0000, 0x825c, 0xa9b8, 0x00e8, 0x0000,
2465 0x0000, 0x0000, 0x6c90, 0xc9ad, 0x00c6, 0x0000, 0x0000, 0x0000, 0x4d00, 0xd1bb,
2466 0x00aa, 0x0000, 0x0000, 0x0000, 0xa4a0, 0xee01, 0x0087, 0x0000, 0x0000, 0x0000,
2467 0x89a8, 0xbe9f, 0x006b, 0x0000, 0x0000, 0x0000, 0x038e, 0xc80c, 0x004d, 0x0000,
2468 0x0000, 0x0000, 0xfe26, 0x8384, 0x002e, 0x0000, 0x0000, 0x0000, 0xcd90, 0xca57,
2469 0x000e, 0x0000
2470 };
2471
2472 void MacroAssembler::libm_reduce_pi04l(Register eax, Register ecx, Register edx, Register ebx, Register esi, Register edi, Register ebp, Register esp) {
2473 Label B1_1, B1_2, B1_3, B1_4, B1_5, B1_6, B1_7, B1_8, B1_9, B1_10, B1_11, B1_12;
2474 Label B1_13, B1_14, B1_15;
2475
2476 assert_different_registers(ebx, eax, ecx, edx, esi, edi, ebp, esp);
2477
2478 address zero_none = (address)_zero_none;
2479 address _4onpi_d = (address)__4onpi_d;
2480 address TWO_32H = (address)_TWO_32H;
2481 address pi04_3d = (address)_pi04_3d;
2482 address pi04_5d = (address)_pi04_5d;
2483 address SCALE = (address)_SCALE;
2484 address zeros = (address)_zeros;
2485 address pi04_2d = (address)_pi04_2d;
2486 address TWO_12H = (address)_TWO_12H;
2487 address _4onpi_31l = (address)__4onpi_31l;
2488
2489 bind(B1_1);
2490 push(ebp);
2491 movl(ebp, esp);
2492 andl(esp, -16);
2493 push(esi);
2494 push(edi);
2495 push(ebx);
2496 subl(esp, 20);
2497 movzwl(ebx, Address(ebp, 16));
2498 andl(ebx, 32767);
2499 movl(eax, Address(ebp, 20));
2500 cmpl(ebx, 16413);
2501 movl(esi, Address(ebp, 24));
2502 movl(Address(esp, 4), eax);
2503 jcc(Assembler::greaterEqual, B1_8);
2504
2505 bind(B1_2);
2506 fld_x(Address(ebp, 8));
2507 fld_d(ExternalAddress(_4onpi_d)); //0x6dc9c883UL, 0x3ff45f30UL
2508 fmul(1);
2509 fstp_x(Address(esp, 8));
2510 movzwl(ecx, Address(esp, 16));
2511 negl(ecx);
2512 addl(ecx, 30);
2513 movl(eax, Address(esp, 12));
2514 shrl(eax);
2515 cmpl(Address(esp, 4), 0);
2516 jcc(Assembler::notEqual, B1_4);
2517
2518 bind(B1_3);
2519 lea(ecx, Address(eax, 1));
2520 andl(ecx, -2);
2521 jmp(B1_5);
2522
2523 bind(B1_4);
2524 movl(ecx, eax);
2525 addl(eax, Address(esp, 4));
2526 movl(edx, eax);
2527 andl(edx, 1);
2528 addl(ecx, edx);
2529
2530 bind(B1_5);
2531 fld_d(ExternalAddress(TWO_32H)); //0x00000000UL, 0x41f80000UL
2532 cmpl(ebx, 16400);
2533 movl(Address(esp, 0), ecx);
2534 fild_s(Address(esp, 0));
2535 jcc(Assembler::greaterEqual, B1_7);
2536
2537 bind(B1_6);
2538 fld_d(ExternalAddress(pi04_3d)); //0x54442d00UL, 0x3fe921fbUL
2539 fmul(1);
2540 fsubp(3);
2541 fxch(1);
2542 fmul(2);
2543 fld_s(2);
2544 fadd(1);
2545 fsubrp(1);
2546 fld_s(0);
2547 fxch(1);
2548 fsuba(3);
2549 fld_d(ExternalAddress(8 + pi04_3d)); //0x98cc5180UL, 0x3ce84698UL
2550 fmul(3);
2551 fsuba(2);
2552 fxch(1);
2553 fsub(2);
2554 fsubrp(1);
2555 faddp(3);
2556 fld_d(ExternalAddress(16 + pi04_3d)); //0xcbb5bf6cUL, 0xb9dfc8f8UL
2557 fmulp(2);
2558 fld_s(1);
2559 fsubr(1);
2560 fsuba(1);
2561 fxch(2);
2562 fsubp(1);
2563 faddp(2);
2564 fxch(1);
2565 jmp(B1_15);
2566
2567 bind(B1_7);
2568 fld_d(ExternalAddress(pi04_5d)); //0x54400000UL, 0x3fe921fbUL
2569 fmul(1);
2570 fsubp(3);
2571 fxch(1);
2572 fmul(2);
2573 fld_s(2);
2574 fadd(1);
2575 fsubrp(1);
2576 fld_s(0);
2577 fxch(1);
2578 fsuba(3);
2579 fld_d(ExternalAddress(8 + pi04_5d)); //0x1a600000UL, 0x3dc0b461UL
2580 fmul(3);
2581 fsuba(2);
2582 fxch(1);
2583 fsub(2);
2584 fsubrp(1);
2585 faddp(3);
2586 fld_d(ExternalAddress(16 + pi04_5d)); //0x2e000000UL, 0x3b93198aUL
2587 fmul(2);
2588 fld_s(0);
2589 fsubr(2);
2590 fsuba(2);
2591 fxch(1);
2592 fsubp(2);
2593 fxch(1);
2594 faddp(3);
2595 fld_d(ExternalAddress(24 + pi04_5d)); //0x25200000UL, 0x396b839aUL
2596 fmul(2);
2597 fld_s(0);
2598 fsubr(2);
2599 fsuba(2);
2600 fxch(1);
2601 fsubp(2);
2602 fxch(1);
2603 faddp(3);
2604 fld_d(ExternalAddress(32 + pi04_5d)); //0x533e63a0UL, 0x37027044UL
2605 fmulp(2);
2606 fld_s(1);
2607 fsubr(1);
2608 fsuba(1);
2609 fxch(2);
2610 fsubp(1);
2611 faddp(2);
2612 fxch(1);
2613 jmp(B1_15);
2614
2615 bind(B1_8);
2616 fld_x(Address(ebp, 8));
2617 addl(ebx, -16417);
2618 fmul_d(as_Address(ExternalAddress(SCALE))); //0x00000000UL, 0x32600000UL
2619 movl(eax, -2078209981);
2620 imull(ebx);
2621 addl(edx, ebx);
2622 movl(ecx, ebx);
2623 sarl(edx, 4);
2624 sarl(ecx, 31);
2625 subl(edx, ecx);
2626 movl(eax, edx);
2627 shll(eax, 5);
2628 fstp_x(Address(ebp, 8));
2629 fld_x(Address(ebp, 8));
2630 subl(eax, edx);
2631 movl(Address(ebp, 8), 0);
2632 subl(ebx, eax);
2633 fld_x(Address(ebp, 8));
2634 cmpl(ebx, 17);
2635 fsuba(1);
2636 jcc(Assembler::less, B1_10);
2637
2638 bind(B1_9);
2639 lea(eax, Address(noreg, edx, Address::times_8));
2640 lea(ecx, Address(eax, edx, Address::times_4));
2641 incl(edx);
2642 fld_x(Address(_4onpi_31l, RelocationHolder::none).plus_disp(ecx, Address::times_1));
2643 fmul(2);
2644 fld_x(Address(12 + _4onpi_31l, RelocationHolder::none).plus_disp(ecx, Address::times_1));
2645 fmul(2);
2646 fld_s(0);
2647 fadd(2);
2648 fsuba(2);
2649 fxch(1);
2650 faddp(2);
2651 fld_s(1);
2652 fadd(1);
2653 fstp_x(Address(esp, 8));
2654 andl(Address(esp, 8), -16777216);
2655 fld_x(Address(esp, 8));
2656 fsubp(1);
2657 jmp(B1_11);
2658
2659 bind(B1_10);
2660 fld_d(ExternalAddress(zeros)); //0x00000000UL, 0x00000000UL
2661 fld_s(0);
2662
2663 bind(B1_11);
2664 fld_s(0);
2665 lea(eax, Address(noreg, edx, Address::times_8));
2666 fld_s(3);
2667 lea(edx, Address(eax, edx, Address::times_4));
2668 fld_x(Address(_4onpi_31l, RelocationHolder::none).plus_disp(edx, Address::times_1));
2669 fmul(6);
2670 movl(Address(esp, 0), edx);
2671 fadda(2);
2672 fxch(2);
2673 fsuba(3);
2674 fxch(2);
2675 faddp(3);
2676 fxch(2);
2677 faddp(3);
2678 fld_x(Address(12 + _4onpi_31l, RelocationHolder::none).plus_disp(edx, Address::times_1));
2679 fmula(2);
2680 fld_s(2);
2681 fadd(2);
2682 fld_s(0);
2683 fxch(1);
2684 fsubra(3);
2685 fxch(3);
2686 fchs();
2687 faddp(4);
2688 fxch(3);
2689 faddp(4);
2690 fxch(2);
2691 fadd(3);
2692 fxch(2);
2693 fmul(5);
2694 fadda(2);
2695 fld_s(4);
2696 fld_x(Address(24 + _4onpi_31l, RelocationHolder::none).plus_disp(edx, Address::times_1));
2697 fmula(1);
2698 fxch(1);
2699 fadda(4);
2700 fxch(4);
2701 fstp_x(Address(esp, 8));
2702 movzwl(ebx, Address(esp, 16));
2703 andl(ebx, 32767);
2704 cmpl(ebx, 16415);
2705 jcc(Assembler::greaterEqual, B1_13);
2706
2707 bind(B1_12);
2708 negl(ebx);
2709 addl(ebx, 30);
2710 movl(ecx, ebx);
2711 movl(eax, Address(esp, 12));
2712 shrl(eax);
2713 shll(eax);
2714 movl(Address(esp, 12), eax);
2715 movl(Address(esp, 8), 0);
2716 shrl(eax);
2717 jmp(B1_14);
2718
2719 bind(B1_13);
2720 negl(ebx);
2721 addl(ebx, 30);
2722 movl(ecx, ebx);
2723 movl(edx, Address(esp, 8));
2724 shrl(edx);
2725 shll(edx);
2726 negl(ecx);
2727 movl(eax, Address(esp, 12));
2728 shll(eax);
2729 movl(ecx, ebx);
2730 movl(Address(esp, 8), edx);
2731 shrl(edx);
2732 orl(eax, edx);
2733
2734 bind(B1_14);
2735 fld_x(Address(esp, 8));
2736 addl(eax, Address(esp, 4));
2737 fsubp(3);
2738 fmul(6);
2739 fld_s(4);
2740 movl(edx, eax);
2741 andl(edx, 1);
2742 fadd(3);
2743 movl(ecx, Address(esp, 0));
2744 fsuba(3);
2745 fxch(3);
2746 faddp(5);
2747 fld_s(1);
2748 fxch(3);
2749 fadd_d(Address(zero_none, RelocationHolder::none).plus_disp(edx, Address::times_8));
2750 fadda(3);
2751 fsub(3);
2752 faddp(2);
2753 fxch(1);
2754 faddp(4);
2755 fld_s(2);
2756 fadd(2);
2757 fsuba(2);
2758 fxch(3);
2759 faddp(2);
2760 fxch(1);
2761 faddp(3);
2762 fld_s(0);
2763 fadd(2);
2764 fsuba(2);
2765 fxch(1);
2766 faddp(2);
2767 fxch(1);
2768 faddp(2);
2769 fld_s(2);
2770 fld_x(Address(36 + _4onpi_31l, RelocationHolder::none).plus_disp(ecx, Address::times_1));
2771 fmula(1);
2772 fld_s(1);
2773 fadd(3);
2774 fsuba(3);
2775 fxch(2);
2776 faddp(3);
2777 fxch(2);
2778 faddp(3);
2779 fxch(1);
2780 fmul(4);
2781 fld_s(0);
2782 fadd(2);
2783 fsuba(2);
2784 fxch(1);
2785 faddp(2);
2786 fxch(1);
2787 faddp(2);
2788 fld_s(2);
2789 fld_x(Address(48 + _4onpi_31l, RelocationHolder::none).plus_disp(ecx, Address::times_1));
2790 fmula(1);
2791 fld_s(1);
2792 fadd(3);
2793 fsuba(3);
2794 fxch(2);
2795 faddp(3);
2796 fxch(2);
2797 faddp(3);
2798 fld_s(3);
2799 fxch(2);
2800 fmul(5);
2801 fld_x(Address(60 + _4onpi_31l, RelocationHolder::none).plus_disp(ecx, Address::times_1));
2802 fmula(3);
2803 fxch(3);
2804 faddp(1);
2805 fld_s(0);
