1// Copyright 2014 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//go:build !plan9
6
7#include "go_asm.h"
8#include "textflag.h"
9
10// See memclrNoHeapPointers Go doc for important implementation constraints.
11
12// func memclrNoHeapPointers(ptr unsafe.Pointer, n uintptr)
13TEXT runtime·memclrNoHeapPointers(SB), NOSPLIT, $0-8
14	MOVL	ptr+0(FP), DI
15	MOVL	n+4(FP), BX
16	XORL	AX, AX
17
18	// MOVOU seems always faster than REP STOSL.
19tail:
20	// BSR+branch table make almost all memmove/memclr benchmarks worse. Not worth doing.
21	TESTL	BX, BX
22	JEQ	_0
23	CMPL	BX, $2
24	JBE	_1or2
25	CMPL	BX, $4
26	JB	_3
27	JE	_4
28	CMPL	BX, $8
29	JBE	_5through8
30	CMPL	BX, $16
31	JBE	_9through16
32#ifdef GO386_softfloat
33	JMP	nosse2
34#endif
35	PXOR	X0, X0
36	CMPL	BX, $32
37	JBE	_17through32
38	CMPL	BX, $64
39	JBE	_33through64
40	CMPL	BX, $128
41	JBE	_65through128
42	CMPL	BX, $256
43	JBE	_129through256
44
45loop:
46	MOVOU	X0, 0(DI)
47	MOVOU	X0, 16(DI)
48	MOVOU	X0, 32(DI)
49	MOVOU	X0, 48(DI)
50	MOVOU	X0, 64(DI)
51	MOVOU	X0, 80(DI)
52	MOVOU	X0, 96(DI)
53	MOVOU	X0, 112(DI)
54	MOVOU	X0, 128(DI)
55	MOVOU	X0, 144(DI)
56	MOVOU	X0, 160(DI)
57	MOVOU	X0, 176(DI)
58	MOVOU	X0, 192(DI)
59	MOVOU	X0, 208(DI)
60	MOVOU	X0, 224(DI)
61	MOVOU	X0, 240(DI)
62	SUBL	$256, BX
63	ADDL	$256, DI
64	CMPL	BX, $256
65	JAE	loop
66	JMP	tail
67
68_1or2:
69	MOVB	AX, (DI)
70	MOVB	AX, -1(DI)(BX*1)
71	RET
72_0:
73	RET
74_3:
75	MOVW	AX, (DI)
76	MOVB	AX, 2(DI)
77	RET
78_4:
79	// We need a separate case for 4 to make sure we clear pointers atomically.
80	MOVL	AX, (DI)
81	RET
82_5through8:
83	MOVL	AX, (DI)
84	MOVL	AX, -4(DI)(BX*1)
85	RET
86_9through16:
87	MOVL	AX, (DI)
88	MOVL	AX, 4(DI)
89	MOVL	AX, -8(DI)(BX*1)
90	MOVL	AX, -4(DI)(BX*1)
91	RET
92_17through32:
93	MOVOU	X0, (DI)
94	MOVOU	X0, -16(DI)(BX*1)
95	RET
96_33through64:
97	MOVOU	X0, (DI)
98	MOVOU	X0, 16(DI)
99	MOVOU	X0, -32(DI)(BX*1)
100	MOVOU	X0, -16(DI)(BX*1)
101	RET
102_65through128:
103	MOVOU	X0, (DI)
104	MOVOU	X0, 16(DI)
105	MOVOU	X0, 32(DI)
106	MOVOU	X0, 48(DI)
107	MOVOU	X0, -64(DI)(BX*1)
108	MOVOU	X0, -48(DI)(BX*1)
109	MOVOU	X0, -32(DI)(BX*1)
110	MOVOU	X0, -16(DI)(BX*1)
111	RET
112_129through256:
113	MOVOU	X0, (DI)
114	MOVOU	X0, 16(DI)
115	MOVOU	X0, 32(DI)
116	MOVOU	X0, 48(DI)
117	MOVOU	X0, 64(DI)
118	MOVOU	X0, 80(DI)
119	MOVOU	X0, 96(DI)
120	MOVOU	X0, 112(DI)
121	MOVOU	X0, -128(DI)(BX*1)
122	MOVOU	X0, -112(DI)(BX*1)
123	MOVOU	X0, -96(DI)(BX*1)
124	MOVOU	X0, -80(DI)(BX*1)
125	MOVOU	X0, -64(DI)(BX*1)
126	MOVOU	X0, -48(DI)(BX*1)
127	MOVOU	X0, -32(DI)(BX*1)
128	MOVOU	X0, -16(DI)(BX*1)
129	RET
130nosse2:
131	MOVL	BX, CX
132	SHRL	$2, CX
133	REP
134	STOSL
135	ANDL	$3, BX
136	JNE	tail
137	RET
138