(unknown charset) bitcount32: replace lengthy comment with SWAR reference

From: (unknown charset) Andriy Gapon <avg_at_FreeBSD.org>
Date: Sat, 07 May 2011 12:58:11 +0300
I think that the SWAR reference should be more concise and should lead an
interested reader to more information on the topic (and more up-to-date as well).

    bitcount32: replace lengthy comment with SWAR reference

diff --git a/sys/sys/systm.h b/sys/sys/systm.h
index 48bb33c..5218988 100644
--- a/sys/sys/systm.h
+++ b/sys/sys/systm.h
_at__at_ -383,44 +383,8 _at__at_ int alloc_unrl(struct unrhdr *uh);
 void free_unr(struct unrhdr *uh, u_int item);

 /*
- * This is about as magic as it gets.  fortune(1) has got similar code
- * for reversing bits in a word.  Who thinks up this stuff??
- *
- * Yes, it does appear to be consistently faster than:
- * while (i = ffs(m)) {
- *	m >>= i;
- *	bits++;
- * }
- * and
- * while (lsb = (m & -m)) {	// This is magic too
- * 	m &= ~lsb;		// or: m ^= lsb
- *	bits++;
- * }
- * Both of these latter forms do some very strange things on gcc-3.1 with
- * -mcpu=pentiumpro and/or -march=pentiumpro and/or -O or -O2.
- * There is probably an SSE or MMX popcnt instruction.
- *
- * I wonder if this should be in libkern?
- *
- * XXX Stop the presses!  Another one:
- * static __inline u_int32_t
- * popcnt1(u_int32_t v)
- * {
- *	v -= ((v >> 1) & 0x55555555);
- *	v = (v & 0x33333333) + ((v >> 2) & 0x33333333);
- *	v = (v + (v >> 4)) & 0x0F0F0F0F;
- *	return (v * 0x01010101) >> 24;
- * }
- * The downside is that it has a multiply.  With a pentium3 with
- * -mcpu=pentiumpro and -march=pentiumpro then gcc-3.1 will use
- * an imull, and in that case it is faster.  In most other cases
- * it appears slightly slower.
- *
- * Another variant (also from fortune):
- * #define BITCOUNT(x) (((BX_(x)+(BX_(x)>>4)) & 0x0F0F0F0F) % 255)
- * #define  BX_(x)     ((x) - (((x)>>1)&0x77777777)            \
- *                          - (((x)>>2)&0x33333333)            \
- *                          - (((x)>>3)&0x11111111))
+ * Population count algorithm using SWAR approach
+ * - "SIMD Within A Register".
  */
 static __inline uint32_t
 bitcount32(uint32_t x)

-- 
Andriy Gapon
Received on Sat May 07 2011 - 07:58:14 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:13 UTC