github.com/likebike/go--@v0.0.0-20190911215757-0bd925d16e96/go/src/math/tanh_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  // Minimax polynomial approximations
     8  DATA tanhrodataL18<>+0(SB)/8, $-1.0
     9  DATA tanhrodataL18<>+8(SB)/8, $-2.0
    10  DATA tanhrodataL18<>+16(SB)/8, $1.0
    11  DATA tanhrodataL18<>+24(SB)/8, $2.0
    12  DATA tanhrodataL18<>+32(SB)/8, $0.20000000000000011868E+01
    13  DATA tanhrodataL18<>+40(SB)/8, $0.13333333333333341256E+01
    14  DATA tanhrodataL18<>+48(SB)/8, $0.26666666663549111502E+00
    15  DATA tanhrodataL18<>+56(SB)/8, $0.66666666658721844678E+00
    16  DATA tanhrodataL18<>+64(SB)/8, $0.88890217768964374821E-01
    17  DATA tanhrodataL18<>+72(SB)/8, $0.25397199429103821138E-01
    18  DATA tanhrodataL18<>+80(SB)/8, $-.346573590279972643E+00
    19  DATA tanhrodataL18<>+88(SB)/8, $20.E0
    20  GLOBL tanhrodataL18<>+0(SB), RODATA, $96
    21  
    22  // Constants
    23  DATA tanhrlog2<>+0(SB)/8, $0x4007154760000000
    24  GLOBL tanhrlog2<>+0(SB), RODATA, $8
    25  DATA tanhxadd<>+0(SB)/8, $0xc2f0000100003ff0
    26  GLOBL tanhxadd<>+0(SB), RODATA, $8
    27  DATA tanhxmone<>+0(SB)/8, $-1.0
    28  GLOBL tanhxmone<>+0(SB), RODATA, $8
    29  DATA tanhxzero<>+0(SB)/8, $0
    30  GLOBL tanhxzero<>+0(SB), RODATA, $8
    31  
    32  // Polynomial coefficients
    33  DATA tanhtab<>+0(SB)/8, $0.000000000000000000E+00
    34  DATA tanhtab<>+8(SB)/8, $-.171540871271399150E-01
    35  DATA tanhtab<>+16(SB)/8, $-.306597931864376363E-01
    36  DATA tanhtab<>+24(SB)/8, $-.410200970469965021E-01
    37  DATA tanhtab<>+32(SB)/8, $-.486343079978231466E-01
    38  DATA tanhtab<>+40(SB)/8, $-.538226193725835820E-01
    39  DATA tanhtab<>+48(SB)/8, $-.568439602538111520E-01
    40  DATA tanhtab<>+56(SB)/8, $-.579091847395528847E-01
    41  DATA tanhtab<>+64(SB)/8, $-.571909584179366341E-01
    42  DATA tanhtab<>+72(SB)/8, $-.548312665987204407E-01
    43  DATA tanhtab<>+80(SB)/8, $-.509471843643441085E-01
    44  DATA tanhtab<>+88(SB)/8, $-.456353588448863359E-01
    45  DATA tanhtab<>+96(SB)/8, $-.389755254243262365E-01
    46  DATA tanhtab<>+104(SB)/8, $-.310332908285244231E-01
    47  DATA tanhtab<>+112(SB)/8, $-.218623539150173528E-01
    48  DATA tanhtab<>+120(SB)/8, $-.115062908917949451E-01
    49  GLOBL tanhtab<>+0(SB), RODATA, $128
    50  
    51  // Tanh returns the hyperbolic tangent of the argument.
    52  //
    53  // Special cases are:
    54  //      Tanh(±0) = ±0
    55  //      Tanh(±Inf) = ±1
    56  //      Tanh(NaN) = NaN
    57  // The algorithm used is minimax polynomial approximation using a table of
    58  // polynomial coefficients determined with a Remez exchange algorithm.
