| blob | b10fd2195cb132451cbe3e9188b933ed4aaa1d68 |
1 .file "acos.s"
4 // Copyright (c) 2000 - 2003 Intel Corporation
5 // All rights reserved.
6 //
7 //
8 // Redistribution and use in source and binary forms, with or without
9 // modification, are permitted provided that the following conditions are
10 // met:
11 //
12 // * Redistributions of source code must retain the above copyright
13 // notice, this list of conditions and the following disclaimer.
14 //
15 // * Redistributions in binary form must reproduce the above copyright
16 // notice, this list of conditions and the following disclaimer in the
17 // documentation and/or other materials provided with the distribution.
18 //
19 // * The name of Intel Corporation may not be used to endorse or promote
20 // products derived from this software without specific prior written
21 // permission.
23 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
26 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
27 // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
28 // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
29 // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
30 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
31 // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
32 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
33 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34 //
35 // Intel Corporation is the author of this code, and requests that all
36 // problem reports or change requests be submitted to it directly at
37 // http://www.intel.com/software/products/opensource/libraries/num.htm.
39 // History
40 //==============================================================
41 // 02/02/00 Initial version
42 // 08/17/00 New and much faster algorithm.
43 // 08/30/00 Avoided bank conflicts on loads, shortened |x|=1 and x=0 paths,
44 // fixed mfb split issue stalls.
45 // 05/20/02 Cleaned up namespace and sf0 syntax
46 // 08/02/02 New and much faster algorithm II
47 // 02/06/03 Reordered header: .section, .global, .proc, .align
49 // Description
50 //=========================================
51 // The acos function computes the principal value of the arc cosine of x.
52 // acos(0) returns Pi/2, acos(1) returns 0, acos(-1) returns Pi.
53 // A domain error occurs for arguments not in the range [-1,+1].
54 //
55 // The acos function returns the arc cosine in the range [0, Pi] radians.
56 //
57 // There are 8 paths:
58 // 1. x = +/-0.0
59 // Return acos(x) = Pi/2 + x
60 //
61 // 2. 0.0 < |x| < 0.625
62 // Return acos(x) = Pi/2 - x - x^3 *PolA(x^2)
63 // where PolA(x^2) = A3 + A5*x^2 + A7*x^4 +...+ A35*x^32
64 //
65 // 3. 0.625 <=|x| < 1.0
66 // Return acos(x) = Pi/2 - asin(x) =
67 // = Pi/2 - sign(x) * ( Pi/2 - sqrt(R) * PolB(R))
68 // Where R = 1 - |x|,
69 // PolB(R) = B0 + B1*R + B2*R^2 +...+B12*R^12
