4 // Copyright (c) 2000 - 2005, 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 // 05/07/01 Reworked to improve speed of all paths
46 // 05/20/02 Cleaned up namespace and sf0 syntax
47 // 11/15/02 Improved speed with new algorithm
48 // 03/31/05 Reformatted delimiters between data tables
51 //==============================================================
52 // double cosh(double)
54 // Overview of operation
55 //==============================================================
56 // Case 1: 0 < |x| < 0.25
57 // Evaluate cosh(x) by a 12th order polynomial
58 // Care is take for the order of multiplication; and A2 is not exactly 1/4!,
59 // A3 is not exactly 1/6!, etc.
60 // cosh(x) = 1 + (A1*x^2 + A2*x^4 + A3*x^6 + A4*x^8 + A5*x^10 + A6*x^12)
62 // Case 2: 0.25 < |x| < 710.47586
63 // Algorithm is based on the identity cosh(x) = ( exp(x) + exp(-x) ) / 2.
64 // The algorithm for exp is described as below. There are a number of
65 // economies from evaluating both exp(x) and exp(-x). Although we
66 // are evaluating both quantities, only where the quantities diverge do we
67 // duplicate the computations. The basic algorithm for exp(x) is described
70 // Take the input x. w is "how many log2/128 in x?"
73 // x = n log2/128 + r + delta
75 // n = 128M + index_1 + 2^4 index_2
76 // x = M log2 + (log2/128) index_1 + (log2/8) index_2 + r + delta
78 // exp(x) = 2^M 2^(index_1/128) 2^(index_2/8) exp(r) exp(delta)
80 // Get 2^(index_1/128) from table_1;
81 // Get 2^(index_2/8) from table_2;
82 // Calculate exp(r) by 5th order polynomial
83 // r = x - n (log2/128)_high
84 // delta = - n (log2/128)_low
85 // Calculate exp(delta) as 1 + delta
89 //==============================================================
93 // cosh(+qnan) = +qnan
94 // cosh(-qnan) = -qnan
95 // cosh(+snan) = +qnan
96 // cosh(-snan) = -qnan
101 // Overflow and Underflow
102 //=======================
103 // cosh(x) = largest double normal when
104 // x = 710.47586 = 0x408633ce8fb9f87d
106 // There is no underflow.
109 //==============================================================
110 // Floating Point registers used:
112 // f6 -> f15, f32 -> f61
114 // General registers used:
117 // Predicate registers used:
121 //==============================================================
136 rExp_bias_minus_1 = r23
140 rIndex_2_16_neg = r24
159 GR_Parameter_RESULT = r39
160 GR_Parameter_TAG = r40
195 fMIN_DBL_OFLOW_ARG = f45
196 fMAX_DBL_NORM_ARG = f46
234 //==============================================================
239 // ************* DO NOT CHANGE ORDER OF THESE TABLES ********************
241 // double-extended 1/ln(2)
242 // 3fff b8aa 3b29 5c17 f0bb be87fed0691d3e88
243 // 3fff b8aa 3b29 5c17 f0bc
244 // For speed the significand will be loaded directly with a movl and setf.sig
245 // and the exponent will be bias+63 instead of bias+0. Thus subsequent
246 // computations need to scale appropriately.
247 // The constant 128/ln(2) is needed for the computation of w. This is also
248 // obtained by scaling the computations.
250 // Two shifting constants are loaded directly with movl and setf.d.
251 // 1. fRSHF_2TO56 = 1.1000..00 * 2^(63-7)
252 // This constant is added to x*1/ln2 to shift the integer part of
253 // x*128/ln2 into the rightmost bits of the significand.
254 // The result of this fma is fW_2TO56_RSH.
255 // 2. fRSHF = 1.1000..00 * 2^(63)
256 // This constant is subtracted from fW_2TO56_RSH * 2^(-56) to give
257 // the integer part of w, n, as a floating-point number.
258 // The result of this fms is fNfloat.
