]> git.saurik.com Git - wxWidgets.git/blobdiff - src/png/pnggccrd.c
Linker fixes for OS/2 build.
[wxWidgets.git] / src / png / pnggccrd.c
index 2d1d1d3b369572b61c38856d0225039e27860fac..a7248d6ca64026a0b447cc936a5ed6a3b67ce658 100644 (file)
-/* pnggccrd.c - mixed C/assembler version of utilities to read a PNG file
- *
- * For Intel x86 CPU (Pentium-MMX or later) and GNU C compiler.
- *
- *     See http://www.intel.com/drg/pentiumII/appnotes/916/916.htm
- *     and http://www.intel.com/drg/pentiumII/appnotes/923/923.htm
- *     for Intel's performance analysis of the MMX vs. non-MMX code.
- *
- * libpng version 1.2.4 - July 8, 2002
- * For conditions of distribution and use, see copyright notice in png.h
- * Copyright (c) 1998-2002 Glenn Randers-Pehrson
- * Copyright (c) 1998, Intel Corporation
- *
- * Based on MSVC code contributed by Nirav Chhatrapati, Intel Corp., 1998.
- * Interface to libpng contributed by Gilles Vollant, 1999.
- * GNU C port by Greg Roelofs, 1999-2001.
- *
- * Lines 2350-4300 converted in place with intel2gas 1.3.1:
- *
- *   intel2gas -mdI pnggccrd.c.partially-msvc -o pnggccrd.c
- *
- * and then cleaned up by hand.  See http://hermes.terminal.at/intel2gas/ .
- *
- * NOTE:  A sufficiently recent version of GNU as (or as.exe under DOS/Windows)
- *        is required to assemble the newer MMX instructions such as movq.
- *        For djgpp, see
- *
- *           ftp://ftp.simtel.net/pub/simtelnet/gnu/djgpp/v2gnu/bnu281b.zip
- *
- *        (or a later version in the same directory).  For Linux, check your
- *        distribution's web site(s) or try these links:
- *
- *           http://rufus.w3.org/linux/RPM/binutils.html
- *           http://www.debian.org/Packages/stable/devel/binutils.html
- *           ftp://ftp.slackware.com/pub/linux/slackware/slackware/slakware/d1/
- *             binutils.tgz
- *
- *        For other platforms, see the main GNU site:
- *
- *           ftp://ftp.gnu.org/pub/gnu/binutils/
- *
- *        Version 2.5.2l.15 is definitely too old...
- */
+/* pnggccrd.c was removed from libpng-1.2.20. */
 
-/*
- * TEMPORARY PORTING NOTES AND CHANGELOG (mostly by Greg Roelofs)
- * =====================================
- *
- * 19991006:
- *  - fixed sign error in post-MMX cleanup code (16- & 32-bit cases)
- *
- * 19991007:
- *  - additional optimizations (possible or definite):
- *     x [DONE] write MMX code for 64-bit case (pixel_bytes == 8) [not tested]
- *     - write MMX code for 48-bit case (pixel_bytes == 6)
- *     - figure out what's up with 24-bit case (pixel_bytes == 3):
- *        why subtract 8 from width_mmx in the pass 4/5 case?
- *        (only width_mmx case) (near line 1606)
- *     x [DONE] replace pixel_bytes within each block with the true
- *        constant value (or are compilers smart enough to do that?)
- *     - rewrite all MMX interlacing code so it's aligned with
- *        the *beginning* of the row buffer, not the end.  This
- *        would not only allow one to eliminate half of the memory
- *        writes for odd passes (that is, pass == odd), it may also
- *        eliminate some unaligned-data-access exceptions (assuming
- *        there's a penalty for not aligning 64-bit accesses on
- *        64-bit boundaries).  The only catch is that the "leftover"
- *        pixel(s) at the end of the row would have to be saved,
- *        but there are enough unused MMX registers in every case,
- *        so this is not a problem.  A further benefit is that the
- *        post-MMX cleanup code (C code) in at least some of the
- *        cases could be done within the assembler block.
- *  x [DONE] the "v3 v2 v1 v0 v7 v6 v5 v4" comments are confusing,
- *     inconsistent, and don't match the MMX Programmer's Reference
- *     Manual conventions anyway.  They should be changed to
- *     "b7 b6 b5 b4 b3 b2 b1 b0," where b0 indicates the byte that
- *     was lowest in memory (e.g., corresponding to a left pixel)
- *     and b7 is the byte that was highest (e.g., a right pixel).
- *
- * 19991016:
- *  - Brennan's Guide notwithstanding, gcc under Linux does *not*
- *     want globals prefixed by underscores when referencing them--
- *     i.e., if the variable is const4, then refer to it as const4,
- *     not _const4.  This seems to be a djgpp-specific requirement.
- *     Also, such variables apparently *must* be declared outside
- *     of functions; neither static nor automatic variables work if
- *     defined within the scope of a single function, but both
- *     static and truly global (multi-module) variables work fine.
- *
- * 19991023:
- *  - fixed png_combine_row() non-MMX replication bug (odd passes only?)
- *  - switched from string-concatenation-with-macros to cleaner method of
- *     renaming global variables for djgpp--i.e., always use prefixes in
- *     inlined assembler code (== strings) and conditionally rename the
- *     variables, not the other way around.  Hence _const4, _mask8_0, etc.
- *
- * 19991024:
- *  - fixed mmxsupport()/png_do_read_interlace() first-row bug
- *     This one was severely weird:  even though mmxsupport() doesn't touch
- *     ebx (where "row" pointer was stored), it nevertheless managed to zero
- *     the register (even in static/non-fPIC code--see below), which in turn
- *     caused png_do_read_interlace() to return prematurely on the first row of
- *     interlaced images (i.e., without expanding the interlaced pixels).
- *     Inspection of the generated assembly code didn't turn up any clues,
- *     although it did point at a minor optimization (i.e., get rid of
- *     mmx_supported_local variable and just use eax).  Possibly the CPUID
- *     instruction is more destructive than it looks?  (Not yet checked.)
- *  - "info gcc" was next to useless, so compared fPIC and non-fPIC assembly
- *     listings...  Apparently register spillage has to do with ebx, since
- *     it's used to index the global offset table.  Commenting it out of the
- *     input-reg lists in png_combine_row() eliminated compiler barfage, so
- *     ifdef'd with __PIC__ macro:  if defined, use a global for unmask
- *
- * 19991107:
- *  - verified CPUID clobberage:  12-char string constant ("GenuineIntel",
- *     "AuthenticAMD", etc.) placed in ebx:ecx:edx.  Still need to polish.
- *
- * 19991120:
- *  - made "diff" variable (now "_dif") global to simplify conversion of
- *     filtering routines (running out of regs, sigh).  "diff" is still used
- *     in interlacing routines, however.
- *  - fixed up both versions of mmxsupport() (ORIG_THAT_USED_TO_CLOBBER_EBX
- *     macro determines which is used); original not yet tested.
- *
- * 20000213:
- *  - when compiling with gcc, be sure to use  -fomit-frame-pointer
- *
- * 20000319:
- *  - fixed a register-name typo in png_do_read_interlace(), default (MMX) case,
- *     pass == 4 or 5, that caused visible corruption of interlaced images
- *
- * 20000623:
- *  - Various problems were reported with gcc 2.95.2 in the Cygwin environment,
- *     many of the form "forbidden register 0 (ax) was spilled for class AREG."
- *     This is explained at http://gcc.gnu.org/fom_serv/cache/23.html, and
- *     Chuck Wilson supplied a patch involving dummy output registers.  See
- *     http://sourceforge.net/bugs/?func=detailbug&bug_id=108741&group_id=5624
- *     for the original (anonymous) SourceForge bug report.
- *
- * 20000706:
- *  - Chuck Wilson passed along these remaining gcc 2.95.2 errors:
- *       pnggccrd.c: In function `png_combine_row':
- *       pnggccrd.c:525: more than 10 operands in `asm'
- *       pnggccrd.c:669: more than 10 operands in `asm'
- *       pnggccrd.c:828: more than 10 operands in `asm'
- *       pnggccrd.c:994: more than 10 operands in `asm'
- *       pnggccrd.c:1177: more than 10 operands in `asm'
- *     They are all the same problem and can be worked around by using the
- *     global _unmask variable unconditionally, not just in the -fPIC case.
- *     Reportedly earlier versions of gcc also have the problem with more than
- *     10 operands; they just don't report it.  Much strangeness ensues, etc.
- *
- * 20000729:
- *  - enabled png_read_filter_row_mmx_up() (shortest remaining unconverted
- *     MMX routine); began converting png_read_filter_row_mmx_sub()
- *  - to finish remaining sections:
- *     - clean up indentation and comments
- *     - preload local variables
- *     - add output and input regs (order of former determines numerical
- *        mapping of latter)
- *     - avoid all usage of ebx (including bx, bh, bl) register [20000823]
- *     - remove "$" from addressing of Shift and Mask variables [20000823]
- *
- * 20000731:
- *  - global union vars causing segfaults in png_read_filter_row_mmx_sub()?
- *
- * 20000822:
- *  - ARGH, stupid png_read_filter_row_mmx_sub() segfault only happens with
- *     shared-library (-fPIC) version!  Code works just fine as part of static
- *     library.  Damn damn damn damn damn, should have tested that sooner.
- *     ebx is getting clobbered again (explicitly this time); need to save it
- *     on stack or rewrite asm code to avoid using it altogether.  Blargh!
- *
- * 20000823:
- *  - first section was trickiest; all remaining sections have ebx -> edx now.
- *     (-fPIC works again.)  Also added missing underscores to various Shift*
- *     and *Mask* globals and got rid of leading "$" signs.
- *
- * 20000826:
- *  - added visual separators to help navigate microscopic printed copies
- *     (http://pobox.com/~newt/code/gpr-latest.zip, mode 10); started working
- *     on png_read_filter_row_mmx_avg()
- *
- * 20000828:
- *  - finished png_read_filter_row_mmx_avg():  only Paeth left! (930 lines...)
- *     What the hell, did png_read_filter_row_mmx_paeth(), too.  Comments not
- *     cleaned up/shortened in either routine, but functionality is complete
- *     and seems to be working fine.
- *
- * 20000829:
- *  - ahhh, figured out last(?) bit of gcc/gas asm-fu:  if register is listed
- *     as an input reg (with dummy output variables, etc.), then it *cannot*
- *     also appear in the clobber list or gcc 2.95.2 will barf.  The solution
- *     is simple enough...
- *
- * 20000914:
- *  - bug in png_read_filter_row_mmx_avg():  16-bit grayscale not handled
- *     correctly (but 48-bit RGB just fine)
- *
- * 20000916:
- *  - fixed bug in png_read_filter_row_mmx_avg(), bpp == 2 case; three errors:
- *     - "_ShiftBpp.use = 24;"      should have been   "_ShiftBpp.use = 16;"
- *     - "_ShiftRem.use = 40;"      should have been   "_ShiftRem.use = 48;"
- *     - "psllq _ShiftRem, %%mm2"   should have been   "psrlq _ShiftRem, %%mm2"
- *
- * 20010101:
- *  - added new png_init_mmx_flags() function (here only because it needs to
- *     call mmxsupport(), which should probably become global png_mmxsupport());
- *     modified other MMX routines to run conditionally (png_ptr->asm_flags)
- *
- * 20010103:
- *  - renamed mmxsupport() to png_mmx_support(), with auto-set of mmx_supported,
- *     and made it public; moved png_init_mmx_flags() to png.c as internal func
- *
- * 20010104:
- *  - removed dependency on png_read_filter_row_c() (C code already duplicated
- *     within MMX version of png_read_filter_row()) so no longer necessary to
- *     compile it into pngrutil.o
- *
- * 20010310:
- *  - fixed buffer-overrun bug in png_combine_row() C code (non-MMX)
- *
- * 20020304:
- *  - eliminated incorrect use of width_mmx in pixel_bytes == 8 case
- *
- * STILL TO DO:
- *     - test png_do_read_interlace() 64-bit case (pixel_bytes == 8)
- *     - write MMX code for 48-bit case (pixel_bytes == 6)
- *     - figure out what's up with 24-bit case (pixel_bytes == 3):
- *        why subtract 8 from width_mmx in the pass 4/5 case?
- *        (only width_mmx case) (near line 1606)
- *     - rewrite all MMX interlacing code so it's aligned with beginning
- *        of the row buffer, not the end (see 19991007 for details)
- *     x pick one version of mmxsupport() and get rid of the other
- *     - add error messages to any remaining bogus default cases
- *     - enable pixel_depth == 8 cases in png_read_filter_row()? (test speed)
- *     x add support for runtime enable/disable/query of various MMX routines
- */
+/* This code snippet is for use by configure's compilation test. */
 
-#define PNG_INTERNAL
-#include "png.h"
+#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && \
+    defined(PNG_MMX_CODE_SUPPORTED)
+int PNGAPI png_dummy_mmx_support(void);
 
-#if defined(PNG_USE_PNGGCCRD)
+static int _mmx_supported = 2; // 0: no MMX; 1: MMX supported; 2: not tested
 
-int PNGAPI png_mmx_support(void);
-
-#ifdef PNG_USE_LOCAL_ARRAYS
-static const int FARDATA png_pass_start[7] = {0, 4, 0, 2, 0, 1, 0};
-static const int FARDATA png_pass_inc[7]   = {8, 8, 4, 4, 2, 2, 1};
-static const int FARDATA png_pass_width[7] = {8, 4, 4, 2, 2, 1, 1};
-#endif
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED)
-/* djgpp, Win32, and Cygwin add their own underscores to global variables,
- * so define them without: */
-#if defined(__DJGPP__) || defined(WIN32) || defined(__CYGWIN__)
-#  define _mmx_supported  mmx_supported
-#  define _const4         const4
-#  define _const6         const6
-#  define _mask8_0        mask8_0
-#  define _mask16_1       mask16_1
-#  define _mask16_0       mask16_0
-#  define _mask24_2       mask24_2
-#  define _mask24_1       mask24_1
-#  define _mask24_0       mask24_0
-#  define _mask32_3       mask32_3
-#  define _mask32_2       mask32_2
-#  define _mask32_1       mask32_1
-#  define _mask32_0       mask32_0
-#  define _mask48_5       mask48_5
-#  define _mask48_4       mask48_4
-#  define _mask48_3       mask48_3
-#  define _mask48_2       mask48_2
-#  define _mask48_1       mask48_1
-#  define _mask48_0       mask48_0
-#  define _LBCarryMask    LBCarryMask
-#  define _HBClearMask    HBClearMask
-#  define _ActiveMask     ActiveMask
-#  define _ActiveMask2    ActiveMask2
-#  define _ActiveMaskEnd  ActiveMaskEnd
-#  define _ShiftBpp       ShiftBpp
-#  define _ShiftRem       ShiftRem
-#ifdef PNG_THREAD_UNSAFE_OK
-#  define _unmask         unmask
-#  define _FullLength     FullLength
-#  define _MMXLength      MMXLength
-#  define _dif            dif
-#  define _patemp         patemp
-#  define _pbtemp         pbtemp
-#  define _pctemp         pctemp
-#endif
-#endif
-
-
-/* These constants are used in the inlined MMX assembly code.
-   Ignore gcc's "At top level: defined but not used" warnings. */
-
-/* GRR 20000706:  originally _unmask was needed only when compiling with -fPIC,
- *  since that case uses the %ebx register for indexing the Global Offset Table
- *  and there were no other registers available.  But gcc 2.95 and later emit
- *  "more than 10 operands in `asm'" errors when %ebx is used to preload unmask
- *  in the non-PIC case, so we'll just use the global unconditionally now.
- */
-#ifdef PNG_THREAD_UNSAFE_OK
-static int _unmask;
-#endif
-
-static unsigned long long _mask8_0  = 0x0102040810204080LL;
-
-static unsigned long long _mask16_1 = 0x0101020204040808LL;
-static unsigned long long _mask16_0 = 0x1010202040408080LL;
-
-static unsigned long long _mask24_2 = 0x0101010202020404LL;
-static unsigned long long _mask24_1 = 0x0408080810101020LL;
-static unsigned long long _mask24_0 = 0x2020404040808080LL;
-
-static unsigned long long _mask32_3 = 0x0101010102020202LL;
-static unsigned long long _mask32_2 = 0x0404040408080808LL;
-static unsigned long long _mask32_1 = 0x1010101020202020LL;
-static unsigned long long _mask32_0 = 0x4040404080808080LL;
-
-static unsigned long long _mask48_5 = 0x0101010101010202LL;
-static unsigned long long _mask48_4 = 0x0202020204040404LL;
-static unsigned long long _mask48_3 = 0x0404080808080808LL;
-static unsigned long long _mask48_2 = 0x1010101010102020LL;
-static unsigned long long _mask48_1 = 0x2020202040404040LL;
-static unsigned long long _mask48_0 = 0x4040808080808080LL;
-
-static unsigned long long _const4   = 0x0000000000FFFFFFLL;
-//static unsigned long long _const5 = 0x000000FFFFFF0000LL;     // NOT USED
-static unsigned long long _const6   = 0x00000000000000FFLL;
-
-// These are used in the row-filter routines and should/would be local
-//  variables if not for gcc addressing limitations.
-// WARNING: Their presence probably defeats the thread safety of libpng.
-
-#ifdef PNG_THREAD_UNSAFE_OK
-static png_uint_32  _FullLength;
-static png_uint_32  _MMXLength;
-static int          _dif;
-static int          _patemp; // temp variables for Paeth routine
-static int          _pbtemp;
-static int          _pctemp;
-#endif
-
-void /* PRIVATE */
-png_squelch_warnings(void)
-{
-#ifdef PNG_THREAD_UNSAFE_OK
-   _dif = _dif;
-   _patemp = _patemp;
-   _pbtemp = _pbtemp;
-   _pctemp = _pctemp;
-   _MMXLength = _MMXLength;
-#endif
-   _const4  = _const4;
-   _const6  = _const6;
-   _mask8_0  = _mask8_0;
-   _mask16_1 = _mask16_1;
-   _mask16_0 = _mask16_0;
-   _mask24_2 = _mask24_2;
-   _mask24_1 = _mask24_1;
-   _mask24_0 = _mask24_0;
-   _mask32_3 = _mask32_3;
-   _mask32_2 = _mask32_2;
-   _mask32_1 = _mask32_1;
-   _mask32_0 = _mask32_0;
-   _mask48_5 = _mask48_5;
-   _mask48_4 = _mask48_4;
-   _mask48_3 = _mask48_3;
-   _mask48_2 = _mask48_2;
-   _mask48_1 = _mask48_1;
-   _mask48_0 = _mask48_0;
-}
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-
-
-static int _mmx_supported = 2;
-
-/*===========================================================================*/
-/*                                                                           */
-/*                       P N G _ C O M B I N E _ R O W                       */
-/*                                                                           */
-/*===========================================================================*/
-
-#if defined(PNG_HAVE_ASSEMBLER_COMBINE_ROW)
-
-#define BPP2  2
-#define BPP3  3 /* bytes per pixel (a.k.a. pixel_bytes) */
-#define BPP4  4
-#define BPP6  6 /* (defined only to help avoid cut-and-paste errors) */
-#define BPP8  8
-
-/* Combines the row recently read in with the previous row.
-   This routine takes care of alpha and transparency if requested.
-   This routine also handles the two methods of progressive display
-   of interlaced images, depending on the mask value.
-   The mask value describes which pixels are to be combined with
-   the row.  The pattern always repeats every 8 pixels, so just 8
-   bits are needed.  A one indicates the pixel is to be combined; a
-   zero indicates the pixel is to be skipped.  This is in addition
-   to any alpha or transparency value associated with the pixel.
-   If you want all pixels to be combined, pass 0xff (255) in mask. */
-
-/* Use this routine for the x86 platform - it uses a faster MMX routine
-   if the machine supports MMX. */
-
-void /* PRIVATE */
-png_combine_row(png_structp png_ptr, png_bytep row, int mask)
-{
-   png_debug(1, "in png_combine_row (pnggccrd.c)\n");
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED)
-   if (_mmx_supported == 2) {
-       /* this should have happened in png_init_mmx_flags() already */
-       png_warning(png_ptr, "asm_flags may not have been initialized");
-       png_mmx_support();
-   }
-#endif
-
-   if (mask == 0xff)
-   {
-      png_debug(2,"mask == 0xff:  doing single png_memcpy()\n");
-      png_memcpy(row, png_ptr->row_buf + 1,
-       (png_size_t)((png_ptr->width * png_ptr->row_info.pixel_depth + 7) >> 3));
-   }
-   else   /* (png_combine_row() is never called with mask == 0) */
-   {
-      switch (png_ptr->row_info.pixel_depth)
-      {
-         case 1:        /* png_ptr->row_info.pixel_depth */
-         {
-            png_bytep sp;
-            png_bytep dp;
-            int s_inc, s_start, s_end;
-            int m;
-            int shift;
-            png_uint_32 i;
-
-            sp = png_ptr->row_buf + 1;
-            dp = row;
-            m = 0x80;
-#if defined(PNG_READ_PACKSWAP_SUPPORTED)
-            if (png_ptr->transformations & PNG_PACKSWAP)
-            {
-                s_start = 0;
-                s_end = 7;
-                s_inc = 1;
-            }
-            else
-#endif
-            {
-                s_start = 7;
-                s_end = 0;
-                s_inc = -1;
-            }
-
-            shift = s_start;
-
-            for (i = 0; i < png_ptr->width; i++)
-            {
-               if (m & mask)
-               {
-                  int value;
-
-                  value = (*sp >> shift) & 0x1;
-                  *dp &= (png_byte)((0x7f7f >> (7 - shift)) & 0xff);
-                  *dp |= (png_byte)(value << shift);
-               }
-
-               if (shift == s_end)
-               {
-                  shift = s_start;
-                  sp++;
-                  dp++;
-               }
-               else
-                  shift += s_inc;
-
-               if (m == 1)
-                  m = 0x80;
-               else
-                  m >>= 1;
-            }
-            break;
-         }
-
-         case 2:        /* png_ptr->row_info.pixel_depth */
-         {
-            png_bytep sp;
-            png_bytep dp;
-            int s_start, s_end, s_inc;
-            int m;
-            int shift;
-            png_uint_32 i;
-            int value;
-
-            sp = png_ptr->row_buf + 1;
-            dp = row;
-            m = 0x80;
-#if defined(PNG_READ_PACKSWAP_SUPPORTED)
-            if (png_ptr->transformations & PNG_PACKSWAP)
-            {
-               s_start = 0;
-               s_end = 6;
-               s_inc = 2;
-            }
-            else
-#endif
-            {
-               s_start = 6;
-               s_end = 0;
-               s_inc = -2;
-            }
-
-            shift = s_start;
-
-            for (i = 0; i < png_ptr->width; i++)
-            {
-               if (m & mask)
-               {
-                  value = (*sp >> shift) & 0x3;
-                  *dp &= (png_byte)((0x3f3f >> (6 - shift)) & 0xff);
-                  *dp |= (png_byte)(value << shift);
-               }
-
-               if (shift == s_end)
-               {
-                  shift = s_start;
-                  sp++;
-                  dp++;
-               }
-               else
-                  shift += s_inc;
-               if (m == 1)
-                  m = 0x80;
-               else
-                  m >>= 1;
-            }
-            break;
-         }
-
-         case 4:        /* png_ptr->row_info.pixel_depth */
-         {
-            png_bytep sp;
-            png_bytep dp;
-            int s_start, s_end, s_inc;
-            int m;
-            int shift;
-            png_uint_32 i;
-            int value;
-
-            sp = png_ptr->row_buf + 1;
-            dp = row;
-            m = 0x80;
-#if defined(PNG_READ_PACKSWAP_SUPPORTED)
-            if (png_ptr->transformations & PNG_PACKSWAP)
-            {
-               s_start = 0;
-               s_end = 4;
-               s_inc = 4;
-            }
-            else
-#endif
-            {
-               s_start = 4;
-               s_end = 0;
-               s_inc = -4;
-            }
-            shift = s_start;
-
-            for (i = 0; i < png_ptr->width; i++)
-            {
-               if (m & mask)
-               {
-                  value = (*sp >> shift) & 0xf;
-                  *dp &= (png_byte)((0xf0f >> (4 - shift)) & 0xff);
-                  *dp |= (png_byte)(value << shift);
-               }
-
-               if (shift == s_end)
-               {
-                  shift = s_start;
-                  sp++;
-                  dp++;
-               }
-               else
-                  shift += s_inc;
-               if (m == 1)
-                  m = 0x80;
-               else
-                  m >>= 1;
-            }
-            break;
-         }
-
-         case 8:        /* png_ptr->row_info.pixel_depth */
-         {
-            png_bytep srcptr;
-            png_bytep dstptr;
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-            if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
-                /* && _mmx_supported */ )
-#else
-            if (_mmx_supported)
-#endif
-            {
-               png_uint_32 len;
-               int diff;
-               int dummy_value_a;   // fix 'forbidden register spilled' error
-               int dummy_value_d;
-               int dummy_value_c;
-               int dummy_value_S;
-               int dummy_value_D;
-               _unmask = ~mask;            // global variable for -fPIC version
-               srcptr = png_ptr->row_buf + 1;
-               dstptr = row;
-               len  = png_ptr->width &~7;  // reduce to multiple of 8
-               diff = (int) (png_ptr->width & 7);  // amount lost
-
-               __asm__ __volatile__ (
-                  "movd      _unmask, %%mm7  \n\t" // load bit pattern
-                  "psubb     %%mm6, %%mm6    \n\t" // zero mm6
-                  "punpcklbw %%mm7, %%mm7    \n\t"
-                  "punpcklwd %%mm7, %%mm7    \n\t"
-                  "punpckldq %%mm7, %%mm7    \n\t" // fill reg with 8 masks
-
-                  "movq      _mask8_0, %%mm0 \n\t"
-                  "pand      %%mm7, %%mm0    \n\t" // nonzero if keep byte
-                  "pcmpeqb   %%mm6, %%mm0    \n\t" // zeros->1s, v versa
-
-// preload        "movl      len, %%ecx      \n\t" // load length of line
-// preload        "movl      srcptr, %%esi   \n\t" // load source
-// preload        "movl      dstptr, %%edi   \n\t" // load dest
-
-                  "cmpl      $0, %%ecx       \n\t" // len == 0 ?
