3 // Copyright (C) 2000, 2001, Intel Corporation
4 // All rights reserved.
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.
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://developer.intel.com/opensource.
41 //==============================================================
42 // 2/02/00 Initial version
43 // 4/04/00 Unwind support added
44 // 8/15/00 Bundle added after call to __libm_error_support to properly
45 // set [the previously overwritten] GR_Parameter_RESULT.
46 // 10/12/00 Update to set denormal operand and underflow flags
47 // 1/22/01 Fixed to set inexact flag for small args.
50 //==============================================================
51 // float = sinhf(float)
52 // input floating point f8
53 // output floating point f8
56 //==============================================================
59 // predicate registers used:
61 // floating-point registers used:
62 // f9 -> f15; f32 -> f45;
63 // f8 has input, then output
65 // Overview of operation
66 //==============================================================
67 // There are four paths
68 // 1. |x| < 0.25 SINH_BY_POLY
69 // 2. |x| < 32 SINH_BY_TBL
70 // 3. |x| < 2^14 SINH_BY_EXP
71 // 4. |x_ >= 2^14 SINH_HUGE
73 // For double extended we get infinity for x >= 400c b174 ddc0 31ae c0ea
74 // >= 1.0110001.... x 2^13
77 // But for double we get infinity for x >= 408633ce8fb9f87e
81 // And for single we get infinity for x >= 42b3a496
85 // SAFE: If there is danger of overflow set SAFE to 0
86 // NOT implemented: if there is danger of underflow, set SAFE to 0
87 // SAFE for all paths listed below
91 // If |x| is less than the tiny threshold, then clear SAFE
92 // For double, the tiny threshold is -1022 = -0x3fe => -3fe + ffff = fc01
93 // register-biased, this is fc01
94 // For single, the tiny threshold is -126 = -7e => -7e + ffff = ff81
95 // If |x| < tiny threshold, set SAFE = 0
99 // SAFE: SAFE is always 1 for TBL;
103 // There is a danger of double-extended overflow if N-1 > 16382 = 0x3ffe
104 // r34 has N-1; 16382 is in register biased form, 0x13ffd
105 // There is danger of double overflow if N-1 > 0x3fe
106 // in register biased form, 0x103fd
107 // Analagously, there is danger of single overflow if N-1 > 0x7e
108 // in register biased form, 0x1007d
109 // SAFE: If there is danger of overflow set SAFE to 0
113 // SAFE: SAFE is always 0 for HUGE
116 #include "libm_support.h"
119 //==============================================================
124 sinh_FR_all_ones = f45
127 sinh_FR_Inv_log2by64 = f9
128 sinh_FR_log2by64_lo = f11
129 sinh_FR_log2by64_hi = f10
147 sinh_FR_peven_temp1 = f34
148 sinh_FR_peven_temp2 = f35
151 sinh_FR_podd_temp1 = f34
152 sinh_FR_podd_temp2 = f35
155 sinh_FR_poly_podd_temp1 = f11
156 sinh_FR_poly_podd_temp2 = f13
157 sinh_FR_poly_peven_temp1 = f11
158 sinh_FR_poly_peven_temp2 = f13
165 sinh_FR_N_temp1 = f11
166 sinh_FR_N_temp2 = f12
182 sinh_GR_all_ones = r40
185 sinh_FR_S_hi_temp = f10
186 sinh_FR_S_lo_temp1 = f11
187 sinh_FR_S_lo_temp2 = f12
188 sinh_FR_S_lo_temp3 = f13
193 sinh_FR_C_hi_temp1 = f10
195 sinh_FR_Y_lo_temp = f12
206 sinh_FR_TINY_THRESH = f9
208 sinh_FR_SINH_temp = f10
211 sinh_FR_signed_hi_lo = f10
220 GR_Parameter_RESULT = r46
223 //==============================================================
232 double_sinh_arg_reduction:
