On Sat, Jan 1, 2011 at 11:18 AM, <stef...@apache.org> wrote:
> Author: stefan2
> Date: Sat Jan 1 16:18:55 2011
> New Revision: 1054250
>
> URL: http://svn.apache.org/viewvc?rev=1054250&view=rev
> Log:
> About 3x speed up of svn_eol__find_eol_start() on 64 bit machines
> (20 .. 40% on 32 bits). This function is a significant contributor
> to our diff algorithm.
>
> * subversion/libsvn_subr/eol.c
> (svn_eol__find_eol_start): use some masking magic to process
> the data in larger chunks
>
> Modified:
> subversion/trunk/subversion/libsvn_subr/eol.c
>
> Modified: subversion/trunk/subversion/libsvn_subr/eol.c
> URL:
> http://svn.apache.org/viewvc/subversion/trunk/subversion/libsvn_subr/eol.c?rev=1054250&r1=1054249&r2=1054250&view=diff
> ==============================================================================
> --- subversion/trunk/subversion/libsvn_subr/eol.c (original)
> +++ subversion/trunk/subversion/libsvn_subr/eol.c Sat Jan 1 16:18:55 2011
> @@ -29,14 +29,68 @@
> #include "svn_io.h"
> #include "private/svn_eol_private.h"
>
> +/* Machine-word-sized masks used in svn_eol__find_eol_start.
> + */
> +#if APR_SIZEOF_VOIDP == 8
> +# define LOWER_7BITS_SET 0x7f7f7f7f7f7f7f7f
> +# define BIT_7_SET 0x8080808080808080
> +# define R_MASK 0x0a0a0a0a0a0a0a0a
> +# define N_MASK 0x0d0d0d0d0d0d0d0d
> +#else
> +# define LOWER_7BITS_SET 0x7f7f7f7f
> +# define BIT_7_SET 0x80808080
> +# define R_MASK 0x0a0a0a0a
> +# define N_MASK 0x0d0d0d0d
> +#endif
> +
> char *
> svn_eol__find_eol_start(char *buf, apr_size_t len)
> {
> +#if !SVN_UNALIGNED_ACCESS_IS_OK
> +
> + /* On some systems, we need to make sure that buf is properly aligned
> + * for chunky data access. This overhead is still justified because
> + * only lines tend to be tens of chars long.
> + */
> + for (; (len > 0) && ((apr_uintptr_t)buf) & (sizeof(apr_uintptr_t)-1)
> + ; ++buf, --len)
> + {
> + if (*buf == '\n' || *buf == '\r')
> + return buf;
> + }
> +
> +#endif
> +
> + /* Scan the input one machine word at a time. */
> + for (; len > sizeof(apr_uintptr_t)
> + ; buf += sizeof(apr_uintptr_t), len -= sizeof(apr_uintptr_t))
> + {
> + /* This is a variant of the well-known strlen test: */
> + apr_uintptr_t chunk = *(const apr_uintptr_t *)buf;
This type isn't in APR 0.9.x, which is causing build failures on at
least two of the build slaves.
> +
> + /* A byte in R_TEST is \0, iff it was \r in *BUF.
> + * Similarly, N_TEST is an indicator for \n. */
> + apr_uintptr_t r_test = chunk ^ R_MASK;
> + apr_uintptr_t n_test = chunk ^ N_MASK;
> +
> + /* A byte in R_TEST can by < 0x80, iff it has been \0 before
> + * (i.e. \r in *BUF). Dito for N_TEST. */
> + r_test |= (r_test & LOWER_7BITS_SET) + LOWER_7BITS_SET;
> + n_test |= (n_test & LOWER_7BITS_SET) + LOWER_7BITS_SET;
> +
> + /* Check whether at least one of the words contains a byte <0x80
> + * (if one is detected, there was a \r or \n in CHUNK). */
> + if ((r_test & n_test & BIT_7_SET) != BIT_7_SET)
> + break;
> + }
> +
> + /* The remaining odd bytes will be examined the naive way: */
> for (; len > 0; ++buf, --len)
> {
> if (*buf == '\n' || *buf == '\r')
> return buf;
> }
> +
> return NULL;
> }
>
>
>
>
