4 // Copyright (c) 2000 - 2003, Intel Corporation
5 // All rights reserved.
8 // Redistribution and use in source and binary forms, with or without
9 // modification, are permitted provided that the following conditions are
12 // * Redistributions of source code must retain the above copyright
13 // notice, this list of conditions and the following disclaimer.
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.
19 // * The name of Intel Corporation may not be used to endorse or promote
20 // products derived from this software without specific prior written
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.
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.
40 //==============================================================
41 // 02/02/00 Initial version
42 // 04/04/00 Unwind support added
43 // 08/15/00 Bundle added after call to __libm_error_support to properly
44 // set [the previously overwritten] GR_Parameter_RESULT.
45 // 08/17/00 Changed predicate register macro-usage to direct predicate
46 // names due to an assembler bug.
47 // 09/28/00 Updated to set invalid on SNaN inputs
48 // 01/19/01 Fixed flags for small results
49 // 04/13/01 Rescheduled to make all paths faster
50 // 05/20/02 Cleaned up namespace and sf0 syntax
51 // 08/20/02 Corrected inexact flag and directed rounding symmetry bugs
52 // 02/06/03 Reordered header: .section, .global, .proc, .align
53 // 04/17/03 Added missing mutex directive
54 // 12/23/03 atan2(NaN1,NaN2) now QNaN1, for consistency with atan2f, atan2l
57 //==============================================================
58 // double atan2(double Y, double X)
60 // Overview of operation
61 //==============================================================
63 // The atan2 function returns values in the interval [-pi,+pi].
65 // There are two basic paths: swap true and swap false.
66 // atan2(Y,X) ==> atan2(V/U) where U >= V. If Y > X, we must swap.
68 // p6 swap True |Y| > |X|
69 // p7 swap False |Y| <= |X|
70 // p8 X+ (If swap=True p8=p9=0)
73 // all the other predicates p10 thru p15 are false for the main path
75 // Simple trigonometric identities show
76 // Region 1 (-45 to +45 degrees):
77 // X>0, |Y|<=X, V=Y, U=X atan2(Y,X) = sgnY * (0 + atan(V/U))
79 // Region 2 (-90 to -45 degrees, and +45 to +90 degrees):
80 // X>0, |Y|>X, V=X, U=Y atan2(Y,X) = sgnY * (pi/2 - atan(V/U))
82 // Region 3 (-135 to -90 degrees, and +90 to +135 degrees):
83 // X<0, |Y|>X, V=X, U=Y atan2(Y,X) = sgnY * (pi/2 + atan(V/U))
85 // Region 4 (-180 to -135 degrees, and +135 to +180 degrees):
86 // X<0, |Y|<=X, V=Y, U=X atan2(Y,X) = sgnY * (pi - atan(V/U))
88 // So the result is always of the form atan2(Y,X) = P + sgnXY * atan(V/U)
90 // We compute atan(V/U) from the identity
91 // atan(z) + atan([(V/U)-z] / [1+(V/U)z])
92 // where z is a limited precision approximation (16 bits) to V/U
94 // z is calculated with the assistance of the frcpa instruction.
96 // atan(z) is calculated by a polynomial z + z^3 * p(w), w=z^2
97 // where p(w) = P0+P1*w+...+P22*w^22
99 // Let d = [(V/U)-z] / [1+(V/U)z]) = (V-U*z)/(U+V*z)
101 // Approximate atan(d) by d + P0*d^3
102 // Let F = 1/(U+V*z) * (1-a), where |a|< 2^-8.8.
103 // Compute q(a) = 1 + a + ... + a^5.
104 // Then F*q(a) approximates the reciprocal to more than 50 bits.