70 //
71 // sqrt(R) is approximated using the following sequence:
72 // y0 = (1 + eps)/sqrt(R) - initial approximation by frsqrta,
73 // |eps| < 2^(-8)
74 // Then 3 iterations are used to refine the result:
75 // H0 = 0.5*y0
76 // S0 = R*y0
77 //
78 // d0 = 0.5 - H0*S0
79 // H1 = H0 + d0*H0
80 // S1 = S0 + d0*S0
81 //
82 // d1 = 0.5 - H1*S1
83 // H2 = H1 + d0*H1
84 // S2 = S1 + d0*S1
85 //
86 // d2 = 0.5 - H2*S2
87 // S3 = S3 + d2*S3
88 //
89 // S3 approximates sqrt(R) with enough accuracy for this algorithm
90 //
91 // So, the result should be reconstracted as follows:
92 // acos(x) = Pi/2 - sign(x) * (Pi/2 - S3*PolB(R))
93 //
94 // But for optimization purposes the reconstruction step is slightly
95 // changed:
96 // acos(x) = Cpi + sign(x)*PolB(R)*S2 - sign(x)*d2*S2*PolB(R)
97 // where Cpi = 0 if x > 0 and Cpi = Pi if x < 0
98 //
99 // 4. |x| = 1.0
100 // Return acos(1.0) = 0.0, acos(-1.0) = Pi
101 //
102 // 5. 1.0 < |x| <= +INF
103 // A domain error occurs for arguments not in the range [-1,+1]
104 //
105 // 6. x = [S,Q]NaN
106 // Return acos(x) = QNaN
107 //
108 // 7. x is denormal
109 // Return acos(x) = Pi/2 - x,
110 //
111 // 8. x is unnormal
112 // Normalize input in f8 and return to the very beginning of the function
113 //
114 // Registers used
115 //==============================================================
116 // Floating Point registers used:
117 // f8, input, output
118 // f6, f7, f9 -> f15, f32 -> f64
120 // General registers used:
121 // r3, r21 -> r31, r32 -> r38
123 // Predicate registers used:
124 // p0, p6 -> p14
126 //
127 // Assembly macros
128 //=========================================
129 // integer registers used
130 // scratch
131 rTblAddr = r3
133 rPiBy2Ptr = r21
134 rTmpPtr3 = r22
135 rDenoBound = r23
136 rOne = r24
137 rAbsXBits = r25
138 rHalf = r26
139 r0625 = r27
140 rSign = r28
141 rXBits = r29
142 rTmpPtr2 = r30
143 rTmpPtr1 = r31
145 // stacked
146 GR_SAVE_PFS = r32
147 GR_SAVE_B0 = r33
148 GR_SAVE_GP = r34
149 GR_Parameter_X = r35
150 GR_Parameter_Y = r36
151 GR_Parameter_RESULT = r37
152 GR_Parameter_TAG = r38
154 // floating point registers used
155 FR_X = f10
156 FR_Y = f1
157 FR_RESULT = f8
160 // scratch
161 fXSqr = f6
162 fXCube = f7
163 fXQuadr = f9
164 f1pX = f10
165 f1mX = f11
166 f1pXRcp = f12
167 f1mXRcp = f13
168 fH = f14
169 fS = f15
170 // stacked
171 fA3 = f32
172 fB1 = f32
173 fA5 = f33
174 fB2 = f33
175 fA7 = f34
176 fPiBy2 = f34
177 fA9 = f35
178 fA11 = f36
179 fB10 = f35
180 fB11 = f36
181 fA13 = f37
182 fA15 = f38
183 fB4 = f37
184 fB5 = f38
185 fA17 = f39
186 fA19 = f40
187 fB6 = f39
188 fB7 = f40
189 fA21 = f41
190 fA23 = f42
191 fB3 = f41
192 fB8 = f42
193 fA25 = f43
194 fA27 = f44
195 fB9 = f43
196 fB12 = f44
197 fA29 = f45
198 fA31 = f46
199 fA33 = f47
200 fA35 = f48
201 fBaseP = f49
202 fB0 = f50