261 LOCAL_OBJECT_START(exp_table_1)
262 data8 0x408633ce8fb9f87e // smallest dbl overflow arg
263 data8 0x408633ce8fb9f87d // largest dbl arg to give normal dbl result
264 data8 0xb17217f7d1cf79ab , 0x00003ff7 // ln2/128 hi
265 data8 0xc9e3b39803f2f6af , 0x00003fb7 // ln2/128 lo
267 // Table 1 is 2^(index_1/128) where
268 // index_1 goes from 0 to 15
270 data8 0x8000000000000000 , 0x00003FFF
271 data8 0x80B1ED4FD999AB6C , 0x00003FFF
272 data8 0x8164D1F3BC030773 , 0x00003FFF
273 data8 0x8218AF4373FC25EC , 0x00003FFF
274 data8 0x82CD8698AC2BA1D7 , 0x00003FFF
275 data8 0x8383594EEFB6EE37 , 0x00003FFF
276 data8 0x843A28C3ACDE4046 , 0x00003FFF
277 data8 0x84F1F656379C1A29 , 0x00003FFF
278 data8 0x85AAC367CC487B15 , 0x00003FFF
279 data8 0x8664915B923FBA04 , 0x00003FFF
280 data8 0x871F61969E8D1010 , 0x00003FFF
281 data8 0x87DB357FF698D792 , 0x00003FFF
282 data8 0x88980E8092DA8527 , 0x00003FFF
283 data8 0x8955EE03618E5FDD , 0x00003FFF
284 data8 0x8A14D575496EFD9A , 0x00003FFF
285 data8 0x8AD4C6452C728924 , 0x00003FFF
286 LOCAL_OBJECT_END(exp_table_1)
288 // Table 2 is 2^(index_1/8) where
289 // index_2 goes from 0 to 7
290 LOCAL_OBJECT_START(exp_table_2)
291 data8 0x8000000000000000 , 0x00003FFF
292 data8 0x8B95C1E3EA8BD6E7 , 0x00003FFF
293 data8 0x9837F0518DB8A96F , 0x00003FFF
294 data8 0xA5FED6A9B15138EA , 0x00003FFF
295 data8 0xB504F333F9DE6484 , 0x00003FFF
296 data8 0xC5672A115506DADD , 0x00003FFF
297 data8 0xD744FCCAD69D6AF4 , 0x00003FFF
298 data8 0xEAC0C6E7DD24392F , 0x00003FFF
299 LOCAL_OBJECT_END(exp_table_2)
301 LOCAL_OBJECT_START(exp_p_table)
302 data8 0x3f8111116da21757 //P5
303 data8 0x3fa55555d787761c //P4
304 data8 0x3fc5555555555414 //P3
305 data8 0x3fdffffffffffd6a //P2
306 LOCAL_OBJECT_END(exp_p_table)
308 LOCAL_OBJECT_START(cosh_p_table)
309 data8 0x8FA02AC65BCBD5BC, 0x00003FE2 // A6
310 data8 0xD00D00D1021D7370, 0x00003FEF // A4
311 data8 0xAAAAAAAAAAAAAB80, 0x00003FFA // A2
312 data8 0x93F27740C0C2F1CC, 0x00003FE9 // A5
313 data8 0xB60B60B60B4FE884, 0x00003FF5 // A3
314 data8 0x8000000000000000, 0x00003FFE // A1
315 LOCAL_OBJECT_END(cosh_p_table)
319 GLOBAL_IEEE754_ENTRY(cosh)
322 getf.exp rSignexp_x = f8 // Must recompute if x unorm
323 movl rSig_inv_ln2 = 0xb8aa3b295c17f0bc // significand of 1/ln2
326 addl rAD_TB1 = @ltoff(exp_table_1), gp
327 movl rRshf_2to56 = 0x4768000000000000 // 1.10000 2^(63+56)
332 ld8 rAD_TB1 = [rAD_TB1]
333 fclass.m p6,p0 = f8,0x0b // Test for x=unorm
334 mov rExp_mask = 0x1ffff
337 mov rExp_bias = 0xffff
339 mov rExp_2tom56 = 0xffff-56
343 // Form two constants we need