2806 fadd(2);
2807 fsuba(2);
2808 fxch(1);
2809 faddp(2);
2810 fxch(1);
2811 faddp(3);
2812 fld_s(3);
2813 fxch(2);
2814 fmul(5);
2815 fld_x(Address(72 + _4onpi_31l, RelocationHolder::none).plus_disp(ecx, Address::times_1));
2816 fmula(3);
2817 fxch(3);
2818 faddp(1);
2819 fld_s(0);
2820 fadd(2);
2821 fsuba(2);
2822 fxch(1);
2823 faddp(2);
2824 fxch(1);
2825 faddp(3);
2826 fxch(1);
2827 fmulp(4);
2828 fld_x(Address(84 + _4onpi_31l, RelocationHolder::none).plus_disp(ecx, Address::times_1));
2829 fmulp(3);
2830 fxch(2);
2831 faddp(3);
2832 fld_s(2);
2833 fadd(2);
2834 fld_d(ExternalAddress(TWO_32H)); //0x00000000UL, 0x41f80000UL
2835 fmul(1);
2836 fadda(1);
2837 fsubp(1);
2838 fsuba(2);
2839 fxch(3);
2840 faddp(2);
2841 faddp(1);
2842 fld_d(ExternalAddress(pi04_2d)); //0x54400000UL, 0x3fe921fbUL
2843 fld_s(0);
2844 fmul(2);
2845 fxch(2);
2846 fadd(3);
2847 fxch(1);
2848 fmulp(3);
2849 fmul_d(as_Address(ExternalAddress(8 + pi04_2d))); //0x1a626331UL, 0x3dc0b461UL
2850 faddp(1);
2851
2852 bind(B1_15);
2853 fld_d(ExternalAddress(TWO_12H)); //0x00000000UL, 0x40b80000UL
2854 fld_s(2);
2855 fadd(2);
2856 fmula(1);
2857 fstp_x(Address(esp, 8));
2858 fld_x(Address(esp, 8));
2859 fadd(1);
2860 fsubrp(1);
2861 fst_d(Address(esi, 0));
2862 fsubp(2);
2863 faddp(1);
2864 fstp_d(Address(esi, 8));
2865 addl(esp, 20);
2866 pop(ebx);
2867 pop(edi);
2868 pop(esi);
2869 movl(esp, ebp);
2870 pop(ebp);
2871 ret(0);
2872 }
2873
2874 ALIGNED_(16) juint _L_2il0floatpacket_0[] =
2875 {
2876 0xffffffffUL, 0x7fffffffUL, 0x00000000UL, 0x00000000UL
2877 };
2878
2879 ALIGNED_(16) juint _Pi4Inv[] =
2880 {
2881 0x6dc9c883UL, 0x3ff45f30UL
2882 };
2883
2884 ALIGNED_(16) juint _Pi4x3[] =
2885 {
2886 0x54443000UL, 0xbfe921fbUL, 0x3b39a000UL, 0x3d373dcbUL, 0xe0e68948UL,
2887 0xba845c06UL
2888 };
2889
2890 ALIGNED_(16) juint _Pi4x4[] =
2891 {
2892 0x54400000UL, 0xbfe921fbUL, 0x1a600000UL, 0xbdc0b461UL, 0x2e000000UL,
2893 0xbb93198aUL, 0x252049c1UL, 0xb96b839aUL
2894 };
2895
2896 ALIGNED_(16) jushort _SP[] =
2897 {
2898 0xaaab, 0xaaaa, 0xaaaa, 0xaaaa, 0xbffc, 0x0000, 0x8887, 0x8888, 0x8888, 0x8888,
2899 0x3ff8, 0x0000, 0xc527, 0x0d00, 0x00d0, 0xd00d, 0xbff2, 0x0000, 0x45f6, 0xb616,
2900 0x1d2a, 0xb8ef, 0x3fec, 0x0000, 0x825b, 0x3997, 0x2b3f, 0xd732, 0xbfe5, 0x0000,
2901 0xbf33, 0x8bb4, 0x2fda, 0xb092, 0x3fde, 0x0000, 0x44a6, 0xed1a, 0x29ef, 0xd73e,
2902 0xbfd6, 0x0000, 0x8610, 0x307f, 0x62a1, 0xc921, 0x3fce, 0x0000
2903 };
2904
2905 ALIGNED_(16) jushort _CP[] =
2906 {
2907 0x0000, 0x0000, 0x0000, 0x8000, 0xbffe, 0x0000, 0xaaa5, 0xaaaa, 0xaaaa, 0xaaaa,
2908 0x3ffa, 0x0000, 0x9c2f, 0x0b60, 0x60b6, 0xb60b, 0xbff5, 0x0000, 0xf024, 0x0cac,
2909 0x00d0, 0xd00d, 0x3fef, 0x0000, 0x03fe, 0x3f65, 0x7dbb, 0x93f2, 0xbfe9, 0x0000,
2910 0xd84d, 0xadee, 0xc698, 0x8f76, 0x3fe2, 0x0000, 0xdaba, 0xfe79, 0xea36, 0xc9c9,
2911 0xbfda, 0x0000, 0x3ac6, 0x0ba0, 0x07ce, 0xd585, 0x3fd2, 0x0000
2912 };
2913
2914 ALIGNED_(16) juint _ones[] =
2915 {
2916 0x00000000UL, 0x3ff00000UL, 0x00000000UL, 0xbff00000UL
2917 };
2918
2919 void MacroAssembler::libm_sincos_huge(XMMRegister xmm0, XMMRegister xmm1, Register eax, Register ecx, Register edx, Register ebx, Register esi, Register edi, Register ebp, Register esp) {
2920 Label B1_1, B1_2, B1_3, B1_4, B1_5, B1_6, B1_7, B1_8, B1_9, B1_10, B1_11, B1_12;
2921 Label B1_13, B1_14, B1_15, B1_16, B1_17, B1_18, B1_19, B1_20, B1_21, B1_22, B1_23;
2922 Label B1_24, B1_25, B1_26, B1_27, B1_28, B1_29, B1_30, B1_31, B1_32, B1_33, B1_34;
2923 Label B1_35, B1_36, B1_37, B1_38, B1_39, B1_40, B1_41, B1_42, B1_43, B1_44, B1_45, B1_46;
2924
2925 assert_different_registers(ebx, eax, ecx, edx, esi, edi, ebp, esp);
2926
2927 address L_2il0floatpacket_0 = (address)_L_2il0floatpacket_0;
2928 address Pi4Inv = (address)_Pi4Inv;
2929 address Pi4x3 = (address)_Pi4x3;
2930 address Pi4x4 = (address)_Pi4x4;
2931 address ones = (address)_ones;
2932 address CP = (address)_CP;
2933 address SP = (address)_SP;
2934
2935 bind(B1_1);
2936 push(ebp);
2937 movl(ebp, esp);
2938 andl(esp, -64);
2939 push(esi);
2940 push(edi);
2941 push(ebx);
2942 subl(esp, 52);
2943 movl(eax, Address(ebp, 16));
2944 movl(edx, Address(ebp, 20));
2945 movl(Address(esp, 32), eax);
2946 movl(Address(esp, 36), edx);
2947
2948 bind(B1_2);
2949 fnstcw(Address(esp, 30));
2950
2951 bind(B1_3);
2952 movsd(xmm1, Address(ebp, 8));
2953 movl(esi, Address(ebp, 12));
2954 movl(eax, esi);
2955 andl(eax, 2147483647);
2956 andps(xmm1, ExternalAddress(L_2il0floatpacket_0)); //0xffffffffUL, 0x7fffffffUL, 0x00000000UL, 0x00000000UL
2957 shrl(esi, 31);
2958 movl(Address(esp, 40), eax);
2959 cmpl(eax, 1104150528);
2960 movsd(Address(ebp, 8), xmm1);
2961 jcc(Assembler::aboveEqual, B1_11);
2962
2963 bind(B1_4);
2964 movsd(xmm0, ExternalAddress(Pi4Inv)); //0x6dc9c883UL, 0x3ff45f30UL
2965 mulsd(xmm0, xmm1);
2966 movzwl(edx, Address(esp, 30));
2967 movl(eax, edx);
2968 andl(eax, 768);
2969 movsd(Address(esp, 0), xmm0);
2970 cmpl(eax, 768);
2971 jcc(Assembler::equal, B1_42);
2972
2973 bind(B1_5);
2974 orl(edx, -64768);
2975 movw(Address(esp, 28), edx);
2976
2977 bind(B1_6);
2978 fldcw(Address(esp, 28));
2979
2980 bind(B1_7);
2981 movsd(xmm1, Address(ebp, 8));
2982 movl(ebx, 1);
2983
2984 bind(B1_8);
2985 movl(Address(esp, 12), ebx);
2986 movl(ebx, Address(esp, 4));
2987 movl(eax, ebx);
2988 movl(Address(esp, 8), esi);
2989 movl(esi, ebx);
2990 shrl(esi, 20);
2991 andl(eax, 1048575);
2992 movl(ecx, esi);
2993 orl(eax, 1048576);
2994 negl(ecx);
2995 movl(edx, eax);
2996 addl(ecx, 19);
2997 addl(esi, 13);
2998 movl(Address(esp, 24), ecx);
2999 shrl(edx);
3000 movl(ecx, esi);
3001 shll(eax);
3002 movl(ecx, Address(esp, 24));
3003 movl(esi, Address(esp, 0));
3004 shrl(esi);
3005 orl(eax, esi);
3006 cmpl(ebx, 1094713344);
3007 movsd(Address(esp, 16), xmm1);
3008 fld_d(Address(esp, 16));
3009 cmov32(Assembler::below, eax, edx);
3010 movl(esi, Address(esp, 8));
3011 lea(edx, Address(eax, 1));
3012 movl(ebx, edx);
3013 andl(ebx, -2);
3014 movl(Address(esp, 16), ebx);
3015 fild_s(Address(esp, 16));
3016 movl(ebx, Address(esp, 12));
3017 cmpl(Address(esp, 40), 1094713344);
3018 jcc(Assembler::aboveEqual, B1_10);
3019
3020 bind(B1_9);
3021 fld_d(ExternalAddress(Pi4x3)); //0x54443000UL, 0xbfe921fbUL
3022 fmul(1);
3023 faddp(2);
3024 fld_d(ExternalAddress(8 + Pi4x3)); //0x3b39a000UL, 0x3d373dcbUL
3025 fmul(1);
3026 faddp(2);
3027 fld_d(ExternalAddress(16 + Pi4x3)); //0xe0e68948UL, 0xba845c06UL
3028 fmulp(1);
3029 faddp(1);
3030 jmp(B1_17);
3031
3032 bind(B1_10);
3033 fld_d(ExternalAddress(Pi4x4)); //0x54400000UL, 0xbfe921fbUL
3034 fmul(1);
3035 faddp(2);
3036 fld_d(ExternalAddress(8 + Pi4x4)); //0x1a600000UL, 0xbdc0b461UL
3037 fmul(1);
3038 faddp(2);
3039 fld_d(ExternalAddress(16 + Pi4x4)); //0x2e000000UL, 0xbb93198aUL
3040 fmul(1);
3041 faddp(2);
3042 fld_d(ExternalAddress(24 + Pi4x4)); //0x252049c1UL, 0xb96b839aUL
3043 fmulp(1);
3044 faddp(1);
3045 jmp(B1_17);
3046
3047 bind(B1_11);
3048 movzwl(edx, Address(esp, 30));
3049 movl(eax, edx);
3050 andl(eax, 768);
3051 cmpl(eax, 768);
3052 jcc(Assembler::equal, B1_43);
3053 bind(B1_12);
3054 orl(edx, -64768);
3055 movw(Address(esp, 28), edx);
3056
3057 bind(B1_13);
3058 fldcw(Address(esp, 28));
3059
3060 bind(B1_14);
3061 movsd(xmm1, Address(ebp, 8));
3062 movl(ebx, 1);
3063
3064 bind(B1_15);
3065 movsd(Address(esp, 16), xmm1);
3066 fld_d(Address(esp, 16));
3067 addl(esp, -32);
3068 lea(eax, Address(esp, 32));
3069 fstp_x(Address(esp, 0));
3070 movl(Address(esp, 12), 0);
3071 movl(Address(esp, 16), eax);
3072 call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::dlibm_reduce_pi04l())));
3073
3074 bind(B1_46);
3075 addl(esp, 32);
3076
3077 bind(B1_16);
3078 fld_d(Address(esp, 0));
3079 lea(edx, Address(eax, 1));
3080 fld_d(Address(esp, 8));
3081 faddp(1);
3082
3083 bind(B1_17);
3084 movl(ecx, edx);
3085 addl(eax, 3);
3086 shrl(ecx, 2);
3087 andl(ecx, 1);
3088 shrl(eax, 2);
3089 xorl(esi, ecx);
3090 movl(ecx, Address(esp, 36));
3091 andl(eax, 1);
3092 andl(ecx, 3);
3093 cmpl(ecx, 3);
3094 jcc(Assembler::notEqual, B1_25);
3095
3096 bind(B1_18);
3097 fld_x(ExternalAddress(84 + SP)); //0x8610, 0x307f, 0x62
3098 fld_s(1);
3099 fmul((2));
3100 testb(edx, 2);
3101 fmula((1));
3102 fld_x(ExternalAddress(72 + SP)); //0x44a6, 0xed1a, 0x29
3103 faddp(2);