107 //==============================================================
114 // +inf +number +pi/2
115 // -inf +number -pi/2
116 // +inf -number +pi/2
117 // -inf -number -pi/2
144 // Nan anything quiet Y
145 // Not NaN NaN quiet X
147 // atan2(+-0/+-0) sets double error tag to 37
150 //==============================================================
152 // predicate registers used:
155 // floating-point registers used:
159 // general registers used
163 //==============================================================
176 GR_Parameter_RESULT = r40
228 atan2_sgn_pi_by_2 = f69
265 atan2_alpha_sq = f100
271 atan2_alpha_cub = f105
285 atan2_sig_near_one = f116
286 atan2_near_one = f116
289 atan2_3pi_by_4 = f118
293 /////////////////////////////////////////////////////////////
300 LOCAL_OBJECT_START(atan2_tb1)
301 data8 0xA21922DC45605EA1 , 0x00003FFA // P11
302 data8 0xB199DD6D2675C40F , 0x0000BFFA // P10
303 data8 0xC2F01E5DDD100DBE , 0x00003FFA // P9
304 data8 0xD78F28FC2A592781 , 0x0000BFFA // P8
305 data8 0xF0F03ADB3FC930D3 , 0x00003FFA // P7
306 data8 0x88887EBB209E3543 , 0x0000BFFB // P6
307 data8 0x9D89D7D55C3287A5 , 0x00003FFB // P5
308 data8 0xBA2E8B9793955C77 , 0x0000BFFB // P4
309 data8 0xE38E38E320A8A098 , 0x00003FFB // P3
310 data8 0x9249249247E37913 , 0x0000BFFC // P2
311 data8 0xCCCCCCCCCCC906CD , 0x00003FFC // P1
312 data8 0xAAAAAAAAAAAAA8A9 , 0x0000BFFD // P0
313 data8 0xC90FDAA22168C235 , 0x00004000 // pi
314 LOCAL_OBJECT_END(atan2_tb1)
316 LOCAL_OBJECT_START(atan2_tb2)
317 data8 0xCE585A259BD8374C , 0x00003FF0 // P21
318 data8 0x9F90FB984D8E39D0 , 0x0000BFF3 // P20
319 data8 0x9D3436AABE218776 , 0x00003FF5 // P19
320 data8 0xDEC343E068A6D2A8 , 0x0000BFF6 // P18
321 data8 0xF396268151CFB11C , 0x00003FF7 // P17
322 data8 0xD818B4BB43D84BF2 , 0x0000BFF8 // P16
323 data8 0xA2270D30A90AA220 , 0x00003FF9 // P15
324 data8 0xD5F4F2182E7A8725 , 0x0000BFF9 // P14
325 data8 0x80D601879218B53A , 0x00003FFA // P13
326 data8 0x9297B23CCFFB291F , 0x0000BFFA // P12
327 data8 0xFE7E52D2A89995B3 , 0x0000BFEC // P22
328 data8 0xC90FDAA22168C235 , 0x00003FFF // pi/2
329 data8 0xC90FDAA22168C235 , 0x00003FFE // pi/4
330 data8 0x96cbe3f9990e91a8 , 0x00004000 // 3pi/4
331 LOCAL_OBJECT_END(atan2_tb2)
337 GLOBAL_IEEE754_ENTRY(atan2)
340 alloc r32 = ar.pfs,1,5,4,0
341 frcpa.s1 atan2_u1_X,p6 = f1,atan2_X
345 addl EXP_AD_P1 = @ltoff(atan2_tb1), gp
346 fma.s1 atan2_two = f1,f1,f1
352 ld8 EXP_AD_P1 = [EXP_AD_P1]
353 frcpa.s1 atan2_u1_Y,p7 = f1,atan2_Y
358 fma.s1 atan2_xsq = atan2_X,atan2_X,f0
365 fclass.m p10,p0 = atan2_Y, 0xc3 // Test for y=nan
370 fma.s1 atan2_ysq = atan2_Y,atan2_Y,f0
376 add EXP_AD_P2 = 0xd0,EXP_AD_P1
377 fclass.m p12,p0 = atan2_X, 0xc3 // Test for x nan
383 // p10 Y NAN, quiet and return
385 ldfe atan2_P11 = [EXP_AD_P1],16
386 fmerge.s atan2_sgnY = atan2_Y,f1
390 ldfe atan2_P21 = [EXP_AD_P2],16
391 (p10) fma.d.s0 f8 = atan2_X,atan2_Y,f0 // If y=nan, result quietized y
392 (p10) br.ret.spnt b0 // Exit if y=nan
398 ldfe atan2_P10 = [EXP_AD_P1],16
399 fma.s1 atan2_z1_X = atan2_u1_X, atan2_Y, f0
403 ldfe atan2_P20 = [EXP_AD_P2],16
404 fnma.s1 atan2_B1X = atan2_u1_X, atan2_X, atan2_two
410 ldfe atan2_P9 = [EXP_AD_P1],16
411 fma.s1 atan2_z1_Y = atan2_u1_Y, atan2_X, f0
415 ldfe atan2_P19 = [EXP_AD_P2],16
416 fnma.s1 atan2_B1Y = atan2_u1_Y, atan2_Y, atan2_two
422 ldfe atan2_P8 = [EXP_AD_P1],16
423 fma.s1 atan2_z2_X = atan2_u1_X, atan2_ysq, f0
427 ldfe atan2_P18 = [EXP_AD_P2],16
428 fma.s1 atan2_z2_Y = atan2_u1_Y, atan2_xsq, f0
434 // p11 ==> x !inf y ?
