github.com/mattn/go@v0.0.0-20171011075504-07f7db3ea99f/src/math/sin_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  // Various constants
     8  DATA sincosxnan<>+0(SB)/8, $0x7ff8000000000000
     9  GLOBL sincosxnan<>+0(SB), RODATA, $8
    10  DATA sincosxlim<>+0(SB)/8, $0x432921fb54442d19
    11  GLOBL sincosxlim<>+0(SB), RODATA, $8
    12  DATA sincosxadd<>+0(SB)/8, $0xc338000000000000
    13  GLOBL sincosxadd<>+0(SB), RODATA, $8
    14  DATA sincosxpi2l<>+0(SB)/8, $0.108285667392191389e-31
    15  GLOBL sincosxpi2l<>+0(SB), RODATA, $8
    16  DATA sincosxpi2m<>+0(SB)/8, $0.612323399573676480e-16
    17  GLOBL sincosxpi2m<>+0(SB), RODATA, $8
    18  DATA sincosxpi2h<>+0(SB)/8, $0.157079632679489656e+01
    19  GLOBL sincosxpi2h<>+0(SB), RODATA, $8
    20  DATA sincosrpi2<>+0(SB)/8, $0.636619772367581341e+00
    21  GLOBL sincosrpi2<>+0(SB), RODATA, $8
    22  
    23  // Minimax polynomial approximations
    24  DATA sincosc0<>+0(SB)/8, $0.100000000000000000E+01
    25  GLOBL sincosc0<>+0(SB), RODATA, $8
    26  DATA sincosc1<>+0(SB)/8, $-.499999999999999833E+00
    27  GLOBL sincosc1<>+0(SB), RODATA, $8
    28  DATA sincosc2<>+0(SB)/8, $0.416666666666625843E-01
    29  GLOBL sincosc2<>+0(SB), RODATA, $8
    30  DATA sincosc3<>+0(SB)/8, $-.138888888885498984E-02
    31  GLOBL sincosc3<>+0(SB), RODATA, $8
    32  DATA sincosc4<>+0(SB)/8, $0.248015871681607202E-04
    33  GLOBL sincosc4<>+0(SB), RODATA, $8
    34  DATA sincosc5<>+0(SB)/8, $-.275572911309937875E-06
    35  GLOBL sincosc5<>+0(SB), RODATA, $8
    36  DATA sincosc6<>+0(SB)/8, $0.208735047247632818E-08
    37  GLOBL sincosc6<>+0(SB), RODATA, $8
    38  DATA sincosc7<>+0(SB)/8, $-.112753632738365317E-10
    39  GLOBL sincosc7<>+0(SB), RODATA, $8
    40  DATA sincoss0<>+0(SB)/8, $0.100000000000000000E+01
    41  GLOBL sincoss0<>+0(SB), RODATA, $8
    42  DATA sincoss1<>+0(SB)/8, $-.166666666666666657E+00
    43  GLOBL sincoss1<>+0(SB), RODATA, $8
    44  DATA sincoss2<>+0(SB)/8, $0.833333333333309209E-02
    45  GLOBL sincoss2<>+0(SB), RODATA, $8
    46  DATA sincoss3<>+0(SB)/8, $-.198412698410701448E-03
    47  GLOBL sincoss3<>+0(SB), RODATA, $8
    48  DATA sincoss4<>+0(SB)/8, $0.275573191453906794E-05
    49  GLOBL sincoss4<>+0(SB), RODATA, $8
    50  DATA sincoss5<>+0(SB)/8, $-.250520918387633290E-07
    51  GLOBL sincoss5<>+0(SB), RODATA, $8
    52  DATA sincoss6<>+0(SB)/8, $0.160571285514715856E-09
    53  GLOBL sincoss6<>+0(SB), RODATA, $8
    54  DATA sincoss7<>+0(SB)/8, $-.753213484933210972E-12
    55  GLOBL sincoss7<>+0(SB), RODATA, $8
    56  
    57  // Sin returns the sine of the radian argument x.
    58  //
    59  // Special cases are:
    60  //      Sin(±0) = ±0
    61  //      Sin(±Inf) = NaN
    62  //      Sin(NaN) = NaN
    63  // The algorithm used is minimax polynomial approximation.
    64  // with coefficients determined with a Remez exchange algorithm.
