From ebe678443fcff7735e4d926e918256b5e33b7ad1 Mon Sep 17 00:00:00 2001
From: Ryan Oldenburg <guzba8@gmail.com>
Date: Sun, 8 Nov 2020 20:21:04 -0600
Subject: [PATCH] faster compress, ~5%

---
 src/zippy.nim         |  4 ++--
 src/zippy/common.nim  |  3 +++
 src/zippy/deflate.nim | 13 +++++++++----
 tests/benchmark.nim   |  3 ++-
 4 files changed, 16 insertions(+), 7 deletions(-)

diff --git a/src/zippy.nim b/src/zippy.nim
index f6865a2..fbc507e 100644
--- a/src/zippy.nim
+++ b/src/zippy.nim
@@ -114,11 +114,11 @@ func uncompress(
 
     inflate(src[pos ..< ^8], dst)
 
-    let checksum = cast[ptr uint32](src[src.len - 8].unsafeAddr)[]
+    let checksum = read32(src[src.len - 8].unsafeAddr)
     if checksum != crc32(dst):
       raise newException(ZippyError, "Checksum verification failed")
 
-    let isize = cast[ptr uint32](src[src.len - 4].unsafeAddr)[]
+    let isize = read32(src[src.len - 4].unsafeAddr)
     if isize != dst.len.uint32:
       raise newException(ZippyError, "Size verification failed")
   of dfZlib:
diff --git a/src/zippy/common.nim b/src/zippy/common.nim
index 0d74ad5..f5473a4 100644
--- a/src/zippy/common.nim
+++ b/src/zippy/common.nim
@@ -101,6 +101,9 @@ const
 when defined(release):
   {.push checks: off.}
 
+template read32*(p: pointer): uint32 =
+  cast[ptr uint32](p)[]
+
 func adler32*(data: seq[uint8]): uint32 =
   ## See https://github.com/madler/zlib/blob/master/adler32.c
 
diff --git a/src/zippy/deflate.nim b/src/zippy/deflate.nim
index 902c864..4ca8fa7 100644
--- a/src/zippy/deflate.nim
+++ b/src/zippy/deflate.nim
@@ -284,10 +284,15 @@ func lz77Encode(src: seq[uint8]): (seq[uint16], seq[int], seq[int], int) =
       prevOffset = offset
 
       var matchLen: int
-      for i in 0 ..< stop - pos:
-        if src[pos - offset + i] != src[pos + i]:
-          break
-        inc matchLen
+      if (
+        (read32(src[pos - offset].unsafeAddr) xor read32(src[pos].unsafeAddr)
+      ) shl 8) == 0:
+        # The first 3 bytes match (the hash of them got us here so usually do)
+        inc(matchLen, 3)
+        for i in 3 ..< stop - pos:
+          if src[pos - offset + i] != src[pos + i]:
+            break
+          inc matchLen
 
       if matchLen > longestMatchLen:
         longestMatchLen = matchLen
diff --git a/tests/benchmark.nim b/tests/benchmark.nim
index 8f38a70..89bc5c0 100644
--- a/tests/benchmark.nim
+++ b/tests/benchmark.nim
@@ -1,4 +1,5 @@
-import miniz, nimPNG/nimz, std/monotimes, strformat, zip/zlib, zippy
+import std/monotimes, strformat, zip/zlib, zippy
+# import miniz, nimPNG/nimz,
 
 const
   zs = [