版主，你好。memcmp.S文件内容如下
#include <machine/cpu-features.h>
.text
.global memcmp
.type memcmp, %function
.align 4
/*
* Optimized memcmp() for ARM9.
* This would not be optimal on XScale or ARM11, where more prefetching
* and use of PLD will be needed.
* The 2 major optimzations here are
* (1) The main loop compares 16 bytes at a time
* (2) The loads are scheduled in a way they won't stall
*/
memcmp:
.fnstart
PLD (r0, #0)
PLD (r1, #0)
/* take of the case where length is 0 or the buffers are the same */
cmp r0, r1
cmpne r2, #0
moveq r0, #0
bxeq lr
.save {r4, lr}
/* save registers */
stmfd sp!, {r4, lr}
PLD (r0, #32)
PLD (r1, #32)
/* since r0 hold the result, move the first source
* pointer somewhere else
*/
mov r4, r0
/* make sure we have at least 8+4 bytes, this simplify things below
* and avoid some overhead for small blocks
*/
cmp r2, #(8+4)
bmi 8f
/* align first pointer to word boundary
* offset = -src & 3
*/
rsb r3, r4, #0
ands r3, r3, #3
beq 0f
/* align first pointer */
sub r2, r2, r3
1: ldrb r0, [r4], #1
ldrb ip, [r1], #1
subs r0, r0, ip
bne 9f
subs r3, r3, #1
bne 1b
0: /* here the first pointer is aligned, and we have at least 4 bytes
* to process.
*/
/* see if the pointers are congruent */
eor r0, r4, r1
ands r0, r0, #3
bne 5f
/* congruent case, 32 bytes per iteration
* We need to make sure there are at least 32+4 bytes left
* because we effectively read ahead one word, and we could
* read past the buffer (and segfault) if we're not careful.
*/
ldr ip, [r1]
subs r2, r2, #(32 + 4)
bmi 1f
0: PLD (r4, #64)
PLD (r1, #64)
ldr r0, [r4], #4
ldr lr, [r1, #4]!
eors r0, r0, ip
ldreq r0, [r4], #4
ldreq ip, [r1, #4]!
eoreqs r0, r0, lr
ldreq r0, [r4], #4
ldreq lr, [r1, #4]!
eoreqs r0, r0, ip
ldreq r0, [r4], #4
ldreq ip, [r1, #4]!
eoreqs r0, r0, lr
ldreq r0, [r4], #4
ldreq lr, [r1, #4]!
eoreqs r0, r0, ip
ldreq r0, [r4], #4
ldreq ip, [r1, #4]!
eoreqs r0, r0, lr
ldreq r0, [r4], #4
ldreq lr, [r1, #4]!
eoreqs r0, r0, ip
ldreq r0, [r4], #4
ldreq ip, [r1, #4]!
eoreqs r0, r0, lr
bne 2f
subs r2, r2, #32
bhs 0b
/* do we have at least 4 bytes left? */
1: adds r2, r2, #(32 - 4 + 4)
bmi 4f
/* finish off 4 bytes at a time */
3: ldr r0, [r4], #4
ldr ip, [r1], #4
eors r0, r0, ip
bne 2f
subs r2, r2, #4
bhs 3b
/* are we done? */
4: adds r2, r2, #4
moveq r0, #0
beq 9f
/* finish off the remaining bytes */
b 8f
2: /* the last 4 bytes are different, restart them */
sub r4, r4, #4
sub r1, r1, #4
mov r2, #4
/* process the last few bytes */
8: ldrb r0, [r4], #1
ldrb ip, [r1], #1
// stall
subs r0, r0, ip
bne 9f
subs r2, r2, #1
bne 8b
9: /* restore registers and return */
ldmfd sp!, {r4, lr}
bx lr
.fnend
5: /*************** non-congruent case ***************/
and r0, r1, #3
cmp r0, #2
bne 4f
/* here, offset is 2 (16-bits aligned, special cased) */
/* make sure we have at least 16 bytes to process */
subs r2, r2, #16
addmi r2, r2, #16
bmi 8b
/* align the unaligned pointer */
bic r1, r1, #3
ldr lr, [r1], #4
6: PLD (r1, #64)
PLD (r4, #64)
mov ip, lr, lsr #16
ldr lr, [r1], #4
ldr r0, [r4], #4
orr ip, ip, lr, lsl #16
eors r0, r0, ip
moveq ip, lr, lsr #16
ldreq lr, [r1], #4
ldreq r0, [r4], #4
orreq ip, ip, lr, lsl #16
eoreqs r0, r0, ip
moveq ip, lr, lsr #16
ldreq lr, [r1], #4
ldreq r0, [r4], #4
orreq ip, ip, lr, lsl #16
eoreqs r0, r0, ip
moveq ip, lr, lsr #16
ldreq lr, [r1], #4
ldreq r0, [r4], #4
orreq ip, ip, lr, lsl #16
eoreqs r0, r0, ip
bne 7f
subs r2, r2, #16
bhs 6b
sub r1, r1, #2
/* are we done? */
adds r2, r2, #16
moveq r0, #0
beq 9b
/* finish off the remaining bytes */
b 8b
7: /* fix up the 2 pointers and fallthrough... */
sub r1, r1, #(4+2)
sub r4, r4, #4
mov r2, #4
b 8b
4: /*************** offset is 1 or 3 (less optimized) ***************/
stmfd sp!, {r5, r6, r7}
// r5 = rhs
// r6 = lhs
// r7 = scratch
mov r5, r0, lsl #3 /* r5 = right shift */
rsb r6, r5, #32 /* r6 = left shift */
/* align the unaligned pointer */
bic r1, r1, #3
ldr r7, [r1], #4
sub r2, r2, #8
6: mov ip, r7, lsr r5
ldr r7, [r1], #4
ldr r0, [r4], #4
orr ip, ip, r7, lsl r6
eors r0, r0, ip
moveq ip, r7, lsr r5
ldreq r7, [r1], #4
ldreq r0, [r4], #4
orreq ip, ip, r7, lsl r6
eoreqs r0, r0, ip
bne 7f
subs r2, r2, #8
bhs 6b
sub r1, r1, r6, lsr #3
ldmfd sp!, {r5, r6, r7}
/* are we done? */
adds r2, r2, #8
moveq r0, #0
beq 9b
/* finish off the remaining bytes */
b 8b
7: /* fix up the 2 pointers and fallthrough... */
sub r1, r1, #4
sub r1, r1, r6, lsr #3
sub r4, r4, #4
mov r2, #4
ldmfd sp!, {r5, r6, r7}
b 8b
不知道是不是我缺乏什么库文件呢？请版主赐教。。感激不尽。。
