]> git.ipfire.org Git - thirdparty/glibc.git/blob - sysdeps/x86_64/fpu/multiarch/svml_d_sinh2_core_sse4.S
Update copyright dates with scripts/update-copyrights
[thirdparty/glibc.git] / sysdeps / x86_64 / fpu / multiarch / svml_d_sinh2_core_sse4.S
1 /* Function sinh vectorized with SSE4.
2 Copyright (C) 2021-2022 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
9
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
14
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, see
17 https://www.gnu.org/licenses/. */
18
19 /*
20 * ALGORITHM DESCRIPTION:
21 *
22 * Compute sinh(x) as (exp(x)-exp(-x))/2,
23 * where exp is calculated as
24 * exp(M*ln2 + ln2*(j/2^k) + r) = 2^M * 2^(j/2^k) * exp(r)
25 *
26 * Special cases:
27 *
28 * sinh(NaN) = quiet NaN, and raise invalid exception
29 * sinh(INF) = that INF
30 * sinh(x) = x for subnormals
31 * sinh(x) overflows for big x and returns MAXLOG+log(2)
32 *
33 */
34
35 /* Offsets for data table __svml_dsinh_data_internal
36 */
37 #define _dbInvLn2 0
38 #define _dbLn2hi 16
39 #define _dbLn2lo 32
40 #define _dSign 48
41 #define _dbT 64
42 #define _dbShifter 2112
43 #define _iDomainRange 2128
44 #define _dPC2 2144
45 #define _dPC3 2160
46 #define _dPC4 2176
47 #define _dPC5 2192
48 #define _lIndexMask 2208
49
50 #include <sysdep.h>
51
52 .text
53 .section .text.sse4,"ax",@progbits
54 ENTRY(_ZGVbN2v_sinh_sse4)
55 subq $72, %rsp
56 cfi_def_cfa_offset(80)
57 movaps %xmm0, %xmm2
58
59 /* Abs argument */
60 movups _dSign+__svml_dsinh_data_internal(%rip), %xmm0
61 lea _dbT+8+__svml_dsinh_data_internal(%rip), %rsi
62 andps %xmm2, %xmm0
63 movaps %xmm0, %xmm1
64
65 /*
66 * Load argument
67 * dM = x*2^K/log(2) + RShifter
68 */
69 movups _dbInvLn2+__svml_dsinh_data_internal(%rip), %xmm10
70 pxor %xmm2, %xmm1
71 mulpd %xmm1, %xmm10
72 movups _dbShifter+__svml_dsinh_data_internal(%rip), %xmm5
73 addpd %xmm5, %xmm10
74
75 /*
76 * R
77 * dN = dM - RShifter
78 */
79 movaps %xmm10, %xmm7
80 subpd %xmm5, %xmm7
81
82 /* dR = dX - dN*Log2_hi/2^K */
83 movups _dbLn2hi+__svml_dsinh_data_internal(%rip), %xmm6
84 mulpd %xmm7, %xmm6
85
86 /* dR = (dX - dN*Log2_hi/2^K) - dN*Log2_lo/2^K */
87 movups _dbLn2lo+__svml_dsinh_data_internal(%rip), %xmm8
88 mulpd %xmm7, %xmm8
89
90 /*
91 * Check for overflow\underflow
92 *
93 */
94 pshufd $221, %xmm1, %xmm4
95 subpd %xmm6, %xmm1
96 subpd %xmm8, %xmm1
97
98 /* VLOAD_CONST( D, dPC[0], TAB._dPC1 ); */
99 movq _iDomainRange+__svml_dsinh_data_internal(%rip), %xmm3
100 pcmpgtd %xmm3, %xmm4
101
102 /* dR2 = dR^2 */
103 movaps %xmm1, %xmm3
104 mulpd %xmm1, %xmm3
105 movmskps %xmm4, %edx
106
107 /*
108 * sinh(r) = r*((a1=1)+r^2*(a3+r^2*a5)) = r + r*(r^2*(a3+r^2*a5)) ....
