diff --git a/README.arithmetic b/README.arithmetic
new file mode 100644
index 0000000..0dd3b9a
--- /dev/null
+++ b/README.arithmetic
@@ -0,0 +1,215 @@
+JPEG arithmetic encoding and decoding portable software implementation
+======================================================================
+
+Release of 28-Mar-98 by Guido Vollbeding <guido@jpegclub.org>
+=============================================================
+
+Primary URLs:
+
+	http://sylvana.net/jpeg-ari/
+	(directory containing the actual archive files:)
+
+	http://sylvana.net/jpeg-ari/jpeg-ari-28mar98.tar.gz
+
+	http://sylvana.net/jpeg-ari/jpeg-ari.zip
+
+
+DISCLAIMER
+==========
+
+This package is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+
+It is possible that certain products which can be built using this
+software modules might form inventions protected by patent rights in
+some countries (e.g. by patents about arithmetic coding algorithms
+owned by IBM and AT&T in the USA). Provision of this software by the
+author does NOT include any licenses for any patents.
+In those countries where a patent license is required for certain
+applications of this software modules, you will have to obtain such
+a license yourself.
+
+See Annex L in the JPEG spec for further information
+and a list of relevant patents.
+
+
+What is it?
+===========
+
+This is my implementation of the arithmetic encoding and decoding
+back-end for JPEG as specified in the
+
+  ISO/IEC International Standard 10918-1 and CCITT Recommendation
+  ITU-T T.81, "Information Technology - Digital Compression and
+  Coding of Continuous-tone Still Images, Part 1: Requirements
+  and Guidelines".
+
+Arithmetic coding is a state-of-the-art lossless entropy data
+compression method which offers better compression performance
+than the well-established Huffman entropy coding process.
+
+The JPEG standard specifies a particular arithmetic coding scheme
+to be used optionally as alternative to Huffman coding.
+
+
+Who needs it?
+=============
+
+This package might be of interest for people who are looking for
+enhanced state-of-the-art image compression technologies.
+
+It is intended to provide a reasonable tool for experimental,
+comparison and evaluation purposes.
+
+See the Disclaimer above for restricted conditions of usage.
+
+
+How does it work?
+=================
+
+This distribution is organized as add-on to the widespread
+Independent JPEG Group's JPEG software.
+
+Thus, once you managed to install the IJG software distribution
+successfully, there should be no additional problems (portability
+issues etc.) to incorporate this package into the library,
+and usage is straightforward.
+
+Transcode given JPEG files simply with a command like
+
+  jpegtran -arithmetic [-progressive] < orig.jpg > arit.jpg
+
+into an arithmetic coded version LOSSLESSLY! Since there are
+practically no applications in existence which can handle such
+files, you can only transform it back with the same tool
+
+  jpegtran [-optimize] [-progressive] < arit.jpg > orig2.jpg
+
+to verify correct operation.
+
+Thus, you can easily verify the enhanced compression performance
+of the arithmetic coding version compared to the Huffman (with
+fixed or custom tables) version.
+
+The claim to evaluate was that arithmetic coding gives an average
+5-10% compression improvement against Huffman.
+Early tests with this implementation support this claim, and you
+can perform tests with own material.
+
+Here are some actual results:
+
+% ./jpegtran -optimize < testorig.jpg > testopt.jpg
+% ./jpegtran -arithmetic < testorig.jpg > testarit.jpg
+% ./jpegtran < testarit.jpg > testorig2.jpg
+% ./jpegtran -arithmetic -progressive < testorig.jpg > testaritp.jpg
+% ./jpegtran < testaritp.jpg > testorig3.jpg
+% ./jpegtran -optimize < ../butterfly.jpg > ../buttopt.jpg
+% ./jpegtran -progressive < ../butterfly.jpg > ../buttprog.jpg
+% ./jpegtran -arithmetic < ../butterfly.jpg > ../buttarit.jpg
+% ./jpegtran < ../buttarit.jpg > ../butterfly2.jpg
+% ./jpegtran -arithmetic -progressive < ../butterfly.jpg > ../buttaritp.jpg
+% ./jpegtran < ../buttaritp.jpg > ../butterfly3.jpg
+% ls -l test*.jpg
+-rw-r--r--  1 guivol       5153 Apr 13 18:51 testarit.jpg
+-rw-r--r--  1 guivol       5186 Apr 13 18:51 testaritp.jpg
+-rw-r--r--  1 guivol       5756 Apr  2 15:10 testimg.jpg
+-rw-r--r--  1 guivol       5645 Apr  2 15:10 testimgp.jpg
+-rw-r--r--  1 guivol       5463 Apr 13 18:51 testopt.jpg
+-rw-r--r--  1 guivol       5770 Apr  2 15:10 testorig.jpg
+-rw-r--r--  1 guivol       5770 Apr 13 18:51 testorig2.jpg
+-rw-r--r--  1 guivol       5770 Apr 13 18:51 testorig3.jpg
+-rw-r--r--  1 guivol       5655 Apr  2 15:10 testprog.jpg
+% ls -l ../butt*.jpg
+-rw-r--r--  1 guivol     460091 Apr 13 18:52 ../buttarit.jpg
+-rw-r--r--  1 guivol     453703 Apr 13 18:52 ../buttaritp.jpg
+-rw-r--r--  1 guivol     527823 Nov 19 18:41 ../butterfly.jpg
+-rw-r--r--  1 guivol     527823 Apr 13 18:52 ../butterfly2.jpg
+-rw-r--r--  1 guivol     527823 Apr 13 18:52 ../butterfly3.jpg
+-rw-r--r--  1 guivol     511834 Apr 13 18:52 ../buttopt.jpg
+-rw-r--r--  1 guivol     492237 Apr 13 18:52 ../buttprog.jpg
+% 
+
+Note that arithmetic coding requires only a single processing
+pass due to its fully-adaptive nature, and compared to one-pass
+(fixed tables) Huffman the arithmetic coded version consistently
+achieves 10% compression improvement.
+Compared with two-pass (custom tables) Huffman the improvement
+is 5-10%.
+
+Note that I wasn't able yet to cross-check interoperability of
+the produced files with other implementations.
+Thus, I can't be sure that the files are compliant to the spec,
+but I hope so and the tests support it.
+The encoding and decoding processes should be correct anyway,
+however, in the sense that they are complementary to each other
+and thus retain data integrity.
+
+I would appreciate any indications for compliance or interoperability
+with other implementations from somebody.
+Please let me know if you are able to cross-check something.
+
+
+Installation
+============
+
+The installation is a 2-stage procedure:
+
+1. Preparing the IJG package for potential incorporation
+   of the arithmetic coding feature.
+
+2. Incorporation of the actual arithmetic coding modules
+   and enabling the feature for usage.
+
+The reason for this 2-stage process is the hope to make
+step 1 obsolete in future IJG releases.
+The actual implementation should remain separate IMHO due
+to the different usage conditions.
+
+Step 1:
+
+1.1. Copy all files from the subdirectory 'patchv6b' into
+     the IJG software's v6b source directory.
+     This includes minor patches to some files and 3 extra
+     files which hold place for the actual implementation.
+
+1.2. Update your Makefile/Projectfile for the inclusion of
+     the 3 extra files. This will be done automatically
+     if you use a configure-generated makefile and type
+     './configure' (reconfigure).
+
+1.3. Recompile ('make').
+
+See the file 'PATCHES' in 'patchv6b' for details.
+
+Step 2:
+
+2.1. Replace the 3 placeholder files by the actual implementation
+     modules.
+
+2.2. Enable application support of the new features by #defining
+     C_ARITH_CODING_SUPPORTED and D_ARITH_CODING_SUPPORTED
+     in 'jmorecfg.h'.
+
+2.3. Recompile ('make').
+
+Note that I suggest to add 3 placeholder files to the IJG
+distribution. This would remove the need for system-dependent
+changes (Makefiles) and thus considerably simplify the actual
+installation for systems without a configure-generated makefile.
+
+
+References
+==========
+
+- The Independent JPEG Group's software
+
+- JBIG-KIT lossless image compression library by Markus Kuhn
+
+- William B. Pennebaker, Joan L. Mitchell:
+  "JPEG Still Image Data Compression Standard",
+  Van Nostrand Reinhold, 1993, ISBN 0-442-01272-1.
+
+- jpeg-faq (http://www.faqs.org/faqs/jpeg-faq/)
+
+- compression-faq (http://www.faqs.org/faqs/compression-faq/)
diff --git a/jaricom.c b/jaricom.c
new file mode 100644
index 0000000..e4292b0
--- /dev/null
+++ b/jaricom.c
@@ -0,0 +1,149 @@
+/*
+ * jaricom.c
+ *
+ * Copyright (C) 1997, Guido Vollbeding <guivol@esc.de>.
+ * This file is NOT part of the Independent JPEG Group's software
+ * for legal reasons.
+ * See the accompanying README file for conditions of distribution and use.
+ *
+ * This file contains probability estimation tables for common use in
+ * arithmetic entropy encoding and decoding routines.
+ *
+ * This data represents Table D.2 in the JPEG spec (ISO/IEC IS 10918-1
+ * and CCITT Recommendation ITU-T T.81) and Table 24 in the JBIG spec
+ * (ISO/IEC IS 11544 and CCITT Recommendation ITU-T T.82).
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+/* The following #define specifies the packing of the four components
+ * into the compact INT32 representation.
+ * Note that this formula must match the actual arithmetic encoder
+ * and decoder implementation. The implementation has to be changed
+ * if this formula is changed.
+ * The current organisation is leaned on Markus Kuhn's JBIG
+ * implementation (jbig_tab.c).