233 ASM_TYPE_DIRECTIVE(double_sinh_arg_reduction,@object)
234 data8 0xB8AA3B295C17F0BC, 0x00004005
235 data8 0xB17217F7D1000000, 0x00003FF8
236 data8 0xCF79ABC9E3B39804, 0x00003FD0
237 ASM_SIZE_DIRECTIVE(double_sinh_arg_reduction)
240 ASM_TYPE_DIRECTIVE(double_sinh_p_table,@object)
241 data8 0xAAAAAAAAAAAAAAAB, 0x00003FFC
242 data8 0x8888888888888412, 0x00003FF8
243 data8 0xD00D00D00D4D39F2, 0x00003FF2
244 data8 0xB8EF1D28926D8891, 0x00003FEC
245 data8 0xD732377688025BE9, 0x00003FE5
246 data8 0xB08AF9AE78C1239F, 0x00003FDE
247 ASM_SIZE_DIRECTIVE(double_sinh_p_table)
249 double_sinh_ab_table:
250 ASM_TYPE_DIRECTIVE(double_sinh_ab_table,@object)
251 data8 0xAAAAAAAAAAAAAAAC, 0x00003FFC
252 data8 0x88888888884ECDD5, 0x00003FF8
253 data8 0xD00D0C6DCC26A86B, 0x00003FF2
254 data8 0x8000000000000002, 0x00003FFE
255 data8 0xAAAAAAAAAA402C77, 0x00003FFA
256 data8 0xB60B6CC96BDB144D, 0x00003FF5
257 ASM_SIZE_DIRECTIVE(double_sinh_ab_table)
260 ASM_TYPE_DIRECTIVE(double_sinh_j_table,@object)
261 data8 0xB504F333F9DE6484, 0x00003FFE, 0x1EB2FB13, 0x00000000
262 data8 0xB6FD91E328D17791, 0x00003FFE, 0x1CE2CBE2, 0x00000000
263 data8 0xB8FBAF4762FB9EE9, 0x00003FFE, 0x1DDC3CBC, 0x00000000
264 data8 0xBAFF5AB2133E45FB, 0x00003FFE, 0x1EE9AA34, 0x00000000
265 data8 0xBD08A39F580C36BF, 0x00003FFE, 0x9EAEFDC1, 0x00000000
266 data8 0xBF1799B67A731083, 0x00003FFE, 0x9DBF517B, 0x00000000
267 data8 0xC12C4CCA66709456, 0x00003FFE, 0x1EF88AFB, 0x00000000
268 data8 0xC346CCDA24976407, 0x00003FFE, 0x1E03B216, 0x00000000
269 data8 0xC5672A115506DADD, 0x00003FFE, 0x1E78AB43, 0x00000000
270 data8 0xC78D74C8ABB9B15D, 0x00003FFE, 0x9E7B1747, 0x00000000
271 data8 0xC9B9BD866E2F27A3, 0x00003FFE, 0x9EFE3C0E, 0x00000000
272 data8 0xCBEC14FEF2727C5D, 0x00003FFE, 0x9D36F837, 0x00000000
273 data8 0xCE248C151F8480E4, 0x00003FFE, 0x9DEE53E4, 0x00000000
274 data8 0xD06333DAEF2B2595, 0x00003FFE, 0x9E24AE8E, 0x00000000
275 data8 0xD2A81D91F12AE45A, 0x00003FFE, 0x1D912473, 0x00000000
276 data8 0xD4F35AABCFEDFA1F, 0x00003FFE, 0x1EB243BE, 0x00000000
277 data8 0xD744FCCAD69D6AF4, 0x00003FFE, 0x1E669A2F, 0x00000000
278 data8 0xD99D15C278AFD7B6, 0x00003FFE, 0x9BBC610A, 0x00000000
279 data8 0xDBFBB797DAF23755, 0x00003FFE, 0x1E761035, 0x00000000
280 data8 0xDE60F4825E0E9124, 0x00003FFE, 0x9E0BE175, 0x00000000
281 data8 0xE0CCDEEC2A94E111, 0x00003FFE, 0x1CCB12A1, 0x00000000
282 data8 0xE33F8972BE8A5A51, 0x00003FFE, 0x1D1BFE90, 0x00000000
283 data8 0xE5B906E77C8348A8, 0x00003FFE, 0x1DF2F47A, 0x00000000
284 data8 0xE8396A503C4BDC68, 0x00003FFE, 0x1EF22F22, 0x00000000
285 data8 0xEAC0C6E7DD24392F, 0x00003FFE, 0x9E3F4A29, 0x00000000
286 data8 0xED4F301ED9942B84, 0x00003FFE, 0x1EC01A5B, 0x00000000
287 data8 0xEFE4B99BDCDAF5CB, 0x00003FFE, 0x1E8CAC3A, 0x00000000
288 data8 0xF281773C59FFB13A, 0x00003FFE, 0x9DBB3FAB, 0x00000000
289 data8 0xF5257D152486CC2C, 0x00003FFE, 0x1EF73A19, 0x00000000
290 data8 0xF7D0DF730AD13BB9, 0x00003FFE, 0x9BB795B5, 0x00000000
291 data8 0xFA83B2DB722A033A, 0x00003FFE, 0x1EF84B76, 0x00000000
292 data8 0xFD3E0C0CF486C175, 0x00003FFE, 0x9EF5818B, 0x00000000
293 data8 0x8000000000000000, 0x00003FFF, 0x00000000, 0x00000000
294 data8 0x8164D1F3BC030773, 0x00003FFF, 0x1F77CACA, 0x00000000
295 data8 0x82CD8698AC2BA1D7, 0x00003FFF, 0x1EF8A91D, 0x00000000
296 data8 0x843A28C3ACDE4046, 0x00003FFF, 0x1E57C976, 0x00000000
