1// Copyright 2017 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 coefficients and other constants
8DATA ·cbrtrodataL9<> + 0(SB)/8, $-.00016272731015974436E+00
9DATA ·cbrtrodataL9<> + 8(SB)/8, $0.66639548758285293179E+00
10DATA ·cbrtrodataL9<> + 16(SB)/8, $0.55519402697349815993E+00
11DATA ·cbrtrodataL9<> + 24(SB)/8, $0.49338566048766782004E+00
12DATA ·cbrtrodataL9<> + 32(SB)/8, $0.45208160036325611486E+00
13DATA ·cbrtrodataL9<> + 40(SB)/8, $0.43099892837778637816E+00
14DATA ·cbrtrodataL9<> + 48(SB)/8, $1.000244140625
15DATA ·cbrtrodataL9<> + 56(SB)/8, $0.33333333333333333333E+00
16DATA ·cbrtrodataL9<> + 64(SB)/8, $79228162514264337593543950336.
17GLOBL ·cbrtrodataL9<> + 0(SB), RODATA, $72
18
19// Index tables
20DATA ·cbrttab32069<> + 0(SB)/8, $0x404030303020202
21DATA ·cbrttab32069<> + 8(SB)/8, $0x101010101000000
22DATA ·cbrttab32069<> + 16(SB)/8, $0x808070706060605
23DATA ·cbrttab32069<> + 24(SB)/8, $0x505040404040303
24DATA ·cbrttab32069<> + 32(SB)/8, $0xe0d0c0c0b0b0b0a
25DATA ·cbrttab32069<> + 40(SB)/8, $0xa09090908080808
26DATA ·cbrttab32069<> + 48(SB)/8, $0x11111010100f0f0f
27DATA ·cbrttab32069<> + 56(SB)/8, $0xe0e0e0e0e0d0d0d
28DATA ·cbrttab32069<> + 64(SB)/8, $0x1515141413131312
29DATA ·cbrttab32069<> + 72(SB)/8, $0x1212111111111010
30GLOBL ·cbrttab32069<> + 0(SB), RODATA, $80
31
32DATA ·cbrttab22068<> + 0(SB)/8, $0x151015001420141
33DATA ·cbrttab22068<> + 8(SB)/8, $0x140013201310130
34DATA ·cbrttab22068<> + 16(SB)/8, $0x122012101200112
35DATA ·cbrttab22068<> + 24(SB)/8, $0x111011001020101
36DATA ·cbrttab22068<> + 32(SB)/8, $0x10000f200f100f0
37DATA ·cbrttab22068<> + 40(SB)/8, $0xe200e100e000d2
38DATA ·cbrttab22068<> + 48(SB)/8, $0xd100d000c200c1
39DATA ·cbrttab22068<> + 56(SB)/8, $0xc000b200b100b0
40DATA ·cbrttab22068<> + 64(SB)/8, $0xa200a100a00092
41DATA ·cbrttab22068<> + 72(SB)/8, $0x91009000820081
42DATA ·cbrttab22068<> + 80(SB)/8, $0x80007200710070
43DATA ·cbrttab22068<> + 88(SB)/8, $0x62006100600052
44DATA ·cbrttab22068<> + 96(SB)/8, $0x51005000420041
45DATA ·cbrttab22068<> + 104(SB)/8, $0x40003200310030
46DATA ·cbrttab22068<> + 112(SB)/8, $0x22002100200012
47DATA ·cbrttab22068<> + 120(SB)/8, $0x11001000020001
48GLOBL ·cbrttab22068<> + 0(SB), RODATA, $128
49
50DATA ·cbrttab12067<> + 0(SB)/8, $0x53e1529051324fe1
51DATA ·cbrttab12067<> + 8(SB)/8, $0x4e904d324be14a90
52DATA ·cbrttab12067<> + 16(SB)/8, $0x493247e146904532
53DATA ·cbrttab12067<> + 24(SB)/8, $0x43e1429041323fe1
54DATA ·cbrttab12067<> + 32(SB)/8, $0x3e903d323be13a90
55DATA ·cbrttab12067<> + 40(SB)/8, $0x393237e136903532
56DATA ·cbrttab12067<> + 48(SB)/8, $0x33e1329031322fe1
57DATA ·cbrttab12067<> + 56(SB)/8, $0x2e902d322be12a90
58DATA ·cbrttab12067<> + 64(SB)/8, $0xd3e1d290d132cfe1
59DATA ·cbrttab12067<> + 72(SB)/8, $0xce90cd32cbe1ca90
60DATA ·cbrttab12067<> + 80(SB)/8, $0xc932c7e1c690c532
61DATA ·cbrttab12067<> + 88(SB)/8, $0xc3e1c290c132bfe1
62DATA ·cbrttab12067<> + 96(SB)/8, $0xbe90bd32bbe1ba90
63DATA ·cbrttab12067<> + 104(SB)/8, $0xb932b7e1b690b532
64DATA ·cbrttab12067<> + 112(SB)/8, $0xb3e1b290b132afe1
65DATA ·cbrttab12067<> + 120(SB)/8, $0xae90ad32abe1aa90
66GLOBL ·cbrttab12067<> + 0(SB), RODATA, $128
67
68// Cbrt returns the cube root of the argument.