-                  "je        mainloop8end    \n\t"
-
-                "mainloop8:                  \n\t"
-                  "movq      (%%esi), %%mm4  \n\t" // *srcptr
-                  "pand      %%mm0, %%mm4    \n\t"
-                  "movq      %%mm0, %%mm6    \n\t"
-                  "pandn     (%%edi), %%mm6  \n\t" // *dstptr
-                  "por       %%mm6, %%mm4    \n\t"
-                  "movq      %%mm4, (%%edi)  \n\t"
-                  "addl      $8, %%esi       \n\t" // inc by 8 bytes processed
-                  "addl      $8, %%edi       \n\t"
-                  "subl      $8, %%ecx       \n\t" // dec by 8 pixels processed
-                  "ja        mainloop8       \n\t"
-
-                "mainloop8end:               \n\t"
-// preload        "movl      diff, %%ecx     \n\t" // (diff is in eax)
-                  "movl      %%eax, %%ecx    \n\t"
-                  "cmpl      $0, %%ecx       \n\t"
-                  "jz        end8            \n\t"
-// preload        "movl      mask, %%edx     \n\t"
-                  "sall      $24, %%edx      \n\t" // make low byte, high byte
-
-                "secondloop8:                \n\t"
-                  "sall      %%edx           \n\t" // move high bit to CF
-                  "jnc       skip8           \n\t" // if CF = 0
-                  "movb      (%%esi), %%al   \n\t"
-                  "movb      %%al, (%%edi)   \n\t"
-
-                "skip8:                      \n\t"
-                  "incl      %%esi           \n\t"
-                  "incl      %%edi           \n\t"
-                  "decl      %%ecx           \n\t"
-                  "jnz       secondloop8     \n\t"
-
-                "end8:                       \n\t"
-                  "EMMS                      \n\t"  // DONE
-
-                  : "=a" (dummy_value_a),           // output regs (dummy)
-                    "=d" (dummy_value_d),
-                    "=c" (dummy_value_c),
-                    "=S" (dummy_value_S),
-                    "=D" (dummy_value_D)
-
-                  : "3" (srcptr),      // esi       // input regs
-                    "4" (dstptr),      // edi
-                    "0" (diff),        // eax
-// was (unmask)     "b"    RESERVED    // ebx       // Global Offset Table idx
-                    "2" (len),         // ecx
-                    "1" (mask)         // edx
-
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-                  : "%mm0", "%mm4", "%mm6", "%mm7"  // clobber list
-#endif
-               );
-            }
-            else /* mmx _not supported - Use modified C routine */
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-            {
-               register png_uint_32 i;
-               png_uint_32 initial_val = png_pass_start[png_ptr->pass];
-                 /* png.c:  png_pass_start[] = {0, 4, 0, 2, 0, 1, 0}; */
-               register int stride = png_pass_inc[png_ptr->pass];
-                 /* png.c:  png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1}; */
-               register int rep_bytes = png_pass_width[png_ptr->pass];
-                 /* png.c:  png_pass_width[] = {8, 4, 4, 2, 2, 1, 1}; */
-               png_uint_32 len = png_ptr->width &~7;  /* reduce to mult. of 8 */
-               int diff = (int) (png_ptr->width & 7); /* amount lost */
-               register png_uint_32 final_val = len;  /* GRR bugfix */
-
-               srcptr = png_ptr->row_buf + 1 + initial_val;
-               dstptr = row + initial_val;
-
-               for (i = initial_val; i < final_val; i += stride)
-               {
-                  png_memcpy(dstptr, srcptr, rep_bytes);
-                  srcptr += stride;
-                  dstptr += stride;
-               }
-               if (diff)  /* number of leftover pixels:  3 for pngtest */
-               {
-                  final_val+=diff /* *BPP1 */ ;
-                  for (; i < final_val; i += stride)
-                  {
-                     if (rep_bytes > (int)(final_val-i))
-                        rep_bytes = (int)(final_val-i);
-                     png_memcpy(dstptr, srcptr, rep_bytes);
-                     srcptr += stride;
-                     dstptr += stride;
-                  }
-               }
-
-            } /* end of else (_mmx_supported) */
-
-            break;
-         }       /* end 8 bpp */
-
-         case 16:       /* png_ptr->row_info.pixel_depth */
-         {
-            png_bytep srcptr;
-            png_bytep dstptr;
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-            if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
-                /* && _mmx_supported */ )
-#else
-            if (_mmx_supported)
-#endif
-            {
-               png_uint_32 len;
-               int diff;
-               int dummy_value_a;   // fix 'forbidden register spilled' error
-               int dummy_value_d;
-               int dummy_value_c;
-               int dummy_value_S;
-               int dummy_value_D;
-               _unmask = ~mask;            // global variable for -fPIC version
-               srcptr = png_ptr->row_buf + 1;
-               dstptr = row;
-               len  = png_ptr->width &~7;  // reduce to multiple of 8
-               diff = (int) (png_ptr->width & 7); // amount lost //
-
-               __asm__ __volatile__ (
-                  "movd      _unmask, %%mm7   \n\t" // load bit pattern
-                  "psubb     %%mm6, %%mm6     \n\t" // zero mm6
-                  "punpcklbw %%mm7, %%mm7     \n\t"
-                  "punpcklwd %%mm7, %%mm7     \n\t"
-                  "punpckldq %%mm7, %%mm7     \n\t" // fill reg with 8 masks
-
-                  "movq      _mask16_0, %%mm0 \n\t"
-                  "movq      _mask16_1, %%mm1 \n\t"
-
-                  "pand      %%mm7, %%mm0     \n\t"
-                  "pand      %%mm7, %%mm1     \n\t"
-
-                  "pcmpeqb   %%mm6, %%mm0     \n\t"
-                  "pcmpeqb   %%mm6, %%mm1     \n\t"
-
-// preload        "movl      len, %%ecx       \n\t" // load length of line
-// preload        "movl      srcptr, %%esi    \n\t" // load source
-// preload        "movl      dstptr, %%edi    \n\t" // load dest
-
-                  "cmpl      $0, %%ecx        \n\t"
-                  "jz        mainloop16end    \n\t"
-
-                "mainloop16:                  \n\t"
-                  "movq      (%%esi), %%mm4   \n\t"
-                  "pand      %%mm0, %%mm4     \n\t"
-                  "movq      %%mm0, %%mm6     \n\t"
-                  "movq      (%%edi), %%mm7   \n\t"
-                  "pandn     %%mm7, %%mm6     \n\t"
-                  "por       %%mm6, %%mm4     \n\t"
-                  "movq      %%mm4, (%%edi)   \n\t"
-
-                  "movq      8(%%esi), %%mm5  \n\t"
-                  "pand      %%mm1, %%mm5     \n\t"
-                  "movq      %%mm1, %%mm7     \n\t"
-                  "movq      8(%%edi), %%mm6  \n\t"
-                  "pandn     %%mm6, %%mm7     \n\t"
-                  "por       %%mm7, %%mm5     \n\t"
-                  "movq      %%mm5, 8(%%edi)  \n\t"
-
-                  "addl      $16, %%esi       \n\t" // inc by 16 bytes processed
-                  "addl      $16, %%edi       \n\t"
-                  "subl      $8, %%ecx        \n\t" // dec by 8 pixels processed
-                  "ja        mainloop16       \n\t"
-
-                "mainloop16end:               \n\t"
-// preload        "movl      diff, %%ecx      \n\t" // (diff is in eax)
-                  "movl      %%eax, %%ecx     \n\t"
-                  "cmpl      $0, %%ecx        \n\t"
-                  "jz        end16            \n\t"
-// preload        "movl      mask, %%edx      \n\t"
-                  "sall      $24, %%edx       \n\t" // make low byte, high byte
-
-                "secondloop16:                \n\t"
-                  "sall      %%edx            \n\t" // move high bit to CF
-                  "jnc       skip16           \n\t" // if CF = 0
-                  "movw      (%%esi), %%ax    \n\t"
-                  "movw      %%ax, (%%edi)    \n\t"
-
-                "skip16:                      \n\t"
-                  "addl      $2, %%esi        \n\t"
-                  "addl      $2, %%edi        \n\t"
-                  "decl      %%ecx            \n\t"
-                  "jnz       secondloop16     \n\t"
-
-                "end16:                       \n\t"
-                  "EMMS                       \n\t" // DONE
-
-                  : "=a" (dummy_value_a),           // output regs (dummy)
-                    "=c" (dummy_value_c),
-                    "=d" (dummy_value_d),
-                    "=S" (dummy_value_S),
-                    "=D" (dummy_value_D)
-
-                  : "0" (diff),        // eax       // input regs
-// was (unmask)     " "    RESERVED    // ebx       // Global Offset Table idx
-                    "1" (len),         // ecx
-                    "2" (mask),        // edx
-                    "3" (srcptr),      // esi
-                    "4" (dstptr)       // edi
-
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-                  : "%mm0", "%mm1", "%mm4"          // clobber list
-                  , "%mm5", "%mm6", "%mm7"
-#endif
-               );
-            }
-            else /* mmx _not supported - Use modified C routine */
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-            {
-               register png_uint_32 i;
-               png_uint_32 initial_val = BPP2 * png_pass_start[png_ptr->pass];
-                 /* png.c:  png_pass_start[] = {0, 4, 0, 2, 0, 1, 0}; */
-               register int stride = BPP2 * png_pass_inc[png_ptr->pass];
-                 /* png.c:  png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1}; */
-               register int rep_bytes = BPP2 * png_pass_width[png_ptr->pass];
-                 /* png.c:  png_pass_width[] = {8, 4, 4, 2, 2, 1, 1}; */
-               png_uint_32 len = png_ptr->width &~7;  /* reduce to mult. of 8 */
-               int diff = (int) (png_ptr->width & 7); /* amount lost */
-               register png_uint_32 final_val = BPP2 * len;   /* GRR bugfix */
-
-               srcptr = png_ptr->row_buf + 1 + initial_val;
-               dstptr = row + initial_val;
-
-               for (i = initial_val; i < final_val; i += stride)
-               {
-                  png_memcpy(dstptr, srcptr, rep_bytes);
-                  srcptr += stride;
-                  dstptr += stride;
-               }
-               if (diff)  /* number of leftover pixels:  3 for pngtest */
-               {
-                  final_val+=diff*BPP2;
-                  for (; i < final_val; i += stride)
-                  {
-                     if (rep_bytes > (int)(final_val-i))
-                        rep_bytes = (int)(final_val-i);
-                     png_memcpy(dstptr, srcptr, rep_bytes);
-                     srcptr += stride;
-                     dstptr += stride;
-                  }
-               }
-            } /* end of else (_mmx_supported) */
-
-            break;
-         }       /* end 16 bpp */
-
-         case 24:       /* png_ptr->row_info.pixel_depth */
-         {
-            png_bytep srcptr;
-            png_bytep dstptr;
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-            if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
-                /* && _mmx_supported */ )
-#else
-            if (_mmx_supported)
-#endif
-            {
-               png_uint_32 len;
-               int diff;
-               int dummy_value_a;   // fix 'forbidden register spilled' error
-               int dummy_value_d;
-               int dummy_value_c;
-               int dummy_value_S;
-               int dummy_value_D;
-               _unmask = ~mask;            // global variable for -fPIC version
-               srcptr = png_ptr->row_buf + 1;
-               dstptr = row;
-               len  = png_ptr->width &~7;  // reduce to multiple of 8
-               diff = (int) (png_ptr->width & 7); // amount lost //
-
-               __asm__ __volatile__ (
-                  "movd      _unmask, %%mm7   \n\t" // load bit pattern
-                  "psubb     %%mm6, %%mm6     \n\t" // zero mm6
-                  "punpcklbw %%mm7, %%mm7     \n\t"
-                  "punpcklwd %%mm7, %%mm7     \n\t"
-                  "punpckldq %%mm7, %%mm7     \n\t" // fill reg with 8 masks
-
-                  "movq      _mask24_0, %%mm0 \n\t"
-                  "movq      _mask24_1, %%mm1 \n\t"
-                  "movq      _mask24_2, %%mm2 \n\t"
-
-                  "pand      %%mm7, %%mm0     \n\t"
-                  "pand      %%mm7, %%mm1     \n\t"
-                  "pand      %%mm7, %%mm2     \n\t"
-
-                  "pcmpeqb   %%mm6, %%mm0     \n\t"
-                  "pcmpeqb   %%mm6, %%mm1     \n\t"
-                  "pcmpeqb   %%mm6, %%mm2     \n\t"
-
-// preload        "movl      len, %%ecx       \n\t" // load length of line
-// preload        "movl      srcptr, %%esi    \n\t" // load source
-// preload        "movl      dstptr, %%edi    \n\t" // load dest
-
-                  "cmpl      $0, %%ecx        \n\t"
-                  "jz        mainloop24end    \n\t"
-
-                "mainloop24:                  \n\t"
-                  "movq      (%%esi), %%mm4   \n\t"
-                  "pand      %%mm0, %%mm4     \n\t"
-                  "movq      %%mm0, %%mm6     \n\t"
-                  "movq      (%%edi), %%mm7   \n\t"
-                  "pandn     %%mm7, %%mm6     \n\t"
-                  "por       %%mm6, %%mm4     \n\t"
-                  "movq      %%mm4, (%%edi)   \n\t"
-
-                  "movq      8(%%esi), %%mm5  \n\t"
-                  "pand      %%mm1, %%mm5     \n\t"
-                  "movq      %%mm1, %%mm7     \n\t"
-                  "movq      8(%%edi), %%mm6  \n\t"
-                  "pandn     %%mm6, %%mm7     \n\t"
-                  "por       %%mm7, %%mm5     \n\t"
-                  "movq      %%mm5, 8(%%edi)  \n\t"
-
-                  "movq      16(%%esi), %%mm6 \n\t"
-                  "pand      %%mm2, %%mm6     \n\t"
-                  "movq      %%mm2, %%mm4     \n\t"
-                  "movq      16(%%edi), %%mm7 \n\t"
-                  "pandn     %%mm7, %%mm4     \n\t"
-                  "por       %%mm4, %%mm6     \n\t"
-                  "movq      %%mm6, 16(%%edi) \n\t"
-
-                  "addl      $24, %%esi       \n\t" // inc by 24 bytes processed
-                  "addl      $24, %%edi       \n\t"
-                  "subl      $8, %%ecx        \n\t" // dec by 8 pixels processed
-
-                  "ja        mainloop24       \n\t"
-
-                "mainloop24end:               \n\t"
-// preload        "movl      diff, %%ecx      \n\t" // (diff is in eax)
-                  "movl      %%eax, %%ecx     \n\t"
-                  "cmpl      $0, %%ecx        \n\t"
-                  "jz        end24            \n\t"
-// preload        "movl      mask, %%edx      \n\t"
-                  "sall      $24, %%edx       \n\t" // make low byte, high byte
-
-                "secondloop24:                \n\t"
-                  "sall      %%edx            \n\t" // move high bit to CF
-                  "jnc       skip24           \n\t" // if CF = 0
-                  "movw      (%%esi), %%ax    \n\t"
-                  "movw      %%ax, (%%edi)    \n\t"
-                  "xorl      %%eax, %%eax     \n\t"
-                  "movb      2(%%esi), %%al   \n\t"
-                  "movb      %%al, 2(%%edi)   \n\t"
-
-                "skip24:                      \n\t"
-                  "addl      $3, %%esi        \n\t"
-                  "addl      $3, %%edi        \n\t"
-                  "decl      %%ecx            \n\t"
-                  "jnz       secondloop24     \n\t"
-
-                "end24:                       \n\t"
-                  "EMMS                       \n\t" // DONE
-
-                  : "=a" (dummy_value_a),           // output regs (dummy)
-                    "=d" (dummy_value_d),
-                    "=c" (dummy_value_c),
-                    "=S" (dummy_value_S),
-                    "=D" (dummy_value_D)
-
-                  : "3" (srcptr),      // esi       // input regs
-                    "4" (dstptr),      // edi
-                    "0" (diff),        // eax
-// was (unmask)     "b"    RESERVED    // ebx       // Global Offset Table idx
-                    "2" (len),         // ecx
-                    "1" (mask)         // edx
-
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-                  : "%mm0", "%mm1", "%mm2"          // clobber list
-                  , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-               );
-            }
-            else /* mmx _not supported - Use modified C routine */
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-            {
-               register png_uint_32 i;
-               png_uint_32 initial_val = BPP3 * png_pass_start[png_ptr->pass];
-                 /* png.c:  png_pass_start[] = {0, 4, 0, 2, 0, 1, 0}; */
-               register int stride = BPP3 * png_pass_inc[png_ptr->pass];
-                 /* png.c:  png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1}; */
-               register int rep_bytes = BPP3 * png_pass_width[png_ptr->pass];
-                 /* png.c:  png_pass_width[] = {8, 4, 4, 2, 2, 1, 1}; */
-               png_uint_32 len = png_ptr->width &~7;  /* reduce to mult. of 8 */
-               int diff = (int) (png_ptr->width & 7); /* amount lost */
-               register png_uint_32 final_val = BPP3 * len;   /* GRR bugfix */
-
-               srcptr = png_ptr->row_buf + 1 + initial_val;
-               dstptr = row + initial_val;
-
-               for (i = initial_val; i < final_val; i += stride)
-               {
-                  png_memcpy(dstptr, srcptr, rep_bytes);
-                  srcptr += stride;
-                  dstptr += stride;
-               }
-               if (diff)  /* number of leftover pixels:  3 for pngtest */
-               {
-                  final_val+=diff*BPP3;
-                  for (; i < final_val; i += stride)
-                  {
-                     if (rep_bytes > (int)(final_val-i))
-                        rep_bytes = (int)(final_val-i);
-                     png_memcpy(dstptr, srcptr, rep_bytes);
-                     srcptr += stride;
-                     dstptr += stride;
-                  }
-               }
-            } /* end of else (_mmx_supported) */
-
-            break;
-         }       /* end 24 bpp */
-
-         case 32:       /* png_ptr->row_info.pixel_depth */
-         {
-            png_bytep srcptr;
-            png_bytep dstptr;
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-            if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
-                /* && _mmx_supported */ )
-#else
-            if (_mmx_supported)
-#endif
-            {
-               png_uint_32 len;
-               int diff;
-               int dummy_value_a;   // fix 'forbidden register spilled' error
-               int dummy_value_d;
-               int dummy_value_c;
-               int dummy_value_S;
-               int dummy_value_D;
-               _unmask = ~mask;            // global variable for -fPIC version
-               srcptr = png_ptr->row_buf + 1;
-               dstptr = row;
-               len  = png_ptr->width &~7;  // reduce to multiple of 8
-               diff = (int) (png_ptr->width & 7); // amount lost //
-
-               __asm__ __volatile__ (
-                  "movd      _unmask, %%mm7   \n\t" // load bit pattern
-                  "psubb     %%mm6, %%mm6     \n\t" // zero mm6
-                  "punpcklbw %%mm7, %%mm7     \n\t"
-                  "punpcklwd %%mm7, %%mm7     \n\t"
-                  "punpckldq %%mm7, %%mm7     \n\t" // fill reg with 8 masks
-
-                  "movq      _mask32_0, %%mm0 \n\t"
-                  "movq      _mask32_1, %%mm1 \n\t"
-                  "movq      _mask32_2, %%mm2 \n\t"
-                  "movq      _mask32_3, %%mm3 \n\t"
-
-                  "pand      %%mm7, %%mm0     \n\t"
-                  "pand      %%mm7, %%mm1     \n\t"
-                  "pand      %%mm7, %%mm2     \n\t"
-                  "pand      %%mm7, %%mm3     \n\t"
-
-                  "pcmpeqb   %%mm6, %%mm0     \n\t"
-                  "pcmpeqb   %%mm6, %%mm1     \n\t"
-                  "pcmpeqb   %%mm6, %%mm2     \n\t"
-                  "pcmpeqb   %%mm6, %%mm3     \n\t"
-
-// preload        "movl      len, %%ecx       \n\t" // load length of line
-// preload        "movl      srcptr, %%esi    \n\t" // load source
-// preload        "movl      dstptr, %%edi    \n\t" // load dest
-
-                  "cmpl      $0, %%ecx        \n\t" // lcr
-                  "jz        mainloop32end    \n\t"
-
-                "mainloop32:                  \n\t"
-                  "movq      (%%esi), %%mm4   \n\t"
-                  "pand      %%mm0, %%mm4     \n\t"
-                  "movq      %%mm0, %%mm6     \n\t"
-                  "movq      (%%edi), %%mm7   \n\t"
-                  "pandn     %%mm7, %%mm6     \n\t"
-                  "por       %%mm6, %%mm4     \n\t"
-                  "movq      %%mm4, (%%edi)   \n\t"
-
-                  "movq      8(%%esi), %%mm5  \n\t"
-                  "pand      %%mm1, %%mm5     \n\t"
-                  "movq      %%mm1, %%mm7     \n\t"
-                  "movq      8(%%edi), %%mm6  \n\t"
-                  "pandn     %%mm6, %%mm7     \n\t"
-                  "por       %%mm7, %%mm5     \n\t"
-                  "movq      %%mm5, 8(%%edi)  \n\t"
-
-                  "movq      16(%%esi), %%mm6 \n\t"
-                  "pand      %%mm2, %%mm6     \n\t"
-                  "movq      %%mm2, %%mm4     \n\t"
-                  "movq      16(%%edi), %%mm7 \n\t"
-                  "pandn     %%mm7, %%mm4     \n\t"
-                  "por       %%mm4, %%mm6     \n\t"
-                  "movq      %%mm6, 16(%%edi) \n\t"
-
-                  "movq      24(%%esi), %%mm7 \n\t"
-                  "pand      %%mm3, %%mm7     \n\t"
-                  "movq      %%mm3, %%mm5     \n\t"
-                  "movq      24(%%edi), %%mm4 \n\t"
-                  "pandn     %%mm4, %%mm5     \n\t"
-                  "por       %%mm5, %%mm7     \n\t"
-                  "movq      %%mm7, 24(%%edi) \n\t"
-
-                  "addl      $32, %%esi       \n\t" // inc by 32 bytes processed
-                  "addl      $32, %%edi       \n\t"
-                  "subl      $8, %%ecx        \n\t" // dec by 8 pixels processed
-                  "ja        mainloop32       \n\t"
-
-                "mainloop32end:               \n\t"
-// preload        "movl      diff, %%ecx      \n\t" // (diff is in eax)
-                  "movl      %%eax, %%ecx     \n\t"
-                  "cmpl      $0, %%ecx        \n\t"
-                  "jz        end32            \n\t"
-// preload        "movl      mask, %%edx      \n\t"
-                  "sall      $24, %%edx       \n\t" // low byte => high byte
-
-                "secondloop32:                \n\t"
-                  "sall      %%edx            \n\t" // move high bit to CF
-                  "jnc       skip32           \n\t" // if CF = 0
-                  "movl      (%%esi), %%eax   \n\t"
-                  "movl      %%eax, (%%edi)   \n\t"
-
-                "skip32:                      \n\t"
-                  "addl      $4, %%esi        \n\t"
-                  "addl      $4, %%edi        \n\t"
-                  "decl      %%ecx            \n\t"
-                  "jnz       secondloop32     \n\t"
-
-                "end32:                       \n\t"
-                  "EMMS                       \n\t" // DONE
-
-                  : "=a" (dummy_value_a),           // output regs (dummy)
-                    "=d" (dummy_value_d),
-                    "=c" (dummy_value_c),
-                    "=S" (dummy_value_S),
-                    "=D" (dummy_value_D)
-
-                  : "3" (srcptr),      // esi       // input regs
-                    "4" (dstptr),      // edi
-                    "0" (diff),        // eax
-// was (unmask)     "b"    RESERVED    // ebx       // Global Offset Table idx
-                    "2" (len),         // ecx
-                    "1" (mask)         // edx
-
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-                  : "%mm0", "%mm1", "%mm2", "%mm3"  // clobber list
-                  , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-               );
-            }
-            else /* mmx _not supported - Use modified C routine */
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-            {
-               register png_uint_32 i;
-               png_uint_32 initial_val = BPP4 * png_pass_start[png_ptr->pass];
-                 /* png.c:  png_pass_start[] = {0, 4, 0, 2, 0, 1, 0}; */
-               register int stride = BPP4 * png_pass_inc[png_ptr->pass];
-                 /* png.c:  png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1}; */
-               register int rep_bytes = BPP4 * png_pass_width[png_ptr->pass];
-                 /* png.c:  png_pass_width[] = {8, 4, 4, 2, 2, 1, 1}; */
