| blob | 5649670d1993ccfc241f2e71bde60df79e61b59f |
1 .file "libm_atan2_reg.s"
3 // Copyright (C) 2000, 2001, Intel Corporation
4 // All rights reserved.
5 //
6 // Contributed 2/2/2000 by John Harrison, Ted Kubaska, Bob Norin, Shane Story,
7 // and Ping Tak Peter Tang of the Computational Software Lab, 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.
23 //
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://developer.intel.com/opensource.
39 //
40 // History
41 //==============================================================
42 // 2/02/00: Initial version
43 // 4/04/00 Unwind support added
45 #include "libm_support.h"
47 .data
49 .align 64
50 ASM_TYPE_DIRECTIVE(Constants_atan#,@object)
51 Constants_atan:
52 data4 0x54442D18, 0x3FF921FB, 0x248D3132, 0x3E000000
53 // double pi/2, single lo_pi/2, two**(-3)
54 data4 0xAAAAAAA3, 0xAAAAAAAA, 0x0000BFFD, 0x00000000 // P_1
55 data4 0xCCCC54B2, 0xCCCCCCCC, 0x00003FFC, 0x00000000 // P_2
56 data4 0x47E4D0C2, 0x92492492, 0x0000BFFC, 0x00000000 // P_3
57 data4 0x58870889, 0xE38E38E0, 0x00003FFB, 0x00000000 // P_4
58 data4 0x290149F8, 0xBA2E895B, 0x0000BFFB, 0x00000000 // P_5
59 data4 0x250F733D, 0x9D88E6D4, 0x00003FFB, 0x00000000 // P_6
60 data4 0xFB8745A0, 0x884E51FF, 0x0000BFFB, 0x00000000 // P_7
61 data4 0x394396BD, 0xE1C7412B, 0x00003FFA, 0x00000000 // P_8
62 data4 0xAAAAA52F, 0xAAAAAAAA, 0x0000BFFD, 0x00000000 // Q_1
63 data4 0xC75B60D3, 0xCCCCCCCC, 0x00003FFC, 0x00000000 // Q_2
64 data4 0x011F1940, 0x924923AD, 0x0000BFFC, 0x00000000 // Q_3
65 data4 0x2A5F89BD, 0xE36F716D, 0x00003FFB, 0x00000000 // Q_4
66 // Entries Tbl_hi (double precision)
67 // B = 1+Index/16+1/32 Index = 0
68 // Entries Tbl_lo (single precision)
69 // B = 1+Index/16+1/32 Index = 0
70 data4 0xA935BD8E, 0x3FE9A000, 0x23ACA08F, 0x00000000
71 // Entries Tbl_hi (double precision) Index = 0,1,...,15
72 // B = 2^(-1)*(1+Index/16+1/32)
73 // Entries Tbl_lo (single precision)
74 // Index = 0,1,...,15 B = 2^(-1)*(1+Index/16+1/32)
75 data4 0x7F175A34, 0x3FDE77EB, 0x238729EE, 0x00000000
76 data4 0x73C1A40B, 0x3FE0039C, 0x249334DB, 0x00000000
77 data4 0x5B5B43DA, 0x3FE0C614, 0x22CBA7D1, 0x00000000
78 data4 0x88BE7C13, 0x3FE1835A, 0x246310E7, 0x00000000
79 data4 0xE2CC9E6A, 0x3FE23B71, 0x236210E5, 0x00000000
80 data4 0x8406CBCA, 0x3FE2EE62, 0x2462EAF5, 0x00000000
81 data4 0x1CD41719, 0x3FE39C39, 0x24B73EF3, 0x00000000
82 data4 0x5B795B55, 0x3FE44506, 0x24C11260, 0x00000000
83 data4 0x5BB6EC04, 0x3FE4E8DE, 0x242519EE, 0x00000000
84 data4 0x1F732FBA, 0x3FE587D8, 0x24D4346C, 0x00000000
85 data4 0x115D7B8D, 0x3FE6220D, 0x24ED487B, 0x00000000
86 data4 0x920B3D98, 0x3FE6B798, 0x2495FF1E, 0x00000000
87 data4 0x8FBA8E0F, 0x3FE74897, 0x223D9531, 0x00000000
88 data4 0x289FA093, 0x3FE7D528, 0x242B0411, 0x00000000
89 data4 0x576CC2C5, 0x3FE85D69, 0x2335B374, 0x00000000
90 data4 0xA99CC05D, 0x3FE8E17A, 0x24C27CFB, 0x00000000
91 //
92 // Entries Tbl_hi (double precision) Index = 0,1,...,15
93 // B = 2^(-2)*(1+Index/16+1/32)
94 // Entries Tbl_lo (single precision)
95 // Index = 0,1,...,15 B = 2^(-2)*(1+Index/16+1/32)
96 //
97 data4 0x510665B5, 0x3FD025FA, 0x24263482, 0x00000000
98 data4 0x362431C9, 0x3FD1151A, 0x242C8DC9, 0x00000000
99 data4 0x67E47C95, 0x3FD20255, 0x245CF9BA, 0x00000000
100 data4 0x7A823CFE, 0x3FD2ED98, 0x235C892C, 0x00000000
101 data4 0x29271134, 0x3FD3D6D1, 0x2389BE52, 0x00000000
102 data4 0x586890E6, 0x3FD4BDEE, 0x24436471, 0x00000000
103 data4 0x175E0F4E, 0x3FD5A2E0, 0x2389DBD4, 0x00000000
104 data4 0x9F5FA6FD, 0x3FD68597, 0x2476D43F, 0x00000000
105 data4 0x52817501, 0x3FD76607, 0x24711774, 0x00000000
106 data4 0xB8DF95D7, 0x3FD84422, 0x23EBB501, 0x00000000
107 data4 0x7CD0C662, 0x3FD91FDE, 0x23883A0C, 0x00000000
108 data4 0x66168001, 0x3FD9F930, 0x240DF63F, 0x00000000
109 data4 0x5422058B, 0x3FDAD00F, 0x23FE261A, 0x00000000
110 data4 0x378624A5, 0x3FDBA473, 0x23A8CD0E, 0x00000000
111 data4 0x0AAD71F8, 0x3FDC7655, 0x2422D1D0, 0x00000000