+ */
+
+#define V(a,b,c,d) (((INT32)a << 16) | ((INT32)c << 8) | ((INT32)d << 7) | b)
+
+const INT32 jaritab[113] = {
+/*
+ * Index, Qe_Value, Next_Index_LPS, Next_Index_MPS, Switch_MPS
+ */
+/*   0 */  V( 0x5a1d,   1,   1, 1 ),
+/*   1 */  V( 0x2586,  14,   2, 0 ),
+/*   2 */  V( 0x1114,  16,   3, 0 ),
+/*   3 */  V( 0x080b,  18,   4, 0 ),
+/*   4 */  V( 0x03d8,  20,   5, 0 ),
+/*   5 */  V( 0x01da,  23,   6, 0 ),
+/*   6 */  V( 0x00e5,  25,   7, 0 ),
+/*   7 */  V( 0x006f,  28,   8, 0 ),
+/*   8 */  V( 0x0036,  30,   9, 0 ),
+/*   9 */  V( 0x001a,  33,  10, 0 ),
+/*  10 */  V( 0x000d,  35,  11, 0 ),
+/*  11 */  V( 0x0006,   9,  12, 0 ),
+/*  12 */  V( 0x0003,  10,  13, 0 ),
+/*  13 */  V( 0x0001,  12,  13, 0 ),
+/*  14 */  V( 0x5a7f,  15,  15, 1 ),
+/*  15 */  V( 0x3f25,  36,  16, 0 ),
+/*  16 */  V( 0x2cf2,  38,  17, 0 ),
+/*  17 */  V( 0x207c,  39,  18, 0 ),
+/*  18 */  V( 0x17b9,  40,  19, 0 ),
+/*  19 */  V( 0x1182,  42,  20, 0 ),
+/*  20 */  V( 0x0cef,  43,  21, 0 ),
+/*  21 */  V( 0x09a1,  45,  22, 0 ),
+/*  22 */  V( 0x072f,  46,  23, 0 ),
+/*  23 */  V( 0x055c,  48,  24, 0 ),
+/*  24 */  V( 0x0406,  49,  25, 0 ),
+/*  25 */  V( 0x0303,  51,  26, 0 ),
+/*  26 */  V( 0x0240,  52,  27, 0 ),
+/*  27 */  V( 0x01b1,  54,  28, 0 ),
+/*  28 */  V( 0x0144,  56,  29, 0 ),
+/*  29 */  V( 0x00f5,  57,  30, 0 ),
+/*  30 */  V( 0x00b7,  59,  31, 0 ),
+/*  31 */  V( 0x008a,  60,  32, 0 ),
+/*  32 */  V( 0x0068,  62,  33, 0 ),
+/*  33 */  V( 0x004e,  63,  34, 0 ),
+/*  34 */  V( 0x003b,  32,  35, 0 ),
+/*  35 */  V( 0x002c,  33,   9, 0 ),
+/*  36 */  V( 0x5ae1,  37,  37, 1 ),
+/*  37 */  V( 0x484c,  64,  38, 0 ),
+/*  38 */  V( 0x3a0d,  65,  39, 0 ),
+/*  39 */  V( 0x2ef1,  67,  40, 0 ),
+/*  40 */  V( 0x261f,  68,  41, 0 ),
+/*  41 */  V( 0x1f33,  69,  42, 0 ),
+/*  42 */  V( 0x19a8,  70,  43, 0 ),
+/*  43 */  V( 0x1518,  72,  44, 0 ),
+/*  44 */  V( 0x1177,  73,  45, 0 ),
+/*  45 */  V( 0x0e74,  74,  46, 0 ),
+/*  46 */  V( 0x0bfb,  75,  47, 0 ),
+/*  47 */  V( 0x09f8,  77,  48, 0 ),
+/*  48 */  V( 0x0861,  78,  49, 0 ),
+/*  49 */  V( 0x0706,  79,  50, 0 ),
+/*  50 */  V( 0x05cd,  48,  51, 0 ),
+/*  51 */  V( 0x04de,  50,  52, 0 ),
+/*  52 */  V( 0x040f,  50,  53, 0 ),
+/*  53 */  V( 0x0363,  51,  54, 0 ),
+/*  54 */  V( 0x02d4,  52,  55, 0 ),
+/*  55 */  V( 0x025c,  53,  56, 0 ),
+/*  56 */  V( 0x01f8,  54,  57, 0 ),
+/*  57 */  V( 0x01a4,  55,  58, 0 ),
+/*  58 */  V( 0x0160,  56,  59, 0 ),
+/*  59 */  V( 0x0125,  57,  60, 0 ),
+/*  60 */  V( 0x00f6,  58,  61, 0 ),
+/*  61 */  V( 0x00cb,  59,  62, 0 ),
+/*  62 */  V( 0x00ab,  61,  63, 0 ),
+/*  63 */  V( 0x008f,  61,  32, 0 ),
+/*  64 */  V( 0x5b12,  65,  65, 1 ),
+/*  65 */  V( 0x4d04,  80,  66, 0 ),
+/*  66 */  V( 0x412c,  81,  67, 0 ),
+/*  67 */  V( 0x37d8,  82,  68, 0 ),
+/*  68 */  V( 0x2fe8,  83,  69, 0 ),
+/*  69 */  V( 0x293c,  84,  70, 0 ),
+/*  70 */  V( 0x2379,  86,  71, 0 ),
+/*  71 */  V( 0x1edf,  87,  72, 0 ),
+/*  72 */  V( 0x1aa9,  87,  73, 0 ),
+/*  73 */  V( 0x174e,  72,  74, 0 ),
+/*  74 */  V( 0x1424,  72,  75, 0 ),
+/*  75 */  V( 0x119c,  74,  76, 0 ),
+/*  76 */  V( 0x0f6b,  74,  77, 0 ),
+/*  77 */  V( 0x0d51,  75,  78, 0 ),
+/*  78 */  V( 0x0bb6,  77,  79, 0 ),
+/*  79 */  V( 0x0a40,  77,  48, 0 ),
+/*  80 */  V( 0x5832,  80,  81, 1 ),
+/*  81 */  V( 0x4d1c,  88,  82, 0 ),
+/*  82 */  V( 0x438e,  89,  83, 0 ),
+/*  83 */  V( 0x3bdd,  90,  84, 0 ),
+/*  84 */  V( 0x34ee,  91,  85, 0 ),
+/*  85 */  V( 0x2eae,  92,  86, 0 ),
+/*  86 */  V( 0x299a,  93,  87, 0 ),
+/*  87 */  V( 0x2516,  86,  71, 0 ),
+/*  88 */  V( 0x5570,  88,  89, 1 ),
+/*  89 */  V( 0x4ca9,  95,  90, 0 ),
+/*  90 */  V( 0x44d9,  96,  91, 0 ),
+/*  91 */  V( 0x3e22,  97,  92, 0 ),
+/*  92 */  V( 0x3824,  99,  93, 0 ),
+/*  93 */  V( 0x32b4,  99,  94, 0 ),
+/*  94 */  V( 0x2e17,  93,  86, 0 ),
+/*  95 */  V( 0x56a8,  95,  96, 1 ),
+/*  96 */  V( 0x4f46, 101,  97, 0 ),
+/*  97 */  V( 0x47e5, 102,  98, 0 ),
+/*  98 */  V( 0x41cf, 103,  99, 0 ),
+/*  99 */  V( 0x3c3d, 104, 100, 0 ),
+/* 100 */  V( 0x375e,  99,  93, 0 ),
+/* 101 */  V( 0x5231, 105, 102, 0 ),
+/* 102 */  V( 0x4c0f, 106, 103, 0 ),
+/* 103 */  V( 0x4639, 107, 104, 0 ),
+/* 104 */  V( 0x415e, 103,  99, 0 ),
+/* 105 */  V( 0x5627, 105, 106, 1 ),
+/* 106 */  V( 0x50e7, 108, 107, 0 ),
+/* 107 */  V( 0x4b85, 109, 103, 0 ),
+/* 108 */  V( 0x5597, 110, 109, 0 ),
+/* 109 */  V( 0x504f, 111, 107, 0 ),
+/* 110 */  V( 0x5a10, 110, 111, 1 ),
+/* 111 */  V( 0x5522, 112, 109, 0 ),
+/* 112 */  V( 0x59eb, 112, 111, 1 )
+};
diff --git a/jcarith.c b/jcarith.c
new file mode 100644
index 0000000..ae1b23c
--- /dev/null
+++ b/jcarith.c
@@ -0,0 +1,922 @@
+/*
+ * jcarith.c
+ *
+ * Copyright (C) 1997, Guido Vollbeding <guivol@esc.de>.
+ * This file is NOT part of the Independent JPEG Group's software
+ * for legal reasons.
+ * See the accompanying README file for conditions of distribution and use.
+ *
+ * This file contains portable arithmetic entropy encoding routines for JPEG
+ * (implementing the ISO/IEC IS 10918-1 and CCITT Recommendation ITU-T T.81).
+ *
+ * Both sequential and progressive modes are supported in this single module.
+ *
+ * Suspension is not currently supported in this module.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Expanded entropy encoder object for arithmetic encoding. */
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  INT32 c; /* C register, base of coding interval, layout as in sec. D.1.3 */
+  INT32 a;               /* A register, normalized size of coding interval */
+  INT32 sc;        /* counter for stacked 0xFF values which might overflow */
+  INT32 zc;          /* counter for pending 0x00 output values which might *
+                          * be discarded at the end ("Pacman" termination) */
+  int ct;  /* bit shift counter, determines when next byte will be written */
+  int buffer;                /* buffer for most recent output byte != 0xFF */
+
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+  int dc_context[MAX_COMPS_IN_SCAN]; /* context index for DC conditioning */
+
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+  int next_restart_num;		/* next restart number to write (0-7) */
+
+  /* Pointers to statistics areas (these workspaces have image lifespan) */
+  unsigned char * dc_stats[NUM_ARITH_TBLS];
+  unsigned char * ac_stats[NUM_ARITH_TBLS];
+} arith_entropy_encoder;
+
+typedef arith_entropy_encoder * arith_entropy_ptr;
+
+/* The following two definitions specify the allocation chunk size
+ * for the statistics area.
+ * According to sections F.1.4.4.1.3 and F.1.4.4.2, we need at least
+ * 49 statistics bins for DC, and 245 statistics bins for AC coding.
+ * Note that we use one additional AC bin for codings with fixed
+ * probability (0.5), thus the minimum number for AC is 246.
+ *
+ * We use a compact representation with 1 byte per statistics bin,
+ * thus the numbers directly represent byte sizes.