-               png_uint_32 len = png_ptr->width &~7;  /* reduce to mult. of 8 */
-               int diff = (int) (png_ptr->width & 7); /* amount lost */
-               register png_uint_32 final_val = BPP4 * len;   /* GRR bugfix */
-
-               srcptr = png_ptr->row_buf + 1 + initial_val;
-               dstptr = row + initial_val;
-
-               for (i = initial_val; i < final_val; i += stride)
-               {
-                  png_memcpy(dstptr, srcptr, rep_bytes);
-                  srcptr += stride;
-                  dstptr += stride;
-               }
-               if (diff)  /* number of leftover pixels:  3 for pngtest */
-               {
-                  final_val+=diff*BPP4;
-                  for (; i < final_val; i += stride)
-                  {
-                     if (rep_bytes > (int)(final_val-i))
-                        rep_bytes = (int)(final_val-i);
-                     png_memcpy(dstptr, srcptr, rep_bytes);
-                     srcptr += stride;
-                     dstptr += stride;
-                  }
-               }
-            } /* end of else (_mmx_supported) */
-
-            break;
-         }       /* end 32 bpp */
-
-         case 48:       /* png_ptr->row_info.pixel_depth */
-         {
-            png_bytep srcptr;
-            png_bytep dstptr;
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-            if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
-                /* && _mmx_supported */ )
-#else
-            if (_mmx_supported)
-#endif
-            {
-               png_uint_32 len;
-               int diff;
-               int dummy_value_a;   // fix 'forbidden register spilled' error
-               int dummy_value_d;
-               int dummy_value_c;
-               int dummy_value_S;
-               int dummy_value_D;
-               _unmask = ~mask;            // global variable for -fPIC version
-               srcptr = png_ptr->row_buf + 1;
-               dstptr = row;
-               len  = png_ptr->width &~7;  // reduce to multiple of 8
-               diff = (int) (png_ptr->width & 7); // amount lost //
-
-               __asm__ __volatile__ (
-                  "movd      _unmask, %%mm7   \n\t" // load bit pattern
-                  "psubb     %%mm6, %%mm6     \n\t" // zero mm6
-                  "punpcklbw %%mm7, %%mm7     \n\t"
-                  "punpcklwd %%mm7, %%mm7     \n\t"
-                  "punpckldq %%mm7, %%mm7     \n\t" // fill reg with 8 masks
-
-                  "movq      _mask48_0, %%mm0 \n\t"
-                  "movq      _mask48_1, %%mm1 \n\t"
-                  "movq      _mask48_2, %%mm2 \n\t"
-                  "movq      _mask48_3, %%mm3 \n\t"
-                  "movq      _mask48_4, %%mm4 \n\t"
-                  "movq      _mask48_5, %%mm5 \n\t"
-
-                  "pand      %%mm7, %%mm0     \n\t"
-                  "pand      %%mm7, %%mm1     \n\t"
-                  "pand      %%mm7, %%mm2     \n\t"
-                  "pand      %%mm7, %%mm3     \n\t"
-                  "pand      %%mm7, %%mm4     \n\t"
-                  "pand      %%mm7, %%mm5     \n\t"
-
-                  "pcmpeqb   %%mm6, %%mm0     \n\t"
-                  "pcmpeqb   %%mm6, %%mm1     \n\t"
-                  "pcmpeqb   %%mm6, %%mm2     \n\t"
-                  "pcmpeqb   %%mm6, %%mm3     \n\t"
-                  "pcmpeqb   %%mm6, %%mm4     \n\t"
-                  "pcmpeqb   %%mm6, %%mm5     \n\t"
-
-// preload        "movl      len, %%ecx       \n\t" // load length of line
-// preload        "movl      srcptr, %%esi    \n\t" // load source
-// preload        "movl      dstptr, %%edi    \n\t" // load dest
-
-                  "cmpl      $0, %%ecx        \n\t"
-                  "jz        mainloop48end    \n\t"
-
-                "mainloop48:                  \n\t"
-                  "movq      (%%esi), %%mm7   \n\t"
-                  "pand      %%mm0, %%mm7     \n\t"
-                  "movq      %%mm0, %%mm6     \n\t"
-                  "pandn     (%%edi), %%mm6   \n\t"
-                  "por       %%mm6, %%mm7     \n\t"
-                  "movq      %%mm7, (%%edi)   \n\t"
-
-                  "movq      8(%%esi), %%mm6  \n\t"
-                  "pand      %%mm1, %%mm6     \n\t"
-                  "movq      %%mm1, %%mm7     \n\t"
-                  "pandn     8(%%edi), %%mm7  \n\t"
-                  "por       %%mm7, %%mm6     \n\t"
-                  "movq      %%mm6, 8(%%edi)  \n\t"
-
-                  "movq      16(%%esi), %%mm6 \n\t"
-                  "pand      %%mm2, %%mm6     \n\t"
-                  "movq      %%mm2, %%mm7     \n\t"
-                  "pandn     16(%%edi), %%mm7 \n\t"
-                  "por       %%mm7, %%mm6     \n\t"
-                  "movq      %%mm6, 16(%%edi) \n\t"
-
-                  "movq      24(%%esi), %%mm7 \n\t"
-                  "pand      %%mm3, %%mm7     \n\t"
-                  "movq      %%mm3, %%mm6     \n\t"
-                  "pandn     24(%%edi), %%mm6 \n\t"
-                  "por       %%mm6, %%mm7     \n\t"
-                  "movq      %%mm7, 24(%%edi) \n\t"
-
-                  "movq      32(%%esi), %%mm6 \n\t"
-                  "pand      %%mm4, %%mm6     \n\t"
-                  "movq      %%mm4, %%mm7     \n\t"
-                  "pandn     32(%%edi), %%mm7 \n\t"
-                  "por       %%mm7, %%mm6     \n\t"
-                  "movq      %%mm6, 32(%%edi) \n\t"
-
-                  "movq      40(%%esi), %%mm7 \n\t"
-                  "pand      %%mm5, %%mm7     \n\t"
-                  "movq      %%mm5, %%mm6     \n\t"
-                  "pandn     40(%%edi), %%mm6 \n\t"
-                  "por       %%mm6, %%mm7     \n\t"
-                  "movq      %%mm7, 40(%%edi) \n\t"
-
-                  "addl      $48, %%esi       \n\t" // inc by 48 bytes processed
-                  "addl      $48, %%edi       \n\t"
-                  "subl      $8, %%ecx        \n\t" // dec by 8 pixels processed
-
-                  "ja        mainloop48       \n\t"
-
-                "mainloop48end:               \n\t"
-// preload        "movl      diff, %%ecx      \n\t" // (diff is in eax)
-                  "movl      %%eax, %%ecx     \n\t"
-                  "cmpl      $0, %%ecx        \n\t"
-                  "jz        end48            \n\t"
-// preload        "movl      mask, %%edx      \n\t"
-                  "sall      $24, %%edx       \n\t" // make low byte, high byte
-
-                "secondloop48:                \n\t"
-                  "sall      %%edx            \n\t" // move high bit to CF
-                  "jnc       skip48           \n\t" // if CF = 0
-                  "movl      (%%esi), %%eax   \n\t"
-                  "movl      %%eax, (%%edi)   \n\t"
-
-                "skip48:                      \n\t"
-                  "addl      $4, %%esi        \n\t"
-                  "addl      $4, %%edi        \n\t"
-                  "decl      %%ecx            \n\t"
-                  "jnz       secondloop48     \n\t"
-
-                "end48:                       \n\t"
-                  "EMMS                       \n\t" // DONE
-
-                  : "=a" (dummy_value_a),           // output regs (dummy)
-                    "=d" (dummy_value_d),
-                    "=c" (dummy_value_c),
-                    "=S" (dummy_value_S),
-                    "=D" (dummy_value_D)
-
-                  : "3" (srcptr),      // esi       // input regs
-                    "4" (dstptr),      // edi
-                    "0" (diff),        // eax
-// was (unmask)     "b"    RESERVED    // ebx       // Global Offset Table idx
-                    "2" (len),         // ecx
-                    "1" (mask)         // edx
-
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-                  : "%mm0", "%mm1", "%mm2", "%mm3"  // clobber list
-                  , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-               );
-            }
-            else /* mmx _not supported - Use modified C routine */
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-            {
-               register png_uint_32 i;
-               png_uint_32 initial_val = BPP6 * png_pass_start[png_ptr->pass];
-                 /* png.c:  png_pass_start[] = {0, 4, 0, 2, 0, 1, 0}; */
-               register int stride = BPP6 * png_pass_inc[png_ptr->pass];
-                 /* png.c:  png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1}; */
-               register int rep_bytes = BPP6 * png_pass_width[png_ptr->pass];
-                 /* png.c:  png_pass_width[] = {8, 4, 4, 2, 2, 1, 1}; */
-               png_uint_32 len = png_ptr->width &~7;  /* reduce to mult. of 8 */
-               int diff = (int) (png_ptr->width & 7); /* amount lost */
-               register png_uint_32 final_val = BPP6 * len;   /* GRR bugfix */
-
-               srcptr = png_ptr->row_buf + 1 + initial_val;
-               dstptr = row + initial_val;
-
-               for (i = initial_val; i < final_val; i += stride)
-               {
-                  png_memcpy(dstptr, srcptr, rep_bytes);
-                  srcptr += stride;
-                  dstptr += stride;
-               }
-               if (diff)  /* number of leftover pixels:  3 for pngtest */
-               {
-                  final_val+=diff*BPP6;
-                  for (; i < final_val; i += stride)
-                  {
-                     if (rep_bytes > (int)(final_val-i))
-                        rep_bytes = (int)(final_val-i);
-                     png_memcpy(dstptr, srcptr, rep_bytes);
-                     srcptr += stride;
-                     dstptr += stride;
-                  }
-               }
-            } /* end of else (_mmx_supported) */
-
-            break;
-         }       /* end 48 bpp */
-
-         case 64:       /* png_ptr->row_info.pixel_depth */
-         {
-            png_bytep srcptr;
-            png_bytep dstptr;
-            register png_uint_32 i;
-            png_uint_32 initial_val = BPP8 * png_pass_start[png_ptr->pass];
-              /* png.c:  png_pass_start[] = {0, 4, 0, 2, 0, 1, 0}; */
-            register int stride = BPP8 * png_pass_inc[png_ptr->pass];
-              /* png.c:  png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1}; */
-            register int rep_bytes = BPP8 * png_pass_width[png_ptr->pass];
-              /* png.c:  png_pass_width[] = {8, 4, 4, 2, 2, 1, 1}; */
-            png_uint_32 len = png_ptr->width &~7;  /* reduce to mult. of 8 */
-            int diff = (int) (png_ptr->width & 7); /* amount lost */
-            register png_uint_32 final_val = BPP8 * len;   /* GRR bugfix */
-
-            srcptr = png_ptr->row_buf + 1 + initial_val;
-            dstptr = row + initial_val;
-
-            for (i = initial_val; i < final_val; i += stride)
-            {
-               png_memcpy(dstptr, srcptr, rep_bytes);
-               srcptr += stride;
-               dstptr += stride;
-            }
-            if (diff)  /* number of leftover pixels:  3 for pngtest */
-            {
-               final_val+=diff*BPP8;
-               for (; i < final_val; i += stride)
-               {
-                  if (rep_bytes > (int)(final_val-i))
-                     rep_bytes = (int)(final_val-i);
-                  png_memcpy(dstptr, srcptr, rep_bytes);
-                  srcptr += stride;
-                  dstptr += stride;
-               }
-            }
-
-            break;
-         }       /* end 64 bpp */
-
-         default: /* png_ptr->row_info.pixel_depth != 1,2,4,8,16,24,32,48,64 */
-         {
-            /* this should never happen */
-            png_warning(png_ptr, "Invalid row_info.pixel_depth in pnggccrd");
-            break;
-         }
-      } /* end switch (png_ptr->row_info.pixel_depth) */
-
-   } /* end if (non-trivial mask) */
-
-} /* end png_combine_row() */
-
-#endif /* PNG_HAVE_ASSEMBLER_COMBINE_ROW */
-
-
-
-
-/*===========================================================================*/
-/*                                                                           */
-/*                 P N G _ D O _ R E A D _ I N T E R L A C E                 */
-/*                                                                           */
-/*===========================================================================*/
-
-#if defined(PNG_READ_INTERLACING_SUPPORTED)
-#if defined(PNG_HAVE_ASSEMBLER_READ_INTERLACE)
-
-/* png_do_read_interlace() is called after any 16-bit to 8-bit conversion
- * has taken place.  [GRR: what other steps come before and/or after?]
- */
-
-void /* PRIVATE */
-png_do_read_interlace(png_structp png_ptr)
-{
-   png_row_infop row_info = &(png_ptr->row_info);
-   png_bytep row = png_ptr->row_buf + 1;
-   int pass = png_ptr->pass;
-#if defined(PNG_READ_PACKSWAP_SUPPORTED)
-   png_uint_32 transformations = png_ptr->transformations;
-#endif
-
-   png_debug(1, "in png_do_read_interlace (pnggccrd.c)\n");
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED)
-   if (_mmx_supported == 2) {
-#if !defined(PNG_1_0_X)
-       /* this should have happened in png_init_mmx_flags() already */
-       png_warning(png_ptr, "asm_flags may not have been initialized");
-#endif
-       png_mmx_support();
-   }
-#endif
-
-   if (row != NULL && row_info != NULL)
-   {
-      png_uint_32 final_width;
-
-      final_width = row_info->width * png_pass_inc[pass];
-
-      switch (row_info->pixel_depth)
-      {
-         case 1:
-         {
-            png_bytep sp, dp;
-            int sshift, dshift;
-            int s_start, s_end, s_inc;
-            png_byte v;
-            png_uint_32 i;
-            int j;
-
-            sp = row + (png_size_t)((row_info->width - 1) >> 3);
-            dp = row + (png_size_t)((final_width - 1) >> 3);
-#if defined(PNG_READ_PACKSWAP_SUPPORTED)
-            if (transformations & PNG_PACKSWAP)
-            {
-               sshift = (int)((row_info->width + 7) & 7);
-               dshift = (int)((final_width + 7) & 7);
-               s_start = 7;
-               s_end = 0;
-               s_inc = -1;
-            }
-            else
-#endif
-            {
-               sshift = 7 - (int)((row_info->width + 7) & 7);
-               dshift = 7 - (int)((final_width + 7) & 7);
-               s_start = 0;
-               s_end = 7;
-               s_inc = 1;
-            }
-
-            for (i = row_info->width; i; i--)
-            {
-               v = (png_byte)((*sp >> sshift) & 0x1);
-               for (j = 0; j < png_pass_inc[pass]; j++)
-               {
-                  *dp &= (png_byte)((0x7f7f >> (7 - dshift)) & 0xff);
-                  *dp |= (png_byte)(v << dshift);
-                  if (dshift == s_end)
-                  {
-                     dshift = s_start;
-                     dp--;
-                  }
-                  else
-                     dshift += s_inc;
-               }
-               if (sshift == s_end)
-               {
-                  sshift = s_start;
-                  sp--;
-               }
-               else
-                  sshift += s_inc;
-            }
-            break;
-         }
-
-         case 2:
-         {
-            png_bytep sp, dp;
-            int sshift, dshift;
-            int s_start, s_end, s_inc;
-            png_uint_32 i;
-
-            sp = row + (png_size_t)((row_info->width - 1) >> 2);
-            dp = row + (png_size_t)((final_width - 1) >> 2);
-#if defined(PNG_READ_PACKSWAP_SUPPORTED)
-            if (transformations & PNG_PACKSWAP)
-            {
-               sshift = (png_size_t)(((row_info->width + 3) & 3) << 1);
-               dshift = (png_size_t)(((final_width + 3) & 3) << 1);
-               s_start = 6;
-               s_end = 0;
-               s_inc = -2;
-            }
-            else
-#endif
-            {
-               sshift = (png_size_t)((3 - ((row_info->width + 3) & 3)) << 1);
-               dshift = (png_size_t)((3 - ((final_width + 3) & 3)) << 1);
-               s_start = 0;
-               s_end = 6;
-               s_inc = 2;
-            }
-
-            for (i = row_info->width; i; i--)
-            {
-               png_byte v;
-               int j;
-
-               v = (png_byte)((*sp >> sshift) & 0x3);
-               for (j = 0; j < png_pass_inc[pass]; j++)
-               {
-                  *dp &= (png_byte)((0x3f3f >> (6 - dshift)) & 0xff);
-                  *dp |= (png_byte)(v << dshift);
-                  if (dshift == s_end)
-                  {
-                     dshift = s_start;
-                     dp--;
-                  }
-                  else
-                     dshift += s_inc;
-               }
-               if (sshift == s_end)
-               {
-                  sshift = s_start;
-                  sp--;
-               }
-               else
-                  sshift += s_inc;
-            }
-            break;
-         }
-
-         case 4:
-         {
-            png_bytep sp, dp;
-            int sshift, dshift;
-            int s_start, s_end, s_inc;
-            png_uint_32 i;
-
-            sp = row + (png_size_t)((row_info->width - 1) >> 1);
-            dp = row + (png_size_t)((final_width - 1) >> 1);
-#if defined(PNG_READ_PACKSWAP_SUPPORTED)
-            if (transformations & PNG_PACKSWAP)
-            {
-               sshift = (png_size_t)(((row_info->width + 1) & 1) << 2);
-               dshift = (png_size_t)(((final_width + 1) & 1) << 2);
-               s_start = 4;
-               s_end = 0;
-               s_inc = -4;
-            }
-            else
-#endif
-            {
-               sshift = (png_size_t)((1 - ((row_info->width + 1) & 1)) << 2);
-               dshift = (png_size_t)((1 - ((final_width + 1) & 1)) << 2);
-               s_start = 0;
-               s_end = 4;
-               s_inc = 4;
-            }
-
-            for (i = row_info->width; i; i--)
-            {
-               png_byte v;
-               int j;
-
-               v = (png_byte)((*sp >> sshift) & 0xf);
-               for (j = 0; j < png_pass_inc[pass]; j++)
-               {
-                  *dp &= (png_byte)((0xf0f >> (4 - dshift)) & 0xff);
-                  *dp |= (png_byte)(v << dshift);
-                  if (dshift == s_end)
-                  {
-                     dshift = s_start;
-                     dp--;
-                  }
-                  else
-                     dshift += s_inc;
-               }
-               if (sshift == s_end)
-               {
-                  sshift = s_start;
-                  sp--;
-               }
-               else
-                  sshift += s_inc;
-            }
-            break;
-         }
-
-       /*====================================================================*/
-
-         default: /* 8-bit or larger (this is where the routine is modified) */
-         {
-#if 0
-//          static unsigned long long _const4 = 0x0000000000FFFFFFLL;  no good
-//          static unsigned long long const4 = 0x0000000000FFFFFFLL;   no good
-//          unsigned long long _const4 = 0x0000000000FFFFFFLL;         no good
-//          unsigned long long const4 = 0x0000000000FFFFFFLL;          no good
-#endif
-            png_bytep sptr, dp;
-            png_uint_32 i;
-            png_size_t pixel_bytes;
-            int width = (int)row_info->width;
-
-            pixel_bytes = (row_info->pixel_depth >> 3);
-
-            /* point sptr at the last pixel in the pre-expanded row: */
-            sptr = row + (width - 1) * pixel_bytes;
-
-            /* point dp at the last pixel position in the expanded row: */
-            dp = row + (final_width - 1) * pixel_bytes;
-
-            /* New code by Nirav Chhatrapati - Intel Corporation */
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED)
-#if !defined(PNG_1_0_X)
-            if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_INTERLACE)
-                /* && _mmx_supported */ )
-#else
-            if (_mmx_supported)
-#endif
-            {
-               //--------------------------------------------------------------
-               if (pixel_bytes == 3)
-               {
-                  if (((pass == 0) || (pass == 1)) && width)
-                  {
-                     int dummy_value_c;   // fix 'forbidden register spilled'
-                     int dummy_value_S;
-                     int dummy_value_D;
-
-                     __asm__ __volatile__ (
-                        "subl $21, %%edi         \n\t"
-                                     // (png_pass_inc[pass] - 1)*pixel_bytes
-
-                     ".loop3_pass0:              \n\t"
-                        "movd (%%esi), %%mm0     \n\t" // x x x x x 2 1 0
-                        "pand _const4, %%mm0     \n\t" // z z z z z 2 1 0
-                        "movq %%mm0, %%mm1       \n\t" // z z z z z 2 1 0
-                        "psllq $16, %%mm0        \n\t" // z z z 2 1 0 z z
-                        "movq %%mm0, %%mm2       \n\t" // z z z 2 1 0 z z
-                        "psllq $24, %%mm0        \n\t" // 2 1 0 z z z z z
-                        "psrlq $8, %%mm1         \n\t" // z z z z z z 2 1
-                        "por %%mm2, %%mm0        \n\t" // 2 1 0 2 1 0 z z
-                        "por %%mm1, %%mm0        \n\t" // 2 1 0 2 1 0 2 1
-                        "movq %%mm0, %%mm3       \n\t" // 2 1 0 2 1 0 2 1
-                        "psllq $16, %%mm0        \n\t" // 0 2 1 0 2 1 z z
-                        "movq %%mm3, %%mm4       \n\t" // 2 1 0 2 1 0 2 1
-                        "punpckhdq %%mm0, %%mm3  \n\t" // 0 2 1 0 2 1 0 2
-                        "movq %%mm4, 16(%%edi)   \n\t"
-                        "psrlq $32, %%mm0        \n\t" // z z z z 0 2 1 0
-                        "movq %%mm3, 8(%%edi)    \n\t"
-                        "punpckldq %%mm4, %%mm0  \n\t" // 1 0 2 1 0 2 1 0
-                        "subl $3, %%esi          \n\t"
-                        "movq %%mm0, (%%edi)     \n\t"
-                        "subl $24, %%edi         \n\t"
-                        "decl %%ecx              \n\t"
-                        "jnz .loop3_pass0        \n\t"
-                        "EMMS                    \n\t" // DONE
-
-                        : "=c" (dummy_value_c),        // output regs (dummy)
-                          "=S" (dummy_value_S),
-                          "=D" (dummy_value_D)
-
-                        : "1" (sptr),      // esi      // input regs
-                          "2" (dp),        // edi
-                          "0" (width)      // ecx
-// doesn't work           "i" (0x0000000000FFFFFFLL)   // %1 (a.k.a. _const4)
-
-#if 0  /* %mm0, ..., %mm4 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                        : "%mm0", "%mm1", "%mm2"       // clobber list
-                        , "%mm3", "%mm4"
-#endif
-                     );
-                  }
-                  else if (((pass == 2) || (pass == 3)) && width)
-                  {
-                     int dummy_value_c;   // fix 'forbidden register spilled'
-                     int dummy_value_S;
-                     int dummy_value_D;
-
-                     __asm__ __volatile__ (
-                        "subl $9, %%edi          \n\t"
-                                     // (png_pass_inc[pass] - 1)*pixel_bytes
-
-                     ".loop3_pass2:              \n\t"
-                        "movd (%%esi), %%mm0     \n\t" // x x x x x 2 1 0
-                        "pand _const4, %%mm0     \n\t" // z z z z z 2 1 0
-                        "movq %%mm0, %%mm1       \n\t" // z z z z z 2 1 0
-                        "psllq $16, %%mm0        \n\t" // z z z 2 1 0 z z
-                        "movq %%mm0, %%mm2       \n\t" // z z z 2 1 0 z z
-                        "psllq $24, %%mm0        \n\t" // 2 1 0 z z z z z
-                        "psrlq $8, %%mm1         \n\t" // z z z z z z 2 1
-                        "por %%mm2, %%mm0        \n\t" // 2 1 0 2 1 0 z z
-                        "por %%mm1, %%mm0        \n\t" // 2 1 0 2 1 0 2 1
-                        "movq %%mm0, 4(%%edi)    \n\t"
-                        "psrlq $16, %%mm0        \n\t" // z z 2 1 0 2 1 0
-                        "subl $3, %%esi          \n\t"
-                        "movd %%mm0, (%%edi)     \n\t"
-                        "subl $12, %%edi         \n\t"
-                        "decl %%ecx              \n\t"
-                        "jnz .loop3_pass2        \n\t"
-                        "EMMS                    \n\t" // DONE
-
-                        : "=c" (dummy_value_c),        // output regs (dummy)
-                          "=S" (dummy_value_S),
-                          "=D" (dummy_value_D)
-
-                        : "1" (sptr),      // esi      // input regs
-                          "2" (dp),        // edi
-                          "0" (width)      // ecx
-
-#if 0  /* %mm0, ..., %mm2 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                        : "%mm0", "%mm1", "%mm2"       // clobber list
-#endif
-                     );
-                  }
-                  else if (width) /* && ((pass == 4) || (pass == 5)) */
-                  {
-                     int width_mmx = ((width >> 1) << 1) - 8;   // GRR:  huh?
