| blob | 4da68c7f0832b8e5c8e837daeba359d0784c4504 |
1 .file "atanf.s"
4 // Copyright (c) 2000 - 2003, Intel Corporation
5 // All rights reserved.
6 //
7 // Contributed 2000 by the Intel Numerics Group, Intel Corporation
8 //
9 // Redistribution and use in source and binary forms, with or without
10 // modification, are permitted provided that the following conditions are
11 // met:
12 //
13 // * Redistributions of source code must retain the above copyright
14 // notice, this list of conditions and the following disclaimer.
15 //
16 // * Redistributions in binary form must reproduce the above copyright
17 // notice, this list of conditions and the following disclaimer in the
18 // documentation and/or other materials provided with the distribution.
19 //
20 // * The name of Intel Corporation may not be used to endorse or promote
21 // products derived from this software without specific prior written
22 // permission.
24 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
27 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
28 // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
29 // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
30 // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
31 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
32 // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
33 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
34 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35 //
36 // Intel Corporation is the author of this code, and requests that all
37 // problem reports or change requests be submitted to it directly at
38 // http://www.intel.com/software/products/opensource/libraries/num.htm.
41 // History
42 //==============================================================
43 // 02/20/00 Initial version
44 // 08/17/00 Changed predicate register macro-usage to direct predicate
45 // names due to an assembler bug.
46 // 02/06/02 Corrected .section statement
47 // 05/20/02 Cleaned up namespace and sf0 syntax
48 // 02/06/03 Reordered header: .section, .global, .proc, .align;
49 // added missing bundling
51 //
52 // Assembly macros
53 //==============================================================
55 // integer registers used
56 EXP_Addr1 = r33
57 EXP_Addr2 = r34
59 // floating point registers used
60 atanf_coeff_R4 = f32
61 atanf_coeff_R5 = f33
62 atanf_coeff_R1 = f34
63 atanf_coeff_R2 = f35
65 atanf_coeff_R3 = f36
66 atanf_coeff_P1 = f37
67 atanf_coeff_Q6 = f38
68 atanf_coeff_Q7 = f39
69 atanf_coeff_Q8 = f40
71 atanf_coeff_Q9 = f41
72 atanf_coeff_Q4 = f42
73 atanf_coeff_Q5 = f43
74 atanf_coeff_Q2 = f44
75 atanf_coeff_Q3 = f45
77 atanf_coeff_P5 = f46
78 atanf_coeff_P6 = f47
79 atanf_coeff_Q0 = f48
80 atanf_coeff_Q1 = f49
81 atanf_coeff_P7 = f50
83 atanf_coeff_P8 = f51
84 atanf_coeff_P3 = f52
85 atanf_coeff_P4 = f53
86 atanf_coeff_P9 = f54
87 atanf_coeff_P10 = f55
89 atanf_coeff_P2 = f56
90 atanf_piby2 = f57
91 atanf_z = f58
92 atanf_b = f59