436 ldfe atan2_P7 = [EXP_AD_P1],16
437 fclass.m p10,p11 = atan2_X, 0x23 // test for x inf
441 ldfe atan2_P17 = [EXP_AD_P2],16
442 (p12) fma.d.s0 f8 = atan2_X,atan2_Y,f0 // If x nan, result quiet x
443 (p12) br.ret.spnt b0 // Exit for x nan
447 // p6 true if swap, means |y| > |x| or ysq > xsq
448 // p7 true if no swap, means |x| >= |y| or xsq >= ysq
450 ldfe atan2_P6 = [EXP_AD_P1],16
451 ldfe atan2_P16 = [EXP_AD_P2],16
452 fcmp.ge.s1 p7,p6 = atan2_xsq, atan2_ysq
457 ldfe atan2_P5 = [EXP_AD_P1],16
458 fma.s1 atan2_wp_X = atan2_z1_X, atan2_z1_X, f0
462 ldfe atan2_P15 = [EXP_AD_P2],16
463 fma.s1 atan2_B1sq_X = atan2_B1X, atan2_B1X, f0
469 ldfe atan2_P4 = [EXP_AD_P1],16
470 (p6) fma.s1 atan2_wp_Y = atan2_z1_Y, atan2_z1_Y, f0
474 ldfe atan2_P14 = [EXP_AD_P2],16
475 (p6) fma.s1 atan2_B1sq_Y = atan2_B1Y, atan2_B1Y, f0
481 ldfe atan2_P3 = [EXP_AD_P1],16
482 (p6) fma.s1 atan2_E = atan2_z2_Y, atan2_B1Y, atan2_Y
486 ldfe atan2_P13 = [EXP_AD_P2],16
487 (p7) fma.s1 atan2_E = atan2_z2_X, atan2_B1X, atan2_X
494 ldfe atan2_P2 = [EXP_AD_P1],16
495 (p6) fma.s1 atan2_z = atan2_z1_Y, atan2_B1Y, f0
499 ldfe atan2_P12 = [EXP_AD_P2],16
500 (p7) fma.s1 atan2_z = atan2_z1_X, atan2_B1X, f0
507 ldfe atan2_P1 = [EXP_AD_P1],16
508 fcmp.eq.s0 p14,p15=atan2_X,atan2_Y // Dummy for denorm and invalid
512 ldfe atan2_P22 = [EXP_AD_P2],16
513 movl rsig_near_one = 0x8000000000000001 // signif near 1.0
518 // p12 ==> x inf y inf
519 // p13 ==> x inf y !inf
521 ldfe atan2_P0 = [EXP_AD_P1],16
522 ldfe atan2_pi_by_2 = [EXP_AD_P2],16
523 (p10) fclass.m.unc p12,p13 = atan2_Y, 0x23 // x inf, test if y inf
528 ldfe atan2_pi = [EXP_AD_P1],16
529 (p6) fma.s1 atan2_w = atan2_wp_Y, atan2_B1sq_Y,f0
533 ldfe atan2_pi_by_4 = [EXP_AD_P2],16
534 (p7) fma.s1 atan2_w = atan2_wp_X, atan2_B1sq_X,f0
540 ldfe atan2_3pi_by_4 = [EXP_AD_P2],16
541 (p11) fclass.m.unc p9,p0 = atan2_Y, 0x23 // x not inf, test if y inf
547 setf.sig atan2_sig_near_one = rsig_near_one
548 (p12) fcmp.gt.unc.s1 p10,p11 = atan2_X,f0 // x inf, y inf, test if x +inf
553 (p6) fnma.s1 atan2_gV = atan2_Y, atan2_z, atan2_X
560 frcpa.s1 atan2_F,p0 = f1, atan2_E
565 (p7) fnma.s1 atan2_gV = atan2_X, atan2_z, atan2_Y
570 // p13 ==> x inf y !inf
573 (p13) fcmp.gt.unc.s1 p14,p15 = atan2_X,f0 // x inf, y !inf, test if x +inf
578 (p9) fma.d.s0 f8 = atan2_sgnY, atan2_pi_by_2, f0 // +-pi/2 if x !inf, y inf
579 (p9) br.ret.spnt b0 // exit if x not inf, y inf, result is +-pi/2
585 fma.s1 atan2_V13 = atan2_w, atan2_P11, atan2_P10
590 fma.s1 atan2_W11 = atan2_w, atan2_P21, atan2_P20
597 fma.s1 atan2_V11 = atan2_w, atan2_P9, atan2_P8
602 fma.s1 atan2_V12 = atan2_w, atan2_w, f0
609 fma.s1 atan2_V8 = atan2_w, atan2_P7 , atan2_P6
614 fma.s1 atan2_W8 = atan2_w, atan2_P19, atan2_P18
621 fnma.s1 atan2_alpha = atan2_E, atan2_F, f1
626 fnma.s1 atan2_alpha_1 = atan2_E, atan2_F, atan2_two