112 data4 0xC9EC862B, 0x3FDD45AE, 0x2344A109, 0x00000000
113 //
114 // Entries Tbl_hi (double precision) Index = 0,1,...,15
115 // B = 2^(-3)*(1+Index/16+1/32)
116 // Entries Tbl_lo (single precision)
117 // Index = 0,1,...,15 B = 2^(-3)*(1+Index/16+1/32)
118 //
119 data4 0x84212B3D, 0x3FC068D5, 0x239874B6, 0x00000000
120 data4 0x41060850, 0x3FC16465, 0x2335E774, 0x00000000
121 data4 0x171A535C, 0x3FC25F6E, 0x233E36BE, 0x00000000
122 data4 0xEDEB99A3, 0x3FC359E8, 0x239680A3, 0x00000000
123 data4 0xC6092A9E, 0x3FC453CE, 0x230FB29E, 0x00000000
124 data4 0xBA11570A, 0x3FC54D18, 0x230C1418, 0x00000000
125 data4 0xFFB3AA73, 0x3FC645BF, 0x23F0564A, 0x00000000
126 data4 0xE8A7D201, 0x3FC73DBD, 0x23D4A5E1, 0x00000000
127 data4 0xE398EBC7, 0x3FC8350B, 0x23D4ADDA, 0x00000000
128 data4 0x7D050271, 0x3FC92BA3, 0x23BCB085, 0x00000000
129 data4 0x601081A5, 0x3FCA217E, 0x23BC841D, 0x00000000
130 data4 0x574D780B, 0x3FCB1696, 0x23CF4A8E, 0x00000000
131 data4 0x4D768466, 0x3FCC0AE5, 0x23BECC90, 0x00000000
132 data4 0x4E1D5395, 0x3FCCFE65, 0x2323DCD2, 0x00000000
133 data4 0x864C9D9D, 0x3FCDF110, 0x23F53F3A, 0x00000000
134 data4 0x451D980C, 0x3FCEE2E1, 0x23CCB11F, 0x00000000
135 data4 0x54442D18, 0x400921FB, 0x33145C07, 0x3CA1A626 // I two doubles
136 data4 0x54442D18, 0x3FF921FB, 0x33145C07, 0x3C91A626 // I_by_2 two dbls
137 data4 0x54442D18, 0x3FE921FB, 0x33145C07, 0x3C81A626 // I_by_4 two dbls
138 data4 0x7F3321D2, 0x4002D97C, 0x4C9E8A0A, 0x3C9A7939 // 3I_by_4 two dbls
139 ASM_SIZE_DIRECTIVE(Constants_atan#)
140 .section .text
142 .proc __libm_atan2_reg#
143 .global __libm_atan2_reg#
144 .align 64
145 __libm_atan2_reg:
148 { .mfi
149 alloc r32 = ar.pfs,0,20,4,0
150 (p0) mov f32 = f8
151 nop.i 0
152 }
153 { .mmi
154 nop.m 0
155 (p0) addl r39 = @ltoff(Constants_atan#), gp
156 nop.i 999
157 }
158 ;;
160 { .mmi
161 ld8 r39 = [r39]
162 nop.m 999
163 nop.i 999
164 }
165 ;;
167 { .mfi
168 nop 999 // EMbo added ...
169 (p0) mov f33 = f9
170 nop.i 0
171 } { .mfi
172 nop 999 // EMbo added ...
173 (p0) fclass.nm.unc p9,p0 = f32 ,0x1FF
174 nop 999;; // EMbo added ...
175 } { .mfi
176 nop 999 // EMbo added ...
177 (p0) fclass.nm.unc p8,p0 = f33 ,0x1FF
178 nop 999 // EMbo added ...
179 } { .mfi
180 nop 999 // EMbo added ...
181 (p0) fclass.m.unc p6,p0 = f33 ,0x103
182 nop 999;; // EMbo added ...
183 } { .mfi
184 nop 999 // EMbo added ...
185 (p0) fclass.m.unc p7,p0 = f32 ,0x103
186 nop 999 // EMbo added ...
187 } { .mfi
188 nop 999 // EMbo added ...
189 (p0) fclass.m.unc p12,p0 = f33 ,0x0C3
190 nop 999;; // EMbo added ...
191 } { .mfb
192 nop 999 // EMbo added ...
193 //
194 // Check for NatVals.
195 // Check for EM Unsupporteds
196 // Check for NaNs.
197 //
198 (p0) fclass.m.unc p13,p0 = f32 ,0x0C3
199 (p6) br.cond.sptk L(ATAN_NATVAL);;
200 } { .mbb
201 nop 999 // EMbo added ...
202 (p7) br.cond.sptk L(ATAN_NATVAL)
203 (p8) br.cond.sptk L(ATAN_UNSUPPORTED);;
204 } { .mib
205 (p0) add r40 = 96, r39
206 nop 999 // EMbo added ...
207 (p9) br.cond.sptk L(ATAN_UNSUPPORTED);;
208 } { .mib
209 (p0) ldfd f50 = [r39],8
210 nop 999 // EMbo added ...
211 (p12) br.cond.sptk L(ATAN_NAN);;
212 } { .mfb
213 nop 999 // EMbo added ...
214 (p0) fnorm.s1 f33 = f33
215 (p13) br.cond.sptk L(ATAN_NAN);;
216 } { .mfi
217 (p0) ldfs f51 = [r39],4
218 //
219 // Remove sign bits from exponents
220 // Load 2**(-3)
221 // Normalize the input argument.
222 //
223 (p0) fnorm.s1 f32 = f32
224 nop 999 // EMbo added ...
225 } { .mfi
226 nop 999 // EMbo added ...
227 (p0) mov f82 = f1
228 nop 999;; // EMbo added ...
229 } { .mmi
230 nop 999;; // EMbo added ...
231 (p0) ldfs f78 = [r39],180
232 nop 999;; // EMbo added ...
233 } { .mmi
234 (p0) getf.exp r36 = f33;;
235 //
236 // Get exp and sign of ArgX
237 // Get exp and sign of ArgY
238 // Load 2**(-3) and increment ptr to Q_4.
239 //
240 (p0) getf.exp r37 = f32
241 (p0) shr.u r36 = r36,17;;
242 } { .mfi
243 nop 999 // EMbo added ...