297 data8 0x85AAC367CC487B15, 0x00003FFF, 0x9EE8DA92, 0x00000000
298 data8 0x871F61969E8D1010, 0x00003FFF, 0x1EE85C9F, 0x00000000
299 data8 0x88980E8092DA8527, 0x00003FFF, 0x1F3BF1AF, 0x00000000
300 data8 0x8A14D575496EFD9A, 0x00003FFF, 0x1D80CA1E, 0x00000000
301 data8 0x8B95C1E3EA8BD6E7, 0x00003FFF, 0x9D0373AF, 0x00000000
302 data8 0x8D1ADF5B7E5BA9E6, 0x00003FFF, 0x9F167097, 0x00000000
303 data8 0x8EA4398B45CD53C0, 0x00003FFF, 0x1EB70051, 0x00000000
304 data8 0x9031DC431466B1DC, 0x00003FFF, 0x1F6EB029, 0x00000000
305 data8 0x91C3D373AB11C336, 0x00003FFF, 0x1DFD6D8E, 0x00000000
306 data8 0x935A2B2F13E6E92C, 0x00003FFF, 0x9EB319B0, 0x00000000
307 data8 0x94F4EFA8FEF70961, 0x00003FFF, 0x1EBA2BEB, 0x00000000
308 data8 0x96942D3720185A00, 0x00003FFF, 0x1F11D537, 0x00000000
309 data8 0x9837F0518DB8A96F, 0x00003FFF, 0x1F0D5A46, 0x00000000
310 data8 0x99E0459320B7FA65, 0x00003FFF, 0x9E5E7BCA, 0x00000000
311 data8 0x9B8D39B9D54E5539, 0x00003FFF, 0x9F3AAFD1, 0x00000000
312 data8 0x9D3ED9A72CFFB751, 0x00003FFF, 0x9E86DACC, 0x00000000
313 data8 0x9EF5326091A111AE, 0x00003FFF, 0x9F3EDDC2, 0x00000000
314 data8 0xA0B0510FB9714FC2, 0x00003FFF, 0x1E496E3D, 0x00000000
315 data8 0xA27043030C496819, 0x00003FFF, 0x9F490BF6, 0x00000000
316 data8 0xA43515AE09E6809E, 0x00003FFF, 0x1DD1DB48, 0x00000000
317 data8 0xA5FED6A9B15138EA, 0x00003FFF, 0x1E65EBFB, 0x00000000
318 data8 0xA7CD93B4E965356A, 0x00003FFF, 0x9F427496, 0x00000000
319 data8 0xA9A15AB4EA7C0EF8, 0x00003FFF, 0x1F283C4A, 0x00000000
320 data8 0xAB7A39B5A93ED337, 0x00003FFF, 0x1F4B0047, 0x00000000
321 data8 0xAD583EEA42A14AC6, 0x00003FFF, 0x1F130152, 0x00000000
322 data8 0xAF3B78AD690A4375, 0x00003FFF, 0x9E8367C0, 0x00000000
323 data8 0xB123F581D2AC2590, 0x00003FFF, 0x9F705F90, 0x00000000
324 data8 0xB311C412A9112489, 0x00003FFF, 0x1EFB3C53, 0x00000000
325 data8 0xB504F333F9DE6484, 0x00003FFF, 0x1F32FB13, 0x00000000
326 ASM_SIZE_DIRECTIVE(double_sinh_j_table)
337 .global __ieee754_sinhf
338 .type __ieee754_sinhf,@function
342 // X infinity or NAN?
343 // Take invalid fault if enabled
347 alloc r32 = ar.pfs,0,12,4,0
348 (p0) fclass.m.unc p6,p0 = f8, 0xe3 //@qnan | @snan | @inf
349 mov sinh_GR_all_ones = -1
356 (p6) fma.s.s0 f8 = f8,f1,f8
357 (p6) br.ret.spnt b0 ;;
360 // Put 0.25 in f9; p6 true if x < 0.25
361 // Make constant that will generate inexact when squared
363 setf.sig sinh_FR_all_ones = sinh_GR_all_ones
364 (p0) movl r32 = 0x000000000000fffd ;;
368 (p0) setf.exp f9 = r32
369 (p0) fclass.m.unc p7,p0 = f8, 0x07 //@zero
375 (p0) fmerge.s sinh_FR_X = f0,f8
376 (p7) br.ret.spnt b0 ;;
379 // Identify denormal operands.
382 fclass.m.unc p10,p0 = f8, 0x09 // + denorm
387 fclass.m.unc p11,p0 = f8, 0x0a // - denorm
393 (p0) fmerge.s sinh_FR_SGNX = f8,f1
399 (p0) fcmp.lt.unc.s1 p0,p7 = sinh_FR_X,f9
406 (p7) br.cond.sptk L(SINH_BY_TBL) ;;
412 // POLY cannot overflow so there is no need to call __libm_error_support
413 // Set tiny_SAFE (p7) to 1(0) if answer is not tiny
414 // Currently we do not use tiny_SAFE. So the setting of tiny_SAFE is
416 //(p0) movl r32 = 0x000000000000fc01
417 //(p0) setf.exp f10 = r32
418 //(p0) fcmp.lt.unc.s1 p6,p7 = f8,f10
419 // Here is essentially the algorithm for SINH_BY_POLY. Care is take for the order
420 // of multiplication; and P_1 is not exactly 1/3!, P_2 is not exactly 1/5!, etc.