203 fSignedS = f51
204 fD = f52
205 fHalf = f53
206 fR = f54
207 fCloseTo1Pol = f55
208 fSignX = f56
209 fDenoBound = f57
210 fNormX = f58
211 fX8 = f59
212 fRSqr = f60
213 fRQuadr = f61
214 fR8 = f62
215 fX16 = f63
216 fCpi = f64
218 // Data tables
219 //==============================================================
220 RODATA
221 .align 16
222 LOCAL_OBJECT_START(acos_base_range_table)
223 // Ai: Polynomial coefficients for the acos(x), |x| < .625000
224 // Bi: Polynomial coefficients for the acos(x), |x| > .625000
225 data8 0xBFDAAB56C01AE468 //A29
226 data8 0x3FE1C470B76A5B2B //A31
227 data8 0xBFDC5FF82A0C4205 //A33
228 data8 0x3FC71FD88BFE93F0 //A35
229 data8 0xB504F333F9DE6487, 0x00003FFF //B0
230 data8 0xAAAAAAAAAAAAFC18, 0x00003FFC //A3
231 data8 0x3F9F1C71BC4A7823 //A9
232 data8 0x3F96E8BBAAB216B2 //A11
233 data8 0x3F91C4CA1F9F8A98 //A13
234 data8 0x3F8C9DDCEDEBE7A6 //A15
235 data8 0x3F877784442B1516 //A17
236 data8 0x3F859C0491802BA2 //A19
237 data8 0x9999999998C88B8F, 0x00003FFB //A5
238 data8 0x3F6BD7A9A660BF5E //A21
239 data8 0x3F9FC1659340419D //A23
240 data8 0xB6DB6DB798149BDF, 0x00003FFA //A7
241 data8 0xBFB3EF18964D3ED3 //A25
242 data8 0x3FCD285315542CF2 //A27
243 data8 0xF15BEEEFF7D2966A, 0x00003FFB //B1
244 data8 0x3EF0DDA376D10FB3 //B10
245 data8 0xBEB83CAFE05EBAC9 //B11
246 data8 0x3F65FFB67B513644 //B4
247 data8 0x3F5032FBB86A4501 //B5
248 data8 0x3F392162276C7CBA //B6
249 data8 0x3F2435949FD98BDF //B7
250 data8 0xD93923D7FA08341C, 0x00003FF9 //B2
251 data8 0x3F802995B6D90BDB //B3
252 data8 0x3F10DF86B341A63F //B8
253 data8 0xC90FDAA22168C235, 0x00003FFF // Pi/2
254 data8 0x3EFA3EBD6B0ECB9D //B9
255 data8 0x3EDE18BA080E9098 //B12
256 LOCAL_OBJECT_END(acos_base_range_table)
258 .section .text
259 GLOBAL_LIBM_ENTRY(acos)
260 acos_unnormal_back:
261 { .mfi
262 getf.d rXBits = f8 // grab bits of input value
263 // set p12 = 1 if x is a NaN, denormal, or zero
264 fclass.m p12, p0 = f8, 0xcf
265 adds rSign = 1, r0
266 }
267 { .mfi
268 addl rTblAddr = @ltoff(acos_base_range_table),gp
269 // 1 - x = 1 - |x| for positive x
270 fms.s1 f1mX = f1, f1, f8
271 addl rHalf = 0xFFFE, r0 // exponent of 1/2
272 }
273 ;;
274 { .mfi
275 addl r0625 = 0x3FE4, r0 // high 16 bits of 0.625
276 // set p8 = 1 if x < 0
277 fcmp.lt.s1 p8, p9 = f8, f0
278 shl rSign = rSign, 63 // sign bit
279 }
280 { .mfi
281 // point to the beginning of the table
282 ld8 rTblAddr = [rTblAddr]
283 // 1 + x = 1 - |x| for negative x
284 fma.s1 f1pX = f1, f1, f8
285 adds rOne = 0x3FF, r0
286 }
287 ;;
288 { .mfi
289 andcm rAbsXBits = rXBits, rSign // bits of |x|
290 fmerge.s fSignX = f8, f1 // signum(x)
291 shl r0625 = r0625, 48 // bits of DP representation of 0.625
292 }
293 { .mfb
294 setf.exp fHalf = rHalf // load A2 to FP reg
295 fma.s1 fXSqr = f8, f8, f0 // x^2
296 // branch on special path if x is a NaN, denormal, or zero