244 (p0) fmerge.s f84 = f1,f32
245 (p0) shr.u r37 = r37,17;;
246 } { .mfi
247 nop 999 // EMbo added ...
248 //
249 // ArgX_abs = |ArgX|
250 // ArgY_abs = |ArgY|
251 // sign_X is sign bit of ArgX
252 // sign_Y is sign bit of ArgY
253 //
254 (p0) fmerge.s f83 = f1,f33
255 (p0) cmp.eq.unc p8,p9 = 0x00000, r37;;
256 } { .mfi
257 nop 999 // EMbo added ...
258 (p8) fadd.s1 f34 = f0, f1
259 nop 999;; // EMbo added ...
260 } { .mfi
261 nop 999 // EMbo added ...
262 (p9) fsub.s1 f34 = f0, f1
263 nop 999;; // EMbo added ...
264 } { .mfi
265 nop 999 // EMbo added ...
266 (p0) fmin.s1 f36 = f83, f84
267 nop 999 // EMbo added ...
268 } { .mfi
269 nop 999 // EMbo added ...
270 (p0) fmax.s1 f35 = f83, f84
271 nop 999;; // EMbo added ...
272 } { .mfi
273 nop 999 // EMbo added ...
274 //
275 // Is ArgX_abs >= ArgY_abs
276 // Is sign_Y == 0?
277 //
278 (p0) fcmp.ge.s1 p6,p7 = f83,f84
279 nop 999;; // EMbo added ...
280 } { .mii
281 (p6) cmp.eq.unc p10, p11 = 0x00000, r36
282 (p6) add r38 = r0, r0;;
283 //
284 // U = max(ArgX_abs,ArgY_abs)
285 // V = min(ArgX_abs,ArgY_abs)
286 // if p6, swap = 0
287 // if p7, swap = 1
288 //
289 //
290 // Let M = 1.0
291 // if p8, s_Y = 1.0
292 // if p9, s_Y = -1.0
293 //
294 (p7) add r38 = 1,r0;;
295 } { .mfi
296 nop 999 // EMbo added ...
297 (p0) frcpa.s1 f37, p6 = f36, f35
298 nop 999;; // EMbo added ...
299 } { .mfb
300 nop 999 // EMbo added ...
301 //
302 // E = frcpa(V,U)
303 //
304 (p10) fsub.s1 f82 = f82, f1
305 (p6) br.cond.sptk L(ATAN_STEP2);;
306 } { .mib
307 nop 999 // EMbo added ...
308 nop 999 // EMbo added ...
309 // /**************************************************/
310 // /********************* STEP2 **********************/
311 // /**************************************************/
312 (p0) br.cond.spnt L(ATAN_SPECIAL_HANDLING);;
313 }
314 L(ATAN_STEP2):
315 { .mlx
316 nop 999 // EMbo added ...
317 (p0) movl r47 = 0x8400000000000000
318 } { .mlx
319 nop 999 // EMbo added ...
320 (p0) movl r48 = 0x0000000000000100;;
321 } { .mfi
322 nop 999 // EMbo added ...
323 (p0) fmpy.s1 f38 = f37, f36
324 nop 999 // EMbo added ...
325 } { .mfi
326 nop 999 // EMbo added ...
327 (p0) fcmp.lt.unc.s0 p0,p9 = f9,f1
328 nop 999;; // EMbo added ...
329 } { .mfi
330 nop 999 // EMbo added ...
331 (p0) fcmp.lt.unc.s0 p0,p8 = f8,f1
332 nop 999 // EMbo added ...
333 } { .mfi
334 nop 999 // EMbo added ...
335 //
336 // Q = E * V
337 //
338 (p11) fadd.s1 f82 = f82, f1
339 nop 999;; // EMbo added ...
340 } { .mfi
341 (p0) getf.sig r46 = f38
342 (p0) fcmp.lt.unc p6,p7 = f38,f78
343 nop 999;; // EMbo added ...
344 } { .mfi
345 nop 999 // EMbo added ...
346 (p0) fmpy.s1 f38 = f37, f36
347 (p0) extr.u r42 = r46, 59, 4;;
348 } { .mfi
349 nop 999 // EMbo added ...
350 (p0) fmpy.s1 f50 = f82, f50
351 (p0) dep r47 = r42, r47, 59, 4
352 } { .mfi
353 nop 999 // EMbo added ...
354 (p0) fmpy.s1 f51 = f82, f51
355 nop 999;; // EMbo added ...
356 } { .mmi
357 nop 999;; // EMbo added ...
358 //
359 // Is Q < 2**(-3)?
360 //
361 //
362 // Do fcmp to raise any denormal operand
363 // exceptions.
364 //
365 (p0) getf.exp r45 = f38
366 nop 999;; // EMbo added ...
367 } { .mib
368 //
369 // lookup = b_1 b_2 b_3 B_4
370 //
371 //
372 // Generate 1.b_1 b_2 b_3 b_4 1 0 0 0 ... 0
373 //
374 (p0) andcm r41 = 0x0003, r45
375 nop 999 // EMbo added ...
376 //
377 // We waited a few extra cycles so P_lo and P_hi could be calculated.
378 // Load the constant 256 for loading up table entries.
379 //
380 // /**************************************************/
381 // /********************* STEP3 **********************/
382 // /**************************************************/
383 (p6) br.cond.spnt L(ATAN_POLY);;
384 } { .mii
385 (p0) setf.sig f39 = r47
386 (p0) cmp.eq.unc p8, p9 = 0x0000, r41
387 //
388 // z_hi = s exp 1.b_1 b_2 b_3 b_4 1 0 0 0 ... 0
389 // point to beginning of Tbl_hi entries - k = 0.
390 //
391 (p0) add r40 = 16, r39
392 } { .mmi
393 (p0) ldfe f73 = [r39],-16;;
394 (p9) sub r41 = r41,r0,1
395 (p9) add r40 = 16,r40
396 } { .mfi
397 (p8) ldfd f48 = [r40],8
398 (p0) fmpy.s1 f50 = f34, f50
399 (p0) xor r38 = r36,r38;;
400 } { .mmi
401 (p0) ldfe f71 = [r39],-16;;
402 (p8) ldfs f49 = [r40],8
403 (p9) pmpy2.r r41 = r41,r48;;
404 } { .mfi
405 (p0) ldfe f69 = [r39],-16
406 //
407 // Let z_hi have exponent and sign of original Q
408 // Load the Tbl_hi(0) else, increment pointer.