3104 fmula(1);
3105 fld_x(ExternalAddress(60 + SP)); //0xbf33, 0x8bb4, 0x2f
3106 faddp(2);
3107 fmula(1);
3108 fld_x(ExternalAddress(48 + SP)); //0x825b, 0x3997, 0x2b
3109 faddp(2);
3110 fmula(1);
3111 fld_x(ExternalAddress(36 + SP)); //0x45f6, 0xb616, 0x1d
3112 faddp(2);
3113 fmula(1);
3114 fld_x(ExternalAddress(24 + SP)); //0xc527, 0x0d00, 0x00
3115 faddp(2);
3116 fmula(1);
3117 fld_x(ExternalAddress(12 + SP)); //0x8887, 0x8888, 0x88
3118 faddp(2);
3119 fmula(1);
3120 fld_x(ExternalAddress(SP)); //0xaaab, 0xaaaa, 0xaa
3121 faddp(2);
3122 fmula(1);
3123 fld_x(ExternalAddress(84 + CP)); //0x3ac6, 0x0ba0, 0x07
3124 fmul(1);
3125 fld_x(ExternalAddress(72 + CP)); //0xdaba, 0xfe79, 0xea
3126 faddp(1);
3127 fmul(1);
3128 fld_x(ExternalAddress(62 + CP)); //0xd84d, 0xadee, 0xc6
3129 faddp(1);
3130 fmul(1);
3131 fld_x(ExternalAddress(48 + CP)); //0x03fe, 0x3f65, 0x7d
3132 faddp(1);
3133 fmul(1);
3134 fld_x(ExternalAddress(36 + CP)); //0xf024, 0x0cac, 0x00
3135 faddp(1);
3136 fmul(1);
3137 fld_x(ExternalAddress(24 + CP)); //0x9c2f, 0x0b60, 0x60
3138 faddp(1);
3139 fmul(1);
3140 fld_x(ExternalAddress(12 + CP)); //0xaaa5, 0xaaaa, 0xaa
3141 faddp(1);
3142 fmul(1);
3143 fld_x(ExternalAddress(CP)); //0x0000, 0x0000, 0x00
3144 faddp(1);
3145 fmulp(1);
3146 fld_d(Address(ones, RelocationHolder::none).plus_disp(esi, Address::times_8));
3147 fld_d(Address(ones, RelocationHolder::none).plus_disp(eax, Address::times_8));
3148 jcc(Assembler::equal, B1_22);
3149
3150 bind(B1_19);
3151 fmulp(4);
3152 testl(ebx, ebx);
3153 fxch(2);
3154 fmul(3);
3155 movl(eax, Address(esp, 2));
3156 faddp(3);
3157 fxch(2);
3158 fstp_d(Address(eax, 0));
3159 fmula(1);
3160 faddp(1);
3161 fstp_d(Address(eax, 8));
3162 jcc(Assembler::equal, B1_21);
3163
3164 bind(B1_20);
3165 fldcw(Address(esp, 30));
3166
3167 bind(B1_21);
3168 addl(esp, 52);
3169 pop(ebx);
3170 pop(edi);
3171 pop(esi);
3172 movl(esp, ebp);
3173 pop(ebp);
3174 ret(0);
3175
3176 bind(B1_22);
3177 fxch(1);
3178 fmulp(4);
3179 testl(ebx, ebx);
3180 fxch(2);
3181 fmul(3);
3182 movl(eax, Address(esp, 32));
3183 faddp(3);
3184 fxch(2);
3185 fstp_d(Address(eax, 8));
3186 fmula(1);
3187 faddp(1);
3188 fstp_d(Address(eax, 0));
3189 jcc(Assembler::equal, B1_24);
3190
3191 bind(B1_23);
3192 fldcw(Address(esp, 30));
3193
3194 bind(B1_24);
3195 addl(esp, 52);
3196 pop(ebx);
3197 pop(edi);
3198 pop(esi);
3199 movl(esp, ebp);
3200 pop(ebp);
3201 ret(0);
3202
3203 bind(B1_25);
3204 testb(Address(esp, 36), 2);
3205 jcc(Assembler::equal, B1_33);
3206
3207 bind(B1_26);
3208 fld_s(0);
3209 testb(edx, 2);
3210 fmul(1);
3211 fld_s(0);
3212 fmul(1);
3213 jcc(Assembler::equal, B1_30);
3214
3215 bind(B1_27);
3216 fstp_d(2);
3217 fld_x(ExternalAddress(84 + CP)); //0x3ac6, 0x0ba0, 0x07
3218 testl(ebx, ebx);
3219 fmul(2);
3220 fld_x(ExternalAddress(72 + CP)); //0xdaba, 0xfe79, 0xea
3221 fmul(3);
3222 fld_x(ExternalAddress(60 + CP)); //0xd84d, 0xadee, 0xc6
3223 movl(eax, Address(rsp, 32));
3224 faddp(2);
3225 fxch(1);
3226 fmul(3);
3227 fld_x(ExternalAddress(48 + CP)); //0x03fe, 0x3f65, 0x7d
3228 faddp(2);
3229 fxch(1);
3230 fmul(3);
3231 fld_x(ExternalAddress(36 + CP)); //0xf024, 0x0cac, 0x00
3232 faddp(2);
3233 fxch(1);
3234 fmul(3);
3235 fld_x(ExternalAddress(24 + CP)); //0x9c2f, 0x0b60, 0x60
3236 faddp(2);
3237 fxch(1);
3238 fmul(3);
3239 fld_x(ExternalAddress(12 + CP)); //0xaaa5, 0xaaaa, 0xaa
3240 faddp(2);
3241 fxch(1);
3242 fmulp(3);
3243 fld_x(ExternalAddress(CP)); //0x0000, 0x0000, 0x00
3244 faddp(1);
3245 fmulp(1);
3246 faddp(1);
3247 fld_d(Address(ones, RelocationHolder::none).plus_disp(rsi, Address::times_8));
3248 fmula(1);
3249 faddp(1);
3250 fstp_d(Address(eax, 8));
3251 jcc(Assembler::equal, B1_29);
3252
3253 bind(B1_28);
3254 fldcw(Address(esp, 30));
3255
3256 bind(B1_29);
3257 addl(esp, 52);
3258 pop(ebx);
3259 pop(edi);
3260 pop(esi);
3261 movl(esp, ebp);
3262 pop(ebp);
3263 ret(0);
3264
3265 bind(B1_30);
3266 fld_x(ExternalAddress(84 + SP)); //0x8610, 0x307f, 0x62
3267 testl(ebx, ebx);
3268 fmul(1);
3269 fld_x(ExternalAddress(72 + SP)); //0x44a6, 0xed1a, 0x29
3270 fmul(2);
3271 fld_x(ExternalAddress(60 + SP)); //0xbf33, 0x8bb4, 0x2f
3272 movl(eax, Address(rsp, 32));
3273 faddp(2);
3274 fxch(1);
3275 fmul(2);
3276 fld_x(ExternalAddress(48 + SP)); //0x825b, 0x3997, 0x2b
3277 faddp(2);
3278 fxch(1);
3279 fmul(2);
3280 fld_x(ExternalAddress(36 + SP)); //0x45f6, 0xb616, 0x1d
3281 faddp(2);
3282 fxch(1);
3283 fmul(2);
3284 fld_x(ExternalAddress(24 + SP)); //0xc527, 0x0d00, 0x00
3285 faddp(2);
3286 fxch(1);
3287 fmul(2);
3288 fld_x(ExternalAddress(12 + SP)); //0x8887, 0x8888, 0x88
3289 faddp(2);
3290 fxch(1);
3291 fmulp(2);
3292 fld_x(ExternalAddress(SP)); //0xaaab, 0xaaaa, 0xaa
3293 faddp(1);
3294 fmulp(2);
3295 faddp(1);
3296 fld_d(Address(ones, RelocationHolder::none).plus_disp(rsi, Address::times_8));
3297 fmulp(2);
3298 fmul(1);
3299 faddp(1);
3300 fstp_d(Address(eax, 8));
3301 jcc(Assembler::equal, B1_32);
3302
3303 bind(B1_31);
3304 fldcw(Address(esp, 30));
3305
3306 bind(B1_32);
3307 addl(esp, 52);
3308 pop(ebx);
3309 pop(edi);
3310 pop(esi);
3311 movl(esp, ebp);
3312 pop(ebp);
3313 ret(0);
3314
3315 bind(B1_33);
3316 testb(Address(esp, 36), 1);
3317 jcc(Assembler::equal, B1_41);
3318
3319 bind(B1_34);
3320 fld_s(0);
3321 testb(edx, 2);
3322 fmul(1);
3323 fld_s(0);
3324 fmul(1);
3325 jcc(Assembler::equal, B1_38);
3326
3327 bind(B1_35);
3328 fld_x(ExternalAddress(84 + SP)); //0x8610, 0x307f, 0x62
3329 testl(ebx, ebx);
3330 fmul(1);
3331 fld_x(ExternalAddress(72 + SP)); //0x44a6, 0xed1a, 0x29
3332 fmul(2);
3333 fld_x(ExternalAddress(60 + SP)); //0xbf33, 0x8bb4, 0x2f
3334 faddp(2);
3335 fxch(1);
3336 fmul(2);
3337 fld_x(ExternalAddress(48 + SP)); //0x825b, 0x3997, 0x2b
3338 faddp(2);
3339 fxch(1);
3340 fmul(2);
3341 fld_x(ExternalAddress(36 + SP)); //0x45f6, 0xb616, 0x1d
3342 faddp(2);
3343 fxch(1);
3344 fmul(2);
3345 fld_x(ExternalAddress(24 + SP)); //0xc527, 0x0d00, 0x00
3346 faddp(2);
3347 fxch(1);
3348 fmul(2);
3349 fld_x(ExternalAddress(12 + SP)); //0x8887, 0x8888, 0x88
3350 faddp(2);
3351 fxch(1);
3352 fmulp(2);
3353 fld_x(ExternalAddress(SP)); //0xaaab, 0xaaaa, 0xaa
3354 faddp(1);
3355 fmulp(2);
3356 faddp(1);
3357 fld_d(Address(ones, RelocationHolder::none).plus_disp(eax, Address::times_8));
3358 fmulp(2);
3359 fmul(1);
3360 movl(eax, Address(esp, 32));
3361 faddp(1);
3362 fstp_d(Address(eax, 0));
3363 jcc(Assembler::equal, B1_37);
3364
3365 bind(B1_36);
3366 fldcw(Address(esp, 30));
3367
3368 bind(B1_37);
3369 addl(esp, 52);
3370 pop(ebx);
3371 pop(edi);
3372 pop(esi);
3373 movl(esp, ebp);
3374 pop(ebp);
3375 ret(0);
3376
3377 bind(B1_38);
3378 fstp_d(2);
3379 fld_x(ExternalAddress(84 + CP)); //0x3ac6, 0x0ba0, 0x07
3380 testl(ebx, ebx);
3381 fmul(2);
3382 fld_x(ExternalAddress(72 + CP)); //0xdaba, 0xfe79, 0xea
3383 fmul(3);
3384 fld_x(ExternalAddress(60 + CP)); //0xd84d, 0xadee, 0xc6
3385 faddp(2);
3386 fxch(1);
3387 fmul(3);
3388 fld_x(ExternalAddress(48 + CP)); //0x03fe, 0x3f65, 0x7d
3389 faddp(2);
3390 fxch(1);
3391 fmul(3);
3392 fld_x(ExternalAddress(36 + CP)); //0xf024, 0x0cac, 0x00
3393 faddp(2);
3394 fxch(1);
3395 fmul(3);
3396 fld_x(ExternalAddress(24 + CP)); //0x9c2f, 0x0b60, 0x60
3397 faddp(2);
3398 fxch(1);
3399 fmul(3);
3400 fld_x(ExternalAddress(12 + CP)); //0xaaa5, 0xaaaa, 0xaa
3401 faddp(2);
3402 fxch(1);
3403 fmulp(3);
3404 fld_x(ExternalAddress(CP)); //0x0000, 0x0000, 0x00
3405 faddp(1);
3406 fmulp(1);
3407 faddp(1);
3408 fld_d(Address(ones, RelocationHolder::none).plus_disp(eax, Address::times_8));
3409 fmula(1);
3410 movl(eax, Address(esp, 32));
3411 faddp(1);
3412 fstp_d(Address(eax, 0));
3413 jcc(Assembler::equal, B1_40);
3414
3415 bind(B1_39);
3416 fldcw(Address(esp, 30));
3417 bind(B1_40);
3418 addl(esp, 52);
3419 pop(ebx);
3420 pop(edi);
3421 pop(esi);
3422 movl(esp, ebp);
3423 pop(ebp);
3424 ret(0);
3425 bind(B1_41);
3426 fstp_d(0);
3427 addl(esp, 52);
3428 pop(ebx);
3429 pop(edi);
3430 pop(esi);
3431 movl(esp, ebp);
3432 pop(ebp);
3433 ret(0);
3434 bind(B1_42);
3435 xorl(ebx, ebx);
3436 jmp(B1_8);
3437 bind(B1_43);
3438 xorl(ebx, ebx);
3439 jmp(B1_15);
3440 }
3441
3442 ALIGNED_(16) juint _static_const_table_sin[] =
3443 {
3444 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
3445 0x00000000UL, 0x00000000UL, 0x3ff00000UL, 0x176d6d31UL, 0xbf73b92eUL,