421 // Note that ax = |x|
422 // sinh(x) = sign * (series(e^x) - series(e^-x))/2
423 // = sign * (ax + ax^3/3! + ax^5/5! + ax^7/7! + ax^9/9! + ax^11/11! + ax^13/13!)
424 // = sign * (ax + ax * ( ax^2 * (1/3! + ax^4 * (1/7! + ax^4*1/11!)) )
425 // + ax * ( ax^4 * (1/5! + ax^4 * (1/9! + ax^4*1/13!)) ) )
426 // = sign * (ax + ax*p_odd + (ax*p_even))
427 // = sign * (ax + Y_lo)
428 // sinh(x) = sign * (Y_hi + Y_lo)
429 // Get the values of P_x from the table
431 (p0) addl r34 = @ltoff(double_sinh_p_table), gp
432 (p10) fma.s.s0 f8 = f8,f8,f8
439 (p11) fnma.s.s0 f8 = f8,f8,f8
444 // Calculate sinh_FR_X2 = ax*ax and sinh_FR_X4 = ax*ax*ax*ax
447 (p0) ldfe sinh_FR_P1 = [r34],16
448 (p0) fma.s1 sinh_FR_X2 = sinh_FR_X, sinh_FR_X, f0 ;;
452 (p0) ldfe sinh_FR_P2 = [r34],16 ;;
453 (p0) ldfe sinh_FR_P3 = [r34],16
458 (p0) ldfe sinh_FR_P4 = [r34],16 ;;
459 (p0) ldfe sinh_FR_P5 = [r34],16
464 (p0) ldfe sinh_FR_P6 = [r34],16
465 (p0) fma.s1 sinh_FR_X4 = sinh_FR_X2, sinh_FR_X2, f0
469 // Calculate sinh_FR_podd = p_odd and sinh_FR_peven = p_even
472 (p0) fma.s1 sinh_FR_poly_podd_temp1 = sinh_FR_X4, sinh_FR_P5, sinh_FR_P3
478 (p0) fma.s1 sinh_FR_poly_podd_temp2 = sinh_FR_X4, sinh_FR_poly_podd_temp1, sinh_FR_P1
484 (p0) fma.s1 sinh_FR_poly_peven_temp1 = sinh_FR_X4, sinh_FR_P6, sinh_FR_P4
490 (p0) fma.s1 sinh_FR_podd = sinh_FR_X2, sinh_FR_poly_podd_temp2, f0
496 (p0) fma.s1 sinh_FR_poly_peven_temp2 = sinh_FR_X4, sinh_FR_poly_peven_temp1, sinh_FR_P2
502 (p0) fma.s1 sinh_FR_peven = sinh_FR_X4, sinh_FR_poly_peven_temp2, f0
506 // Calculate sinh_FR_Y_lo = ax*p_odd + (ax*p_even)
509 (p0) fma.s1 sinh_FR_Y_lo_temp = sinh_FR_X, sinh_FR_peven, f0
515 (p0) fma.s1 sinh_FR_Y_lo = sinh_FR_X, sinh_FR_podd, sinh_FR_Y_lo_temp
519 // Calculate sinh_FR_SINH = Y_hi + Y_lo. Note that ax = Y_hi
522 (p0) fma.s1 sinh_FR_SINH = sinh_FR_X, f1, sinh_FR_Y_lo
525 // Dummy multiply to generate inexact
528 (p0) fmpy.s0 sinh_FR_tmp = sinh_FR_all_ones, sinh_FR_all_ones
532 // Calculate f8 = sign * (Y_hi + Y_lo)
536 (p0) fma.s.s0 f8 = sinh_FR_SGNX,sinh_FR_SINH,f0
537 (p0) br.ret.sptk b0 ;;
543 // Now that we are at TBL; so far all we know is that |x| >= 0.25.
544 // The first two steps are the same for TBL and EXP, but if we are HUGE
545 // we want to leave now.
547 // Go to HUGE if |x| >= 2^14, 1000d (register-biased) is e = 14 (true)
549 // Go to HUGE if |x| >= 2^10, 10009 (register-biased) is e = 10 (true)
551 // Go to HUGE if |x| >= 2^7, 10006 (register-biased) is e = 7 (true)
555 (p0) movl r32 = 0x0000000000010006 ;;
559 (p0) setf.exp f9 = r32
566 (p0) fcmp.ge.unc.s1 p6,p7 = sinh_FR_X,f9
573 (p6) br.cond.spnt L(SINH_HUGE) ;;
582 // TBL can never overflow
583 // sinh(x) = sinh(B+R)
584 // = sinh(B)cosh(R) + cosh(B)sinh(R)
586 // ax = |x| = M*log2/64 + R
589 // We will calcualte M and get N as (M-j)/64
590 // The division is a shift.