409 //
410 (p0) fmerge.se f39 = f38,f39
411 (p9) shladd r42 = r42,0x0004,r41;;
412 } { .mmi
413 (p9) add r40 = r40, r42;;
414 (p9) ldfd f48 = [r40],8
415 nop 999;; // EMbo added ...
416 } { .mmi
417 (p0) ldfe f67 = [r39],-16;;
418 (p9) ldfs f49 = [r40],8
419 nop 999 // EMbo added ...
420 } { .mfi
421 nop 999 // EMbo added ...
422 //
423 // U_prime_hi = U + V * z_hi
424 // Load the Tbl_lo(0)
425 //
426 (p0) fma.s1 f40 = f36, f39, f35
427 nop 999;; // EMbo added ...
428 } { .mfi
429 nop 999 // EMbo added ...
430 (p0) fnma.s1 f42 = f35, f39, f36
431 nop 999 // EMbo added ...
432 } { .mfi
433 nop 999 // EMbo added ...
434 (p0) mov f52 = f48
435 nop 999;; // EMbo added ...
436 } { .mfi
437 nop 999 // EMbo added ...
438 (p0) frcpa.s1 f43, p6 = f1, f40
439 nop 999;; // EMbo added ...
440 } { .mfi
441 nop 999 // EMbo added ...
442 //
443 // U_prime_lo = U - U_prime_hi
444 // k = k * 256 - result can be 0, 256, or 512.
445 //
446 (p0) fsub.s1 f41 = f35, f40
447 (p0) cmp.eq.unc p7, p6 = 0x00000, r38
448 } { .mfi
449 nop 999 // EMbo added ...
450 (p0) fmpy.s1 f52 = f34, f52
451 nop 999;; // EMbo added ...
452 } { .mfi
453 nop 999 // EMbo added ...
454 (p7) fadd.s1 f54 = f0, f1
455 nop 999;; // EMbo added ...
456 } { .mfi
457 nop 999 // EMbo added ...
458 (p6) fsub.s1 f54 = f0, f1
459 nop 999;; // EMbo added ...
460 } { .mfi
461 nop 999 // EMbo added ...
462 (p0) fnma.s1 f80 = f43, f40, f1
463 nop 999;; // EMbo added ...
464 } { .mfi
465 nop 999 // EMbo added ...
466 (p0) fadd.s1 f79 = f41, f40
467 nop 999 // EMbo added ...
468 } { .mfi
469 nop 999 // EMbo added ...
470 (p0) fma.s1 f41 = f36, f39, f41
471 nop 999;; // EMbo added ...
472 } { .mfi
473 nop 999 // EMbo added ...
474 (p0) fma.s1 f56 = f54, f52, f50
475 nop 999;; // EMbo added ...
476 } { .mfi
477 nop 999 // EMbo added ...
478 (p0) fma.s1 f43 = f80, f43, f43
479 nop 999;; // EMbo added ...
480 } { .mfi
481 nop 999 // EMbo added ...
482 //
483 // U_prime_lo = U - U_hold
484 // lookup -> lookup * 16 + k
485 //
486 //
487 // V_prime = V - U * z_hi
488 // U_prime_lo = V * z_hi + U_prime_lo
489 //
490 (p0) fsub.s1 f79 = f35, f79
491 nop 999;; // EMbo added ...
492 } { .mfi
493 nop 999 // EMbo added ...
494 (p0) fnma.s1 f80 = f43, f40, f1
495 nop 999;; // EMbo added ...
496 } { .mfi
497 nop 999 // EMbo added ...
498 //
499 // C_hi = frcpa(1,U_prime_hi)
500 // U_prime_lo = U_prime_lo + U_hold
501 //
502 //
503 // C_hi_hold = 1 - C_hi * U_prime_hi (1)
504 //
505 //
506 // C_hi = C_hi + C_hi * C_hi_hold (1)
507 //
508 //
509 // C_hi_hold = 1 - C_hi * U_prime_hi (2)
510 //
511 (p0) fadd.s1 f41 = f41, f79
512 nop 999;; // EMbo added ...
513 } { .mfi
514 nop 999 // EMbo added ...
515 //
516 // C_hi = C_hi + C_hi * C_hi_hold (2)
517 //
518 (p0) fma.s1 f43 = f80, f43, f43
519 nop 999;; // EMbo added ...
520 } { .mfi
521 nop 999 // EMbo added ...
522 //
523 // C_hi_hold = 1 - C_hi * U_prime_hi (3)
524 //
525 (p0) fnma.s1 f80 = f43, f40, f1
526 nop 999;; // EMbo added ...
527 } { .mfi
528 nop 999 // EMbo added ...
529 //
530 // C_hi = C_hi + C_hi * C_hi_hold (3)
531 //
532 (p0) fma.s1 f43 = f80, f43, f43
533 nop 999;; // EMbo added ...
534 } { .mfi
535 nop 999 // EMbo added ...
536 //
537 // w_hi = V_prime * C_hi
538 //
539 (p0) fmpy.s1 f44 = f42, f43
540 nop 999;; // EMbo added ...
541 } { .mfi
542 nop 999 // EMbo added ...
543 (p0) fmpy.s1 f46 = f44, f44
544 nop 999 // EMbo added ...
545 } { .mfi
546 nop 999 // EMbo added ...
547 //
548 // wsq = w_hi * w_hi
549 // w_lo = = V_prime - w_hi * U_prime_hi
550 //
551 (p0) fnma.s1 f45 = f44, f40, f42
552 nop 999;; // EMbo added ...
553 } { .mfi
554 nop 999 // EMbo added ...
555 (p0) fma.s1 f47 = f46, f73, f71
556 nop 999 // EMbo added ...
557 } { .mfi
558 nop 999 // EMbo added ...
559 //
560 // poly = Q_3 + wsq * Q_4
561 // w_lo = = w_lo - w_hi * U_prime_lo
562 //
563 (p0) fnma.s1 f45 = f44, f41, f45
564 nop 999;; // EMbo added ...
565 } { .mfi
566 nop 999 // EMbo added ...
567 (p0) fma.s1 f47 = f46, f47, f69
568 nop 999 // EMbo added ...