3446 0xbc29b42cUL, 0x3fb917a6UL, 0xe0000000UL, 0xbc3e2718UL, 0x00000000UL,
3447 0x3ff00000UL, 0x011469fbUL, 0xbf93ad06UL, 0x3c69a60bUL, 0x3fc8f8b8UL,
3448 0xc0000000UL, 0xbc626d19UL, 0x00000000UL, 0x3ff00000UL, 0x939d225aUL,
3449 0xbfa60beaUL, 0x2ed59f06UL, 0x3fd29406UL, 0xa0000000UL, 0xbc75d28dUL,
3450 0x00000000UL, 0x3ff00000UL, 0x866b95cfUL, 0xbfb37ca1UL, 0xa6aea963UL,
3451 0x3fd87de2UL, 0xe0000000UL, 0xbc672cedUL, 0x00000000UL, 0x3ff00000UL,
3452 0x73fa1279UL, 0xbfbe3a68UL, 0x3806f63bUL, 0x3fde2b5dUL, 0x20000000UL,
3453 0x3c5e0d89UL, 0x00000000UL, 0x3ff00000UL, 0x5bc57974UL, 0xbfc59267UL,
3454 0x39ae68c8UL, 0x3fe1c73bUL, 0x20000000UL, 0x3c8b25ddUL, 0x00000000UL,
3455 0x3ff00000UL, 0x53aba2fdUL, 0xbfcd0dfeUL, 0x25091dd6UL, 0x3fe44cf3UL,
3456 0x20000000UL, 0x3c68076aUL, 0x00000000UL, 0x3ff00000UL, 0x99fcef32UL,
3457 0x3fca8279UL, 0x667f3bcdUL, 0x3fe6a09eUL, 0x20000000UL, 0xbc8bdd34UL,
3458 0x00000000UL, 0x3fe00000UL, 0x94247758UL, 0x3fc133ccUL, 0x6b151741UL,
3459 0x3fe8bc80UL, 0x20000000UL, 0xbc82c5e1UL, 0x00000000UL, 0x3fe00000UL,
3460 0x9ae68c87UL, 0x3fac73b3UL, 0x290ea1a3UL, 0x3fea9b66UL, 0xe0000000UL,
3461 0x3c39f630UL, 0x00000000UL, 0x3fe00000UL, 0x7f909c4eUL, 0xbf9d4a2cUL,
3462 0xf180bdb1UL, 0x3fec38b2UL, 0x80000000UL, 0xbc76e0b1UL, 0x00000000UL,
3463 0x3fe00000UL, 0x65455a75UL, 0xbfbe0875UL, 0xcf328d46UL, 0x3fed906bUL,
3464 0x20000000UL, 0x3c7457e6UL, 0x00000000UL, 0x3fe00000UL, 0x76acf82dUL,
3465 0x3fa4a031UL, 0x56c62ddaUL, 0x3fee9f41UL, 0xe0000000UL, 0x3c8760b1UL,
3466 0x00000000UL, 0x3fd00000UL, 0x0e5967d5UL, 0xbfac1d1fUL, 0xcff75cb0UL,
3467 0x3fef6297UL, 0x20000000UL, 0x3c756217UL, 0x00000000UL, 0x3fd00000UL,
3468 0x0f592f50UL, 0xbf9ba165UL, 0xa3d12526UL, 0x3fefd88dUL, 0x40000000UL,
3469 0xbc887df6UL, 0x00000000UL, 0x3fc00000UL, 0x00000000UL, 0x00000000UL,
3470 0x00000000UL, 0x3ff00000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
3471 0x00000000UL, 0x0f592f50UL, 0x3f9ba165UL, 0xa3d12526UL, 0x3fefd88dUL,
3472 0x40000000UL, 0xbc887df6UL, 0x00000000UL, 0xbfc00000UL, 0x0e5967d5UL,
3473 0x3fac1d1fUL, 0xcff75cb0UL, 0x3fef6297UL, 0x20000000UL, 0x3c756217UL,
3474 0x00000000UL, 0xbfd00000UL, 0x76acf82dUL, 0xbfa4a031UL, 0x56c62ddaUL,
3475 0x3fee9f41UL, 0xe0000000UL, 0x3c8760b1UL, 0x00000000UL, 0xbfd00000UL,
3476 0x65455a75UL, 0x3fbe0875UL, 0xcf328d46UL, 0x3fed906bUL, 0x20000000UL,
3477 0x3c7457e6UL, 0x00000000UL, 0xbfe00000UL, 0x7f909c4eUL, 0x3f9d4a2cUL,
3478 0xf180bdb1UL, 0x3fec38b2UL, 0x80000000UL, 0xbc76e0b1UL, 0x00000000UL,
3479 0xbfe00000UL, 0x9ae68c87UL, 0xbfac73b3UL, 0x290ea1a3UL, 0x3fea9b66UL,
3480 0xe0000000UL, 0x3c39f630UL, 0x00000000UL, 0xbfe00000UL, 0x94247758UL,
3481 0xbfc133ccUL, 0x6b151741UL, 0x3fe8bc80UL, 0x20000000UL, 0xbc82c5e1UL,
3482 0x00000000UL, 0xbfe00000UL, 0x99fcef32UL, 0xbfca8279UL, 0x667f3bcdUL,
3483 0x3fe6a09eUL, 0x20000000UL, 0xbc8bdd34UL, 0x00000000UL, 0xbfe00000UL,
3484 0x53aba2fdUL, 0x3fcd0dfeUL, 0x25091dd6UL, 0x3fe44cf3UL, 0x20000000UL,
3485 0x3c68076aUL, 0x00000000UL, 0xbff00000UL, 0x5bc57974UL, 0x3fc59267UL,
3486 0x39ae68c8UL, 0x3fe1c73bUL, 0x20000000UL, 0x3c8b25ddUL, 0x00000000UL,
3487 0xbff00000UL, 0x73fa1279UL, 0x3fbe3a68UL, 0x3806f63bUL, 0x3fde2b5dUL,
3488 0x20000000UL, 0x3c5e0d89UL, 0x00000000UL, 0xbff00000UL, 0x866b95cfUL,
3489 0x3fb37ca1UL, 0xa6aea963UL, 0x3fd87de2UL, 0xe0000000UL, 0xbc672cedUL,
3490 0x00000000UL, 0xbff00000UL, 0x939d225aUL, 0x3fa60beaUL, 0x2ed59f06UL,
3491 0x3fd29406UL, 0xa0000000UL, 0xbc75d28dUL, 0x00000000UL, 0xbff00000UL,
3492 0x011469fbUL, 0x3f93ad06UL, 0x3c69a60bUL, 0x3fc8f8b8UL, 0xc0000000UL,
3493 0xbc626d19UL, 0x00000000UL, 0xbff00000UL, 0x176d6d31UL, 0x3f73b92eUL,
3494 0xbc29b42cUL, 0x3fb917a6UL, 0xe0000000UL, 0xbc3e2718UL, 0x00000000UL,
3495 0xbff00000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
3496 0x00000000UL, 0x00000000UL, 0x00000000UL, 0xbff00000UL, 0x176d6d31UL,
3497 0x3f73b92eUL, 0xbc29b42cUL, 0xbfb917a6UL, 0xe0000000UL, 0x3c3e2718UL,
3498 0x00000000UL, 0xbff00000UL, 0x011469fbUL, 0x3f93ad06UL, 0x3c69a60bUL,
3499 0xbfc8f8b8UL, 0xc0000000UL, 0x3c626d19UL, 0x00000000UL, 0xbff00000UL,
3500 0x939d225aUL, 0x3fa60beaUL, 0x2ed59f06UL, 0xbfd29406UL, 0xa0000000UL,
3501 0x3c75d28dUL, 0x00000000UL, 0xbff00000UL, 0x866b95cfUL, 0x3fb37ca1UL,
3502 0xa6aea963UL, 0xbfd87de2UL, 0xe0000000UL, 0x3c672cedUL, 0x00000000UL,
3503 0xbff00000UL, 0x73fa1279UL, 0x3fbe3a68UL, 0x3806f63bUL, 0xbfde2b5dUL,
3504 0x20000000UL, 0xbc5e0d89UL, 0x00000000UL, 0xbff00000UL, 0x5bc57974UL,
3505 0x3fc59267UL, 0x39ae68c8UL, 0xbfe1c73bUL, 0x20000000UL, 0xbc8b25ddUL,
3506 0x00000000UL, 0xbff00000UL, 0x53aba2fdUL, 0x3fcd0dfeUL, 0x25091dd6UL,
3507 0xbfe44cf3UL, 0x20000000UL, 0xbc68076aUL, 0x00000000UL, 0xbff00000UL,
3508 0x99fcef32UL, 0xbfca8279UL, 0x667f3bcdUL, 0xbfe6a09eUL, 0x20000000UL,
3509 0x3c8bdd34UL, 0x00000000UL, 0xbfe00000UL, 0x94247758UL, 0xbfc133ccUL,
3510 0x6b151741UL, 0xbfe8bc80UL, 0x20000000UL, 0x3c82c5e1UL, 0x00000000UL,
3511 0xbfe00000UL, 0x9ae68c87UL, 0xbfac73b3UL, 0x290ea1a3UL, 0xbfea9b66UL,
3512 0xe0000000UL, 0xbc39f630UL, 0x00000000UL, 0xbfe00000UL, 0x7f909c4eUL,
3513 0x3f9d4a2cUL, 0xf180bdb1UL, 0xbfec38b2UL, 0x80000000UL, 0x3c76e0b1UL,
3514 0x00000000UL, 0xbfe00000UL, 0x65455a75UL, 0x3fbe0875UL, 0xcf328d46UL,
3515 0xbfed906bUL, 0x20000000UL, 0xbc7457e6UL, 0x00000000UL, 0xbfe00000UL,
3516 0x76acf82dUL, 0xbfa4a031UL, 0x56c62ddaUL, 0xbfee9f41UL, 0xe0000000UL,
3517 0xbc8760b1UL, 0x00000000UL, 0xbfd00000UL, 0x0e5967d5UL, 0x3fac1d1fUL,
3518 0xcff75cb0UL, 0xbfef6297UL, 0x20000000UL, 0xbc756217UL, 0x00000000UL,
3519 0xbfd00000UL, 0x0f592f50UL, 0x3f9ba165UL, 0xa3d12526UL, 0xbfefd88dUL,
3520 0x40000000UL, 0x3c887df6UL, 0x00000000UL, 0xbfc00000UL, 0x00000000UL,
3521 0x00000000UL, 0x00000000UL, 0xbff00000UL, 0x00000000UL, 0x00000000UL,
3522 0x00000000UL, 0x00000000UL, 0x0f592f50UL, 0xbf9ba165UL, 0xa3d12526UL,
3523 0xbfefd88dUL, 0x40000000UL, 0x3c887df6UL, 0x00000000UL, 0x3fc00000UL,
3524 0x0e5967d5UL, 0xbfac1d1fUL, 0xcff75cb0UL, 0xbfef6297UL, 0x20000000UL,
3525 0xbc756217UL, 0x00000000UL, 0x3fd00000UL, 0x76acf82dUL, 0x3fa4a031UL,
3526 0x56c62ddaUL, 0xbfee9f41UL, 0xe0000000UL, 0xbc8760b1UL, 0x00000000UL,
3527 0x3fd00000UL, 0x65455a75UL, 0xbfbe0875UL, 0xcf328d46UL, 0xbfed906bUL,
3528 0x20000000UL, 0xbc7457e6UL, 0x00000000UL, 0x3fe00000UL, 0x7f909c4eUL,
3529 0xbf9d4a2cUL, 0xf180bdb1UL, 0xbfec38b2UL, 0x80000000UL, 0x3c76e0b1UL,
3530 0x00000000UL, 0x3fe00000UL, 0x9ae68c87UL, 0x3fac73b3UL, 0x290ea1a3UL,
3531 0xbfea9b66UL, 0xe0000000UL, 0xbc39f630UL, 0x00000000UL, 0x3fe00000UL,
3532 0x94247758UL, 0x3fc133ccUL, 0x6b151741UL, 0xbfe8bc80UL, 0x20000000UL,
3533 0x3c82c5e1UL, 0x00000000UL, 0x3fe00000UL, 0x99fcef32UL, 0x3fca8279UL,
3534 0x667f3bcdUL, 0xbfe6a09eUL, 0x20000000UL, 0x3c8bdd34UL, 0x00000000UL,
3535 0x3fe00000UL, 0x53aba2fdUL, 0xbfcd0dfeUL, 0x25091dd6UL, 0xbfe44cf3UL,
3536 0x20000000UL, 0xbc68076aUL, 0x00000000UL, 0x3ff00000UL, 0x5bc57974UL,
3537 0xbfc59267UL, 0x39ae68c8UL, 0xbfe1c73bUL, 0x20000000UL, 0xbc8b25ddUL,
3538 0x00000000UL, 0x3ff00000UL, 0x73fa1279UL, 0xbfbe3a68UL, 0x3806f63bUL,
3539 0xbfde2b5dUL, 0x20000000UL, 0xbc5e0d89UL, 0x00000000UL, 0x3ff00000UL,
3540 0x866b95cfUL, 0xbfb37ca1UL, 0xa6aea963UL, 0xbfd87de2UL, 0xe0000000UL,
3541 0x3c672cedUL, 0x00000000UL, 0x3ff00000UL, 0x939d225aUL, 0xbfa60beaUL,
3542 0x2ed59f06UL, 0xbfd29406UL, 0xa0000000UL, 0x3c75d28dUL, 0x00000000UL,
3543 0x3ff00000UL, 0x011469fbUL, 0xbf93ad06UL, 0x3c69a60bUL, 0xbfc8f8b8UL,
3544 0xc0000000UL, 0x3c626d19UL, 0x00000000UL, 0x3ff00000UL, 0x176d6d31UL,
3545 0xbf73b92eUL, 0xbc29b42cUL, 0xbfb917a6UL, 0xe0000000UL, 0x3c3e2718UL,
3546 0x00000000UL, 0x3ff00000UL, 0x55555555UL, 0xbfc55555UL, 0x00000000UL,
3547 0xbfe00000UL, 0x11111111UL, 0x3f811111UL, 0x55555555UL, 0x3fa55555UL,
3548 0x1a01a01aUL, 0xbf2a01a0UL, 0x16c16c17UL, 0xbf56c16cUL, 0xa556c734UL,
3549 0x3ec71de3UL, 0x1a01a01aUL, 0x3efa01a0UL, 0x1a600000UL, 0x3d90b461UL,
3550 0x1a600000UL, 0x3d90b461UL, 0x54400000UL, 0x3fb921fbUL, 0x00000000UL,
3551 0x00000000UL, 0x2e037073UL, 0x3b63198aUL, 0x00000000UL, 0x00000000UL,
3552 0x6dc9c883UL, 0x40245f30UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
3553 0x43380000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x43600000UL,
3554 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x3c800000UL, 0x00000000UL,