344 // 1/ln2 * 2^63 to compute w = x * 1/ln2 * 128
345 // 1.1000..000 * 2^(63+63-7) to right shift int(w) into the significand
348 setf.sig fINV_LN2_2TO63 = rSig_inv_ln2 // form 1/ln2 * 2^63
349 fclass.m p8,p0 = f8,0x07 // Test for x=0
353 setf.d fRSHF_2TO56 = rRshf_2to56 // Form const 1.100 * 2^(63+56)
354 movl rRshf = 0x43e8000000000000 // 1.10000 2^63 for right shift
359 ldfpd fMIN_DBL_OFLOW_ARG, fMAX_DBL_NORM_ARG = [rAD_TB1],16
360 fclass.m p10,p0 = f8,0x1e3 // Test for x=inf, nan, NaT
364 setf.exp f2TOM56 = rExp_2tom56 // form 2^-56 for scaling Nfloat
366 (p6) br.cond.spnt COSH_UNORM // Branch if x=unorm
372 ldfe fLn2_by_128_hi = [rAD_TB1],16
377 setf.d fRSHF = rRshf // Form right shift const 1.100 * 2^63
378 (p8) fma.d.s0 f8 = f1,f1,f0 // quick exit for x=0
384 ldfe fLn2_by_128_lo = [rAD_TB1],16
389 and rExp_x = rExp_mask, rSignexp_x // Biased exponent of x
390 (p10) fma.d.s0 f8 = f8,f8,f0 // Result if x=inf, nan, NaT
391 (p10) br.ret.spnt b0 // quick exit for x=inf, nan, NaT
395 // After that last load rAD_TB1 points to the beginning of table 1
398 fcmp.eq.s0 p6,p0 = f8, f0 // Dummy to set D
399 sub rExp_x = rExp_x, rExp_bias // True exponent of x
405 fmerge.s fAbsX = f0, fNormX // Form |x|
409 cmp.gt p7, p0 = -2, rExp_x // Test |x| < 2^(-2)
410 fma.s1 fXsq = fNormX, fNormX, f0 // x*x for small path
411 (p7) br.cond.spnt COSH_SMALL // Branch if 0 < |x| < 2^-2
415 // W = X * Inv_log2_by_128
416 // By adding 1.10...0*2^63 we shift and get round_int(W) in significand.
417 // We actually add 1.10...0*2^56 to X * Inv_log2 to do the same thing.
420 add rAD_P = 0x180, rAD_TB1
421 fma.s1 fW_2TO56_RSH = fNormX, fINV_LN2_2TO63, fRSHF_2TO56
422 add rAD_TB2 = 0x100, rAD_TB1
426 // Divide arguments into the following categories:
427 // Certain Safe - 0.25 <= |x| <= MAX_DBL_NORM_ARG
428 // Possible Overflow p14 - MAX_DBL_NORM_ARG < |x| < MIN_DBL_OFLOW_ARG
429 // Certain Overflow p15 - MIN_DBL_OFLOW_ARG <= |x| < +inf
431 // If the input is really a double arg, then there will never be
432 // "Possible Overflow" arguments.
436 ldfpd fP5, fP4 = [rAD_P] ,16
437 fcmp.ge.s1 p15,p14 = fAbsX,fMIN_DBL_OFLOW_ARG
442 // Nfloat = round_int(W)
443 // The signficand of fW_2TO56_RSH contains the rounded integer part of W,
444 // as a twos complement number in the lower bits (that is, it may be negative).
445 // That twos complement number (called N) is put into rN.
447 // Since fW_2TO56_RSH is scaled by 2^56, it must be multiplied by 2^-56
448 // before the shift constant 1.10000 * 2^63 is subtracted to yield fNfloat.
