Manual browser: __builtin_prefetch(3)

Section:
Page:
__BUILTIN_PREFETCH(3) Library Functions Manual __BUILTIN_PREFETCH(3)

NAME

__builtin_prefetchGNU extension to prefetch memory

SYNOPSIS

void
__builtin_prefetch(const void *addr, ...);

DESCRIPTION

The __builtin_prefetch() function prefetches memory from addr. The rationale is to minimize cache-miss latency by trying to move data into a cache before accessing the data. Possible use cases include frequently called sections of code in which it is known that the data in a given address is likely to be accessed soon.

In addition to addr, there are two optional stdarg(3) arguments, rw and locality. The value of the latter should be a compile-time constant integer between 0 and 3. The higher the value, the higher the temporal locality in the data. When locality is 0, it is assumed that there is little or no temporal locality in the data; after access, it is not necessary to leave the data in the cache. The default value is 3. The value of rw is either 0 or 1, corresponding with read and write prefetch, respectively. The default value of rw is 0. Also rw must be a compile-time constant integer.

The __builtin_prefetch() function translates into prefetch instructions only if the architecture has support for these. If there is no support, addr is evaluated only if it includes side effects, although no warnings are issued by gcc(1).

EXAMPLES

The following optimization appears in the heavily used cpu_in_cksum() function that calculates checksums for the inet(4) headers:

while (mlen >= 32) { 
	__builtin_prefetch(data + 32); 
	partial += *(uint16_t *)data; 
	partial += *(uint16_t *)(data + 2); 
	partial += *(uint16_t *)(data + 4); 
 
	... 
 
	partial += *(uint16_t *)(data + 28); 
	partial += *(uint16_t *)(data + 30); 
 
	data += 32; 
	mlen -= 32; 
 
	...

SEE ALSO

gcc(1), attribute(3)

Ulrich Drepper, What Every Programmer Should Know About Memory, http://www.akkadia.org/drepper/cpumemory.pdf, November 21, 2007.

CAVEATS

This is a non-standard, compiler-specific extension.
December 22, 2010 NetBSD 7.0