Use 64-bit-clean addition/subtraction opcodes for pointer arithmetic iff
_LP64 is defined.
This commit is contained in:
sekiya
parent
7f549a8679
commit
77aa58c5b4
@ -1,10 +1,18 @@
|
||||
/* $NetBSD: memcpy.S,v 1.7 2001/10/16 15:40:53 uch Exp $ */
|
||||
/* $NetBSD: memcpy.S,v 1.8 2004/09/29 04:45:26 sekiya Exp $ */
|
||||
/* XXXX We need to define this in a way which supports multiple architectures */
|
||||
|
||||
#include <machine/cdefs.h> /* Get SZREG correct */
|
||||
#include <mips/asm.h>
|
||||
#include <machine/endian.h>
|
||||
|
||||
#if defined(_LP64)
|
||||
#define ADDU daddu
|
||||
#define SUBU dsubu
|
||||
#else
|
||||
#define ADDU addu
|
||||
#define SUBU subu
|
||||
#endif
|
||||
|
||||
.set push
|
||||
.set noreorder
|
||||
|
||||
@ -32,7 +40,7 @@ XLEAF(memcpy)
|
||||
* Make sure we can copy forwards.
|
||||
*/
|
||||
sltu t0,a1,a0 # t0 == a1 < a0
|
||||
addu a3,a1,a2 # a3 == end of source
|
||||
ADDU a3,a1,a2 # a3 == end of source
|
||||
sltu t1,a0,a3 # t1 == a0 < a1+a2
|
||||
and t2,t0,t1 # overlap -- copy backwards
|
||||
bne t2,zero,backcopy
|
||||
@ -61,10 +69,10 @@ XLEAF(memcpy)
|
||||
*/
|
||||
li AT,-32 # BDSLOT
|
||||
and t0,a2,AT # count truncated to multiple of 32
|
||||
addu a3,a1,t0 # run fast loop up to this address
|
||||
ADDU a3,a1,t0 # run fast loop up to this address
|
||||
sltu AT,a1,a3 # any work to do?
|
||||
beq AT,zero,wordcopy
|
||||
subu a2,t0 # BDSLOT
|
||||
SUBU a2,t0 # BDSLOT
|
||||
|
||||
/*
|
||||
* loop body
|
||||
@ -81,13 +89,13 @@ cp8:
|
||||
ld v1,8(a1)
|
||||
ld t0,16(a1)
|
||||
ld t1,24(a1)
|
||||
addu a1,32
|
||||
ADDU a1,32
|
||||
sd t3,0(a0)
|
||||
sd v1,8(a0)
|
||||
sd t0,16(a0)
|
||||
sd t1,24(a0)
|
||||
bne a1,a3,cp8
|
||||
addu a0,32 # BDSLOT
|
||||
ADDU a0,32 # BDSLOT
|
||||
|
||||
b wordcopy
|
||||
nop # BDSLOT
|
||||
@ -97,7 +105,7 @@ cp:
|
||||
lw v1,4(a1)
|
||||
lw t0,8(a1)
|
||||
lw t1,12(a1)
|
||||
addu a1,32
|
||||
ADDU a1,32
|
||||
sw t3,0(a0)
|
||||
sw v1,4(a0)
|
||||
sw t0,8(a0)
|
||||
@ -106,7 +114,7 @@ cp:
|
||||
lw t0,-8(a1)
|
||||
lw v1,-12(a1)
|
||||
lw t3,-16(a1)
|
||||
addu a0,32
|
||||
ADDU a0,32
|
||||
sw t1,-4(a0)
|
||||
sw t0,-8(a0)
|
||||
sw v1,-12(a0)
|
||||
@ -118,34 +126,34 @@ cp:
|
||||
*/
|
||||
wordcopy:
|
||||
andi t2,a2,3 # get byte count / 4
|
||||
subu t2,a2,t2 # t2 = number of words to copy * 4
|
||||
SUBU t2,a2,t2 # t2 = number of words to copy * 4
|
||||
beq t2,zero,bytecopy
|
||||
addu t0,a1,t2 # BDSLOT stop at t0
|
||||
subu a2,a2,t2
|
||||
ADDU t0,a1,t2 # BDSLOT stop at t0
|
||||
SUBU a2,a2,t2
|
||||
1:
|
||||
lw t3,0(a1)
|
||||
addu a1,4
|
||||
ADDU a1,4
|
||||
sw t3,0(a0)
|
||||
#ifdef MIPS3_5900
|
||||
nop
|
||||
nop
|
||||
#endif
|
||||
bne a1,t0,1b
|
||||
addu a0,4 # BDSLOT
|
||||
ADDU a0,4 # BDSLOT
|
||||
|
||||
bytecopy:
|
||||
beq a2,zero,copydone # nothing left to do?