569 } { .mfi
570 nop 999 // EMbo added ...
571 //
572 // poly = Q_2 + wsq * poly
573 // w_lo = = w_lo * C_hi
574 //
575 (p0) fmpy.s1 f45 = f43, f45
576 nop 999;; // EMbo added ...
577 } { .mfi
578 nop 999 // EMbo added ...
579 (p0) fma.s1 f47 = f46, f47, f67
580 nop 999 // EMbo added ...
581 } { .mfi
582 nop 999 // EMbo added ...
583 //
584 // poly = Q_1 + wsq * poly
585 // A_lo = Tbl_lo + w_lo
586 // swap = xor(swap,sign_X)
587 //
588 (p0) fadd.s1 f53 = f49, f45
589 nop 999;; // EMbo added ...
590 } { .mfi
591 nop 999 // EMbo added ...
592 //
593 // Is (swap) != 0 ?
594 // poly = wsq * poly
595 // A_hi = Tbl_hi
596 //
597 (p0) fmpy.s1 f47 = f46, f47
598 nop 999;; // EMbo added ...
599 } { .mfi
600 nop 999 // EMbo added ...
601 //
602 // poly = wsq * poly
603 //
604 //
605 // if (p6) sigma = -1.0
606 // if (p7) sigma = 1.0
607 //
608 (p0) fmpy.s1 f47 = f44, f47
609 nop 999;; // EMbo added ...
610 } { .mfi
611 nop 999 // EMbo added ...
612 //
613 // P_hi = s_Y * P_hi
614 // A_lo = A_lo + poly
615 //
616 (p0) fadd.s1 f53 = f53, f47
617 nop 999;; // EMbo added ...
618 } { .mfi
619 nop 999 // EMbo added ...
620 //
621 // A_lo = A_lo + w_hi
622 // A_hi = s_Y * A_hi
623 //
624 (p0) fadd.s1 f53 = f53, f44
625 nop 999;; // EMbo added ...
626 } { .mfb
627 nop 999 // EMbo added ...
628 //
629 // result_hi = P_hi + sigma * A_hi
630 // result_lo = P_lo + sigma * A_lo
631 //
632 (p0) fma.s1 f55 = f54, f53, f51
633 (p0) br.cond.sptk L(RETURN_ATAN);;
634 }
635 //
636 // result = result_hi + result_lo * s_Y (User Supplied Rounding Mode)
637 //
638 // (p0) fma.d.s0 f57 = f55, f34, f56
639 //
640 // /**************************************************/
641 // /********************* STEP4 **********************/
642 // /**************************************************/
643 //
644 L(ATAN_POLY):
645 { .mmi
646 (p0) xor r38 = r36,r38
647 (p0) addl r39 = @ltoff(Constants_atan#), gp
648 nop.i 999
649 }
650 ;;
652 { .mmi
653 ld8 r39 = [r39]
654 nop.m 999
655 nop.i 999
656 }
657 ;;
660 { .mlx
661 nop 999 // EMbo added ...
662 (p0) movl r47 = 0x24005;;
663 } { .mfi
664 (p0) add r39 = 128, r39
665 (p0) fnma.s1 f81 = f37, f35, f1
666 (p0) cmp.eq.unc p7, p6 = 0x00000, r38;;
667 } { .mmf
668 nop 999 // EMbo added ...
669 (p0) ldfe f77 = [r39],-16
670 //
671 // Iterate 3 times E = E + E*(1.0 - E*U)
672 // Also load P_8, P_7, P_6, P_5, P_4
673 // E_hold = 1.0 - E * U (1)
674 // A_temp = Q
675 //
676 (p0) mov f85 = f38;;
677 } { .mmf
678 nop 999 // EMbo added ...
679 (p0) ldfe f76 = [r39],-16
680 (p6) fsub.s1 f54 = f0, f1;;
681 } { .mmf
682 nop 999 // EMbo added ...
683 (p0) ldfe f75 = [r39],-16
684 //
685 // E = E + E_hold*E (1)
686 // Point to P_8.
687 //
688 (p0) fma.s1 f37 = f37, f81, f37;;
689 } { .mmf
690 nop 999 // EMbo added ...
691 (p0) ldfe f74 = [r39],-16
692 (p0) fnma.s1 f64 = f85, f35, f36;;
693 } { .mmf
694 nop 999 // EMbo added ...
695 (p0) ldfe f72 = [r39],-16
696 (p7) fadd.s1 f54 = f0, f1;;
697 } { .mmf
698 nop 999 // EMbo added ...
699 (p0) ldfe f70 = [r39],-16
700 //
701 // E_hold = 1.0 - E * U (2)
702 //
703 (p0) fnma.s1 f81 = f37, f35, f1;;
704 } { .mmf
705 nop 999 // EMbo added ...
706 (p0) ldfe f68 = [r39],-16
707 (p0) fmpy.s1 f50 = f34, f50;;
708 } { .mmf
709 nop 999 // EMbo added ...
710 (p0) ldfe f66 = [r39],-16
711 (p0) fmpy.d.s0 f67 = f67, f67
712 } { .mfi
713 nop 999 // EMbo added ...
714 //
715 // E = E + E_hold*E (2)
716 //
717 (p0) fma.s1 f37 = f37, f81, f37
718 nop 999;; // EMbo added ...
719 } { .mfi
720 nop 999 // EMbo added ...
721 //
722 // E_hold = 1.0 - E * U (3)
723 //
724 (p0) fnma.s1 f81 = f37, f35, f1
725 nop 999;; // EMbo added ...
726 } { .mfi
727 nop 999 // EMbo added ...
728 //
729 // E = E + E_hold*E (3)
730 // At this point E approximates 1/U to roughly working precision
731 // z = V*E approximates V/U
732 //
733 (p0) fma.s1 f37 = f37, f81, f37
734 nop 999;; // EMbo added ...
735 } { .mfi
736 nop 999 // EMbo added ...
737 //
738 // z = V * E
739 //
740 (p0) fmpy.s1 f59 = f36, f37
741 nop 999 // EMbo added ...
742 } { .mfi
743 nop 999 // EMbo added ...
744 (p0) fmpy.s1 f64 = f64, f37
745 nop 999;; // EMbo added ...
746 } { .mfi
747 nop 999 // EMbo added ...
