Commit 02cdce53 authored Mar 27, 2010 by YOSHIFUJI Hideaki / 吉藤英明 Committed by David S. Miller Mar 30, 2010

ipv6 fib: Use "Sweezle" to optimize addr_bit_test().



addr_bit_test() is used in various places in IPv6 routing table
subsystem.  It checks if the given fn_bit is set,
where fn_bit counts bits from MSB in words in network-order.

 fn_bit        :   0 .... 31 32 .... 64 65 .... 95 96 ....127

fn_bit >> 5 gives offset of word, and (~fn_bit & 0x1f) gives
count from LSB in the network-endian word in question.

 fn_bit >> 5   :       0          1          2          3
 ~fn_bit & 0x1f:  31 ....  0 31 ....  0 31 ....  0 31 ....  0

Thus, the mask was generated as htonl(1 << (~fn_bit & 0x1f)).
This can be optimized by "sweezle" (See include/asm-generic/bitops/le.h).

In little-endian,
  htonl(1 << bit) = 1 << (bit ^ BITOP_BE32_SWIZZLE)
where
  BITOP_BE32_SWIZZLE is (0x1f & ~7)
So,
  htonl(1 << (~fn_bit & 0x1f)) = 1 << ((~fn_bit & 0x1f) ^ (0x1f & ~7))
                               = 1 << ((~fn_bit ^ ~7) & 0x1f)
                               = 1 << ((~fn_bit ^ BITOP_BE32_SWIZZLE) & 0x1f)

In big-endian, BITOP_BE32_SWIZZLE is equal to 0.
  1 << ((~fn_bit ^ BITOP_BE32_SWIZZLE) & 0x1f)
                               = 1 << ((~fn_bit) & 0x1f)
                               = htonl(1 << (~fn_bit & 0x1f))

Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

parent de7737e0

net/ipv6/ip6_fib.c

+13 −2

Original line number	Diff line number	Diff line
		@@ -127,12 +127,23 @@ static __inline__ u32 fib6_new_sernum(void)
		/*
		* test bit
		*/
		#if defined(__LITTLE_ENDIAN)
		# define BITOP_BE32_SWIZZLE (0x1F & ~7)
		#else
		# define BITOP_BE32_SWIZZLE 0
		#endif

		static __inline__ __be32 addr_bit_set(void *token, int fn_bit)
		{
		__be32 *addr = token;

		return htonl(1 << ((~fn_bit)&0x1F)) & addr[fn_bit>>5];
		/*
		* Here,
		* 1 << ((~fn_bit ^ BITOP_BE32_SWIZZLE) & 0x1f)
		* is optimized version of
		* htonl(1 << ((~fn_bit)&0x1F))
		* See include/asm-generic/bitops/le.h.
		*/
		return (1 << ((~fn_bit ^ BITOP_BE32_SWIZZLE) & 0x1f)) & addr[fn_bit >> 5];
		}

		static __inline__ struct fib6_node * node_alloc(void)