1 files changed, 84 insertions, 0 deletions
diff --git a/arch/microblaze/lib/uaccess_old.S b/arch/microblaze/lib/uaccess_old.S
index d09f2dce648d..142492ec270f 100644
--- a/arch/microblaze/lib/uaccess_old.S
+++ b/arch/microblaze/lib/uaccess_old.S
@@ -10,6 +10,7 @@
 #include <linux/errno.h>
 #include <linux/linkage.h>
+#include <asm/page.h>
 /*
 * int __strncpy_user(char *to, char *from, int len);
@@ -102,6 +103,49 @@ __strnlen_user:
        .section        __ex_table,"a"
        .word   1b,4b
+/* Loop unrolling for __copy_tofrom_user */
+#define COPY(offset)    \
+1:      lwi     r4 , r6, 0x0000 + offset;       \
+2:      lwi     r19, r6, 0x0004 + offset;       \
+3:      lwi     r20, r6, 0x0008 + offset;       \
+4:      lwi     r21, r6, 0x000C + offset;       \
+5:      lwi     r22, r6, 0x0010 + offset;       \
+6:      lwi     r23, r6, 0x0014 + offset;       \
+7:      lwi     r24, r6, 0x0018 + offset;       \
+8:      lwi     r25, r6, 0x001C + offset;       \
+9:      swi     r4 , r5, 0x0000 + offset;       \
+10:     swi     r19, r5, 0x0004 + offset;       \
+11:     swi     r20, r5, 0x0008 + offset;       \
+12:     swi     r21, r5, 0x000C + offset;       \
+13:     swi     r22, r5, 0x0010 + offset;       \
+14:     swi     r23, r5, 0x0014 + offset;       \
+15:     swi     r24, r5, 0x0018 + offset;       \
+16:     swi     r25, r5, 0x001C + offset;       \
+        .section __ex_table,"a";                \
+        .word   1b, 0f;                         \
+        .word   2b, 0f;                         \
+        .word   3b, 0f;                         \
+        .word   4b, 0f;                         \
+        .word   5b, 0f;                         \
+        .word   6b, 0f;                         \
+        .word   7b, 0f;                         \
+        .word   8b, 0f;                         \
+        .word   9b, 0f;                         \
+        .word   10b, 0f;                        \
+        .word   11b, 0f;                        \
+        .word   12b, 0f;                        \
+        .word   13b, 0f;                        \
+        .word   14b, 0f;                        \
+        .word   15b, 0f;                        \
+        .word   16b, 0f;                        \
+        .text
+#define COPY_80(offset) \
+        COPY(0x00 + offset);\
+        COPY(0x20 + offset);\
+        COPY(0x40 + offset);\
+        COPY(0x60 + offset);
 /*
 * int __copy_tofrom_user(char *to, char *from, int len)
 * Return:
@@ -126,6 +170,10 @@ __copy_tofrom_user:
        bneid   r3, bu1 /* if r3 is not zero then byte copying */
        or      r3, r0, r0
+        rsubi   r3, r7, PAGE_SIZE /* detect PAGE_SIZE */
+        beqid   r3, page;
+        or      r3, r0, r0
 w1:     lw      r4, r6, r3 /* at least one 4 byte copy */
 w2:     sw      r4, r5, r3
        addik   r7, r7, -4
@@ -140,6 +188,42 @@ w2:	sw	r4, r5, r3
        .word   w2, 0f;
        .text
+.align 4 /* Alignment is important to keep icache happy */
+page:   /* Create room on stack and save registers for storign values */
+        addik   r1, r1, -32
+        swi     r19, r1, 4
+        swi     r20, r1, 8
+        swi     r21, r1, 12
+        swi     r22, r1, 16
+        swi     r23, r1, 20
+        swi     r24, r1, 24
+        swi     r25, r1, 28
+loop:   /* r4, r19, r20, r21, r22, r23, r24, r25 are used for storing values */
+        /* Loop unrolling to get performance boost */
+        COPY_80(0x000);
+        COPY_80(0x080);
+        COPY_80(0x100);
+        COPY_80(0x180);
+        /* copy loop */
+        addik   r6, r6, 0x200
+        addik   r7, r7, -0x200
+        bneid   r7, loop
+        addik   r5, r5, 0x200
+        /* Restore register content */
+        lwi     r19, r1, 4
+        lwi     r20, r1, 8
+        lwi     r21, r1, 12
+        lwi     r22, r1, 16
+        lwi     r23, r1, 20
+        lwi     r24, r1, 24
+        lwi     r25, r1, 28
+        addik   r1, r1, 32
+        /* return back */
+        addik   r3, r7, 0