297 (p12) br.cond.spnt acos_special
298 }
299 ;;
300 { .mfi
301 adds rPiBy2Ptr = 272, rTblAddr
302 nop.f 0
303 shl rOne = rOne, 52 // bits of 1.0
304 }
305 { .mfi
306 adds rTmpPtr1 = 16, rTblAddr
307 nop.f 0
308 // set p6 = 1 if |x| < 0.625
309 cmp.lt p6, p7 = rAbsXBits, r0625
310 }
311 ;;
312 { .mfi
313 ldfpd fA29, fA31 = [rTblAddr] // A29, fA31
314 // 1 - x = 1 - |x| for positive x
315 (p9) fms.s1 fR = f1, f1, f8
316 // point to coefficient of "near 1" polynomial
317 (p7) adds rTmpPtr2 = 176, rTblAddr
318 }
319 { .mfi
320 ldfpd fA33, fA35 = [rTmpPtr1], 16 // A33, fA35
321 // 1 + x = 1 - |x| for negative x
322 (p8) fma.s1 fR = f1, f1, f8
323 (p6) adds rTmpPtr2 = 48, rTblAddr
324 }
325 ;;
326 { .mfi
327 ldfe fB0 = [rTmpPtr1], 16 // B0
328 nop.f 0
329 nop.i 0
330 }
331 { .mib
332 adds rTmpPtr3 = 16, rTmpPtr2
333 // set p10 = 1 if |x| = 1.0
334 cmp.eq p10, p0 = rAbsXBits, rOne
335 // branch on special path for |x| = 1.0
336 (p10) br.cond.spnt acos_abs_1
337 }
338 ;;
339 { .mfi
340 ldfe fA3 = [rTmpPtr2], 48 // A3 or B1
341 nop.f 0
342 adds rTmpPtr1 = 64, rTmpPtr3
343 }
344 { .mib
345 ldfpd fA9, fA11 = [rTmpPtr3], 16 // A9, A11 or B10, B11
346 // set p11 = 1 if |x| > 1.0
347 cmp.gt p11, p0 = rAbsXBits, rOne
348 // branch on special path for |x| > 1.0
349 (p11) br.cond.spnt acos_abs_gt_1
350 }
351 ;;
352 { .mfi
353 ldfpd fA17, fA19 = [rTmpPtr2], 16 // A17, A19 or B6, B7
354 // initial approximation of 1 / sqrt(1 - x)
355 frsqrta.s1 f1mXRcp, p0 = f1mX
356 nop.i 0
357 }
358 { .mfi
359 ldfpd fA13, fA15 = [rTmpPtr3] // A13, A15 or B4, B5
360 fma.s1 fXCube = fXSqr, f8, f0 // x^3
361 nop.i 0
362 }
363 ;;
364 { .mfi
365 ldfe fA5 = [rTmpPtr2], 48 // A5 or B2
366 // initial approximation of 1 / sqrt(1 + x)
367 frsqrta.s1 f1pXRcp, p0 = f1pX
368 nop.i 0
369 }
370 { .mfi
371 ldfpd fA21, fA23 = [rTmpPtr1], 16 // A21, A23 or B3, B8
372 fma.s1 fXQuadr = fXSqr, fXSqr, f0 // x^4
373 nop.i 0
374 }
375 ;;
376 { .mfi
377 ldfe fA7 = [rTmpPtr1] // A7 or Pi/2
378 fma.s1 fRSqr = fR, fR, f0 // R^2
379 nop.i 0
380 }
381 { .mfb
382 ldfpd fA25, fA27 = [rTmpPtr2] // A25, A27 or B9, B12
383 nop.f 0
384 (p6) br.cond.spnt acos_base_range;
385 }
386 ;;
388 { .mfi
389 nop.m 0
390 (p9) fma.s1 fH = fHalf, f1mXRcp, f0 // H0 for x > 0
391 nop.i 0
392 }
393 { .mfi
394 nop.m 0
395 (p9) fma.s1 fS = f1mX, f1mXRcp, f0 // S0 for x > 0
396 nop.i 0
397 }
398 ;;
399 { .mfi
400 nop.m 0
401 (p8) fma.s1 fH = fHalf, f1pXRcp, f0 // H0 for x < 0
402 nop.i 0
403 }
404 { .mfi
405 nop.m 0
406 (p8) fma.s1 fS = f1pX, f1pXRcp, f0 // S0 for x > 0
407 nop.i 0
408 }
409 ;;
410 { .mfi
411 nop.m 0
412 fma.s1 fRQuadr = fRSqr, fRSqr, f0 // R^4
413 nop.i 0
414 }
415 ;;
416 { .mfi
417 nop.m 0
418 fma.s1 fB11 = fB11, fR, fB10
419 nop.i 0
420 }
421 { .mfi
422 nop.m 0
423 fma.s1 fB1 = fB1, fR, fB0
424 nop.i 0
425 }
426 ;;
427 { .mfi
