2 * arch/alpha/lib/ev6-csum_ipv6_magic.S
3 * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
5 * unsigned short csum_ipv6_magic(struct in6_addr *saddr,
6 * struct in6_addr *daddr,
8 * unsigned short proto,
11 * Much of the information about 21264 scheduling/coding comes from:
12 * Compiler Writer's Guide for the Alpha 21264
13 * abbreviated as 'CWG' in other comments here
14 * ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
15 * Scheduling notation:
17 * U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1
18 * L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1
19 * Try not to change the actual algorithm if possible for consistency.
20 * Determining actual stalls (other than slotting) doesn't appear to be easy to do.
22 * unsigned short csum_ipv6_magic(struct in6_addr *saddr,
23 * struct in6_addr *daddr,
25 * unsigned short proto,
28 * Swap <proto> (takes form 0xaabb)
29 * Then shift it left by 48, so result is:
31 * Then turn it back into a sign extended 32-bit item
34 * Swap <len> (an unsigned int) using Mike Burrows' 7-instruction sequence
35 * (we can't hide the 3-cycle latency of the unpkbw in the 6-instruction sequence)
36 * Assume input takes form 0xAABBCCDD
38 * Finally, original 'folding' approach is to split the long into 4 unsigned shorts
39 * add 4 ushorts, resulting in ushort/carry
40 * add carry bits + ushort --> ushort
41 * add carry bits + ushort --> ushort (in case the carry results in an overflow)
42 * Truncate to a ushort. (took 13 instructions)
43 * From doing some testing, using the approach in checksum.c:from64to16()
44 * results in the same outcome:
45 * split into 2 uints, add those, generating a ulong
46 * add the 3 low ushorts together, generating a uint
47 * a final add of the 2 lower ushorts
48 * truncating the result.
50 * Misalignment handling added by Ivan Kokshaysky <ink@jurassic.park.msu.ru>
51 * The cost is 16 instructions (~8 cycles), including two extra loads which
52 * may cause additional delay in rare cases (load-load replay traps).
55 .globl csum_ipv6_magic
62 ldq_u $0,0($16) # L : Latency: 3
63 inslh $18,7,$4 # U : 0000000000AABBCC
64 ldq_u $1,8($16) # L : Latency: 3
65 sll $19,8,$7 # U : U L U L : 0x00000000 00aabb00
67 and $16,7,$6 # E : src misalignment
68 ldq_u $5,15($16) # L : Latency: 3
69 zapnot $20,15,$20 # U : zero extend incoming csum
70 ldq_u $2,0($17) # L : U L U L : Latency: 3
74 ldq_u $3,8($17) # L : Latency: 3
75 sll $19,24,$19 # U : U U L U : 0x000000aa bb000000
77 cmoveq $6,$31,$22 # E : src aligned?
78 ldq_u $23,15($17) # L : Latency: 3
79 inswl $18,3,$18 # U : 000000CCDD000000
80 addl $19,$7,$19 # E : U L U L : <sign bits>bbaabb00
82 or $0,$22,$0 # E : 1st src word complete
84 or $18,$4,$18 # E : 000000CCDDAABBCC
85 extqh $5,$6,$5 # U : L U L U
87 and $17,7,$6 # E : dst misalignment
89 or $1,$5,$1 # E : 2nd src word complete
90 extqh $3,$6,$22 # U : L U L U :
92 cmoveq $6,$31,$22 # E : dst aligned?
94 addq $20,$0,$20 # E : begin summing the words
95 extqh $23,$6,$23 # U : L U L U :
97 srl $18,16,$4 # U : 0000000000CCDDAA
98 or $2,$22,$2 # E : 1st dst word complete
99 zap $19,0x3,$19 # U : <sign bits>bbaa0000
100 or $3,$23,$3 # E : U L U L : 2nd dst word complete
102 cmpult $20,$0,$0 # E :
103 addq $20,$1,$20 # E :
104 zapnot $18,0xa,$18 # U : 00000000DD00BB00
105 zap $4,0xa,$4 # U : U U L L : 0000000000CC00AA
107 or $18,$4,$18 # E : 00000000DDCCBBAA
109 cmpult $20,$1,$1 # E :
110 addq $20,$2,$20 # E : U L U L
112 cmpult $20,$2,$2 # E :
113 addq $20,$3,$20 # E :
114 cmpult $20,$3,$3 # E : (1 cycle stall on $20)
115 addq $20,$18,$20 # E : U L U L (1 cycle stall on $20)
117 cmpult $20,$18,$18 # E :
118 addq $20,$19,$20 # E : (1 cycle stall on $20)
119 addq $0,$1,$0 # E : merge the carries back into the csum
122 cmpult $20,$19,$19 # E :
123 addq $18,$19,$18 # E : (1 cycle stall on $19)
125 addq $20,$18,$20 # E : U L U L :
126 /* (1 cycle stall on $18, 2 cycles on $20) */
129 zapnot $0,15,$1 # U : Start folding output (1 cycle stall on $0)
131 srl $0,32,$0 # U : U L U L : (1 cycle stall on $0)
133 addq $1,$0,$1 # E : Finished generating ulong
134 extwl $1,2,$2 # U : ushort[1] (1 cycle stall on $1)
135 zapnot $1,3,$0 # U : ushort[0] (1 cycle stall on $1)
136 extwl $1,4,$1 # U : ushort[2] (1 cycle stall on $1)
139 addq $0,$1,$3 # E : Finished generating uint
140 /* (1 cycle stall on $0) */
141 extwl $3,2,$1 # U : ushort[1] (1 cycle stall on $3)
144 addq $1,$3,$0 # E : Final carry
145 not $0,$4 # E : complement (1 cycle stall on $0)
146 zapnot $4,3,$0 # U : clear upper garbage bits
147 /* (1 cycle stall on $4) */