4 // Copyright (c) 2000 - 2005, Intel Corporation
5 // All rights reserved.
7 // Contributed 2000 by the Intel Numerics Group, Intel Corporation
9 // Redistribution and use in source and binary forms, with or without
10 // modification, are permitted provided that the following conditions are
13 // * Redistributions of source code must retain the above copyright
14 // notice, this list of conditions and the following disclaimer.
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.
20 // * The name of Intel Corporation may not be used to endorse or promote
21 // products derived from this software without specific prior written
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.
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 //==============================================================
42 // 02/02/00 Initial version
43 // 04/04/00 Unwind support added
44 // 12/27/00 Improved speed
45 // 02/21/01 Updated to call tanl
46 // 05/30/02 Improved speed, added cotf.
47 // 11/25/02 Added explicit completer on fnorm
48 // 02/10/03 Reordered header: .section, .global, .proc, .align
49 // 04/17/03 Eliminated redundant stop bits
50 // 03/31/05 Reformatted delimiters between data tables
53 //==============================================================
57 // Algorithm Description for tanf
58 //==============================================================
59 // The tanf function computes the principle value of the tangent of x,
60 // where x is radian argument.
64 // Return tanf(x) = +/-0.0
67 // Return tanf(x) = QNaN
70 // Return tanf(x) = QNaN
72 // 4. x = r + (Pi/2)*N, N = RoundInt(x*(2/Pi)), N is even, |r|<Pi/4
73 // Return tanf(x) = P19(r) = A1*r + A3*r^3 + A5*r^5 + ... + A19*r^19 =
74 // = r*(A1 + A3*t + A5*t^2 + ... + A19*t^9) = r*P9(t), where t = r^2
76 // 5. x = r + (Pi/2)*N, N = RoundInt(x*(2/Pi)), N is odd, |r|<Pi/4
77 // Return tanf(x) = -1/r + P11(r) = -1/r + B1*r + B3*r^3 + ... + B11*r^11 =
78 // = -1/r + r*(B1 + B3*t + B5*t^2 + ... + B11*t^5) = -1/r + r*P11(t),
81 // Algorithm Description for cotf
82 //==============================================================
83 // The cotf function computes the principle value of the cotangent of x,
84 // where x is radian argument.
88 // Return cotf(x) = +/-Inf and error handling is called
91 // Return cotf(x) = QNaN
94 // Return cotf(x) = QNaN
96 // 4. x = r + (Pi/2)*N, N = RoundInt(x*(2/Pi)), N is odd, |r|<Pi/4
97 // Return cotf(x) = P19(-r) = A1*(-r) + A3*(-r^3) + ... + A19*(-r^19) =
98 // = -r*(A1 + A3*t + A5*t^2 + ... + A19*t^9) = -r*P9(t), where t = r^2
100 // 5. x = r + (Pi/2)*N, N = RoundInt(x*(2/Pi)), N is even, |r|<Pi/4
101 // Return cotf(x) = 1/r + P11(-r) = 1/r + B1*(-r) + ... + B11*(-r^11) =
102 // = 1/r - r*(B1 + B3*t + B5*t^2 + ... + B11*t^5) = 1/r - r*P11(t),
105 // We set p10 and clear p11 if computing tanf, vice versa for cotf.