428 nop.m 0
429 fma.s1 fB5 = fB5, fR, fB4
430 nop.i 0
431 }
432 { .mfi
433 nop.m 0
434 fma.s1 fB7 = fB7, fR, fB6
435 nop.i 0
436 }
437 ;;
438 { .mfi
439 nop.m 0
440 fma.s1 fB3 = fB3, fR, fB2
441 nop.i 0
442 }
443 ;;
444 { .mfi
445 nop.m 0
446 fnma.s1 fD = fH, fS, fHalf // d0 = 1/2 - H0*S0
447 nop.i 0
448 }
449 ;;
450 { .mfi
451 nop.m 0
452 fma.s1 fR8 = fRQuadr, fRQuadr, f0 // R^4
453 nop.i 0
454 }
455 { .mfi
456 nop.m 0
457 fma.s1 fB9 = fB9, fR, fB8
458 nop.i 0
459 }
460 ;;
461 {.mfi
462 nop.m 0
463 fma.s1 fB12 = fB12, fRSqr, fB11
464 nop.i 0
465 }
466 {.mfi
467 nop.m 0
468 fma.s1 fB7 = fB7, fRSqr, fB5
469 nop.i 0
470 }
471 ;;
472 {.mfi
473 nop.m 0
474 fma.s1 fB3 = fB3, fRSqr, fB1
475 nop.i 0
476 }
477 ;;
478 { .mfi
479 nop.m 0
480 fma.s1 fH = fH, fD, fH // H1 = H0 + H0*d0
481 nop.i 0
482 }
483 { .mfi
484 nop.m 0
485 fma.s1 fS = fS, fD, fS // S1 = S0 + S0*d0
486 nop.i 0
487 }
488 ;;
489 {.mfi
490 nop.m 0
491 (p9) fma.s1 fCpi = f1, f0, f0 // Cpi = 0 if x > 0
492 nop.i 0
493 }
494 { .mfi
495 nop.m 0
496 (p8) fma.s1 fCpi = fPiBy2, f1, fPiBy2 // Cpi = Pi if x < 0
497 nop.i 0
498 }
499 ;;
500 { .mfi
501 nop.m 0
502 fma.s1 fB12 = fB12, fRSqr, fB9
503 nop.i 0
504 }
505 { .mfi
506 nop.m 0
507 fma.s1 fB7 = fB7, fRQuadr, fB3
508 nop.i 0
509 }
510 ;;
511 {.mfi
512 nop.m 0
513 fnma.s1 fD = fH, fS, fHalf // d1 = 1/2 - H1*S1
514 nop.i 0
515 }
516 { .mfi
517 nop.m 0
518 fnma.s1 fSignedS = fSignX, fS, f0 // -signum(x)*S1
519 nop.i 0
520 }
521 ;;
522 { .mfi
523 nop.m 0
524 fma.s1 fCloseTo1Pol = fB12, fR8, fB7
525 nop.i 0
526 }
527 ;;
528 { .mfi
529 nop.m 0
530 fma.s1 fH = fH, fD, fH // H2 = H1 + H1*d1
531 nop.i 0
532 }
533 { .mfi
534 nop.m 0
535 fma.s1 fS = fS, fD, fS // S2 = S1 + S1*d1
536 nop.i 0
537 }
538 ;;
539 { .mfi
540 nop.m 0
541 // -signum(x)* S2 = -signum(x)*(S1 + S1*d1)
542 fma.s1 fSignedS = fSignedS, fD, fSignedS
543 nop.i 0
544 }
545 ;;
546 {.mfi
547 nop.m 0
548 fnma.s1 fD = fH, fS, fHalf // d2 = 1/2 - H2*S2
549 nop.i 0
550 }
551 ;;
552 { .mfi
553 nop.m 0
554 // Cpi + signum(x)*PolB*S2
555 fnma.s1 fCpi = fSignedS, fCloseTo1Pol, fCpi
556 nop.i 0
557 }
558 { .mfi
559 nop.m 0
560 // signum(x)*PolB * S2
561 fnma.s1 fCloseTo1Pol = fSignedS, fCloseTo1Pol, f0
562 nop.i 0
563 }
564 ;;
565 { .mfb
566 nop.m 0
567 // final result for 0.625 <= |x| < 1
568 fma.d.s0 f8 = fCloseTo1Pol, fD, fCpi
569 // exit here for 0.625 <= |x| < 1
570 br.ret.sptk b0
571 }
572 ;;
575 // here if |x| < 0.625
576 .align 32
577 acos_base_range:
578 { .mfi
579 ldfe fCpi = [rPiBy2Ptr] // Pi/2
580 fma.s1 fA33 = fA33, fXSqr, fA31
581 nop.i 0
582 }
583 { .mfi
584 nop.m 0
585 fma.s1 fA15 = fA15, fXSqr, fA13
586 nop.i 0
587 }
588 ;;
589 { .mfi
590 nop.m 0
591 fma.s1 fA29 = fA29, fXSqr, fA27
592 nop.i 0
593 }
594 { .mfi
595 nop.m 0
596 fma.s1 fA25 = fA25, fXSqr, fA23
597 nop.i 0
598 }
599 ;;
600 { .mfi
601 nop.m 0