93 atanf_zsq = f60
95 atanf_sgn_x = f61
96 atanf_sgnx_piby2 = f62
97 atanf_abs_x = f63
98 atanf_t = f64
99 atanf_xcub = f65
101 atanf_tsq = f66
102 atanf_t4 = f67
103 atanf_x5 = f68
104 atanf_x6 = f69
105 atanf_x11 = f70
107 atanf_poly_p1 = f71
108 atanf_poly_p2 = f72
109 atanf_poly_p3 = f73
110 atanf_poly_p4 = f74
111 atanf_poly_p5 = f75
113 atanf_poly_q1 = f76
114 atanf_poly_q2 = f77
115 atanf_poly_q3 = f78
116 atanf_poly_q4 = f79
117 atanf_poly_q5 = f80
119 atanf_poly_q = f81
120 atanf_poly_r1 = f81
121 atanf_poly_r2 = f82
122 atanf_poly_r3 = f83
123 atanf_bsq = f84
124 atanf_z4 = f85
126 atanf_z5 = f86
127 atanf_z8 = f87
128 atanf_z13 = f88
129 atanf_poly_r2 = f89
130 atanf_poly_r1 = f90
132 atanf_z8_bsq = f91
133 atanf_poly_r = f92
134 atanf_z21_poly_r = f93
135 atanf_answer = f8
138 // predicate registers used
139 //atanf_pred_LE1 = p6
140 //atanf_pred_GT1 = p7
143 RODATA
145 .align 16
147 LOCAL_OBJECT_START(atanf_coeff_1_table)
148 data8 0x40c4c241be751ff2 // r4
149 data8 0x40e9f300c2f3070b // r5
150 data8 0x409babffef772075 // r3
151 data8 0xbfd5555512191621 // p1
152 data8 0x3fc9997e7afbff4e // p2 = q8
153 data8 0xbfd5555512191621 // p1 = q9
154 data8 0x3f97105b4160f86b // p8 = q2
155 data8 0xbfa6e10ba401393f // p7 = q3
156 data8 0x3f522e5d33bc9baa // p10 = q0
157 data8 0xbf7deaadaa336451 // p9 = q1
158 data8 0xbfc2473c5145ee38 // p3
159 data8 0x3fbc4f512b1865f5 // p4
160 data8 0x3fc9997e7afbff4e // p2
161 data8 0x3ff921fb54442d18 // pi/2
162 LOCAL_OBJECT_END(atanf_coeff_1_table)
166 LOCAL_OBJECT_START(atanf_coeff_2_table)
167 data8 0x4035000000004284 // r1
168 data8 0x406cdffff336a59b // r2
169 data8 0x3fbc4f512b1865f5 // p4 = q6
170 data8 0xbfc2473c5145ee38 // p3 = q7
171 data8 0x3fb142a73d7c54e3 // p6 = q4
172 data8 0xbfb68eed6a8cfa32 // p5 = q5
173 data8 0xbfb68eed6a8cfa32 // p5
174 data8 0x3fb142a73d7c54e3 // p6
175 data8 0xbfa6e10ba401393f // p7
176 data8 0x3f97105b4160f86b // p8
177 data8 0xbf7deaadaa336451 // p9
178 data8 0x3f522e5d33bc9baa // p10
179 LOCAL_OBJECT_END(atanf_coeff_2_table)
183 .section .text
184 GLOBAL_LIBM_ENTRY(atanf)
186 { .mfi
187 alloc r32 = ar.pfs,1,2,0,0
188 frcpa.s1 atanf_z,p0 = f1,f8
189 addl EXP_Addr2 = @ltoff(atanf_coeff_2_table),gp
190 }
191 { .mfi
192 addl EXP_Addr1 = @ltoff(atanf_coeff_1_table),gp
193 fma.s1 atanf_t = f8,f8,f0
194 nop.i 999;;
195 }
198 { .mfi
199 nop.m 999
200 fmerge.s atanf_sgn_x = f8,f1
201 nop.i 999;;
202 }
204 { .mfi
205 ld8 EXP_Addr1 = [EXP_Addr1]
206 fmerge.s atanf_abs_x = f1,f8
207 nop.i 999
208 }
209 { .mfi
210 ld8 EXP_Addr2 = [EXP_Addr2]
211 nop.f 999
212 nop.i 999;;
213 }
216 { .mfi
217 nop.m 999
218 fclass.m p8,p0 = f8,0x7 // @zero
219 nop.i 999;;
220 }
222 { .mfi
223 nop.m 999
224 fcmp.eq.unc.s0 p9,p10 = f8,f1
225 nop.i 999;;
226 }
228 { .mfi
229 ldfpd atanf_coeff_R4,atanf_coeff_R5 = [EXP_Addr1],16
230 fnma.s1 atanf_b = f8,atanf_z,f1
231 nop.i 999
232 }
233 { .mfi
234 ldfpd atanf_coeff_R1,atanf_coeff_R2 = [EXP_Addr2],16