634 fma.s1 atan2_V7 = atan2_w, atan2_P5 , atan2_P4
639 fma.s1 atan2_W7 = atan2_w, atan2_P17, atan2_P16
646 fma.s1 atan2_V4 = atan2_w, atan2_P3 , atan2_P2
651 fma.s1 atan2_W4 = atan2_w, atan2_P15, atan2_P14
658 fma.s1 atan2_V3 = atan2_w, atan2_P1 , atan2_P0
663 fma.s1 atan2_W3 = atan2_w, atan2_P13, atan2_P12
670 fma.s1 atan2_V10 = atan2_V12, atan2_V13, atan2_V11
675 fma.s1 atan2_gVF = atan2_gV, atan2_F, f0
682 fma.s1 atan2_alpha_sq = atan2_alpha, atan2_alpha, f0
687 fma.s1 atan2_Cp = atan2_alpha, atan2_alpha_1, f1
694 fma.s1 atan2_V9 = atan2_V12, atan2_V12, f0
699 fma.s1 atan2_W10 = atan2_V12, atan2_P22 , atan2_W11
706 fma.s1 atan2_V6 = atan2_V12, atan2_V8 , atan2_V7
711 fma.s1 atan2_W6 = atan2_V12, atan2_W8 , atan2_W7
718 fma.s1 atan2_V2 = atan2_V12, atan2_V4 , atan2_V3
723 fma.s1 atan2_W2 = atan2_V12, atan2_W4 , atan2_W3
732 fclass.m p8,p9 = atan2_Y, 0x07 // Test for y=0
737 fma.s1 atan2_zcub = atan2_z, atan2_w, f0
744 fma.s1 atan2_alpha_cub = atan2_alpha, atan2_alpha_sq, f0
749 fma.s1 atan2_C = atan2_gVF, atan2_Cp, f0
758 (p8) fclass.m.unc p12,p13 = atan2_X, 0x07 // y=0, test if x is 0
763 fma.s1 atan2_W12 = atan2_V9, atan2_V9, f0
770 fma.s1 atan2_V5 = atan2_V9, atan2_V10, atan2_V6
775 fma.s1 atan2_W5 = atan2_V9, atan2_W10, atan2_W6
784 (p9) fclass.m.unc p9,p0 = atan2_X, 0x07 // y not 0, test if x is 0
787 // p10 ==> X +INF, Y +-INF
790 (p10) fma.d.s0 f8 = atan2_sgnY, atan2_pi_by_4, f0 // x=+inf, y=inf
791 (p10) br.ret.spnt b0 // Exit for x=+inf, y=inf, result is +-pi/4
795 .pred.rel "mutex",p11,p14
798 (p14) fmerge.s f8 = atan2_sgnY, f0 // x=+inf, y !inf, result +-0
801 // p11 ==> X -INF, Y +-INF
804 (p11) fma.d.s0 f8 = atan2_sgnY, atan2_3pi_by_4, f0 // x=-inf, y=inf
805 (p11) br.ret.spnt b0 // Exit for x=-inf, y=inf, result is +-3pi/4
811 (p13) fcmp.gt.unc.s1 p10,p11 = atan2_X,f0 // x not 0, y=0, test if x>0
816 fma.s1 atan2_d = atan2_alpha_cub, atan2_C, atan2_C
817 (p14) br.ret.spnt b0 // Exit if x=+inf, y !inf, result +-0
823 fma.s1 atan2_W12 = atan2_V9, atan2_W12, f0
828 (p9) fma.d.s0 f8 = atan2_sgnY, atan2_pi_by_2, f0 // x=0, y not 0
829 (p9) br.ret.spnt b0 // Exit if x=0 and y not 0, result is +-pi/2
835 fma.s1 atan2_V1 = atan2_V9, atan2_V5, atan2_V2
840 fma.s1 atan2_W1 = atan2_V9, atan2_W5, atan2_W2
841 (p12) br.spnt ATAN2_ERROR // Branch if x=0 and y=0
847 (p10) fmerge.s f8 = atan2_sgnY, f0 // +-0 if x>0, y=0
852 (p11) fma.d.s0 f8 = atan2_sgnY, atan2_pi, f0 // +-pi if x<0, y=0
853 (p13) br.ret.spnt b0 // Exit if x!0 and y=0
860 fma.s1 atan2_pd = atan2_P0, atan2_d, f0
865 fma.s1 atan2_dsq = atan2_d, atan2_d, f0
873 fmerge.se atan2_near_one = f1, atan2_sig_near_one // Const ~1.0
878 fma.s1 atan2_Pp = atan2_W12, atan2_W1, atan2_V1
883 // p8 true if no swap and X positive
884 // p9 true if no swap and X negative
885 // both are false is swap is true
888 (p7) fcmp.ge.unc.s1 p8,p9 = atan2_X,f0