3555 0x00000000UL, 0xffffffffUL, 0x3fefffffUL, 0x00000000UL, 0x00000000UL,
3556 0x00000000UL, 0x80000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
3557 0x80000000UL, 0x00000000UL, 0x80000000UL, 0x00000000UL, 0x3fe00000UL,
3558 0x00000000UL, 0x3fe00000UL
3559 };
3560
3561 void MacroAssembler::fast_sin(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ebx, Register edx) {
3562
3563 Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
3564 Label L_2TAG_PACKET_4_0_2, start;
3565 assert_different_registers(eax, ebx, edx);
3566 address static_const_table_sin = (address)_static_const_table_sin;
3567
3568 bind(start);
3569 subl(rsp, 120);
3570 movl(Address(rsp, 56), ebx);
3571 lea(ebx, ExternalAddress(static_const_table_sin));
3572 movsd(xmm0, Address(rsp, 128));
3573 pextrw(eax, xmm0, 3);
3574 andl(eax, 32767);
3575 subl(eax, 12336);
3576 cmpl(eax, 4293);
3577 jcc(Assembler::above, L_2TAG_PACKET_0_0_2);
3578 movsd(xmm1, Address(ebx, 2160));
3579 mulsd(xmm1, xmm0);
3580 movsd(xmm5, Address(ebx, 2272));
3581 movdqu(xmm4, Address(ebx, 2256));
3582 pand(xmm4, xmm0);
3583 por(xmm5, xmm4);
3584 movsd(xmm3, Address(ebx, 2128));
3585 movdqu(xmm2, Address(ebx, 2112));
3586 addpd(xmm1, xmm5);
3587 cvttsd2sil(edx, xmm1);
3588 cvtsi2sdl(xmm1, edx);
3589 mulsd(xmm3, xmm1);
3590 unpcklpd(xmm1, xmm1);
3591 addl(edx, 1865216);
3592 movdqu(xmm4, xmm0);
3593 andl(edx, 63);
3594 movdqu(xmm5, Address(ebx, 2096));
3595 lea(eax, Address(ebx, 0));
3596 shll(edx, 5);
3597 addl(eax, edx);
3598 mulpd(xmm2, xmm1);
3599 subsd(xmm0, xmm3);
3600 mulsd(xmm1, Address(ebx, 2144));
3601 subsd(xmm4, xmm3);
3602 movsd(xmm7, Address(eax, 8));
3603 unpcklpd(xmm0, xmm0);
3604 movapd(xmm3, xmm4);
3605 subsd(xmm4, xmm2);
3606 mulpd(xmm5, xmm0);
3607 subpd(xmm0, xmm2);
3608 movdqu(xmm6, Address(ebx, 2064));
3609 mulsd(xmm7, xmm4);
3610 subsd(xmm3, xmm4);
3611 mulpd(xmm5, xmm0);
3612 mulpd(xmm0, xmm0);
3613 subsd(xmm3, xmm2);
3614 movdqu(xmm2, Address(eax, 0));
3615 subsd(xmm1, xmm3);
3616 movsd(xmm3, Address(eax, 24));
3617 addsd(xmm2, xmm3);
3618 subsd(xmm7, xmm2);
3619 mulsd(xmm2, xmm4);
3620 mulpd(xmm6, xmm0);
3621 mulsd(xmm3, xmm4);
3622 mulpd(xmm2, xmm0);
3623 mulpd(xmm0, xmm0);
3624 addpd(xmm5, Address(ebx, 2080));
3625 mulsd(xmm4, Address(eax, 0));
3626 addpd(xmm6, Address(ebx, 2048));
3627 mulpd(xmm5, xmm0);
3628 movapd(xmm0, xmm3);
3629 addsd(xmm3, Address(eax, 8));
3630 mulpd(xmm1, xmm7);
3631 movapd(xmm7, xmm4);
3632 addsd(xmm4, xmm3);
3633 addpd(xmm6, xmm5);
3634 movsd(xmm5, Address(eax, 8));
3635 subsd(xmm5, xmm3);
3636 subsd(xmm3, xmm4);
3637 addsd(xmm1, Address(eax, 16));
3638 mulpd(xmm6, xmm2);
3639 addsd(xmm5, xmm0);
3640 addsd(xmm3, xmm7);
3641 addsd(xmm1, xmm5);
3642 addsd(xmm1, xmm3);
3643 addsd(xmm1, xmm6);
3644 unpckhpd(xmm6, xmm6);
3645 addsd(xmm1, xmm6);
3646 addsd(xmm4, xmm1);
3647 movsd(Address(rsp, 0), xmm4);
3648 fld_d(Address(rsp, 0));
3649 jmp(L_2TAG_PACKET_1_0_2);
3650
3651 bind(L_2TAG_PACKET_0_0_2);
3652 jcc(Assembler::greater, L_2TAG_PACKET_2_0_2);
3653 shrl(eax, 4);
3654 cmpl(eax, 268434685);
3655 jcc(Assembler::notEqual, L_2TAG_PACKET_3_0_2);
3656 movsd(Address(rsp, 0), xmm0);
3657 fld_d(Address(rsp, 0));
3658 jmp(L_2TAG_PACKET_1_0_2);
3659
3660 bind(L_2TAG_PACKET_3_0_2);
3661 movsd(xmm3, Address(ebx, 2192));
3662 mulsd(xmm3, xmm0);
3663 subsd(xmm3, xmm0);
3664 mulsd(xmm3, Address(ebx, 2208));
3665 movsd(Address(rsp, 0), xmm0);
3666 fld_d(Address(rsp, 0));
3667 jmp(L_2TAG_PACKET_1_0_2);
3668
3669 bind(L_2TAG_PACKET_2_0_2);
3670 movl(eax, Address(rsp, 132));
3671 andl(eax, 2146435072);
3672 cmpl(eax, 2146435072);
3673 jcc(Assembler::equal, L_2TAG_PACKET_4_0_2);
3674 subl(rsp, 32);
3675 movsd(Address(rsp, 0), xmm0);
3676 lea(eax, Address(rsp, 40));
3677 movl(Address(rsp, 8), eax);
3678 movl(eax, 2);
3679 movl(Address(rsp, 12), eax);
3680 call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::dlibm_sin_cos_huge())));
3681 addl(rsp, 32);
3682 fld_d(Address(rsp, 16));
3683 jmp(L_2TAG_PACKET_1_0_2);
3684 bind(L_2TAG_PACKET_4_0_2);
3685 fld_d(Address(rsp, 128));
3686 fmul_d(Address(ebx, 2240));
3687 bind(L_2TAG_PACKET_1_0_2);
3688 movl(ebx, Address(rsp, 56));
3689 }
3690
3691 #endif
3692
3693 /******************************************************************************/
3694 // ALGORITHM DESCRIPTION - COS()
3695 // ---------------------
3696 //
3697 // 1. RANGE REDUCTION
3698 //
3699 // We perform an initial range reduction from X to r with
3700 //
3701 // X =~= N * pi/32 + r
3702 //
3703 // so that |r| <= pi/64 + epsilon. We restrict inputs to those
3704 // where |N| <= 932560. Beyond this, the range reduction is
3705 // insufficiently accurate. For extremely small inputs,
3706 // denormalization can occur internally, impacting performance.
3707 // This means that the main path is actually only taken for
3708 // 2^-252 <= |X| < 90112.
3709 //
3710 // To avoid branches, we perform the range reduction to full
3711 // accuracy each time.
3712 //
3713 // X - N * (P_1 + P_2 + P_3)
3714 //
3715 // where P_1 and P_2 are 32-bit numbers (so multiplication by N
3716 // is exact) and P_3 is a 53-bit number. Together, these
3717 // approximate pi well enough for all cases in the restricted
3718 // range.
3719 //
3720 // The main reduction sequence is:
3721 //
3722 // y = 32/pi * x
3723 // N = integer(y)
3724 // (computed by adding and subtracting off SHIFTER)
3725 //
3726 // m_1 = N * P_1
3727 // m_2 = N * P_2
3728 // r_1 = x - m_1
3729 // r = r_1 - m_2
3730 // (this r can be used for most of the calculation)
3731 //
3732 // c_1 = r_1 - r
3733 // m_3 = N * P_3
3734 // c_2 = c_1 - m_2
3735 // c = c_2 - m_3
3736 //
3737 // 2. MAIN ALGORITHM
3738 //
3739 // The algorithm uses a table lookup based on B = M * pi / 32
3740 // where M = N mod 64. The stored values are:
3741 // sigma closest power of 2 to cos(B)
3742 // C_hl 53-bit cos(B) - sigma
3743 // S_hi + S_lo 2 * 53-bit sin(B)
3744 //
3745 // The computation is organized as follows:
3746 //
3747 // sin(B + r + c) = [sin(B) + sigma * r] +
3748 // r * (cos(B) - sigma) +
3749 // sin(B) * [cos(r + c) - 1] +
3750 // cos(B) * [sin(r + c) - r]
3751 //
3752 // which is approximately:
3753 //
3754 // [S_hi + sigma * r] +
3755 // C_hl * r +
3756 // S_lo + S_hi * [(cos(r) - 1) - r * c] +
3757 // (C_hl + sigma) * [(sin(r) - r) + c]
3758 //
3759 // and this is what is actually computed. We separate this sum
3760 // into four parts:
3761 //
3762 // hi + med + pols + corr
3763 //
3764 // where
3765 //
3766 // hi = S_hi + sigma r
3767 // med = C_hl * r
3768 // pols = S_hi * (cos(r) - 1) + (C_hl + sigma) * (sin(r) - r)
3769 // corr = S_lo + c * ((C_hl + sigma) - S_hi * r)
3770 //
3771 // 3. POLYNOMIAL
3772 //
3773 // The polynomial S_hi * (cos(r) - 1) + (C_hl + sigma) *
3774 // (sin(r) - r) can be rearranged freely, since it is quite
3775 // small, so we exploit parallelism to the fullest.
3776 //
3777 // psc4 = SC_4 * r_1
3778 // msc4 = psc4 * r
3779 // r2 = r * r
3780 // msc2 = SC_2 * r2
3781 // r4 = r2 * r2
3782 // psc3 = SC_3 + msc4
3783 // psc1 = SC_1 + msc2
3784 // msc3 = r4 * psc3
3785 // sincospols = psc1 + msc3
3786 // pols = sincospols *
3787 // <S_hi * r^2 | (C_hl + sigma) * r^3>
3788 //
3789 // 4. CORRECTION TERM
3790 //
3791 // This is where the "c" component of the range reduction is
3792 // taken into account; recall that just "r" is used for most of
3793 // the calculation.
3794 //
3795 // -c = m_3 - c_2
3796 // -d = S_hi * r - (C_hl + sigma)
3797 // corr = -c * -d + S_lo
3798 //
3799 // 5. COMPENSATED SUMMATIONS
3800 //
3801 // The two successive compensated summations add up the high
3802 // and medium parts, leaving just the low parts to add up at
3803 // the end.
3804 //
3805 // rs = sigma * r
3806 // res_int = S_hi + rs
3807 // k_0 = S_hi - res_int
3808 // k_2 = k_0 + rs
3809 // med = C_hl * r
3810 // res_hi = res_int + med
3811 // k_1 = res_int - res_hi
3812 // k_3 = k_1 + med
3813 //
3814 // 6. FINAL SUMMATION
3815 //
3816 // We now add up all the small parts:
3817 //
3818 // res_lo = pols(hi) + pols(lo) + corr + k_1 + k_3
3819 //
3820 // Now the overall result is just:
3821 //
3822 // res_hi + res_lo
3823 //
3824 // 7. SMALL ARGUMENTS
3825 //
3826 // Inputs with |X| < 2^-252 are treated specially as
3827 // 1 - |x|.
