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 // 1/23/01 Set inexact flag for large args.
49 //==============================================================
50 // float = cosh(float)
51 // double = cosh(double)
52 // long double = coshl(long double)
53 // input floating point f8
54 // output floating point f8
57 // Overview of operation
58 //==============================================================
59 // There are four paths
61 // 1. |x| < 0.25 COSH_BY_POLY
62 // 2. |x| < 32 COSH_BY_TBL
63 // 3. |x| < 2^14 COSH_BY_EXP
64 // 4. |x| >= 2^14 COSH_HUGE
66 // For paths 1, and 2 SAFE is always 1.
67 // For path 4, Safe is always 0.
68 // SAFE = 1 means we cannot overflow.
70 #include "libm_support.h"
73 //==============================================================
77 cosh_FR_all_ones = f45
81 cosh_FR_Inv_log2by64 = f9
82 cosh_FR_log2by64_lo = f11
83 cosh_FR_log2by64_hi = f10
102 cosh_FR_peven_temp1 = f34
103 cosh_FR_peven_temp2 = f35
106 cosh_FR_podd_temp1 = f34
107 cosh_FR_podd_temp2 = f35
115 cosh_FR_N_temp1 = f11
116 cosh_FR_N_temp2 = f12
133 cosh_GR_all_ones = r40
140 GR_Parameter_RESULT = r46
141 GR_Parameter_TAG = r47
144 cosh_FR_C_hi_temp = f10
145 cosh_FR_C_lo_temp1 = f11
146 cosh_FR_C_lo_temp2 = f12
147 cosh_FR_C_lo_temp3 = f13
152 cosh_FR_S_hi_temp1 = f10
154 cosh_FR_Y_lo_temp = f12
168 cosh_FR_TINY_THRESH = f9
170 cosh_FR_COSH_temp = f10
175 cosh_FR_poly_podd_temp1 = f11
176 cosh_FR_poly_podd_temp2 = f13
177 cosh_FR_poly_peven_temp1 = f11
178 cosh_FR_poly_peven_temp2 = f13
181 //==============================================================
190 double_cosh_arg_reduction:
191 ASM_TYPE_DIRECTIVE(double_cosh_arg_reduction,@object)
192 data8 0xB8AA3B295C17F0BC, 0x00004005
193 data8 0xB17217F7D1000000, 0x00003FF8
194 data8 0xCF79ABC9E3B39804, 0x00003FD0
195 ASM_SIZE_DIRECTIVE(double_cosh_arg_reduction)
198 ASM_TYPE_DIRECTIVE(double_cosh_p_table,@object)
199 data8 0x8000000000000000, 0x00003FFE
200 data8 0xAAAAAAAAAAAAAB80, 0x00003FFA
201 data8 0xB60B60B60B4FE884, 0x00003FF5
202 data8 0xD00D00D1021D7370, 0x00003FEF
203 data8 0x93F27740C0C2F1CC, 0x00003FE9
204 data8 0x8FA02AC65BCBD5BC, 0x00003FE2
205 ASM_SIZE_DIRECTIVE(double_cosh_p_table)
207 double_cosh_ab_table:
208 ASM_TYPE_DIRECTIVE(double_cosh_ab_table,@object)
209 data8 0xAAAAAAAAAAAAAAAC, 0x00003FFC
210 data8 0x88888888884ECDD5, 0x00003FF8
211 data8 0xD00D0C6DCC26A86B, 0x00003FF2
212 data8 0x8000000000000002, 0x00003FFE
213 data8 0xAAAAAAAAAA402C77, 0x00003FFA
214 data8 0xB60B6CC96BDB144D, 0x00003FF5
215 ASM_SIZE_DIRECTIVE(double_cosh_ab_table)