748 //
749 // zsq = z * z
750 // Also load P_3
751 //
752 (p0) fmpy.s1 f60 = f59, f59
753 nop 999 // EMbo added ...
754 } { .mfi
755 nop 999 // EMbo added ...
756 (p0) fadd.s1 f52 = f85, f64
757 nop 999;; // EMbo added ...
758 } { .mfi
759 nop 999 // EMbo added ...
760 (p0) fma.s1 f62 = f60, f77, f76
761 nop 999 // EMbo added ...
762 } { .mfi
763 nop 999 // EMbo added ...
764 (p0) fma.s1 f63 = f60, f70, f68
765 nop 999;; // EMbo added ...
766 } { .mfi
767 nop 999 // EMbo added ...
768 //
769 // z8 = zsq * zsq
770 // Also load P_2
771 //
772 (p0) fmpy.s1 f61 = f60, f60
773 nop 999 // EMbo added ...
774 } { .mfi
775 nop 999 // EMbo added ...
776 (p0) fsub.s1 f85 = f85, f52
777 nop 999;; // EMbo added ...
778 } { .mfi
779 nop 999 // EMbo added ...
780 (p0) fmerge.s f65 = f52,f52
781 nop 999;; // EMbo added ...
782 } { .mfi
783 nop 999 // EMbo added ...
784 (p0) fma.s1 f62 = f60, f62, f75
785 nop 999 // EMbo added ...
786 } { .mfi
787 nop 999 // EMbo added ...
788 (p0) fma.s1 f63 = f60, f63, f66
789 nop 999;; // EMbo added ...
790 } { .mfi
791 nop 999 // EMbo added ...
792 //
793 // z8 = z8 * z8
794 // Also load P_1
795 // poly1 = _4 + zsq*(P_5 + zsq*(P_6 + zsq*(P_7 + zsq*P_8)))
796 // poly2 = zsq*(P_1 + zsq*(P_2 + zsq*P_3))
797 //
798 //
799 // poly1 = P_7 + zsq * P_8
800 // poly2 = P_2 + zsq * P_3
801 // poly1 = P_4 + zsq*(P_5 + zsq*(P_6 + zsq*poly1))
802 // poly2 = zsq*(P_1 + zsq*poly2)
803 //
804 //
805 // poly1 = P_6 + zsq * poly1
806 // poly2 = P_1 + zsq * poly2
807 // poly1 = P_4 + zsq*(P_5 + zsq*poly1)
808 // poly2 = zsq*poly2
809 //
810 (p0) fmpy.s1 f61 = f61, f61
811 nop 999 // EMbo added ...
812 } { .mfi
813 nop 999 // EMbo added ...
814 (p0) fadd.s1 f64 = f85, f64
815 nop 999;; // EMbo added ...
816 } { .mfi
817 nop 999 // EMbo added ...
818 (p0) fma.s1 f62 = f60, f62, f74
819 nop 999 // EMbo added ...
820 } { .mfi
821 nop 999 // EMbo added ...
822 //
823 // poly1 = P_5 + zsq * poly1
824 // poly2 = zsq * poly2
825 // poly1 = P_4 + zsq*poly1
826 //
827 (p0) fmpy.s1 f63 = f63, f60
828 nop 999;; // EMbo added ...
829 } { .mfi
830 nop 999 // EMbo added ...
831 //
832 // poly1 = P_4 + zsq * poly1
833 // swap = xor(swap,sign_X)
834 //
835 (p0) fma.s1 f62 = f60, f62, f72
836 nop 999;; // EMbo added ...
837 } { .mfi
838 nop 999 // EMbo added ...
839 //
840 // poly = z8*poly1 + poly2 (Typo in writeup)
841 // Is (swap) != 0 ?
842 //
843 //
844 // z_lo = V - A_temp * U
845 // if (p7) sigma = 1.0
846 // Writeup shows A_temp as A_hi
847 //
848 //
849 // z_lo = z_lo * E
850 // if (p6) sigma = -1.0
851 // z_lo = (V - A_temp * U) *E
852 //
853 //
854 // Fixup added to force inexact later -
855 // A_hi = A_temp + z_lo
856 // z_lo = (A_temp - A_hi) + z_lo
857 // z_lo = A_hi - z_lo -A_hi + z_lo = about 0
858 //
859 (p0) fma.s1 f47 = f61, f62, f63
860 nop 999;; // EMbo added ...
861 } { .mfi
862 nop 999 // EMbo added ...
863 //
864 // A_lo = z * poly + z_lo
865 //
866 (p0) fma.s1 f53 = f59, f47, f64
867 nop 999;; // EMbo added ...
868 } { .mfi
869 nop 999 // EMbo added ...
870 (p0) fadd.s1 f52 = f65, f53
871 nop 999;; // EMbo added ...
872 } { .mfi
873 nop 999 // EMbo added ...
874 (p0) fsub.s1 f65 = f65, f52
875 nop 999 // EMbo added ...
876 } { .mfi
877 nop 999 // EMbo added ...
878 (p0) fmpy.s1 f52 = f34, f52
879 nop 999;; // EMbo added ...
880 } { .mfi
881 nop 999 // EMbo added ...
882 (p0) fadd.s1 f53 = f65, f53
883 nop 999 // EMbo added ...
884 } { .mfi
885 (p0) setf.exp f65 = r47
886 (p0) fma.s1 f56 = f54, f52, f50
887 nop 999;; // EMbo added ...
888 } { .mfi
889 nop 999 // EMbo added ...
890 (p0) fclass.m.unc p6,p0 = f53,0x007
891 nop 999;; // EMbo added ...
892 } { .mfi
893 nop 999 // EMbo added ...
894 //
895 // P_hi = s_Y * P_hi
896 // A_hi = s_Y * A_hi
897 //
898 //
899 // result_hi = P_hi + sigma * A_hi
900 //
901 (p6) mov f53 = f65
902 nop 999 // EMbo added ...
903 } { .mfi
904 nop 999 // EMbo added ...
905 //
906 // tmp = P_hi - result_hi
907 //
908 (p0) fsub.s1 f65 = f50, f56
909 nop 999;; // EMbo added ...
910 } { .mfi
911 nop 999 // EMbo added ...
912 (p0) fma.s1 f65 = f52, f54, f65
913 nop 999 // EMbo added ...