+        rtsd    r15, 8
+        nop
+.align 4 /* Alignment is important to keep icache happy */
 bu1:    lbu     r4,r6,r3
 bu2:    sb      r4,r5,r3
        addik   r7,r7,-1

diff --git a/arch/microblaze/lib/uaccess_old.S b/arch/microblaze/lib/uaccess_old.S index d09f2dce648d..142492ec270f 100644 --- a/arch/microblaze/lib/uaccess_old.S +++ b/arch/microblaze/lib/uaccess_old.S
@@ -10,6 +10,7 @@
10		10
11	#include <linux/errno.h>	11	#include <linux/errno.h>
12	#include <linux/linkage.h>	12	#include <linux/linkage.h>
		13	#include <asm/page.h>
13		14
14	/*	15	/*
15	* int __strncpy_user(char to, char from, int len);	16	* int __strncpy_user(char to, char from, int len);
@@ -102,6 +103,49 @@ __strnlen_user:
102	.section __ex_table,"a"	103	.section __ex_table,"a"
103	.word 1b,4b	104	.word 1b,4b
104		105
		106	/* Loop unrolling for __copy_tofrom_user */
		107	#define COPY(offset) \
		108	1: lwi r4 , r6, 0x0000 + offset; \
		109	2: lwi r19, r6, 0x0004 + offset; \
		110	3: lwi r20, r6, 0x0008 + offset; \
		111	4: lwi r21, r6, 0x000C + offset; \
		112	5: lwi r22, r6, 0x0010 + offset; \
		113	6: lwi r23, r6, 0x0014 + offset; \
		114	7: lwi r24, r6, 0x0018 + offset; \
		115	8: lwi r25, r6, 0x001C + offset; \
		116	9: swi r4 , r5, 0x0000 + offset; \
		117	10: swi r19, r5, 0x0004 + offset; \
		118	11: swi r20, r5, 0x0008 + offset; \
		119	12: swi r21, r5, 0x000C + offset; \
		120	13: swi r22, r5, 0x0010 + offset; \
		121	14: swi r23, r5, 0x0014 + offset; \
		122	15: swi r24, r5, 0x0018 + offset; \
		123	16: swi r25, r5, 0x001C + offset; \
		124	.section __ex_table,"a"; \
		125	.word 1b, 0f; \
		126	.word 2b, 0f; \
		127	.word 3b, 0f; \
		128	.word 4b, 0f; \
		129	.word 5b, 0f; \
		130	.word 6b, 0f; \
		131	.word 7b, 0f; \
		132	.word 8b, 0f; \
		133	.word 9b, 0f; \
		134	.word 10b, 0f; \
		135	.word 11b, 0f; \
		136	.word 12b, 0f; \
		137	.word 13b, 0f; \
		138	.word 14b, 0f; \
		139	.word 15b, 0f; \
		140	.word 16b, 0f; \
		141	.text
		142
		143	#define COPY_80(offset) \
		144	COPY(0x00 + offset);\
		145	COPY(0x20 + offset);\
		146	COPY(0x40 + offset);\
		147	COPY(0x60 + offset);
		148
105	/*	149	/*
106	* int __copy_tofrom_user(char to, char from, int len)	150	* int __copy_tofrom_user(char to, char from, int len)
107	* Return:	151	* Return:
@@ -126,6 +170,10 @@ __copy_tofrom_user:
126	bneid r3, bu1 /* if r3 is not zero then byte copying */	170	bneid r3, bu1 /* if r3 is not zero then byte copying */
127	or r3, r0, r0	171	or r3, r0, r0
128		172
		173	rsubi r3, r7, PAGE_SIZE /* detect PAGE_SIZE */
		174	beqid r3, page;
		175	or r3, r0, r0
		176
129	w1: lw r4, r6, r3 /* at least one 4 byte copy */	177	w1: lw r4, r6, r3 /* at least one 4 byte copy */
130	w2: sw r4, r5, r3	178	w2: sw r4, r5, r3
131	addik r7, r7, -4	179	addik r7, r7, -4
@@ -140,6 +188,42 @@ w2: sw r4, r5, r3
140	.word w2, 0f;	188	.word w2, 0f;
141	.text	189	.text
142		190
		191	.align 4 /* Alignment is important to keep icache happy */
		192	page: /* Create room on stack and save registers for storign values */
		193	addik r1, r1, -32
		194	swi r19, r1, 4
		195	swi r20, r1, 8
		196	swi r21, r1, 12
		197	swi r22, r1, 16
		198	swi r23, r1, 20
		199	swi r24, r1, 24
		200	swi r25, r1, 28
		201	loop: /* r4, r19, r20, r21, r22, r23, r24, r25 are used for storing values */
		202	/* Loop unrolling to get performance boost */
		203	COPY_80(0x000);
		204	COPY_80(0x080);
		205	COPY_80(0x100);
		206	COPY_80(0x180);
		207	/* copy loop */
		208	addik r6, r6, 0x200
		209	addik r7, r7, -0x200
		210	bneid r7, loop
		211	addik r5, r5, 0x200
		212	/* Restore register content */
		213	lwi r19, r1, 4
		214	lwi r20, r1, 8
		215	lwi r21, r1, 12
		216	lwi r22, r1, 16
		217	lwi r23, r1, 20
		218	lwi r24, r1, 24
		219	lwi r25, r1, 28
		220	addik r1, r1, 32
		221	/* return back */
		222	addik r3, r7, 0
		223	rtsd r15, 8
		224	nop
		225
		226	.align 4 /* Alignment is important to keep icache happy */
143	bu1: lbu r4,r6,r3	227	bu1: lbu r4,r6,r3
144	bu2: sb r4,r5,r3	228	bu2: sb r4,r5,r3
145	addik r7,r7,-1	229	addik r7,r7,-1