-                     if (width_mmx < 0)
-                         width_mmx = 0;
-                     width -= width_mmx;        // 8 or 9 pix, 24 or 27 bytes
-                     if (width_mmx)
-                     {
-                        // png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1};
-                        // sptr points at last pixel in pre-expanded row
-                        // dp points at last pixel position in expanded row
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $3, %%esi          \n\t"
-                           "subl $9, %%edi          \n\t"
-                                        // (png_pass_inc[pass] + 1)*pixel_bytes
-
-                        ".loop3_pass4:              \n\t"
-                           "movq (%%esi), %%mm0     \n\t" // x x 5 4 3 2 1 0
-                           "movq %%mm0, %%mm1       \n\t" // x x 5 4 3 2 1 0
-                           "movq %%mm0, %%mm2       \n\t" // x x 5 4 3 2 1 0
-                           "psllq $24, %%mm0        \n\t" // 4 3 2 1 0 z z z
-                           "pand _const4, %%mm1     \n\t" // z z z z z 2 1 0
-                           "psrlq $24, %%mm2        \n\t" // z z z x x 5 4 3
-                           "por %%mm1, %%mm0        \n\t" // 4 3 2 1 0 2 1 0
-                           "movq %%mm2, %%mm3       \n\t" // z z z x x 5 4 3
-                           "psllq $8, %%mm2         \n\t" // z z x x 5 4 3 z
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "psrlq $16, %%mm3        \n\t" // z z z z z x x 5
-                           "pand _const6, %%mm3     \n\t" // z z z z z z z 5
-                           "por %%mm3, %%mm2        \n\t" // z z x x 5 4 3 5
-                           "subl $6, %%esi          \n\t"
-                           "movd %%mm2, 8(%%edi)    \n\t"
-                           "subl $12, %%edi         \n\t"
-                           "subl $2, %%ecx          \n\t"
-                           "jnz .loop3_pass4        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0, ..., %mm3 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0", "%mm1"               // clobber list
-                           , "%mm2", "%mm3"
-#endif
-                        );
-                     }
-
-                     sptr -= width_mmx*3;
-                     dp -= width_mmx*6;
-                     for (i = width; i; i--)
-                     {
-                        png_byte v[8];
-                        int j;
-
-                        png_memcpy(v, sptr, 3);
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           png_memcpy(dp, v, 3);
-                           dp -= 3;
-                        }
-                        sptr -= 3;
-                     }
-                  }
-               } /* end of pixel_bytes == 3 */
-
-               //--------------------------------------------------------------
-               else if (pixel_bytes == 1)
-               {
-                  if (((pass == 0) || (pass == 1)) && width)
-                  {
-                     int width_mmx = ((width >> 2) << 2);
-                     width -= width_mmx;        // 0-3 pixels => 0-3 bytes
-                     if (width_mmx)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $3, %%esi          \n\t"
-                           "subl $31, %%edi         \n\t"
-
-                        ".loop1_pass0:              \n\t"
-                           "movd (%%esi), %%mm0     \n\t" // x x x x 3 2 1 0
-                           "movq %%mm0, %%mm1       \n\t" // x x x x 3 2 1 0
-                           "punpcklbw %%mm0, %%mm0  \n\t" // 3 3 2 2 1 1 0 0
-                           "movq %%mm0, %%mm2       \n\t" // 3 3 2 2 1 1 0 0
-                           "punpcklwd %%mm0, %%mm0  \n\t" // 1 1 1 1 0 0 0 0
-                           "movq %%mm0, %%mm3       \n\t" // 1 1 1 1 0 0 0 0
-                           "punpckldq %%mm0, %%mm0  \n\t" // 0 0 0 0 0 0 0 0
-                           "punpckhdq %%mm3, %%mm3  \n\t" // 1 1 1 1 1 1 1 1
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "punpckhwd %%mm2, %%mm2  \n\t" // 3 3 3 3 2 2 2 2
-                           "movq %%mm3, 8(%%edi)    \n\t"
-                           "movq %%mm2, %%mm4       \n\t" // 3 3 3 3 2 2 2 2
-                           "punpckldq %%mm2, %%mm2  \n\t" // 2 2 2 2 2 2 2 2
-                           "punpckhdq %%mm4, %%mm4  \n\t" // 3 3 3 3 3 3 3 3
-                           "movq %%mm2, 16(%%edi)   \n\t"
-                           "subl $4, %%esi          \n\t"
-                           "movq %%mm4, 24(%%edi)   \n\t"
-                           "subl $32, %%edi         \n\t"
-                           "subl $4, %%ecx          \n\t"
-                           "jnz .loop1_pass0        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0, ..., %mm4 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0", "%mm1", "%mm2"       // clobber list
-                           , "%mm3", "%mm4"
-#endif
-                        );
-                     }
-
-                     sptr -= width_mmx;
-                     dp -= width_mmx*8;
-                     for (i = width; i; i--)
-                     {
-                        int j;
-
-                       /* I simplified this part in version 1.0.4e
-                        * here and in several other instances where
-                        * pixel_bytes == 1  -- GR-P
-                        *
-                        * Original code:
-                        *
-                        * png_byte v[8];
-                        * png_memcpy(v, sptr, pixel_bytes);
-                        * for (j = 0; j < png_pass_inc[pass]; j++)
-                        * {
-                        *    png_memcpy(dp, v, pixel_bytes);
-                        *    dp -= pixel_bytes;
-                        * }
-                        * sptr -= pixel_bytes;
-                        *
-                        * Replacement code is in the next three lines:
-                        */
-
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           *dp-- = *sptr;
-                        }
-                        --sptr;
-                     }
-                  }
-                  else if (((pass == 2) || (pass == 3)) && width)
-                  {
-                     int width_mmx = ((width >> 2) << 2);
-                     width -= width_mmx;        // 0-3 pixels => 0-3 bytes
-                     if (width_mmx)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $3, %%esi          \n\t"
-                           "subl $15, %%edi         \n\t"
-
-                        ".loop1_pass2:              \n\t"
-                           "movd (%%esi), %%mm0     \n\t" // x x x x 3 2 1 0
-                           "punpcklbw %%mm0, %%mm0  \n\t" // 3 3 2 2 1 1 0 0
-                           "movq %%mm0, %%mm1       \n\t" // 3 3 2 2 1 1 0 0
-                           "punpcklwd %%mm0, %%mm0  \n\t" // 1 1 1 1 0 0 0 0
-                           "punpckhwd %%mm1, %%mm1  \n\t" // 3 3 3 3 2 2 2 2
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "subl $4, %%esi          \n\t"
-                           "movq %%mm1, 8(%%edi)    \n\t"
-                           "subl $16, %%edi         \n\t"
-                           "subl $4, %%ecx          \n\t"
-                           "jnz .loop1_pass2        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0, %mm1 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0", "%mm1"               // clobber list
-#endif
-                        );
-                     }
-
-                     sptr -= width_mmx;
-                     dp -= width_mmx*4;
-                     for (i = width; i; i--)
-                     {
-                        int j;
-
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           *dp-- = *sptr;
-                        }
-                        --sptr;
-                     }
-                  }
-                  else if (width)  /* && ((pass == 4) || (pass == 5)) */
-                  {
-                     int width_mmx = ((width >> 3) << 3);
-                     width -= width_mmx;        // 0-3 pixels => 0-3 bytes
-                     if (width_mmx)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $7, %%esi          \n\t"
-                           "subl $15, %%edi         \n\t"
-
-                        ".loop1_pass4:              \n\t"
-                           "movq (%%esi), %%mm0     \n\t" // 7 6 5 4 3 2 1 0
-                           "movq %%mm0, %%mm1       \n\t" // 7 6 5 4 3 2 1 0
-                           "punpcklbw %%mm0, %%mm0  \n\t" // 3 3 2 2 1 1 0 0
-                           "punpckhbw %%mm1, %%mm1  \n\t" // 7 7 6 6 5 5 4 4
-                           "movq %%mm1, 8(%%edi)    \n\t"
-                           "subl $8, %%esi          \n\t"
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "subl $16, %%edi         \n\t"
-                           "subl $8, %%ecx          \n\t"
-                           "jnz .loop1_pass4        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (none)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0, %mm1 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0", "%mm1"               // clobber list
-#endif
-                        );
-                     }
-
-                     sptr -= width_mmx;
-                     dp -= width_mmx*2;
-                     for (i = width; i; i--)
-                     {
-                        int j;
-
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           *dp-- = *sptr;
-                        }
-                        --sptr;
-                     }
-                  }
-               } /* end of pixel_bytes == 1 */
-
-               //--------------------------------------------------------------
-               else if (pixel_bytes == 2)
-               {
-                  if (((pass == 0) || (pass == 1)) && width)
-                  {
-                     int width_mmx = ((width >> 1) << 1);
-                     width -= width_mmx;        // 0,1 pixels => 0,2 bytes
-                     if (width_mmx)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $2, %%esi          \n\t"
-                           "subl $30, %%edi         \n\t"
-
-                        ".loop2_pass0:              \n\t"
-                           "movd (%%esi), %%mm0     \n\t" // x x x x 3 2 1 0
-                           "punpcklwd %%mm0, %%mm0  \n\t" // 3 2 3 2 1 0 1 0
-                           "movq %%mm0, %%mm1       \n\t" // 3 2 3 2 1 0 1 0
-                           "punpckldq %%mm0, %%mm0  \n\t" // 1 0 1 0 1 0 1 0
-                           "punpckhdq %%mm1, %%mm1  \n\t" // 3 2 3 2 3 2 3 2
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "movq %%mm0, 8(%%edi)    \n\t"
-                           "movq %%mm1, 16(%%edi)   \n\t"
-                           "subl $4, %%esi          \n\t"
-                           "movq %%mm1, 24(%%edi)   \n\t"
-                           "subl $32, %%edi         \n\t"
-                           "subl $2, %%ecx          \n\t"
-                           "jnz .loop2_pass0        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0, %mm1 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0", "%mm1"               // clobber list
-#endif
-                        );
-                     }
-
-                     sptr -= (width_mmx*2 - 2); // sign fixed
-                     dp -= (width_mmx*16 - 2);  // sign fixed
-                     for (i = width; i; i--)
-                     {
-                        png_byte v[8];
-                        int j;
-                        sptr -= 2;
-                        png_memcpy(v, sptr, 2);
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           dp -= 2;
-                           png_memcpy(dp, v, 2);
-                        }
-                     }
-                  }
-                  else if (((pass == 2) || (pass == 3)) && width)
-                  {
-                     int width_mmx = ((width >> 1) << 1) ;
-                     width -= width_mmx;        // 0,1 pixels => 0,2 bytes
-                     if (width_mmx)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $2, %%esi          \n\t"
-                           "subl $14, %%edi         \n\t"
-
-                        ".loop2_pass2:              \n\t"
-                           "movd (%%esi), %%mm0     \n\t" // x x x x 3 2 1 0
-                           "punpcklwd %%mm0, %%mm0  \n\t" // 3 2 3 2 1 0 1 0
-                           "movq %%mm0, %%mm1       \n\t" // 3 2 3 2 1 0 1 0
-                           "punpckldq %%mm0, %%mm0  \n\t" // 1 0 1 0 1 0 1 0
-                           "punpckhdq %%mm1, %%mm1  \n\t" // 3 2 3 2 3 2 3 2
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "subl $4, %%esi          \n\t"
-                           "movq %%mm1, 8(%%edi)    \n\t"
-                           "subl $16, %%edi         \n\t"
-                           "subl $2, %%ecx          \n\t"
-                           "jnz .loop2_pass2        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0, %mm1 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0", "%mm1"               // clobber list
-#endif
-                        );
-                     }
-
-                     sptr -= (width_mmx*2 - 2); // sign fixed
-                     dp -= (width_mmx*8 - 2);   // sign fixed
-                     for (i = width; i; i--)
-                     {
-                        png_byte v[8];
-                        int j;
-                        sptr -= 2;
-                        png_memcpy(v, sptr, 2);
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           dp -= 2;
-                           png_memcpy(dp, v, 2);
-                        }
-                     }
-                  }
-                  else if (width)  // pass == 4 or 5
-                  {
-                     int width_mmx = ((width >> 1) << 1) ;
-                     width -= width_mmx;        // 0,1 pixels => 0,2 bytes
-                     if (width_mmx)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $2, %%esi          \n\t"
-                           "subl $6, %%edi          \n\t"
-
-                        ".loop2_pass4:              \n\t"
-                           "movd (%%esi), %%mm0     \n\t" // x x x x 3 2 1 0
-                           "punpcklwd %%mm0, %%mm0  \n\t" // 3 2 3 2 1 0 1 0
-                           "subl $4, %%esi          \n\t"
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "subl $8, %%edi          \n\t"
-                           "subl $2, %%ecx          \n\t"
-                           "jnz .loop2_pass4        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0"                       // clobber list
-#endif
-                        );
-                     }
-
-                     sptr -= (width_mmx*2 - 2); // sign fixed
-                     dp -= (width_mmx*4 - 2);   // sign fixed
-                     for (i = width; i; i--)
-                     {
-                        png_byte v[8];
-                        int j;
-                        sptr -= 2;
-                        png_memcpy(v, sptr, 2);
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           dp -= 2;
-                           png_memcpy(dp, v, 2);
-                        }
-                     }
-                  }
-               } /* end of pixel_bytes == 2 */
-
-               //--------------------------------------------------------------
-               else if (pixel_bytes == 4)
-               {
-                  if (((pass == 0) || (pass == 1)) && width)
-                  {
-                     int width_mmx = ((width >> 1) << 1);
-                     width -= width_mmx;        // 0,1 pixels => 0,4 bytes
-                     if (width_mmx)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $4, %%esi          \n\t"
-                           "subl $60, %%edi         \n\t"
-
-                        ".loop4_pass0:              \n\t"
-                           "movq (%%esi), %%mm0     \n\t" // 7 6 5 4 3 2 1 0
-                           "movq %%mm0, %%mm1       \n\t" // 7 6 5 4 3 2 1 0
-                           "punpckldq %%mm0, %%mm0  \n\t" // 3 2 1 0 3 2 1 0
-                           "punpckhdq %%mm1, %%mm1  \n\t" // 7 6 5 4 7 6 5 4
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "movq %%mm0, 8(%%edi)    \n\t"
-                           "movq %%mm0, 16(%%edi)   \n\t"
-                           "movq %%mm0, 24(%%edi)   \n\t"
-                           "movq %%mm1, 32(%%edi)   \n\t"
-                           "movq %%mm1, 40(%%edi)   \n\t"
-                           "movq %%mm1, 48(%%edi)   \n\t"
-                           "subl $8, %%esi          \n\t"
-                           "movq %%mm1, 56(%%edi)   \n\t"
-                           "subl $64, %%edi         \n\t"
-                           "subl $2, %%ecx          \n\t"
-                           "jnz .loop4_pass0        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0, %mm1 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0", "%mm1"               // clobber list
-#endif
-                        );
-                     }
-
-                     sptr -= (width_mmx*4 - 4); // sign fixed
-                     dp -= (width_mmx*32 - 4);  // sign fixed
-                     for (i = width; i; i--)
-                     {
-                        png_byte v[8];
-                        int j;
-                        sptr -= 4;
-                        png_memcpy(v, sptr, 4);
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           dp -= 4;
-                           png_memcpy(dp, v, 4);
-                        }
-                     }
-                  }
-                  else if (((pass == 2) || (pass == 3)) && width)
-                  {
-                     int width_mmx = ((width >> 1) << 1);
-                     width -= width_mmx;        // 0,1 pixels => 0,4 bytes
-                     if (width_mmx)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $4, %%esi          \n\t"
-                           "subl $28, %%edi         \n\t"
-
-                        ".loop4_pass2:              \n\t"
-                           "movq (%%esi), %%mm0     \n\t" // 7 6 5 4 3 2 1 0
-                           "movq %%mm0, %%mm1       \n\t" // 7 6 5 4 3 2 1 0
-                           "punpckldq %%mm0, %%mm0  \n\t" // 3 2 1 0 3 2 1 0
-                           "punpckhdq %%mm1, %%mm1  \n\t" // 7 6 5 4 7 6 5 4
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "movq %%mm0, 8(%%edi)    \n\t"
-                           "movq %%mm1, 16(%%edi)   \n\t"
-                           "movq %%mm1, 24(%%edi)   \n\t"
-                           "subl $8, %%esi          \n\t"
-                           "subl $32, %%edi         \n\t"
-                           "subl $2, %%ecx          \n\t"
-                           "jnz .loop4_pass2        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0, %mm1 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0", "%mm1"               // clobber list
-#endif
-                        );
-                     }
-
-                     sptr -= (width_mmx*4 - 4); // sign fixed
-                     dp -= (width_mmx*16 - 4);  // sign fixed
-                     for (i = width; i; i--)
-                     {
-                        png_byte v[8];
-                        int j;
-                        sptr -= 4;
-                        png_memcpy(v, sptr, 4);
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           dp -= 4;
-                           png_memcpy(dp, v, 4);
-                        }
-                     }
-                  }
-                  else if (width)  // pass == 4 or 5
-                  {
-                     int width_mmx = ((width >> 1) << 1) ;
-                     width -= width_mmx;        // 0,1 pixels => 0,4 bytes
-                     if (width_mmx)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $4, %%esi          \n\t"
-                           "subl $12, %%edi         \n\t"
-
-                        ".loop4_pass4:              \n\t"
-                           "movq (%%esi), %%mm0     \n\t" // 7 6 5 4 3 2 1 0
-                           "movq %%mm0, %%mm1       \n\t" // 7 6 5 4 3 2 1 0
-                           "punpckldq %%mm0, %%mm0  \n\t" // 3 2 1 0 3 2 1 0
-                           "punpckhdq %%mm1, %%mm1  \n\t" // 7 6 5 4 7 6 5 4
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "subl $8, %%esi          \n\t"
-                           "movq %%mm1, 8(%%edi)    \n\t"
-                           "subl $16, %%edi         \n\t"
-                           "subl $2, %%ecx          \n\t"
-                           "jnz .loop4_pass4        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width_mmx)  // ecx
-
-#if 0  /* %mm0, %mm1 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0", "%mm1"               // clobber list
-#endif
-                        );
-                     }
-
-                     sptr -= (width_mmx*4 - 4); // sign fixed
-                     dp -= (width_mmx*8 - 4);   // sign fixed
-                     for (i = width; i; i--)
-                     {
-                        png_byte v[8];
-                        int j;
-                        sptr -= 4;
-                        png_memcpy(v, sptr, 4);
-                        for (j = 0; j < png_pass_inc[pass]; j++)
-                        {
-                           dp -= 4;
-                           png_memcpy(dp, v, 4);
-                        }
-                     }
-                  }
-               } /* end of pixel_bytes == 4 */
-
-               //--------------------------------------------------------------
-               else if (pixel_bytes == 8)
-               {
-// GRR TEST:  should work, but needs testing (special 64-bit version of rpng2?)
-                  // GRR NOTE:  no need to combine passes here!
-                  if (((pass == 0) || (pass == 1)) && width)
-                  {
-                     int dummy_value_c;  // fix 'forbidden register spilled'
-                     int dummy_value_S;
-                     int dummy_value_D;
-
-                     // source is 8-byte RRGGBBAA
-                     // dest is 64-byte RRGGBBAA RRGGBBAA RRGGBBAA RRGGBBAA ...