109 * dSinh_r = (a3+r^2*a5)
110 */
111 movups _dPC5+__svml_dsinh_data_internal(%rip), %xmm12
112
113 /*
114 * poly(r) = (dG2+dG1)+dG3*sinh(dR)+dG1*sinh(dR)+(dG1+dG2)*dR2*(a2 +a4*dR2)
115 * dOut = (a2 +a4*dR2)
116 */
117 movups _dPC4+__svml_dsinh_data_internal(%rip), %xmm13
118 mulpd %xmm3, %xmm12
119 mulpd %xmm3, %xmm13
120 addpd _dPC3+__svml_dsinh_data_internal(%rip), %xmm12
121 addpd _dPC2+__svml_dsinh_data_internal(%rip), %xmm13
122
123 /* dSinh_r = r^2*(a3+r^2*a5) */
124 mulpd %xmm3, %xmm12
125
126 /* dOut = dR2*(a2 +a4*dR2) */
127 mulpd %xmm13, %xmm3
128
129 /* dSinh_r = r + r*(r^2*(a3+r^2*a5)) */
130 mulpd %xmm1, %xmm12
131
132 /*
133 * Index and lookup
134 * j
135 */
136 movups _lIndexMask+__svml_dsinh_data_internal(%rip), %xmm9
137 andps %xmm10, %xmm9
138 movd %xmm9, %eax
139
140 /* split j and N */
141 pxor %xmm9, %xmm10
142
143 /*
144 * G1,G2,G3: dTdif,dTn * 2^N,2^(-N)
145 * lM now is an EXP(2^N)
146 */
147 psllq $45, %xmm10
148
149 /* */
150 movaps %xmm10, %xmm4
151 pextrw $4, %xmm9, %ecx
152 addpd %xmm12, %xmm1
153 shll $4, %eax
154 shll $4, %ecx
155 movq (%rax,%rsi), %xmm11
156 movhpd (%rcx,%rsi), %xmm11
157 paddq %xmm11, %xmm4
158
159 /* */
160 psubq %xmm10, %xmm11
161
162 /* dG3 = dTn*2^N + dTn*2^-N */
163 movdqa %xmm4, %xmm14
164 addpd %xmm11, %xmm14
165
166 /* dG2 = dTn*2^N - dTn*2^-N */
167 subpd %xmm11, %xmm4
168 movq -8(%rax,%rsi), %xmm15
169 movhpd -8(%rcx,%rsi), %xmm15
170 paddq %xmm10, %xmm15
171
172 /* dG2 += dG1 */
173 addpd %xmm15, %xmm4
174
175 /* dG1 += dG3 */
176 addpd %xmm14, %xmm15
177
178 /* dOut = dG2*dR2*(a2 +a4*dR2) */
179 mulpd %xmm4, %xmm3
180
181 /* dOut = dG1*sinh(dR)+dG2*dR2*(a2 +a4*dR2) */
182 mulpd %xmm15, %xmm1
183 addpd %xmm1, %xmm3
184
185 /* dOut = dG2 + dG1*sinh(dR)+dG2*dR2*(a2 +a4*dR2) */
186 addpd %xmm3, %xmm4
187
188 /* Ret H */
189 orps %xmm4, %xmm0
190 andl $3, %edx
191
192 /* Go to special inputs processing branch */
193 jne L(SPECIAL_VALUES_BRANCH)
194 # LOE rbx rbp r12 r13 r14 r15 edx xmm0 xmm2
195
196 /* Restore registers
197 * and exit the function
198 */
199
200 L(EXIT):
201 addq $72, %rsp
202 cfi_def_cfa_offset(8)
203 ret
204 cfi_def_cfa_offset(80)
205
206 /* Branch to process
207 * special inputs
208 */
209
210 L(SPECIAL_VALUES_BRANCH):
211 movups %xmm2, 32(%rsp)
212 movups %xmm0, 48(%rsp)
213 # LOE rbx rbp r12 r13 r14 r15 edx xmm0
214
215 xorl %eax, %eax
216 movq %r12, 16(%rsp)
217 cfi_offset(12, -64)
218 movl %eax, %r12d
219 movq %r13, 8(%rsp)
220 cfi_offset(13, -72)
221 movl %edx, %r13d
222 movq %r14, (%rsp)
223 cfi_offset(14, -80)
224 # LOE rbx rbp r15 r12d r13d
225
226 /* Range mask
227 * bits check
228 */
229
230 L(RANGEMASK_CHECK):
231 btl %r12d, %r13d