|
||||
nop
|
||||
2:
|
||||
lb t3,0(a1)
|
||||
addu a1,1
|
||||
ADDU a1,1
|
||||
sb t3,0(a0)
|
||||
subu a2,1
|
||||
SUBU a2,1
|
||||
#ifdef MIPS3_5900
|
||||
nop
|
||||
#endif
|
||||
bgtz a2,2b
|
||||
addu a0,1 # BDSLOT
|
||||
ADDU a0,1 # BDSLOT
|
||||
|
||||
copydone:
|
||||
j ra
|
||||
@ -155,9 +163,9 @@ copydone:
|
||||
*/
|
||||
destaligned:
|
||||
andi t0,a2,3 # t0 = bytecount mod 4
|
||||
subu a3,a2,t0 # number of words to transfer
|
||||
SUBU a3,a2,t0 # number of words to transfer
|
||||
beq a3,zero,bytecopy
|
||||
addu a3,a1,a3 # BDSLOT: stop point for destaligned
|
||||
ADDU a3,a1,a3 # BDSLOT: stop point for destaligned
|
||||
move a2,t0 # this many to do after we are done
|
||||
|
||||
3:
|
||||
@ -179,18 +187,18 @@ destaligned:
|
||||
*/
|
||||
backcopy:
|
||||
blez a2,copydone # nothing left to do?
|
||||
addu t0,a1,a2 # BDSLOT: end of source
|
||||
addu t1,a0,a2 # end of destination
|
||||
ADDU t0,a1,a2 # BDSLOT: end of source
|
||||
ADDU t1,a0,a2 # end of destination
|
||||
4:
|
||||
lb t3,-1(t0)
|
||||
subu t0,1
|
||||
SUBU t0,1
|
||||
sb t3,-1(t1)
|
||||
#ifdef MIPS3_5900
|
||||
nop
|
||||
nop
|
||||
#endif
|
||||
bne t0,a1,4b
|
||||
subu t1,1 # BDSLOT
|
||||
SUBU t1,1 # BDSLOT
|
||||
j ra
|
||||
nop
|
||||
|
||||
|
||||
@ -1,8 +1,16 @@
|
||||
/* $NetBSD: memset.S,v 1.3 2001/10/16 15:40:53 uch Exp $ */
|
||||
/* $NetBSD: memset.S,v 1.4 2004/09/29 04:45:26 sekiya Exp $ */
|
||||
#include <machine/cdefs.h>
|
||||
#include <mips/asm.h>
|
||||
#include <machine/endian.h>
|
||||
|
||||
#if defined(_LP64)
|
||||
#define ADDU daddu
|
||||
#define SUBU dsubu
|
||||
#else
|
||||
#define ADDU addu
|
||||
#define SUBU subu
|
||||
#endif
|
||||
|
||||
.set noreorder
|
||||
|
||||
|
||||
@ -19,19 +27,19 @@ LEAF(memset)
|
||||
sll t2, t1, 16 # shift that left 16
|
||||
or t1, t2, t1 # or together
|
||||
|
||||
subu t0, zero, a0 # compute # bytes to word align address
|
||||
SUBU t0, zero, a0 # compute # bytes to word align address
|
||||
and t0, t0, 3
|
||||
beq t0, zero, 1f # skip if word aligned
|
||||
subu a2, a2, t0 # subtract from remaining count
|
||||
SUBU a2, a2, t0 # subtract from remaining count
|
||||
SWHI t1, 0(a0) # store 1, 2, or 3 bytes to align
|
||||
addu a0, a0, t0
|
||||
ADDU a0, a0, t0
|
||||
1:
|
||||
and v1, a2, 3 # compute number of whole words left
|
||||
subu t0, a2, v1
|
||||
subu a2, a2, t0
|
||||