3828 //
3829 // Special cases:
3830 // cos(NaN) = quiet NaN, and raise invalid exception
3831 // cos(INF) = NaN and raise invalid exception
3832 // cos(0) = 1
3833 //
3834 /******************************************************************************/
3835
3836 #ifdef _LP64
3837
3838 ALIGNED_(8) juint _ONE[] =
3839 {
3840 0x00000000UL, 0x3ff00000UL
3841 };
3842
3843 void MacroAssembler::fast_cos(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register r8, Register r9, Register r10, Register r11) {
3844 Label L_2TAG_PACKET_0_0_1, L_2TAG_PACKET_1_0_1, L_2TAG_PACKET_2_0_1, L_2TAG_PACKET_3_0_1;
3845 Label L_2TAG_PACKET_4_0_1, L_2TAG_PACKET_5_0_1, L_2TAG_PACKET_6_0_1, L_2TAG_PACKET_7_0_1;
3846 Label L_2TAG_PACKET_8_0_1, L_2TAG_PACKET_9_0_1, L_2TAG_PACKET_10_0_1, L_2TAG_PACKET_11_0_1;
3847 Label L_2TAG_PACKET_12_0_1, L_2TAG_PACKET_13_0_1, B1_2, B1_3, B1_4, B1_5, start;
3848
3849 assert_different_registers(r8, r9, r10, r11, eax, ecx, edx);
3850
3851 address ONEHALF = (address)_ONEHALF;
3852 address P_2 = (address)_P_2;
3853 address SC_4 = (address)_SC_4;
3854 address Ctable = (address)_Ctable;
3855 address SC_2 = (address)_SC_2;
3856 address SC_3 = (address)_SC_3;
3857 address SC_1 = (address)_SC_1;
3858 address PI_INV_TABLE = (address)_PI_INV_TABLE;
3859 address PI_4 = (address)_PI_4;
3860 address PI32INV = (address)_PI32INV;
3861 address SIGN_MASK = (address)_SIGN_MASK;
3862 address P_1 = (address)_P_1;
3863 address P_3 = (address)_P_3;
3864 address ONE = (address)_ONE;
3865 address NEG_ZERO = (address)_NEG_ZERO;
3866
3867 bind(start);
3868 push(rbx);
3869 subq(rsp, 16);
3870 movsd(Address(rsp, 8), xmm0);
3871
3872 bind(B1_2);
3873 movl(eax, Address(rsp, 12));
3874 movq(xmm1, ExternalAddress(PI32INV)); //0x6dc9c883UL, 0x40245f30UL
3875 andl(eax, 2147418112);
3876 subl(eax, 808452096);
3877 cmpl(eax, 281346048);
3878 jcc(Assembler::above, L_2TAG_PACKET_0_0_1);
3879 mulsd(xmm1, xmm0);
3880 movdqu(xmm5, ExternalAddress(ONEHALF)); //0x00000000UL, 0x3fe00000UL, 0x00000000UL, 0x3fe00000UL
3881 movq(xmm4, ExternalAddress(SIGN_MASK)); //0x00000000UL, 0x80000000UL
3882 pand(xmm4, xmm0);
3883 por(xmm5, xmm4);
3884 addpd(xmm1, xmm5);
3885 cvttsd2sil(edx, xmm1);
3886 cvtsi2sdl(xmm1, edx);
3887 movdqu(xmm2, ExternalAddress(P_2)); //0x1a600000UL, 0x3d90b461UL, 0x1a600000UL, 0x3d90b461UL
3888 movq(xmm3, ExternalAddress(P_1)); //0x54400000UL, 0x3fb921fbUL
3889 mulsd(xmm3, xmm1);
3890 unpcklpd(xmm1, xmm1);
3891 addq(rdx, 1865232);
3892 movdqu(xmm4, xmm0);
3893 andq(rdx, 63);
3894 movdqu(xmm5, ExternalAddress(SC_4)); //0xa556c734UL, 0x3ec71de3UL, 0x1a01a01aUL, 0x3efa01a0UL
3895 lea(rax, ExternalAddress(Ctable));
3896 shlq(rdx, 5);
3897 addq(rax, rdx);
3898 mulpd(xmm2, xmm1);
3899 subsd(xmm0, xmm3);
3900 mulsd(xmm1, ExternalAddress(P_3)); //0x2e037073UL, 0x3b63198aUL
3901 subsd(xmm4, xmm3);
3902 movq(xmm7, Address(rax, 8));
3903 unpcklpd(xmm0, xmm0);
3904 movdqu(xmm3, xmm4);
3905 subsd(xmm4, xmm2);
3906 mulpd(xmm5, xmm0);
3907 subpd(xmm0, xmm2);
3908 movdqu(xmm6, ExternalAddress(SC_2)); //0x11111111UL, 0x3f811111UL, 0x55555555UL, 0x3fa55555UL
3909 mulsd(xmm7, xmm4);
3910 subsd(xmm3, xmm4);
3911 mulpd(xmm5, xmm0);
3912 mulpd(xmm0, xmm0);
3913 subsd(xmm3, xmm2);
3914 movdqu(xmm2, Address(rax, 0));
3915 subsd(xmm1, xmm3);
3916 movq(xmm3, Address(rax, 24));
3917 addsd(xmm2, xmm3);
3918 subsd(xmm7, xmm2);
3919 mulsd(xmm2, xmm4);
3920 mulpd(xmm6, xmm0);
3921 mulsd(xmm3, xmm4);
3922 mulpd(xmm2, xmm0);
3923 mulpd(xmm0, xmm0);
3924 addpd(xmm5, ExternalAddress(SC_3)); //0x1a01a01aUL, 0xbf2a01a0UL, 0x16c16c17UL, 0xbf56c16cUL
3925 mulsd(xmm4, Address(rax, 0));
3926 addpd(xmm6, ExternalAddress(SC_1)); //0x55555555UL, 0xbfc55555UL, 0x00000000UL, 0xbfe00000UL
3927 mulpd(xmm5, xmm0);
3928 movdqu(xmm0, xmm3);
3929 addsd(xmm3, Address(rax, 8));
3930 mulpd(xmm1, xmm7);
3931 movdqu(xmm7, xmm4);
3932 addsd(xmm4, xmm3);
3933 addpd(xmm6, xmm5);
3934 movq(xmm5, Address(rax, 8));
3935 subsd(xmm5, xmm3);
3936 subsd(xmm3, xmm4);
3937 addsd(xmm1, Address(rax, 16));
3938 mulpd(xmm6, xmm2);
3939 addsd(xmm0, xmm5);
3940 addsd(xmm3, xmm7);
3941 addsd(xmm0, xmm1);
3942 addsd(xmm0, xmm3);
3943 addsd(xmm0, xmm6);
3944 unpckhpd(xmm6, xmm6);
3945 addsd(xmm0, xmm6);
3946 addsd(xmm0, xmm4);
3947 jmp(B1_4);
3948
3949 bind(L_2TAG_PACKET_0_0_1);
3950 jcc(Assembler::greater, L_2TAG_PACKET_1_0_1);
3951 pextrw(eax, xmm0, 3);
3952 andl(eax, 32767);
3953 pinsrw(xmm0, eax, 3);
3954 movq(xmm1, ExternalAddress(ONE)); //0x00000000UL, 0x3ff00000UL
3955 subsd(xmm1, xmm0);
3956 movdqu(xmm0, xmm1);
3957 jmp(B1_4);
3958
3959 bind(L_2TAG_PACKET_1_0_1);
3960 pextrw(eax, xmm0, 3);
3961 andl(eax, 32752);
3962 cmpl(eax, 32752);
3963 jcc(Assembler::equal, L_2TAG_PACKET_2_0_1);
3964 pextrw(ecx, xmm0, 3);
3965 andl(ecx, 32752);
3966 subl(ecx, 16224);
3967 shrl(ecx, 7);
3968 andl(ecx, 65532);
3969 lea(r11, ExternalAddress(PI_INV_TABLE));
3970 addq(rcx, r11);
3971 movdq(rax, xmm0);
3972 movl(r10, Address(rcx, 20));
3973 movl(r8, Address(rcx, 24));
3974 movl(edx, eax);
3975 shrq(rax, 21);
3976 orl(eax, INT_MIN);
3977 shrl(eax, 11);
3978 movl(r9, r10);
3979 imulq(r10, rdx);
3980 imulq(r9, rax);
3981 imulq(r8, rax);
3982 movl(rsi, Address(rcx, 16));
3983 movl(rdi, Address(rcx, 12));
3984 movl(r11, r10);
3985 shrq(r10, 32);
3986 addq(r9, r10);
3987 addq(r11, r8);
3988 movl(r8, r11);
3989 shrq(r11, 32);
3990 addq(r9, r11);
3991 movl(r10, rsi);
3992 imulq(rsi, rdx);
3993 imulq(r10, rax);
3994 movl(r11, rdi);
3995 imulq(rdi, rdx);
3996 movl(rbx, rsi);
3997 shrq(rsi, 32);
3998 addq(r9, rbx);
3999 movl(rbx, r9);
4000 shrq(r9, 32);
4001 addq(r10, rsi);
4002 addq(r10, r9);
4003 shlq(rbx, 32);
4004 orq(r8, rbx);
4005 imulq(r11, rax);
4006 movl(r9, Address(rcx, 8));
4007 movl(rsi, Address(rcx, 4));
4008 movl(rbx, rdi);
4009 shrq(rdi, 32);
4010 addq(r10, rbx);
4011 movl(rbx, r10);
4012 shrq(r10, 32);
4013 addq(r11, rdi);
4014 addq(r11, r10);
4015 movq(rdi, r9);
4016 imulq(r9, rdx);
4017 imulq(rdi, rax);
4018 movl(r10, r9);
4019 shrq(r9, 32);
4020 addq(r11, r10);
4021 movl(r10, r11);
4022 shrq(r11, 32);
4023 addq(rdi, r9);
4024 addq(rdi, r11);
4025 movq(r9, rsi);
4026 imulq(rsi, rdx);
4027 imulq(r9, rax);
4028 shlq(r10, 32);
4029 orq(r10, rbx);
4030 movl(eax, Address(rcx, 0));
4031 movl(r11, rsi);
4032 shrq(rsi, 32);
4033 addq(rdi, r11);
4034 movl(r11, rdi);
4035 shrq(rdi, 32);
4036 addq(r9, rsi);
4037 addq(r9, rdi);
4038 imulq(rdx, rax);
4039 pextrw(rbx, xmm0, 3);
4040 lea(rdi, ExternalAddress(PI_INV_TABLE));
4041 subq(rcx, rdi);
4042 addl(ecx, ecx);
4043 addl(ecx, ecx);
4044 addl(ecx, ecx);
4045 addl(ecx, 19);
4046 movl(rsi, 32768);
4047 andl(rsi, rbx);
4048 shrl(rbx, 4);
4049 andl(rbx, 2047);
4050 subl(rbx, 1023);
4051 subl(ecx, rbx);
4052 addq(r9, rdx);
4053 movl(edx, ecx);
4054 addl(edx, 32);
4055 cmpl(ecx, 1);
4056 jcc(Assembler::less, L_2TAG_PACKET_3_0_1);
4057 negl(ecx);
4058 addl(ecx, 29);
4059 shll(r9);
4060 movl(rdi, r9);
4061 andl(r9, 536870911);
4062 testl(r9, 268435456);
4063 jcc(Assembler::notEqual, L_2TAG_PACKET_4_0_1);
4064 shrl(r9);
4065 movl(rbx, 0);
4066 shlq(r9, 32);
4067 orq(r9, r11);
4068
4069 bind(L_2TAG_PACKET_5_0_1);
4070
4071 bind(L_2TAG_PACKET_6_0_1);
4072 cmpq(r9, 0);
4073 jcc(Assembler::equal, L_2TAG_PACKET_7_0_1);
4074
4075 bind(L_2TAG_PACKET_8_0_1);
4076 bsrq(r11, r9);
4077 movl(ecx, 29);
4078 subl(ecx, r11);
4079 jcc(Assembler::lessEqual, L_2TAG_PACKET_9_0_1);
4080 shlq(r9);
4081 movq(rax, r10);
4082 shlq(r10);
4083 addl(edx, ecx);
4084 negl(ecx);
4085 addl(ecx, 64);
4086 shrq(rax);
4087 shrq(r8);
4088 orq(r9, rax);
4089 orq(r10, r8);
4090
4091 bind(L_2TAG_PACKET_10_0_1);
4092 cvtsi2sdq(xmm0, r9);
4093 shrq(r10, 1);
4094 cvtsi2sdq(xmm3, r10);
4095 xorpd(xmm4, xmm4);
4096 shll(edx, 4);
4097 negl(edx);
4098 addl(edx, 16368);
4099 orl(edx, rsi);
4100 xorl(edx, rbx);
4101 pinsrw(xmm4, edx, 3);
4102 movq(xmm2, ExternalAddress(PI_4)); //0x40000000UL, 0x3fe921fbUL, 0x18469899UL, 0x3e64442dUL
4103 movq(xmm6, ExternalAddress(8 + PI_4)); //0x3fe921fbUL, 0x18469899UL, 0x3e64442dUL
4104 xorpd(xmm5, xmm5);
4105 subl(edx, 1008);
4106 pinsrw(xmm5, edx, 3);
4107 mulsd(xmm0, xmm4);
4108 shll(rsi, 16);
4109 sarl(rsi, 31);
4110 mulsd(xmm3, xmm5);
4111 movdqu(xmm1, xmm0);
4112 mulsd(xmm0, xmm2);
4113 shrl(rdi, 29);
4114 addsd(xmm1, xmm3);
4115 mulsd(xmm3, xmm2);
4116 addl(rdi, rsi);
4117 xorl(rdi, rsi);
4118 mulsd(xmm6, xmm1);
4119 movl(eax, rdi);
4120 addsd(xmm6, xmm3);
4121 movdqu(xmm2, xmm0);
4122 addsd(xmm0, xmm6);
4123 subsd(xmm2, xmm0);
4124 addsd(xmm6, xmm2);
4125
4126 bind(L_2TAG_PACKET_11_0_1);
4127 movq(xmm1, ExternalAddress(PI32INV)); //0x6dc9c883UL, 0x40245f30UL
