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.
51 .globl csum_ipv6_magic
58 ldq $0,0($16) # L : Latency: 3
59 inslh $18,7,$4 # U : 0000000000AABBCC
60 ldq $1,8($16) # L : Latency: 3
61 sll $19,8,$7 # U : U L U L : 0x00000000 00aabb00
63 zapnot $20,15,$20 # U : zero extend incoming csum
64 ldq $2,0($17) # L : Latency: 3
65 sll $19,24,$19 # U : U L L U : 0x000000aa bb000000
66 inswl $18,3,$18 # U : 000000CCDD000000
68 ldq $3,8($17) # L : Latency: 3
69 bis $18,$4,$18 # E : 000000CCDDAABBCC
70 addl $19,$7,$19 # E : <sign bits>bbaabb00
73 addq $20,$0,$20 # E : begin summing the words
74 srl $18,16,$4 # U : 0000000000CCDDAA
75 zap $19,0x3,$19 # U : <sign bits>bbaa0000
78 cmpult $20,$0,$0 # E :
80 zapnot $18,0xa,$18 # U : 00000000DD00BB00
81 zap $4,0xa,$4 # U : U U L L : 0000000000CC00AA
83 or $18,$4,$18 # E : 00000000DDCCBBAA
85 cmpult $20,$1,$1 # E :
86 addq $20,$2,$20 # E : U L U L
88 cmpult $20,$2,$2 # E :
90 cmpult $20,$3,$3 # E : (1 cycle stall on $20)
91 addq $20,$18,$20 # E : U L U L (1 cycle stall on $20)
93 cmpult $20,$18,$18 # E :
94 addq $20,$19,$20 # E : (1 cycle stall on $20)
95 addq $0,$1,$0 # E : merge the carries back into the csum
98 cmpult $20,$19,$19 # E :
99 addq $18,$19,$18 # E : (1 cycle stall on $19)
101 addq $20,$18,$20 # E : U L U L :
102 /* (1 cycle stall on $18, 2 cycles on $20) */
105 zapnot $0,15,$1 # U : Start folding output (1 cycle stall on $0)
107 srl $0,32,$0 # U : U L U L : (1 cycle stall on $0)
109 addq $1,$0,$1 # E : Finished generating ulong
110 extwl $1,2,$2 # U : ushort[1] (1 cycle stall on $1)
111 zapnot $1,3,$0 # U : ushort[0] (1 cycle stall on $1)
112 extwl $1,4,$1 # U : ushort[2] (1 cycle stall on $1)
115 addq $0,$1,$3 # E : Finished generating uint
116 /* (1 cycle stall on $0) */
117 extwl $3,2,$1 # U : ushort[1] (1 cycle stall on $3)
120 addq $1,$3,$0 # E : Final carry
121 not $0,$4 # E : complement (1 cycle stall on $0)
122 zapnot $4,3,$0 # U : clear upper garbage bits
123 /* (1 cycle stall on $4) */