go.image/tiff: fix (TIFF-specific) LZW decompression.

The blue-purple-pink.lzwcompressed.tiff image was created by
"bmp2tiff -c lzw".

LGTM=djd
R=djd
CC=bsiegert, golang-codereviews
https://golang.org/cl/105750045

tiff/lzw/reader.go

@@ -6,13 +6,32 @@
 // described in T. A. Welch, ``A Technique for High-Performance Data
 // Compression'', Computer, 17(6) (June 1984), pp 8-19.
 //
-// In particular, it implements LZW as used by the GIF, TIFF and PDF file
+// In particular, it implements LZW as used by the TIFF file format, including
-// formats, which means variable-width codes up to 12 bits and the first
+// an "off by one" algorithmic difference when compared to standard LZW.
-// two non-literal codes are a clear code and an EOF code.
 package lzw
 
-// TODO(nigeltao): check that TIFF and PDF use LZW in the same way as GIF,
+/*
-// modulo LSB/MSB packing order.
+This file was branched from src/pkg/compress/lzw/reader.go in the
+standard library. Differences from the original are marked with "NOTE".
+
+The tif_lzw.c file in the libtiff C library has this comment:
+
+----
+The 5.0 spec describes a different algorithm than Aldus
+implements.  Specifically, Aldus does code length transitions
+one code earlier than should be done (for real LZW).
+Earlier versions of this library implemented the correct
+LZW algorithm, but emitted codes in a bit order opposite
+to the TIFF spec.  Thus, to maintain compatibility w/ Aldus
+we interpret MSB-LSB ordered codes to be images written w/
+old versions of this library, but otherwise adhere to the
+Aldus "off by one" algorithm.
+----
+
+The Go code doesn't read (invalid) TIFF files written by old versions of
+libtiff, but the LZW algorithm in this package still differs from the one in
+Go's standard package library to accomodate this "off by one" in valid TIFFs.
+*/
 
 import (
 	"bufio"
@@ -53,7 +72,7 @@ type decoder struct {
 	// The next two codes mean clear and EOF.
 	// Other valid codes are in the range [lo, hi] where lo := clear + 2,
 	// with the upper bound incrementing on each code seen.
-	// overflow is the code at which hi overflows the code width.
+	// overflow is the code at which hi overflows the code width. NOTE: TIFF's LZW is "off by one".
 	// last is the most recently seen code, or decoderInvalidCode.
 	clear, eof, hi, overflow, last uint16
 
@@ -189,7 +208,7 @@ func (d *decoder) decode() {
 			return
 		}
 		d.last, d.hi = code, d.hi+1
-		if d.hi >= d.overflow {
+		if d.hi+1 >= d.overflow { // NOTE: the "+1" is where TIFF's LZW differs from the standard algorithm.
 			if d.width == maxWidth {
 				d.last = decoderInvalidCode
 			} else {

tiff/reader.go

@@ -8,7 +8,6 @@
 package tiff
 
 import (
-	"compress/lzw"
 	"compress/zlib"
 	"encoding/binary"
 	"fmt"
@@ -16,6 +15,8 @@ import (
 	"image/color"
 	"io"
 	"io/ioutil"
+
+	"code.google.com/p/go.image/tiff/lzw"
 )
 
 // A FormatError reports that the input is not a valid TIFF image.

tiff/reader_test.go

@@ -125,6 +125,21 @@ func TestDecode(t *testing.T) {
 	compare(t, img0, img4)
 }
 
+// TestDecodeLZW tests that decoding a PNG image and a LZW-compressed TIFF image
+// result in the same pixel data.
+func TestDecodeLZW(t *testing.T) {
+	img0, err := load("blue-purple-pink.png")
+	if err != nil {
+		t.Fatal(err)
+	}
+	img1, err := load("blue-purple-pink.lzwcompressed.tiff")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	compare(t, img0, img1)
+}
+
 // TestDecompress tests that decoding some TIFF images that use different
 // compression formats result in the same pixel data.
 func TestDecompress(t *testing.T) {