109 //==============================================================
110 // Floating Point registers used:
114 // General registers used:
115 // r14 -> r23, r32 -> r39
117 // Predicate registers used:
121 //==============================================================
139 GR_Parameter_RESULT = r38
140 GR_Parameter_Tag = r39
142 //==============================================================
143 // floating point registers
193 //==============================================================
199 LOCAL_OBJECT_START(coeff_A)
200 data8 0x3FF0000000000000 // A1 = 1.00000000000000000000e+00
201 data8 0x3FD5555556BCE758 // A3 = 3.33333334641442641606e-01
202 data8 0x3FC111105C2DAE48 // A5 = 1.33333249100689099175e-01
203 data8 0x3FABA1F876341060 // A7 = 5.39701122561673229739e-02
204 data8 0x3F965FB86D12A38D // A9 = 2.18495194027670719750e-02
205 data8 0x3F8265F62415F9D6 // A11 = 8.98353860497717439465e-03
206 data8 0x3F69E3AE64CCF58D // A13 = 3.16032468108912746342e-03
207 data8 0x3F63920D09D0E6F6 // A15 = 2.38897844840557235331e-03
208 LOCAL_OBJECT_END(coeff_A)
210 LOCAL_OBJECT_START(coeff_B)
211 data8 0xC90FDAA22168C235, 0x3FFF // pi/2
212 data8 0x3FD55555555358DB // B1 = 3.33333333326107426583e-01
213 data8 0x3F96C16C252F643F // B3 = 2.22222230621336129239e-02
214 data8 0x3F61566243AB3C60 // B5 = 2.11638633968606896785e-03
215 data8 0x3F2BC1169BD4438B // B7 = 2.11748132564551094391e-04
216 data8 0x3EF611B4CEA056A1 // B9 = 2.10467959860990200942e-05
217 data8 0x3EC600F9E32194BF // B11 = 2.62305891234274186608e-06
218 data8 0xBF42BA7BCC177616 // A17 =-5.71546981685324877205e-04
219 data8 0x3F4F2614BC6D3BB8 // A19 = 9.50584530849832782542e-04
220 LOCAL_OBJECT_END(coeff_B)
225 LOCAL_LIBM_ENTRY(cotf)
228 getf.exp rExp = f8 // ***** Get 2ˆ17 * s + E
229 movl rSigRcpPiby2= 0xA2F9836E4E44152A // significand of 2/Pi
232 addl rCoeffA = @ltoff(coeff_A), gp
233 movl rScRshf = 0x47e8000000000000 // 1.5*2^(63+63+1)
238 alloc r32 = ar.pfs, 0, 4, 4, 0
239 fclass.m p9, p0 = f8, 0xc3 // Test for x=nan
240 cmp.eq p11, p10 = r0, r0 // if p11=1 we compute cotf
243 ld8 rCoeffA = [rCoeffA]
244 mov rExpCut = 0x10009 // cutoff for exponent
245 br.cond.sptk Common_Path
252 GLOBAL_IEEE754_ENTRY(tanf)
255 getf.exp rExp = f8 // ***** Get 2ˆ17 * s + E
256 movl rSigRcpPiby2= 0xA2F9836E4E44152A // significand of 2/Pi
259 addl rCoeffA = @ltoff(coeff_A), gp
260 movl rScRshf = 0x47e8000000000000 // 1.5*2^(63+63+1)
265 alloc r32 = ar.pfs, 0, 4, 4, 0
266 fclass.m p9, p0 = f8, 0xc3 // Test for x=nan
267 cmp.eq p10, p11 = r0, r0 // if p10=1 we compute tandf
270 ld8 rCoeffA = [rCoeffA]
271 mov rExpCut = 0x10009 // cutoff for exponent
276 // Below is common path for both tandf and cotdf
279 setf.sig fScRcpPiby2 = rSigRcpPiby2 // 2^(63+1)*(2/Pi)
280 fclass.m p8, p0 = f8, 0x23 // Test for x=inf
