Text file
src/math/tanh_s390x.s
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 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 RISBGZ $57, $60, $3, R2, R4
80 WORD $0xED105058 //cdb %f1,.L19-.L18(%r5)
81 BYTE $0x00
82 BYTE $0x19
83 RISBGN $0, $15, $48, R2, R1
84 WORD $0x68543000 //ld %f5,0(%r4,%r3)
85 LDGR R1, F6
86 BLT L3
87 MOVD $tanhxzero<>+0(SB), R1
88 FMOVD 0(R1), F2
89 WFCHDBS V0, V2, V4
90 BEQ L9
91 WFCHDBS V2, V0, V2
92 BNE L1
93 MOVD $tanhxmone<>+0(SB), R1
94 FMOVD 0(R1), F0
95 FMOVD F0, ret+8(FP)
96 RET
97
98 L3:
99 FADD F4, F2
100 FMOVD tanhrodataL18<>+80(SB), F4
101 FMADD F4, F2, F0
102 FMOVD tanhrodataL18<>+72(SB), F1
103 WFMDB V0, V0, V3
104 FMOVD tanhrodataL18<>+64(SB), F2
105 WFMADB V0, V1, V2, V1
106 FMOVD tanhrodataL18<>+56(SB), F4
107 FMOVD tanhrodataL18<>+48(SB), F2
108 WFMADB V1, V3, V4, V1
109 FMOVD tanhrodataL18<>+40(SB), F4
110 WFMADB V3, V2, V4, V2
111 FMOVD tanhrodataL18<>+32(SB), F4
112 WORD $0xB9270022 //lhr %r2,%r2
113 WFMADB V3, V1, V4, V1
114 FMOVD tanhrodataL18<>+24(SB), F4
115 WFMADB V3, V2, V4, V3
116 WFMADB V0, V5, V0, V2
117 WFMADB V0, V1, V3, V0
118 WORD $0xA7183ECF //lhi %r1,16079
119 WFMADB V0, V2, V5, V2
120 FMUL F6, F2
121 MOVW R2, R10
122 MOVW R1, R11
123 CMPBLE R10, R11, L16
124 FMOVD F6, F0
125 WORD $0xED005010 //adb %f0,.L28-.L18(%r5)
126 BYTE $0x00
127 BYTE $0x1A
128 WORD $0xA7184330 //lhi %r1,17200
129 FADD F2, F0
130 MOVW R2, R10
131 MOVW R1, R11
132 CMPBGT R10, R11, L17
133 WORD $0xED605010 //sdb %f6,.L28-.L18(%r5)
134 BYTE $0x00
135 BYTE $0x1B
136 FADD F6, F2
137 WFDDB V0, V2, V0
138 FMOVD F0, ret+8(FP)
139 RET
140
141 L9:
142 FMOVD tanhrodataL18<>+16(SB), F0
143 L1:
144 FMOVD F0, ret+8(FP)
145 RET
146
147 L15:
148 FNEG F0, F1
149 BR L2
150 L16:
151 FADD F6, F2
152 FMOVD tanhrodataL18<>+8(SB), F0
153 FMADD F4, F2, F0
154 FMOVD tanhrodataL18<>+0(SB), F4
155 FNEG F0, F0
156 WFMADB V0, V2, V4, V0
157 FMOVD F0, ret+8(FP)
158 RET
159
160 L17:
161 WFDDB V0, V4, V0
162 FMOVD tanhrodataL18<>+16(SB), F2
163 WFSDB V0, V2, V0
164 FMOVD F0, ret+8(FP)
165 RET
166
167 tanhIsZero: //return ±0
168 FMOVD F0, ret+8(FP)
169 RET
170
View as plain text