aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--arch/cris/arch-v32/lib/checksum.S72
-rw-r--r--arch/cris/arch-v32/lib/checksumcopy.S69
2 files changed, 43 insertions, 98 deletions
diff --git a/arch/cris/arch-v32/lib/checksum.S b/arch/cris/arch-v32/lib/checksum.S
index 32e66181b826..87f3fd71ab10 100644
--- a/arch/cris/arch-v32/lib/checksum.S
+++ b/arch/cris/arch-v32/lib/checksum.S
@@ -1,6 +1,6 @@
1/* 1/*
2 * A fast checksum routine using movem 2 * A fast checksum routine using movem
3 * Copyright (c) 1998-2001, 2003 Axis Communications AB 3 * Copyright (c) 1998-2007 Axis Communications AB
4 * 4 *
5 * csum_partial(const unsigned char * buff, int len, unsigned int sum) 5 * csum_partial(const unsigned char * buff, int len, unsigned int sum)
6 */ 6 */
@@ -12,30 +12,23 @@ csum_partial:
12 ;; r11 - length 12 ;; r11 - length
13 ;; r12 - checksum 13 ;; r12 - checksum
14 14
15 ;; check for breakeven length between movem and normal word looping versions 15 ;; Optimized for large packets
16 ;; we also do _NOT_ want to compute a checksum over more than the 16 subq 10*4, $r11
17 ;; actual length when length < 40 17 blt _word_loop
18 18 move.d $r11, $acr
19 cmpu.w 80,$r11
20 blo _word_loop
21 nop
22
23 ;; need to save the registers we use below in the movem loop
24 ;; this overhead is why we have a check above for breakeven length
25 ;; only r0 - r8 have to be saved, the other ones are clobber-able
26 ;; according to the ABI
27 19
28 subq 9*4,$sp 20 subq 9*4,$sp
29 subq 10*4,$r11 ; update length for the first loop 21 clearf c
30 movem $r8,[$sp] 22 movem $r8,[$sp]
31 23
32 ;; do a movem checksum 24 ;; do a movem checksum
33 25
34_mloop: movem [$r10+],$r9 ; read 10 longwords 26_mloop: movem [$r10+],$r9 ; read 10 longwords
35 27 ;; Loop count without touching the c flag.
28 addoq -10*4, $acr, $acr
36 ;; perform dword checksumming on the 10 longwords 29 ;; perform dword checksumming on the 10 longwords
37 30
38 add.d $r0,$r12 31 addc $r0,$r12
39 addc $r1,$r12 32 addc $r1,$r12
40 addc $r2,$r12 33 addc $r2,$r12
41 addc $r3,$r12 34 addc $r3,$r12
@@ -46,60 +39,41 @@ _mloop: movem [$r10+],$r9 ; read 10 longwords
46 addc $r8,$r12 39 addc $r8,$r12
47 addc $r9,$r12 40 addc $r9,$r12
48 41
49 ;; fold the carry into the checksum, to avoid having to loop the carry 42 ;; test $acr without trashing carry.
50 ;; back into the top 43 move.d $acr, $acr
51 44 bpl _mloop
52 addc 0,$r12 45 ;; r11 <= acr is not really needed in the mloop, just using the dslot
53 addc 0,$r12 ; do it again, since we might have generated a carry 46 ;; to prepare for what is needed after mloop.
54 47 move.d $acr, $r11
55 subq 10*4,$r11
56 bge _mloop
57 nop
58
59 addq 10*4,$r11 ; compensate for last loop underflowing length
60 48
49 ;; fold the last carry into r13
50 addc 0, $r12
61 movem [$sp+],$r8 ; restore regs 51 movem [$sp+],$r8 ; restore regs
62 52
63_word_loop: 53_word_loop:
64 ;; only fold if there is anything to fold. 54 addq 10*4,$r11 ; compensate for last loop underflowing length
65
66 cmpq 0,$r12
67 beq _no_fold
68
69 ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below.