218 ASM_TYPE_DIRECTIVE(double_cosh_j_table,@object)
219 data8 0xB504F333F9DE6484, 0x00003FFE, 0x1EB2FB13, 0x00000000
220 data8 0xB6FD91E328D17791, 0x00003FFE, 0x1CE2CBE2, 0x00000000
221 data8 0xB8FBAF4762FB9EE9, 0x00003FFE, 0x1DDC3CBC, 0x00000000
222 data8 0xBAFF5AB2133E45FB, 0x00003FFE, 0x1EE9AA34, 0x00000000
223 data8 0xBD08A39F580C36BF, 0x00003FFE, 0x9EAEFDC1, 0x00000000
224 data8 0xBF1799B67A731083, 0x00003FFE, 0x9DBF517B, 0x00000000
225 data8 0xC12C4CCA66709456, 0x00003FFE, 0x1EF88AFB, 0x00000000
226 data8 0xC346CCDA24976407, 0x00003FFE, 0x1E03B216, 0x00000000
227 data8 0xC5672A115506DADD, 0x00003FFE, 0x1E78AB43, 0x00000000
228 data8 0xC78D74C8ABB9B15D, 0x00003FFE, 0x9E7B1747, 0x00000000
229 data8 0xC9B9BD866E2F27A3, 0x00003FFE, 0x9EFE3C0E, 0x00000000
230 data8 0xCBEC14FEF2727C5D, 0x00003FFE, 0x9D36F837, 0x00000000
231 data8 0xCE248C151F8480E4, 0x00003FFE, 0x9DEE53E4, 0x00000000
232 data8 0xD06333DAEF2B2595, 0x00003FFE, 0x9E24AE8E, 0x00000000
233 data8 0xD2A81D91F12AE45A, 0x00003FFE, 0x1D912473, 0x00000000
234 data8 0xD4F35AABCFEDFA1F, 0x00003FFE, 0x1EB243BE, 0x00000000
235 data8 0xD744FCCAD69D6AF4, 0x00003FFE, 0x1E669A2F, 0x00000000
236 data8 0xD99D15C278AFD7B6, 0x00003FFE, 0x9BBC610A, 0x00000000
237 data8 0xDBFBB797DAF23755, 0x00003FFE, 0x1E761035, 0x00000000
238 data8 0xDE60F4825E0E9124, 0x00003FFE, 0x9E0BE175, 0x00000000
239 data8 0xE0CCDEEC2A94E111, 0x00003FFE, 0x1CCB12A1, 0x00000000
240 data8 0xE33F8972BE8A5A51, 0x00003FFE, 0x1D1BFE90, 0x00000000
241 data8 0xE5B906E77C8348A8, 0x00003FFE, 0x1DF2F47A, 0x00000000
242 data8 0xE8396A503C4BDC68, 0x00003FFE, 0x1EF22F22, 0x00000000
243 data8 0xEAC0C6E7DD24392F, 0x00003FFE, 0x9E3F4A29, 0x00000000
244 data8 0xED4F301ED9942B84, 0x00003FFE, 0x1EC01A5B, 0x00000000
245 data8 0xEFE4B99BDCDAF5CB, 0x00003FFE, 0x1E8CAC3A, 0x00000000
246 data8 0xF281773C59FFB13A, 0x00003FFE, 0x9DBB3FAB, 0x00000000
247 data8 0xF5257D152486CC2C, 0x00003FFE, 0x1EF73A19, 0x00000000
248 data8 0xF7D0DF730AD13BB9, 0x00003FFE, 0x9BB795B5, 0x00000000
249 data8 0xFA83B2DB722A033A, 0x00003FFE, 0x1EF84B76, 0x00000000
250 data8 0xFD3E0C0CF486C175, 0x00003FFE, 0x9EF5818B, 0x00000000
251 data8 0x8000000000000000, 0x00003FFF, 0x00000000, 0x00000000
252 data8 0x8164D1F3BC030773, 0x00003FFF, 0x1F77CACA, 0x00000000
253 data8 0x82CD8698AC2BA1D7, 0x00003FFF, 0x1EF8A91D, 0x00000000
254 data8 0x843A28C3ACDE4046, 0x00003FFF, 0x1E57C976, 0x00000000
255 data8 0x85AAC367CC487B15, 0x00003FFF, 0x9EE8DA92, 0x00000000
256 data8 0x871F61969E8D1010, 0x00003FFF, 0x1EE85C9F, 0x00000000
257 data8 0x88980E8092DA8527, 0x00003FFF, 0x1F3BF1AF, 0x00000000