-                     __asm__ __volatile__ (
-                        "subl $56, %%edi         \n\t" // start of last block
-
-                     ".loop8_pass0:              \n\t"
-                        "movq (%%esi), %%mm0     \n\t" // 7 6 5 4 3 2 1 0
-                        "movq %%mm0, (%%edi)     \n\t"
-                        "movq %%mm0, 8(%%edi)    \n\t"
-                        "movq %%mm0, 16(%%edi)   \n\t"
-                        "movq %%mm0, 24(%%edi)   \n\t"
-                        "movq %%mm0, 32(%%edi)   \n\t"
-                        "movq %%mm0, 40(%%edi)   \n\t"
-                        "movq %%mm0, 48(%%edi)   \n\t"
-                        "subl $8, %%esi          \n\t"
-                        "movq %%mm0, 56(%%edi)   \n\t"
-                        "subl $64, %%edi         \n\t"
-                        "decl %%ecx              \n\t"
-                        "jnz .loop8_pass0        \n\t"
-                        "EMMS                    \n\t" // DONE
-
-                        : "=c" (dummy_value_c),        // output regs (dummy)
-                          "=S" (dummy_value_S),
-                          "=D" (dummy_value_D)
-
-                        : "1" (sptr),      // esi      // input regs
-                          "2" (dp),        // edi
-                          "0" (width)      // ecx
-
-#if 0  /* %mm0 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                        : "%mm0"                       // clobber list
-#endif
-                     );
-                  }
-                  else if (((pass == 2) || (pass == 3)) && width)
-                  {
-                     // source is 8-byte RRGGBBAA
-                     // dest is 32-byte RRGGBBAA RRGGBBAA RRGGBBAA RRGGBBAA
-                     // (recall that expansion is _in place_:  sptr and dp
-                     //  both point at locations within same row buffer)
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $24, %%edi         \n\t" // start of last block
-
-                        ".loop8_pass2:              \n\t"
-                           "movq (%%esi), %%mm0     \n\t" // 7 6 5 4 3 2 1 0
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "movq %%mm0, 8(%%edi)    \n\t"
-                           "movq %%mm0, 16(%%edi)   \n\t"
-                           "subl $8, %%esi          \n\t"
-                           "movq %%mm0, 24(%%edi)   \n\t"
-                           "subl $32, %%edi         \n\t"
-                           "decl %%ecx              \n\t"
-                           "jnz .loop8_pass2        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width)      // ecx
-
-#if 0  /* %mm0 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0"                       // clobber list
-#endif
-                        );
-                     }
-                  }
-                  else if (width)  // pass == 4 or 5
-                  {
-                     // source is 8-byte RRGGBBAA
-                     // dest is 16-byte RRGGBBAA RRGGBBAA
-                     {
-                        int dummy_value_c;  // fix 'forbidden register spilled'
-                        int dummy_value_S;
-                        int dummy_value_D;
-
-                        __asm__ __volatile__ (
-                           "subl $8, %%edi          \n\t" // start of last block
-
-                        ".loop8_pass4:              \n\t"
-                           "movq (%%esi), %%mm0     \n\t" // 7 6 5 4 3 2 1 0
-                           "movq %%mm0, (%%edi)     \n\t"
-                           "subl $8, %%esi          \n\t"
-                           "movq %%mm0, 8(%%edi)    \n\t"
-                           "subl $16, %%edi         \n\t"
-                           "decl %%ecx              \n\t"
-                           "jnz .loop8_pass4        \n\t"
-                           "EMMS                    \n\t" // DONE
-
-                           : "=c" (dummy_value_c),        // output regs (dummy)
-                             "=S" (dummy_value_S),
-                             "=D" (dummy_value_D)
-
-                           : "1" (sptr),      // esi      // input regs
-                             "2" (dp),        // edi
-                             "0" (width)      // ecx
-
-#if 0  /* %mm0 not supported by gcc 2.7.2.3 or egcs 1.1 */
-                           : "%mm0"                       // clobber list
-#endif
-                        );
-                     }
-                  }
-
-               } /* end of pixel_bytes == 8 */
-
-               //--------------------------------------------------------------
-               else if (pixel_bytes == 6)
-               {
-                  for (i = width; i; i--)
-                  {
-                     png_byte v[8];
-                     int j;
-                     png_memcpy(v, sptr, 6);
-                     for (j = 0; j < png_pass_inc[pass]; j++)
-                     {
-                        png_memcpy(dp, v, 6);
-                        dp -= 6;
-                     }
-                     sptr -= 6;
-                  }
-               } /* end of pixel_bytes == 6 */
-
-               //--------------------------------------------------------------
-               else
-               {
-                  for (i = width; i; i--)
-                  {
-                     png_byte v[8];
-                     int j;
-                     png_memcpy(v, sptr, pixel_bytes);
-                     for (j = 0; j < png_pass_inc[pass]; j++)
-                     {
-                        png_memcpy(dp, v, pixel_bytes);
-                        dp -= pixel_bytes;
-                     }
-                     sptr-= pixel_bytes;
-                  }
-               }
-            } // end of _mmx_supported ========================================
-
-            else /* MMX not supported:  use modified C code - takes advantage
-                  *   of inlining of png_memcpy for a constant */
-                 /* GRR 19991007:  does it?  or should pixel_bytes in each
-                  *   block be replaced with immediate value (e.g., 1)? */
-                 /* GRR 19991017:  replaced with constants in each case */
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-            {
-               if (pixel_bytes == 1)
-               {
-                  for (i = width; i; i--)
-                  {
-                     int j;
-                     for (j = 0; j < png_pass_inc[pass]; j++)
-                     {
-                        *dp-- = *sptr;
-                     }
-                     --sptr;
-                  }
-               }
-               else if (pixel_bytes == 3)
-               {
-                  for (i = width; i; i--)
-                  {
-                     png_byte v[8];
-                     int j;
-                     png_memcpy(v, sptr, 3);
-                     for (j = 0; j < png_pass_inc[pass]; j++)
-                     {
-                        png_memcpy(dp, v, 3);
-                        dp -= 3;
-                     }
-                     sptr -= 3;
-                  }
-               }
-               else if (pixel_bytes == 2)
-               {
-                  for (i = width; i; i--)
-                  {
-                     png_byte v[8];
-                     int j;
-                     png_memcpy(v, sptr, 2);
-                     for (j = 0; j < png_pass_inc[pass]; j++)
-                     {
-                        png_memcpy(dp, v, 2);
-                        dp -= 2;
-                     }
-                     sptr -= 2;
-                  }
-               }
-               else if (pixel_bytes == 4)
-               {
-                  for (i = width; i; i--)
-                  {
-                     png_byte v[8];
-                     int j;
-                     png_memcpy(v, sptr, 4);
-                     for (j = 0; j < png_pass_inc[pass]; j++)
-                     {
-#ifdef PNG_DEBUG
-                        if (dp < row || dp+3 > row+png_ptr->row_buf_size)
-                        {
-                           printf("dp out of bounds: row=%d, dp=%d, rp=%d\n",
-                             row, dp, row+png_ptr->row_buf_size);
-                           printf("row_buf=%d\n",png_ptr->row_buf_size);
-                        }
-#endif
-                        png_memcpy(dp, v, 4);
-                        dp -= 4;
-                     }
-                     sptr -= 4;
-                  }
-               }
-               else if (pixel_bytes == 6)
-               {
-                  for (i = width; i; i--)
-                  {
-                     png_byte v[8];
-                     int j;
-                     png_memcpy(v, sptr, 6);
-                     for (j = 0; j < png_pass_inc[pass]; j++)
-                     {
-                        png_memcpy(dp, v, 6);
-                        dp -= 6;
-                     }
-                     sptr -= 6;
-                  }
-               }
-               else if (pixel_bytes == 8)
-               {
-                  for (i = width; i; i--)
-                  {
-                     png_byte v[8];
-                     int j;
-                     png_memcpy(v, sptr, 8);
-                     for (j = 0; j < png_pass_inc[pass]; j++)
-                     {
-                        png_memcpy(dp, v, 8);
-                        dp -= 8;
-                     }
-                     sptr -= 8;
-                  }
-               }
-               else     /* GRR:  should never be reached */
-               {
-                  for (i = width; i; i--)
-                  {
-                     png_byte v[8];
-                     int j;
-                     png_memcpy(v, sptr, pixel_bytes);
-                     for (j = 0; j < png_pass_inc[pass]; j++)
-                     {
-                        png_memcpy(dp, v, pixel_bytes);
-                        dp -= pixel_bytes;
-                     }
-                     sptr -= pixel_bytes;
-                  }
-               }
-
-            } /* end if (MMX not supported) */
-            break;
-         }
-      } /* end switch (row_info->pixel_depth) */
-
-      row_info->width = final_width;
-      row_info->rowbytes = ((final_width *
-         (png_uint_32)row_info->pixel_depth + 7) >> 3);
-   }
-
-} /* end png_do_read_interlace() */
-
-#endif /* PNG_HAVE_ASSEMBLER_READ_INTERLACE */
-#endif /* PNG_READ_INTERLACING_SUPPORTED */
-
-
-
-#if defined(PNG_HAVE_ASSEMBLER_READ_FILTER_ROW)
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED)
-
-// These variables are utilized in the functions below.  They are declared
-// globally here to ensure alignment on 8-byte boundaries.
-
-union uAll {
-   long long use;
-   double  align;
-} _LBCarryMask = {0x0101010101010101LL},
-  _HBClearMask = {0x7f7f7f7f7f7f7f7fLL},
-  _ActiveMask, _ActiveMask2, _ActiveMaskEnd, _ShiftBpp, _ShiftRem;
-
-#ifdef PNG_THREAD_UNSAFE_OK
-//===========================================================================//
-//                                                                           //
-//           P N G _ R E A D _ F I L T E R _ R O W _ M M X _ A V G           //
-//                                                                           //
-//===========================================================================//
-
-// Optimized code for PNG Average filter decoder
-
-static void /* PRIVATE */
-png_read_filter_row_mmx_avg(png_row_infop row_info, png_bytep row,
-                            png_bytep prev_row)
-{
-   int bpp;
-   int dummy_value_c;   // fix 'forbidden register 2 (cx) was spilled' error
-   int dummy_value_S;
-   int dummy_value_D;
-
-   bpp = (row_info->pixel_depth + 7) >> 3;  // get # bytes per pixel
-   _FullLength  = row_info->rowbytes;       // # of bytes to filter
-
-   __asm__ __volatile__ (
-      // initialize address pointers and offset
-#ifdef __PIC__
-      "pushl %%ebx                 \n\t" // save index to Global Offset Table
-#endif
-//pre "movl row, %%edi             \n\t" // edi:  Avg(x)
-      "xorl %%ebx, %%ebx           \n\t" // ebx:  x
-      "movl %%edi, %%edx           \n\t"
-//pre "movl prev_row, %%esi        \n\t" // esi:  Prior(x)
-//pre "subl bpp, %%edx             \n\t" // (bpp is preloaded into ecx)
-      "subl %%ecx, %%edx           \n\t" // edx:  Raw(x-bpp)
-
-      "xorl %%eax,%%eax            \n\t"
-
-      // Compute the Raw value for the first bpp bytes
-      //    Raw(x) = Avg(x) + (Prior(x)/2)
-   "avg_rlp:                       \n\t"
-      "movb (%%esi,%%ebx,),%%al    \n\t" // load al with Prior(x)
-      "incl %%ebx                  \n\t"
-      "shrb %%al                   \n\t" // divide by 2
-      "addb -1(%%edi,%%ebx,),%%al  \n\t" // add Avg(x); -1 to offset inc ebx
-//pre "cmpl bpp, %%ebx             \n\t" // (bpp is preloaded into ecx)
-      "cmpl %%ecx, %%ebx           \n\t"
-      "movb %%al,-1(%%edi,%%ebx,)  \n\t" // write Raw(x); -1 to offset inc ebx
-      "jb avg_rlp                  \n\t" // mov does not affect flags
-
-      // get # of bytes to alignment
-      "movl %%edi, _dif            \n\t" // take start of row
-      "addl %%ebx, _dif            \n\t" // add bpp
-      "addl $0xf, _dif             \n\t" // add 7+8 to incr past alignment bdry
-      "andl $0xfffffff8, _dif      \n\t" // mask to alignment boundary
-      "subl %%edi, _dif            \n\t" // subtract from start => value ebx at
-      "jz avg_go                   \n\t" //  alignment
-
-      // fix alignment
-      // Compute the Raw value for the bytes up to the alignment boundary
-      //    Raw(x) = Avg(x) + ((Raw(x-bpp) + Prior(x))/2)
-      "xorl %%ecx, %%ecx           \n\t"
-
-   "avg_lp1:                       \n\t"
-      "xorl %%eax, %%eax           \n\t"
-      "movb (%%esi,%%ebx,), %%cl   \n\t" // load cl with Prior(x)
-      "movb (%%edx,%%ebx,), %%al   \n\t" // load al with Raw(x-bpp)
-      "addw %%cx, %%ax             \n\t"
-      "incl %%ebx                  \n\t"
-      "shrw %%ax                   \n\t" // divide by 2
-      "addb -1(%%edi,%%ebx,), %%al \n\t" // add Avg(x); -1 to offset inc ebx
-      "cmpl _dif, %%ebx            \n\t" // check if at alignment boundary
-      "movb %%al, -1(%%edi,%%ebx,) \n\t" // write Raw(x); -1 to offset inc ebx
-      "jb avg_lp1                  \n\t" // repeat until at alignment boundary
-
-   "avg_go:                        \n\t"
-      "movl _FullLength, %%eax     \n\t"
-      "movl %%eax, %%ecx           \n\t"
-      "subl %%ebx, %%eax           \n\t" // subtract alignment fix
-      "andl $0x00000007, %%eax     \n\t" // calc bytes over mult of 8
-      "subl %%eax, %%ecx           \n\t" // drop over bytes from original length
-      "movl %%ecx, _MMXLength      \n\t"
-#ifdef __PIC__
-      "popl %%ebx                  \n\t" // restore index to Global Offset Table
-#endif
-
-      : "=c" (dummy_value_c),            // output regs (dummy)
-        "=S" (dummy_value_S),
-        "=D" (dummy_value_D)
-
-      : "0" (bpp),       // ecx          // input regs
-        "1" (prev_row),  // esi
-        "2" (row)        // edi
-
-      : "%eax", "%edx"                   // clobber list
-#ifndef __PIC__
-      , "%ebx"
-#endif
-      // GRR: INCLUDE "memory" as clobbered? (_dif, _MMXLength)
-      // (seems to work fine without...)
-   );
-
-   // now do the math for the rest of the row
-   switch (bpp)
-   {
-      case 3:
-      {
-         _ActiveMask.use  = 0x0000000000ffffffLL;
-         _ShiftBpp.use = 24;    // == 3 * 8
-         _ShiftRem.use = 40;    // == 64 - 24
-
-         __asm__ __volatile__ (
-            // re-init address pointers and offset
-            "movq _ActiveMask, %%mm7      \n\t"
-            "movl _dif, %%ecx             \n\t" // ecx:  x = offset to
-            "movq _LBCarryMask, %%mm5     \n\t" //  alignment boundary
-// preload  "movl row, %%edi              \n\t" // edi:  Avg(x)
-            "movq _HBClearMask, %%mm4     \n\t"
-// preload  "movl prev_row, %%esi         \n\t" // esi:  Prior(x)
-
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%ecx,), %%mm2 \n\t" // load previous aligned 8 bytes
-                                                // (correct pos. in loop below)
-         "avg_3lp:                        \n\t"
-            "movq (%%edi,%%ecx,), %%mm0   \n\t" // load mm0 with Avg(x)
-            "movq %%mm5, %%mm3            \n\t"
-            "psrlq _ShiftRem, %%mm2       \n\t" // correct position Raw(x-bpp)
-                                                // data
-            "movq (%%esi,%%ecx,), %%mm1   \n\t" // load mm1 with Prior(x)
-            "movq %%mm7, %%mm6            \n\t"
-            "pand %%mm1, %%mm3            \n\t" // get lsb for each prev_row byte
-            "psrlq $1, %%mm1              \n\t" // divide prev_row bytes by 2
-            "pand  %%mm4, %%mm1           \n\t" // clear invalid bit 7 of each
-                                                // byte
-            "paddb %%mm1, %%mm0           \n\t" // add (Prev_row/2) to Avg for
-                                                // each byte
-            // add 1st active group (Raw(x-bpp)/2) to average with LBCarry
-            "movq %%mm3, %%mm1            \n\t" // now use mm1 for getting
-                                                // LBCarrys
-            "pand %%mm2, %%mm1            \n\t" // get LBCarrys for each byte
-                                                // where both
-                               // lsb's were == 1 (only valid for active group)
-            "psrlq $1, %%mm2              \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm2           \n\t" // clear invalid bit 7 of each
-                                                // byte
-            "paddb %%mm1, %%mm2           \n\t" // add LBCarrys to (Raw(x-bpp)/2)
-                                                // for each byte
-            "pand %%mm6, %%mm2            \n\t" // leave only Active Group 1
-                                                // bytes to add to Avg
-            "paddb %%mm2, %%mm0           \n\t" // add (Raw/2) + LBCarrys to
-                                                // Avg for each Active
-                               //  byte
-            // add 2nd active group (Raw(x-bpp)/2) to average with _LBCarry
-            "psllq _ShiftBpp, %%mm6       \n\t" // shift the mm6 mask to cover
-                                                // bytes 3-5
-            "movq %%mm0, %%mm2            \n\t" // mov updated Raws to mm2
-            "psllq _ShiftBpp, %%mm2       \n\t" // shift data to pos. correctly
-            "movq %%mm3, %%mm1            \n\t" // now use mm1 for getting
-                                                // LBCarrys
-            "pand %%mm2, %%mm1            \n\t" // get LBCarrys for each byte
-                                                // where both
-                               // lsb's were == 1 (only valid for active group)
-            "psrlq $1, %%mm2              \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm2           \n\t" // clear invalid bit 7 of each
-                                                // byte
-            "paddb %%mm1, %%mm2           \n\t" // add LBCarrys to (Raw(x-bpp)/2)
-                                                // for each byte
-            "pand %%mm6, %%mm2            \n\t" // leave only Active Group 2
-                                                // bytes to add to Avg
-            "paddb %%mm2, %%mm0           \n\t" // add (Raw/2) + LBCarrys to
-                                                // Avg for each Active
-                               //  byte
-
-            // add 3rd active group (Raw(x-bpp)/2) to average with _LBCarry
-            "psllq _ShiftBpp, %%mm6       \n\t" // shift mm6 mask to cover last
-                                                // two
-                                 // bytes
-            "movq %%mm0, %%mm2            \n\t" // mov updated Raws to mm2
-            "psllq _ShiftBpp, %%mm2       \n\t" // shift data to pos. correctly
-                              // Data only needs to be shifted once here to
-                              // get the correct x-bpp offset.