232
233 /* Call scalar math function */
234 jc L(SCALAR_MATH_CALL)
235 # LOE rbx rbp r15 r12d r13d
236
237 /* Special inputs
238 * processing loop
239 */
240
241 L(SPECIAL_VALUES_LOOP):
242 incl %r12d
243 cmpl $2, %r12d
244
245 /* Check bits in range mask */
246 jl L(RANGEMASK_CHECK)
247 # LOE rbx rbp r15 r12d r13d
248
249 movq 16(%rsp), %r12
250 cfi_restore(12)
251 movq 8(%rsp), %r13
252 cfi_restore(13)
253 movq (%rsp), %r14
254 cfi_restore(14)
255 movups 48(%rsp), %xmm0
256
257 /* Go to exit */
258 jmp L(EXIT)
259 cfi_offset(12, -64)
260 cfi_offset(13, -72)
261 cfi_offset(14, -80)
262 # LOE rbx rbp r12 r13 r14 r15 xmm0
263
264 /* Scalar math fucntion call
265 * to process special input
266 */
267
268 L(SCALAR_MATH_CALL):
269 movl %r12d, %r14d
270 movsd 32(%rsp,%r14,8), %xmm0
271 call sinh@PLT
272 # LOE rbx rbp r14 r15 r12d r13d xmm0
273
274 movsd %xmm0, 48(%rsp,%r14,8)
275
276 /* Process special inputs in loop */
277 jmp L(SPECIAL_VALUES_LOOP)
278 # LOE rbx rbp r15 r12d r13d
279 END(_ZGVbN2v_sinh_sse4)
280
281 .section .rodata, "a"
282 .align 16
283
284 #ifdef __svml_dsinh_data_internal_typedef
285 typedef unsigned int VUINT32;
286 typedef struct
287 {
288 __declspec(align(16)) VUINT32 _dbInvLn2[2][2];
289 __declspec(align(16)) VUINT32 _dbLn2hi[2][2];
290 __declspec(align(16)) VUINT32 _dbLn2lo[2][2];
291 __declspec(align(16)) VUINT32 _dSign[2][2]; //0x8000000000000000
292 __declspec(align(16)) VUINT32 _dbT[(1<<7)][2][2]; //precalc poly coeff
293 __declspec(align(16)) VUINT32 _dbShifter[2][2];
294 __declspec(align(16)) VUINT32 _iDomainRange[4][1];
295 __declspec(align(16)) VUINT32 _dPC2[2][2];
296 __declspec(align(16)) VUINT32 _dPC3[2][2];
297 __declspec(align(16)) VUINT32 _dPC4[2][2];
298 __declspec(align(16)) VUINT32 _dPC5[2][2];
299 __declspec(align(16)) VUINT32 _lIndexMask[2][2];
300 } __svml_dsinh_data_internal;
301 #endif
302 __svml_dsinh_data_internal:
303 .quad 0x3FF71547652B82FE, 0x3FF71547652B82FE /* _dbInvLn2 = 1/log(2) */
304 .align 16
305 .quad 0x3FE62E42FEFA0000, 0x3FE62E42FEFA0000 /* _dbLn2hi = log(2) hi*/
306 .align 16
307 .quad 0x3D7CF79ABC9E3B3A, 0x3D7CF79ABC9E3B3A /* _dbLn2lo = log(2) lo*/
308 .align 16
309 .quad 0x8000000000000000, 0x8000000000000000 /* _dSign */
310 //_dbT
311 .align 16
312 .quad 0x0000000000000000, 0x3FE0000000000000 //2^( 0 /128-1) - 2^(- 0 /128-1), 2^(- 0 /128-1)
313 .quad 0x3F762E4A19BD1E74, 0x3FDFD3C22B8F71F1 //2^( 1 /128-1) - 2^(- 1 /128-1), 2^(- 1 /128-1)
314 .quad 0x3F862E5F6A0DFD36, 0x3FDFA7C1819E90D8 //2^( 2 /128-1) - 2^(- 2 /128-1), 2^(- 2 /128-1)
315 .quad 0x3F90A2E234040F5F, 0x3FDF7BFDAD9CBE14 //2^( 3 /128-1) - 2^(- 3 /128-1), 2^(- 3 /128-1)