+ * This 1 byte per statistics bin contains the meaning of the MPS
+ * (more probable symbol) in the highest bit (mask 0x80), and the
+ * index into the probability estimation state machine table
+ * in the lower bits (mask 0x7F).
+ */
+
+#define DC_STAT_BINS 64
+#define AC_STAT_BINS 256
+
+/* NOTE: Uncomment the following #define if you want to use the
+ * given formula for calculating the AC conditioning parameter Kx
+ * for spectral selection progressive coding in section G.1.3.2
+ * of the spec (Kx = Kmin + SRL (8 + Se - Kmin) 4).
+ * Although the spec and P&M authors claim that this "has proven
+ * to give good results for 8 bit precision samples", I'm not
+ * convinced yet that this is really beneficial.
+ * Early tests gave only very marginal compression enhancements
+ * (a few - around 5 or so - bytes even for very large files),
+ * which would turn out rather negative if we'd suppress the
+ * DAC (Define Arithmetic Conditioning) marker segments for
+ * the default parameters in the future.
+ * Note that currently the marker writing module emits 12-byte
+ * DAC segments for a full-component scan in a color image.
+ * This is not worth worrying about IMHO. However, since the
+ * spec defines the default values to be used if the tables
+ * are omitted (unlike Huffman tables, which are required
+ * anyway), one might optimize this behaviour in the future,
+ * and then it would be disadvantageous to use custom tables if
+ * they don't provide sufficient gain to exceed the DAC size.
+ *
+ * On the other hand, I'd consider it as a reasonable result
+ * that the conditioning has no significant influence on the
+ * compression performance. This means that the basic
+ * statistical model is already rather stable.
+ *
+ * Thus, at the moment, we use the default conditioning values
+ * anyway, and do not use the custom formula.
+ *
+#define CALCULATE_SPECTRAL_CONDITIONING
+ */
+
+/* IRIGHT_SHIFT is like RIGHT_SHIFT, but works on int rather than INT32.
+ * We assume that int right shift is unsigned if INT32 right shift is,
+ * which should be safe.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS	int ishift_temp;
+#define IRIGHT_SHIFT(x,shft)  \
+	((ishift_temp = (x)) < 0 ? \
+	 (ishift_temp >> (shft)) | ((~0) << (16-(shft))) : \
+	 (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+
+LOCAL(void)
+emit_byte (int val, j_compress_ptr cinfo)
+/* Write next output byte; we do not support suspension in this module. */
+{
+  struct jpeg_destination_mgr * dest = cinfo->dest;
+
+  *dest->next_output_byte++ = (JOCTET) val;
+  if (--dest->free_in_buffer == 0)
+    if (! (*dest->empty_output_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+}
+
+
+/*
+ * Finish up at the end of an arithmetic-compressed scan.
+ */
+
+METHODDEF(void)
+finish_pass (j_compress_ptr cinfo)
+{
+  arith_entropy_ptr e = (arith_entropy_ptr) cinfo->entropy;
+  INT32 temp;
+
+  /* Section D.1.8: Termination of encoding */
+
+  /* Find the e->c in the coding interval with the largest
+   * number of trailing zero bits */
+  if ((temp = (e->a - 1 + e->c) & 0xFFFF0000L) < e->c)
+    e->c = temp + 0x8000L;
+  else
+    e->c = temp;
+  /* Send remaining bytes to output */
+  e->c <<= e->ct;
+  if (e->c & 0xF8000000L) {
+    /* One final overflow has to be handled */
+    if (e->buffer >= 0) {
+      if (e->zc)
+	do emit_byte(0x00, cinfo);
+	while (--e->zc);
+      emit_byte(e->buffer + 1, cinfo);
+      if (e->buffer + 1 == 0xFF)
+	emit_byte(0x00, cinfo);
+    }
+    e->zc += e->sc;  /* carry-over converts stacked 0xFF bytes to 0x00 */
+    e->sc = 0;
+  } else {
+    if (e->buffer == 0)
+      ++e->zc;
+    else if (e->buffer >= 0) {
+      if (e->zc)
+	do emit_byte(0x00, cinfo);
+	while (--e->zc);
+      emit_byte(e->buffer, cinfo);
+    }
+    if (e->sc) {
+      if (e->zc)
+	do emit_byte(0x00, cinfo);
+	while (--e->zc);
+      do {
+	emit_byte(0xFF, cinfo);
+	emit_byte(0x00, cinfo);
+      } while (--e->sc);
+    }
+  }
+  /* Output final bytes only if they are not 0x00 */
+  if (e->c & 0x7FFF800L) {
+    if (e->zc)  /* output final pending zero bytes */
+      do emit_byte(0x00, cinfo);
+      while (--e->zc);
+    emit_byte((e->c >> 19) & 0xFF, cinfo);
+    if (((e->c >> 19) & 0xFF) == 0xFF)
+      emit_byte(0x00, cinfo);
+    if (e->c & 0x7F800L) {
+      emit_byte((e->c >> 11) & 0xFF, cinfo);
+      if (((e->c >> 11) & 0xFF) == 0xFF)
+	emit_byte(0x00, cinfo);
+    }
+  }
+}
+
+
+/*
+ * The core arithmetic encoding routine (common in JPEG and JBIG).
+ * This needs to go as fast as possible.
+ * Machine-dependent optimization facilities
+ * are not utilized in this portable implementation.
+ * However, this code should be fairly efficient and
+ * may be a good base for further optimizations anyway.
+ *
+ * Parameter 'val' to be encoded may be 0 or 1 (binary decision).
+ *
+ * Note: I've added full "Pacman" termination support to the
+ * byte output routines, which is equivalent to the optional
+ * Discard_final_zeros procedure (Figure D.15) in the spec.
+ * Thus, we always produce the shortest possible output
+ * stream compliant to the spec (no trailing zero bytes,
+ * except for FF stuffing).
+ *
+ * I've also introduced a new scheme for accessing
+ * the probability estimation state machine table,
+ * derived from Markus Kuhn's JBIG implementation.
+ */
+
+LOCAL(void)
+arith_encode (j_compress_ptr cinfo, unsigned char *st, int val) 
+{
+  extern const INT32 jaritab[];
+  register arith_entropy_ptr e = (arith_entropy_ptr) cinfo->entropy;
+  register unsigned char nl, nm;
+  register INT32 qe, temp;
+  register int sv;
+
+  /* Fetch values from our compact representation of Table D.2:
+   * Qe values and probability estimation state machine
+   */
+  sv = *st;
+  qe = jaritab[sv & 0x7F];	/* => Qe_Value */
+  nl = qe & 0xFF; qe >>= 8;	/* Next_Index_LPS + Switch_MPS */
+  nm = qe & 0xFF; qe >>= 8;	/* Next_Index_MPS */
+
+  /* Encode & estimation procedures per sections D.1.4 & D.1.5 */
+  e->a -= qe;
+  if (val != (sv >> 7)) {
+    /* Encode the less probable symbol */
+    if (e->a >= qe) {
+      /* If the interval size (qe) for the less probable symbol (LPS)
+       * is larger than the interval size for the MPS, then exchange
+       * the two symbols for coding efficiency, otherwise code the LPS
+       * as usual: */
+      e->c += e->a;
+      e->a = qe;
+    }
+    *st = (sv & 0x80) ^ nl;	/* Estimate_after_LPS */
+  } else {
+    /* Encode the more probable symbol */
+    if (e->a >= 0x8000L)
+      return;  /* A >= 0x8000 -> ready, no renormalization required */
+    if (e->a < qe) {
+      /* If the interval size (qe) for the less probable symbol (LPS)
+       * is larger than the interval size for the MPS, then exchange
+       * the two symbols for coding efficiency: */
+      e->c += e->a;
+      e->a = qe;
+    }
+    *st = (sv & 0x80) ^ nm;	/* Estimate_after_MPS */
+  }
+
+  /* Renormalization & data output per section D.1.6 */
+  do {
+    e->a <<= 1;
+    e->c <<= 1;
+    if (--e->ct == 0) {
+      /* Another byte is ready for output */
+      temp = e->c >> 19;
+      if (temp > 0xFF) {
+	/* Handle overflow over all stacked 0xFF bytes */
+	if (e->buffer >= 0) {
+	  if (e->zc)
+	    do emit_byte(0x00, cinfo);
+	    while (--e->zc);
+	  emit_byte(e->buffer + 1, cinfo);
+	  if (e->buffer + 1 == 0xFF)
+	    emit_byte(0x00, cinfo);
+	}
+	e->zc += e->sc;  /* carry-over converts stacked 0xFF bytes to 0x00 */
+	e->sc = 0;
+	/* Note: The 3 spacer bits in the C register guarantee
+	 * that the new buffer byte can't be 0xFF here
+	 * (see page 160 in the P&M JPEG book). */
+	e->buffer = temp & 0xFF;  /* new output byte, might overflow later */
+      } else if (temp == 0xFF) {
+	++e->sc;  /* stack 0xFF byte (which might overflow later) */
+      } else {
+	/* Output all stacked 0xFF bytes, they will not overflow any more */
+	if (e->buffer == 0)
+	  ++e->zc;
+	else if (e->buffer >= 0) {
+	  if (e->zc)
+	    do emit_byte(0x00, cinfo);
+	    while (--e->zc);
+	  emit_byte(e->buffer, cinfo);
+	}
+	if (e->sc) {
+	  if (e->zc)
+	    do emit_byte(0x00, cinfo);
+	    while (--e->zc);
+	  do {
+	    emit_byte(0xFF, cinfo);
+	    emit_byte(0x00, cinfo);
+	  } while (--e->sc);
+	}
+	e->buffer = temp & 0xFF;  /* new output byte (can still overflow) */
+      }
+      e->c &= 0x7FFFFL;
+      e->ct += 8;
+    }
+  } while (e->a < 0x8000L);
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(void)
+emit_restart (j_compress_ptr cinfo, int restart_num)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  int ci;
+  jpeg_component_info * compptr;
+
+  finish_pass(cinfo);
+
+  emit_byte(0xFF, cinfo);
+  emit_byte(JPEG_RST0 + restart_num, cinfo);
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Re-initialize statistics areas */
+    if (cinfo->progressive_mode == 0 || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      MEMZERO(entropy->dc_stats[compptr->dc_tbl_no], DC_STAT_BINS);
+      /* Reset DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    if (cinfo->progressive_mode == 0 || cinfo->Ss) {
+      MEMZERO(entropy->ac_stats[compptr->ac_tbl_no], AC_STAT_BINS);
+    }
+  }
+
+  /* Reset arithmetic encoding variables */
+  entropy->c = 0;
+  entropy->a = 0x10000L;
+  entropy->sc = 0;
+  entropy->zc = 0;
+  entropy->ct = 11;
+  entropy->buffer = -1;  /* empty */
+}
+
+
+/*
+ * MCU encoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl;
+  int v, v2, m;
+  ISHIFT_TEMPS
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    tbl = cinfo->cur_comp_info[ci]->dc_tbl_no;
+
+    /* Compute the DC value after the required point transform by Al.