258 data8 0x8A14D575496EFD9A, 0x00003FFF, 0x1D80CA1E, 0x00000000
259 data8 0x8B95C1E3EA8BD6E7, 0x00003FFF, 0x9D0373AF, 0x00000000
260 data8 0x8D1ADF5B7E5BA9E6, 0x00003FFF, 0x9F167097, 0x00000000
261 data8 0x8EA4398B45CD53C0, 0x00003FFF, 0x1EB70051, 0x00000000
262 data8 0x9031DC431466B1DC, 0x00003FFF, 0x1F6EB029, 0x00000000
263 data8 0x91C3D373AB11C336, 0x00003FFF, 0x1DFD6D8E, 0x00000000
264 data8 0x935A2B2F13E6E92C, 0x00003FFF, 0x9EB319B0, 0x00000000
265 data8 0x94F4EFA8FEF70961, 0x00003FFF, 0x1EBA2BEB, 0x00000000
266 data8 0x96942D3720185A00, 0x00003FFF, 0x1F11D537, 0x00000000
267 data8 0x9837F0518DB8A96F, 0x00003FFF, 0x1F0D5A46, 0x00000000
268 data8 0x99E0459320B7FA65, 0x00003FFF, 0x9E5E7BCA, 0x00000000
269 data8 0x9B8D39B9D54E5539, 0x00003FFF, 0x9F3AAFD1, 0x00000000
270 data8 0x9D3ED9A72CFFB751, 0x00003FFF, 0x9E86DACC, 0x00000000
271 data8 0x9EF5326091A111AE, 0x00003FFF, 0x9F3EDDC2, 0x00000000
272 data8 0xA0B0510FB9714FC2, 0x00003FFF, 0x1E496E3D, 0x00000000
273 data8 0xA27043030C496819, 0x00003FFF, 0x9F490BF6, 0x00000000
274 data8 0xA43515AE09E6809E, 0x00003FFF, 0x1DD1DB48, 0x00000000
275 data8 0xA5FED6A9B15138EA, 0x00003FFF, 0x1E65EBFB, 0x00000000
276 data8 0xA7CD93B4E965356A, 0x00003FFF, 0x9F427496, 0x00000000
277 data8 0xA9A15AB4EA7C0EF8, 0x00003FFF, 0x1F283C4A, 0x00000000
278 data8 0xAB7A39B5A93ED337, 0x00003FFF, 0x1F4B0047, 0x00000000
279 data8 0xAD583EEA42A14AC6, 0x00003FFF, 0x1F130152, 0x00000000
280 data8 0xAF3B78AD690A4375, 0x00003FFF, 0x9E8367C0, 0x00000000
281 data8 0xB123F581D2AC2590, 0x00003FFF, 0x9F705F90, 0x00000000
282 data8 0xB311C412A9112489, 0x00003FFF, 0x1EFB3C53, 0x00000000
283 data8 0xB504F333F9DE6484, 0x00003FFF, 0x1F32FB13, 0x00000000
284 ASM_SIZE_DIRECTIVE(double_cosh_j_table)
296 .global __ieee754_coshl#
297 .proc __ieee754_coshl#
304 alloc r32 = ar.pfs,0,12,4,0
305 (p0) fclass.m.unc p6,p7 = f8, 0xc3
306 mov cosh_GR_all_ones = -1
309 // This is more than we need but it is in preparation
310 // for the values we add for error support. We push three
311 // addresses on the stack (3*8) = 24 bytes and one tag
315 (p6) fma.s0 f8 = f8,f1,f8
316 (p6) br.ret.spnt b0 ;;
320 // Make constant that will generate inexact when squared
323 setf.sig cosh_FR_all_ones = cosh_GR_all_ones
324 (p0) fclass.m.unc p6,p0 = f8, 0x23
330 (p6) fmerge.s f8 = f0,f8
331 (p6) br.ret.spnt b0 ;;
336 // Put 0.25 in f9; p6 true if x < 0.25
339 (p0) movl r32 = 0x000000000000fffd ;;
343 (p0) setf.exp f9 = r32
350 (p0) fmerge.s cosh_FR_X = f0,f8
356 (p0) fmerge.s cosh_FR_SGNX = f8,f1