316 .quad 0x3F962EB4ABCC5A81, 0x3FDF50765B6E4540 //2^( 4 /128-1) - 2^(- 4 /128-1), 2^(- 4 /128-1)
317 .quad 0x3F9BBAB1C5033244, 0x3FDF252B376BBA97 //2^( 5 /128-1) - 2^(- 5 /128-1), 2^(- 5 /128-1)
318 .quad 0x3FA0A372144EEB45, 0x3FDEFA1BEE615A27 //2^( 6 /128-1) - 2^(- 6 /128-1), 2^(- 6 /128-1)
319 .quad 0x3FA369AB3FFBF8B0, 0x3FDECF482D8E67F1 //2^( 7 /128-1) - 2^(- 7 /128-1), 2^(- 7 /128-1)
320 .quad 0x3FA63009BA740A2A, 0x3FDEA4AFA2A490DA //2^( 8 /128-1) - 2^(- 8 /128-1), 2^(- 8 /128-1)
321 .quad 0x3FA8F692D8EA1B5A, 0x3FDE7A51FBC74C83 //2^( 9 /128-1) - 2^(- 9 /128-1), 2^(- 9 /128-1)
322 .quad 0x3FABBD4BF0E31A6F, 0x3FDE502EE78B3FF6 //2^( 10 /128-1) - 2^(- 10 /128-1), 2^(- 10 /128-1)
323 .quad 0x3FAE843A5840286A, 0x3FDE264614F5A129 //2^( 11 /128-1) - 2^(- 11 /128-1), 2^(- 11 /128-1)
324 .quad 0x3FB0A5B1B2A46D0A, 0x3FDDFC97337B9B5F //2^( 12 /128-1) - 2^(- 12 /128-1), 2^(- 12 /128-1)
325 .quad 0x3FB20966375ABCDF, 0x3FDDD321F301B460 //2^( 13 /128-1) - 2^(- 13 /128-1), 2^(- 13 /128-1)
326 .quad 0x3FB36D3D65DCA4E8, 0x3FDDA9E603DB3285 //2^( 14 /128-1) - 2^(- 14 /128-1), 2^(- 14 /128-1)
327 .quad 0x3FB4D139EA06642A, 0x3FDD80E316C98398 //2^( 15 /128-1) - 2^(- 15 /128-1), 2^(- 15 /128-1)
328 .quad 0x3FB6355E6FFBF9BA, 0x3FDD5818DCFBA487 //2^( 16 /128-1) - 2^(- 16 /128-1), 2^(- 16 /128-1)
329 .quad 0x3FB799ADA42E4788, 0x3FDD2F87080D89F2 //2^( 17 /128-1) - 2^(- 17 /128-1), 2^(- 17 /128-1)
330 .quad 0x3FB8FE2A336035BC, 0x3FDD072D4A07897C //2^( 18 /128-1) - 2^(- 18 /128-1), 2^(- 18 /128-1)
331 .quad 0x3FBA62D6CAABD6B6, 0x3FDCDF0B555DC3FA //2^( 19 /128-1) - 2^(- 19 /128-1), 2^(- 19 /128-1)
332 .quad 0x3FBBC7B617878BAF, 0x3FDCB720DCEF9069 //2^( 20 /128-1) - 2^(- 20 /128-1), 2^(- 20 /128-1)
333 .quad 0x3FBD2CCAC7CB2A11, 0x3FDC8F6D9406E7B5 //2^( 21 /128-1) - 2^(- 21 /128-1), 2^(- 21 /128-1)
334 .quad 0x3FBE921789B52185, 0x3FDC67F12E57D14B //2^( 22 /128-1) - 2^(- 22 /128-1), 2^(- 22 /128-1)
335 .quad 0x3FBFF79F0BEFA2C7, 0x3FDC40AB5FFFD07A //2^( 23 /128-1) - 2^(- 23 /128-1), 2^(- 23 /128-1)
336 .quad 0x3FC0AEB1FECAE3A9, 0x3FDC199BDD85529C //2^( 24 /128-1) - 2^(- 24 /128-1), 2^(- 24 /128-1)
337 .quad 0x3FC161B4871C5CEC, 0x3FDBF2C25BD71E09 //2^( 25 /128-1) - 2^(- 25 /128-1), 2^(- 25 /128-1)
338 .quad 0x3FC214D876F26FD0, 0x3FDBCC1E904BC1D2 //2^( 26 /128-1) - 2^(- 26 /128-1), 2^(- 26 /128-1)
339 .quad 0x3FC2C81F2693816F, 0x3FDBA5B030A1064A //2^( 27 /128-1) - 2^(- 27 /128-1), 2^(- 27 /128-1)
340 .quad 0x3FC37B89EE88BEF7, 0x3FDB7F76F2FB5E47 //2^( 28 /128-1) - 2^(- 28 /128-1), 2^(- 28 /128-1)
341 .quad 0x3FC42F1A27A0B3CD, 0x3FDB59728DE5593A //2^( 29 /128-1) - 2^(- 29 /128-1), 2^(- 29 /128-1)
342 .quad 0x3FC4E2D12AF1E037, 0x3FDB33A2B84F15FB //2^( 30 /128-1) - 2^(- 30 /128-1), 2^(- 30 /128-1)