+     * This is simply an arithmetic right shift.
+     */
+    m = IRIGHT_SHIFT((int) ((*block)[0]), cinfo->Al);
+
+    /* Sections F.1.4.1 & F.1.4.4.1: Encoding of DC coefficients */
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.4: Encode_DC_DIFF */
+    if ((v = m - entropy->last_dc_val[ci]) == 0) {
+      arith_encode(cinfo, st, 0);
+      entropy->dc_context[ci] = 0;	/* zero diff category */
+    } else {
+      entropy->last_dc_val[ci] = m;
+      arith_encode(cinfo, st, 1);
+      /* Figure F.6: Encoding nonzero value v */
+      /* Figure F.7: Encoding the sign of v */
+      if (v > 0) {
+	arith_encode(cinfo, st + 1, 0);	/* Table F.4: SS = S0 + 1 */
+	st += 2;			/* Table F.4: SP = S0 + 2 */
+	entropy->dc_context[ci] = 4;	/* small positive diff category */
+      } else {
+	v = -v;
+	arith_encode(cinfo, st + 1, 1);	/* Table F.4: SS = S0 + 1 */
+	st += 3;			/* Table F.4: SN = S0 + 3 */
+	entropy->dc_context[ci] = 8;	/* small negative diff category */
+      }
+      /* Figure F.8: Encoding the magnitude category of v */
+      m = 0;
+      if (v -= 1) {
+	arith_encode(cinfo, st, 1);
+	m = 1;
+	v2 = v;
+	st = entropy->dc_stats[tbl] + 20; /* Table F.4: X1 = 20 */
+	while (v2 >>= 1) {
+	  arith_encode(cinfo, st, 1);
+	  m <<= 1;
+	  st += 1;
+	}
+      }
+      arith_encode(cinfo, st, 0);
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int) (((INT32) 1 << cinfo->arith_dc_L[tbl]) >> 1))
+	entropy->dc_context[ci] = 0;	/* zero diff category */
+      else if (m > (int) (((INT32) 1 << cinfo->arith_dc_U[tbl]) >> 1))
+	entropy->dc_context[ci] += 8;	/* large diff category */
+      /* Figure F.9: Encoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	arith_encode(cinfo, st, (m & v) ? 1 : 0);
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int tbl, k, ke;
+  int v, v2, m;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  /* Sections F.1.4.2 & F.1.4.4.2: Encoding of AC coefficients */
+
+  /* Establish EOB (end-of-block) index */
+  for (ke = cinfo->Se + 1; ke > 1; ke--)
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value.
+     */
+    if ((v = (*block)[jpeg_natural_order[ke - 1]]) >= 0) {
+      if (v >>= cinfo->Al) break;
+    } else {
+      v = -v;
+      if (v >>= cinfo->Al) break;
+    }
+
+  /* Figure F.5: Encode_AC_Coefficients */
+  for (k = cinfo->Ss; k < ke; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    arith_encode(cinfo, st, 0);		/* EOB decision */
+    entropy->ac_stats[tbl][245] = 0;
+    for (;;) {
+      if ((v = (*block)[jpeg_natural_order[k]]) >= 0) {
+	if (v >>= cinfo->Al) {
+	  arith_encode(cinfo, st + 1, 1);
+	  arith_encode(cinfo, entropy->ac_stats[tbl] + 245, 0);
+	  break;
+	}
+      } else {
+	v = -v;
+	if (v >>= cinfo->Al) {
+	  arith_encode(cinfo, st + 1, 1);
+	  arith_encode(cinfo, entropy->ac_stats[tbl] + 245, 1);
+	  break;
+	}
+      }
+      arith_encode(cinfo, st + 1, 0); st += 3; k++;
+    }
+    st += 2;
+    /* Figure F.8: Encoding the magnitude category of v */
+    m = 0;
+    if (v -= 1) {
+      arith_encode(cinfo, st, 1);
+      m = 1;
+      v2 = v;
+      if (v2 >>= 1) {
+	arith_encode(cinfo, st, 1);
+	m <<= 1;
+	st = entropy->ac_stats[tbl] +
+	     (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+	while (v2 >>= 1) {
+	  arith_encode(cinfo, st, 1);
+	  m <<= 1;
+	  st += 1;
+	}
+      }
+    }
+    arith_encode(cinfo, st, 0);
+    /* Figure F.9: Encoding the magnitude bit pattern of v */
+    st += 14;
+    while (m >>= 1)
+      arith_encode(cinfo, st, (m & v) ? 1 : 0);
+  }
+  /* Encode EOB decision only if k <= cinfo->Se */
+  if (k <= cinfo->Se) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    arith_encode(cinfo, st, 1);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for DC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  unsigned char st[4];
+  int Al, blkn;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  Al = cinfo->Al;
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    st[0] = 0;	/* use fixed probability estimation */
+    /* We simply emit the Al'th bit of the DC coefficient value. */
+    arith_encode(cinfo, st, (MCU_data[blkn][0][0] >> Al) & 1);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int tbl, k, ke, kex;
+  int v;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  /* Section G.1.3.3: Encoding of AC coefficients */
+
+  /* Establish EOB (end-of-block) index */
+  for (ke = cinfo->Se + 1; ke > 1; ke--)
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value.
+     */
+    if ((v = (*block)[jpeg_natural_order[ke - 1]]) >= 0) {
+      if (v >>= cinfo->Al) break;
+    } else {
+      v = -v;
+      if (v >>= cinfo->Al) break;
+    }
+
+  /* Establish EOBx (previous stage end-of-block) index */
+  for (kex = ke; kex > 1; kex--)
+    if ((v = (*block)[jpeg_natural_order[kex - 1]]) >= 0) {
+      if (v >>= cinfo->Ah) break;
+    } else {
+      v = -v;
+      if (v >>= cinfo->Ah) break;
+    }
+
+  /* Figure G.10: Encode_AC_Coefficients_SA */
+  for (k = cinfo->Ss; k < ke; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    if (k >= kex)
+      arith_encode(cinfo, st, 0);	/* EOB decision */
+    entropy->ac_stats[tbl][245] = 0;
+    for (;;) {
+      if ((v = (*block)[jpeg_natural_order[k]]) >= 0) {
+	if (v >>= cinfo->Al) {
+	  if (v >> 1)		/* previously nonzero coef */
+	    arith_encode(cinfo, st + 2, (v & 1));
+	  else {		/* newly nonzero coef */
+	    arith_encode(cinfo, st + 1, 1);
+	    arith_encode(cinfo, entropy->ac_stats[tbl] + 245, 0);
+	  }
+	  break;
+	}
+      } else {
+	v = -v;
+	if (v >>= cinfo->Al) {
+	  if (v >> 1)		/* previously nonzero coef */
+	    arith_encode(cinfo, st + 2, (v & 1));
+	  else {		/* newly nonzero coef */
+	    arith_encode(cinfo, st + 1, 1);
+	    arith_encode(cinfo, entropy->ac_stats[tbl] + 245, 1);
+	  }
+	  break;
+	}
+      }
+      arith_encode(cinfo, st + 1, 0); st += 3; k++;
+    }
+  }
+  /* Encode EOB decision only if k <= cinfo->Se */
+  if (k <= cinfo->Se) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    arith_encode(cinfo, st, 1);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Encode and output one MCU's worth of arithmetic-compressed coefficients.
