forked from rrcarlosr/Jetpack
127 lines
2.4 KiB
ArmAsm
127 lines
2.4 KiB
ArmAsm
|
/*
|
||
|
|
* A fast checksum+copy routine using movem
|
||
|
|
* Copyright (c) 1998, 2001 Axis Communications AB
|
||
|
|
*
|
||
|
|
* Authors: Bjorn Wesen
|
||
|
|
*
|
||
|
|
* csum_partial_copy_nocheck(const char *src, char *dst,
|
||
|
|
* int len, unsigned int sum)
|
||
|
|
*/
|
||
|
|
|
||
|
|
.globl csum_partial_copy_nocheck
|
||
|
|
csum_partial_copy_nocheck:
|
||
|
|
|
||
|
|
;; r10 - src
|
||
|
|
;; r11 - dst
|
||
|
|
;; r12 - length
|
||
|
|
;; r13 - checksum
|
||
|
|
|
||
|
|
;; check for breakeven length between movem and normal word looping versions
|
||
|
|
;; we also do _NOT_ want to compute a checksum over more than the
|
||
|
|
;; actual length when length < 40
|
||
|
|
|
||
|
|
cmpu.w 80, $r12
|
||
|
|
blo _word_loop
|
||
|
|
nop
|
||
|
|
|
||
|
|
;; need to save the registers we use below in the movem loop
|
||
|
|
;; this overhead is why we have a check above for breakeven length
|
||
|
|
;; only r0 - r8 have to be saved, the other ones are clobber-able
|
||
|
|
;; according to the ABI
|
||
|
|
|
||
|
|
subq 9*4, $sp
|
||
|
|
movem $r8, [$sp]
|
||
|
|
|
||
|
|
;; do a movem copy and checksum
|
||
|
|
|
||
|
|
subq 10*4, $r12 ; update length for the first loop
|
||
|
|
|
||
|
|
_mloop: movem [$r10+],$r9 ; read 10 longwords
|
||
|
|
1: ;; A failing userspace access will have this as PC.
|
||
|
|
movem $r9,[$r11+] ; write 10 longwords
|
||
|
|
|
||
|
|
;; perform dword checksumming on the 10 longwords
|
||
|
|
|
||
|
|
add.d $r0,$r13
|
||
|
|
ax
|
||
|
|
add.d $r1,$r13
|
||
|
|
ax
|
||
|
|
add.d $r2,$r13
|
||
|
|
ax
|
||
|
|
add.d $r3,$r13
|
||
|
|
ax
|
||
|
|
add.d $r4,$r13
|
||
|
|
ax
|
||
|
|
add.d $r5,$r13
|
||
|
|
ax
|
||
|
|
add.d $r6,$r13
|
||
|
|
ax
|
||
|
|
add.d $r7,$r13
|
||
|
|
ax
|
||
|
|
add.d $r8,$r13
|
||
|
|
ax
|
||
|
|
add.d $r9,$r13
|
||
|
|
|
||
|
|
;; fold the carry into the checksum, to avoid having to loop the carry
|
||
|
|
;; back into the top
|
||
|
|
|
||
|
|
ax
|
||
|
|
addq 0,$r13
|
||
|
|
|
||
|
|
subq 10*4,$r12
|
||
|
|
bge _mloop
|
||
|
|
nop
|
||
|
|
|
||
|
|
addq 10*4,$r12 ; compensate for last loop underflowing length
|
||
|
|
|
||
|
|
movem [$sp+],$r8 ; restore regs
|
||
|
|
|
||
|
|
_word_loop:
|
||
|
|
;; only fold if there is anything to fold.
|
||
|
|
|
||
|
|
cmpq 0,$r13
|
||
|
|
beq _no_fold
|
||
|
|
|
||
|
|
;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below
|
||
|
|
;; r9 can be used as temporary.
|
||
|
|
|
||
|
|
move.d $r13,$r9
|
||
|
|
lsrq 16,$r9 ; r0 = checksum >> 16
|
||
|
|
and.d 0xffff,$r13 ; checksum = checksum & 0xffff
|
||
|
|
add.d $r9,$r13 ; checksum += r0
|
||
|
|
|
||
|
|
_no_fold:
|
||
|
|
cmpq 2,$r12
|
||
|
|
blt _no_words
|
||
|
|
nop
|
||
|
|
|
||
|
|
;; copy and checksum the rest of the words
|
||
|
|
|
||
|
|
subq 2,$r12
|
||
|
|
|
||
|
|
_wloop: move.w [$r10+],$r9
|
||
|
|
2: ;; A failing userspace access will have this as PC.
|
||
|
|
addu.w $r9,$r13
|
||
|
|
subq 2,$r12
|
||
|
|
bge _wloop
|
||
|
|
move.w $r9,[$r11+]
|
||
|
|
|
||
|
|
addq 2,$r12
|
||
|
|
|
||
|
|
_no_words:
|
||
|
|
;; see if we have one odd byte more
|
||
|
|
cmpq 1,$r12
|
||
|
|
beq _do_byte
|
||
|
|
nop
|
||
|
|
ret
|
||
|
|
move.d $r13, $r10
|
||
|
|
|
||
|
|
_do_byte:
|
||
|
|
;; copy and checksum the last byte
|
||
|
|
move.b [$r10],$r9
|
||
|
|
3: ;; A failing userspace access will have this as PC.
|
||
|
|
addu.b $r9,$r13
|
||
|
|
move.b $r9,[$r11]
|
||
|
|
ret
|
||
|
|
move.d $r13, $r10
|