602 fma.s1 fA21 = fA21, fXSqr, fA19
603 nop.i 0
604 }
605 { .mfi
606 nop.m 0
607 fma.s1 fA9 = fA9, fXSqr, fA7
608 nop.i 0
609 }
610 ;;
611 { .mfi
612 nop.m 0
613 fma.s1 fA5 = fA5, fXSqr, fA3
614 nop.i 0
615 }
616 ;;
617 { .mfi
618 nop.m 0
619 fma.s1 fA35 = fA35, fXQuadr, fA33
620 nop.i 0
621 }
622 { .mfi
623 nop.m 0
624 fma.s1 fA17 = fA17, fXQuadr, fA15
625 nop.i 0
626 }
627 ;;
628 { .mfi
629 nop.m 0
630 fma.s1 fX8 = fXQuadr, fXQuadr, f0 // x^8
631 nop.i 0
632 }
633 { .mfi
634 nop.m 0
635 fma.s1 fA25 = fA25, fXQuadr, fA21
636 nop.i 0
637 }
638 ;;
639 { .mfi
640 nop.m 0
641 fma.s1 fA9 = fA9, fXQuadr, fA5
642 nop.i 0
643 }
644 ;;
645 { .mfi
646 nop.m 0
647 fms.s1 fCpi = fCpi, f1, f8 // Pi/2 - x
648 nop.i 0
649 }
650 ;;
651 { .mfi
652 nop.m 0
653 fma.s1 fA35 = fA35, fXQuadr, fA29
654 nop.i 0
655 }
656 { .mfi
657 nop.m 0
658 fma.s1 fA17 = fA17, fXSqr, fA11
659 nop.i 0
660 }
661 ;;
662 { .mfi
663 nop.m 0
664 fma.s1 fX16 = fX8, fX8, f0 // x^16
665 nop.i 0
666 }
667 ;;
668 { .mfi
669 nop.m 0
670 fma.s1 fA35 = fA35, fX8, fA25
671 nop.i 0
672 }
673 { .mfi
674 nop.m 0
675 fma.s1 fA17 = fA17, fX8, fA9
676 nop.i 0
677 }
678 ;;
679 { .mfi
680 nop.m 0
681 fma.s1 fBaseP = fA35, fX16, fA17
682 nop.i 0
683 }
684 ;;
685 { .mfb
686 nop.m 0
687 // final result for |x| < 0.625
688 fnma.d.s0 f8 = fBaseP, fXCube, fCpi
689 // exit here for |x| < 0.625 path
690 br.ret.sptk b0
691 }
692 ;;
694 // here if |x| = 1
695 // acos(1) = 0
696 // acos(-1) = Pi
697 .align 32
698 acos_abs_1:
699 { .mfi
700 ldfe fPiBy2 = [rPiBy2Ptr] // Pi/2
701 nop.f 0
702 nop.i 0
703 }
704 ;;
705 .pred.rel "mutex", p8, p9
706 { .mfi
707 nop.m 0
708 // result for x = 1.0
709 (p9) fma.d.s0 f8 = f1, f0, f0 // 0.0
710 nop.i 0
711 }
712 {.mfb
713 nop.m 0
714 // result for x = -1.0
715 (p8) fma.d.s0 f8 = fPiBy2, f1, fPiBy2 // Pi
716 // exit here for |x| = 1.0
717 br.ret.sptk b0
718 }
719 ;;
721 // here if x is a NaN, denormal, or zero
722 .align 32
723 acos_special:
724 { .mfi
725 // point to Pi/2
726 adds rPiBy2Ptr = 272, rTblAddr
727 // set p12 = 1 if x is a NaN
728 fclass.m p12, p0 = f8, 0xc3
729 nop.i 0
730 }
731 { .mlx
732 nop.m 0
733 // smallest positive DP normalized number
734 movl rDenoBound = 0x0010000000000000
735 }
736 ;;
737 { .mfi
738 ldfe fPiBy2 = [rPiBy2Ptr] // Pi/2
739 // set p13 = 1 if x = 0.0
740 fclass.m p13, p0 = f8, 0x07
741 nop.i 0
742 }
743 { .mfi
744 nop.m 0
745 fnorm.s1 fNormX = f8
746 nop.i 0
747 }
748 ;;
749 { .mfb
750 // load smallest normal to FP reg
751 setf.d fDenoBound = rDenoBound
752 // answer if x is a NaN
753 (p12) fma.d.s0 f8 = f8,f1,f0
754 // exit here if x is a NaN
755 (p12) br.ret.spnt b0
756 }
757 ;;
758 { .mfi
759 nop.m 0
760 // absolute value of normalized x
761 fmerge.s fNormX = f1, fNormX
762 nop.i 0
763 }
764 ;;
765 { .mfb
766 nop.m 0
767 // final result for x = 0
768 (p13) fma.d.s0 f8 = fPiBy2, f1, f8