+ */
+
+METHODDEF(boolean)
+encode_mcu (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  jpeg_component_info * compptr;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl, k, ke;
+  int v, v2, m;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Sections F.1.4.1 & F.1.4.4.1: Encoding of DC coefficients */
+
+    tbl = compptr->dc_tbl_no;
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.4: Encode_DC_DIFF */
+    if ((v = (*block)[0] - entropy->last_dc_val[ci]) == 0) {
+      arith_encode(cinfo, st, 0);
+      entropy->dc_context[ci] = 0;	/* zero diff category */
+    } else {
+      entropy->last_dc_val[ci] = (*block)[0];
+      arith_encode(cinfo, st, 1);
+      /* Figure F.6: Encoding nonzero value v */
+      /* Figure F.7: Encoding the sign of v */
+      if (v > 0) {
+	arith_encode(cinfo, st + 1, 0);	/* Table F.4: SS = S0 + 1 */
+	st += 2;			/* Table F.4: SP = S0 + 2 */
+	entropy->dc_context[ci] = 4;	/* small positive diff category */
+      } else {
+	v = -v;
+	arith_encode(cinfo, st + 1, 1);	/* Table F.4: SS = S0 + 1 */
+	st += 3;			/* Table F.4: SN = S0 + 3 */
+	entropy->dc_context[ci] = 8;	/* small negative diff category */
+      }
+      /* Figure F.8: Encoding the magnitude category of v */
+      m = 0;
+      if (v -= 1) {
+	arith_encode(cinfo, st, 1);
+	m = 1;
+	v2 = v;
+	st = entropy->dc_stats[tbl] + 20; /* Table F.4: X1 = 20 */
+	while (v2 >>= 1) {
+	  arith_encode(cinfo, st, 1);
+	  m <<= 1;
+	  st += 1;
+	}
+      }
+      arith_encode(cinfo, st, 0);
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int) (((INT32) 1 << cinfo->arith_dc_L[tbl]) >> 1))
+	entropy->dc_context[ci] = 0;	/* zero diff category */
+      else if (m > (int) (((INT32) 1 << cinfo->arith_dc_U[tbl]) >> 1))
+	entropy->dc_context[ci] += 8;	/* large diff category */
+      /* Figure F.9: Encoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	arith_encode(cinfo, st, (m & v) ? 1 : 0);
+    }
+
+    /* Sections F.1.4.2 & F.1.4.4.2: Encoding of AC coefficients */
+
+    tbl = compptr->ac_tbl_no;
+
+    /* Establish EOB (end-of-block) index */
+    for (ke = DCTSIZE2; ke > 1; ke--)
+      if ((*block)[jpeg_natural_order[ke - 1]]) break;
+
+    /* Figure F.5: Encode_AC_Coefficients */
+    for (k = 1; k < ke; k++) {
+      st = entropy->ac_stats[tbl] + 3 * (k - 1);
+      arith_encode(cinfo, st, 0);	/* EOB decision */
+      while ((v = (*block)[jpeg_natural_order[k]]) == 0) {
+	arith_encode(cinfo, st + 1, 0); st += 3; k++;
+      }
+      arith_encode(cinfo, st + 1, 1);
+      /* Figure F.6: Encoding nonzero value v */
+      /* Figure F.7: Encoding the sign of v */
+      entropy->ac_stats[tbl][245] = 0;
+      if (v > 0) {
+	arith_encode(cinfo, entropy->ac_stats[tbl] + 245, 0);
+      } else {
+	v = -v;
+	arith_encode(cinfo, entropy->ac_stats[tbl] + 245, 1);
+      }
+      st += 2;
+      /* Figure F.8: Encoding the magnitude category of v */
+      m = 0;
+      if (v -= 1) {
+	arith_encode(cinfo, st, 1);
+	m = 1;
+	v2 = v;
+	if (v2 >>= 1) {
+	  arith_encode(cinfo, st, 1);
+	  m <<= 1;
+	  st = entropy->ac_stats[tbl] +
+	       (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+	  while (v2 >>= 1) {
+	    arith_encode(cinfo, st, 1);
+	    m <<= 1;
+	    st += 1;
+	  }
+	}
+      }
+      arith_encode(cinfo, st, 0);
+      /* Figure F.9: Encoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	arith_encode(cinfo, st, (m & v) ? 1 : 0);
+    }
+    /* Encode EOB decision only if k < DCTSIZE2 */
+    if (k < DCTSIZE2) {
+      st = entropy->ac_stats[tbl] + 3 * (k - 1);
+      arith_encode(cinfo, st, 1);
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Initialize for an arithmetic-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass (j_compress_ptr cinfo, boolean gather_statistics)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+
+  if (gather_statistics)
+    /* Make sure to avoid that in the master control logic!
+     * We are fully adaptive here and need no extra
+     * statistics gathering pass!
+     */
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+
+  /* We assume jcmaster.c already validated the progressive scan parameters. */
+
+  /* Select execution routines */
+  if (cinfo->progressive_mode) {
+    if (cinfo->Ah == 0) {
+      if (cinfo->Ss == 0)
+	entropy->pub.encode_mcu = encode_mcu_DC_first;
+      else
+	entropy->pub.encode_mcu = encode_mcu_AC_first;
+    } else {
+      if (cinfo->Ss == 0)
+	entropy->pub.encode_mcu = encode_mcu_DC_refine;
+      else
+	entropy->pub.encode_mcu = encode_mcu_AC_refine;
+    }
+  } else
+    entropy->pub.encode_mcu = encode_mcu;
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Allocate & initialize requested statistics areas */
+    if (cinfo->progressive_mode == 0 || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      tbl = compptr->dc_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+	ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->dc_stats[tbl] == NULL)
+	entropy->dc_stats[tbl] = (unsigned char *) (*cinfo->mem->alloc_small)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, DC_STAT_BINS);
+      MEMZERO(entropy->dc_stats[tbl], DC_STAT_BINS);
+      /* Initialize DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    if (cinfo->progressive_mode == 0 || cinfo->Ss) {
+      tbl = compptr->ac_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+	ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->ac_stats[tbl] == NULL)
+	entropy->ac_stats[tbl] = (unsigned char *) (*cinfo->mem->alloc_small)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, AC_STAT_BINS);
+      MEMZERO(entropy->ac_stats[tbl], AC_STAT_BINS);
+#ifdef CALCULATE_SPECTRAL_CONDITIONING
+      if (cinfo->progressive_mode)
+	/* Section G.1.3.2: Set appropriate arithmetic conditioning value Kx */
+	cinfo->arith_ac_K[tbl] = cinfo->Ss + ((8 + cinfo->Se - cinfo->Ss) >> 4);
+#endif
+    }
+  }
+
+  /* Initialize arithmetic encoding variables */
+  entropy->c = 0;
+  entropy->a = 0x10000L;
+  entropy->sc = 0;
+  entropy->zc = 0;
+  entropy->ct = 11;
+  entropy->buffer = -1;  /* empty */
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/*
+ * Module initialization routine for arithmetic entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_arith_encoder (j_compress_ptr cinfo)
+{
+  arith_entropy_ptr entropy;
+  int i;
+
+  entropy = (arith_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(arith_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
+  entropy->pub.start_pass = start_pass;
+  entropy->pub.finish_pass = finish_pass;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    entropy->dc_stats[i] = NULL;
+    entropy->ac_stats[i] = NULL;
+  }
+}
diff --git a/jcinit.c b/jcinit.c
index 5efffe3..a0f3e4f 100644
--- a/jcinit.c
+++ b/jcinit.c
@@ -41,9 +41,9 @@
   /* Forward DCT */
   jinit_forward_dct(cinfo);
   /* Entropy encoding: either Huffman or arithmetic coding. */
-  if (cinfo->arith_code) {
-    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
-  } else {
+  if (cinfo->arith_code)
+    jinit_arith_encoder(cinfo);
+  else {
     if (cinfo->progressive_mode) {
 #ifdef C_PROGRESSIVE_SUPPORTED
       jinit_phuff_encoder(cinfo);
diff --git a/jcmarker.c b/jcmarker.c
index 3d1e6c6..2d7bdc3 100644
--- a/jcmarker.c
+++ b/jcmarker.c
@@ -529,7 +529,10 @@
 
   /* Emit the proper SOF marker */
   if (cinfo->arith_code) {
-    emit_sof(cinfo, M_SOF9);	/* SOF code for arithmetic coding */
+    if (cinfo->progressive_mode)
+      emit_sof(cinfo, M_SOF10); /* SOF code for progressive arithmetic */
+    else
+      emit_sof(cinfo, M_SOF9);  /* SOF code for sequential arithmetic */
   } else {
     if (cinfo->progressive_mode)
       emit_sof(cinfo, M_SOF2);	/* SOF code for progressive Huffman */
diff --git a/jcmaster.c b/jcmaster.c
index aab4020..366f7c0 100644
--- a/jcmaster.c
+++ b/jcmaster.c
@@ -433,7 +433,7 @@
     /* Do Huffman optimization for a scan after the first one. */
     select_scan_parameters(cinfo);
     per_scan_setup(cinfo);
-    if (cinfo->Ss != 0 || cinfo->Ah == 0 || cinfo->arith_code) {
+    if (cinfo->Ss != 0 || cinfo->Ah == 0) {
       (*cinfo->entropy->start_pass) (cinfo, TRUE);
       (*cinfo->coef->start_pass) (cinfo, JBUF_CRANK_DEST);
       master->pub.call_pass_startup = FALSE;
@@ -567,7 +567,7 @@
     cinfo->num_scans = 1;
   }
 
-  if (cinfo->progressive_mode)	/*  TEMPORARY HACK ??? */
+  if (cinfo->progressive_mode && cinfo->arith_code == 0)	/*  TEMPORARY HACK ??? */
     cinfo->optimize_coding = TRUE; /* assume default tables no good for progressive mode */
 
   /* Initialize my private state */
diff --git a/jctrans.c b/jctrans.c
index 0e6d707..9753a43 100644
--- a/jctrans.c
+++ b/jctrans.c
@@ -167,7 +167,7 @@
 
   /* Entropy encoding: either Huffman or arithmetic coding. */
   if (cinfo->arith_code) {
-    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+    jinit_arith_encoder(cinfo);
   } else {
     if (cinfo->progressive_mode) {
 #ifdef C_PROGRESSIVE_SUPPORTED
diff --git a/jdarith.c b/jdarith.c
new file mode 100644
index 0000000..1ef513e
--- /dev/null
+++ b/jdarith.c
@@ -0,0 +1,762 @@
+/*
+ * jdarith.c
+ *
+ * Copyright (C) 1997, Guido Vollbeding <guivol@esc.de>.
+ * This file is NOT part of the Independent JPEG Group's software
+ * for legal reasons.
+ * See the accompanying README file for conditions of distribution and use.
+ *
+ * This file contains portable arithmetic entropy decoding routines for JPEG
+ * (implementing the ISO/IEC IS 10918-1 and CCITT Recommendation ITU-T T.81).
+ *
+ * Both sequential and progressive modes are supported in this single module.
+ *
+ * Suspension is not currently supported in this module.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Expanded entropy decoder object for arithmetic decoding. */
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  INT32 c;       /* C register, base of coding interval + input bit buffer */
+  INT32 a;               /* A register, normalized size of coding interval */
+  int ct;     /* bit shift counter, # of bits left in bit buffer part of C */
+                                                         /* init: ct = -16 */
+                                                         /* run: ct = 0..7 */
+                                                         /* error: ct = -1 */
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+  int dc_context[MAX_COMPS_IN_SCAN]; /* context index for DC conditioning */
+
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+
+  /* Pointers to statistics areas (these workspaces have image lifespan) */
+  unsigned char * dc_stats[NUM_ARITH_TBLS];
+  unsigned char * ac_stats[NUM_ARITH_TBLS];
+} arith_entropy_decoder;
+
+typedef arith_entropy_decoder * arith_entropy_ptr;
+
+/* The following two definitions specify the allocation chunk size
+ * for the statistics area.