4128 mulsd(xmm1, xmm0);
4129 movq(xmm5, ExternalAddress(ONEHALF)); //0x00000000UL, 0x3fe00000UL, 0x00000000UL, 0x3fe00000UL
4130 movq(xmm4, ExternalAddress(SIGN_MASK)); //0x00000000UL, 0x80000000UL
4131 pand(xmm4, xmm0);
4132 por(xmm5, xmm4);
4133 addpd(xmm1, xmm5);
4134 cvttsd2siq(rdx, xmm1);
4135 cvtsi2sdq(xmm1, rdx);
4136 movq(xmm3, ExternalAddress(P_1)); //0x54400000UL, 0x3fb921fbUL
4137 movdqu(xmm2, ExternalAddress(P_2)); //0x1a600000UL, 0x3d90b461UL, 0x1a600000UL, 0x3d90b461UL
4138 mulsd(xmm3, xmm1);
4139 unpcklpd(xmm1, xmm1);
4140 shll(eax, 3);
4141 addl(edx, 1865232);
4142 movdqu(xmm4, xmm0);
4143 addl(edx, eax);
4144 andl(edx, 63);
4145 movdqu(xmm5, ExternalAddress(SC_4)); //0xa556c734UL, 0x3ec71de3UL, 0x1a01a01aUL, 0x3efa01a0UL
4146 lea(rax, ExternalAddress(Ctable));
4147 shll(edx, 5);
4148 addq(rax, rdx);
4149 mulpd(xmm2, xmm1);
4150 subsd(xmm0, xmm3);
4151 mulsd(xmm1, ExternalAddress(P_3)); //0x2e037073UL, 0x3b63198aUL
4152 subsd(xmm4, xmm3);
4153 movq(xmm7, Address(rax, 8));
4154 unpcklpd(xmm0, xmm0);
4155 movdqu(xmm3, xmm4);
4156 subsd(xmm4, xmm2);
4157 mulpd(xmm5, xmm0);
4158 subpd(xmm0, xmm2);
4159 mulsd(xmm7, xmm4);
4160 subsd(xmm3, xmm4);
4161 mulpd(xmm5, xmm0);
4162 mulpd(xmm0, xmm0);
4163 subsd(xmm3, xmm2);
4164 movdqu(xmm2, Address(rax, 0));
4165 subsd(xmm1, xmm3);
4166 movq(xmm3, Address(rax, 24));
4167 addsd(xmm2, xmm3);
4168 subsd(xmm7, xmm2);
4169 subsd(xmm1, xmm6);
4170 movdqu(xmm6, ExternalAddress(SC_2)); //0x11111111UL, 0x3f811111UL, 0x55555555UL, 0x3fa55555UL
4171 mulsd(xmm2, xmm4);
4172 mulpd(xmm6, xmm0);
4173 mulsd(xmm3, xmm4);
4174 mulpd(xmm2, xmm0);
4175 mulpd(xmm0, xmm0);
4176 addpd(xmm5, ExternalAddress(SC_3)); //0x1a01a01aUL, 0xbf2a01a0UL, 0x16c16c17UL, 0xbf56c16cUL
4177 mulsd(xmm4, Address(rax, 0));
4178 addpd(xmm6, ExternalAddress(SC_1)); //0x55555555UL, 0xbfc55555UL, 0x00000000UL, 0xbfe00000UL
4179 mulpd(xmm5, xmm0);
4180 movdqu(xmm0, xmm3);
4181 addsd(xmm3, Address(rax, 8));
4182 mulpd(xmm1, xmm7);
4183 movdqu(xmm7, xmm4);
4184 addsd(xmm4, xmm3);
4185 addpd(xmm6, xmm5);
4186 movq(xmm5, Address(rax, 8));
4187 subsd(xmm5, xmm3);
4188 subsd(xmm3, xmm4);
4189 addsd(xmm1, Address(rax, 16));
4190 mulpd(xmm6, xmm2);
4191 addsd(xmm5, xmm0);
4192 addsd(xmm3, xmm7);
4193 addsd(xmm1, xmm5);
4194 addsd(xmm1, xmm3);
4195 addsd(xmm1, xmm6);
4196 unpckhpd(xmm6, xmm6);
4197 movdqu(xmm0, xmm4);
4198 addsd(xmm1, xmm6);
4199 addsd(xmm0, xmm1);
4200 jmp(B1_4);
4201
4202 bind(L_2TAG_PACKET_7_0_1);
4203 addl(edx, 64);
4204 movq(r9, r10);
4205 movq(r10, r8);
4206 movl(r8, 0);
4207 cmpq(r9, 0);
4208 jcc(Assembler::notEqual, L_2TAG_PACKET_8_0_1);
4209 addl(edx, 64);
4210 movq(r9, r10);
4211 movq(r10, r8);
4212 cmpq(r9, 0);
4213 jcc(Assembler::notEqual, L_2TAG_PACKET_8_0_1);
4214 xorpd(xmm0, xmm0);
4215 xorpd(xmm6, xmm6);
4216 jmp(L_2TAG_PACKET_11_0_1);
4217
4218 bind(L_2TAG_PACKET_9_0_1);
4219 jcc(Assembler::equal, L_2TAG_PACKET_10_0_1);
4220 negl(ecx);
4221 shrq(r10);
4222 movq(rax, r9);
4223 shrq(r9);
4224 subl(edx, ecx);
4225 negl(ecx);
4226 addl(ecx, 64);
4227 shlq(rax);
4228 orq(r10, rax);
4229 jmp(L_2TAG_PACKET_10_0_1);
4230 bind(L_2TAG_PACKET_3_0_1);
4231 negl(ecx);
4232 shlq(r9, 32);
4233 orq(r9, r11);
4234 shlq(r9);
4235 movq(rdi, r9);
4236 testl(r9, INT_MIN);
4237 jcc(Assembler::notEqual, L_2TAG_PACKET_12_0_1);
4238 shrl(r9);
4239 movl(rbx, 0);
4240 shrq(rdi, 3);
4241 jmp(L_2TAG_PACKET_6_0_1);
4242
4243 bind(L_2TAG_PACKET_4_0_1);
4244 shrl(r9);
4245 movl(rbx, 536870912);
4246 shrl(rbx);
4247 shlq(r9, 32);
4248 orq(r9, r11);
4249 shlq(rbx, 32);
4250 addl(rdi, 536870912);
4251 movl(rcx, 0);
4252 movl(r11, 0);
4253 subq(rcx, r8);
4254 sbbq(r11, r10);
4255 sbbq(rbx, r9);
4256 movq(r8, rcx);
4257 movq(r10, r11);
4258 movq(r9, rbx);
4259 movl(rbx, 32768);
4260 jmp(L_2TAG_PACKET_5_0_1);
4261
4262 bind(L_2TAG_PACKET_12_0_1);
4263 shrl(r9);
4264 mov64(rbx, 0x100000000);
4265 shrq(rbx);
4266 movl(rcx, 0);
4267 movl(r11, 0);
4268 subq(rcx, r8);
4269 sbbq(r11, r10);
4270 sbbq(rbx, r9);
4271 movq(r8, rcx);
4272 movq(r10, r11);
4273 movq(r9, rbx);
4274 movl(rbx, 32768);
4275 shrq(rdi, 3);
4276 addl(rdi, 536870912);
4277 jmp(L_2TAG_PACKET_6_0_1);
4278
4279 bind(L_2TAG_PACKET_2_0_1);
4280 movsd(xmm0, Address(rsp, 8));
4281 mulsd(xmm0, ExternalAddress(NEG_ZERO)); //0x00000000UL, 0x80000000UL
4282 movq(Address(rsp, 0), xmm0);
4283
4284 bind(L_2TAG_PACKET_13_0_1);
4285
4286 bind(B1_4);
4287 addq(rsp, 16);
4288 pop(rbx);
4289 }
4290
4291 #endif
4292
4293 #ifndef _LP64
4294
4295 ALIGNED_(16) juint _static_const_table_cos[] =
4296 {
4297 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
4298 0x00000000UL, 0x00000000UL, 0x3ff00000UL, 0x176d6d31UL, 0xbf73b92eUL,
4299 0xbc29b42cUL, 0x3fb917a6UL, 0xe0000000UL, 0xbc3e2718UL, 0x00000000UL,
4300 0x3ff00000UL, 0x011469fbUL, 0xbf93ad06UL, 0x3c69a60bUL, 0x3fc8f8b8UL,
4301 0xc0000000UL, 0xbc626d19UL, 0x00000000UL, 0x3ff00000UL, 0x939d225aUL,
4302 0xbfa60beaUL, 0x2ed59f06UL, 0x3fd29406UL, 0xa0000000UL, 0xbc75d28dUL,
4303 0x00000000UL, 0x3ff00000UL, 0x866b95cfUL, 0xbfb37ca1UL, 0xa6aea963UL,
4304 0x3fd87de2UL, 0xe0000000UL, 0xbc672cedUL, 0x00000000UL, 0x3ff00000UL,
4305 0x73fa1279UL, 0xbfbe3a68UL, 0x3806f63bUL, 0x3fde2b5dUL, 0x20000000UL,
4306 0x3c5e0d89UL, 0x00000000UL, 0x3ff00000UL, 0x5bc57974UL, 0xbfc59267UL,
4307 0x39ae68c8UL, 0x3fe1c73bUL, 0x20000000UL, 0x3c8b25ddUL, 0x00000000UL,
4308 0x3ff00000UL, 0x53aba2fdUL, 0xbfcd0dfeUL, 0x25091dd6UL, 0x3fe44cf3UL,
4309 0x20000000UL, 0x3c68076aUL, 0x00000000UL, 0x3ff00000UL, 0x99fcef32UL,
4310 0x3fca8279UL, 0x667f3bcdUL, 0x3fe6a09eUL, 0x20000000UL, 0xbc8bdd34UL,
4311 0x00000000UL, 0x3fe00000UL, 0x94247758UL, 0x3fc133ccUL, 0x6b151741UL,
4312 0x3fe8bc80UL, 0x20000000UL, 0xbc82c5e1UL, 0x00000000UL, 0x3fe00000UL,
4313 0x9ae68c87UL, 0x3fac73b3UL, 0x290ea1a3UL, 0x3fea9b66UL, 0xe0000000UL,
4314 0x3c39f630UL, 0x00000000UL, 0x3fe00000UL, 0x7f909c4eUL, 0xbf9d4a2cUL,
4315 0xf180bdb1UL, 0x3fec38b2UL, 0x80000000UL, 0xbc76e0b1UL, 0x00000000UL,
4316 0x3fe00000UL, 0x65455a75UL, 0xbfbe0875UL, 0xcf328d46UL, 0x3fed906bUL,
4317 0x20000000UL, 0x3c7457e6UL, 0x00000000UL, 0x3fe00000UL, 0x76acf82dUL,
4318 0x3fa4a031UL, 0x56c62ddaUL, 0x3fee9f41UL, 0xe0000000UL, 0x3c8760b1UL,
4319 0x00000000UL, 0x3fd00000UL, 0x0e5967d5UL, 0xbfac1d1fUL, 0xcff75cb0UL,
4320 0x3fef6297UL, 0x20000000UL, 0x3c756217UL, 0x00000000UL, 0x3fd00000UL,
4321 0x0f592f50UL, 0xbf9ba165UL, 0xa3d12526UL, 0x3fefd88dUL, 0x40000000UL,
4322 0xbc887df6UL, 0x00000000UL, 0x3fc00000UL, 0x00000000UL, 0x00000000UL,
4323 0x00000000UL, 0x3ff00000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
4324 0x00000000UL, 0x0f592f50UL, 0x3f9ba165UL, 0xa3d12526UL, 0x3fefd88dUL,
4325 0x40000000UL, 0xbc887df6UL, 0x00000000UL, 0xbfc00000UL, 0x0e5967d5UL,
4326 0x3fac1d1fUL, 0xcff75cb0UL, 0x3fef6297UL, 0x20000000UL, 0x3c756217UL,
4327 0x00000000UL, 0xbfd00000UL, 0x76acf82dUL, 0xbfa4a031UL, 0x56c62ddaUL,
4328 0x3fee9f41UL, 0xe0000000UL, 0x3c8760b1UL, 0x00000000UL, 0xbfd00000UL,
4329 0x65455a75UL, 0x3fbe0875UL, 0xcf328d46UL, 0x3fed906bUL, 0x20000000UL,
4330 0x3c7457e6UL, 0x00000000UL, 0xbfe00000UL, 0x7f909c4eUL, 0x3f9d4a2cUL,
4331 0xf180bdb1UL, 0x3fec38b2UL, 0x80000000UL, 0xbc76e0b1UL, 0x00000000UL,
4332 0xbfe00000UL, 0x9ae68c87UL, 0xbfac73b3UL, 0x290ea1a3UL, 0x3fea9b66UL,
4333 0xe0000000UL, 0x3c39f630UL, 0x00000000UL, 0xbfe00000UL, 0x94247758UL,
4334 0xbfc133ccUL, 0x6b151741UL, 0x3fe8bc80UL, 0x20000000UL, 0xbc82c5e1UL,
4335 0x00000000UL, 0xbfe00000UL, 0x99fcef32UL, 0xbfca8279UL, 0x667f3bcdUL,
4336 0x3fe6a09eUL, 0x20000000UL, 0xbc8bdd34UL, 0x00000000UL, 0xbfe00000UL,
4337 0x53aba2fdUL, 0x3fcd0dfeUL, 0x25091dd6UL, 0x3fe44cf3UL, 0x20000000UL,
4338 0x3c68076aUL, 0x00000000UL, 0xbff00000UL, 0x5bc57974UL, 0x3fc59267UL,
4339 0x39ae68c8UL, 0x3fe1c73bUL, 0x20000000UL, 0x3c8b25ddUL, 0x00000000UL,
4340 0xbff00000UL, 0x73fa1279UL, 0x3fbe3a68UL, 0x3806f63bUL, 0x3fde2b5dUL,
4341 0x20000000UL, 0x3c5e0d89UL, 0x00000000UL, 0xbff00000UL, 0x866b95cfUL,
4342 0x3fb37ca1UL, 0xa6aea963UL, 0x3fd87de2UL, 0xe0000000UL, 0xbc672cedUL,
4343 0x00000000UL, 0xbff00000UL, 0x939d225aUL, 0x3fa60beaUL, 0x2ed59f06UL,
4344 0x3fd29406UL, 0xa0000000UL, 0xbc75d28dUL, 0x00000000UL, 0xbff00000UL,
4345 0x011469fbUL, 0x3f93ad06UL, 0x3c69a60bUL, 0x3fc8f8b8UL, 0xc0000000UL,
4346 0xbc626d19UL, 0x00000000UL, 0xbff00000UL, 0x176d6d31UL, 0x3f73b92eUL,
4347 0xbc29b42cUL, 0x3fb917a6UL, 0xe0000000UL, 0xbc3e2718UL, 0x00000000UL,
4348 0xbff00000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
4349 0x00000000UL, 0x00000000UL, 0x00000000UL, 0xbff00000UL, 0x176d6d31UL,
4350 0x3f73b92eUL, 0xbc29b42cUL, 0xbfb917a6UL, 0xe0000000UL, 0x3c3e2718UL,