362 (p0) fcmp.lt.unc p0,p7 = cosh_FR_X,f9
369 (p7) br.cond.sptk L(COSH_BY_TBL)
375 // POLY cannot overflow so there is no need to call __libm_error_support
376 // Get the values of P_x from the table
380 (p0) addl r34 = @ltoff(double_cosh_p_table), gp
393 // Calculate cosh_FR_X2 = ax*ax and cosh_FR_X4 = ax*ax*ax*ax
396 (p0) ldfe cosh_FR_P1 = [r34],16
397 (p0) fma.s1 cosh_FR_X2 = cosh_FR_X, cosh_FR_X, f0 ;;
401 (p0) ldfe cosh_FR_P2 = [r34],16 ;;
402 (p0) ldfe cosh_FR_P3 = [r34],16
407 (p0) ldfe cosh_FR_P4 = [r34],16 ;;
408 (p0) ldfe cosh_FR_P5 = [r34],16
413 (p0) ldfe cosh_FR_P6 = [r34],16
414 (p0) fma.s1 cosh_FR_X4 = cosh_FR_X2, cosh_FR_X2, f0
418 // Calculate cosh_FR_podd = x4 *(x4 * P_5 + P_3) + P_1
421 (p0) fma.s1 cosh_FR_poly_podd_temp1 = cosh_FR_X4, cosh_FR_P5, cosh_FR_P3
427 (p0) fma.s1 cosh_FR_podd = cosh_FR_X4, cosh_FR_poly_podd_temp1, cosh_FR_P1
431 // Calculate cosh_FR_peven = p_even = x4 *(x4 * (x4 * P_6 + P_4) + P_2)
434 (p0) fma.s1 cosh_FR_poly_peven_temp1 = cosh_FR_X4, cosh_FR_P6, cosh_FR_P4
440 (p0) fma.s1 cosh_FR_poly_peven_temp2 = cosh_FR_X4, cosh_FR_poly_peven_temp1, cosh_FR_P2
446 (p0) fma.s1 cosh_FR_peven = cosh_FR_X4, cosh_FR_poly_peven_temp2, f0
450 // Y_lo = x2*p_odd + p_even
451 // Calculate f8 = Y_hi + Y_lo
454 (p0) fma.s1 cosh_FR_Y_lo = cosh_FR_X2, cosh_FR_podd, cosh_FR_peven
460 (p0) fma.s0 f8 = f1, f1, cosh_FR_Y_lo
461 (p0) br.ret.sptk b0 ;;
467 // Now that we are at TBL; so far all we know is that |x| >= 0.25.
468 // The first two steps are the same for TBL and EXP, but if we are HUGE
470 // Go to HUGE if |x| >= 2^14, 1000d (register-biased) is e = 14 (true)
472 // Go to HUGE if |x| >= 2^10, 10009 (register-biased) is e = 10 (true)
474 // Go to HUGE if |x| >= 2^7, 10006 (register-biased) is e = 7 (true)
475 // we want to leave now. Go to HUGE if |x| >= 2^14
476 // 1000d (register-biased) is e = 14 (true)
480 (p0) movl r32 = 0x000000000001000d ;;
484 (p0) setf.exp f9 = r32
491 (p0) fcmp.ge.unc p6,p7 = cosh_FR_X,f9
498 (p6) br.cond.spnt L(COSH_HUGE) ;;
507 // TBL can never overflow
508 // cosh(x) = cosh(B+R)
509 // = cosh(B) cosh(R) + sinh(B) sinh(R)
510 // cosh(R) can be approximated by 1 + p_even
511 // sinh(R) can be approximated by p_odd
513 // ******************************************************
514 // STEP 1 (TBL and EXP)
515 // ******************************************************
516 // Get the following constants.
522 (p0) adds r32 = 0x1,r0
523 (p0) addl r34 = @ltoff(double_cosh_arg_reduction), gp
528 // We want 2^(N-1) and 2^(-N-1). So bias N-1 and -N-1 and
529 // put them in an exponent.