769 // exit here if x = 0.0
770 (p13) br.ret.spnt b0
771 }
772 ;;
773 // if we still here then x is denormal or unnormal
774 { .mfi
775 nop.m 0
776 // set p14 = 1 if normalized x is greater than or
777 // equal to the smallest denormalized value
778 // So, if p14 is set to 1 it means that we deal with
779 // unnormal rather than with "true" denormal
780 fcmp.ge.s1 p14, p0 = fNormX, fDenoBound
781 nop.i 0
782 }
783 ;;
784 { .mfi
785 nop.m 0
786 (p14) fcmp.eq.s0 p6, p0 = f8, f0 // Set D flag if x unnormal
787 nop.i 0
788 }
789 { .mfb
790 nop.m 0
791 // normalize unnormal input
792 (p14) fnorm.s1 f8 = f8
793 // return to the main path
794 (p14) br.cond.sptk acos_unnormal_back
795 }
796 ;;
797 // if we still here it means that input is "true" denormal
798 { .mfb
799 nop.m 0
800 // final result if x is denormal
801 fms.d.s0 f8 = fPiBy2, f1, f8 // Pi/2 - x
802 // exit here if x is denormal
803 br.ret.sptk b0
804 }
805 ;;
807 // here if |x| > 1.0
808 // error handler should be called
809 .align 32
810 acos_abs_gt_1:
811 { .mfi
812 alloc r32 = ar.pfs, 0, 3, 4, 0 // get some registers
813 fmerge.s FR_X = f8,f8
814 nop.i 0
815 }
816 { .mfb
817 mov GR_Parameter_TAG = 58 // error code
818 frcpa.s0 FR_RESULT, p0 = f0,f0
819 // call error handler routine
820 br.cond.sptk __libm_error_region
821 }
822 ;;
823 GLOBAL_LIBM_END(acos)
824 libm_alias_double_other (acos, acos)
828 LOCAL_LIBM_ENTRY(__libm_error_region)
829 .prologue
830 { .mfi
831 add GR_Parameter_Y=-32,sp // Parameter 2 value
832 nop.f 0
833 .save ar.pfs,GR_SAVE_PFS
834 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
835 }
836 { .mfi
837 .fframe 64
838 add sp=-64,sp // Create new stack
839 nop.f 0
840 mov GR_SAVE_GP=gp // Save gp
841 };;
842 { .mmi
843 stfd [GR_Parameter_Y] = FR_Y,16 // STORE Parameter 2 on stack
844 add GR_Parameter_X = 16,sp // Parameter 1 address
845 .save b0, GR_SAVE_B0
846 mov GR_SAVE_B0=b0 // Save b0
847 };;
848 .body
849 { .mib
850 stfd [GR_Parameter_X] = FR_X // STORE Parameter 1 on stack
851 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
852 nop.b 0
853 }
854 { .mib
855 stfd [GR_Parameter_Y] = FR_RESULT // STORE Parameter 3 on stack
856 add GR_Parameter_Y = -16,GR_Parameter_Y
857 br.call.sptk b0=__libm_error_support# // Call error handling function
858 };;
859 { .mmi
860 add GR_Parameter_RESULT = 48,sp
861 nop.m 0
862 nop.i 0
863 };;
864 { .mmi
865 ldfd f8 = [GR_Parameter_RESULT] // Get return result off stack
866 .restore sp
867 add sp = 64,sp // Restore stack pointer
868 mov b0 = GR_SAVE_B0 // Restore return address
869 };;
870 { .mib
871 mov gp = GR_SAVE_GP // Restore gp
872 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
873 br.ret.sptk b0 // Return
874 };;
876 LOCAL_LIBM_END(__libm_error_region)
877 .type __libm_error_support#,@function
878 .global __libm_error_support#