914 } { .mfi
915 nop 999 // EMbo added ...
916 //
917 // tmp = sigma * A_hi + tmp
918 // sigma = A_lo * sigma + P_lo
919 //
920 (p0) fma.s1 f54 = f53, f54, f51
921 nop 999;; // EMbo added ...
922 } { .mfi
923 nop 999 // EMbo added ...
924 //
925 // result_lo = s_Y * sigma + tmp
926 //
927 (p0) fma.s1 f55 = f34, f54, f65
928 nop 999;; // EMbo added ...
929 } { .mfb
930 nop.m 0
931 mov f34 = f1
932 (p0) br.cond.sptk L(RETURN_ATAN);;
933 }
934 //
935 // result = result_hi + result_lo (User Supplied Rounding Mode)
936 //
937 // (p0) fadd.d.s0 f57 = f55, f56
938 L(ATAN_UNSUPPORTED):
939 L(ATAN_NATVAL):
940 { .mfb
941 nop 999 // EMbo added ...
942 //
943 // Deal with the NatVal and unsupported cases.
944 // Raise invalid if warrented.
945 //
946 (p0) fmpy.d.s0 f57 = f8, f9
947 br.cond.sptk L(RETURN_ATAN);;
948 }
949 L(ATAN_NAN):
950 { .mfb
951 nop 999 // EMbo added ...
952 //
953 // If only one NaN, then generate the resulting
954 // NaN and return - may raise invalid.
955 //
956 (p0) fmpy.d.s0 f57 = f8, f9
957 (p0) br.cond.sptk L(RETURN_ATAN);;
958 }
959 L(ATAN_SPECIAL_HANDLING):
961 { .mmf
962 (p0) addl r39 = @ltoff(Constants_atan#), gp
963 nop.m 999
964 (p0) fcmp.lt.s0 p0,p7 = f8,f1
965 }
966 ;;
968 //
969 // Raise denormal operand faults if necessary
970 //
972 { .mfi
973 ld8 r39 = [r39]
974 (p0) fcmp.lt.s0 p0,p6 = f9,f1
975 nop 999;; // EMbo added ...
976 }
977 ;;
981 { .mfi
982 nop 999 // EMbo added ...
983 (p0) fclass.m.unc p6,p7 = f32,0x007
984 nop 999;; // EMbo added ...
985 } { .mlx
986 nop 999 // EMbo added ...
987 (p0) movl r47 = 992;;
988 } { .mib
989 (p0) add r39 = r39, r47
990 nop 999 // EMbo added ...
991 (p7) br.cond.sptk L(ATAN_ArgY_Not_ZERO);;
992 } { .mfi
993 nop 999 // EMbo added ...
994 (p6) fclass.m.unc p14,p0 = f33,0x035
995 nop 999 // EMbo added ...
996 } { .mfi
997 nop 999 // EMbo added ...
998 (p6) fclass.m.unc p15,p0 = f33,0x036
999 nop 999;; // EMbo added ...
1000 } { .mfi
1001 nop 999 // EMbo added ...
1002 (p6) fclass.m.unc p13,p0 = f33,0x007
1003 nop 999 // EMbo added ...
1004 } { .mfi
1005 (p0) ldfd f56 = [r39],8
1006 nop 999 // EMbo added ...
1007 nop 999;; // EMbo added ...
1008 } { .mfi
1009 (p0) ldfd f55 = [r39],-8
1010 (p14) fmerge.s f56 = f32,f0
1011 nop 999;; // EMbo added ...
1012 } { .mfi
1013 nop 999 // EMbo added ...
1014 //
1015 // Return sign_Y * 0 when Y = +/-0 and X > 0
1016 //
1017 (p14) fmerge.s f55 = f32,f0
1018 nop 999;; // EMbo added ...
1019 } { .mfi
1020 nop 999 // EMbo added ...
1021 (p15) fmerge.s f56 = f32,f56
1022 nop 999;; // EMbo added ...
1023 } { .mfi
1024 nop 999 // EMbo added ...
1025 //
1026 // Return sign_Y * PI when X < -0
1027 //
1028 //
1029 (p15) fmerge.s f55 = f32,f55
1030 nop 999;; // EMbo added ...
1031 } { .mfi
1032 nop 999 // EMbo added ...
1033 (p0) fadd.d.s0 f57 = f56,f55
1034 nop.i 0
1035 } { .bbb
1036 //
1037 // Call error support function for atan(0,0)
1038 // - expected value already computed.
1039 //
1040 nop.b 0
1041 nop.b 0
1042 (p0) br.cond.sptk L(RETURN_ATAN)
1043 }
1044 L(ATAN_ArgY_Not_ZERO):
1045 { .mfi
1046 nop 999 // EMbo added ...
1047 (p0) fclass.m.unc p9,p10 = f32,0x023
1048 nop 999;; // EMbo added ...
1049 } { .mfb
1050 nop 999 // EMbo added ...
1051 (p9) fclass.m.unc p6,p0 = f33,0x017
1052 (p10) br.cond.sptk L(ATAN_ArgY_Not_INF);;
1053 } { .mfi
1054 (p6) add r39 = 16,r39
1055 (p9) fclass.m.unc p7,p0 = f33,0x021
1056 nop 999;; // EMbo added ...
1057 } { .mmf
1058 nop 999 // EMbo added ...
1059 (p0) ldfd f56 = [r39],8
1060 (p9) fclass.m.unc p8,p0 = f33,0x022;;
1061 } { .mbb
1062 (p0) ldfd f55 = [r39],-8
1063 nop 999 // EMbo added ...
1064 nop 999;; // EMbo added ...
1065 } { .mfi
1066 nop 999 // EMbo added ...
1067 (p6) fmerge.s f56 = f32,f56
1068 nop 999;; // EMbo added ...
1069 } { .mfi
1070 nop 999 // EMbo added ...
1071 (p6) fmerge.s f55 = f32,f55
1072 nop 999;; // EMbo added ...
1073 } { .mfb
1074 nop 999 // EMbo added ...
1075 //
1076 // Load I/2 and adjust its sign.