530 // cosh_FR_spos = 2^(N-1) and cosh_FR_sneg = 2^(-N-1)
531 // r39 = 0xffff + (N-1) = 0xffff +N -1
532 // r40 = 0xffff - (N +1) = 0xffff -N -1
536 (p0) movl r38 = 0x000000000000fffe ;;
540 (p0) ldfe cosh_FR_Inv_log2by64 = [r34],16 ;;
541 (p0) ldfe cosh_FR_log2by64_hi = [r34],16
546 (p0) ldfe cosh_FR_log2by64_lo = [r34],16
551 // Get the A coefficients
558 (p0) addl r34 = @ltoff(double_cosh_ab_table), gp
571 // Calculate M and keep it as integer and floating point.
572 // M = round-to-integer(x*Inv_log2by64)
573 // cosh_FR_M = M = truncate(ax/(log2/64))
574 // Put the significand of M in r35
575 // and the floating point representation of M in cosh_FR_M
579 (p0) fma.s1 cosh_FR_M = cosh_FR_X, cosh_FR_Inv_log2by64, f0
584 (p0) ldfe cosh_FR_A1 = [r34],16
591 (p0) fcvt.fx.s1 cosh_FR_M_temp = cosh_FR_M
597 (p0) fnorm.s1 cosh_FR_M = cosh_FR_M_temp
602 (p0) getf.sig r35 = cosh_FR_M_temp
607 // M is still in r35. Calculate j. j is the signed extension of the six lsb of M. It
608 // has a range of -32 thru 31.
614 (p0) and r36 = 0x3f, r35 ;;
618 // f13 = f44 - f12*f10 = x - M*log2by64_hi
619 // f14 = f13 - f8*f11 = R = (x - M*log2by64_hi) - M*log2by64_lo
623 (p0) fnma.s1 cosh_FR_R_temp = cosh_FR_M, cosh_FR_log2by64_hi, cosh_FR_X
628 (p0) ldfe cosh_FR_A2 = [r34],16
635 (p0) fnma.s1 cosh_FR_R = cosh_FR_M, cosh_FR_log2by64_lo, cosh_FR_R_temp
639 // Get the B coefficients
645 (p0) ldfe cosh_FR_A3 = [r34],16 ;;
646 (p0) ldfe cosh_FR_B1 = [r34],16
651 (p0) ldfe cosh_FR_B2 = [r34],16 ;;
652 (p0) ldfe cosh_FR_B3 = [r34],16
658 (p0) shl r34 = r36, 0x2 ;;
659 (p0) sxt1 r37 = r34 ;;
662 // ******************************************************
663 // STEP 2 (TBL and EXP)
664 // ******************************************************
665 // Calculate Rsquared and Rcubed in preparation for p_even and p_odd
668 // f14 = R <== from above
672 (p0) fma.s1 cosh_FR_Rsq = cosh_FR_R, cosh_FR_R, f0
673 (p0) shr r36 = r37, 0x2 ;;
676 // r34 = M-j = r35 - r36
677 // r35 = N = (M-j)/64
680 (p0) sub r34 = r35, r36
682 (p0) shr r35 = r34, 0x6 ;;
686 (p0) sub r40 = r38, r35
687 (p0) adds r37 = 0x1, r35
688 (p0) add r39 = r38, r35 ;;
691 // Get the address of the J table, add the offset,
692 // addresses are sinh_AD_mJ and sinh_AD_J, get the T value
699 (p0) sub r34 = r35, r32
700 (p0) addl r37 = @ltoff(double_cosh_j_table), gp
707 (p0) fma.s1 cosh_FR_Rcub = cosh_FR_Rsq, cosh_FR_R, f0
711 // ******************************************************
712 // STEP 3 Now decide if we need to branch to EXP
713 // ******************************************************
714 // Put 32 in f9; p6 true if x < 32