69//
70// Special cases are:
71//      Cbrt(±0) = ±0
72//      Cbrt(±Inf) = ±Inf
73//      Cbrt(NaN) = NaN
74// The algorithm used is minimax polynomial approximation
75// with coefficients determined with a Remez exchange algorithm.
76
77TEXT	·cbrtAsm(SB), NOSPLIT, $0-16
78	FMOVD	x+0(FP), F0
79	MOVDcbrtrodataL9<>+0(SB), R9
80	LGDR	F0, R2
81	WORD	$0xC039000F	//iilf	%r3,1048575
82	BYTE	$0xFF
83	BYTE	$0xFF
84	SRAD	$32, R2
85	WORD	$0xB9170012	//llgtr	%r1,%r2
86	MOVW	R1, R6
87	MOVW	R3, R7
88	CMPBLE	R6, R7, L2
89	WORD	$0xC0397FEF	//iilf	%r3,2146435071
90	BYTE	$0xFF
91	BYTE	$0xFF
92	MOVW	R3, R7
93	CMPBLE	R6, R7, L8
94L1:
95	FMOVD	F0, ret+8(FP)
96	RET
97L3:
98L2:
99	LTDBR	F0, F0
100	BEQ	L1
101	FMOVD	F0, F2
102	WORD	$0xED209040	//mdb	%f2,.L10-.L9(%r9)
103	BYTE	$0x00
104	BYTE	$0x1C
105	MOVH	$0x200, R4
106	LGDR	F2, R2
107	SRAD	$32, R2
108L4:
109	RISBGZ	$57, $62, $39, R2, R3
110	MOVDcbrttab12067<>+0(SB), R1
111	WORD	$0x48131000	//lh	%r1,0(%r3,%r1)
112	RISBGZ	$57, $62, $45, R2, R3
113	MOVDcbrttab22068<>+0(SB), R5
114	RISBGNZ	$60, $63, $48, R2, R2
115	WORD	$0x4A135000	//ah	%r1,0(%r3,%r5)
116	BYTE	$0x18	//lr	%r3,%r1
117	BYTE	$0x31
118	MOVDcbrttab32069<>+0(SB), R1
119	FMOVD	56(R9), F1
120	FMOVD	48(R9), F5
121	WORD	$0xEC23393B	//rosbg	%r2,%r3,57,59,4
122	BYTE	$0x04
123	BYTE	$0x56
124	WORD	$0xE3121000	//llc	%r1,0(%r2,%r1)
125	BYTE	$0x00
126	BYTE	$0x94
127	ADDW	R3, R1
128	ADDW	R4, R1
129	SLW	$16, R1, R1
130	SLD	$32, R1, R1
131	LDGR	R1, F2
132	WFMDB	V2, V2, V4
133	WFMDB	V4, V0, V6
134	WFMSDB	V4, V6, V2, V4
135	FMOVD	40(R9), F6
136	FMSUB	F1, F4, F2
137	FMOVD	32(R9), F4
138	WFMDB	V2, V2, V3
139	FMOVD	24(R9), F1
140	FMUL	F3, F0
141	FMOVD	16(R9), F3
142	WFMADB	V2, V0, V5, V2
143	FMOVD	8(R9), F5
144	FMADD	F6, F2, F4
145	WFMADB	V2, V1, V3, V1
146	WFMDB	V2, V2, V6
147	FMOVD	0(R9), F3
148	WFMADB	V4, V6, V1, V4
149	WFMADB	V2, V5, V3, V2
150	FMADD	F4, F6, F2
151	FMADD	F2, F0, F0
152	FMOVD	F0, ret+8(FP)
153	RET
154L8:
155	MOVH	$0x0, R4
156	BR	L4
157