281 mov rSignMask = 0x1ffff // mask for sign bit
284 setf.d fScRshf = rScRshf // 1.5*2^(63+63+1)
285 movl rRshf = 0x43e8000000000000 // 1.5 2^63 for right shift
290 and rSignMask = rSignMask, rExp // clear sign bit
291 (p10) fclass.m.unc p7, p0 = f8, 0x07 // Test for x=0 (for tanf)
292 mov rScFctrExp = 0xffff-64 // exp of scaling factor
295 adds rCoeffB = coeff_B - coeff_A, rCoeffA
296 (p9) fma.s.s0 f8 = f8, f1, f8 // Set qnan if x=nan
297 (p9) br.ret.spnt b0 // Exit for x=nan
302 cmp.ge p6, p0 = rSignMask, rExpCut // p6 = (E => 0x10009)
303 (p8) frcpa.s0 f8, p0 = f0, f0 // Set qnan indef if x=inf
304 mov GR_Parameter_Tag = 227 // (cotf)
307 ldfe fPiby2 = [rCoeffB], 16
308 (p8) br.ret.spnt b0 // Exit for x=inf
309 (p6) br.cond.spnt Huge_Argument // Branch if |x|>=2^10
315 (p11) fclass.m.unc p6, p0 = f8, 0x07 // Test for x=0 (for cotf)
320 fnorm.s0 fNormArg = f8
321 (p7) br.ret.spnt b0 // Exit for x=0 (for tanf)
326 ldfpd fA01, fA03 = [rCoeffA], 16
327 ldfpd fB01, fB03 = [rCoeffB], 16
328 fmerge.s f10 = f8, f8 // Save input for error call
333 setf.exp fScFctr = rScFctrExp // get as real
334 setf.d fRshf = rRshf // get right shifter as real
335 (p6) frcpa.s0 f8, p0 = f1, f8 // cotf(+-0) = +-Inf
340 ldfpd fA05, fA07 = [rCoeffA], 16
341 ldfpd fB05, fB07 = [rCoeffB], 16
342 (p6) br.cond.spnt __libm_error_region // call error support if cotf(+-0)
347 ldfpd fA09, fA11 = [rCoeffA], 16
348 ldfpd fB09, fB11 = [rCoeffB], 16
355 fma.s1 fShiftedN = fNormArg,fScRcpPiby2,fScRshf // x*2^70*(2/Pi)+ScRshf
362 fms.s1 fN = fShiftedN, fScFctr, fRshf // N = Y*2^(-70) - Rshf
367 .pred.rel "mutex", p10, p11
369 getf.sig rIntN = fShiftedN // get N as integer
370 (p10) fnma.s1 fR = fN, fPiby2, fNormArg // R = x - (Pi/2)*N (tanf)
375 (p11) fms.s1 fR = fN, fPiby2, fNormArg // R = (Pi/2)*N - x (cotf)
381 ldfpd fA13, fA15 = [rCoeffA], 16
382 ldfpd fA17, fA19 = [rCoeffB], 16
390 fma.s1 fRp2 = fR, fR, f0 // R^2
391 (p11) add rIntN = 0x1, rIntN // N = N + 1 (cotf)
397 frcpa.s1 fY0, p0 = f1, fR // Y0 ~ 1/R
398 tbit.z p8, p9 = rIntN, 0 // p8=1 if N is even
402 // Below are mixed polynomial calculations (mixed for even and odd N)
405 (p9) fma.s1 fB03_01 = fRp2, fB03, fB01 // R^2*B3 + B1
410 fma.s1 fRp4 = fRp2, fRp2, f0 // R^4
417 (p8) fma.s1 fA15_13 = fRp2, fA15, fA13 // R^2*A15 + A13
422 (p8) fma.s1 fA19_17 = fRp2, fA19, fA17 // R^2*A19 + A17
429 (p8) fma.s1 fA07_05 = fRp2, fA07, fA05 // R^2*A7 + A5
434 (p8) fma.s1 fA11_09 = fRp2, fA11, fA09 // R^2*A11 + A9
441 (p9) fma.s1 fB07_05 = fRp2, fB07, fB05 // R^2*B7 + B5
446 (p9) fma.s1 fB11_09 = fRp2, fB11, fB09 // R^2*B11 + B9
453 (p9) fnma.s1 fD = fR, fY0, f1 // D = 1 - R*Y0
458 (p8) fma.s1 fA03_01 = fRp2, fA03, fA01 // R^2*A3 + A1
465 fma.s1 fRp8 = fRp4, fRp4, f0 // R^8
470 fma.s1 fRp5 = fR, fRp4, f0 // R^5
477 (p8) fma.s1 fA11_05 = fRp4, fA11_09, fA07_05 // R^4*(R^2*A11 + A9) + ...