    65  
    66  TEXT ·sinAsm(SB),NOSPLIT,$0-16
    67  	FMOVD   x+0(FP), F0
    68  	//special case Sin(±0) = ±0
    69  	FMOVD   $(0.0), F1
    70  	FCMPU   F0, F1
    71  	BEQ     sinIsZero
    72  	WORD    $0xB3120000     //ltdbr %f0,%f0
    73  	BLTU    L17
    74  	FMOVD   F0, F5
    75  L2:
    76  	MOVD    $sincoss7<>+0(SB), R1
    77  	FMOVD   0(R1), F4
    78  	MOVD    $sincoss6<>+0(SB), R1
    79  	FMOVD   0(R1), F1
    80  	MOVD    $sincoss5<>+0(SB), R1
    81  	VLEG    $0, 0(R1), V18
    82  	MOVD    $sincoss4<>+0(SB), R1
    83  	FMOVD   0(R1), F6
    84  	MOVD    $sincoss2<>+0(SB), R1
    85  	VLEG    $0, 0(R1), V16
    86  	MOVD    $sincoss3<>+0(SB), R1
    87  	FMOVD   0(R1), F7
    88  	MOVD    $sincoss1<>+0(SB), R1
    89  	FMOVD   0(R1), F3
    90  	MOVD    $sincoss0<>+0(SB), R1
    91  	FMOVD   0(R1), F2
    92  	WFCHDBS V2, V5, V2
    93  	BEQ     L18
    94  	MOVD    $sincosrpi2<>+0(SB), R1
    95  	FMOVD   0(R1), F3
    96  	MOVD    $sincosxadd<>+0(SB), R1
    97  	FMOVD   0(R1), F2
    98  	WFMSDB  V0, V3, V2, V3
    99  	FMOVD   0(R1), F6
   100  	FADD    F3, F6
   101  	MOVD    $sincosxpi2h<>+0(SB), R1
   102  	FMOVD   0(R1), F2
   103  	FMSUB   F2, F6, F0
   104  	MOVD    $sincosxpi2m<>+0(SB), R1
   105  	FMOVD   0(R1), F4
   106  	FMADD   F4, F6, F0
   107  	MOVD    $sincosxpi2l<>+0(SB), R1
   108  	WFMDB   V0, V0, V1
   109  	FMOVD   0(R1), F7
   110  	WFMDB   V1, V1, V2
   111  	LGDR    F3, R1
   112  	MOVD    $sincosxlim<>+0(SB), R2
   113  	WORD    $0xA7110001     //tmll  %r1,1
   114  	BEQ     L6
   115  	FMOVD   0(R2), F0
   116  	WFCHDBS V0, V5, V0
   117  	BNE     L14
   118  	MOVD    $sincosc7<>+0(SB), R2
   119  	FMOVD   0(R2), F0
   120  	MOVD    $sincosc6<>+0(SB), R2
   121  	FMOVD   0(R2), F4
   122  	MOVD    $sincosc5<>+0(SB), R2
   123  	WFMADB  V1, V0, V4, V0
   124  	FMOVD   0(R2), F6
   125  	MOVD    $sincosc4<>+0(SB), R2
   126  	WFMADB  V1, V0, V6, V0
   127  	FMOVD   0(R2), F4
   128  	MOVD    $sincosc2<>+0(SB), R2
   129  	FMOVD   0(R2), F6
   130  	WFMADB  V2, V4, V6, V4
   131  	MOVD    $sincosc3<>+0(SB), R2
   132  	FMOVD   0(R2), F3
   133  	MOVD    $sincosc1<>+0(SB), R2
   134  	WFMADB  V2, V0, V3, V0
   135  	FMOVD   0(R2), F6
   136  	WFMADB  V1, V4, V6, V4
   137  	WORD    $0xA7110002     //tmll  %r1,2
   138  	WFMADB  V2, V0, V4, V0
   139  	MOVD    $sincosc0<>+0(SB), R1
   140  	FMOVD   0(R1), F2
   141  	WFMADB  V1, V0, V2, V0
   142  	BNE     L15
   143  	FMOVD   F0, ret+8(FP)
   144  	RET
   145  
   146  L6:
   147  	FMOVD   0(R2), F4
   148  	WFCHDBS V4, V5, V4
   149  	BNE     L14
   150  	MOVD    $sincoss7<>+0(SB), R2
   151  	FMOVD   0(R2), F4
   152  	MOVD    $sincoss6<>+0(SB), R2
   153  	FMOVD   0(R2), F3
   154  	MOVD    $sincoss5<>+0(SB), R2
   155  	WFMADB  V1, V4, V3, V4
   156  	WFMADB  V6, V7, V0, V6
   157  	FMOVD   0(R2), F0