    59  
    60  TEXT ·tanhAsm(SB),NOSPLIT,$0-16
    61  	FMOVD   x+0(FP), F0
    62  	// special case Tanh(±0) = ±0
    63  	FMOVD   $(0.0), F1
    64  	FCMPU   F0, F1
    65  	BEQ     tanhIsZero
    66  	MOVD    $tanhrodataL18<>+0(SB), R5
    67  	WORD    $0xB3120000     //ltdbr %f0,%f0
    68  	MOVD    $0x4034000000000000, R1
    69  	BLTU    L15
    70  	FMOVD   F0, F1
    71  L2:
    72  	MOVD    $tanhxadd<>+0(SB), R2
    73  	FMOVD   0(R2), F2
    74  	MOVD    tanhrlog2<>+0(SB), R2
    75  	LDGR    R2, F4
    76  	WFMSDB  V0, V4, V2, V4
    77  	MOVD    $tanhtab<>+0(SB), R3
    78  	LGDR    F4, R2
    79  	WORD    $0xEC4239BC     //risbg %r4,%r2,57,128+60,3
    80  	BYTE    $0x03
    81  	BYTE    $0x55
    82  	WORD    $0xED105058     //cdb %f1,.L19-.L18(%r5)
    83  	BYTE    $0x00
    84  	BYTE    $0x19
    85  	WORD    $0xEC12000F     //risbgn %r1,%r2,64-64+0,64-64+0+16-1,64-0-16
    86  	BYTE    $0x30
    87  	BYTE    $0x59
    88  	WORD    $0x68543000     //ld %f5,0(%r4,%r3)
    89  	LDGR    R1, F6
    90  	BLT     L3
    91  	MOVD    $tanhxzero<>+0(SB), R1
    92  	FMOVD   0(R1), F2
    93  	WFCHDBS V0, V2, V4
    94  	BEQ     L9
    95  	WFCHDBS V2, V0, V2
    96  	BNE     L1
    97  	MOVD    $tanhxmone<>+0(SB), R1
    98  	FMOVD   0(R1), F0
    99  	FMOVD   F0, ret+8(FP)
   100  	RET
   101  
   102  L3:
   103  	FADD    F4, F2
   104  	FMOVD   tanhrodataL18<>+80(SB), F4
   105  	FMADD   F4, F2, F0
   106  	FMOVD   tanhrodataL18<>+72(SB), F1
   107  	WFMDB   V0, V0, V3
   108  	FMOVD   tanhrodataL18<>+64(SB), F2
   109  	WFMADB  V0, V1, V2, V1
   110  	FMOVD   tanhrodataL18<>+56(SB), F4
   111  	FMOVD   tanhrodataL18<>+48(SB), F2
   112  	WFMADB  V1, V3, V4, V1
   113  	FMOVD   tanhrodataL18<>+40(SB), F4
   114  	WFMADB  V3, V2, V4, V2
   115  	FMOVD   tanhrodataL18<>+32(SB), F4
   116  	WORD    $0xB9270022     //lhr %r2,%r2
   117  	WFMADB  V3, V1, V4, V1
   118  	FMOVD   tanhrodataL18<>+24(SB), F4
   119  	WFMADB  V3, V2, V4, V3
   120  	WFMADB  V0, V5, V0, V2
   121  	WFMADB  V0, V1, V3, V0
   122  	WORD    $0xA7183ECF     //lhi %r1,16079
   123  	WFMADB  V0, V2, V5, V2
   124  	FMUL    F6, F2
   125  	MOVW    R2, R10
   126  	MOVW    R1, R11
   127  	CMPBLE  R10, R11, L16
   128  	FMOVD   F6, F0
   129  	WORD    $0xED005010     //adb %f0,.L28-.L18(%r5)
   130  	BYTE    $0x00
   131  	BYTE    $0x1A
   132  	WORD    $0xA7184330     //lhi %r1,17200
   133  	FADD    F2, F0
   134  	MOVW    R2, R10
   135  	MOVW    R1, R11
   136  	CMPBGT  R10, R11, L17
   137  	WORD    $0xED605010     //sdb %f6,.L28-.L18(%r5)
   138  	BYTE    $0x00
   139  	BYTE    $0x1B
   140  	FADD    F6, F2
   141  	WFDDB   V0, V2, V0
   142  	FMOVD   F0, ret+8(FP)
   143  	RET
   144  
   145  L9:
   146  	FMOVD   tanhrodataL18<>+16(SB), F0
   147  L1:
   148  	FMOVD   F0, ret+8(FP)
   149  	RET
   150  
   151  L15:
   152  	FNEG    F0, F1
   153  	BR      L2
   154  L16:
   155  	FADD    F6, F2
   156  	FMOVD   tanhrodataL18<>+8(SB), F0
   157  	FMADD   F4, F2, F0
   158  	FMOVD   tanhrodataL18<>+0(SB), F4
   159  	FNEG    F0, F0
   160  	WFMADB  V0, V2, V4, V0
   161  	FMOVD   F0, ret+8(FP)
   162  	RET
   163  
   164  L17:
   165  	WFDDB   V0, V4, V0
   166  	FMOVD   tanhrodataL18<>+16(SB), F2
   167  	WFSDB   V0, V2, V0
   168  	FMOVD   F0, ret+8(FP)
   169  	RET
   170  
   171  tanhIsZero:      //return ±0
   172  	FMOVD   F0, ret+8(FP)
   173  	RET