718 (p0) movl r32 = 0x0000000000010004 ;;
722 // f34 = B_2 + Rsq *B_3
723 // f35 = B_1 + Rsq*f34 = B_1 + Rsq * (B_2 + Rsq *B_3)
724 // f36 = peven = Rsq * f35 = Rsq * (B_1 + Rsq * (B_2 + Rsq *B_3))
728 (p0) fma.s1 cosh_FR_peven_temp1 = cosh_FR_Rsq, cosh_FR_B3, cosh_FR_B2
734 (p0) fma.s1 cosh_FR_peven_temp2 = cosh_FR_Rsq, cosh_FR_peven_temp1, cosh_FR_B1
739 // f34 = A_2 + Rsq *A_3
740 // f35 = A_1 + Rsq * (A_2 + Rsq *A_3)
741 // f37 = podd = R + Rcub * (A_1 + Rsq * (A_2 + Rsq *A_3))
745 (p0) fma.s1 cosh_FR_podd_temp1 = cosh_FR_Rsq, cosh_FR_A3, cosh_FR_A2
750 (p0) setf.exp cosh_FR_N_temp1 = r39
757 (p0) fma.s1 cosh_FR_peven = cosh_FR_Rsq, cosh_FR_peven_temp2, f0
763 (p0) fma.s1 cosh_FR_podd_temp2 = cosh_FR_Rsq, cosh_FR_podd_temp1, cosh_FR_A1
768 (p0) setf.exp f9 = r32
775 (p0) fma.s1 cosh_FR_podd = cosh_FR_podd_temp2, cosh_FR_Rcub, cosh_FR_R
779 // sinh_GR_mj contains the table offset for -j
780 // sinh_GR_j contains the table offset for +j
781 // p6 is true when j <= 0
784 (p0) setf.exp cosh_FR_N_temp2 = r40
785 (p0) movl r40 = 0x0000000000000020 ;;
789 (p0) sub GR_mJ = r40, r36
790 (p0) fmerge.se cosh_FR_spos = cosh_FR_N_temp1, f1
791 (p0) adds GR_J = 0x20, r36 ;;
796 (p0) shl GR_mJ = GR_mJ, 5 ;;
797 (p0) add AD_mJ = r37, GR_mJ ;;
802 (p0) ldfe cosh_FR_Tmjhi = [AD_mJ],16
803 (p0) shl GR_J = GR_J, 5 ;;
807 (p0) ldfs cosh_FR_Tmjlo = [AD_mJ],16
808 (p0) fcmp.lt.unc.s1 p6,p7 = cosh_FR_X,f9
809 (p0) add AD_J = r37, GR_J ;;
813 (p0) ldfe cosh_FR_Tjhi = [AD_J],16 ;;
814 (p0) ldfs cosh_FR_Tjlo = [AD_J],16
820 (p0) fmerge.se cosh_FR_sneg = cosh_FR_N_temp2, f1
821 (p7) br.cond.spnt L(COSH_BY_EXP) ;;
824 // ******************************************************
825 // If NOT branch to EXP
826 // ******************************************************
828 // ******************************************************
829 // cosh_FR_C_hi_temp = cosh_FR_sneg * cosh_FR_Tmjhi
830 // cosh_FR_C_hi = cosh_FR_spos * cosh_FR_Tjhi + (cosh_FR_sneg * cosh_FR_Tmjhi)
834 (p0) fma.s1 cosh_FR_C_hi_temp = cosh_FR_sneg, cosh_FR_Tmjhi, f0
840 (p0) fma.s1 cosh_FR_C_hi = cosh_FR_spos, cosh_FR_Tjhi, cosh_FR_C_hi_temp
844 // ******************************************************
846 // ******************************************************
847 // cosh_FR_S_hi_temp1 = cosh_FR_sneg * cosh_FR_Tmjhi
848 // cosh_FR_S_hi = cosh_FR_spos * cosh_FR_Tjhi - cosh_FR_C_hi_temp1
852 (p0) fma.s1 cosh_FR_S_hi_temp1 = cosh_FR_sneg, cosh_FR_Tmjhi, f0
856 // ******************************************************
858 // ******************************************************
859 // cosh_FR_C_lo_temp1 = cosh_FR_spos * cosh_FR_Tjhi - cosh_FR_C_hi