343 .quad 0x3FC596B051DD508D, 0x3FDB0E07298DB666 //2^( 31 /128-1) - 2^(- 31 /128-1), 2^(- 31 /128-1)
344 .quad 0x3FC64AB8F61134FA, 0x3FDAE89F995AD3AD //2^( 32 /128-1) - 2^(- 32 /128-1), 2^(- 32 /128-1)
345 .quad 0x3FC6FEEC718B79D1, 0x3FDAC36BBFD3F37A //2^( 33 /128-1) - 2^(- 33 /128-1), 2^(- 33 /128-1)
346 .quad 0x3FC7B34C1E9C607F, 0x3FDA9E6B5579FDBF //2^( 34 /128-1) - 2^(- 34 /128-1), 2^(- 34 /128-1)
347 .quad 0x3FC867D957E91912, 0x3FDA799E1330B358 //2^( 35 /128-1) - 2^(- 35 /128-1), 2^(- 35 /128-1)
348 .quad 0x3FC91C95786E5C72, 0x3FDA5503B23E255D //2^( 36 /128-1) - 2^(- 36 /128-1), 2^(- 36 /128-1)
349 .quad 0x3FC9D181DB83072F, 0x3FDA309BEC4A2D33 //2^( 37 /128-1) - 2^(- 37 /128-1), 2^(- 37 /128-1)
350 .quad 0x3FCA869FDCDAB512, 0x3FDA0C667B5DE565 //2^( 38 /128-1) - 2^(- 38 /128-1), 2^(- 38 /128-1)
351 .quad 0x3FCB3BF0D8885D4C, 0x3FD9E86319E32323 //2^( 39 /128-1) - 2^(- 39 /128-1), 2^(- 39 /128-1)
352 .quad 0x3FCBF1762B00EF69, 0x3FD9C49182A3F090 //2^( 40 /128-1) - 2^(- 40 /128-1), 2^(- 40 /128-1)
353 .quad 0x3FCCA731311DF0FB, 0x3FD9A0F170CA07BA //2^( 41 /128-1) - 2^(- 41 /128-1), 2^(- 41 /128-1)
354 .quad 0x3FCD5D2348201C09, 0x3FD97D829FDE4E50 //2^( 42 /128-1) - 2^(- 42 /128-1), 2^(- 42 /128-1)
355 .quad 0x3FCE134DCDB1FE3E, 0x3FD95A44CBC8520F //2^( 43 /128-1) - 2^(- 43 /128-1), 2^(- 43 /128-1)
356 .quad 0x3FCEC9B21FEA98EA, 0x3FD93737B0CDC5E5 //2^( 44 /128-1) - 2^(- 44 /128-1), 2^(- 44 /128-1)
357 .quad 0x3FCF80519D5001D3, 0x3FD9145B0B91FFC6 //2^( 45 /128-1) - 2^(- 45 /128-1), 2^(- 45 /128-1)
358 .quad 0x3FD01B96D26D026A, 0x3FD8F1AE99157736 //2^( 46 /128-1) - 2^(- 46 /128-1), 2^(- 46 /128-1)
359 .quad 0x3FD07723CAFA6331, 0x3FD8CF3216B5448C //2^( 47 /128-1) - 2^(- 47 /128-1), 2^(- 47 /128-1)
360 .quad 0x3FD0D2D06841B373, 0x3FD8ACE5422AA0DB //2^( 48 /128-1) - 2^(- 48 /128-1), 2^(- 48 /128-1)
361 .quad 0x3FD12E9D5A715381, 0x3FD88AC7D98A6699 //2^( 49 /128-1) - 2^(- 49 /128-1), 2^(- 49 /128-1)
362 .quad 0x3FD18A8B51F5C661, 0x3FD868D99B4492ED //2^( 50 /128-1) - 2^(- 50 /128-1), 2^(- 50 /128-1)
363 .quad 0x3FD1E69AFF7B04D7, 0x3FD8471A4623C7AD //2^( 51 /128-1) - 2^(- 51 /128-1), 2^(- 51 /128-1)
364 .quad 0x3FD242CD13EDD0F1, 0x3FD82589994CCE13 //2^( 52 /128-1) - 2^(- 52 /128-1), 2^(- 52 /128-1)
365 .quad 0x3FD29F22407D0A0C, 0x3FD80427543E1A12 //2^( 53 /128-1) - 2^(- 53 /128-1), 2^(- 53 /128-1)
366 .quad 0x3FD2FB9B369B0153, 0x3FD7E2F336CF4E62 //2^( 54 /128-1) - 2^(- 54 /128-1), 2^(- 54 /128-1)
367 .quad 0x3FD35838A7FECEC8, 0x3FD7C1ED0130C132 //2^( 55 /128-1) - 2^(- 55 /128-1), 2^(- 55 /128-1)
368 .quad 0x3FD3B4FB46A5A6CC, 0x3FD7A11473EB0187 //2^( 56 /128-1) - 2^(- 56 /128-1), 2^(- 56 /128-1)