235 fma.s1 atanf_zsq = atanf_z,atanf_z,f0
236 nop.i 999;;
237 }
240 { .mfi
241 ldfpd atanf_coeff_R3,atanf_coeff_P1 = [EXP_Addr1],16
242 fma.s1 atanf_xcub = f8,atanf_t,f0
243 nop.i 999
244 }
245 { .mfi
246 ldfpd atanf_coeff_Q6,atanf_coeff_Q7 = [EXP_Addr2],16
247 fma.s1 atanf_tsq = atanf_t,atanf_t,f0
248 nop.i 999;;
249 }
252 { .mfi
253 ldfpd atanf_coeff_Q8,atanf_coeff_Q9 = [EXP_Addr1],16
254 // fcmp.le.s1 atanf_pred_LE1,atanf_pred_GT1 = atanf_abs_x,f1
255 fcmp.le.s1 p6,p7 = atanf_abs_x,f1
256 nop.i 999
257 }
258 { .mfi
259 ldfpd atanf_coeff_Q4,atanf_coeff_Q5 = [EXP_Addr2],16
260 nop.f 999
261 nop.i 999;;
262 }
265 { .mfi
266 ldfpd atanf_coeff_Q2,atanf_coeff_Q3 = [EXP_Addr1],16
267 fclass.m p8,p0 = f8,0xe7 // @inf|@qnan|@snan|@zero
268 nop.i 999
269 }
270 { .mfi
271 ldfpd atanf_coeff_P5,atanf_coeff_P6 = [EXP_Addr2],16
272 nop.f 999
273 nop.i 999;;
274 }
277 { .mfi
278 ldfpd atanf_coeff_Q0,atanf_coeff_Q1 = [EXP_Addr1],16
279 nop.f 999
280 nop.i 999
281 }
282 { .mfi
283 ldfpd atanf_coeff_P7,atanf_coeff_P8 = [EXP_Addr2],16
284 nop.f 999
285 nop.i 999;;
286 }
289 { .mfi
290 ldfpd atanf_coeff_P3,atanf_coeff_P4 = [EXP_Addr1],16
291 fma.s1 atanf_bsq = atanf_b,atanf_b,f0
292 nop.i 999
293 }
294 { .mfi
295 ldfpd atanf_coeff_P9,atanf_coeff_P10 = [EXP_Addr2]
296 fma.s1 atanf_z4 = atanf_zsq,atanf_zsq,f0
297 nop.i 999;;
298 }
301 { .mfi
302 ldfpd atanf_coeff_P2,atanf_piby2 = [EXP_Addr1]
303 fma.s1 atanf_x6 = atanf_t,atanf_tsq,f0
304 nop.i 999
305 }
306 { .mfi
307 nop.m 999
308 fma.s1 atanf_t4 = atanf_tsq,atanf_tsq,f0
309 nop.i 999;;
310 }
313 { .mfb
314 nop.m 999
315 fma.s1 atanf_x5 = atanf_t,atanf_xcub,f0
316 (p8) br.cond.spnt ATANF_X_INF_NAN_ZERO
317 }
318 ;;
320 { .mfi
321 nop.m 999
322 fma.s1 atanf_poly_r1 = atanf_b,atanf_coeff_R1,f1
323 nop.i 999
324 }
325 { .mfi
326 nop.m 999
327 fma.s1 atanf_poly_r3 = atanf_b,atanf_coeff_R5,atanf_coeff_R4
328 nop.i 999;;
329 }
332 { .mfi
333 nop.m 999
334 fma.s1 atanf_poly_r2 = atanf_b,atanf_coeff_R3,atanf_coeff_R2
335 nop.i 999
336 }
337 { .mfi
338 nop.m 999
339 fma.s1 atanf_z8 = atanf_z4,atanf_z4,f0
340 nop.i 999;;
341 }
344 { .mfi
345 nop.m 999
346 fma.s1 atanf_poly_q2 = atanf_t,atanf_coeff_Q5,atanf_coeff_Q4
347 nop.i 999
348 }
349 { .mfi
350 nop.m 999
351 fma.s1 atanf_poly_q3 = atanf_t,atanf_coeff_Q7,atanf_coeff_Q6
352 nop.i 999;;
353 }
356 { .mfi
357 nop.m 999
358 fma.s1 atanf_z5 = atanf_z,atanf_z4,f0
359 nop.i 999
360 }
361 { .mfi
362 nop.m 999
363 fma.s1 atanf_poly_q1 = atanf_t,atanf_coeff_Q9,atanf_coeff_Q8
364 nop.i 999;;
365 }
368 { .mfi
369 nop.m 999
370 fma.s1 atanf_poly_q4 = atanf_t,atanf_coeff_Q1,atanf_coeff_Q0
371 nop.i 999
372 }
373 { .mfi
374 nop.m 999
375 fma.s1 atanf_poly_q5 = atanf_t,atanf_coeff_Q3,atanf_coeff_Q2
376 nop.i 999;;
377 }
380 { .mfi
381 nop.m 999
382 fma.s1 atanf_poly_p4 = f8,atanf_coeff_P1,f0
383 nop.i 999
384 }
385 { .mfi
386 nop.m 999
387 fma.s1 atanf_poly_p5 = atanf_t,atanf_coeff_P4,atanf_coeff_P3
388 nop.i 999;;
389 }
