mirror of
https://github.com/AuxXxilium/linux_dsm_epyc7002.git
synced 2024-12-25 16:05:34 +07:00
83af58f806
Add performance-optimized versions of some string functions. Signed-off-by: Helge Deller <deller@gmx.de> Tested-by: Sven Schnelle <svens@stackframe.org>
137 lines
2.4 KiB
ArmAsm
137 lines
2.4 KiB
ArmAsm
// SPDX-License-Identifier: GPL-2.0
|
|
/*
|
|
* PA-RISC assembly string functions
|
|
*
|
|
* Copyright (C) 2019 Helge Deller <deller@gmx.de>
|
|
*/
|
|
|
|
#include <asm/assembly.h>
|
|
#include <linux/linkage.h>
|
|
|
|
.section .text.hot
|
|
.level PA_ASM_LEVEL
|
|
|
|
t0 = r20
|
|
t1 = r21
|
|
t2 = r22
|
|
|
|
ENTRY_CFI(strlen, frame=0,no_calls)
|
|
or,COND(<>) arg0,r0,ret0
|
|
b,l,n .Lstrlen_null_ptr,r0
|
|
depwi 0,31,2,ret0
|
|
cmpb,COND(<>) arg0,ret0,.Lstrlen_not_aligned
|
|
ldw,ma 4(ret0),t0
|
|
cmpib,tr 0,r0,.Lstrlen_loop
|
|
uxor,nbz r0,t0,r0
|
|
.Lstrlen_not_aligned:
|
|
uaddcm arg0,ret0,t1
|
|
shladd t1,3,r0,t1
|
|
mtsar t1
|
|
depwi -1,%sar,32,t0
|
|
uxor,nbz r0,t0,r0
|
|
.Lstrlen_loop:
|
|
b,l,n .Lstrlen_end_loop,r0
|
|
ldw,ma 4(ret0),t0
|
|
cmpib,tr 0,r0,.Lstrlen_loop
|
|
uxor,nbz r0,t0,r0
|
|
.Lstrlen_end_loop:
|
|
extrw,u,<> t0,7,8,r0
|
|
addib,tr,n -3,ret0,.Lstrlen_out
|
|
extrw,u,<> t0,15,8,r0
|
|
addib,tr,n -2,ret0,.Lstrlen_out
|
|
extrw,u,<> t0,23,8,r0
|
|
addi -1,ret0,ret0
|
|
.Lstrlen_out:
|
|
bv r0(rp)
|
|
uaddcm ret0,arg0,ret0
|
|
.Lstrlen_null_ptr:
|
|
bv,n r0(rp)
|
|
ENDPROC_CFI(strlen)
|
|
|
|
|
|
ENTRY_CFI(strcpy, frame=0,no_calls)
|
|
ldb 0(arg1),t0
|
|
stb t0,0(arg0)
|
|
ldo 0(arg0),ret0
|
|
ldo 1(arg1),t1
|
|
cmpb,= r0,t0,2f
|
|
ldo 1(arg0),t2
|
|
1: ldb 0(t1),arg1
|
|
stb arg1,0(t2)
|
|
ldo 1(t1),t1
|
|
cmpb,<> r0,arg1,1b
|
|
ldo 1(t2),t2
|
|
2: bv,n r0(rp)
|
|
ENDPROC_CFI(strcpy)
|
|
|
|
|
|
ENTRY_CFI(strncpy, frame=0,no_calls)
|
|
ldb 0(arg1),t0
|
|
stb t0,0(arg0)
|
|
ldo 1(arg1),t1
|
|
ldo 0(arg0),ret0
|
|
cmpb,= r0,t0,2f
|
|
ldo 1(arg0),arg1
|
|
1: ldo -1(arg2),arg2
|
|
cmpb,COND(=),n r0,arg2,2f
|
|
ldb 0(t1),arg0
|
|
stb arg0,0(arg1)
|
|
ldo 1(t1),t1
|
|
cmpb,<> r0,arg0,1b
|
|
ldo 1(arg1),arg1
|
|
2: bv,n r0(rp)
|
|
ENDPROC_CFI(strncpy)
|
|
|
|
|
|
ENTRY_CFI(strcat, frame=0,no_calls)
|
|
ldb 0(arg0),t0
|
|
cmpb,= t0,r0,2f
|
|
ldo 0(arg0),ret0
|
|
ldo 1(arg0),arg0
|
|
1: ldb 0(arg0),t1
|
|
cmpb,<>,n r0,t1,1b
|
|
ldo 1(arg0),arg0
|
|
2: ldb 0(arg1),t2
|
|
stb t2,0(arg0)
|
|
ldo 1(arg0),arg0
|
|
ldb 0(arg1),t0
|
|
cmpb,<> r0,t0,2b
|
|
ldo 1(arg1),arg1
|
|
bv,n r0(rp)
|
|
ENDPROC_CFI(strcat)
|
|
|
|
|
|
ENTRY_CFI(memset, frame=0,no_calls)
|
|
copy arg0,ret0
|
|
cmpb,COND(=) r0,arg0,4f
|
|
copy arg0,t2
|
|
cmpb,COND(=) r0,arg2,4f
|
|
ldo -1(arg2),arg3
|
|
subi -1,arg3,t0
|
|
subi 0,t0,t1
|
|
cmpiclr,COND(>=) 0,t1,arg2
|
|
ldo -1(t1),arg2
|
|
extru arg2,31,2,arg0
|
|
2: stb arg1,0(t2)
|
|
ldo 1(t2),t2
|
|
addib,>= -1,arg0,2b
|
|
ldo -1(arg3),arg3
|
|
cmpiclr,COND(<=) 4,arg2,r0
|
|
b,l,n 4f,r0
|
|
#ifdef CONFIG_64BIT
|
|
depd,* r0,63,2,arg2
|
|
#else
|
|
depw r0,31,2,arg2
|
|
#endif
|
|
ldo 1(t2),t2
|
|
3: stb arg1,-1(t2)
|
|
stb arg1,0(t2)
|
|
stb arg1,1(t2)
|
|
stb arg1,2(t2)
|
|
addib,COND(>) -4,arg2,3b
|
|
ldo 4(t2),t2
|
|
4: bv,n r0(rp)
|
|
ENDPROC_CFI(memset)
|
|
|
|
.end
|