github.com/kdevb0x/go@v0.0.0-20180115030120-39687051e9e7/src/math/cosh_s390x.s (about)

     1  // Copyright 2016 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  #include "textflag.h"
     6  
     7  // Constants
     8  DATA coshrodataL23<>+0(SB)/8, $0.231904681384629956E-16
     9  DATA coshrodataL23<>+8(SB)/8, $0.693147180559945286E+00
    10  DATA coshrodataL23<>+16(SB)/8, $0.144269504088896339E+01
    11  DATA coshrodataL23<>+24(SB)/8, $704.E0
    12  GLOBL coshrodataL23<>+0(SB), RODATA, $32
    13  DATA coshxinf<>+0(SB)/8, $0x7FF0000000000000
    14  GLOBL coshxinf<>+0(SB), RODATA, $8
    15  DATA coshxlim1<>+0(SB)/8, $800.E0
    16  GLOBL coshxlim1<>+0(SB), RODATA, $8
    17  DATA coshxaddhy<>+0(SB)/8, $0xc2f0000100003fdf
    18  GLOBL coshxaddhy<>+0(SB), RODATA, $8
    19  DATA coshx4ff<>+0(SB)/8, $0x4ff0000000000000
    20  GLOBL coshx4ff<>+0(SB), RODATA, $8
    21  DATA coshe1<>+0(SB)/8, $0x3ff000000000000a
    22  GLOBL coshe1<>+0(SB), RODATA, $8
    23  
    24  // Log multiplier table
    25  DATA coshtab<>+0(SB)/8, $0.442737824274138381E-01
    26  DATA coshtab<>+8(SB)/8, $0.263602189790660309E-01
    27  DATA coshtab<>+16(SB)/8, $0.122565642281703586E-01
    28  DATA coshtab<>+24(SB)/8, $0.143757052860721398E-02
    29  DATA coshtab<>+32(SB)/8, $-.651375034121276075E-02
    30  DATA coshtab<>+40(SB)/8, $-.119317678849450159E-01
    31  DATA coshtab<>+48(SB)/8, $-.150868749549871069E-01
    32  DATA coshtab<>+56(SB)/8, $-.161992609578469234E-01
    33  DATA coshtab<>+64(SB)/8, $-.154492360403337917E-01
    34  DATA coshtab<>+72(SB)/8, $-.129850717389178721E-01
    35  DATA coshtab<>+80(SB)/8, $-.892902649276657891E-02
    36  DATA coshtab<>+88(SB)/8, $-.338202636596794887E-02
    37  DATA coshtab<>+96(SB)/8, $0.357266307045684762E-02
    38  DATA coshtab<>+104(SB)/8, $0.118665304327406698E-01
    39  DATA coshtab<>+112(SB)/8, $0.214434994118118914E-01
    40  DATA coshtab<>+120(SB)/8, $0.322580645161290314E-01
    41  GLOBL coshtab<>+0(SB), RODATA, $128
    42  
    43  // Minimax polynomial approximations
    44  DATA coshe2<>+0(SB)/8, $0.500000000000004237e+00
    45  GLOBL coshe2<>+0(SB), RODATA, $8
    46  DATA coshe3<>+0(SB)/8, $0.166666666630345592e+00
    47  GLOBL coshe3<>+0(SB), RODATA, $8
    48  DATA coshe4<>+0(SB)/8, $0.416666664838056960e-01
    49  GLOBL coshe4<>+0(SB), RODATA, $8
    50  DATA coshe5<>+0(SB)/8, $0.833349307718286047e-02
    51  GLOBL coshe5<>+0(SB), RODATA, $8
    52  DATA coshe6<>+0(SB)/8, $0.138926439368309441e-02
    53  GLOBL coshe6<>+0(SB), RODATA, $8
    54  
    55  // Cosh returns the hyperbolic cosine of x.
    56  //
    57  // Special cases are:
    58  //      Cosh(±0) = 1
    59  //      Cosh(±Inf) = +Inf
    60  //      Cosh(NaN) = NaN
    61  // The algorithm used is minimax polynomial approximation
    62  // with coefficients determined with a Remez exchange algorithm.