-            "movq %%mm3, %%mm1            \n\t" // now use mm1 for getting
-                                                // LBCarrys
-            "pand %%mm2, %%mm1            \n\t" // get LBCarrys for each byte
-                                                // where both
-                              // lsb's were == 1 (only valid for active group)
-            "psrlq $1, %%mm2              \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm2           \n\t" // clear invalid bit 7 of each
-                                                // byte
-            "paddb %%mm1, %%mm2           \n\t" // add LBCarrys to (Raw(x-bpp)/2)
-                                                // for each byte
-            "pand %%mm6, %%mm2            \n\t" // leave only Active Group 2
-                                                // bytes to add to Avg
-            "addl $8, %%ecx               \n\t"
-            "paddb %%mm2, %%mm0           \n\t" // add (Raw/2) + LBCarrys to
-                                                // Avg for each Active
-                                                // byte
-            // now ready to write back to memory
-            "movq %%mm0, -8(%%edi,%%ecx,) \n\t"
-            // move updated Raw(x) to use as Raw(x-bpp) for next loop
-            "cmpl _MMXLength, %%ecx       \n\t"
-            "movq %%mm0, %%mm2            \n\t" // mov updated Raw(x) to mm2
-            "jb avg_3lp                   \n\t"
-
-            : "=S" (dummy_value_S),             // output regs (dummy)
-              "=D" (dummy_value_D)
-
-            : "0" (prev_row),  // esi           // input regs
-              "1" (row)        // edi
-
-            : "%ecx"                            // clobber list
-#if 0  /* %mm0, ..., %mm7 not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm2", "%mm3"
-            , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;  // end 3 bpp
-
-      case 6:
-      case 4:
-      //case 7:   // who wrote this?  PNG doesn't support 5 or 7 bytes/pixel
-      //case 5:   // GRR BOGUS
-      {
-         _ActiveMask.use  = 0xffffffffffffffffLL; // use shift below to clear
-                                                  // appropriate inactive bytes
-         _ShiftBpp.use = bpp << 3;
-         _ShiftRem.use = 64 - _ShiftBpp.use;
-
-         __asm__ __volatile__ (
-            "movq _HBClearMask, %%mm4    \n\t"
-
-            // re-init address pointers and offset
-            "movl _dif, %%ecx            \n\t" // ecx:  x = offset to
-                                               // alignment boundary
-
-            // load _ActiveMask and clear all bytes except for 1st active group
-            "movq _ActiveMask, %%mm7     \n\t"
-// preload  "movl row, %%edi             \n\t" // edi:  Avg(x)
-            "psrlq _ShiftRem, %%mm7      \n\t"
-// preload  "movl prev_row, %%esi        \n\t" // esi:  Prior(x)
-            "movq %%mm7, %%mm6           \n\t"
-            "movq _LBCarryMask, %%mm5    \n\t"
-            "psllq _ShiftBpp, %%mm6      \n\t" // create mask for 2nd active
-                                               // group
-
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%ecx,), %%mm2 \n\t" // load previous aligned 8 bytes
-                                          // (we correct pos. in loop below)
-         "avg_4lp:                       \n\t"
-            "movq (%%edi,%%ecx,), %%mm0  \n\t"
-            "psrlq _ShiftRem, %%mm2      \n\t" // shift data to pos. correctly
-            "movq (%%esi,%%ecx,), %%mm1  \n\t"
-            // add (Prev_row/2) to average
-            "movq %%mm5, %%mm3           \n\t"
-            "pand %%mm1, %%mm3           \n\t" // get lsb for each prev_row byte
-            "psrlq $1, %%mm1             \n\t" // divide prev_row bytes by 2
-            "pand  %%mm4, %%mm1          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "paddb %%mm1, %%mm0          \n\t" // add (Prev_row/2) to Avg for
-                                               // each byte
-            // add 1st active group (Raw(x-bpp)/2) to average with _LBCarry
-            "movq %%mm3, %%mm1           \n\t" // now use mm1 for getting
-                                               // LBCarrys
-            "pand %%mm2, %%mm1           \n\t" // get LBCarrys for each byte
-                                               // where both
-                              // lsb's were == 1 (only valid for active group)
-            "psrlq $1, %%mm2             \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm2          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "paddb %%mm1, %%mm2          \n\t" // add LBCarrys to (Raw(x-bpp)/2)
-                                               // for each byte
-            "pand %%mm7, %%mm2           \n\t" // leave only Active Group 1
-                                               // bytes to add to Avg
-            "paddb %%mm2, %%mm0          \n\t" // add (Raw/2) + LBCarrys to Avg
-                                               // for each Active
-                              // byte
-            // add 2nd active group (Raw(x-bpp)/2) to average with _LBCarry
-            "movq %%mm0, %%mm2           \n\t" // mov updated Raws to mm2
-            "psllq _ShiftBpp, %%mm2      \n\t" // shift data to pos. correctly
-            "addl $8, %%ecx              \n\t"
-            "movq %%mm3, %%mm1           \n\t" // now use mm1 for getting
-                                               // LBCarrys
-            "pand %%mm2, %%mm1           \n\t" // get LBCarrys for each byte
-                                               // where both
-                              // lsb's were == 1 (only valid for active group)
-            "psrlq $1, %%mm2             \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm2          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "paddb %%mm1, %%mm2          \n\t" // add LBCarrys to (Raw(x-bpp)/2)
-                                               // for each byte
-            "pand %%mm6, %%mm2           \n\t" // leave only Active Group 2
-                                               // bytes to add to Avg
-            "paddb %%mm2, %%mm0          \n\t" // add (Raw/2) + LBCarrys to
-                                               // Avg for each Active
-                              // byte
-            "cmpl _MMXLength, %%ecx      \n\t"
-            // now ready to write back to memory
-            "movq %%mm0, -8(%%edi,%%ecx,) \n\t"
-            // prep Raw(x-bpp) for next loop
-            "movq %%mm0, %%mm2           \n\t" // mov updated Raws to mm2
-            "jb avg_4lp                  \n\t"
-
-            : "=S" (dummy_value_S),            // output regs (dummy)
-              "=D" (dummy_value_D)
-
-            : "0" (prev_row),  // esi          // input regs
-              "1" (row)        // edi
-
-            : "%ecx"                           // clobber list
-#if 0  /* %mm0, ..., %mm7 not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm2", "%mm3"
-            , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;  // end 4,6 bpp
-
-      case 2:
-      {
-         _ActiveMask.use  = 0x000000000000ffffLL;
-         _ShiftBpp.use = 16;   // == 2 * 8
-         _ShiftRem.use = 48;   // == 64 - 16
-
-         __asm__ __volatile__ (
-            // load _ActiveMask
-            "movq _ActiveMask, %%mm7     \n\t"
-            // re-init address pointers and offset
-            "movl _dif, %%ecx            \n\t" // ecx:  x = offset to alignment
-                                               // boundary
-            "movq _LBCarryMask, %%mm5    \n\t"
-// preload  "movl row, %%edi             \n\t" // edi:  Avg(x)
-            "movq _HBClearMask, %%mm4    \n\t"
-// preload  "movl prev_row, %%esi        \n\t" // esi:  Prior(x)
-
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%ecx,), %%mm2 \n\t" // load previous aligned 8 bytes
-                              // (we correct pos. in loop below)
-         "avg_2lp:                       \n\t"
-            "movq (%%edi,%%ecx,), %%mm0  \n\t"
-            "psrlq _ShiftRem, %%mm2      \n\t" // shift data to pos. correctly
-            "movq (%%esi,%%ecx,), %%mm1  \n\t" //  (GRR BUGFIX:  was psllq)
-            // add (Prev_row/2) to average
-            "movq %%mm5, %%mm3           \n\t"
-            "pand %%mm1, %%mm3           \n\t" // get lsb for each prev_row byte
-            "psrlq $1, %%mm1             \n\t" // divide prev_row bytes by 2
-            "pand  %%mm4, %%mm1          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "movq %%mm7, %%mm6           \n\t"
-            "paddb %%mm1, %%mm0          \n\t" // add (Prev_row/2) to Avg for
-                                               // each byte
-
-            // add 1st active group (Raw(x-bpp)/2) to average with _LBCarry
-            "movq %%mm3, %%mm1           \n\t" // now use mm1 for getting
-                                               // LBCarrys
-            "pand %%mm2, %%mm1           \n\t" // get LBCarrys for each byte
-                                               // where both
-                                               // lsb's were == 1 (only valid
-                                               // for active group)
-            "psrlq $1, %%mm2             \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm2          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "paddb %%mm1, %%mm2          \n\t" // add LBCarrys to (Raw(x-bpp)/2)
-                                               // for each byte
-            "pand %%mm6, %%mm2           \n\t" // leave only Active Group 1
-                                               // bytes to add to Avg
-            "paddb %%mm2, %%mm0          \n\t" // add (Raw/2) + LBCarrys to Avg
-                                               // for each Active byte
-
-            // add 2nd active group (Raw(x-bpp)/2) to average with _LBCarry
-            "psllq _ShiftBpp, %%mm6      \n\t" // shift the mm6 mask to cover
-                                               // bytes 2 & 3
-            "movq %%mm0, %%mm2           \n\t" // mov updated Raws to mm2
-            "psllq _ShiftBpp, %%mm2      \n\t" // shift data to pos. correctly
-            "movq %%mm3, %%mm1           \n\t" // now use mm1 for getting
-                                               // LBCarrys
-            "pand %%mm2, %%mm1           \n\t" // get LBCarrys for each byte
-                                               // where both
-                                               // lsb's were == 1 (only valid
-                                               // for active group)
-            "psrlq $1, %%mm2             \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm2          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "paddb %%mm1, %%mm2          \n\t" // add LBCarrys to (Raw(x-bpp)/2)
-                                               // for each byte
-            "pand %%mm6, %%mm2           \n\t" // leave only Active Group 2
-                                               // bytes to add to Avg
-            "paddb %%mm2, %%mm0          \n\t" // add (Raw/2) + LBCarrys to
-                                               // Avg for each Active byte
-
-            // add 3rd active group (Raw(x-bpp)/2) to average with _LBCarry
-            "psllq _ShiftBpp, %%mm6      \n\t" // shift the mm6 mask to cover
-                                               // bytes 4 & 5
-            "movq %%mm0, %%mm2           \n\t" // mov updated Raws to mm2
-            "psllq _ShiftBpp, %%mm2      \n\t" // shift data to pos. correctly
-            "movq %%mm3, %%mm1           \n\t" // now use mm1 for getting
-                                               // LBCarrys
-            "pand %%mm2, %%mm1           \n\t" // get LBCarrys for each byte
-                                               // where both lsb's were == 1
-                                               // (only valid for active group)
-            "psrlq $1, %%mm2             \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm2          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "paddb %%mm1, %%mm2          \n\t" // add LBCarrys to (Raw(x-bpp)/2)
-                                               // for each byte
-            "pand %%mm6, %%mm2           \n\t" // leave only Active Group 2
-                                               // bytes to add to Avg
-            "paddb %%mm2, %%mm0          \n\t" // add (Raw/2) + LBCarrys to
-                                               // Avg for each Active byte
-
-            // add 4th active group (Raw(x-bpp)/2) to average with _LBCarry
-            "psllq _ShiftBpp, %%mm6      \n\t" // shift the mm6 mask to cover
-                                               // bytes 6 & 7
-            "movq %%mm0, %%mm2           \n\t" // mov updated Raws to mm2
-            "psllq _ShiftBpp, %%mm2      \n\t" // shift data to pos. correctly
-            "addl $8, %%ecx              \n\t"
-            "movq %%mm3, %%mm1           \n\t" // now use mm1 for getting
-                                               // LBCarrys
-            "pand %%mm2, %%mm1           \n\t" // get LBCarrys for each byte
-                                               // where both
-                                               // lsb's were == 1 (only valid
-                                               // for active group)
-            "psrlq $1, %%mm2             \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm2          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "paddb %%mm1, %%mm2          \n\t" // add LBCarrys to (Raw(x-bpp)/2)
-                                               // for each byte
-            "pand %%mm6, %%mm2           \n\t" // leave only Active Group 2
-                                               // bytes to add to Avg
-            "paddb %%mm2, %%mm0          \n\t" // add (Raw/2) + LBCarrys to
-                                               // Avg for each Active byte
-
-            "cmpl _MMXLength, %%ecx      \n\t"
-            // now ready to write back to memory
-            "movq %%mm0, -8(%%edi,%%ecx,) \n\t"
-            // prep Raw(x-bpp) for next loop
-            "movq %%mm0, %%mm2           \n\t" // mov updated Raws to mm2
-            "jb avg_2lp                  \n\t"
-
-            : "=S" (dummy_value_S),            // output regs (dummy)
-              "=D" (dummy_value_D)
-
-            : "0" (prev_row),  // esi          // input regs
-              "1" (row)        // edi
-
-            : "%ecx"                           // clobber list
-#if 0  /* %mm0, ..., %mm7 not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm2", "%mm3"
-            , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;  // end 2 bpp
-
-      case 1:
-      {
-         __asm__ __volatile__ (
-            // re-init address pointers and offset
-#ifdef __PIC__
-            "pushl %%ebx                 \n\t" // save Global Offset Table index
-#endif
-            "movl _dif, %%ebx            \n\t" // ebx:  x = offset to alignment
-                                               // boundary
-// preload  "movl row, %%edi             \n\t" // edi:  Avg(x)
-            "cmpl _FullLength, %%ebx     \n\t" // test if offset at end of array
-            "jnb avg_1end                \n\t"
-            // do Paeth decode for remaining bytes
-// preload  "movl prev_row, %%esi        \n\t" // esi:  Prior(x)
-            "movl %%edi, %%edx           \n\t"
-// preload  "subl bpp, %%edx             \n\t" // (bpp is preloaded into ecx)
-            "subl %%ecx, %%edx           \n\t" // edx:  Raw(x-bpp)
-            "xorl %%ecx, %%ecx           \n\t" // zero ecx before using cl & cx
-                                               //  in loop below
-         "avg_1lp:                       \n\t"
-            // Raw(x) = Avg(x) + ((Raw(x-bpp) + Prior(x))/2)
-            "xorl %%eax, %%eax           \n\t"
-            "movb (%%esi,%%ebx,), %%cl   \n\t" // load cl with Prior(x)
-            "movb (%%edx,%%ebx,), %%al   \n\t" // load al with Raw(x-bpp)
-            "addw %%cx, %%ax             \n\t"
-            "incl %%ebx                  \n\t"
-            "shrw %%ax                   \n\t" // divide by 2
-            "addb -1(%%edi,%%ebx,), %%al \n\t" // add Avg(x); -1 to offset
-                                               // inc ebx
-            "cmpl _FullLength, %%ebx     \n\t" // check if at end of array
-            "movb %%al, -1(%%edi,%%ebx,) \n\t" // write back Raw(x);
-                         // mov does not affect flags; -1 to offset inc ebx
-            "jb avg_1lp                  \n\t"
-
-         "avg_1end:                      \n\t"
-#ifdef __PIC__
-            "popl %%ebx                  \n\t" // Global Offset Table index
-#endif
-
-            : "=c" (dummy_value_c),            // output regs (dummy)
-              "=S" (dummy_value_S),
-              "=D" (dummy_value_D)
-
-            : "0" (bpp),       // ecx          // input regs
-              "1" (prev_row),  // esi
-              "2" (row)        // edi
-
-            : "%eax", "%edx"                   // clobber list
-#ifndef __PIC__
-            , "%ebx"
-#endif
-         );
-      }
-      return;  // end 1 bpp
-
-      case 8:
-      {
-         __asm__ __volatile__ (
-            // re-init address pointers and offset
-            "movl _dif, %%ecx            \n\t" // ecx:  x == offset to alignment
-            "movq _LBCarryMask, %%mm5    \n\t" //            boundary
-// preload  "movl row, %%edi             \n\t" // edi:  Avg(x)
-            "movq _HBClearMask, %%mm4    \n\t"
-// preload  "movl prev_row, %%esi        \n\t" // esi:  Prior(x)
-
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%ecx,), %%mm2 \n\t" // load previous aligned 8 bytes
-                                      // (NO NEED to correct pos. in loop below)
-
-         "avg_8lp:                       \n\t"
-            "movq (%%edi,%%ecx,), %%mm0  \n\t"
-            "movq %%mm5, %%mm3           \n\t"
-            "movq (%%esi,%%ecx,), %%mm1  \n\t"
-            "addl $8, %%ecx              \n\t"
-            "pand %%mm1, %%mm3           \n\t" // get lsb for each prev_row byte
-            "psrlq $1, %%mm1             \n\t" // divide prev_row bytes by 2
-            "pand %%mm2, %%mm3           \n\t" // get LBCarrys for each byte
-                                               //  where both lsb's were == 1
-            "psrlq $1, %%mm2             \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm1          \n\t" // clear invalid bit 7, each byte
-            "paddb %%mm3, %%mm0          \n\t" // add LBCarrys to Avg, each byte
-            "pand  %%mm4, %%mm2          \n\t" // clear invalid bit 7, each byte
-            "paddb %%mm1, %%mm0          \n\t" // add (Prev_row/2) to Avg, each
-            "paddb %%mm2, %%mm0          \n\t" // add (Raw/2) to Avg for each
-            "cmpl _MMXLength, %%ecx      \n\t"
-            "movq %%mm0, -8(%%edi,%%ecx,) \n\t"
-            "movq %%mm0, %%mm2           \n\t" // reuse as Raw(x-bpp)
-            "jb avg_8lp                  \n\t"
-
-            : "=S" (dummy_value_S),            // output regs (dummy)
-              "=D" (dummy_value_D)
-
-            : "0" (prev_row),  // esi          // input regs
-              "1" (row)        // edi
-
-            : "%ecx"                           // clobber list
-#if 0  /* %mm0, ..., %mm5 not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm2"
-            , "%mm3", "%mm4", "%mm5"
-#endif
-         );
-      }
-      break;  // end 8 bpp
-
-      default:                  // bpp greater than 8 (!= 1,2,3,4,[5],6,[7],8)
-      {
-
-#ifdef PNG_DEBUG
-         // GRR:  PRINT ERROR HERE:  SHOULD NEVER BE REACHED
-        png_debug(1,
-        "Internal logic error in pnggccrd (png_read_filter_row_mmx_avg())\n");
-#endif
-
-#if 0
-        __asm__ __volatile__ (
-            "movq _LBCarryMask, %%mm5    \n\t"
-            // re-init address pointers and offset
-            "movl _dif, %%ebx            \n\t" // ebx:  x = offset to
-                                               // alignment boundary
-            "movl row, %%edi             \n\t" // edi:  Avg(x)
-            "movq _HBClearMask, %%mm4    \n\t"
-            "movl %%edi, %%edx           \n\t"
-            "movl prev_row, %%esi        \n\t" // esi:  Prior(x)
-            "subl bpp, %%edx             \n\t" // edx:  Raw(x-bpp)
-         "avg_Alp:                       \n\t"
-            "movq (%%edi,%%ebx,), %%mm0  \n\t"
-            "movq %%mm5, %%mm3           \n\t"
-            "movq (%%esi,%%ebx,), %%mm1  \n\t"
-            "pand %%mm1, %%mm3           \n\t" // get lsb for each prev_row byte
-            "movq (%%edx,%%ebx,), %%mm2  \n\t"
-            "psrlq $1, %%mm1             \n\t" // divide prev_row bytes by 2
-            "pand %%mm2, %%mm3           \n\t" // get LBCarrys for each byte
-                                               // where both lsb's were == 1
-            "psrlq $1, %%mm2             \n\t" // divide raw bytes by 2
-            "pand  %%mm4, %%mm1          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "paddb %%mm3, %%mm0          \n\t" // add LBCarrys to Avg for each
-                                               // byte
-            "pand  %%mm4, %%mm2          \n\t" // clear invalid bit 7 of each
-                                               // byte
-            "paddb %%mm1, %%mm0          \n\t" // add (Prev_row/2) to Avg for
-                                               // each byte
-            "addl $8, %%ebx              \n\t"
-            "paddb %%mm2, %%mm0          \n\t" // add (Raw/2) to Avg for each
-                                               // byte
-            "cmpl _MMXLength, %%ebx      \n\t"
-            "movq %%mm0, -8(%%edi,%%ebx,) \n\t"
-            "jb avg_Alp                  \n\t"
-
-            : // FIXASM: output regs/vars go here, e.g.:  "=m" (memory_var)
-
-            : // FIXASM: input regs, e.g.:  "c" (count), "S" (src), "D" (dest)
-
-            : "%ebx", "%edx", "%edi", "%esi" // CHECKASM: clobber list
-         );
-#endif /* 0 - NEVER REACHED */
-      }
-      break;
-
-   } // end switch (bpp)
-
-   __asm__ __volatile__ (
-      // MMX acceleration complete; now do clean-up
-      // check if any remaining bytes left to decode
-#ifdef __PIC__
-      "pushl %%ebx                 \n\t" // save index to Global Offset Table
-#endif
-      "movl _MMXLength, %%ebx      \n\t" // ebx:  x == offset bytes after MMX
-//pre "movl row, %%edi             \n\t" // edi:  Avg(x)
-      "cmpl _FullLength, %%ebx     \n\t" // test if offset at end of array
-      "jnb avg_end                 \n\t"
-
-      // do Avg decode for remaining bytes
-//pre "movl prev_row, %%esi        \n\t" // esi:  Prior(x)
-      "movl %%edi, %%edx           \n\t"
-//pre "subl bpp, %%edx             \n\t" // (bpp is preloaded into ecx)
-      "subl %%ecx, %%edx           \n\t" // edx:  Raw(x-bpp)
-      "xorl %%ecx, %%ecx           \n\t" // zero ecx before using cl & cx below
-
-   "avg_lp2:                       \n\t"
-      // Raw(x) = Avg(x) + ((Raw(x-bpp) + Prior(x))/2)
-      "xorl %%eax, %%eax           \n\t"
-      "movb (%%esi,%%ebx,), %%cl   \n\t" // load cl with Prior(x)
-      "movb (%%edx,%%ebx,), %%al   \n\t" // load al with Raw(x-bpp)
-      "addw %%cx, %%ax             \n\t"
-      "incl %%ebx                  \n\t"
-      "shrw %%ax                   \n\t" // divide by 2
-      "addb -1(%%edi,%%ebx,), %%al \n\t" // add Avg(x); -1 to offset inc ebx
-      "cmpl _FullLength, %%ebx     \n\t" // check if at end of array
-      "movb %%al, -1(%%edi,%%ebx,) \n\t" // write back Raw(x) [mov does not
-      "jb avg_lp2                  \n\t" //  affect flags; -1 to offset inc ebx]
-
-   "avg_end:                       \n\t"
-      "EMMS                        \n\t" // end MMX; prep for poss. FP instrs.
-#ifdef __PIC__
-      "popl %%ebx                  \n\t" // restore index to Global Offset Table
-#endif
-
-      : "=c" (dummy_value_c),            // output regs (dummy)
-        "=S" (dummy_value_S),
-        "=D" (dummy_value_D)
-
-      : "0" (bpp),       // ecx          // input regs
-        "1" (prev_row),  // esi
-        "2" (row)        // edi
-
-      : "%eax", "%edx"                   // clobber list
-#ifndef __PIC__
-      , "%ebx"
-#endif
-   );
-
-} /* end png_read_filter_row_mmx_avg() */
-#endif
-
-
-
-#ifdef PNG_THREAD_UNSAFE_OK
-//===========================================================================//
-//                                                                           //
-//         P N G _ R E A D _ F I L T E R _ R O W _ M M X _ P A E T H         //
-//                                                                           //
-//===========================================================================//
-
-// Optimized code for PNG Paeth filter decoder
-
-static void /* PRIVATE */
-png_read_filter_row_mmx_paeth(png_row_infop row_info, png_bytep row,
-                              png_bytep prev_row)
-{
-   int bpp;
-   int dummy_value_c;   // fix 'forbidden register 2 (cx) was spilled' error
-   int dummy_value_S;
-   int dummy_value_D;
-
-   bpp = (row_info->pixel_depth + 7) >> 3; // Get # bytes per pixel
-   _FullLength  = row_info->rowbytes; // # of bytes to filter
-
-   __asm__ __volatile__ (
-#ifdef __PIC__
-      "pushl %%ebx                 \n\t" // save index to Global Offset Table
-#endif
-      "xorl %%ebx, %%ebx           \n\t" // ebx:  x offset
-//pre "movl row, %%edi             \n\t"
-      "xorl %%edx, %%edx           \n\t" // edx:  x-bpp offset
-//pre "movl prev_row, %%esi        \n\t"
-      "xorl %%eax, %%eax           \n\t"
-
-      // Compute the Raw value for the first bpp bytes
-      // Note: the formula works out to be always
-      //   Paeth(x) = Raw(x) + Prior(x)      where x < bpp
-   "paeth_rlp:                     \n\t"
-      "movb (%%edi,%%ebx,), %%al   \n\t"
-      "addb (%%esi,%%ebx,), %%al   \n\t"
-      "incl %%ebx                  \n\t"
-//pre "cmpl bpp, %%ebx             \n\t" (bpp is preloaded into ecx)
-      "cmpl %%ecx, %%ebx           \n\t"
-      "movb %%al, -1(%%edi,%%ebx,) \n\t"
-      "jb paeth_rlp                \n\t"
-      // get # of bytes to alignment
-      "movl %%edi, _dif            \n\t" // take start of row
-      "addl %%ebx, _dif            \n\t" // add bpp
-      "xorl %%ecx, %%ecx           \n\t"
-      "addl $0xf, _dif             \n\t" // add 7 + 8 to incr past alignment
-                                         // boundary
-      "andl $0xfffffff8, _dif      \n\t" // mask to alignment boundary
-      "subl %%edi, _dif            \n\t" // subtract from start ==> value ebx
-                                         // at alignment
-      "jz paeth_go                 \n\t"
-      // fix alignment
-
-   "paeth_lp1:                     \n\t"
-      "xorl %%eax, %%eax           \n\t"
-      // pav = p - a = (a + b - c) - a = b - c
-      "movb (%%esi,%%ebx,), %%al   \n\t" // load Prior(x) into al
-      "movb (%%esi,%%edx,), %%cl   \n\t" // load Prior(x-bpp) into cl
-      "subl %%ecx, %%eax           \n\t" // subtract Prior(x-bpp)
-      "movl %%eax, _patemp         \n\t" // Save pav for later use
-      "xorl %%eax, %%eax           \n\t"
-      // pbv = p - b = (a + b - c) - b = a - c
-      "movb (%%edi,%%edx,), %%al   \n\t" // load Raw(x-bpp) into al
-      "subl %%ecx, %%eax           \n\t" // subtract Prior(x-bpp)
-      "movl %%eax, %%ecx           \n\t"
-      // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-      "addl _patemp, %%eax         \n\t" // pcv = pav + pbv
-      // pc = abs(pcv)
-      "testl $0x80000000, %%eax    \n\t"
-      "jz paeth_pca                \n\t"
-      "negl %%eax                  \n\t" // reverse sign of neg values
-
-   "paeth_pca:                     \n\t"
-      "movl %%eax, _pctemp         \n\t" // save pc for later use
-      // pb = abs(pbv)
-      "testl $0x80000000, %%ecx    \n\t"
-      "jz paeth_pba                \n\t"
-      "negl %%ecx                  \n\t" // reverse sign of neg values
-
-   "paeth_pba:                     \n\t"
-      "movl %%ecx, _pbtemp         \n\t" // save pb for later use
-      // pa = abs(pav)
-      "movl _patemp, %%eax         \n\t"
-      "testl $0x80000000, %%eax    \n\t"
-      "jz paeth_paa                \n\t"
-      "negl %%eax                  \n\t" // reverse sign of neg values
-
-   "paeth_paa:                     \n\t"
-      "movl %%eax, _patemp         \n\t" // save pa for later use
-      // test if pa <= pb
-      "cmpl %%ecx, %%eax           \n\t"
-      "jna paeth_abb               \n\t"
-      // pa > pb; now test if pb <= pc
-      "cmpl _pctemp, %%ecx         \n\t"
-      "jna paeth_bbc               \n\t"
-      // pb > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
-      "movb (%%esi,%%edx,), %%cl   \n\t" // load Prior(x-bpp) into cl
-      "jmp paeth_paeth             \n\t"
-
-   "paeth_bbc:                     \n\t"
-      // pb <= pc; Raw(x) = Paeth(x) + Prior(x)
-      "movb (%%esi,%%ebx,), %%cl   \n\t" // load Prior(x) into cl
-      "jmp paeth_paeth             \n\t"
-
-   "paeth_abb:                     \n\t"
-      // pa <= pb; now test if pa <= pc
-      "cmpl _pctemp, %%eax         \n\t"
-      "jna paeth_abc               \n\t"
-      // pa > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
-      "movb (%%esi,%%edx,), %%cl   \n\t" // load Prior(x-bpp) into cl
-      "jmp paeth_paeth             \n\t"
-
-   "paeth_abc:                     \n\t"
-      // pa <= pc; Raw(x) = Paeth(x) + Raw(x-bpp)
-      "movb (%%edi,%%edx,), %%cl   \n\t" // load Raw(x-bpp) into cl
-
-   "paeth_paeth:                   \n\t"
-      "incl %%ebx                  \n\t"
-      "incl %%edx                  \n\t"
-      // Raw(x) = (Paeth(x) + Paeth_Predictor( a, b, c )) mod 256
-      "addb %%cl, -1(%%edi,%%ebx,) \n\t"
-      "cmpl _dif, %%ebx            \n\t"
-      "jb paeth_lp1                \n\t"
-
-   "paeth_go:                      \n\t"
-      "movl _FullLength, %%ecx     \n\t"
-      "movl %%ecx, %%eax           \n\t"
-      "subl %%ebx, %%eax           \n\t" // subtract alignment fix
-      "andl $0x00000007, %%eax     \n\t" // calc bytes over mult of 8
-      "subl %%eax, %%ecx           \n\t" // drop over bytes from original length
-      "movl %%ecx, _MMXLength      \n\t"
-#ifdef __PIC__
-      "popl %%ebx                  \n\t" // restore index to Global Offset Table
-#endif
-
-      : "=c" (dummy_value_c),            // output regs (dummy)
-        "=S" (dummy_value_S),
-        "=D" (dummy_value_D)
-
-      : "0" (bpp),       // ecx          // input regs
-        "1" (prev_row),  // esi
-        "2" (row)        // edi
-
-      : "%eax", "%edx"                   // clobber list
-#ifndef __PIC__
-      , "%ebx"
-#endif
-   );
-
-   // now do the math for the rest of the row
-   switch (bpp)
-   {
-      case 3:
-      {
-         _ActiveMask.use = 0x0000000000ffffffLL;
-         _ActiveMaskEnd.use = 0xffff000000000000LL;
-         _ShiftBpp.use = 24;    // == bpp(3) * 8
-         _ShiftRem.use = 40;    // == 64 - 24
-
-         __asm__ __volatile__ (
-            "movl _dif, %%ecx            \n\t"
-// preload  "movl row, %%edi             \n\t"
-// preload  "movl prev_row, %%esi        \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%ecx,), %%mm1 \n\t"
-         "paeth_3lp:                     \n\t"
-            "psrlq _ShiftRem, %%mm1      \n\t" // shift last 3 bytes to 1st
-                                               // 3 bytes
-            "movq (%%esi,%%ecx,), %%mm2  \n\t" // load b=Prior(x)
-            "punpcklbw %%mm0, %%mm1      \n\t" // unpack High bytes of a
-            "movq -8(%%esi,%%ecx,), %%mm3 \n\t" // prep c=Prior(x-bpp) bytes
-            "punpcklbw %%mm0, %%mm2      \n\t" // unpack High bytes of b
-            "psrlq _ShiftRem, %%mm3      \n\t" // shift last 3 bytes to 1st
-                                               // 3 bytes
-            // pav = p - a = (a + b - c) - a = b - c
-            "movq %%mm2, %%mm4           \n\t"
-            "punpcklbw %%mm0, %%mm3      \n\t" // unpack High bytes of c
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movq %%mm1, %%mm5           \n\t"
-            "psubw %%mm3, %%mm4          \n\t"
-            "pxor %%mm7, %%mm7           \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-            "movq %%mm4, %%mm6           \n\t"
-            "psubw %%mm3, %%mm5          \n\t"
-
-            // pa = abs(p-a) = abs(pav)
-            // pb = abs(p-b) = abs(pbv)
-            // pc = abs(p-c) = abs(pcv)
-            "pcmpgtw %%mm4, %%mm0        \n\t" // create mask pav bytes < 0
-            "paddw %%mm5, %%mm6          \n\t"
-            "pand %%mm4, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "pcmpgtw %%mm5, %%mm7        \n\t" // create mask pbv bytes < 0
-            "psubw %%mm0, %%mm4          \n\t"
-            "pand %%mm5, %%mm7           \n\t" // only pbv bytes < 0 in mm0
-            "psubw %%mm0, %%mm4          \n\t"
-            "psubw %%mm7, %%mm5          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "pcmpgtw %%mm6, %%mm0        \n\t" // create mask pcv bytes < 0
-            "pand %%mm6, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm7, %%mm5          \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            //  test pa <= pb
-            "movq %%mm4, %%mm7           \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            "pcmpgtw %%mm5, %%mm7        \n\t" // pa > pb?