392 { .mfi
393 nop.m 999
394 fma.s1 atanf_poly_r1 = atanf_z8,atanf_poly_r1,f0
395 nop.i 999
396 }
397 { .mfi
398 nop.m 999
399 fma.s1 atanf_z8_bsq = atanf_z8,atanf_bsq,f0
400 nop.i 999;;
401 }
404 { .mfi
405 nop.m 999
406 fma.s1 atanf_poly_q2 = atanf_tsq,atanf_poly_q3,atanf_poly_q2
407 nop.i 999
408 }
409 { .mfi
410 nop.m 999
411 fma.s1 atanf_poly_r2 = atanf_bsq,atanf_poly_r3,atanf_poly_r2
412 nop.i 999;;
413 }
416 { .mfi
417 nop.m 999
418 fma.s1 atanf_poly_p2 = atanf_t,atanf_coeff_P8,atanf_coeff_P7
419 nop.i 999
420 }
421 { .mfi
422 nop.m 999
423 fma.s1 atanf_poly_q1 = atanf_poly_q1,f1,atanf_tsq
424 nop.i 999;;
425 }
428 { .mfi
429 nop.m 999
430 fma.s1 atanf_z13 = atanf_z5,atanf_z8,f0
431 nop.i 999
432 }
433 { .mfi
434 nop.m 999
435 fma.s1 atanf_poly_p1 = atanf_t,atanf_coeff_P10,atanf_coeff_P9
436 nop.i 999;;
437 }
440 { .mfi
441 nop.m 999
442 fma.s1 atanf_poly_p4 = atanf_t,atanf_poly_p4,f8
443 nop.i 999
444 }
445 { .mfi
446 nop.m 999
447 fma.s1 atanf_poly_q4 = atanf_tsq,atanf_poly_q5,atanf_poly_q4
448 nop.i 999;;
449 }
452 { .mfi
453 nop.m 999
454 fma.s1 atanf_poly_p3 = atanf_t,atanf_coeff_P6,atanf_coeff_P5
455 nop.i 999
456 }
457 { .mfi
458 nop.m 999
459 fma.s1 atanf_poly_p5 = atanf_t,atanf_poly_p5,atanf_coeff_P2
460 nop.i 999;;
461 }
464 { .mfi
465 nop.m 999
466 fma.s1 atanf_x11 = atanf_x5,atanf_x6,f0
467 nop.i 999
468 }
469 { .mfi
470 nop.m 999
471 fma.s1 atanf_poly_r = atanf_z8_bsq,atanf_poly_r2,atanf_poly_r1
472 nop.i 999;;
473 }
476 { .mfi
477 nop.m 999
478 fma.s0 atanf_sgnx_piby2 = atanf_sgn_x,atanf_piby2,f0
479 nop.i 999
480 }
481 { .mfi
482 nop.m 999
483 fma.s1 atanf_poly_q2 = atanf_t4,atanf_poly_q1,atanf_poly_q2
484 nop.i 999;;
485 }
488 { .mfi
489 nop.m 999
490 fma.s1 atanf_poly_p1 = atanf_tsq,atanf_poly_p1,atanf_poly_p2
491 nop.i 999;;
492 }
494 { .mfi
495 nop.m 999
496 fma.s1 atanf_poly_p4 = atanf_x5,atanf_poly_p5,atanf_poly_p4
497 nop.i 999;;
498 }
500 { .mfi
501 nop.m 999
502 fma.s1 atanf_z21_poly_r = atanf_z13,atanf_poly_r,f0
503 nop.i 999;;
504 }
506 { .mfi
507 nop.m 999
508 fma.s1 atanf_poly_q = atanf_t4,atanf_poly_q2,atanf_poly_q4
509 nop.i 999;;
510 }
512 { .mfi
513 nop.m 999
514 fma.s1 atanf_poly_p1 = atanf_tsq,atanf_poly_p1,atanf_poly_p3
515 nop.i 999;;
516 }
518 { .mfi
519 nop.m 999
520 //(atanf_pred_GT1) fnma.s atanf_answer = atanf_poly_q,atanf_z21_poly_r,atanf_sgnx_piby2
521 (p7) fnma.s.s0 atanf_answer = atanf_poly_q,atanf_z21_poly_r,atanf_sgnx_piby2
522 nop.i 999;;
523 }
525 { .mfb
526 nop.m 999
527 //(atanf_pred_LE1) fma.s atanf_answer = atanf_x11,atanf_poly_p1,atanf_poly_p4
528 (p6) fma.s.s0 atanf_answer = atanf_x11,atanf_poly_p1,atanf_poly_p4
529 br.ret.sptk b0
530 }
534 ATANF_X_INF_NAN_ZERO:
536 { .mfi
537 nop.m 0
538 fclass.m p8,p9 = f8,0x23 // @inf
539 nop.i 0
540 }
541 ;;
542 { .mfi
543 nop.m 0
544 (p8) fmerge.s f8 = f8, atanf_piby2
545 nop.i 0
546 }
547 ;;
548 { .mfb
549 nop.m 0
550 fnorm.s.s0 f8 = f8
551 br.ret.sptk b0
552 }
553 ;;
555 GLOBAL_LIBM_END(atanf)