70 ;; r9 and r13 can be used as temporaries.
71 55
72 moveq -1,$r9 ; put 0xffff in r9, faster than move.d 0xffff,r9 56 moveq -1,$r9 ; put 0xffff in r9, faster than move.d 0xffff,r9
73 lsrq 16,$r9 57 lsrq 16,$r9
74 58
75 move.d $r12,$r13 59 move.d $r12,$r13
76 lsrq 16,$r13 ; r13 = checksum >> 16 60 lsrq 16,$r13 ; r13 = checksum >> 16
77 and.d $r9,$r12 ; checksum = checksum & 0xffff 61 and.d $r9,$r12 ; checksum = checksum & 0xffff
78 add.d $r13,$r12 ; checksum += r13
79 move.d $r12,$r13 ; do the same again, maybe we got a carry last add
80 lsrq 16,$r13
81 and.d $r9,$r12
82 add.d $r13,$r12
83 62
84_no_fold: 63_no_fold:
85 cmpq 2,$r11 64 subq 2,$r11
86 blt _no_words 65 blt _no_words
87 nop 66 add.d $r13,$r12 ; checksum += r13
88 67
89 ;; checksum the rest of the words 68 ;; checksum the rest of the words
90
91 subq 2,$r11
92
93_wloop: subq 2,$r11 69_wloop: subq 2,$r11
94 bge _wloop 70 bge _wloop
95 addu.w [$r10+],$r12 71 addu.w [$r10+],$r12
96 72
97 addq 2,$r11
98
99_no_words: 73_no_words:
74 addq 2,$r11
100 ;; see if we have one odd byte more 75 ;; see if we have one odd byte more
101 cmpq 1,$r11 76 bne _do_byte
102 beq _do_byte
103 nop 77 nop
104 ret 78 ret
105 move.d $r12,$r10 79 move.d $r12,$r10
diff --git a/arch/cris/arch-v32/lib/checksumcopy.S b/arch/cris/arch-v32/lib/checksumcopy.S
index 9303ccbadc6d..21aabe91489b 100644
--- a/arch/cris/arch-v32/lib/checksumcopy.S
+++ b/arch/cris/arch-v32/lib/checksumcopy.S
@@ -1,6 +1,6 @@
1/* 1/*
2 * A fast checksum+copy routine using movem 2 * A fast checksum+copy routine using movem
3 * Copyright (c) 1998, 2001, 2003 Axis Communications AB 3 * Copyright (c) 1998-2007 Axis Communications AB
4 * 4 *
5 * Authors: Bjorn Wesen 5 * Authors: Bjorn Wesen
6 * 6 *
@@ -16,32 +16,23 @@ csum_partial_copy_nocheck:
16 ;; r12 - length 16 ;; r12 - length
17 ;; r13 - checksum 17 ;; r13 - checksum
18 18
19 ;; check for breakeven length between movem and normal word looping versions 19 ;; Optimized for large packets
20 ;; we also do _NOT_ want to compute a checksum over more than the 20 subq 10*4, $r12
21 ;; actual length when length < 40 21 blt _word_loop
22 22 move.d $r12, $acr
23 cmpu.w 80,$r12
24 blo _word_loop
25 nop
26
27 ;; need to save the registers we use below in the movem loop
28 ;; this overhead is why we have a check above for breakeven length
29 ;; only r0 - r8 have to be saved, the other ones are clobber-able
30 ;; according to the ABI
31 23
32 subq 9*4,$sp 24 subq 9*4,$sp
33 subq 10*4,$r12 ; update length for the first loop 25 clearf c
34 movem $r8,[$sp] 26 movem $r8,[$sp]
35 27
36 ;; do a movem copy and checksum 28 ;; do a movem copy and checksum
37
381: ;; A failing userspace access (the read) will have this as PC. 291: ;; A failing userspace access (the read) will have this as PC.