369 .quad 0x3FD411E3C4D4302F, 0x3FD780694FDE5D3F //2^( 57 /128-1) - 2^(- 57 /128-1), 2^(- 57 /128-1)
370 .quad 0x3FD46EF2D517DAC8, 0x3FD75FEB564267C9 //2^( 58 /128-1) - 2^(- 58 /128-1), 2^(- 58 /128-1)
371 .quad 0x3FD4CC292A48369E, 0x3FD73F9A48A58174 //2^( 59 /128-1) - 2^(- 59 /128-1), 2^(- 59 /128-1)
372 .quad 0x3FD5298777884B96, 0x3FD71F75E8EC5F74 //2^( 60 /128-1) - 2^(- 60 /128-1), 2^(- 60 /128-1)
373 .quad 0x3FD5870E7047F1BC, 0x3FD6FF7DF9519484 //2^( 61 /128-1) - 2^(- 61 /128-1), 2^(- 61 /128-1)
374 .quad 0x3FD5E4BEC8452A1A, 0x3FD6DFB23C651A2F //2^( 62 /128-1) - 2^(- 62 /128-1), 2^(- 62 /128-1)
375 .quad 0x3FD64299338D7827, 0x3FD6C012750BDABF //2^( 63 /128-1) - 2^(- 63 /128-1), 2^(- 63 /128-1)
376 .quad 0x3FD6A09E667F3BCD, 0x3FD6A09E667F3BCD //2^( 64 /128-1) - 2^(- 64 /128-1), 2^(- 64 /128-1)
377 .quad 0x3FD6FECF15CB0C0B, 0x3FD68155D44CA973 //2^( 65 /128-1) - 2^(- 65 /128-1), 2^(- 65 /128-1)
378 .quad 0x3FD75D2BF6751239, 0x3FD6623882552225 //2^( 66 /128-1) - 2^(- 66 /128-1), 2^(- 66 /128-1)
379 .quad 0x3FD7BBB5BDD665E8, 0x3FD6434634CCC320 //2^( 67 /128-1) - 2^(- 67 /128-1), 2^(- 67 /128-1)
380 .quad 0x3FD81A6D219E6963, 0x3FD6247EB03A5585 //2^( 68 /128-1) - 2^(- 68 /128-1), 2^(- 68 /128-1)
381 .quad 0x3FD87952D7D426DF, 0x3FD605E1B976DC09 //2^( 69 /128-1) - 2^(- 69 /128-1), 2^(- 69 /128-1)
382 .quad 0x3FD8D86796D7AE49, 0x3FD5E76F15AD2148 //2^( 70 /128-1) - 2^(- 70 /128-1), 2^(- 70 /128-1)
383 .quad 0x3FD937AC156373C8, 0x3FD5C9268A5946B7 //2^( 71 /128-1) - 2^(- 71 /128-1), 2^(- 71 /128-1)
384 .quad 0x3FD997210A8DAEE4, 0x3FD5AB07DD485429 //2^( 72 /128-1) - 2^(- 72 /128-1), 2^(- 72 /128-1)
385 .quad 0x3FD9F6C72DC9BA68, 0x3FD58D12D497C7FD //2^( 73 /128-1) - 2^(- 73 /128-1), 2^(- 73 /128-1)
386 .quad 0x3FDA569F36E974EA, 0x3FD56F4736B527DA //2^( 74 /128-1) - 2^(- 74 /128-1), 2^(- 74 /128-1)
387 .quad 0x3FDAB6A9DE1EA215, 0x3FD551A4CA5D920F //2^( 75 /128-1) - 2^(- 75 /128-1), 2^(- 75 /128-1)
388 .quad 0x3FDB16E7DBFC4CA3, 0x3FD5342B569D4F82 //2^( 76 /128-1) - 2^(- 76 /128-1), 2^(- 76 /128-1)
389 .quad 0x3FDB7759E9782918, 0x3FD516DAA2CF6642 //2^( 77 /128-1) - 2^(- 77 /128-1), 2^(- 77 /128-1)
390 .quad 0x3FDBD800BFEBF932, 0x3FD4F9B2769D2CA7 //2^( 78 /128-1) - 2^(- 78 /128-1), 2^(- 78 /128-1)
391 .quad 0x3FDC38DD1916F025, 0x3FD4DCB299FDDD0D //2^( 79 /128-1) - 2^(- 79 /128-1), 2^(- 79 /128-1)
392 .quad 0x3FDC99EFAF1F1790, 0x3FD4BFDAD5362A27 //2^( 80 /128-1) - 2^(- 80 /128-1), 2^(- 80 /128-1)
393 .quad 0x3FDCFB393C92B539, 0x3FD4A32AF0D7D3DE //2^( 81 /128-1) - 2^(- 81 /128-1), 2^(- 81 /128-1)
394 .quad 0x3FDD5CBA7C69B19C, 0x3FD486A2B5C13CD0 //2^( 82 /128-1) - 2^(- 82 /128-1), 2^(- 82 /128-1)
395 .quad 0x3FDDBE742A06FF34, 0x3FD46A41ED1D0057 //2^( 83 /128-1) - 2^(- 83 /128-1), 2^(- 83 /128-1)