+ * According to sections F.1.4.4.1.3 and F.1.4.4.2, we need at least
+ * 49 statistics bins for DC, and 245 statistics bins for AC coding.
+ * Note that we use one additional AC bin for codings with fixed
+ * probability (0.5), thus the minimum number for AC is 246.
+ *
+ * We use a compact representation with 1 byte per statistics bin,
+ * thus the numbers directly represent byte sizes.
+ * This 1 byte per statistics bin contains the meaning of the MPS
+ * (more probable symbol) in the highest bit (mask 0x80), and the
+ * index into the probability estimation state machine table
+ * in the lower bits (mask 0x7F).
+ */
+
+#define DC_STAT_BINS 64
+#define AC_STAT_BINS 256
+
+
+LOCAL(int)
+get_byte (j_decompress_ptr cinfo)
+/* Read next input byte; we do not support suspension in this module. */
+{
+  struct jpeg_source_mgr * src = cinfo->src;
+
+  if (src->bytes_in_buffer == 0)
+    if (! (*src->fill_input_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+  src->bytes_in_buffer--;
+  return GETJOCTET(*src->next_input_byte++);
+}
+
+
+/*
+ * The core arithmetic decoding routine (common in JPEG and JBIG).
+ * This needs to go as fast as possible.
+ * Machine-dependent optimization facilities
+ * are not utilized in this portable implementation.
+ * However, this code should be fairly efficient and
+ * may be a good base for further optimizations anyway.
+ *
+ * Return value is 0 or 1 (binary decision).
+ *
+ * Note: I've changed the handling of the code base & bit
+ * buffer register C compared to other implementations
+ * based on the standards layout & procedures.
+ * While it also contains both the actual base of the
+ * coding interval (16 bits) and the next-bits buffer,
+ * the cut-point between these two parts is floating
+ * (instead of fixed) with the bit shift counter CT.
+ * Thus, we also need only one (variable instead of
+ * fixed size) shift for the LPS/MPS decision, and
+ * we can get away with any renormalization update
+ * of C (except for new data insertion, of course).
+ *
+ * I've also introduced a new scheme for accessing
+ * the probability estimation state machine table,
+ * derived from Markus Kuhn's JBIG implementation.
+ */
+
+LOCAL(int)
+arith_decode (j_decompress_ptr cinfo, unsigned char *st)
+{
+  extern const INT32 jaritab[];
+  register arith_entropy_ptr e = (arith_entropy_ptr) cinfo->entropy;
+  register unsigned char nl, nm;
+  register INT32 qe, temp;
+  register int sv, data;
+
+  /* Renormalization & data input per section D.2.6 */
+  while (e->a < 0x8000L) {
+    if (--e->ct < 0) {
+      /* Need to fetch next data byte */
+      if (cinfo->unread_marker)
+	data = 0;		/* stuff zero data */
+      else {
+	data = get_byte(cinfo);	/* read next input byte */
+	if (data == 0xFF) {	/* zero stuff or marker code */
+	  do data = get_byte(cinfo);
+	  while (data == 0xFF);	/* swallow extra 0xFF bytes */
+	  if (data == 0)
+	    data = 0xFF;	/* discard stuffed zero byte */
+	  else {
+	    /* Note: Different from the Huffman decoder, hitting
+	     * a marker while processing the compressed data
+	     * segment is legal in arithmetic coding.
+	     * The convention is to supply zero data
+	     * then until decoding is complete.
+	     */
+	    cinfo->unread_marker = data;
+	    data = 0;
+	  }
+	}
+      }
+      e->c = (e->c << 8) | data; /* insert data into C register */
+      if ((e->ct += 8) < 0)	 /* update bit shift counter */
+	/* Need more initial bytes */
+	if (++e->ct == 0)
+	  /* Got 2 initial bytes -> re-init A and exit loop */
+	  e->a = 0x8000L; /* => e->a = 0x10000L after loop exit */
+    }
+    e->a <<= 1;
+  }
+
+  /* Fetch values from our compact representation of Table D.2:
+   * Qe values and probability estimation state machine
+   */
+  sv = *st;
+  qe = jaritab[sv & 0x7F];	/* => Qe_Value */
+  nl = qe & 0xFF; qe >>= 8;	/* Next_Index_LPS + Switch_MPS */
+  nm = qe & 0xFF; qe >>= 8;	/* Next_Index_MPS */
+
+  /* Decode & estimation procedures per sections D.2.4 & D.2.5 */
+  temp = e->a - qe;
+  e->a = temp;
+  temp <<= e->ct;
+  if (e->c >= temp) {
+    e->c -= temp;
+    /* Conditional LPS (less probable symbol) exchange */
+    if (e->a < qe) {
+      e->a = qe;
+      *st = (sv & 0x80) ^ nm;	/* Estimate_after_MPS */
+    } else {
+      e->a = qe;
+      *st = (sv & 0x80) ^ nl;	/* Estimate_after_LPS */
+      sv ^= 0x80;		/* Exchange LPS/MPS */
+    }
+  } else if (e->a < 0x8000L) {
+    /* Conditional MPS (more probable symbol) exchange */
+    if (e->a < qe) {
+      *st = (sv & 0x80) ^ nl;	/* Estimate_after_LPS */
+      sv ^= 0x80;		/* Exchange LPS/MPS */
+    } else {
+      *st = (sv & 0x80) ^ nm;	/* Estimate_after_MPS */
+    }
+  }
+
+  return sv >> 7;
+}
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ */
+
+LOCAL(void)
+process_restart (j_decompress_ptr cinfo)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  int ci;
+  jpeg_component_info * compptr;
+
+  /* Advance past the RSTn marker */
+  if (! (*cinfo->marker->read_restart_marker) (cinfo))
+    ERREXIT(cinfo, JERR_CANT_SUSPEND);
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Re-initialize statistics areas */
+    if (cinfo->progressive_mode == 0 || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      MEMZERO(entropy->dc_stats[compptr->dc_tbl_no], DC_STAT_BINS);
+      /* Reset DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    if (cinfo->progressive_mode == 0 || cinfo->Ss) {
+      MEMZERO(entropy->ac_stats[compptr->ac_tbl_no], AC_STAT_BINS);
+    }
+  }
+
+  /* Reset arithmetic decoding variables */
+  entropy->c = 0;
+  entropy->a = 0;
+  entropy->ct = -16;	/* force reading 2 initial bytes to fill C */
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Arithmetic MCU decoding.
+ * Each of these routines decodes and returns one MCU's worth of
+ * arithmetic-compressed coefficients.
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA IS INITIALLY ZEROED BY THE CALLER.
+ */
+
+/*
+ * MCU decoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl, sign;
+  int v, m;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;	/* if error do nothing */
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    tbl = cinfo->cur_comp_info[ci]->dc_tbl_no;
+
+    /* Sections F.2.4.1 & F.1.4.4.1: Decoding of DC coefficients */
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.19: Decode_DC_DIFF */
+    if (arith_decode(cinfo, st) == 0)
+      entropy->dc_context[ci] = 0;
+    else {
+      /* Figure F.21: Decoding nonzero value v */
+      /* Figure F.22: Decoding the sign of v */
+      sign = arith_decode(cinfo, st + 1);
+      st += 2; st += sign;
+      /* Figure F.23: Decoding the magnitude category of v */
+      if ((m = arith_decode(cinfo, st)) != 0) {
+	st = entropy->dc_stats[tbl] + 20;	/* Table F.4: X1 = 20 */
+	while (arith_decode(cinfo, st)) {
+	  if ((m <<= 1) == 0x8000) {
+	    WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	    entropy->ct = -1;			/* magnitude overflow */
+	    return TRUE;
+	  }
+	  st += 1;
+	}
+      }
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int) (((INT32) 1 << cinfo->arith_dc_L[tbl]) >> 1))
+	entropy->dc_context[ci] = 0;		   /* zero diff category */
+      else if (m > (int) (((INT32) 1 << cinfo->arith_dc_U[tbl]) >> 1))
+	entropy->dc_context[ci] = 12 + (sign * 4); /* large diff category */
+      else
+	entropy->dc_context[ci] = 4 + (sign * 4);  /* small diff category */
+      v = m;
+      /* Figure F.24: Decoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	if (arith_decode(cinfo, st)) v |= m;
+      v += 1; if (sign) v = -v;
+      entropy->last_dc_val[ci] += v;
+    }
+
+    /* Scale and output the DC coefficient (assumes jpeg_natural_order[0]=0) */
+    (*block)[0] = (JCOEF) (entropy->last_dc_val[ci] << cinfo->Al);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int tbl, sign, k;
+  int v, m;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;	/* if error do nothing */
+
+  /* There is always only one block per MCU */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  /* Sections F.2.4.2 & F.1.4.4.2: Decoding of AC coefficients */
+
+  /* Figure F.20: Decode_AC_coefficients */
+  for (k = cinfo->Ss; k <= cinfo->Se; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    if (arith_decode(cinfo, st)) break;		/* EOB flag */
+    while (arith_decode(cinfo, st + 1) == 0) {
+      st += 3; k++;
+      if (k > cinfo->Se) {
+	WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	entropy->ct = -1;			/* spectral overflow */
+	return TRUE;
+      }
+    }
+    /* Figure F.21: Decoding nonzero value v */
+    /* Figure F.22: Decoding the sign of v */
+    entropy->ac_stats[tbl][245] = 0;
+    sign = arith_decode(cinfo, entropy->ac_stats[tbl] + 245);
+    st += 2;
+    /* Figure F.23: Decoding the magnitude category of v */
+    if ((m = arith_decode(cinfo, st)) != 0) {
+      if (arith_decode(cinfo, st)) {
+	m <<= 1;
+	st = entropy->ac_stats[tbl] +
+	     (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+	while (arith_decode(cinfo, st)) {
+	  if ((m <<= 1) == 0x8000) {
+	    WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	    entropy->ct = -1;			/* magnitude overflow */
+	    return TRUE;
+	  }
+	  st += 1;
+	}
+      }
+    }
+    v = m;
+    /* Figure F.24: Decoding the magnitude bit pattern of v */
+    st += 14;
+    while (m >>= 1)
+      if (arith_decode(cinfo, st)) v |= m;
+    v += 1; if (sign) v = -v;
+    /* Scale and output coefficient in natural (dezigzagged) order */
+    (*block)[jpeg_natural_order[k]] = (JCOEF) (v << cinfo->Al);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for DC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  unsigned char st[4];
+  int p1, blkn;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  p1 = 1 << cinfo->Al;		/* 1 in the bit position being coded */
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    st[0] = 0;	/* use fixed probability estimation */
+    /* Encoded data is simply the next bit of the two's-complement DC value */
+    if (arith_decode(cinfo, st))
+      MCU_data[blkn][0][0] |= p1;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  JCOEFPTR thiscoef;
+  unsigned char *st;
+  int tbl, k, kex;
+  int p1, m1;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;	/* if error do nothing */
+
+  /* There is always only one block per MCU */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  p1 = 1 << cinfo->Al;		/* 1 in the bit position being coded */
+  m1 = (-1) << cinfo->Al;	/* -1 in the bit position being coded */
+
+  /* Establish EOBx (previous stage end-of-block) index */
+  for (kex = cinfo->Se + 1; kex > 1; kex--)
+    if ((*block)[jpeg_natural_order[kex - 1]]) break;
+
+  for (k = cinfo->Ss; k <= cinfo->Se; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    if (k >= kex)
+      if (arith_decode(cinfo, st)) break;	/* EOB flag */
+    for (;;) {
+      thiscoef = *block + jpeg_natural_order[k];
+      if (*thiscoef) {				/* previously nonzero coef */
+	if (arith_decode(cinfo, st + 2))
+	  if (*thiscoef < 0)
+	    *thiscoef += m1;
+	  else
+	    *thiscoef += p1;
+	break;
+      }
+      if (arith_decode(cinfo, st + 1)) {	/* newly nonzero coef */
+	entropy->ac_stats[tbl][245] = 0;
+	if (arith_decode(cinfo, entropy->ac_stats[tbl] + 245))
+	  *thiscoef = m1;
+	else
+	  *thiscoef = p1;
+	break;
+      }
+      st += 3; k++;
+      if (k > cinfo->Se) {
+	WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	entropy->ct = -1;			/* spectral overflow */
+	return TRUE;
+      }
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Decode one MCU's worth of arithmetic-compressed coefficients.