39_mloop: movem [$r10+],$r9 ; read 10 longwords 30_mloop: movem [$r10+],$r9 ; read 10 longwords
31 addoq -10*4, $acr, $acr ; loop counter in latency cycle
40 movem $r9,[$r11+] ; write 10 longwords 32 movem $r9,[$r11+] ; write 10 longwords
41 33
42 ;; perform dword checksumming on the 10 longwords 34 ;; perform dword checksumming on the 10 longwords
43 35 addc $r0,$r13
44 add.d $r0,$r13
45 addc $r1,$r13 36 addc $r1,$r13
46 addc $r2,$r13 37 addc $r2,$r13
47 addc $r3,$r13 38 addc $r3,$r13
@@ -52,47 +43,30 @@ _mloop: movem [$r10+],$r9 ; read 10 longwords
52 addc $r8,$r13 43 addc $r8,$r13
53 addc $r9,$r13 44 addc $r9,$r13
54 45
55 ;; fold the carry into the checksum, to avoid having to loop the carry 46 ;; test $acr, without trashing carry.
56 ;; back into the top 47 move.d $acr, $acr
57 48 bpl _mloop
58 addc 0,$r13 49 ;; r12 <= acr is needed after mloop and in the exception handlers.
59 addc 0,$r13 ; do it again, since we might have generated a carry 50 move.d $acr, $r12
60
61 subq 10*4,$r12
62 bge _mloop
63 nop
64
65 addq 10*4,$r12 ; compensate for last loop underflowing length
66 51
52 ;; fold the last carry into r13
53 addc 0, $r13
67 movem [$sp+],$r8 ; restore regs 54 movem [$sp+],$r8 ; restore regs
68 55
69_word_loop: 56_word_loop:
70 ;; only fold if there is anything to fold. 57 addq 10*4,$r12 ; compensate for last loop underflowing length
71
72 cmpq 0,$r13
73 beq _no_fold
74 58
75 ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below 59 ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below
76 ;; r9 can be used as temporary. 60 ;; r9 can be used as temporary.
77
78 move.d $r13,$r9 61 move.d $r13,$r9
79 lsrq 16,$r9 ; r0 = checksum >> 16 62 lsrq 16,$r9 ; r0 = checksum >> 16
80 and.d 0xffff,$r13 ; checksum = checksum & 0xffff 63 and.d 0xffff,$r13 ; checksum = checksum & 0xffff
81 add.d $r9,$r13 ; checksum += r0
82 move.d $r13,$r9 ; do the same again, maybe we got a carry last add
83 lsrq 16,$r9
84 and.d 0xffff,$r13
85 add.d $r9,$r13
86 64
87_no_fold: 65 subq 2, $r12
88 cmpq 2,$r12
89 blt _no_words 66 blt _no_words
90 nop 67 add.d $r9,$r13 ; checksum += r0
91 68
92 ;; copy and checksum the rest of the words 69 ;; copy and checksum the rest of the words
93
94 subq 2,$r12
95
962: ;; A failing userspace access for the read below will have this as PC. 702: ;; A failing userspace access for the read below will have this as PC.
97_wloop: move.w [$r10+],$r9 71_wloop: move.w [$r10+],$r9
98 addu.w $r9,$r13 72 addu.w $r9,$r13
@@ -100,12 +74,9 @@ _wloop: move.w [$r10+],$r9
100 bge _wloop 74 bge _wloop
101 move.w $r9,[$r11+] 75 move.w $r9,[$r11+]
102 76
103 addq 2,$r12
104
105_no_words: 77_no_words:
106 ;; see if we have one odd byte more 78 addq 2,$r12
107 cmpq 1,$r12 79 bne _do_byte
108 beq _do_byte
109 nop 80 nop
110 ret 81 ret
111 move.d $r13,$r10 82 move.d $r13,$r10