    63  
    64  TEXT ·coshAsm(SB),NOSPLIT,$0-16
    65  	FMOVD   x+0(FP), F0
    66  	MOVD    $coshrodataL23<>+0(SB), R9
    67  	WORD    $0xB3120000     //ltdbr %f0,%f0
    68  	MOVD    $0x4086000000000000, R2
    69  	MOVD    $0x4086000000000000, R3
    70  	BLTU    L19
    71  	FMOVD   F0, F4
    72  L2:
    73  	WORD    $0xED409018     //cdb %f4,.L24-.L23(%r9)
    74  	BYTE    $0x00
    75  	BYTE    $0x19
    76  	BGE     L14     //jnl   .L14
    77  	BVS     L14
    78  	WFCEDBS V4, V4, V2
    79  	BEQ     L20
    80  L1:
    81  	FMOVD   F0, ret+8(FP)
    82  	RET
    83  
    84  L14:
    85  	WFCEDBS V4, V4, V2
    86  	BVS     L1
    87  	MOVD    $coshxlim1<>+0(SB), R1
    88  	FMOVD   0(R1), F2
    89  	WFCHEDBS        V4, V2, V2
    90  	BEQ     L21
    91  	MOVD    $coshxaddhy<>+0(SB), R1
    92  	FMOVD   coshrodataL23<>+16(SB), F5
    93  	FMOVD   0(R1), F2
    94  	WFMSDB  V0, V5, V2, V5
    95  	FMOVD   coshrodataL23<>+8(SB), F3
    96  	FADD    F5, F2
    97  	MOVD    $coshe6<>+0(SB), R1
    98  	WFMSDB  V2, V3, V0, V3
    99  	FMOVD   0(R1), F6
   100  	WFMDB   V3, V3, V1
   101  	MOVD    $coshe4<>+0(SB), R1
   102  	FMOVD   coshrodataL23<>+0(SB), F7
   103  	WFMADB  V2, V7, V3, V2
   104  	FMOVD   0(R1), F3
   105  	MOVD    $coshe5<>+0(SB), R1
   106  	WFMADB  V1, V6, V3, V6
   107  	FMOVD   0(R1), F7
   108  	MOVD    $coshe3<>+0(SB), R1
   109  	FMOVD   0(R1), F3
   110  	WFMADB  V1, V7, V3, V7
   111  	FNEG    F2, F3
   112  	LGDR    F5, R1
   113  	MOVD    $coshe2<>+0(SB), R3
   114  	WFCEDBS V4, V0, V0
   115  	FMOVD   0(R3), F5
   116  	MOVD    $coshe1<>+0(SB), R3
   117  	WFMADB  V1, V6, V5, V6
   118  	FMOVD   0(R3), F5
   119  	WORD    $0xEC21000F     //risbgn %r2,%r1,64-64+0,64-64+0+16-1,64-0-16
   120  	BYTE    $0x30
   121  	BYTE    $0x59
   122  	WFMADB  V1, V7, V5, V1
   123  	BVS     L22
   124  	WORD    $0xEC4139BC     //risbg %r4,%r1,57,128+60,3
   125  	BYTE    $0x03
   126  	BYTE    $0x55
   127  	MOVD    $coshtab<>+0(SB), R3
   128  	WFMADB  V3, V6, V1, V6
   129  	WORD    $0x68043000     //ld    %f0,0(%r4,%r3)
   130  	FMSUB   F0, F3, F2
   131  	WORD    $0xA71AF000     //ahi   %r1,-4096
   132  	WFMADB  V2, V6, V0, V6
   133  L17:
   134  	WORD    $0xEC21000F     //risbgn %r2,%r1,64-64+0,64-64+0+16-1,64-0-16
   135  	BYTE    $0x30
   136  	BYTE    $0x59
   137  	LDGR    R2, F2
   138  	FMADD   F2, F6, F2
   139  	MOVD    $coshx4ff<>+0(SB), R1
   140  	FMOVD   0(R1), F0
   141  	FMUL    F2, F0
   142  	FMOVD   F0, ret+8(FP)
   143  	RET
   144  
   145  L19:
   146  	FNEG    F0, F4
