ARM optimized insert_string

Using a faster hash function yields a considerable performance boost
in compression (average 8% on A53 and 24% on A72).

This change was enabled by previous patch with optimized crc32 using
ARMv8-1 crypto extensions for performing CPU feature detection
(so won't help older ARMv7 SoCs).

Bug: 873759
Change-Id: I88ece549a63d923beef4f96a046acdf09e529784
Reviewed-on: https://chromium-review.googlesource.com/1173262Reviewed-by: Chris Blume <cblume@chromium.org>
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Adenilson Cavalcanti <cavalcantii@chromium.org>
Cr-Commit-Position: refs/heads/master@{#583113}

third_party/zlib/crc32_simd.c

@@ -202,4 +202,24 @@ uint32_t ZLIB_INTERNAL armv8_crc32_little(unsigned long crc,
     return ~c;
 }
 
+Pos ZLIB_INTERNAL insert_string_arm(deflate_state *const s, const Pos str)
+{
+    Pos ret;
+    unsigned *ip, val, h = 0;
+
+    ip = (unsigned *)&s->window[str];
+    val = *ip;
+
+    if (s->level >= 6)
+        val &= 0xFFFFFF;
+
+    h = __crc32w(h, val);
+
+    ret = s->head[h & s->hash_mask];
+    s->head[h & s->hash_mask] = str;
+    s->prev[str & s->w_mask] = ret;
+    return ret;
+}
+
+
 #endif

third_party/zlib/crc32_simd.h

@@ -9,6 +9,7 @@
 
 #include "zconf.h"
 #include "zutil.h"
+#include "deflate.h"
 
 /*
  * crc32_sse42_simd_(): compute the crc32 of the buffer, where the buffer
@@ -33,3 +34,8 @@ uint32_t ZLIB_INTERNAL armv8_crc32_little(unsigned long crc,
                                           const unsigned char* buf,
                                           z_size_t len);
 
+/*
+ * Insert hash string.
+ */
+Pos ZLIB_INTERNAL insert_string_arm(deflate_state *const s, const Pos str);
+

third_party/zlib/deflate.c

@@ -51,9 +51,17 @@
 #include <assert.h>
 #include "deflate.h"
 #include "x86.h"
+
 #if (defined(__ARM_NEON__) || defined(__ARM_NEON))
 #include "contrib/optimizations/slide_hash_neon.h"
 #endif
+/* We need crypto extension crc32 to implement optimized hash in
+ * insert_string.
+ */
+#if defined(CRC32_ARMV8_CRC32)
+#include "arm_features.h"
+#include "crc32_simd.h"
+#endif
 
 const char deflate_copyright[] =
    " deflate 1.2.11 Copyright 1995-2017 Jean-loup Gailly and Mark Adler ";
@@ -207,12 +215,16 @@ local INLINE Pos insert_string_c(deflate_state *const s, const Pos str)
 
 local INLINE Pos insert_string(deflate_state *const s, const Pos str)
 {
+#if defined(CRC32_ARMV8_CRC32)
+    if (arm_cpu_enable_crc32)
+        return insert_string_arm(s, str);
+#endif
     if (x86_cpu_enable_simd)
         return insert_string_sse(s, str);
+
     return insert_string_c(s, str);
 }
 
-
 /* ===========================================================================
  * Initialize the hash table (avoiding 64K overflow for 16 bit systems).
  * prev[] will be initialized on the fly.