+ */
+
+METHODDEF(boolean)
+decode_mcu (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  jpeg_component_info * compptr;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl, sign, k;
+  int v, m;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;	/* if error do nothing */
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Sections F.2.4.1 & F.1.4.4.1: Decoding of DC coefficients */
+
+    tbl = compptr->dc_tbl_no;
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.19: Decode_DC_DIFF */
+    if (arith_decode(cinfo, st) == 0)
+      entropy->dc_context[ci] = 0;
+    else {
+      /* Figure F.21: Decoding nonzero value v */
+      /* Figure F.22: Decoding the sign of v */
+      sign = arith_decode(cinfo, st + 1);
+      st += 2; st += sign;
+      /* Figure F.23: Decoding the magnitude category of v */
+      if ((m = arith_decode(cinfo, st)) != 0) {
+	st = entropy->dc_stats[tbl] + 20;	/* Table F.4: X1 = 20 */
+	while (arith_decode(cinfo, st)) {
+	  if ((m <<= 1) == 0x8000) {
+	    WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	    entropy->ct = -1;			/* magnitude overflow */
+	    return TRUE;
+	  }
+	  st += 1;
+	}
+      }
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int) (((INT32) 1 << cinfo->arith_dc_L[tbl]) >> 1))
+	entropy->dc_context[ci] = 0;		   /* zero diff category */
+      else if (m > (int) (((INT32) 1 << cinfo->arith_dc_U[tbl]) >> 1))
+	entropy->dc_context[ci] = 12 + (sign * 4); /* large diff category */
+      else
+	entropy->dc_context[ci] = 4 + (sign * 4);  /* small diff category */
+      v = m;
+      /* Figure F.24: Decoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	if (arith_decode(cinfo, st)) v |= m;
+      v += 1; if (sign) v = -v;
+      entropy->last_dc_val[ci] += v;
+    }
+
+    (*block)[0] = (JCOEF) entropy->last_dc_val[ci];
+
+    /* Sections F.2.4.2 & F.1.4.4.2: Decoding of AC coefficients */
+
+    tbl = compptr->ac_tbl_no;
+
+    /* Figure F.20: Decode_AC_coefficients */
+    for (k = 1; k < DCTSIZE2; k++) {
+      st = entropy->ac_stats[tbl] + 3 * (k - 1);
+      if (arith_decode(cinfo, st)) break;	/* EOB flag */
+      while (arith_decode(cinfo, st + 1) == 0) {
+	st += 3; k++;
+	if (k >= DCTSIZE2) {
+	  WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	  entropy->ct = -1;			/* spectral overflow */
+	  return TRUE;
+	}
+      }
+      /* Figure F.21: Decoding nonzero value v */
+      /* Figure F.22: Decoding the sign of v */
+      entropy->ac_stats[tbl][245] = 0;
+      sign = arith_decode(cinfo, entropy->ac_stats[tbl] + 245);
+      st += 2;
+      /* Figure F.23: Decoding the magnitude category of v */
+      if ((m = arith_decode(cinfo, st)) != 0) {
+	if (arith_decode(cinfo, st)) {
+	  m <<= 1;
+	  st = entropy->ac_stats[tbl] +
+	       (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+	  while (arith_decode(cinfo, st)) {
+	    if ((m <<= 1) == 0x8000) {
+	      WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	      entropy->ct = -1;			/* magnitude overflow */
+	      return TRUE;
+	    }
+	    st += 1;
+	  }
+	}
+      }
+      v = m;
+      /* Figure F.24: Decoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	if (arith_decode(cinfo, st)) v |= m;
+      v += 1; if (sign) v = -v;
+      (*block)[jpeg_natural_order[k]] = (JCOEF) v;
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Initialize for an arithmetic-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass (j_decompress_ptr cinfo)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+
+  if (cinfo->progressive_mode) {
+    /* Validate progressive scan parameters */
+    if (cinfo->Ss == 0) {
+      if (cinfo->Se != 0)
+	goto bad;
+    } else {
+      /* need not check Ss/Se < 0 since they came from unsigned bytes */
+      if (cinfo->Se < cinfo->Ss || cinfo->Se >= DCTSIZE2)
+	goto bad;
+      /* AC scans may have only one component */
+      if (cinfo->comps_in_scan != 1)
+	goto bad;
+    }
+    if (cinfo->Ah != 0) {
+      /* Successive approximation refinement scan: must have Al = Ah-1. */
+      if (cinfo->Ah-1 != cinfo->Al)
+	goto bad;
+    }
+    if (cinfo->Al > 13) {	/* need not check for < 0 */
+      bad:
+      ERREXIT4(cinfo, JERR_BAD_PROGRESSION,
+	       cinfo->Ss, cinfo->Se, cinfo->Ah, cinfo->Al);
+    }
+    /* Update progression status, and verify that scan order is legal.
+     * Note that inter-scan inconsistencies are treated as warnings
+     * not fatal errors ... not clear if this is right way to behave.
+     */
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      int coefi, cindex = cinfo->cur_comp_info[ci]->component_index;
+      int *coef_bit_ptr = & cinfo->coef_bits[cindex][0];
+      if (cinfo->Ss && coef_bit_ptr[0] < 0) /* AC without prior DC scan */
+	WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, 0);
+      for (coefi = cinfo->Ss; coefi <= cinfo->Se; coefi++) {
+	int expected = (coef_bit_ptr[coefi] < 0) ? 0 : coef_bit_ptr[coefi];
+	if (cinfo->Ah != expected)
+	  WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, coefi);
+	coef_bit_ptr[coefi] = cinfo->Al;
+      }
+    }
+    /* Select MCU decoding routine */
+    if (cinfo->Ah == 0) {
+      if (cinfo->Ss == 0)
+	entropy->pub.decode_mcu = decode_mcu_DC_first;
+      else
+	entropy->pub.decode_mcu = decode_mcu_AC_first;
+    } else {
+      if (cinfo->Ss == 0)
+	entropy->pub.decode_mcu = decode_mcu_DC_refine;
+      else
+	entropy->pub.decode_mcu = decode_mcu_AC_refine;
+    }
+  } else {
+    /* Check that the scan parameters Ss, Se, Ah/Al are OK for sequential JPEG.
+     * This ought to be an error condition, but we make it a warning because
+     * there are some baseline files out there with all zeroes in these bytes.