482 (p8) fma.s1 fA19_13 = fRp4, fA19_17, fA15_13 // R^4*(R^2*A19 + A17) + ..
489 (p9) fma.s1 fB11_05 = fRp4, fB11_09, fB07_05 // R^4*(R^2*B11 + B9) + ...
494 (p9) fma.s1 fRbyB03_01 = fR, fB03_01, f0 // R*(R^2*B3 + B1)
501 (p9) fma.s1 fY1 = fY0, fD, fY0 // Y1 = Y0*D + Y0
506 (p9) fma.s1 fDp2 = fD, fD, f0 // D^2
513 // R^8*(R^6*A19 + R^4*A17 + R^2*A15 + A13) + R^6*A11 + R^4*A9 + R^2*A7 + A5
514 (p8) fma.d.s1 fA19_05 = fRp8, fA19_13, fA11_05
519 (p8) fma.d.s1 fRbyA03_01 = fR, fA03_01, f0 // R*(R^2*A3 + A1)
526 (p9) fma.d.s1 fInvR = fY1, fDp2, fY1 // 1/R = Y1*D^2 + Y1
531 // R^5*(R^6*B11 + R^4*B9 + R^2*B7 + B5) + R^3*B3 + R*B1
532 (p9) fma.d.s1 fRbyB11_01 = fRp5, fB11_05, fRbyB03_01
537 .pred.rel "mutex", p8, p9
540 // Result = R^5*(R^14*A19 + R^12*A17 + R^10*A15 + ...) + R^3*A3 + R*A1
541 (p8) fma.s.s0 f8 = fRp5, fA19_05, fRbyA03_01
546 // Result = -1/R + R^11*B11 + R^9*B9 + R^7*B7 + R^5*B5 + R^3*B3 + R*B1
547 (p9) fnma.s.s0 f8 = f1, fInvR, fRbyB11_01
548 br.ret.sptk b0 // exit for main path
552 GLOBAL_IEEE754_END(tanf)
555 LOCAL_LIBM_ENTRY(__libm_callout)
562 .save ar.pfs,GR_SAVE_PFS
563 mov GR_SAVE_PFS=ar.pfs
578 (p10) br.cond.sptk.many call_tanl ;;
581 // Here if we should call cotl (p10=0, p11=1)
585 br.call.sptk.many b0=__libm_cotl# ;;
597 mov ar.pfs = GR_SAVE_PFS
602 // Here if we should call tanl (p10=1, p11=0)
607 br.call.sptk.many b0=__libm_tanl# ;;
619 mov ar.pfs = GR_SAVE_PFS
624 LOCAL_LIBM_END(__libm_callout)
626 .type __libm_tanl#,@function
628 .type __libm_cotl#,@function
632 LOCAL_LIBM_ENTRY(__libm_error_region)
637 add GR_Parameter_Y=-32,sp // Parameter 2 value
639 .save ar.pfs,GR_SAVE_PFS
640 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
644 add sp=-64,sp // Create new stack
646 mov GR_SAVE_GP=gp // Save gp
651 stfs [GR_Parameter_Y] = f1,16 // STORE Parameter 2 on stack
652 add GR_Parameter_X = 16,sp // Parameter 1 address
654 mov GR_SAVE_B0=b0 // Save b0
660 stfs [GR_Parameter_X] = f10 // STORE Parameter 1 on stack
661 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
665 stfs [GR_Parameter_Y] = f8 // STORE Parameter 3 on stack
666 add GR_Parameter_Y = -16,GR_Parameter_Y
667 br.call.sptk b0=__libm_error_support# // Call error handling function
672 add GR_Parameter_RESULT = 48,sp
677 ldfs f8 = [GR_Parameter_RESULT] // Get return result off stack
679 add sp = 64,sp // Restore stack pointer
680 mov b0 = GR_SAVE_B0 // Restore return address
683 mov gp = GR_SAVE_GP // Restore gp
684 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
685 br.ret.sptk b0 // Return
688 LOCAL_LIBM_END(__libm_error_region)
690 .type __libm_error_support#,@function
691 .global __libm_error_support#