[PATCH] x86-64: Micro optimization to dma_alloc_coherent node lookup
[linux-2.6] / arch / x86_64 / lib / csum-partial.c
1 /*
2  * arch/x86_64/lib/csum-partial.c
3  *
4  * This file contains network checksum routines that are better done
5  * in an architecture-specific manner due to speed.
6  */
7  
8 #include <linux/compiler.h>
9 #include <linux/module.h>
10 #include <asm/checksum.h>
11
12 #define __force_inline inline __attribute__((always_inline))
13
14 static inline unsigned short from32to16(unsigned a) 
15 {
16         unsigned short b = a >> 16; 
17         asm("addw %w2,%w0\n\t"
18             "adcw $0,%w0\n" 
19             : "=r" (b)
20             : "0" (b), "r" (a));
21         return b;
22 }
23
24 /*
25  * Do a 64-bit checksum on an arbitrary memory area.
26  * Returns a 32bit checksum.
27  *
28  * This isn't as time critical as it used to be because many NICs
29  * do hardware checksumming these days.
30  * 
31  * Things tried and found to not make it faster:
32  * Manual Prefetching
33  * Unrolling to an 128 bytes inner loop.
34  * Using interleaving with more registers to break the carry chains.
35  */
36 static __force_inline unsigned do_csum(const unsigned char *buff, unsigned len)
37 {
38         unsigned odd, count;
39         unsigned long result = 0;
40
41         if (unlikely(len == 0))
42                 return result; 
43         odd = 1 & (unsigned long) buff;
44         if (unlikely(odd)) {
45                 result = *buff << 8;
46                 len--;
47                 buff++;
48         }
49         count = len >> 1;               /* nr of 16-bit words.. */
50         if (count) {
51                 if (2 & (unsigned long) buff) {
52                         result += *(unsigned short *)buff;
53                         count--;
54                         len -= 2;
55                         buff += 2;
56                 }
57                 count >>= 1;            /* nr of 32-bit words.. */
58                 if (count) {
59                         unsigned long zero;
60                         unsigned count64;
61                         if (4 & (unsigned long) buff) {
62                                 result += *(unsigned int *) buff;
63                                 count--;
64                                 len -= 4;
65                                 buff += 4;
66                         }
67                         count >>= 1;    /* nr of 64-bit words.. */
68
69                         /* main loop using 64byte blocks */
70                         zero = 0;
71                         count64 = count >> 3;
72                         while (count64) { 
73                                 asm("addq 0*8(%[src]),%[res]\n\t"
74                                     "adcq 1*8(%[src]),%[res]\n\t"
75                                     "adcq 2*8(%[src]),%[res]\n\t"
76                                     "adcq 3*8(%[src]),%[res]\n\t"
77                                     "adcq 4*8(%[src]),%[res]\n\t"
78                                     "adcq 5*8(%[src]),%[res]\n\t"
79                                     "adcq 6*8(%[src]),%[res]\n\t"
80                                     "adcq 7*8(%[src]),%[res]\n\t"
81                                     "adcq %[zero],%[res]"
82                                     : [res] "=r" (result)
83                                     : [src] "r" (buff), [zero] "r" (zero),
84                                     "[res]" (result));
85                                 buff += 64;
86                                 count64--;
87                         }
88
89                         /* last upto 7 8byte blocks */
90                         count %= 8; 
91                         while (count) { 
92                                 asm("addq %1,%0\n\t"
93                                     "adcq %2,%0\n" 
94                                             : "=r" (result)
95                                     : "m" (*(unsigned long *)buff), 
96                                     "r" (zero),  "0" (result));
97                                 --count; 
98                                         buff += 8;
99                         }
100                         result = add32_with_carry(result>>32,
101                                                   result&0xffffffff); 
102
103                         if (len & 4) {
104                                 result += *(unsigned int *) buff;
105                                 buff += 4;
106                         }
107                 }
108                 if (len & 2) {
109                         result += *(unsigned short *) buff;
110                         buff += 2;
111                 }
112         }
113         if (len & 1)
114                 result += *buff;
115         result = add32_with_carry(result>>32, result & 0xffffffff); 
116         if (unlikely(odd)) { 
117                 result = from32to16(result);
118                 result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
119         }
120         return result;
121 }
122
123 /*
124  * computes the checksum of a memory block at buff, length len,
125  * and adds in "sum" (32-bit)
126  *
127  * returns a 32-bit number suitable for feeding into itself
128  * or csum_tcpudp_magic
129  *
130  * this function must be called with even lengths, except
131  * for the last fragment, which may be odd
132  *
133  * it's best to have buff aligned on a 64-bit boundary
134  */
135 unsigned csum_partial(const unsigned char *buff, unsigned len, unsigned sum)
136 {
137         return add32_with_carry(do_csum(buff, len), sum); 
138 }
139
140 EXPORT_SYMBOL(csum_partial);
141
142 /*
143  * this routine is used for miscellaneous IP-like checksums, mainly
144  * in icmp.c
145  */
146 unsigned short ip_compute_csum(unsigned char * buff, int len)
147 {
148         return csum_fold(csum_partial(buff,len,0));
149 }
150