591 // exp(B) = exp(N*log2 + j*log2/64)
592 // = 2^N * 2^(j*log2/64)
593 // sinh(B) = 1/2(e^B -e^-B)
594 // = 1/2(2^N * 2^(j*log2/64) - 2^-N * 2^(-j*log2/64))
595 // sinh(B) = (2^(N-1) * 2^(j*log2/64) - 2^(-N-1) * 2^(-j*log2/64))
596 // cosh(B) = (2^(N-1) * 2^(j*log2/64) + 2^(-N-1) * 2^(-j*log2/64))
597 // 2^(j*log2/64) is stored as Tjhi + Tjlo , j= -32,....,32
598 // Tjhi is double-extended (80-bit) and Tjlo is single(32-bit)
599 // R = ax - M*log2/64
600 // R = ax - M*log2_by_64_hi - M*log2_by_64_lo
601 // exp(R) = 1 + R +R^2(1/2! + R(1/3! + R(1/4! + ... + R(1/n!)...)
602 // = 1 + p_odd + p_even
603 // where the p_even uses the A coefficients and the p_even uses the B coefficients
604 // So sinh(R) = 1 + p_odd + p_even -(1 -p_odd -p_even)/2 = p_odd
605 // cosh(R) = 1 + p_even
606 // sinh(B) = S_hi + S_lo
608 // sinh(x) = sinh(B)cosh(R) + cosh(B)sinh(R)
609 // ******************************************************
610 // STEP 1 (TBL and EXP)
611 // ******************************************************
612 // Get the following constants.
618 (p0) adds r32 = 0x1,r0
619 (p0) addl r34 = @ltoff(double_sinh_arg_reduction), gp
632 // We want 2^(N-1) and 2^(-N-1). So bias N-1 and -N-1 and
633 // put them in an exponent.
634 // sinh_FR_spos = 2^(N-1) and sinh_FR_sneg = 2^(-N-1)
635 // r39 = 0xffff + (N-1) = 0xffff +N -1
636 // r40 = 0xffff - (N +1) = 0xffff -N -1
640 (p0) movl r38 = 0x000000000000fffe ;;
644 (p0) ldfe sinh_FR_Inv_log2by64 = [r34],16 ;;
645 (p0) ldfe sinh_FR_log2by64_hi = [r34],16
650 (p0) ldfe sinh_FR_log2by64_lo = [r34],16
655 // Get the A coefficients
662 (p0) addl r34 = @ltoff(double_sinh_ab_table), gp
675 // Calculate M and keep it as integer and floating point.
676 // f38 = M = round-to-integer(x*Inv_log2by64)
677 // sinh_FR_M = M = truncate(ax/(log2/64))
678 // Put the significand of M in r35
679 // and the floating point representation of M in sinh_FR_M
683 (p0) fma.s1 sinh_FR_M = sinh_FR_X, sinh_FR_Inv_log2by64, f0
688 (p0) ldfe sinh_FR_A1 = [r34],16
695 (p0) fcvt.fx.s1 sinh_FR_M_temp = sinh_FR_M
701 (p0) fnorm.s1 sinh_FR_M = sinh_FR_M_temp
706 (p0) getf.sig r35 = sinh_FR_M_temp
711 // M is still in r35. Calculate j. j is the signed extension of the six lsb of M. It
712 // has a range of -32 thru 31.
719 (p0) and r36 = 0x3f, r35 ;;
723 // f13 = f44 - f12*f10 = ax - M*log2by64_hi
724 // f14 = f13 - f8*f11 = R = (ax - M*log2by64_hi) - M*log2by64_lo
728 (p0) fnma.s1 sinh_FR_R_temp = sinh_FR_M, sinh_FR_log2by64_hi, sinh_FR_X
733 (p0) ldfe sinh_FR_A2 = [r34],16
740 (p0) fnma.s1 sinh_FR_R = sinh_FR_M, sinh_FR_log2by64_lo, sinh_FR_R_temp
744 // Get the B coefficients
750 (p0) ldfe sinh_FR_A3 = [r34],16 ;;
751 (p0) ldfe sinh_FR_B1 = [r34],16
756 (p0) ldfe sinh_FR_B2 = [r34],16 ;;
757 (p0) ldfe sinh_FR_B3 = [r34],16
763 (p0) shl r34 = r36, 0x2 ;;
764 (p0) sxt1 r37 = r34 ;;
767 // ******************************************************
768 // STEP 2 (TBL and EXP)
769 // ******************************************************
770 // Calculate Rsquared and Rcubed in preparation for p_even and p_odd
773 // f14 = R <== from above
777 (p0) fma.s1 sinh_FR_Rsq = sinh_FR_R, sinh_FR_R, f0
778 (p0) shr r36 = r37, 0x2 ;;
781 // r34 = M-j = r35 - r36
782 // r35 = N = (M-j)/64