addu t0, t0, a0 # compute ending address
|
||||
SUBU t0, a2, v1
|
||||
SUBU a2, a2, t0
|
||||
ADDU t0, t0, a0 # compute ending address
|
||||
2:
|
||||
addu a0, a0, 4 # clear words
|
||||
ADDU a0, a0, 4 # clear words
|
||||
#ifdef MIPS3_5900
|
||||
nop
|
||||
nop
|
||||
@ -43,9 +51,9 @@ LEAF(memset)
|
||||
|
||||
smallclr:
|
||||
ble a2, zero, 2f
|
||||
addu t0, a2, a0 # compute ending address
|
||||
ADDU t0, a2, a0 # compute ending address
|
||||
1:
|
||||
addu a0, a0, 1 # clear bytes
|
||||
ADDU a0, a0, 1 # clear bytes
|
||||
#ifdef MIPS3_5900
|
||||
nop
|
||||
nop
|
||||
|
||||
@ -1,7 +1,15 @@
|
||||
/* $NetBSD: strcmp.S,v 1.1 1999/01/15 08:44:27 castor Exp $ */
|
||||
/* $NetBSD: strcmp.S,v 1.2 2004/09/29 04:45:26 sekiya Exp $ */
|
||||
#include <machine/cdefs.h>
|
||||
#include <mips/asm.h>
|
||||
|
||||
#if defined(_LP64)
|
||||
#define ADDU daddu
|
||||
#define SUBU dsubu
|
||||
#else
|
||||
#define ADDU addu
|
||||
#define SUBU subu
|
||||
#endif
|
||||
|
||||
.set noreorder
|
||||
|
||||
/*
|
||||
@ -18,13 +26,13 @@ LEAF(strcmp)
|
||||
lbu t0, 1(a0) # unroll loop
|
||||
lbu t1, 1(a1)
|
||||
beq t0, zero, LessOrEq # end of first string?
|
||||
addu a0, a0, 2
|
||||
ADDU a0, a0, 2
|
||||
beq t0, t1, 1b
|
||||
addu a1, a1, 2
|
||||
ADDU a1, a1, 2
|
||||
NotEq:
|
||||
j ra
|
||||
subu v0, t0, t1
|
||||
SUBU v0, t0, t1
|
||||
LessOrEq:
|
||||
j ra
|
||||
subu v0, zero, t1
|
||||
SUBU v0, zero, t1
|
||||
END(strcmp)
|
||||
|
||||
@ -1,19 +1,27 @@
|
||||
/* $NetBSD: strlen.S,v 1.1 1999/01/15 08:44:27 castor Exp $ */
|
||||
/* $NetBSD: strlen.S,v 1.2 2004/09/29 04:45:26 sekiya Exp $ */
|
||||
#include <machine/cdefs.h>
|
||||
#include <mips/asm.h>
|
||||
|
||||
#if defined(_LP64)
|
||||
#define ADDU daddu
|
||||
#define SUBU dsubu
|
||||
#else
|
||||
#define ADDU addu
|
||||
#define SUBU subu
|
||||
#endif
|
||||
|
||||
.set noreorder
|
||||
|
||||
/*
|
||||
* strlen(str)
|
||||
*/
|
||||
LEAF(strlen)
|
||||
addu v1, a0, 1
|
||||
ADDU v1, a0, 1
|
||||
1:
|
||||
lb v0, 0(a0) # get byte from string
|
||||
addu a0, a0, 1 # increment pointer
|
||||
ADDU a0, a0, 1 # increment pointer
|
||||
bne v0, zero, 1b # continue if not end
|
||||
nop
|
||||
j ra
|
||||
subu v0, a0, v1 # compute length - 1 for '\0' char
|
||||
SUBU v0, a0, v1 # compute length - 1 for '\0' char
|
||||
END(strlen)
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user