+     */
+    if (cinfo->Ss != 0 || cinfo->Se != DCTSIZE2-1 ||
+	cinfo->Ah != 0 || cinfo->Al != 0)
+      WARNMS(cinfo, JWRN_NOT_SEQUENTIAL);
+    /* Select MCU decoding routine */
+    entropy->pub.decode_mcu = decode_mcu;
+  }
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Allocate & initialize requested statistics areas */
+    if (cinfo->progressive_mode == 0 || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      tbl = compptr->dc_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+	ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->dc_stats[tbl] == NULL)
+	entropy->dc_stats[tbl] = (unsigned char *) (*cinfo->mem->alloc_small)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, DC_STAT_BINS);
+      MEMZERO(entropy->dc_stats[tbl], DC_STAT_BINS);
+      /* Initialize DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    if (cinfo->progressive_mode == 0 || cinfo->Ss) {
+      tbl = compptr->ac_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+	ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->ac_stats[tbl] == NULL)
+	entropy->ac_stats[tbl] = (unsigned char *) (*cinfo->mem->alloc_small)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, AC_STAT_BINS);
+      MEMZERO(entropy->ac_stats[tbl], AC_STAT_BINS);
+    }
+  }
+
+  /* Initialize arithmetic decoding variables */
+  entropy->c = 0;
+  entropy->a = 0;
+  entropy->ct = -16;	/* force reading 2 initial bytes to fill C */
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Module initialization routine for arithmetic entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_arith_decoder (j_decompress_ptr cinfo)
+{
+  arith_entropy_ptr entropy;
+  int i;
+
+  entropy = (arith_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(arith_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *) entropy;
+  entropy->pub.start_pass = start_pass;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    entropy->dc_stats[i] = NULL;
+    entropy->ac_stats[i] = NULL;
+  }
+
+  if (cinfo->progressive_mode) {
+    /* Create progression status table */
+    int *coef_bit_ptr, ci;
+    cinfo->coef_bits = (int (*)[DCTSIZE2])
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  cinfo->num_components*DCTSIZE2*SIZEOF(int));
+    coef_bit_ptr = & cinfo->coef_bits[0][0];
+    for (ci = 0; ci < cinfo->num_components; ci++) 
+      for (i = 0; i < DCTSIZE2; i++)
+	*coef_bit_ptr++ = -1;
+  }
+}
diff --git a/jdmaster.c b/jdmaster.c
index 2802c5b..6d23fee 100644
--- a/jdmaster.c
+++ b/jdmaster.c
@@ -373,7 +373,7 @@
   jinit_inverse_dct(cinfo);
   /* Entropy decoding: either Huffman or arithmetic coding. */
   if (cinfo->arith_code) {
-    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+    jinit_arith_decoder(cinfo);
   } else {
     if (cinfo->progressive_mode) {
 #ifdef D_PROGRESSIVE_SUPPORTED
diff --git a/jdtrans.c b/jdtrans.c
index 6c0ab71..7a3b268 100644
--- a/jdtrans.c
+++ b/jdtrans.c
@@ -100,9 +100,9 @@
   cinfo->buffered_image = TRUE;
 
   /* Entropy decoding: either Huffman or arithmetic coding. */
-  if (cinfo->arith_code) {
-    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
-  } else {
+  if (cinfo->arith_code)
+    jinit_arith_decoder(cinfo);
+  else {
     if (cinfo->progressive_mode) {
 #ifdef D_PROGRESSIVE_SUPPORTED
       jinit_phuff_decoder(cinfo);
diff --git a/jerror.h b/jerror.h
index fc2fffe..109e3d3 100644
--- a/jerror.h
+++ b/jerror.h
@@ -93,6 +93,7 @@
 JMESSAGE(JERR_MODE_CHANGE, "Invalid color quantization mode change")
 JMESSAGE(JERR_NOTIMPL, "Not implemented yet")
 JMESSAGE(JERR_NOT_COMPILED, "Requested feature was omitted at compile time")
+JMESSAGE(JERR_NO_ARITH_TABLE, "Arithmetic table 0x%02x was not defined")
 JMESSAGE(JERR_NO_BACKING_STORE, "Backing store not supported")
 JMESSAGE(JERR_NO_HUFF_TABLE, "Huffman table 0x%02x was not defined")
 JMESSAGE(JERR_NO_IMAGE, "JPEG datastream contains no image")
@@ -170,6 +171,7 @@
 JMESSAGE(JTRC_XMS_CLOSE, "Freed XMS handle %u")
 JMESSAGE(JTRC_XMS_OPEN, "Obtained XMS handle %u")
 JMESSAGE(JWRN_ADOBE_XFORM, "Unknown Adobe color transform code %d")
+JMESSAGE(JWRN_ARITH_BAD_CODE, "Corrupt JPEG data: bad arithmetic code")
 JMESSAGE(JWRN_BOGUS_PROGRESSION,
 	 "Inconsistent progression sequence for component %d coefficient %d")
 JMESSAGE(JWRN_EXTRANEOUS_DATA,
diff --git a/jmorecfg.h b/jmorecfg.h
index 54a7d1c..bd6d605 100644
--- a/jmorecfg.h
+++ b/jmorecfg.h
@@ -266,7 +266,7 @@
 
 /* Encoder capability options: */
 
-#undef  C_ARITH_CODING_SUPPORTED    /* Arithmetic coding back end? */
+#define C_ARITH_CODING_SUPPORTED    /* Arithmetic coding back end? */
 #define C_MULTISCAN_FILES_SUPPORTED /* Multiple-scan JPEG files? */
 #define C_PROGRESSIVE_SUPPORTED	    /* Progressive JPEG? (Requires MULTISCAN)*/
 #define ENTROPY_OPT_SUPPORTED	    /* Optimization of entropy coding parms? */
@@ -282,7 +282,7 @@
 
 /* Decoder capability options: */
 
-#undef  D_ARITH_CODING_SUPPORTED    /* Arithmetic coding back end? */
+#define D_ARITH_CODING_SUPPORTED    /* Arithmetic coding back end? */
 #define D_MULTISCAN_FILES_SUPPORTED /* Multiple-scan JPEG files? */
 #define D_PROGRESSIVE_SUPPORTED	    /* Progressive JPEG? (Requires MULTISCAN)*/
 #define SAVE_MARKERS_SUPPORTED	    /* jpeg_save_markers() needed? */
diff --git a/jpegint.h b/jpegint.h
index 95b00d4..3378133 100644
--- a/jpegint.h
+++ b/jpegint.h
@@ -345,6 +345,7 @@
 EXTERN(void) jinit_forward_dct JPP((j_compress_ptr cinfo));
 EXTERN(void) jinit_huff_encoder JPP((j_compress_ptr cinfo));
 EXTERN(void) jinit_phuff_encoder JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_arith_encoder JPP((j_compress_ptr cinfo));
 EXTERN(void) jinit_marker_writer JPP((j_compress_ptr cinfo));
 /* Decompression module initialization routines */
 EXTERN(void) jinit_master_decompress JPP((j_decompress_ptr cinfo));
@@ -358,6 +359,7 @@
 EXTERN(void) jinit_marker_reader JPP((j_decompress_ptr cinfo));
 EXTERN(void) jinit_huff_decoder JPP((j_decompress_ptr cinfo));
 EXTERN(void) jinit_phuff_decoder JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_arith_decoder JPP((j_decompress_ptr cinfo));
 EXTERN(void) jinit_inverse_dct JPP((j_decompress_ptr cinfo));
 EXTERN(void) jinit_upsampler JPP((j_decompress_ptr cinfo));
 EXTERN(void) jinit_color_deconverter JPP((j_decompress_ptr cinfo));
diff --git a/makefile.cfg b/makefile.cfg
index f25e42e..7968319 100644
--- a/makefile.cfg
+++ b/makefile.cfg
@@ -80,7 +80,7 @@
         jdinput.c jdmainct.c jdmarker.c jdmaster.c jdmerge.c jdphuff.c \
         jdpostct.c jdsample.c jdtrans.c jerror.c jfdctflt.c jfdctfst.c \
         jfdctint.c jidctflt.c jidctfst.c jidctint.c jidctred.c jquant1.c \
-        jquant2.c jutils.c jmemmgr.c
+        jquant2.c jutils.c jmemmgr.c jaricom.c jcarith.c jdarith.c
 # memmgr back ends: compile only one of these into a working library
 SYSDEPSOURCES= jmemansi.c jmemname.c jmemnobs.c jmemdos.c jmemmac.c
 # source files: cjpeg/djpeg/jpegtran applications, also rdjpgcom/wrjpgcom
@@ -110,19 +110,19 @@
 DISTFILES= $(DOCS) $(MKFILES) $(CONFIGFILES) $(SOURCES) $(INCLUDES) \
         $(CONFIGUREFILES) $(OTHERFILES) $(TESTFILES)
 # library object files common to compression and decompression
-COMOBJECTS= jcomapi.$(O) jutils.$(O) jerror.$(O) jmemmgr.$(O) $(SYSDEPMEM)
+COMOBJECTS= jcomapi.$(O) jutils.$(O) jerror.$(O) jmemmgr.$(O) jaricom.$(O) $(SYSDEPMEM)
 # compression library object files
 CLIBOBJECTS= jcapimin.$(O) jcapistd.$(O) jctrans.$(O) jcparam.$(O) \
         jdatadst.$(O) jcinit.$(O) jcmaster.$(O) jcmarker.$(O) jcmainct.$(O) \
         jcprepct.$(O) jccoefct.$(O) jccolor.$(O) jcsample.$(O) jchuff.$(O) \
         jcphuff.$(O) jcdctmgr.$(O) jfdctfst.$(O) jfdctflt.$(O) \
-        jfdctint.$(O)
+        jfdctint.$(O) jcarith.$(O)
 # decompression library object files
 DLIBOBJECTS= jdapimin.$(O) jdapistd.$(O) jdtrans.$(O) jdatasrc.$(O) \
         jdmaster.$(O) jdinput.$(O) jdmarker.$(O) jdhuff.$(O) jdphuff.$(O) \
         jdmainct.$(O) jdcoefct.$(O) jdpostct.$(O) jddctmgr.$(O) \
         jidctfst.$(O) jidctflt.$(O) jidctint.$(O) jidctred.$(O) \
-        jdsample.$(O) jdcolor.$(O) jquant1.$(O) jquant2.$(O) jdmerge.$(O)
+        jdsample.$(O) jdcolor.$(O) jquant1.$(O) jquant2.$(O) jdmerge.$(O) jdarith.$(O)
 # These objectfiles are included in libjpeg.a
 LIBOBJECTS= $(CLIBOBJECTS) $(DLIBOBJECTS) $(COMOBJECTS)
 # object files for sample applications (excluding library files)
@@ -317,3 +317,6 @@
 wrbmp.$(O): wrbmp.c cdjpeg.h jinclude.h jconfig.h jpeglib.h jmorecfg.h jerror.h cderror.h
 rdrle.$(O): rdrle.c cdjpeg.h jinclude.h jconfig.h jpeglib.h jmorecfg.h jerror.h cderror.h
 wrrle.$(O): wrrle.c cdjpeg.h jinclude.h jconfig.h jpeglib.h jmorecfg.h jerror.h cderror.h
+jcarith.$(O): jcarith.c jinclude.h jconfig.h jpeglib.h jmorecfg.h jpegint.h jerror.h
+jdarith.$(O): jdarith.c jinclude.h jconfig.h jpeglib.h jmorecfg.h jpegint.h jerror.h
+jaricom.$(O): jaricom.c jinclude.h jconfig.h jpeglib.h jmorecfg.h jpegint.h jerror.h