-            "movq %%mm7, %%mm0           \n\t"
-            // use mm7 mask to merge pa & pb
-            "pand %%mm7, %%mm5           \n\t"
-            // use mm0 mask copy to merge a & b
-            "pand %%mm0, %%mm2           \n\t"
-            "pandn %%mm4, %%mm7          \n\t"
-            "pandn %%mm1, %%mm0          \n\t"
-            "paddw %%mm5, %%mm7          \n\t"
-            "paddw %%mm2, %%mm0          \n\t"
-            //  test  ((pa <= pb)? pa:pb) <= pc
-            "pcmpgtw %%mm6, %%mm7        \n\t" // pab > pc?
-            "pxor %%mm1, %%mm1           \n\t"
-            "pand %%mm7, %%mm3           \n\t"
-            "pandn %%mm0, %%mm7          \n\t"
-            "paddw %%mm3, %%mm7          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "packuswb %%mm1, %%mm7       \n\t"
-            "movq (%%esi,%%ecx,), %%mm3  \n\t" // load c=Prior(x-bpp)
-            "pand _ActiveMask, %%mm7     \n\t"
-            "movq %%mm3, %%mm2           \n\t" // load b=Prior(x) step 1
-            "paddb (%%edi,%%ecx,), %%mm7 \n\t" // add Paeth predictor with Raw(x)
-            "punpcklbw %%mm0, %%mm3      \n\t" // unpack High bytes of c
-            "movq %%mm7, (%%edi,%%ecx,)  \n\t" // write back updated value
-            "movq %%mm7, %%mm1           \n\t" // now mm1 will be used as
-                                               // Raw(x-bpp)
-            // now do Paeth for 2nd set of bytes (3-5)
-            "psrlq _ShiftBpp, %%mm2      \n\t" // load b=Prior(x) step 2
-            "punpcklbw %%mm0, %%mm1      \n\t" // unpack High bytes of a
-            "pxor %%mm7, %%mm7           \n\t"
-            "punpcklbw %%mm0, %%mm2      \n\t" // unpack High bytes of b
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movq %%mm1, %%mm5           \n\t"
-            // pav = p - a = (a + b - c) - a = b - c
-            "movq %%mm2, %%mm4           \n\t"
-            "psubw %%mm3, %%mm5          \n\t"
-            "psubw %%mm3, %%mm4          \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) =
-            //       pav + pbv = pbv + pav
-            "movq %%mm5, %%mm6           \n\t"
-            "paddw %%mm4, %%mm6          \n\t"
-
-            // pa = abs(p-a) = abs(pav)
-            // pb = abs(p-b) = abs(pbv)
-            // pc = abs(p-c) = abs(pcv)
-            "pcmpgtw %%mm5, %%mm0        \n\t" // create mask pbv bytes < 0
-            "pcmpgtw %%mm4, %%mm7        \n\t" // create mask pav bytes < 0
-            "pand %%mm5, %%mm0           \n\t" // only pbv bytes < 0 in mm0
-            "pand %%mm4, %%mm7           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm0, %%mm5          \n\t"
-            "psubw %%mm7, %%mm4          \n\t"
-            "psubw %%mm0, %%mm5          \n\t"
-            "psubw %%mm7, %%mm4          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "pcmpgtw %%mm6, %%mm0        \n\t" // create mask pcv bytes < 0
-            "pand %%mm6, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm0, %%mm6          \n\t"
-            //  test pa <= pb
-            "movq %%mm4, %%mm7           \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            "pcmpgtw %%mm5, %%mm7        \n\t" // pa > pb?
-            "movq %%mm7, %%mm0           \n\t"
-            // use mm7 mask to merge pa & pb
-            "pand %%mm7, %%mm5           \n\t"
-            // use mm0 mask copy to merge a & b
-            "pand %%mm0, %%mm2           \n\t"
-            "pandn %%mm4, %%mm7          \n\t"
-            "pandn %%mm1, %%mm0          \n\t"
-            "paddw %%mm5, %%mm7          \n\t"
-            "paddw %%mm2, %%mm0          \n\t"
-            //  test  ((pa <= pb)? pa:pb) <= pc
-            "pcmpgtw %%mm6, %%mm7        \n\t" // pab > pc?
-            "movq (%%esi,%%ecx,), %%mm2  \n\t" // load b=Prior(x)
-            "pand %%mm7, %%mm3           \n\t"
-            "pandn %%mm0, %%mm7          \n\t"
-            "pxor %%mm1, %%mm1           \n\t"
-            "paddw %%mm3, %%mm7          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "packuswb %%mm1, %%mm7       \n\t"
-            "movq %%mm2, %%mm3           \n\t" // load c=Prior(x-bpp) step 1
-            "pand _ActiveMask, %%mm7     \n\t"
-            "punpckhbw %%mm0, %%mm2      \n\t" // unpack High bytes of b
-            "psllq _ShiftBpp, %%mm7      \n\t" // shift bytes to 2nd group of
-                                               // 3 bytes
-             // pav = p - a = (a + b - c) - a = b - c
-            "movq %%mm2, %%mm4           \n\t"
-            "paddb (%%edi,%%ecx,), %%mm7 \n\t" // add Paeth predictor with Raw(x)
-            "psllq _ShiftBpp, %%mm3      \n\t" // load c=Prior(x-bpp) step 2
-            "movq %%mm7, (%%edi,%%ecx,)  \n\t" // write back updated value
-            "movq %%mm7, %%mm1           \n\t"
-            "punpckhbw %%mm0, %%mm3      \n\t" // unpack High bytes of c
-            "psllq _ShiftBpp, %%mm1      \n\t" // shift bytes
-                                    // now mm1 will be used as Raw(x-bpp)
-            // now do Paeth for 3rd, and final, set of bytes (6-7)
-            "pxor %%mm7, %%mm7           \n\t"
-            "punpckhbw %%mm0, %%mm1      \n\t" // unpack High bytes of a
-            "psubw %%mm3, %%mm4          \n\t"
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movq %%mm1, %%mm5           \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-            "movq %%mm4, %%mm6           \n\t"
-            "psubw %%mm3, %%mm5          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "paddw %%mm5, %%mm6          \n\t"
-
-            // pa = abs(p-a) = abs(pav)
-            // pb = abs(p-b) = abs(pbv)
-            // pc = abs(p-c) = abs(pcv)
-            "pcmpgtw %%mm4, %%mm0        \n\t" // create mask pav bytes < 0
-            "pcmpgtw %%mm5, %%mm7        \n\t" // create mask pbv bytes < 0
-            "pand %%mm4, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "pand %%mm5, %%mm7           \n\t" // only pbv bytes < 0 in mm0
-            "psubw %%mm0, %%mm4          \n\t"
-            "psubw %%mm7, %%mm5          \n\t"
-            "psubw %%mm0, %%mm4          \n\t"
-            "psubw %%mm7, %%mm5          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "pcmpgtw %%mm6, %%mm0        \n\t" // create mask pcv bytes < 0
-            "pand %%mm6, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm0, %%mm6          \n\t"
-            //  test pa <= pb
-            "movq %%mm4, %%mm7           \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            "pcmpgtw %%mm5, %%mm7        \n\t" // pa > pb?
-            "movq %%mm7, %%mm0           \n\t"
-            // use mm0 mask copy to merge a & b
-            "pand %%mm0, %%mm2           \n\t"
-            // use mm7 mask to merge pa & pb
-            "pand %%mm7, %%mm5           \n\t"
-            "pandn %%mm1, %%mm0          \n\t"
-            "pandn %%mm4, %%mm7          \n\t"
-            "paddw %%mm2, %%mm0          \n\t"
-            "paddw %%mm5, %%mm7          \n\t"
-            //  test  ((pa <= pb)? pa:pb) <= pc
-            "pcmpgtw %%mm6, %%mm7        \n\t" // pab > pc?
-            "pand %%mm7, %%mm3           \n\t"
-            "pandn %%mm0, %%mm7          \n\t"
-            "paddw %%mm3, %%mm7          \n\t"
-            "pxor %%mm1, %%mm1           \n\t"
-            "packuswb %%mm7, %%mm1       \n\t"
-            // step ecx to next set of 8 bytes and repeat loop til done
-            "addl $8, %%ecx              \n\t"
-            "pand _ActiveMaskEnd, %%mm1  \n\t"
-            "paddb -8(%%edi,%%ecx,), %%mm1 \n\t" // add Paeth predictor with
-                                                 // Raw(x)
-
-            "cmpl _MMXLength, %%ecx      \n\t"
-            "pxor %%mm0, %%mm0           \n\t" // pxor does not affect flags
-            "movq %%mm1, -8(%%edi,%%ecx,) \n\t" // write back updated value
-                                 // mm1 will be used as Raw(x-bpp) next loop
-                           // mm3 ready to be used as Prior(x-bpp) next loop
-            "jb paeth_3lp                \n\t"
-
-            : "=S" (dummy_value_S),             // output regs (dummy)
-              "=D" (dummy_value_D)
-
-            : "0" (prev_row),  // esi           // input regs
-              "1" (row)        // edi
-
-            : "%ecx"                            // clobber list
-#if 0  /* %mm0, ..., %mm7 not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm2", "%mm3"
-            , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;  // end 3 bpp
-
-      case 6:
-      //case 7:   // GRR BOGUS
-      //case 5:   // GRR BOGUS
-      {
-         _ActiveMask.use  = 0x00000000ffffffffLL;
-         _ActiveMask2.use = 0xffffffff00000000LL;
-         _ShiftBpp.use = bpp << 3;    // == bpp * 8
-         _ShiftRem.use = 64 - _ShiftBpp.use;
-
-         __asm__ __volatile__ (
-            "movl _dif, %%ecx            \n\t"
-// preload  "movl row, %%edi             \n\t"
-// preload  "movl prev_row, %%esi        \n\t"
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%ecx,), %%mm1 \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-
-         "paeth_6lp:                     \n\t"
-            // must shift to position Raw(x-bpp) data
-            "psrlq _ShiftRem, %%mm1      \n\t"
-            // do first set of 4 bytes
-            "movq -8(%%esi,%%ecx,), %%mm3 \n\t" // read c=Prior(x-bpp) bytes
-            "punpcklbw %%mm0, %%mm1      \n\t" // unpack Low bytes of a
-            "movq (%%esi,%%ecx,), %%mm2  \n\t" // load b=Prior(x)
-            "punpcklbw %%mm0, %%mm2      \n\t" // unpack Low bytes of b
-            // must shift to position Prior(x-bpp) data
-            "psrlq _ShiftRem, %%mm3      \n\t"
-            // pav = p - a = (a + b - c) - a = b - c
-            "movq %%mm2, %%mm4           \n\t"
-            "punpcklbw %%mm0, %%mm3      \n\t" // unpack Low bytes of c
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movq %%mm1, %%mm5           \n\t"
-            "psubw %%mm3, %%mm4          \n\t"
-            "pxor %%mm7, %%mm7           \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-            "movq %%mm4, %%mm6           \n\t"
-            "psubw %%mm3, %%mm5          \n\t"
-            // pa = abs(p-a) = abs(pav)
-            // pb = abs(p-b) = abs(pbv)
-            // pc = abs(p-c) = abs(pcv)
-            "pcmpgtw %%mm4, %%mm0        \n\t" // create mask pav bytes < 0
-            "paddw %%mm5, %%mm6          \n\t"
-            "pand %%mm4, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "pcmpgtw %%mm5, %%mm7        \n\t" // create mask pbv bytes < 0
-            "psubw %%mm0, %%mm4          \n\t"
-            "pand %%mm5, %%mm7           \n\t" // only pbv bytes < 0 in mm0
-            "psubw %%mm0, %%mm4          \n\t"
-            "psubw %%mm7, %%mm5          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "pcmpgtw %%mm6, %%mm0        \n\t" // create mask pcv bytes < 0
-            "pand %%mm6, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm7, %%mm5          \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            //  test pa <= pb
-            "movq %%mm4, %%mm7           \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            "pcmpgtw %%mm5, %%mm7        \n\t" // pa > pb?
-            "movq %%mm7, %%mm0           \n\t"
-            // use mm7 mask to merge pa & pb
-            "pand %%mm7, %%mm5           \n\t"
-            // use mm0 mask copy to merge a & b
-            "pand %%mm0, %%mm2           \n\t"
-            "pandn %%mm4, %%mm7          \n\t"
-            "pandn %%mm1, %%mm0          \n\t"
-            "paddw %%mm5, %%mm7          \n\t"
-            "paddw %%mm2, %%mm0          \n\t"
-            //  test  ((pa <= pb)? pa:pb) <= pc
-            "pcmpgtw %%mm6, %%mm7        \n\t" // pab > pc?
-            "pxor %%mm1, %%mm1           \n\t"
-            "pand %%mm7, %%mm3           \n\t"
-            "pandn %%mm0, %%mm7          \n\t"
-            "paddw %%mm3, %%mm7          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "packuswb %%mm1, %%mm7       \n\t"
-            "movq -8(%%esi,%%ecx,), %%mm3 \n\t" // load c=Prior(x-bpp)
-            "pand _ActiveMask, %%mm7     \n\t"
-            "psrlq _ShiftRem, %%mm3      \n\t"
-            "movq (%%esi,%%ecx,), %%mm2  \n\t" // load b=Prior(x) step 1
-            "paddb (%%edi,%%ecx,), %%mm7 \n\t" // add Paeth predictor and Raw(x)
-            "movq %%mm2, %%mm6           \n\t"
-            "movq %%mm7, (%%edi,%%ecx,)  \n\t" // write back updated value
-            "movq -8(%%edi,%%ecx,), %%mm1 \n\t"
-            "psllq _ShiftBpp, %%mm6      \n\t"
-            "movq %%mm7, %%mm5           \n\t"
-            "psrlq _ShiftRem, %%mm1      \n\t"
-            "por %%mm6, %%mm3            \n\t"
-            "psllq _ShiftBpp, %%mm5      \n\t"
-            "punpckhbw %%mm0, %%mm3      \n\t" // unpack High bytes of c
-            "por %%mm5, %%mm1            \n\t"
-            // do second set of 4 bytes
-            "punpckhbw %%mm0, %%mm2      \n\t" // unpack High bytes of b
-            "punpckhbw %%mm0, %%mm1      \n\t" // unpack High bytes of a
-            // pav = p - a = (a + b - c) - a = b - c
-            "movq %%mm2, %%mm4           \n\t"
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movq %%mm1, %%mm5           \n\t"
-            "psubw %%mm3, %%mm4          \n\t"
-            "pxor %%mm7, %%mm7           \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-            "movq %%mm4, %%mm6           \n\t"
-            "psubw %%mm3, %%mm5          \n\t"
-            // pa = abs(p-a) = abs(pav)
-            // pb = abs(p-b) = abs(pbv)
-            // pc = abs(p-c) = abs(pcv)
-            "pcmpgtw %%mm4, %%mm0        \n\t" // create mask pav bytes < 0
-            "paddw %%mm5, %%mm6          \n\t"
-            "pand %%mm4, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "pcmpgtw %%mm5, %%mm7        \n\t" // create mask pbv bytes < 0
-            "psubw %%mm0, %%mm4          \n\t"
-            "pand %%mm5, %%mm7           \n\t" // only pbv bytes < 0 in mm0
-            "psubw %%mm0, %%mm4          \n\t"
-            "psubw %%mm7, %%mm5          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "pcmpgtw %%mm6, %%mm0        \n\t" // create mask pcv bytes < 0
-            "pand %%mm6, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm7, %%mm5          \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            //  test pa <= pb
-            "movq %%mm4, %%mm7           \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            "pcmpgtw %%mm5, %%mm7        \n\t" // pa > pb?
-            "movq %%mm7, %%mm0           \n\t"
-            // use mm7 mask to merge pa & pb
-            "pand %%mm7, %%mm5           \n\t"
-            // use mm0 mask copy to merge a & b
-            "pand %%mm0, %%mm2           \n\t"
-            "pandn %%mm4, %%mm7          \n\t"
-            "pandn %%mm1, %%mm0          \n\t"
-            "paddw %%mm5, %%mm7          \n\t"
-            "paddw %%mm2, %%mm0          \n\t"
-            //  test  ((pa <= pb)? pa:pb) <= pc
-            "pcmpgtw %%mm6, %%mm7        \n\t" // pab > pc?
-            "pxor %%mm1, %%mm1           \n\t"
-            "pand %%mm7, %%mm3           \n\t"
-            "pandn %%mm0, %%mm7          \n\t"
-            "pxor %%mm1, %%mm1           \n\t"
-            "paddw %%mm3, %%mm7          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            // step ecx to next set of 8 bytes and repeat loop til done
-            "addl $8, %%ecx              \n\t"
-            "packuswb %%mm7, %%mm1       \n\t"
-            "paddb -8(%%edi,%%ecx,), %%mm1 \n\t" // add Paeth predictor with Raw(x)
-            "cmpl _MMXLength, %%ecx      \n\t"
-            "movq %%mm1, -8(%%edi,%%ecx,) \n\t" // write back updated value
-                                // mm1 will be used as Raw(x-bpp) next loop
-            "jb paeth_6lp                \n\t"
-
-            : "=S" (dummy_value_S),             // output regs (dummy)
-              "=D" (dummy_value_D)
-
-            : "0" (prev_row),  // esi           // input regs
-              "1" (row)        // edi
-
-            : "%ecx"                            // clobber list
-#if 0  /* %mm0, ..., %mm7 not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm2", "%mm3"
-            , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;  // end 6 bpp
-
-      case 4:
-      {
-         _ActiveMask.use  = 0x00000000ffffffffLL;
-
-         __asm__ __volatile__ (
-            "movl _dif, %%ecx            \n\t"
-// preload  "movl row, %%edi             \n\t"
-// preload  "movl prev_row, %%esi        \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%ecx,), %%mm1 \n\t" // only time should need to read
-                                     //  a=Raw(x-bpp) bytes
-         "paeth_4lp:                     \n\t"
-            // do first set of 4 bytes
-            "movq -8(%%esi,%%ecx,), %%mm3 \n\t" // read c=Prior(x-bpp) bytes
-            "punpckhbw %%mm0, %%mm1      \n\t" // unpack Low bytes of a
-            "movq (%%esi,%%ecx,), %%mm2  \n\t" // load b=Prior(x)
-            "punpcklbw %%mm0, %%mm2      \n\t" // unpack High bytes of b
-            // pav = p - a = (a + b - c) - a = b - c
-            "movq %%mm2, %%mm4           \n\t"
-            "punpckhbw %%mm0, %%mm3      \n\t" // unpack High bytes of c
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movq %%mm1, %%mm5           \n\t"
-            "psubw %%mm3, %%mm4          \n\t"
-            "pxor %%mm7, %%mm7           \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-            "movq %%mm4, %%mm6           \n\t"
-            "psubw %%mm3, %%mm5          \n\t"
-            // pa = abs(p-a) = abs(pav)
-            // pb = abs(p-b) = abs(pbv)
-            // pc = abs(p-c) = abs(pcv)
-            "pcmpgtw %%mm4, %%mm0        \n\t" // create mask pav bytes < 0
-            "paddw %%mm5, %%mm6          \n\t"
-            "pand %%mm4, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "pcmpgtw %%mm5, %%mm7        \n\t" // create mask pbv bytes < 0
-            "psubw %%mm0, %%mm4          \n\t"
-            "pand %%mm5, %%mm7           \n\t" // only pbv bytes < 0 in mm0
-            "psubw %%mm0, %%mm4          \n\t"
-            "psubw %%mm7, %%mm5          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "pcmpgtw %%mm6, %%mm0        \n\t" // create mask pcv bytes < 0
-            "pand %%mm6, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm7, %%mm5          \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            //  test pa <= pb
-            "movq %%mm4, %%mm7           \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            "pcmpgtw %%mm5, %%mm7        \n\t" // pa > pb?
