1/* NGpage.S: Niagara optimize clear and copy page.
2 *
3 * Copyright (C) 2006 (davem@davemloft.net)
4 */
5
6#include <asm/asi.h>
7#include <asm/page.h>
8
9	.text
10	.align	32
11
12	/* This is heavily simplified from the sun4u variants
13	 * because Niagara does not have any D-cache aliasing issues
14	 * and also we don't need to use the FPU in order to implement
15	 * an optimal page copy/clear.
16	 */
17
18NGcopy_user_page:	/* %o0=dest, %o1=src, %o2=vaddr */
19	save		%sp, -192, %sp
20	rd		%asi, %g3
21	wr		%g0, ASI_BLK_INIT_QUAD_LDD_P, %asi
22	set		PAGE_SIZE, %g7
23	prefetch	[%i1 + 0x00], #one_read
24	prefetch	[%i1 + 0x40], #one_read
25
261:	prefetch	[%i1 + 0x80], #one_read
27	prefetch	[%i1 + 0xc0], #one_read
28	ldda		[%i1 + 0x00] %asi, %o2
29	ldda		[%i1 + 0x10] %asi, %o4
30	ldda		[%i1 + 0x20] %asi, %l2
31	ldda		[%i1 + 0x30] %asi, %l4
32	stxa		%o2, [%i0 + 0x00] %asi
33	stxa		%o3, [%i0 + 0x08] %asi
34	stxa		%o4, [%i0 + 0x10] %asi
35	stxa		%o5, [%i0 + 0x18] %asi
36	stxa		%l2, [%i0 + 0x20] %asi
37	stxa		%l3, [%i0 + 0x28] %asi
38	stxa		%l4, [%i0 + 0x30] %asi
39	stxa		%l5, [%i0 + 0x38] %asi
40	ldda		[%i1 + 0x40] %asi, %o2
41	ldda		[%i1 + 0x50] %asi, %o4
42	ldda		[%i1 + 0x60] %asi, %l2
43	ldda		[%i1 + 0x70] %asi, %l4
44	stxa		%o2, [%i0 + 0x40] %asi
45	stxa		%o3, [%i0 + 0x48] %asi
46	stxa		%o4, [%i0 + 0x50] %asi
47	stxa		%o5, [%i0 + 0x58] %asi
48	stxa		%l2, [%i0 + 0x60] %asi
49	stxa		%l3, [%i0 + 0x68] %asi
50	stxa		%l4, [%i0 + 0x70] %asi
51	stxa		%l5, [%i0 + 0x78] %asi
52	add		%i1, 128, %i1
53	subcc		%g7, 128, %g7
54	bne,pt		%xcc, 1b
55	 add		%i0, 128, %i0
56	wr		%g3, 0x0, %asi
57	membar		#Sync
58	ret
59	 restore
60
61	.align		32
62	.globl		NGclear_page
63	.globl		NGclear_user_page
64NGclear_page:		/* %o0=dest */
65NGclear_user_page:	/* %o0=dest, %o1=vaddr */
66	rd		%asi, %g3
67	wr		%g0, ASI_BLK_INIT_QUAD_LDD_P, %asi
68	set		PAGE_SIZE, %g7
69
701:	stxa		%g0, [%o0 + 0x00] %asi
71	stxa		%g0, [%o0 + 0x08] %asi
72	stxa		%g0, [%o0 + 0x10] %asi
73	stxa		%g0, [%o0 + 0x18] %asi
74	stxa		%g0, [%o0 + 0x20] %asi
75	stxa		%g0, [%o0 + 0x28] %asi
76	stxa		%g0, [%o0 + 0x30] %asi
77	stxa		%g0, [%o0 + 0x38] %asi
78	stxa		%g0, [%o0 + 0x40] %asi
79	stxa		%g0, [%o0 + 0x48] %asi
80	stxa		%g0, [%o0 + 0x50] %asi
81	stxa		%g0, [%o0 + 0x58] %asi
82	stxa		%g0, [%o0 + 0x60] %asi
83	stxa		%g0, [%o0 + 0x68] %asi
84	stxa		%g0, [%o0 + 0x70] %asi
85	stxa		%g0, [%o0 + 0x78] %asi
86	stxa		%g0, [%o0 + 0x80] %asi
87	stxa		%g0, [%o0 + 0x88] %asi
88	stxa		%g0, [%o0 + 0x90] %asi
89	stxa		%g0, [%o0 + 0x98] %asi
90	stxa		%g0, [%o0 + 0xa0] %asi
91	stxa		%g0, [%o0 + 0xa8] %asi
92	stxa		%g0, [%o0 + 0xb0] %asi
93	stxa		%g0, [%o0 + 0xb8] %asi
94	stxa		%g0, [%o0 + 0xc0] %asi
95	stxa		%g0, [%o0 + 0xc8] %asi
96	stxa		%g0, [%o0 + 0xd0] %asi
97	stxa		%g0, [%o0 + 0xd8] %asi
98	stxa		%g0, [%o0 + 0xe0] %asi
99	stxa		%g0, [%o0 + 0xe8] %asi
100	stxa		%g0, [%o0 + 0xf0] %asi
101	stxa		%g0, [%o0 + 0xf8] %asi
102	subcc		%g7, 256, %g7
103	bne,pt		%xcc, 1b
104	 add		%o0, 256, %o0
105	wr		%g3, 0x0, %asi
106	membar		#Sync
107	retl
108	 nop
109
110#define BRANCH_ALWAYS	0x10680000
111#define NOP		0x01000000
112#define NG_DO_PATCH(OLD, NEW)	\
113	sethi	%hi(NEW), %g1; \
114	or	%g1, %lo(NEW), %g1; \
115	sethi	%hi(OLD), %g2; \
116	or	%g2, %lo(OLD), %g2; \
117	sub	%g1, %g2, %g1; \
118	sethi	%hi(BRANCH_ALWAYS), %g3; \
119	sll	%g1, 11, %g1; \
120	srl	%g1, 11 + 2, %g1; \
121	or	%g3, %lo(BRANCH_ALWAYS), %g3; \
122	or	%g3, %g1, %g3; \
123	stw	%g3, [%g2]; \
124	sethi	%hi(NOP), %g3; \
125	or	%g3, %lo(NOP), %g3; \
126	stw	%g3, [%g2 + 0x4]; \
127	flush	%g2;
128
129	.globl	niagara_patch_pageops
130	.type	niagara_patch_pageops,#function
131niagara_patch_pageops:
132	NG_DO_PATCH(copy_user_page, NGcopy_user_page)
133	NG_DO_PATCH(_clear_page, NGclear_page)
134	NG_DO_PATCH(clear_user_page, NGclear_user_page)
135	retl
136	 nop
137	.size	niagara_patch_pageops,.-niagara_patch_pageops
138