449 // Thus, fNfloat contains the floating point version of N
452 ldfpd fP3, fP2 = [rAD_P]
453 (p14) fcmp.gt.unc.s1 p14,p0 = fAbsX,fMAX_DBL_NORM_ARG
458 fms.s1 fNfloat = fW_2TO56_RSH, f2TOM56, fRSHF
459 (p15) br.cond.spnt COSH_CERTAIN_OVERFLOW
464 getf.sig rN = fW_2TO56_RSH
466 mov rExp_bias_minus_1 = 0xfffe
470 // rIndex_1 has index_1
471 // rIndex_2_16 has index_2 * 16
475 // r = x - Nfloat * ln2_by_128_hi
476 // f = 1 - Nfloat * ln2_by_128_lo
478 and rIndex_1 = 0x0f, rN
479 fnma.s1 fR = fNfloat, fLn2_by_128_hi, fNormX
483 and rIndex_2_16 = 0x70, rN
484 fnma.s1 fF = fNfloat, fLn2_by_128_lo, f1
490 and rIndex_1_neg = 0x0f, rN_neg
491 add rBiased_M = rExp_bias_minus_1, rM
492 shr rM_neg = rN_neg, 0x7
495 and rIndex_2_16_neg = 0x70, rN_neg
496 add rAD_T2 = rAD_TB2, rIndex_2_16
497 shladd rAD_T1 = rIndex_1, 4, rAD_TB1
501 // rAD_T1 has address of T1
502 // rAD_T2 has address if T2
505 setf.exp f2M = rBiased_M
510 add rBiased_M_neg = rExp_bias_minus_1, rM_neg
511 add rAD_T2_neg = rAD_TB2, rIndex_2_16_neg
512 shladd rAD_T1_neg = rIndex_1_neg, 4, rAD_TB1
516 // Create Scale = 2^M
524 setf.exp f2M_neg = rBiased_M_neg
525 ldfe fT2_neg = [rAD_T2_neg]
526 fma.s1 fF_neg = fNfloat, fLn2_by_128_lo, f1
532 fma.s1 fRsq = fR, fR, f0
536 ldfe fT1_neg = [rAD_T1_neg]
537 fma.s1 fP54 = fR, fP5, fP4
544 fma.s1 fP32 = fR, fP3, fP2
549 fnma.s1 fP54_neg = fR, fP5, fP4
556 fnma.s1 fP32_neg = fR, fP3, fP2
563 fma.s1 fP5432 = fRsq, fP54, fP32
568 fma.s1 fS2 = fF,fT2,f0
575 fma.s1 fS1 = f2M,fT1,f0
580 fma.s1 fP5432_neg = fRsq, fP54_neg, fP32_neg
587 fma.s1 fS1_neg = f2M_neg,fT1_neg,f0
592 fma.s1 fS2_neg = fF_neg,fT2_neg,f0
599 fma.s1 fP = fRsq, fP5432, fR
604 fma.s1 fS = fS1,fS2,f0
611 fms.s1 fP_neg = fRsq, fP5432_neg, fR
616 fma.s1 fS_neg = fS1_neg,fS2_neg,f0
623 fmpy.s0 fTmp = fLn2_by_128_lo, fLn2_by_128_lo // Force inexact
624 (p14) br.cond.spnt COSH_POSSIBLE_OVERFLOW
630 fma.s1 fExp = fS, fP, fS
635 fma.s1 fExp_neg = fS_neg, fP_neg, fS_neg
642 fma.d.s0 f8 = fExp, f1, fExp_neg
643 br.ret.sptk b0 // Normal path exit
647 // Here if 0 < |x| < 0.25
650 add rAD_T1 = 0x1a0, rAD_TB1
651 add rAD_T2 = 0x1d0, rAD_TB1
656 ldfe fA6 = [rAD_T1],16
657 ldfe fA5 = [rAD_T2],16
663 ldfe fA4 = [rAD_T1],16
664 ldfe fA3 = [rAD_T2],16
670 ldfe fA2 = [rAD_T1],16
671 ldfe fA1 = [rAD_T2],16
678 fma.s1 fX4 = fXsq, fXsq, f0
685 fma.s1 fA65 = fXsq, fA6, fA5
690 fma.s1 fA43 = fXsq, fA4, fA3
697 fma.s1 fA21 = fXsq, fA2, fA1
704 fma.s1 fA6543 = fX4, fA65, fA43
711 fma.s1 fA654321 = fX4, fA6543, fA21
716 // Dummy multiply to generate inexact
719 fmpy.s0 fTmp = fA6, fA6
724 fma.d.s0 f8 = fA654321, fXsq, f1
725 br.ret.sptk b0 // Exit if 0 < |x| < 0.25
730 COSH_POSSIBLE_OVERFLOW:
732 // Here if fMAX_DBL_NORM_ARG < |x| < fMIN_DBL_OFLOW_ARG
733 // This cannot happen if input is a double, only if input higher precision.