   158  	MOVD    $sincoss4<>+0(SB), R2
   159  	FMADD   F4, F1, F0
   160  	FMOVD   0(R2), F3
   161  	MOVD    $sincoss2<>+0(SB), R2
   162  	FMOVD   0(R2), F4
   163  	MOVD    $sincoss3<>+0(SB), R2
   164  	WFMADB  V2, V3, V4, V3
   165  	FMOVD   0(R2), F4
   166  	MOVD    $sincoss1<>+0(SB), R2
   167  	WFMADB  V2, V0, V4, V0
   168  	FMOVD   0(R2), F4
   169  	WFMADB  V1, V3, V4, V3
   170  	FNEG    F6, F4
   171  	WFMADB  V2, V0, V3, V2
   172  	WFMDB   V4, V1, V0
   173  	WORD    $0xA7110002     //tmll  %r1,2
   174  	WFMSDB  V0, V2, V6, V0
   175  	BNE     L15
   176  	FMOVD   F0, ret+8(FP)
   177  	RET
   178  
   179  L14:
   180  	MOVD    $sincosxnan<>+0(SB), R1
   181  	FMOVD   0(R1), F0
   182  	FMOVD   F0, ret+8(FP)
   183  	RET
   184  
   185  L18:
   186  	WFMDB   V0, V0, V2
   187  	WFMADB  V2, V4, V1, V4
   188  	WFMDB   V2, V2, V1
   189  	WFMADB  V2, V4, V18, V4
   190  	WFMADB  V1, V6, V16, V6
   191  	WFMADB  V1, V4, V7, V4
   192  	WFMADB  V2, V6, V3, V6
   193  	FMUL    F0, F2
   194  	WFMADB  V1, V4, V6, V4
   195  	FMADD   F4, F2, F0
   196  	FMOVD   F0, ret+8(FP)
   197  	RET
   198  
   199  L17:
   200  	FNEG    F0, F5
   201  	BR      L2
   202  L15:
   203  	FNEG    F0, F0
   204  	FMOVD   F0, ret+8(FP)
   205  	RET
   206  
   207  
   208  sinIsZero:
   209  	FMOVD   F0, ret+8(FP)
   210  	RET
   211  
   212  // Cos returns the cosine of the radian argument.
   213  //
   214  // Special cases are:
   215  //      Cos(±Inf) = NaN
   216  //      Cos(NaN) = NaN
   217  // The algorithm used is minimax polynomial approximation.
   218  // with coefficients determined with a Remez exchange algorithm.
   219  
   220  TEXT ·cosAsm(SB),NOSPLIT,$0-16
   221  	FMOVD   x+0(FP), F0
   222  	WORD    $0xB3120000     //ltdbr %f0,%f0
   223  	BLTU    L35
   224  	FMOVD   F0, F1
   225  L21:
   226  	MOVD    $sincosc7<>+0(SB), R1
   227  	FMOVD   0(R1), F4
   228  	MOVD    $sincosc6<>+0(SB), R1
   229  	VLEG    $0, 0(R1), V20
   230  	MOVD    $sincosc5<>+0(SB), R1
   231  	VLEG    $0, 0(R1), V18
   232  	MOVD    $sincosc4<>+0(SB), R1
   233  	FMOVD   0(R1), F6
   234  	MOVD    $sincosc2<>+0(SB), R1
   235  	VLEG    $0, 0(R1), V16
   236  	MOVD    $sincosc3<>+0(SB), R1
   237  	FMOVD   0(R1), F7
   238  	MOVD    $sincosc1<>+0(SB), R1
   239  	FMOVD   0(R1), F5
   240  	MOVD    $sincosrpi2<>+0(SB), R1
   241  	FMOVD   0(R1), F2
   242  	MOVD    $sincosxadd<>+0(SB), R1
   243  	FMOVD   0(R1), F3
   244  	MOVD    $sincoss0<>+0(SB), R1
   245  	WFMSDB  V0, V2, V3, V2
   246  	FMOVD   0(R1), F3
   247  	WFCHDBS V3, V1, V3
   248  	LGDR    F2, R1
   249  	BEQ     L36
   250  	MOVD    $sincosxadd<>+0(SB), R2
   251  	FMOVD   0(R2), F4
   252  	FADD    F2, F4
   253  	MOVD    $sincosxpi2h<>+0(SB), R2
   254  	FMOVD   0(R2), F2
   255  	WFMSDB  V4, V2, V0, V2
   256  	MOVD    $sincosxpi2m<>+0(SB), R2