785 (p0) sub r34 = r35, r36
787 (p0) shr r35 = r34, 0x6 ;;
791 (p0) sub r40 = r38, r35
792 (p0) adds r37 = 0x1, r35
793 (p0) add r39 = r38, r35 ;;
796 // Get the address of the J table, add the offset,
797 // addresses are sinh_AD_mJ and sinh_AD_J, get the T value
804 (p0) sub r34 = r35, r32
805 (p0) addl r37 = @ltoff(double_sinh_j_table), gp
820 (p0) fma.s1 sinh_FR_Rcub = sinh_FR_Rsq, sinh_FR_R, f0
824 // ******************************************************
825 // STEP 3 Now decide if we need to branch to EXP
826 // ******************************************************
827 // Put 32 in f9; p6 true if x < 32
828 // Go to EXP if |x| >= 32
832 (p0) movl r32 = 0x0000000000010004 ;;
836 // f34 = B_2 + Rsq *B_3
837 // f35 = B_1 + Rsq*f34 = B_1 + Rsq * (B_2 + Rsq *B_3)
838 // f36 = p_even = Rsq * f35 = Rsq * (B_1 + Rsq * (B_2 + Rsq *B_3))
842 (p0) fma.s1 sinh_FR_peven_temp1 = sinh_FR_Rsq, sinh_FR_B3, sinh_FR_B2
848 (p0) fma.s1 sinh_FR_peven_temp2 = sinh_FR_Rsq, sinh_FR_peven_temp1, sinh_FR_B1
853 // f34 = A_2 + Rsq *A_3
854 // f35 = A_1 + Rsq * (A_2 + Rsq *A_3)
855 // f37 = podd = R + Rcub * (A_1 + Rsq * (A_2 + Rsq *A_3))
859 (p0) fma.s1 sinh_FR_podd_temp1 = sinh_FR_Rsq, sinh_FR_A3, sinh_FR_A2
864 (p0) setf.exp sinh_FR_N_temp1 = r39
871 (p0) fma.s1 sinh_FR_peven = sinh_FR_Rsq, sinh_FR_peven_temp2, f0
877 (p0) fma.s1 sinh_FR_podd_temp2 = sinh_FR_Rsq, sinh_FR_podd_temp1, sinh_FR_A1
882 (p0) setf.exp f9 = r32
889 (p0) fma.s1 sinh_FR_podd = sinh_FR_podd_temp2, sinh_FR_Rcub, sinh_FR_R
893 // sinh_GR_mj contains the table offset for -j
894 // sinh_GR_j contains the table offset for +j
895 // p6 is true when j <= 0
898 (p0) setf.exp sinh_FR_N_temp2 = r40
899 (p0) movl r40 = 0x0000000000000020 ;;
903 (p0) sub sinh_GR_mJ = r40, r36
904 (p0) fmerge.se sinh_FR_spos = sinh_FR_N_temp1, f1
905 (p0) adds sinh_GR_J = 0x20, r36 ;;
910 (p0) shl sinh_GR_mJ = sinh_GR_mJ, 5 ;;
911 (p0) add sinh_AD_mJ = r37, sinh_GR_mJ ;;
916 (p0) ldfe sinh_FR_Tmjhi = [sinh_AD_mJ],16
917 (p0) shl sinh_GR_J = sinh_GR_J, 5 ;;
921 (p0) ldfs sinh_FR_Tmjlo = [sinh_AD_mJ],16
922 (p0) fcmp.lt.unc.s1 p0,p7 = sinh_FR_X,f9
923 (p0) add sinh_AD_J = r37, sinh_GR_J ;;
927 (p0) ldfe sinh_FR_Tjhi = [sinh_AD_J],16 ;;
928 (p0) ldfs sinh_FR_Tjlo = [sinh_AD_J],16
934 (p0) fmerge.se sinh_FR_sneg = sinh_FR_N_temp2, f1
935 (p7) br.cond.spnt L(SINH_BY_EXP) ;;
944 // ******************************************************
945 // If NOT branch to EXP
946 // ******************************************************
947 // Calculate S_hi and S_lo
948 // sinh_FR_S_hi_temp = sinh_FR_sneg * sinh_FR_Tmjhi
949 // sinh_FR_S_hi = sinh_FR_spos * sinh_FR_Tjhi - sinh_FR_S_hi_temp
950 // sinh_FR_S_hi = sinh_FR_spos * sinh_FR_Tjhi - (sinh_FR_sneg * sinh_FR_Tmjlo)
954 (p0) fma.s1 sinh_FR_S_hi_temp = sinh_FR_sneg, sinh_FR_Tmjhi, f0
960 (p0) fms.s1 sinh_FR_S_hi = sinh_FR_spos, sinh_FR_Tjhi, sinh_FR_S_hi_temp
965 // sinh_FR_C_hi_temp1 = sinh_FR_sneg * sinh_FR_Tmjhi
966 // sinh_FR_C_hi = sinh_FR_spos * sinh_FR_Tjhi + sinh_FR_C_hi_temp1
970 (p0) fma.s1 sinh_FR_C_hi_temp1 = sinh_FR_sneg, sinh_FR_Tmjhi, f0
974 // sinh_FR_S_lo_temp1 = sinh_FR_spos * sinh_FR_Tjhi - sinh_FR_S_hi
975 // sinh_FR_S_lo_temp2 = -sinh_FR_sneg * sinh_FR_Tmjlo + (sinh_FR_spos * sinh_FR_Tjhi - sinh_FR_S_hi)