860 // cosh_FR_C_lo_temp2 = cosh_FR_sneg * cosh_FR_Tmjlo + (cosh_FR_spos * cosh_FR_Tjhi - cosh_FR_C_hi)
861 // cosh_FR_C_lo_temp1 = cosh_FR_sneg * cosh_FR_Tmjlo
862 // cosh_FR_C_lo_temp3 = cosh_FR_spos * cosh_FR_Tjlo + (cosh_FR_sneg * cosh_FR_Tmjlo)
863 // cosh_FR_C_lo = cosh_FR_C_lo_temp3 + cosh_FR_C_lo_temp2
867 (p0) fms.s1 cosh_FR_C_lo_temp1 = cosh_FR_spos, cosh_FR_Tjhi, cosh_FR_C_hi
873 (p0) fms.s1 cosh_FR_S_hi = cosh_FR_spos, cosh_FR_Tjhi, cosh_FR_S_hi_temp1
879 (p0) fma.s1 cosh_FR_C_lo_temp2 = cosh_FR_sneg, cosh_FR_Tmjhi, cosh_FR_C_lo_temp1
885 (p0) fma.s1 cosh_FR_C_lo_temp1 = cosh_FR_sneg, cosh_FR_Tmjlo, f0
891 (p0) fma.s1 cosh_FR_C_lo_temp3 = cosh_FR_spos, cosh_FR_Tjlo, cosh_FR_C_lo_temp1
897 (p0) fma.s1 cosh_FR_C_lo = cosh_FR_C_lo_temp3, f1, cosh_FR_C_lo_temp2
901 // ******************************************************
902 // cosh_FR_Y_lo_temp = cosh_FR_C_hi * cosh_FR_peven + cosh_FR_C_lo
903 // cosh_FR_Y_lo = cosh_FR_S_hi * cosh_FR_podd + cosh_FR_Y_lo_temp
904 // cosh_FR_COSH = Y_hi + Y_lo
908 (p0) fma.s1 cosh_FR_Y_lo_temp = cosh_FR_C_hi, cosh_FR_peven, cosh_FR_C_lo
914 (p0) fma.s1 cosh_FR_Y_lo = cosh_FR_S_hi, cosh_FR_podd, cosh_FR_Y_lo_temp
920 (p0) fma.s0 f8 = cosh_FR_C_hi, f1, cosh_FR_Y_lo
921 (p0) br.ret.sptk b0 ;;
926 // When p7 is true, we know that an overflow is not going to happen
927 // When p7 is false, we must check for possible overflow
928 // p7 is the over_SAFE flag
929 // f44 = Scale * (Y_hi + Y_lo)
930 // = cosh_FR_spos * (cosh_FR_Tjhi + cosh_FR_Y_lo)
934 (p0) fma.s1 cosh_FR_Y_lo_temp = cosh_FR_peven, f1, cosh_FR_podd
938 // Now we are in EXP. This is the only path where an overflow is possible
939 // but not for certain. So this is the only path where over_SAFE has any use.
941 // There is a danger of double-extended overflow if N-1 > 0x3ffe = 16382
942 // There is a danger of double overflow if N-1 > 0x3fe = 1022
943 // There is a danger of single overflow if N-1 > 0x7e = 126
947 (p0) movl r32 = 0x0000000000003ffe ;;
951 (p0) cmp.gt.unc p0,p7 = r34, r32
958 (p0) fma.s1 cosh_FR_Y_lo = cosh_FR_Tjhi, cosh_FR_Y_lo_temp, cosh_FR_Tjlo
964 (p0) fma.s1 cosh_FR_COSH_temp = cosh_FR_Y_lo, f1, cosh_FR_Tjhi
970 (p0) fma.s0 f44 = cosh_FR_spos, cosh_FR_COSH_temp, f0
974 // Dummy multiply to generate inexact
977 (p7) fmpy.s0 cosh_FR_tmp = cosh_FR_all_ones, cosh_FR_all_ones
981 // If over_SAFE is set, return
984 (p7) fmerge.s f8 = f44,f44
985 (p7) br.ret.sptk b0 ;;
988 // Else see if we overflowed
989 // S0 user supplied status
990 // S2 user supplied status + WRE + TD (Overflows)
991 // If WRE is set then an overflow will not occur in EXP.