734 // Overflow is a possibility, not a certainty.
736 // Recompute result using status field 2 with user's rounding mode,
737 // and wre set. If result is larger than largest double, then we have
741 mov rGt_ln = 0x103ff // Exponent for largest dbl + 1 ulp
742 fsetc.s2 0x7F,0x42 // Get user's round mode, set wre
748 setf.exp fGt_pln = rGt_ln // Create largest double + 1 ulp
749 fma.d.s2 fWre_urm_f8 = fS, fP, fS // Result with wre set
756 fsetc.s2 0x7F,0x40 // Turn off wre in sf2
763 fcmp.ge.s1 p6, p0 = fWre_urm_f8, fGt_pln // Test for overflow
771 (p6) br.cond.spnt COSH_CERTAIN_OVERFLOW // Branch if overflow
777 fma.d.s0 f8 = fS, fP, fS
778 br.ret.sptk b0 // Exit if really no overflow
782 COSH_CERTAIN_OVERFLOW:
784 sub rTmp = rExp_mask, r0, 1
792 alloc r32=ar.pfs,1,4,4,0
793 fmerge.s FR_X = f8,f8
797 mov GR_Parameter_TAG = 64
798 fma.d.s0 FR_RESULT = fTmp, fTmp, f0 // Set I,O and +INF result
799 br.cond.sptk __libm_error_region
806 getf.exp rSignexp_x = fNormX // Must recompute if x unorm
807 fcmp.eq.s0 p6, p0 = f8, f0 // Set D flag
808 br.cond.sptk COSH_COMMON
812 GLOBAL_IEEE754_END(cosh)
813 libm_alias_double_other (__cosh, cosh)
816 LOCAL_LIBM_ENTRY(__libm_error_region)
819 add GR_Parameter_Y=-32,sp // Parameter 2 value
821 .save ar.pfs,GR_SAVE_PFS
822 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
826 add sp=-64,sp // Create new stack
828 mov GR_SAVE_GP=gp // Save gp
831 stfd [GR_Parameter_Y] = FR_Y,16 // STORE Parameter 2 on stack
832 add GR_Parameter_X = 16,sp // Parameter 1 address
834 mov GR_SAVE_B0=b0 // Save b0
838 stfd [GR_Parameter_X] = FR_X // STORE Parameter 1 on stack
839 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
843 stfd [GR_Parameter_Y] = FR_RESULT // STORE Parameter 3 on stack
844 add GR_Parameter_Y = -16,GR_Parameter_Y
845 br.call.sptk b0=__libm_error_support# // Call error handling function
848 add GR_Parameter_RESULT = 48,sp
853 ldfd f8 = [GR_Parameter_RESULT] // Get return result off stack
855 add sp = 64,sp // Restore stack pointer
856 mov b0 = GR_SAVE_B0 // Restore return address
859 mov gp = GR_SAVE_GP // Restore gp
860 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
861 br.ret.sptk b0 // Return
864 LOCAL_LIBM_END(__libm_error_region)
865 .type __libm_error_support#,@function
866 .global __libm_error_support#