[PATCH 2/2] (g)unzip: Optimize inflate_codes()

[Denys Vlasenko]

On Thursday 11 February 2010 08:42, Joakim Tjernlund wrote:
> >
> > Great. But looks like you forgot to send the patch without this debug.
> > The only patch I found has this:
> >
> > +                                       /* Align out addr */
> > +                                       if (e < 3)
> > +                                               fprintf(stderr, "error len:%d\n", e);
> 
> Hehe, here we go then. Looking at the gzip code I think it is crap though.
> The upstream gzip code is old and unoptimized. One should just scrap
> it and redo it with zlib instead.
> 
>    Jocke
> 
> From cb65f50ec0599ec40677b655149491a2918133bc Mon Sep 17 00:00:00 2001
> From: Joakim Tjernlund <Joakim.Tjernlund at transmode.se>
> Date: Mon, 8 Feb 2010 18:46:38 +0100
> Subject: [PATCH] (g)unzip: Optimize inflate_codes()
> 
> Ported the recent optimization from the Linux kernel.
> This will not perform as god as the kernel version as the
> code structure in busybox is different and I had to adopt
> the optimization to it.
> 
> This has seen very little testing and is a RFC only at this point.
> The inflate speed increase in the kernel was 12-15% on ppc.
> ---
> 
> V2: Optimize size a bit:
> function                                             old     new   delta
> inflate_codes                                        624     735    +111
> 
>  archival/libunarchive/decompress_unzip.c |   42 ++++++++++++++++++++++++++++-
>  1 files changed, 40 insertions(+), 2 deletions(-)
> 
> diff --git a/archival/libunarchive/decompress_unzip.c b/archival/libunarchive/decompress_unzip.c
> index c616202..111d7fc 100644
> --- a/archival/libunarchive/decompress_unzip.c
> +++ b/archival/libunarchive/decompress_unzip.c
> @@ -589,11 +589,49 @@ static NOINLINE int inflate_codes(STATE_PARAM_ONLY)
>  					w += e;
>  					dd += e;
>  				} else {
> +					unsigned short *sout;
> +					unsigned int loops;
> +					union uu {
> +						unsigned short us;
> +						unsigned char b[2];
> +					} mm;
>  					/* do it slow to avoid memcpy() overlap */
>  					/* !NOMEMCPY */
> -					do {
> +					/* minimum length is three */
> +					/* Align out addr */
> +					if (w & 1) {
> +						gunzip_window[w++] = gunzip_window[dd++];
> +						--e;
> +					}
> +					/* Use pre increment as it is faster on some arch's */
> +					sout = (unsigned short *) (gunzip_window + w - 2);
> +					if (delta > 2) {
> +						unsigned short *sfrom;
> +
> +						sfrom = (unsigned short *) (gunzip_window + dd - 2);
> +						loops = e >> 1;
> +						do
> +							move_from_unaligned16(*++sout, ++sfrom);
> +						while (--loops);
> +					} else {
> +						unsigned short pat16;
> +
> +						pat16 = *sout;
> +						if (delta == 1) {
> +							/* copy one char pattern to both bytes */
> +							mm.us = pat16;
> +							mm.b[0] = mm.b[1];
> +							pat16 = mm.us;
> +						}

It segfaults on 277 Mb gz file (a source tree of old openoffice version):

# time ./busybox_old gunzip <OOo_2.0.2_src.tar.gz >/dev/null

real    0m13.616s
user    0m13.493s
sys     0m0.116s
# time ./busybox gunzip <OOo_2.0.2_src.tar.gz >/dev/null
/bin/bash: line 1: 15635 Segmentation fault      ./busybox gunzip < OOo_2.0.2_src.tar.gz > /dev/null

real    0m2.198s
user    0m2.186s
sys     0m0.011s

-- 
vda