From 17a154db6774a4acf347cfc5189eaf2cd675e696 Mon Sep 17 00:00:00 2001
From: Adenilson Cavalcanti <adenilson.cavalcanti@arm.com>
Date: Mon, 9 Apr 2018 15:14:19 -0700
Subject: [PATCH 2/3] Porting optimized longest_match
This patch was contributed to zlib-ng and features an improved longest_match
function using the most distant hash code to reduce number of checks
(see: http://www.gildor.org/en/projects/zlib).
Original patch by Jun He.
---
CMakeLists.txt | 3 +-
contrib/arm/arm_longest_match.h | 142 ++++++++++++++++++++++++++++++++
deflate.c | 11 ++-
3 files changed, 152 insertions(+), 4 deletions(-)
create mode 100644 contrib/arm/arm_longest_match.h
diff --git a/CMakeLists.txt b/CMakeLists.txt
index e9a74e9..3826eba 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -141,7 +141,8 @@ if(CMAKE_COMPILER_IS_GNUCC)
set(ZLIB_ARM_NEON_HDRS
contrib/arm/chunkcopy.h
contrib/arm/inffast_chunk.h
- contrib/arm/neon_slide_hash.h)
+ contrib/arm/neon_slide_hash.h
+ contrib/arm/arm_longest_match.h)
set(ZLIB_ARM_NEON contrib/arm/inflate.c contrib/arm/inffast_chunk.c)
add_definitions(-DARM_NEON)
set(COMPILER ${CMAKE_C_COMPILER})
diff --git a/contrib/arm/arm_longest_match.h b/contrib/arm/arm_longest_match.h
new file mode 100644
index 0000000..9e7083f
--- /dev/null
+++ b/contrib/arm/arm_longest_match.h
@@ -0,0 +1,142 @@
+/* Copyright (C) 1995-2011, 2016 Mark Adler
+ * Copyright (C) 2017 ARM Holdings Inc.
+ * Authors: Adenilson Cavalcanti <adenilson.cavalcanti@arm.com>
+ * Jun He <jun.he@arm.com>
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+#ifndef __ARM_LONGEST__MATCH__
+#define __ARM_LONGEST__MATCH__
+
+#if defined(ARM_NEON)
+#include "deflate.h"
+#include <stdint.h>
+static inline long get_match_len(const unsigned char *a, const unsigned char *b, long max)
+{
+ register int len = 0;
+ register unsigned long xor = 0;
+ register int check_loops = max/sizeof(unsigned long);
+ while(check_loops-- > 0) {
+ xor = (*(unsigned long *)(a+len)) ^ (*(unsigned long *)(b+len));
+ if (xor) break;
+ len += sizeof(unsigned long);
+ }
+ if (0 == xor) {
+ while (len < max) {
+ if (a[len] != b[len]) break;
+ len++;
+ }
+ return len;
+ }
+ xor = __builtin_ctzl(xor)>>3;
+ return len + xor;
+}
+
+/*
+ * This implementation is based on algorithm described at:
+ * http://www.gildor.org/en/projects/zlib
+ * It uses the hash chain indexed by the most distant hash code to
+ * reduce number of checks.
+ * This also eliminates the those unnecessary check loops in legacy
+ * longest_match's do..while loop if the "most distant code" is out
+ * of search buffer
+ *
+ */
+static inline unsigned arm_longest_match(deflate_state *const s, IPos cur_match) {
+ unsigned chain_length = s->max_chain_length;/* max hash chain length */
+ unsigned char *scan = s->window + s->strstart; /* current string */
+ unsigned char *match; /* matched string */
+ unsigned int len; /* length of current match */
+ unsigned int best_len = s->prev_length; /* best match length so far */
+ unsigned int nice_match = s->nice_match; /* stop if match long enough */
+ IPos limit = s->strstart > (IPos)MAX_DIST(s) ?
+ s->strstart - (IPos)MAX_DIST(s) : 0;
+ /* Stop when cur_match becomes <= limit. To simplify the code,
+ * we prevent matches with the string of window index 0.