396 .quad 0x3FDE2067013A029D, 0x3FD44E086061892D //2^( 84 /128-1) - 2^(- 84 /128-1), 2^(- 84 /128-1)
397 .quad 0x3FDE8293BE3FFB87, 0x3FD431F5D950A897 //2^( 85 /128-1) - 2^(- 85 /128-1), 2^(- 85 /128-1)
398 .quad 0x3FDEE4FB1DC56E75, 0x3FD4160A21F72E2A //2^( 86 /128-1) - 2^(- 86 /128-1), 2^(- 86 /128-1)
399 .quad 0x3FDF479DDCE78F58, 0x3FD3FA4504AC801C //2^( 87 /128-1) - 2^(- 87 /128-1), 2^(- 87 /128-1)
400 .quad 0x3FDFAA7CB935ACFE, 0x3FD3DEA64C123422 //2^( 88 /128-1) - 2^(- 88 /128-1), 2^(- 88 /128-1)
401 .quad 0x3FE006CC38594EB1, 0x3FD3C32DC313A8E5 //2^( 89 /128-1) - 2^(- 89 /128-1), 2^(- 89 /128-1)
402 .quad 0x3FE03878E0EB1569, 0x3FD3A7DB34E59FF7 //2^( 90 /128-1) - 2^(- 90 /128-1), 2^(- 90 /128-1)
403 .quad 0x3FE06A44B5C74101, 0x3FD38CAE6D05D866 //2^( 91 /128-1) - 2^(- 91 /128-1), 2^(- 91 /128-1)
404 .quad 0x3FE09C3016A0D077, 0x3FD371A7373AA9CB //2^( 92 /128-1) - 2^(- 92 /128-1), 2^(- 92 /128-1)
405 .quad 0x3FE0CE3B63676360, 0x3FD356C55F929FF1 //2^( 93 /128-1) - 2^(- 93 /128-1), 2^(- 93 /128-1)
406 .quad 0x3FE10066FC47F240, 0x3FD33C08B26416FF //2^( 94 /128-1) - 2^(- 94 /128-1), 2^(- 94 /128-1)
407 .quad 0x3FE132B341AD8761, 0x3FD32170FC4CD831 //2^( 95 /128-1) - 2^(- 95 /128-1), 2^(- 95 /128-1)
408 .quad 0x3FE165209441F823, 0x3FD306FE0A31B715 //2^( 96 /128-1) - 2^(- 96 /128-1), 2^(- 96 /128-1)
409 .quad 0x3FE197AF54EE9EBB, 0x3FD2ECAFA93E2F56 //2^( 97 /128-1) - 2^(- 97 /128-1), 2^(- 97 /128-1)
410 .quad 0x3FE1CA5FE4DD1475, 0x3FD2D285A6E4030B //2^( 98 /128-1) - 2^(- 98 /128-1), 2^(- 98 /128-1)
411 .quad 0x3FE1FD32A577EC72, 0x3FD2B87FD0DAD990 //2^( 99 /128-1) - 2^(- 99 /128-1), 2^(- 99 /128-1)
412 .quad 0x3FE23027F86B6ED6, 0x3FD29E9DF51FDEE1 //2^( 100 /128-1) - 2^(- 100 /128-1), 2^(- 100 /128-1)
413 .quad 0x3FE263403FA65489, 0x3FD284DFE1F56381 //2^( 101 /128-1) - 2^(- 101 /128-1), 2^(- 101 /128-1)
414 .quad 0x3FE2967BDD5A8364, 0x3FD26B4565E27CDD //2^( 102 /128-1) - 2^(- 102 /128-1), 2^(- 102 /128-1)
415 .quad 0x3FE2C9DB33FDCAE9, 0x3FD251CE4FB2A63F //2^( 103 /128-1) - 2^(- 103 /128-1), 2^(- 103 /128-1)
416 .quad 0x3FE2FD5EA64AA180, 0x3FD2387A6E756238 //2^( 104 /128-1) - 2^(- 104 /128-1), 2^(- 104 /128-1)
417 .quad 0x3FE331069740E22F, 0x3FD21F49917DDC96 //2^( 105 /128-1) - 2^(- 105 /128-1), 2^(- 105 /128-1)
418 .quad 0x3FE364D36A268AE0, 0x3FD2063B88628CD6 //2^( 106 /128-1) - 2^(- 106 /128-1), 2^(- 106 /128-1)
419 .quad 0x3FE398C582887B27, 0x3FD1ED5022FCD91D //2^( 107 /128-1) - 2^(- 107 /128-1), 2^(- 107 /128-1)
420 .quad 0x3FE3CCDD443B3394, 0x3FD1D4873168B9AA //2^( 108 /128-1) - 2^(- 108 /128-1), 2^(- 108 /128-1)
421 .quad 0x3FE4011B135B9590, 0x3FD1BBE084045CD4 //2^( 109 /128-1) - 2^(- 109 /128-1), 2^(- 109 /128-1)