992 // The input value that would cause a register (WRE) value to overflow is about 2^15
993 // and this input would go into the HUGE path.
994 // Answer with WRE is in f43.
998 (p0) fsetc.s2 0x7F,0x42
1004 (p0) fma.s2 f43 = cosh_FR_spos, cosh_FR_COSH_temp, f0
1008 // 103FF => 103FF -FFFF = 400(true)
1009 // 400 + 3FF = 7FF, which is 1 more than the exponent of the largest
1010 // double (7FE). So 0 103FF 8000000000000000 is one ulp more than
1011 // largest double in register bias
1013 // 13FFF => 13FFF -FFFF = 4000(true)
1015 // Now set p8 if the answer with WRE is greater than or equal this value
1016 // Also set p9 if the answer with WRE is less than or equal to negative this value
1020 (p0) movl r32 = 0x0000000000013fff ;;
1025 (p0) setf.exp f41 = r32
1026 (p0) fsetc.s2 0x7F,0x40 ;;
1031 (p0) fcmp.ge.unc.s1 p8, p0 = f43, f41
1037 (p0) fmerge.ns f42 = f41, f41
1041 // The error tag for overflow is 63
1045 (p8) mov GR_Parameter_TAG = 63 ;;
1050 (p0) fcmp.le.unc.s1 p9, p0 = f43, f42
1051 (p8) br.cond.spnt __libm_error_region ;;
1057 (p9) mov GR_Parameter_TAG = 63
1063 (p9) br.cond.spnt __libm_error_region ;;
1066 // Dummy multiply to generate inexact
1069 (p0) fmpy.s0 cosh_FR_tmp = cosh_FR_all_ones, cosh_FR_all_ones
1075 (p0) fmerge.s f8 = f44,f44
1076 (p0) br.ret.sptk b0 ;;
1080 // for COSH_HUGE, put 24000 in exponent; take sign from input; add 1
1081 // SAFE: SAFE is always 0 for HUGE
1087 (p0) movl r32 = 0x0000000000015dbf ;;
1091 (p0) setf.exp f9 = r32
1098 (p0) fma.s1 cosh_FR_hi_lo = f1, f9, f1
1104 (p0) fma.s0 f44 = f9, cosh_FR_hi_lo, f0
1105 (p0) mov GR_Parameter_TAG = 63
1108 ASM_SIZE_DIRECTIVE(coshl)
1110 .proc __libm_error_region
1111 __libm_error_region:
1114 add GR_Parameter_Y=-32,sp // Parameter 2 value
1116 .save ar.pfs,GR_SAVE_PFS
1117 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
1121 add sp=-64,sp // Create new stack
1123 mov GR_SAVE_GP=gp // Save gp
1126 stfe [GR_Parameter_Y] = FR_Y,16 // Save Parameter 2 on stack
1127 add GR_Parameter_X = 16,sp // Parameter 1 address
1128 .save b0, GR_SAVE_B0
1129 mov GR_SAVE_B0=b0 // Save b0
1133 stfe [GR_Parameter_X] = FR_X // Store Parameter 1 on stack
1134 add GR_Parameter_RESULT = 0,GR_Parameter_Y
1135 nop.b 0 // Parameter 3 address
1138 stfe [GR_Parameter_Y] = FR_RESULT // Store Parameter 3 on stack
1139 add GR_Parameter_Y = -16,GR_Parameter_Y
1140 br.call.sptk b0=__libm_error_support# // Call error handling function
1145 add GR_Parameter_RESULT = 48,sp
1148 ldfe f8 = [GR_Parameter_RESULT] // Get return result off stack
1150 add sp = 64,sp // Restore stack pointer
1151 mov b0 = GR_SAVE_B0 // Restore return address
1154 mov gp = GR_SAVE_GP // Restore gp
1155 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
1156 br.ret.sptk b0 // Return
1159 .endp __libm_error_region
1160 ASM_SIZE_DIRECTIVE(__libm_error_region)
1162 .type __libm_error_support#,@function
1163 .global __libm_error_support#