1077 // Return +I/2 when ArgY = +Inf and ArgX = +/-0,normal
1078 // Return -I/2 when ArgY = -Inf and ArgX = +/-0,normal
1079 //
1080 (p6) fadd.d.s0 f57 = f56, f55
1081 (p6) br.cond.sptk L(RETURN_ATAN);;
1082 } { .mmi
1083 (p7) add r39 = 32,r39;;
1084 (p7) ldfd f56 = [r39],8
1085 nop 999;; // EMbo added ...
1086 } { .mmi
1087 nop 999;; // EMbo added ...
1088 (p7) ldfd f55 = [r39],-8
1089 nop 999;; // EMbo added ...
1090 } { .mfi
1091 nop 999 // EMbo added ...
1092 (p7) fmerge.s f56 = f32,f56
1093 nop 999;; // EMbo added ...
1094 } { .mfi
1095 nop 999 // EMbo added ...
1096 (p7) fmerge.s f55 = f32,f55
1097 nop 999;; // EMbo added ...
1098 } { .mfb
1099 nop 999 // EMbo added ...
1100 //
1101 // Load PI/4 and adjust its sign.
1102 // Return +PI/4 when ArgY = +Inf and ArgX = +Inf
1103 // Return -PI/4 when ArgY = -Inf and ArgX = +Inf
1104 //
1105 (p7) fadd.d.s0 f57 = f56, f55
1106 (p7) br.cond.sptk L(RETURN_ATAN);;
1107 } { .mmi
1108 (p8) add r39 = 48,r39;;
1109 (p8) ldfd f56 =[r39],8
1110 nop 999;; // EMbo added ...
1111 } { .mmi
1112 nop 999;; // EMbo added ...
1113 (p8) ldfd f55 =[r39],-8
1114 nop 999;; // EMbo added ...
1115 } { .mfi
1116 nop 999 // EMbo added ...
1117 (p8) fmerge.s f56 = f32,f56
1118 nop 999;; // EMbo added ...
1119 } { .mfi
1120 nop 999 // EMbo added ...
1121 (p8) fmerge.s f55 = f32,f55
1122 nop 999;; // EMbo added ...
1123 } { .mfb
1124 nop 999 // EMbo added ...
1125 //
1126 // Load I/4 and adjust its sign.
1127 // Return +3I/4 when ArgY = +Inf and ArgX = -Inf
1128 // Return -3I/4 when ArgY = -Inf and ArgX = -Inf
1129 //
1130 (p8) fadd.d.s0 f57 = f56, f55
1131 (p8) br.cond.sptk L(RETURN_ATAN);;
1132 }
1133 L(ATAN_ArgY_Not_INF):
1134 { .mfi
1135 nop 999 // EMbo added ...
1136 (p0) fclass.m.unc p6,p0 = f33,0x007
1137 nop 999 // EMbo added ...
1138 } { .mfi
1139 nop 999 // EMbo added ...
1140 (p0) fclass.m.unc p7,p0 = f33,0x021
1141 nop 999;; // EMbo added ...
1142 } { .mfi
1143 nop 999 // EMbo added ...
1144 (p0) fclass.m.unc p8,p0 = f33,0x022
1145 (p6) add r39 = 16,r39;;
1146 } { .mfi
1147 (p6) ldfd f56 =[r39],8
1148 nop 999 // EMbo added ...
1149 nop 999;; // EMbo added ...
1150 } { .mmi
1151 nop 999;; // EMbo added ...
1152 (p6) ldfd f55 =[r39],-8
1153 nop 999;; // EMbo added ...
1154 } { .mfi
1155 nop 999 // EMbo added ...
1156 (p6) fmerge.s f56 = f32,f56
1157 nop 999;; // EMbo added ...
1158 } { .mfi
1159 nop 999 // EMbo added ...
1160 (p6) fmerge.s f55 = f32,f55
1161 nop 999;; // EMbo added ...
1162 } { .mfb
1163 nop 999 // EMbo added ...
1164 //
1165 // return = sign_Y * I/2 when ArgX = +/-0
1166 //
1167 (p6) fadd.d.s0 f57 = f56, f55
1168 (p6) br.cond.sptk L(RETURN_ATAN);;
1169 } { .mfi
1170 nop 999 // EMbo added ...
1171 (p7) fmerge.s f56 = f32,f0
1172 nop 999 // EMbo added ...
1173 } { .mfi
1174 nop 999 // EMbo added ...
1175 (p7) fmerge.s f55 = f32,f0
1176 nop 999;; // EMbo added ...
1177 } { .mfb
1178 nop 999 // EMbo added ...
1179 //
1180 // return = sign_Y * 0 when ArgX = Inf
1181 //
1182 (p7) fadd.d.s0 f57 = f56, f55
1183 (p7) br.cond.sptk L(RETURN_ATAN);;
1184 } { .mfi
1185 (p8) ldfd f56 = [r39],8
1186 nop 999 // EMbo added ...
1187 nop 999;; // EMbo added ...
1188 } { .mmi
1189 nop 999;; // EMbo added ...
1190 (p8) ldfd f55 = [r39],-8
1191 nop 999;; // EMbo added ...
1192 } { .mfi
1193 nop 999 // EMbo added ...
1194 (p8) fmerge.s f56 = f32,f56
1195 nop 999;; // EMbo added ...
1196 } { .mfi
1197 nop 999 // EMbo added ...
1198 (p8) fmerge.s f55 = f32,f55
1199 nop 999;; // EMbo added ...
1200 } { .mfi
1201 nop 999 // EMbo added ...
1202 //
1203 // return = sign_Y * I when ArgX = -Inf
1204 //
1205 (p8) fadd.d.s0 f57 = f56, f55
1206 nop 999 // EMbo added ...
1207 };;
1208 L(RETURN_ATAN):
1209 // mov f8 = f57 ;;
1210 // The answer is in f57.
1211 // But Z_hi is f56
1212 // Z_lo is f55
1213 // s_Y is f34
1214 // W is in f9 and untouched
1216 { .mfi
1217 nop 999
1218 mov f8 = f56
1219 nop.i 0
1220 };;
1222 { .mfi
1223 nop 999
1224 mov f10 = f55
1225 nop.i 999
1226 }
1227 { .mfb
1228 nop 999
1229 mov f11 = f34
1230 br.ret.sptk b0
1231 };;
1233 .endp __libm_atan2_reg
1234 ASM_SIZE_DIRECTIVE(__libm_atan2_reg)