   257  	FMOVD   0(R2), F0
   258  	WFMADB  V4, V0, V2, V0
   259  	MOVD    $sincosxpi2l<>+0(SB), R2
   260  	WFMDB   V0, V0, V2
   261  	FMOVD   0(R2), F5
   262  	WFMDB   V2, V2, V6
   263  	MOVD    $sincosxlim<>+0(SB), R2
   264  	WORD    $0xA7110001     //tmll %r1,1
   265  	BNE     L25
   266  	FMOVD   0(R2), F0
   267  	WFCHDBS V0, V1, V0
   268  	BNE     L33
   269  	MOVD    $sincosc7<>+0(SB), R2
   270  	FMOVD   0(R2), F0
   271  	MOVD    $sincosc6<>+0(SB), R2
   272  	FMOVD   0(R2), F4
   273  	MOVD    $sincosc5<>+0(SB), R2
   274  	WFMADB  V2, V0, V4, V0
   275  	FMOVD   0(R2), F1
   276  	MOVD    $sincosc4<>+0(SB), R2
   277  	WFMADB  V2, V0, V1, V0
   278  	FMOVD   0(R2), F4
   279  	MOVD    $sincosc2<>+0(SB), R2
   280  	FMOVD   0(R2), F1
   281  	WFMADB  V6, V4, V1, V4
   282  	MOVD    $sincosc3<>+0(SB), R2
   283  	FMOVD   0(R2), F3
   284  	MOVD    $sincosc1<>+0(SB), R2
   285  	WFMADB  V6, V0, V3, V0
   286  	FMOVD   0(R2), F1
   287  	WFMADB  V2, V4, V1, V4
   288  	WORD    $0xA7110002     //tmll %r1,2
   289  	WFMADB  V6, V0, V4, V0
   290  	MOVD    $sincosc0<>+0(SB), R1
   291  	FMOVD   0(R1), F4
   292  	WFMADB  V2, V0, V4, V0
   293  	BNE     L34
   294  	FMOVD   F0, ret+8(FP)
   295  	RET
   296  
   297  L25:
   298  	FMOVD   0(R2), F3
   299  	WFCHDBS V3, V1, V1
   300  	BNE     L33
   301  	MOVD    $sincoss7<>+0(SB), R2
   302  	FMOVD   0(R2), F1
   303  	MOVD    $sincoss6<>+0(SB), R2
   304  	FMOVD   0(R2), F3
   305  	MOVD    $sincoss5<>+0(SB), R2
   306  	WFMADB  V2, V1, V3, V1
   307  	FMOVD   0(R2), F3
   308  	MOVD    $sincoss4<>+0(SB), R2
   309  	WFMADB  V2, V1, V3, V1
   310  	FMOVD   0(R2), F3
   311  	MOVD    $sincoss2<>+0(SB), R2
   312  	FMOVD   0(R2), F7
   313  	WFMADB  V6, V3, V7, V3
   314  	MOVD    $sincoss3<>+0(SB), R2
   315  	FMADD   F5, F4, F0
   316  	FMOVD   0(R2), F4
   317  	MOVD    $sincoss1<>+0(SB), R2
   318  	FMADD   F1, F6, F4
   319  	FMOVD   0(R2), F1
   320  	FMADD   F3, F2, F1
   321  	FMUL    F0, F2
   322  	WFMADB  V6, V4, V1, V6
   323  	WORD    $0xA7110002     //tmll  %r1,2
   324  	FMADD   F6, F2, F0
   325  	BNE     L34
   326  	FMOVD   F0, ret+8(FP)
   327  	RET
   328  
   329  L33:
   330  	MOVD    $sincosxnan<>+0(SB), R1
   331  	FMOVD   0(R1), F0
   332  	FMOVD   F0, ret+8(FP)
   333  	RET
   334  
   335  L36:
   336  	FMUL    F0, F0
   337  	MOVD    $sincosc0<>+0(SB), R1
   338  	WFMDB   V0, V0, V1
   339  	WFMADB  V0, V4, V20, V4
   340  	WFMADB  V1, V6, V16, V6
   341  	WFMADB  V0, V4, V18, V4
   342  	WFMADB  V0, V6, V5, V6
   343  	WFMADB  V1, V4, V7, V4
   344  	FMOVD   0(R1), F2
   345  	WFMADB  V1, V4, V6, V4
   346  	WFMADB  V0, V4, V2, V0
   347  	FMOVD   F0, ret+8(FP)
   348  	RET
   349  
   350  L35:
   351  	FNEG    F0, F1
   352  	BR      L21
   353  L34:
   354  	FNEG    F0, F0
   355  	FMOVD   F0, ret+8(FP)
   356  	RET