976 // sinh_FR_S_lo_temp2 = -sinh_FR_sneg * sinh_FR_Tmjlo + (sinh_FR_S_lo_temp1 )
980 (p0) fms.s1 sinh_FR_S_lo_temp1 = sinh_FR_spos, sinh_FR_Tjhi, sinh_FR_S_hi
986 (p0) fma.s1 sinh_FR_C_hi = sinh_FR_spos, sinh_FR_Tjhi, sinh_FR_C_hi_temp1
992 (p0) fnma.s1 sinh_FR_S_lo_temp2 = sinh_FR_sneg, sinh_FR_Tmjhi, sinh_FR_S_lo_temp1
996 // sinh_FR_S_lo_temp1 = sinh_FR_sneg * sinh_FR_Tmjlo
997 // sinh_FR_S_lo_temp3 = sinh_FR_spos * sinh_FR_Tjlo - sinh_FR_S_lo_temp1
998 // sinh_FR_S_lo_temp3 = sinh_FR_spos * sinh_FR_Tjlo -(sinh_FR_sneg * sinh_FR_Tmjlo)
999 // sinh_FR_S_lo = sinh_FR_S_lo_temp3 + sinh_FR_S_lo_temp2
1003 (p0) fma.s1 sinh_FR_S_lo_temp1 = sinh_FR_sneg, sinh_FR_Tmjlo, f0
1007 /////////// BUG FIX fma to fms -TK
1010 (p0) fms.s1 sinh_FR_S_lo_temp3 = sinh_FR_spos, sinh_FR_Tjlo, sinh_FR_S_lo_temp1
1016 (p0) fma.s1 sinh_FR_S_lo = sinh_FR_S_lo_temp3, f1, sinh_FR_S_lo_temp2
1021 // Y_lo = C_hi*p_odd + (S_hi*p_even + S_lo)
1022 // sinh_FR_Y_lo_temp = sinh_FR_S_hi * sinh_FR_peven + sinh_FR_S_lo
1023 // sinh_FR_Y_lo = sinh_FR_C_hi * sinh_FR_podd + sinh_FR_Y_lo_temp
1027 (p0) fma.s1 sinh_FR_Y_lo_temp = sinh_FR_S_hi, sinh_FR_peven, sinh_FR_S_lo
1033 (p0) fma.s1 sinh_FR_Y_lo = sinh_FR_C_hi, sinh_FR_podd, sinh_FR_Y_lo_temp
1037 // sinh_FR_SINH = Y_hi + Y_lo
1038 // f8 = answer = sinh_FR_SGNX * sinh_FR_SINH
1040 // Dummy multiply to generate inexact
1043 (p0) fmpy.s0 sinh_FR_tmp = sinh_FR_all_ones, sinh_FR_all_ones
1048 (p0) fma.s1 sinh_FR_SINH = sinh_FR_S_hi, f1, sinh_FR_Y_lo
1054 (p0) fma.s.s0 f8 = sinh_FR_SGNX, sinh_FR_SINH,f0
1055 (p0) br.ret.sptk b0 ;;
1061 // When p7 is true, we know that an overflow is not going to happen
1062 // When p7 is false, we must check for possible overflow
1063 // p7 is the over_SAFE flag
1065 // Y_lo = Tjhi * (p_odd + p_even) +Tjlo
1066 // Scale = sign * 2^(N-1)
1067 // sinh_FR_Y_lo = sinh_FR_Tjhi * (sinh_FR_peven + sinh_FR_podd)
1068 // sinh_FR_Y_lo = sinh_FR_Tjhi * (sinh_FR_Y_lo_temp )
1072 (p0) fma.s1 sinh_FR_Y_lo_temp = sinh_FR_peven, f1, sinh_FR_podd
1076 // Now we are in EXP. This is the only path where an overflow is possible
1077 // but not for certain. So this is the only path where over_SAFE has any use.
1078 // r34 still has N-1
1079 // There is a danger of double-extended overflow if N-1 > 16382 = 0x3ffe
1080 // There is a danger of double overflow if N-1 > 0x3fe = 1022
1081 // There is a danger of single overflow if N-1 > 0x7e = 126
1084 (p0) movl r32 = 0x000000000000007e ;;
1088 (p0) cmp.gt.unc p0,p7 = r34, r32
1089 (p0) fmerge.s sinh_FR_SCALE = sinh_FR_SGNX, sinh_FR_spos
1095 (p0) fma.s1 sinh_FR_Y_lo = sinh_FR_Tjhi, sinh_FR_Y_lo_temp, sinh_FR_Tjlo
1099 // f8 = answer = scale * (Y_hi + Y_lo)
1102 (p0) fma.s1 sinh_FR_SINH_temp = sinh_FR_Y_lo, f1, sinh_FR_Tjhi
1108 (p0) fma.s.s0 f44 = sinh_FR_SCALE, sinh_FR_SINH_temp, f0
1112 // Dummy multiply to generate inexact
1115 (p7) fmpy.s0 sinh_FR_tmp = sinh_FR_all_ones, sinh_FR_all_ones
1119 // If over_SAFE is set, return
1122 (p7) fmerge.s f8 = f44,f44
1123 (p7) br.ret.sptk b0 ;;
1126 // Else see if we overflowed
1127 // S0 user supplied status
1128 // S2 user supplied status + WRE + TD (Overflows)
1129 // If WRE is set then an overflow will not occur in EXP.