+ */
+ int offset = 0; /* offset of the head[most_distant_hash] from IN cur_match */
+ Pos *prev = s->prev;
+ unsigned int wmask = s->w_mask;
+ unsigned char *scan_buf_base = s->window;
+
+ /* The code is optimized for HASH_BITS >= 8 and MAX_MATCH-2 multiple of 16.
+ * It is easy to get rid of this optimization if necessary.
+ */
+ Assert(s->hash_bits >= 8 && MAX_MATCH == 258, "Code too clever");
+
+ /* Do not look for matches beyond the end of the input. This is necessary
+ * to make deflate deterministic.
+ */
+ if ((unsigned int)nice_match > s->lookahead) nice_match = s->lookahead;
+
+ Assert((unsigned long)s->strstart <= s->window_size-MIN_LOOKAHEAD, "need lookahead");
+
+ /* find most distant hash code for lazy_match */
+ if (best_len > MIN_MATCH) {
+ /* search for most distant hash code */
+ int i;
+ uint16_t hash = 0;
+ IPos pos;
+
+ UPDATE_HASH(s, hash, scan[1]);
+ UPDATE_HASH(s, hash, scan[2]);
+ for (i = 3; i <= best_len; i++) {
+ UPDATE_HASH(s, hash, scan[i]);
+ /* get head IPos of hash calced by scan[i-2..i] */
+ pos = s->head[hash];
+ /* compare it to current "farthest hash" IPos */
+ if (pos <= cur_match) {
+ /* we have a new "farthest hash" now */
+ offset = i - 2;
+ cur_match = pos;
+ }
+ }
+
+ /* update variables to correspond offset */
+ limit += offset;
+ /*
+ * check if the most distant code's offset is out of search buffer
+ * if it is true, then this means scan[offset..offset+2] are not
+ * presented in the search buffer. So we just return best_len
+ * we've found.
+ */
+ if (cur_match < limit) return best_len;
+
+ scan_buf_base -= offset;
+ /* reduce hash search depth based on best_len */
+ chain_length /= best_len - MIN_MATCH;
+ }
+
+ do {
+ Assert(cur_match < s->strstart, "no future");
+
+ /* Determine matched length at current pos */
+ match = scan_buf_base + cur_match;
+ len = get_match_len(match, scan, MAX_MATCH);
+
+ if (len > best_len) {
+ /* found longer string */
+ s->match_start = cur_match - offset;
+ best_len = len;
+ /* good enough? */
+ if (len >= nice_match) break;
+ }
+ /* move to prev pos in this hash chain */
+ } while ((cur_match = prev[cur_match & wmask]) > limit && --chain_length != 0);
+
+ return (best_len <= s->lookahead)? best_len : s->lookahead;
+}
+
+#endif
+#endif
diff --git a/deflate.c b/deflate.c
index 36f99ac..4c42259 100644
--- a/deflate.c
+++ b/deflate.c
@@ -50,9 +50,6 @@
/* @(#) $Id$ */
#include "deflate.h"
-#if __ARM_NEON
-#include "contrib/arm/neon_slide_hash.h"
-#endif
const char deflate_copyright[] =
" deflate 1.2.11 Copyright 1995-2017 Jean-loup Gailly and Mark Adler ";
@@ -196,6 +193,11 @@ local const config configuration_table[10] = {
s->head[s->hash_size-1] = NIL; \
zmemzero((Bytef *)s->head, (unsigned)(s->hash_size-1)*sizeof(*s->head));
+#if defined(ARM_NEON)
+#include "contrib/arm/arm_longest_match.h"
+#include "contrib/arm/neon_slide_hash.h"
+#endif
+
/* ===========================================================================
* Slide the hash table when sliding the window down (could be avoided with 32
* bit values at the expense of memory usage). We slide even when level == 0 to
@@ -1244,6 +1246,9 @@ local uInt longest_match(s, cur_match)
deflate_state *s;
IPos cur_match; /* current match */
{
+#if defined(ARM_NEON)
+ return arm_longest_match(s, cur_match);
+#endif
unsigned chain_length = s->max_chain_length;/* max hash chain length */
register Bytef *scan = s->window + s->strstart; /* current string */
register Bytef *match; /* matched string */
--
2.19.0