422 .quad 0x3FE4357F544FA3C1, 0x3FD1A35BEB6FCB75 //2^( 110 /128-1) - 2^(- 110 /128-1), 2^(- 110 /128-1)
423 .quad 0x3FE46A0A6BC742FD, 0x3FD18AF9388C8DEA //2^( 111 /128-1) - 2^(- 111 /128-1), 2^(- 111 /128-1)
424 .quad 0x3FE49EBCBEBCFBCA, 0x3FD172B83C7D517B //2^( 112 /128-1) - 2^(- 112 /128-1), 2^(- 112 /128-1)
425 .quad 0x3FE4D396B276BC6F, 0x3FD15A98C8A58E51 //2^( 113 /128-1) - 2^(- 113 /128-1), 2^(- 113 /128-1)
426 .quad 0x3FE50898AC869B96, 0x3FD1429AAEA92DE0 //2^( 114 /128-1) - 2^(- 114 /128-1), 2^(- 114 /128-1)
427 .quad 0x3FE53DC312CB9B7A, 0x3FD12ABDC06C31CC //2^( 115 /128-1) - 2^(- 115 /128-1), 2^(- 115 /128-1)
428 .quad 0x3FE573164B726DB6, 0x3FD11301D0125B51 //2^( 116 /128-1) - 2^(- 116 /128-1), 2^(- 116 /128-1)
429 .quad 0x3FE5A892BCF6379B, 0x3FD0FB66AFFED31B //2^( 117 /128-1) - 2^(- 117 /128-1), 2^(- 117 /128-1)
430 .quad 0x3FE5DE38CE215725, 0x3FD0E3EC32D3D1A2 //2^( 118 /128-1) - 2^(- 118 /128-1), 2^(- 118 /128-1)
431 .quad 0x3FE61408E60E2888, 0x3FD0CC922B7247F7 //2^( 119 /128-1) - 2^(- 119 /128-1), 2^(- 119 /128-1)
432 .quad 0x3FE64A036C27CC52, 0x3FD0B5586CF9890F //2^( 120 /128-1) - 2^(- 120 /128-1), 2^(- 120 /128-1)
433 .quad 0x3FE68028C82AEE2F, 0x3FD09E3ECAC6F383 //2^( 121 /128-1) - 2^(- 121 /128-1), 2^(- 121 /128-1)
434 .quad 0x3FE6B67962268C43, 0x3FD0874518759BC8 //2^( 122 /128-1) - 2^(- 122 /128-1), 2^(- 122 /128-1)
435 .quad 0x3FE6ECF5A27CBF28, 0x3FD0706B29DDF6DE //2^( 123 /128-1) - 2^(- 123 /128-1), 2^(- 123 /128-1)
436 .quad 0x3FE7239DF1E38286, 0x3FD059B0D3158574 //2^( 124 /128-1) - 2^(- 124 /128-1), 2^(- 124 /128-1)
437 .quad 0x3FE75A72B9657E51, 0x3FD04315E86E7F85 //2^( 125 /128-1) - 2^(- 125 /128-1), 2^(- 125 /128-1)
438 .quad 0x3FE791746262D0A8, 0x3FD02C9A3E778061 //2^( 126 /128-1) - 2^(- 126 /128-1), 2^(- 126 /128-1)
439 .quad 0x3FE7C8A35691D856, 0x3FD0163DA9FB3335 //2^( 127 /128-1) - 2^(- 127 /128-1), 2^(- 127 /128-1)
440 .align 16
441 .quad 0x42C8000000000000, 0x42C8000000000000 /* _dbShifter = 1.5 * 2^(52-k)*/
442 .align 16
443 .long 0x40861d99, 0x40861d99, 0x40861d99, 0x40861d99 /* _iDomainRange 0x40861d9ac12a3e85 =(1021*2^K-0.5)*log(2)/2^K -needed for quick exp*/
444 .align 16
445 .quad 0x3FDFFFFFFFFFFDBD, 0x3FDFFFFFFFFFFDBD /* _dPC2 */
446 .align 16
447 .quad 0x3FC55555555554AD, 0x3FC55555555554AD /* _dPC3 */
448 .align 16
449 .quad 0x3FA55555CF16D299, 0x3FA55555CF16D299 /* _dPC4 */
450 .align 16
451 .quad 0x3F8111115712F425, 0x3F8111115712F425 /* _dPC5 */
452 .align 16
453 .quad 0x000000000000007f, 0x000000000000007f /* _lIndexMask */
454 .align 16
455 .type __svml_dsinh_data_internal,@object
456 .size __svml_dsinh_data_internal,.-__svml_dsinh_data_internal