1130 // The input value that would cause a register (WRE) value to overflow is about 2^15
1131 // and this input would go into the HUGE path.
1132 // Answer with WRE is in f43.
1136 (p0) fsetc.s2 0x7F,0x42
1142 (p0) fma.s.s2 f43 = sinh_FR_SCALE, sinh_FR_SINH_temp, f0
1146 // 1007F => 1007F -FFFF = 80(true)
1147 // 80 + 7F = FF, which is 1 more that the exponent of the largest
1148 // double (FE). So 0 1007F 8000000000000000 is one ulp more than
1149 // largest single in register bias
1150 // Now set p8 if the answer with WRE is greater than or equal this value
1151 // Also set p9 if the answer with WRE is less than or equal to negative this value
1155 (p0) movl r32 = 0x0000000001007F ;;
1160 (p0) setf.exp f41 = r32
1161 (p0) fsetc.s2 0x7F,0x40 ;;
1166 (p0) fcmp.ge.unc.s1 p8, p0 = f43, f41
1172 (p0) fmerge.ns f42 = f41, f41
1176 // The error tag for overflow is 128
1180 (p8) mov r47 = 128 ;;
1185 (p0) fcmp.le.unc.s1 p9, p0 = f43, f42
1186 (p8) br.cond.spnt L(SINH_ERROR_SUPPORT) ;;
1198 (p9) br.cond.spnt L(SINH_ERROR_SUPPORT) ;;
1201 // Dummy multiply to generate inexact
1204 (p0) fmpy.s0 sinh_FR_tmp = sinh_FR_all_ones, sinh_FR_all_ones
1210 (p0) fmerge.s f8 = f44,f44
1211 (p0) br.ret.sptk b0 ;;
1216 // for SINH_HUGE, put 24000 in exponent; take sign from input; add 1
1217 // SAFE: SAFE is always 0 for HUGE
1221 (p0) movl r32 = 0x0000000000015dbf ;;
1225 (p0) setf.exp f9 = r32
1232 (p0) fma.s1 sinh_FR_signed_hi_lo = sinh_FR_SGNX, f9, f1
1238 (p0) fma.s.s0 f44 = sinh_FR_signed_hi_lo, f9, f0
1242 ASM_SIZE_DIRECTIVE(sinhf)
1244 ASM_SIZE_DIRECTIVE(__ieee754_sinhf)
1247 // Stack operations when calling error support.
1248 // (1) (2) (3) (call) (4)
1249 // sp -> + psp -> + psp -> + sp -> +
1251 // | | <- GR_Y R3 ->| <- GR_RESULT | -> f8
1253 // | <-GR_Y Y2->| Y2 ->| <- GR_Y |
1255 // | | <- GR_X X1 ->| |
1257 // sp-64 -> + sp -> + sp -> + +
1258 // save ar.pfs save b0 restore gp
1259 // save gp restore ar.pfs
1261 .proc __libm_error_region
1262 __libm_error_region:
1263 L(SINH_ERROR_SUPPORT):
1268 add GR_Parameter_Y=-32,sp // Parameter 2 value
1270 .save ar.pfs,GR_SAVE_PFS
1271 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
1275 add sp=-64,sp // Create new stack
1277 mov GR_SAVE_GP=gp // Save gp
1283 stfs [GR_Parameter_Y] = f0,16 // STORE Parameter 2 on stack
1284 add GR_Parameter_X = 16,sp // Parameter 1 address
1285 .save b0, GR_SAVE_B0
1286 mov GR_SAVE_B0=b0 // Save b0
1292 stfs [GR_Parameter_X] = f8 // STORE Parameter 1 on stack
1293 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
1297 stfs [GR_Parameter_Y] = f44 // STORE Parameter 3 on stack
1298 add GR_Parameter_Y = -16,GR_Parameter_Y
1299 br.call.sptk b0=__libm_error_support# // Call error handling function
1304 add GR_Parameter_RESULT = 48,sp
1309 ldfs f8 = [GR_Parameter_RESULT] // Get return result off stack
1311 add sp = 64,sp // Restore stack pointer
1312 mov b0 = GR_SAVE_B0 // Restore return address
1315 mov gp = GR_SAVE_GP // Restore gp
1316 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
1317 br.ret.sptk b0 // Return
1320 .endp __libm_error_region
1321 ASM_SIZE_DIRECTIVE(__libm_error_region)
1323 .type __libm_error_support#,@function
1324 .global __libm_error_support#