893 (p15) fma.d.s0 f8 = atan2_sgnY, atan2_pi, f0
894 (p15) br.ret.spnt b0 // Exit if x=-inf, y !inf, result +-pi
900 fma.s1 atan2_sgn_pi_by_2 = atan2_pi_by_2, atan2_sgnY, f0
905 fma.s1 atan2_A_lo = atan2_pd, atan2_dsq, atan2_d
913 fma.s1 atan2_sgn_pi = atan2_pi, atan2_sgnY, f0
918 fma.s1 atan2_A_hi = atan2_zcub, atan2_Pp, atan2_z
924 // For |Y| <= |X| and X > 0, force inexact in case A_lo is zero
927 (p8) fmpy.s0 atan2_tmp = atan2_P22, atan2_P22
934 fma.s1 atan2_A = atan2_A_hi, f1, atan2_A_lo
937 // For |Y| <= |X| and X > 0, result is A_hi + A_lo
940 (p8) fma.d.s0 f8 = atan2_A_hi, f1, atan2_A_lo
945 .pred.rel "mutex",p6,p9
946 // We perturb A by multiplying by 1.0+1ulp as we produce the result
947 // in order to get symmetrically rounded results in directed rounding modes.
948 // If we don't do this, there are a few cases where the trailing 11 bits of
949 // the significand of the result, before converting to double, are zero. These
950 // cases do not round symmetrically in round to +infinity or round to -infinity.
951 // The perturbation also insures that the inexact flag is set.
952 // For |Y| > |X|, result is +- pi/2 - (A_hi + A_lo)
955 (p6) fnma.d.s0 f8 = atan2_A, atan2_near_one, atan2_sgn_pi_by_2
958 // For |Y| <= |X|, and X < 0, result is +- pi + (A_hi + A_lo)
961 (p9) fma.d.s0 f8 = atan2_A, atan2_near_one, atan2_sgn_pi
967 // Here if x=0 and y=0
970 fclass.m p10,p11 = atan2_X,0x05 // Test if x=+0
976 mov atan2_GR_tag = 37
977 (p10) fmerge.s f10 = atan2_sgnY, f0 // x=+0, y=0
982 (p11) fma.d.s0 f10 = atan2_sgnY, atan2_pi, f0 // x=-0, y=0
986 GLOBAL_IEEE754_END(atan2)
987 libm_alias_double_other (__atan2, atan2)
990 LOCAL_LIBM_ENTRY(__libm_error_region)
994 add GR_Parameter_Y=-32,sp // Parameter 2 value
996 .save ar.pfs,GR_SAVE_PFS
997 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
1001 add sp=-64,sp // Create new stack
1003 mov GR_SAVE_GP=gp // Save gp
1009 stfd [GR_Parameter_Y] = f8,16 // STORE Parameter 2 on stack
1010 add GR_Parameter_X = 16,sp // Parameter 1 address
1011 .save b0, GR_SAVE_B0
1012 mov GR_SAVE_B0=b0 // Save b0
1018 stfd [GR_Parameter_X] = f9 // STORE Parameter 1 on stack
1019 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
1023 stfd [GR_Parameter_Y] = f10 // STORE Parameter 3 on stack
1024 add GR_Parameter_Y = -16,GR_Parameter_Y
1025 br.call.sptk b0=__libm_error_support# // Call error handling function
1028 add GR_Parameter_RESULT = 48,sp
1035 ldfd f8 = [GR_Parameter_RESULT] // Get return result off stack
1037 add sp = 64,sp // Restore stack pointer
1038 mov b0 = GR_SAVE_B0 // Restore return address
1041 mov gp = GR_SAVE_GP // Restore gp
1042 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
1043 br.ret.sptk b0 // Return
1046 LOCAL_LIBM_END(__libm_error_region)
1048 .type __libm_error_support#,@function
1049 .global __libm_error_support#