4351 0x00000000UL, 0xbff00000UL, 0x011469fbUL, 0x3f93ad06UL, 0x3c69a60bUL,
4352 0xbfc8f8b8UL, 0xc0000000UL, 0x3c626d19UL, 0x00000000UL, 0xbff00000UL,
4353 0x939d225aUL, 0x3fa60beaUL, 0x2ed59f06UL, 0xbfd29406UL, 0xa0000000UL,
4354 0x3c75d28dUL, 0x00000000UL, 0xbff00000UL, 0x866b95cfUL, 0x3fb37ca1UL,
4355 0xa6aea963UL, 0xbfd87de2UL, 0xe0000000UL, 0x3c672cedUL, 0x00000000UL,
4356 0xbff00000UL, 0x73fa1279UL, 0x3fbe3a68UL, 0x3806f63bUL, 0xbfde2b5dUL,
4357 0x20000000UL, 0xbc5e0d89UL, 0x00000000UL, 0xbff00000UL, 0x5bc57974UL,
4358 0x3fc59267UL, 0x39ae68c8UL, 0xbfe1c73bUL, 0x20000000UL, 0xbc8b25ddUL,
4359 0x00000000UL, 0xbff00000UL, 0x53aba2fdUL, 0x3fcd0dfeUL, 0x25091dd6UL,
4360 0xbfe44cf3UL, 0x20000000UL, 0xbc68076aUL, 0x00000000UL, 0xbff00000UL,
4361 0x99fcef32UL, 0xbfca8279UL, 0x667f3bcdUL, 0xbfe6a09eUL, 0x20000000UL,
4362 0x3c8bdd34UL, 0x00000000UL, 0xbfe00000UL, 0x94247758UL, 0xbfc133ccUL,
4363 0x6b151741UL, 0xbfe8bc80UL, 0x20000000UL, 0x3c82c5e1UL, 0x00000000UL,
4364 0xbfe00000UL, 0x9ae68c87UL, 0xbfac73b3UL, 0x290ea1a3UL, 0xbfea9b66UL,
4365 0xe0000000UL, 0xbc39f630UL, 0x00000000UL, 0xbfe00000UL, 0x7f909c4eUL,
4366 0x3f9d4a2cUL, 0xf180bdb1UL, 0xbfec38b2UL, 0x80000000UL, 0x3c76e0b1UL,
4367 0x00000000UL, 0xbfe00000UL, 0x65455a75UL, 0x3fbe0875UL, 0xcf328d46UL,
4368 0xbfed906bUL, 0x20000000UL, 0xbc7457e6UL, 0x00000000UL, 0xbfe00000UL,
4369 0x76acf82dUL, 0xbfa4a031UL, 0x56c62ddaUL, 0xbfee9f41UL, 0xe0000000UL,
4370 0xbc8760b1UL, 0x00000000UL, 0xbfd00000UL, 0x0e5967d5UL, 0x3fac1d1fUL,
4371 0xcff75cb0UL, 0xbfef6297UL, 0x20000000UL, 0xbc756217UL, 0x00000000UL,
4372 0xbfd00000UL, 0x0f592f50UL, 0x3f9ba165UL, 0xa3d12526UL, 0xbfefd88dUL,
4373 0x40000000UL, 0x3c887df6UL, 0x00000000UL, 0xbfc00000UL, 0x00000000UL,
4374 0x00000000UL, 0x00000000UL, 0xbff00000UL, 0x00000000UL, 0x00000000UL,
4375 0x00000000UL, 0x00000000UL, 0x0f592f50UL, 0xbf9ba165UL, 0xa3d12526UL,
4376 0xbfefd88dUL, 0x40000000UL, 0x3c887df6UL, 0x00000000UL, 0x3fc00000UL,
4377 0x0e5967d5UL, 0xbfac1d1fUL, 0xcff75cb0UL, 0xbfef6297UL, 0x20000000UL,
4378 0xbc756217UL, 0x00000000UL, 0x3fd00000UL, 0x76acf82dUL, 0x3fa4a031UL,
4379 0x56c62ddaUL, 0xbfee9f41UL, 0xe0000000UL, 0xbc8760b1UL, 0x00000000UL,
4380 0x3fd00000UL, 0x65455a75UL, 0xbfbe0875UL, 0xcf328d46UL, 0xbfed906bUL,
4381 0x20000000UL, 0xbc7457e6UL, 0x00000000UL, 0x3fe00000UL, 0x7f909c4eUL,
4382 0xbf9d4a2cUL, 0xf180bdb1UL, 0xbfec38b2UL, 0x80000000UL, 0x3c76e0b1UL,
4383 0x00000000UL, 0x3fe00000UL, 0x9ae68c87UL, 0x3fac73b3UL, 0x290ea1a3UL,
4384 0xbfea9b66UL, 0xe0000000UL, 0xbc39f630UL, 0x00000000UL, 0x3fe00000UL,
4385 0x94247758UL, 0x3fc133ccUL, 0x6b151741UL, 0xbfe8bc80UL, 0x20000000UL,
4386 0x3c82c5e1UL, 0x00000000UL, 0x3fe00000UL, 0x99fcef32UL, 0x3fca8279UL,
4387 0x667f3bcdUL, 0xbfe6a09eUL, 0x20000000UL, 0x3c8bdd34UL, 0x00000000UL,
4388 0x3fe00000UL, 0x53aba2fdUL, 0xbfcd0dfeUL, 0x25091dd6UL, 0xbfe44cf3UL,
4389 0x20000000UL, 0xbc68076aUL, 0x00000000UL, 0x3ff00000UL, 0x5bc57974UL,
4390 0xbfc59267UL, 0x39ae68c8UL, 0xbfe1c73bUL, 0x20000000UL, 0xbc8b25ddUL,
4391 0x00000000UL, 0x3ff00000UL, 0x73fa1279UL, 0xbfbe3a68UL, 0x3806f63bUL,
4392 0xbfde2b5dUL, 0x20000000UL, 0xbc5e0d89UL, 0x00000000UL, 0x3ff00000UL,
4393 0x866b95cfUL, 0xbfb37ca1UL, 0xa6aea963UL, 0xbfd87de2UL, 0xe0000000UL,
4394 0x3c672cedUL, 0x00000000UL, 0x3ff00000UL, 0x939d225aUL, 0xbfa60beaUL,
4395 0x2ed59f06UL, 0xbfd29406UL, 0xa0000000UL, 0x3c75d28dUL, 0x00000000UL,
4396 0x3ff00000UL, 0x011469fbUL, 0xbf93ad06UL, 0x3c69a60bUL, 0xbfc8f8b8UL,
4397 0xc0000000UL, 0x3c626d19UL, 0x00000000UL, 0x3ff00000UL, 0x176d6d31UL,
4398 0xbf73b92eUL, 0xbc29b42cUL, 0xbfb917a6UL, 0xe0000000UL, 0x3c3e2718UL,
4399 0x00000000UL, 0x3ff00000UL, 0x55555555UL, 0xbfc55555UL, 0x00000000UL,
4400 0xbfe00000UL, 0x11111111UL, 0x3f811111UL, 0x55555555UL, 0x3fa55555UL,
4401 0x1a01a01aUL, 0xbf2a01a0UL, 0x16c16c17UL, 0xbf56c16cUL, 0xa556c734UL,
4402 0x3ec71de3UL, 0x1a01a01aUL, 0x3efa01a0UL, 0x1a600000UL, 0x3d90b461UL,
4403 0x1a600000UL, 0x3d90b461UL, 0x54400000UL, 0x3fb921fbUL, 0x00000000UL,
4404 0x00000000UL, 0x2e037073UL, 0x3b63198aUL, 0x00000000UL, 0x00000000UL,
4405 0x6dc9c883UL, 0x40245f30UL, 0x00000000UL, 0x00000000UL, 0x00000000UL,
4406 0x43380000UL, 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x3ff00000UL,
4407 0x00000000UL, 0x00000000UL, 0x00000000UL, 0x80000000UL, 0x00000000UL,
4408 0x00000000UL, 0x00000000UL, 0x80000000UL, 0x00000000UL, 0x00000000UL,
4409 0x00000000UL, 0x3fe00000UL, 0x00000000UL, 0x3fe00000UL
4410 };
4411
4412 //registers,
4413 // input: (rbp + 8)
4414 // scratch: xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7
4415 // rax, rdx, rcx, rbx (tmp)
4416
4417 // Code generated by Intel C compiler for LIBM library
4418
4419 void MacroAssembler::fast_cos(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp) {
4420 Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
4421 Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
4422 Label L_2TAG_PACKET_8_0_2, L_2TAG_PACKET_9_0_2, L_2TAG_PACKET_10_0_2, L_2TAG_PACKET_11_0_2;
4423 Label L_2TAG_PACKET_12_0_2, L_2TAG_PACKET_13_0_2, B1_3, B1_5, start;
4424
4425 assert_different_registers(tmp, eax, ecx, edx);
4426
4427 address static_const_table_cos = (address)_static_const_table_cos;
4428
4429 bind(start);
4430 subl(rsp, 120);
4431 movl(Address(rsp, 56), tmp);
4432 lea(tmp, ExternalAddress(static_const_table_cos));
4433 movsd(xmm0, Address(rsp, 128));
4434 pextrw(eax, xmm0, 3);
4435 andl(eax, 32767);
4436 subl(eax, 12336);
4437 cmpl(eax, 4293);
4438 jcc(Assembler::above, L_2TAG_PACKET_0_0_2);
4439 movsd(xmm1, Address(tmp, 2160));
4440 mulsd(xmm1, xmm0);
4441 movdqu(xmm5, Address(tmp, 2240));
4442 movsd(xmm4, Address(tmp, 2224));
4443 pand(xmm4, xmm0);
4444 por(xmm5, xmm4);
4445 movsd(xmm3, Address(tmp, 2128));
4446 movdqu(xmm2, Address(tmp, 2112));
4447 addpd(xmm1, xmm5);
4448 cvttsd2sil(edx, xmm1);
4449 cvtsi2sdl(xmm1, edx);
4450 mulsd(xmm3, xmm1);
4451 unpcklpd(xmm1, xmm1);
4452 addl(edx, 1865232);
4453 movdqu(xmm4, xmm0);
4454 andl(edx, 63);
4455 movdqu(xmm5, Address(tmp, 2096));
4456 lea(eax, Address(tmp, 0));
4457 shll(edx, 5);
4458 addl(eax, edx);
4459 mulpd(xmm2, xmm1);
4460 subsd(xmm0, xmm3);
4461 mulsd(xmm1, Address(tmp, 2144));
4462 subsd(xmm4, xmm3);
4463 movsd(xmm7, Address(eax, 8));
4464 unpcklpd(xmm0, xmm0);
4465 movapd(xmm3, xmm4);
4466 subsd(xmm4, xmm2);
4467 mulpd(xmm5, xmm0);
4468 subpd(xmm0, xmm2);
4469 movdqu(xmm6, Address(tmp, 2064));
4470 mulsd(xmm7, xmm4);
4471 subsd(xmm3, xmm4);
4472 mulpd(xmm5, xmm0);
4473 mulpd(xmm0, xmm0);
4474 subsd(xmm3, xmm2);
4475 movdqu(xmm2, Address(eax, 0));
4476 subsd(xmm1, xmm3);
4477 movsd(xmm3, Address(eax, 24));
4478 addsd(xmm2, xmm3);
4479 subsd(xmm7, xmm2);
4480 mulsd(xmm2, xmm4);
4481 mulpd(xmm6, xmm0);
4482 mulsd(xmm3, xmm4);
4483 mulpd(xmm2, xmm0);
4484 mulpd(xmm0, xmm0);
4485 addpd(xmm5, Address(tmp, 2080));
4486 mulsd(xmm4, Address(eax, 0));
4487 addpd(xmm6, Address(tmp, 2048));
4488 mulpd(xmm5, xmm0);
4489 movapd(xmm0, xmm3);
4490 addsd(xmm3, Address(eax, 8));
4491 mulpd(xmm1, xmm7);
4492 movapd(xmm7, xmm4);
4493 addsd(xmm4, xmm3);
4494 addpd(xmm6, xmm5);
4495 movsd(xmm5, Address(eax, 8));
4496 subsd(xmm5, xmm3);
4497 subsd(xmm3, xmm4);
4498 addsd(xmm1, Address(eax, 16));
4499 mulpd(xmm6, xmm2);
4500 addsd(xmm5, xmm0);
4501 addsd(xmm3, xmm7);
4502 addsd(xmm1, xmm5);
4503 addsd(xmm1, xmm3);
4504 addsd(xmm1, xmm6);
4505 unpckhpd(xmm6, xmm6);
4506 addsd(xmm1, xmm6);
4507 addsd(xmm4, xmm1);
4508 movsd(Address(rsp, 0), xmm4);
4509 fld_d(Address(rsp, 0));
4510 jmp(L_2TAG_PACKET_1_0_2);
4511
4512 bind(L_2TAG_PACKET_0_0_2);
4513 jcc(Assembler::greater, L_2TAG_PACKET_2_0_2);
4514 pextrw(eax, xmm0, 3);
4515 andl(eax, 32767);
4516 pinsrw(xmm0, eax, 3);
4517 movsd(xmm1, Address(tmp, 2192));
4518 subsd(xmm1, xmm0);
4519 movsd(Address(rsp, 0), xmm1);
4520 fld_d(Address(rsp, 0));
4521 jmp(L_2TAG_PACKET_1_0_2);
4522
4523 bind(L_2TAG_PACKET_2_0_2);
4524 movl(eax, Address(rsp, 132));
4525 andl(eax, 2146435072);
4526 cmpl(eax, 2146435072);
4527 jcc(Assembler::equal, L_2TAG_PACKET_3_0_2);
4528 subl(rsp, 32);
4529 movsd(Address(rsp, 0), xmm0);
4530 lea(eax, Address(rsp, 40));
4531 movl(Address(rsp, 8), eax);
4532 movl(eax, 1);
4533 movl(Address(rsp, 12), eax);
4534 call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::dlibm_sin_cos_huge())));
4535 addl(rsp, 32);
4536 fld_d(Address(rsp, 8));
4537 jmp(L_2TAG_PACKET_1_0_2);
4538
4539 bind(L_2TAG_PACKET_3_0_2);
4540 fld_d(Address(rsp, 128));
4541 fmul_d(Address(tmp, 2208));
4542
4543 bind(L_2TAG_PACKET_1_0_2);
4544 movl(tmp, Address(rsp, 56));
4545 }
4546
4547 #endif