   147  	BR      L2
   148  L20:
   149  	MOVD    $coshxaddhy<>+0(SB), R1
   150  	FMOVD   coshrodataL23<>+16(SB), F3
   151  	FMOVD   0(R1), F2
   152  	WFMSDB  V0, V3, V2, V3
   153  	FMOVD   coshrodataL23<>+8(SB), F4
   154  	FADD    F3, F2
   155  	MOVD    $coshe6<>+0(SB), R1
   156  	FMSUB   F4, F2, F0
   157  	FMOVD   0(R1), F6
   158  	WFMDB   V0, V0, V1
   159  	MOVD    $coshe4<>+0(SB), R1
   160  	FMOVD   0(R1), F4
   161  	MOVD    $coshe5<>+0(SB), R1
   162  	FMOVD   coshrodataL23<>+0(SB), F5
   163  	WFMADB  V1, V6, V4, V6
   164  	FMADD   F5, F2, F0
   165  	FMOVD   0(R1), F2
   166  	MOVD    $coshe3<>+0(SB), R1
   167  	FMOVD   0(R1), F4
   168  	WFMADB  V1, V2, V4, V2
   169  	MOVD    $coshe2<>+0(SB), R1
   170  	FMOVD   0(R1), F5
   171  	FNEG    F0, F4
   172  	WFMADB  V1, V6, V5, V6
   173  	MOVD    $coshe1<>+0(SB), R1
   174  	FMOVD   0(R1), F5
   175  	WFMADB  V1, V2, V5, V1
   176  	LGDR    F3, R1
   177  	MOVD    $coshtab<>+0(SB), R5
   178  	WFMADB  V4, V6, V1, V3
   179  	WORD    $0xEC4139BC     //risbg %r4,%r1,57,128+60,3
   180  	BYTE    $0x03
   181  	BYTE    $0x55
   182  	WFMSDB  V4, V6, V1, V6
   183  	WORD    $0x68145000     //ld %f1,0(%r4,%r5)
   184  	WFMSDB  V4, V1, V0, V2
   185  	WORD    $0xA7487FBE     //lhi %r4,32702
   186  	FMADD   F3, F2, F1
   187  	SUBW    R1, R4
   188  	WORD    $0xECC439BC     //risbg %r12,%r4,57,128+60,3
   189  	BYTE    $0x03
   190  	BYTE    $0x55
   191  	WORD    $0x682C5000     //ld %f2,0(%r12,%r5)
   192  	FMSUB   F2, F4, F0
   193  	WORD    $0xEC21000F     //risbgn %r2,%r1,64-64+0,64-64+0+16-1,64-0-16
   194  	BYTE    $0x30
   195  	BYTE    $0x59
   196  	WFMADB  V0, V6, V2, V6
   197  	WORD    $0xEC34000F     //risbgn %r3,%r4,64-64+0,64-64+0+16-1,64-0-16
   198  	BYTE    $0x30
   199  	BYTE    $0x59
   200  	LDGR    R2, F2
   201  	LDGR    R3, F0
   202  	FMADD   F2, F1, F2
   203  	FMADD   F0, F6, F0
   204  	FADD    F2, F0
   205  	FMOVD   F0, ret+8(FP)
   206  	RET
   207  
   208  L22:
   209  	WORD    $0xA7387FBE     //lhi %r3,32702
   210  	MOVD    $coshtab<>+0(SB), R4
   211  	SUBW    R1, R3
   212  	WFMSDB  V3, V6, V1, V6
   213  	WORD    $0xEC3339BC     //risbg %r3,%r3,57,128+60,3
   214  	BYTE    $0x03
   215  	BYTE    $0x55
   216  	WORD    $0x68034000     //ld %f0,0(%r3,%r4)
   217  	FMSUB   F0, F3, F2
   218  	WORD    $0xA7386FBE     //lhi %r3,28606
   219  	WFMADB  V2, V6, V0, V6
   220  	SUBW    R1, R3, R1
   221  	BR      L17
   222  L21:
   223  	MOVD    $coshxinf<>+0(SB), R1
   224  	FMOVD   0(R1), F0
   225  	FMOVD   F0, ret+8(FP)
   226  	RET
   227