-            "movq %%mm7, %%mm0           \n\t"
-            // use mm7 mask to merge pa & pb
-            "pand %%mm7, %%mm5           \n\t"
-            // use mm0 mask copy to merge a & b
-            "pand %%mm0, %%mm2           \n\t"
-            "pandn %%mm4, %%mm7          \n\t"
-            "pandn %%mm1, %%mm0          \n\t"
-            "paddw %%mm5, %%mm7          \n\t"
-            "paddw %%mm2, %%mm0          \n\t"
-            //  test  ((pa <= pb)? pa:pb) <= pc
-            "pcmpgtw %%mm6, %%mm7        \n\t" // pab > pc?
-            "pxor %%mm1, %%mm1           \n\t"
-            "pand %%mm7, %%mm3           \n\t"
-            "pandn %%mm0, %%mm7          \n\t"
-            "paddw %%mm3, %%mm7          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "packuswb %%mm1, %%mm7       \n\t"
-            "movq (%%esi,%%ecx,), %%mm3  \n\t" // load c=Prior(x-bpp)
-            "pand _ActiveMask, %%mm7     \n\t"
-            "movq %%mm3, %%mm2           \n\t" // load b=Prior(x) step 1
-            "paddb (%%edi,%%ecx,), %%mm7 \n\t" // add Paeth predictor with Raw(x)
-            "punpcklbw %%mm0, %%mm3      \n\t" // unpack High bytes of c
-            "movq %%mm7, (%%edi,%%ecx,)  \n\t" // write back updated value
-            "movq %%mm7, %%mm1           \n\t" // now mm1 will be used as Raw(x-bpp)
-            // do second set of 4 bytes
-            "punpckhbw %%mm0, %%mm2      \n\t" // unpack Low bytes of b
-            "punpcklbw %%mm0, %%mm1      \n\t" // unpack Low bytes of a
-            // pav = p - a = (a + b - c) - a = b - c
-            "movq %%mm2, %%mm4           \n\t"
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movq %%mm1, %%mm5           \n\t"
-            "psubw %%mm3, %%mm4          \n\t"
-            "pxor %%mm7, %%mm7           \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-            "movq %%mm4, %%mm6           \n\t"
-            "psubw %%mm3, %%mm5          \n\t"
-            // pa = abs(p-a) = abs(pav)
-            // pb = abs(p-b) = abs(pbv)
-            // pc = abs(p-c) = abs(pcv)
-            "pcmpgtw %%mm4, %%mm0        \n\t" // create mask pav bytes < 0
-            "paddw %%mm5, %%mm6          \n\t"
-            "pand %%mm4, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "pcmpgtw %%mm5, %%mm7        \n\t" // create mask pbv bytes < 0
-            "psubw %%mm0, %%mm4          \n\t"
-            "pand %%mm5, %%mm7           \n\t" // only pbv bytes < 0 in mm0
-            "psubw %%mm0, %%mm4          \n\t"
-            "psubw %%mm7, %%mm5          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "pcmpgtw %%mm6, %%mm0        \n\t" // create mask pcv bytes < 0
-            "pand %%mm6, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm7, %%mm5          \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            //  test pa <= pb
-            "movq %%mm4, %%mm7           \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            "pcmpgtw %%mm5, %%mm7        \n\t" // pa > pb?
-            "movq %%mm7, %%mm0           \n\t"
-            // use mm7 mask to merge pa & pb
-            "pand %%mm7, %%mm5           \n\t"
-            // use mm0 mask copy to merge a & b
-            "pand %%mm0, %%mm2           \n\t"
-            "pandn %%mm4, %%mm7          \n\t"
-            "pandn %%mm1, %%mm0          \n\t"
-            "paddw %%mm5, %%mm7          \n\t"
-            "paddw %%mm2, %%mm0          \n\t"
-            //  test  ((pa <= pb)? pa:pb) <= pc
-            "pcmpgtw %%mm6, %%mm7        \n\t" // pab > pc?
-            "pxor %%mm1, %%mm1           \n\t"
-            "pand %%mm7, %%mm3           \n\t"
-            "pandn %%mm0, %%mm7          \n\t"
-            "pxor %%mm1, %%mm1           \n\t"
-            "paddw %%mm3, %%mm7          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            // step ecx to next set of 8 bytes and repeat loop til done
-            "addl $8, %%ecx              \n\t"
-            "packuswb %%mm7, %%mm1       \n\t"
-            "paddb -8(%%edi,%%ecx,), %%mm1 \n\t" // add predictor with Raw(x)
-            "cmpl _MMXLength, %%ecx      \n\t"
-            "movq %%mm1, -8(%%edi,%%ecx,) \n\t" // write back updated value
-                                // mm1 will be used as Raw(x-bpp) next loop
-            "jb paeth_4lp                \n\t"
-
-            : "=S" (dummy_value_S),             // output regs (dummy)
-              "=D" (dummy_value_D)
-
-            : "0" (prev_row),  // esi           // input regs
-              "1" (row)        // edi
-
-            : "%ecx"                            // clobber list
-#if 0  /* %mm0, ..., %mm7 not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm2", "%mm3"
-            , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;  // end 4 bpp
-
-      case 8:                          // bpp == 8
-      {
-         _ActiveMask.use  = 0x00000000ffffffffLL;
-
-         __asm__ __volatile__ (
-            "movl _dif, %%ecx            \n\t"
-// preload  "movl row, %%edi             \n\t"
-// preload  "movl prev_row, %%esi        \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%ecx,), %%mm1 \n\t" // only time should need to read
-                                       //  a=Raw(x-bpp) bytes
-         "paeth_8lp:                     \n\t"
-            // do first set of 4 bytes
-            "movq -8(%%esi,%%ecx,), %%mm3 \n\t" // read c=Prior(x-bpp) bytes
-            "punpcklbw %%mm0, %%mm1      \n\t" // unpack Low bytes of a
-            "movq (%%esi,%%ecx,), %%mm2  \n\t" // load b=Prior(x)
-            "punpcklbw %%mm0, %%mm2      \n\t" // unpack Low bytes of b
-            // pav = p - a = (a + b - c) - a = b - c
-            "movq %%mm2, %%mm4           \n\t"
-            "punpcklbw %%mm0, %%mm3      \n\t" // unpack Low bytes of c
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movq %%mm1, %%mm5           \n\t"
-            "psubw %%mm3, %%mm4          \n\t"
-            "pxor %%mm7, %%mm7           \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-            "movq %%mm4, %%mm6           \n\t"
-            "psubw %%mm3, %%mm5          \n\t"
-            // pa = abs(p-a) = abs(pav)
-            // pb = abs(p-b) = abs(pbv)
-            // pc = abs(p-c) = abs(pcv)
-            "pcmpgtw %%mm4, %%mm0        \n\t" // create mask pav bytes < 0
-            "paddw %%mm5, %%mm6          \n\t"
-            "pand %%mm4, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "pcmpgtw %%mm5, %%mm7        \n\t" // create mask pbv bytes < 0
-            "psubw %%mm0, %%mm4          \n\t"
-            "pand %%mm5, %%mm7           \n\t" // only pbv bytes < 0 in mm0
-            "psubw %%mm0, %%mm4          \n\t"
-            "psubw %%mm7, %%mm5          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "pcmpgtw %%mm6, %%mm0        \n\t" // create mask pcv bytes < 0
-            "pand %%mm6, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm7, %%mm5          \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            //  test pa <= pb
-            "movq %%mm4, %%mm7           \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            "pcmpgtw %%mm5, %%mm7        \n\t" // pa > pb?
-            "movq %%mm7, %%mm0           \n\t"
-            // use mm7 mask to merge pa & pb
-            "pand %%mm7, %%mm5           \n\t"
-            // use mm0 mask copy to merge a & b
-            "pand %%mm0, %%mm2           \n\t"
-            "pandn %%mm4, %%mm7          \n\t"
-            "pandn %%mm1, %%mm0          \n\t"
-            "paddw %%mm5, %%mm7          \n\t"
-            "paddw %%mm2, %%mm0          \n\t"
-            //  test  ((pa <= pb)? pa:pb) <= pc
-            "pcmpgtw %%mm6, %%mm7        \n\t" // pab > pc?
-            "pxor %%mm1, %%mm1           \n\t"
-            "pand %%mm7, %%mm3           \n\t"
-            "pandn %%mm0, %%mm7          \n\t"
-            "paddw %%mm3, %%mm7          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "packuswb %%mm1, %%mm7       \n\t"
-            "movq -8(%%esi,%%ecx,), %%mm3 \n\t" // read c=Prior(x-bpp) bytes
-            "pand _ActiveMask, %%mm7     \n\t"
-            "movq (%%esi,%%ecx,), %%mm2  \n\t" // load b=Prior(x)
-            "paddb (%%edi,%%ecx,), %%mm7 \n\t" // add Paeth predictor with Raw(x)
-            "punpckhbw %%mm0, %%mm3      \n\t" // unpack High bytes of c
-            "movq %%mm7, (%%edi,%%ecx,)  \n\t" // write back updated value
-            "movq -8(%%edi,%%ecx,), %%mm1 \n\t" // read a=Raw(x-bpp) bytes
-
-            // do second set of 4 bytes
-            "punpckhbw %%mm0, %%mm2      \n\t" // unpack High bytes of b
-            "punpckhbw %%mm0, %%mm1      \n\t" // unpack High bytes of a
-            // pav = p - a = (a + b - c) - a = b - c
-            "movq %%mm2, %%mm4           \n\t"
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movq %%mm1, %%mm5           \n\t"
-            "psubw %%mm3, %%mm4          \n\t"
-            "pxor %%mm7, %%mm7           \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-            "movq %%mm4, %%mm6           \n\t"
-            "psubw %%mm3, %%mm5          \n\t"
-            // pa = abs(p-a) = abs(pav)
-            // pb = abs(p-b) = abs(pbv)
-            // pc = abs(p-c) = abs(pcv)
-            "pcmpgtw %%mm4, %%mm0        \n\t" // create mask pav bytes < 0
-            "paddw %%mm5, %%mm6          \n\t"
-            "pand %%mm4, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "pcmpgtw %%mm5, %%mm7        \n\t" // create mask pbv bytes < 0
-            "psubw %%mm0, %%mm4          \n\t"
-            "pand %%mm5, %%mm7           \n\t" // only pbv bytes < 0 in mm0
-            "psubw %%mm0, %%mm4          \n\t"
-            "psubw %%mm7, %%mm5          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            "pcmpgtw %%mm6, %%mm0        \n\t" // create mask pcv bytes < 0
-            "pand %%mm6, %%mm0           \n\t" // only pav bytes < 0 in mm7
-            "psubw %%mm7, %%mm5          \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            //  test pa <= pb
-            "movq %%mm4, %%mm7           \n\t"
-            "psubw %%mm0, %%mm6          \n\t"
-            "pcmpgtw %%mm5, %%mm7        \n\t" // pa > pb?
-            "movq %%mm7, %%mm0           \n\t"
-            // use mm7 mask to merge pa & pb
-            "pand %%mm7, %%mm5           \n\t"
-            // use mm0 mask copy to merge a & b
-            "pand %%mm0, %%mm2           \n\t"
-            "pandn %%mm4, %%mm7          \n\t"
-            "pandn %%mm1, %%mm0          \n\t"
-            "paddw %%mm5, %%mm7          \n\t"
-            "paddw %%mm2, %%mm0          \n\t"
-            //  test  ((pa <= pb)? pa:pb) <= pc
-            "pcmpgtw %%mm6, %%mm7        \n\t" // pab > pc?
-            "pxor %%mm1, %%mm1           \n\t"
-            "pand %%mm7, %%mm3           \n\t"
-            "pandn %%mm0, %%mm7          \n\t"
-            "pxor %%mm1, %%mm1           \n\t"
-            "paddw %%mm3, %%mm7          \n\t"
-            "pxor %%mm0, %%mm0           \n\t"
-            // step ecx to next set of 8 bytes and repeat loop til done
-            "addl $8, %%ecx              \n\t"
-            "packuswb %%mm7, %%mm1       \n\t"
-            "paddb -8(%%edi,%%ecx,), %%mm1 \n\t" // add Paeth predictor with Raw(x)
-            "cmpl _MMXLength, %%ecx      \n\t"
-            "movq %%mm1, -8(%%edi,%%ecx,) \n\t" // write back updated value
-                            // mm1 will be used as Raw(x-bpp) next loop
-            "jb paeth_8lp                \n\t"
-
-            : "=S" (dummy_value_S),             // output regs (dummy)
-              "=D" (dummy_value_D)
-
-            : "0" (prev_row),  // esi           // input regs
-              "1" (row)        // edi
-
-            : "%ecx"                            // clobber list
-#if 0  /* %mm0, ..., %mm7 not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm2", "%mm3"
-            , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;  // end 8 bpp
-
-      case 1:                // bpp = 1
-      case 2:                // bpp = 2
-      default:               // bpp > 8
-      {
-         __asm__ __volatile__ (
-#ifdef __PIC__
-            "pushl %%ebx                 \n\t" // save Global Offset Table index
-#endif
-            "movl _dif, %%ebx            \n\t"
-            "cmpl _FullLength, %%ebx     \n\t"
-            "jnb paeth_dend              \n\t"
-
-// preload  "movl row, %%edi             \n\t"
-// preload  "movl prev_row, %%esi        \n\t"
-            // do Paeth decode for remaining bytes
-            "movl %%ebx, %%edx           \n\t"
-// preload  "subl bpp, %%edx             \n\t" // (bpp is preloaded into ecx)
-            "subl %%ecx, %%edx           \n\t" // edx = ebx - bpp
-            "xorl %%ecx, %%ecx           \n\t" // zero ecx before using cl & cx
-
-         "paeth_dlp:                     \n\t"
-            "xorl %%eax, %%eax           \n\t"
-            // pav = p - a = (a + b - c) - a = b - c
-            "movb (%%esi,%%ebx,), %%al   \n\t" // load Prior(x) into al
-            "movb (%%esi,%%edx,), %%cl   \n\t" // load Prior(x-bpp) into cl
-            "subl %%ecx, %%eax           \n\t" // subtract Prior(x-bpp)
-            "movl %%eax, _patemp         \n\t" // Save pav for later use
-            "xorl %%eax, %%eax           \n\t"
-            // pbv = p - b = (a + b - c) - b = a - c
-            "movb (%%edi,%%edx,), %%al   \n\t" // load Raw(x-bpp) into al
-            "subl %%ecx, %%eax           \n\t" // subtract Prior(x-bpp)
-            "movl %%eax, %%ecx           \n\t"
-            // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-            "addl _patemp, %%eax         \n\t" // pcv = pav + pbv
-            // pc = abs(pcv)
-            "testl $0x80000000, %%eax    \n\t"
-            "jz paeth_dpca               \n\t"
-            "negl %%eax                  \n\t" // reverse sign of neg values
-
-         "paeth_dpca:                    \n\t"
-            "movl %%eax, _pctemp         \n\t" // save pc for later use
-            // pb = abs(pbv)
-            "testl $0x80000000, %%ecx    \n\t"
-            "jz paeth_dpba               \n\t"
-            "negl %%ecx                  \n\t" // reverse sign of neg values
-
-         "paeth_dpba:                    \n\t"
-            "movl %%ecx, _pbtemp         \n\t" // save pb for later use
-            // pa = abs(pav)
-            "movl _patemp, %%eax         \n\t"
-            "testl $0x80000000, %%eax    \n\t"
-            "jz paeth_dpaa               \n\t"
-            "negl %%eax                  \n\t" // reverse sign of neg values
-
-         "paeth_dpaa:                    \n\t"
-            "movl %%eax, _patemp         \n\t" // save pa for later use
-            // test if pa <= pb
-            "cmpl %%ecx, %%eax           \n\t"
-            "jna paeth_dabb              \n\t"
-            // pa > pb; now test if pb <= pc
-            "cmpl _pctemp, %%ecx         \n\t"
-            "jna paeth_dbbc              \n\t"
-            // pb > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
-            "movb (%%esi,%%edx,), %%cl   \n\t" // load Prior(x-bpp) into cl
-            "jmp paeth_dpaeth            \n\t"
-
-         "paeth_dbbc:                    \n\t"
-            // pb <= pc; Raw(x) = Paeth(x) + Prior(x)
-            "movb (%%esi,%%ebx,), %%cl   \n\t" // load Prior(x) into cl
-            "jmp paeth_dpaeth            \n\t"
-
-         "paeth_dabb:                    \n\t"
-            // pa <= pb; now test if pa <= pc
-            "cmpl _pctemp, %%eax         \n\t"
-            "jna paeth_dabc              \n\t"
-            // pa > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
-            "movb (%%esi,%%edx,), %%cl   \n\t" // load Prior(x-bpp) into cl
-            "jmp paeth_dpaeth            \n\t"
-
-         "paeth_dabc:                    \n\t"
-            // pa <= pc; Raw(x) = Paeth(x) + Raw(x-bpp)
-            "movb (%%edi,%%edx,), %%cl   \n\t" // load Raw(x-bpp) into cl
-
-         "paeth_dpaeth:                  \n\t"
-            "incl %%ebx                  \n\t"
-            "incl %%edx                  \n\t"
-            // Raw(x) = (Paeth(x) + Paeth_Predictor( a, b, c )) mod 256
-            "addb %%cl, -1(%%edi,%%ebx,) \n\t"
-            "cmpl _FullLength, %%ebx     \n\t"
-            "jb paeth_dlp                \n\t"
-
-         "paeth_dend:                    \n\t"
-#ifdef __PIC__
-            "popl %%ebx                  \n\t" // index to Global Offset Table
-#endif
-
-            : "=c" (dummy_value_c),            // output regs (dummy)
-              "=S" (dummy_value_S),
-              "=D" (dummy_value_D)
-
-            : "0" (bpp),       // ecx          // input regs
-              "1" (prev_row),  // esi
-              "2" (row)        // edi
-
-            : "%eax", "%edx"                   // clobber list
-#ifndef __PIC__
-            , "%ebx"
-#endif
-         );
-      }
-      return;                   // No need to go further with this one
-
-   } // end switch (bpp)
-
-   __asm__ __volatile__ (
-      // MMX acceleration complete; now do clean-up
-      // check if any remaining bytes left to decode
-#ifdef __PIC__
-      "pushl %%ebx                 \n\t" // save index to Global Offset Table
-#endif
-      "movl _MMXLength, %%ebx      \n\t"
-      "cmpl _FullLength, %%ebx     \n\t"
-      "jnb paeth_end               \n\t"
-//pre "movl row, %%edi             \n\t"
-//pre "movl prev_row, %%esi        \n\t"
-      // do Paeth decode for remaining bytes
-      "movl %%ebx, %%edx           \n\t"
-//pre "subl bpp, %%edx             \n\t" // (bpp is preloaded into ecx)
-      "subl %%ecx, %%edx           \n\t" // edx = ebx - bpp
-      "xorl %%ecx, %%ecx           \n\t" // zero ecx before using cl & cx below
-
-   "paeth_lp2:                     \n\t"
-      "xorl %%eax, %%eax           \n\t"
-      // pav = p - a = (a + b - c) - a = b - c
-      "movb (%%esi,%%ebx,), %%al   \n\t" // load Prior(x) into al
-      "movb (%%esi,%%edx,), %%cl   \n\t" // load Prior(x-bpp) into cl
-      "subl %%ecx, %%eax           \n\t" // subtract Prior(x-bpp)
-      "movl %%eax, _patemp         \n\t" // Save pav for later use
-      "xorl %%eax, %%eax           \n\t"
-      // pbv = p - b = (a + b - c) - b = a - c
-      "movb (%%edi,%%edx,), %%al   \n\t" // load Raw(x-bpp) into al
-      "subl %%ecx, %%eax           \n\t" // subtract Prior(x-bpp)
-      "movl %%eax, %%ecx           \n\t"
-      // pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
-      "addl _patemp, %%eax         \n\t" // pcv = pav + pbv
-      // pc = abs(pcv)
-      "testl $0x80000000, %%eax    \n\t"
-      "jz paeth_pca2               \n\t"
-      "negl %%eax                  \n\t" // reverse sign of neg values
-
-   "paeth_pca2:                    \n\t"
-      "movl %%eax, _pctemp         \n\t" // save pc for later use
-      // pb = abs(pbv)
-      "testl $0x80000000, %%ecx    \n\t"
-      "jz paeth_pba2               \n\t"
-      "negl %%ecx                  \n\t" // reverse sign of neg values
-
-   "paeth_pba2:                    \n\t"
-      "movl %%ecx, _pbtemp         \n\t" // save pb for later use
-      // pa = abs(pav)
-      "movl _patemp, %%eax         \n\t"
-      "testl $0x80000000, %%eax    \n\t"
-      "jz paeth_paa2               \n\t"
-      "negl %%eax                  \n\t" // reverse sign of neg values
-
-   "paeth_paa2:                    \n\t"
-      "movl %%eax, _patemp         \n\t" // save pa for later use
-      // test if pa <= pb
-      "cmpl %%ecx, %%eax           \n\t"
-      "jna paeth_abb2              \n\t"
-      // pa > pb; now test if pb <= pc
-      "cmpl _pctemp, %%ecx         \n\t"
-      "jna paeth_bbc2              \n\t"
-      // pb > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
-      "movb (%%esi,%%edx,), %%cl   \n\t" // load Prior(x-bpp) into cl
-      "jmp paeth_paeth2            \n\t"
-
-   "paeth_bbc2:                    \n\t"
-      // pb <= pc; Raw(x) = Paeth(x) + Prior(x)
-      "movb (%%esi,%%ebx,), %%cl   \n\t" // load Prior(x) into cl
-      "jmp paeth_paeth2            \n\t"
-
-   "paeth_abb2:                    \n\t"
-      // pa <= pb; now test if pa <= pc
-      "cmpl _pctemp, %%eax         \n\t"
-      "jna paeth_abc2              \n\t"
-      // pa > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
-      "movb (%%esi,%%edx,), %%cl   \n\t" // load Prior(x-bpp) into cl
-      "jmp paeth_paeth2            \n\t"
-
-   "paeth_abc2:                    \n\t"
-      // pa <= pc; Raw(x) = Paeth(x) + Raw(x-bpp)
-      "movb (%%edi,%%edx,), %%cl   \n\t" // load Raw(x-bpp) into cl
-
-   "paeth_paeth2:                  \n\t"
-      "incl %%ebx                  \n\t"
-      "incl %%edx                  \n\t"
-      // Raw(x) = (Paeth(x) + Paeth_Predictor( a, b, c )) mod 256
-      "addb %%cl, -1(%%edi,%%ebx,) \n\t"
-      "cmpl _FullLength, %%ebx     \n\t"
-      "jb paeth_lp2                \n\t"
-
-   "paeth_end:                     \n\t"
-      "EMMS                        \n\t" // end MMX; prep for poss. FP instrs.
-#ifdef __PIC__
-      "popl %%ebx                  \n\t" // restore index to Global Offset Table
-#endif
-
-      : "=c" (dummy_value_c),            // output regs (dummy)
-        "=S" (dummy_value_S),
-        "=D" (dummy_value_D)
-
-      : "0" (bpp),       // ecx          // input regs
-        "1" (prev_row),  // esi
-        "2" (row)        // edi
-
-      : "%eax", "%edx"                   // clobber list (no input regs!)
-#ifndef __PIC__
-      , "%ebx"
-#endif
-   );
-
-} /* end png_read_filter_row_mmx_paeth() */
-#endif
-
-
-
-
-#ifdef PNG_THREAD_UNSAFE_OK
-//===========================================================================//
-//                                                                           //
-//           P N G _ R E A D _ F I L T E R _ R O W _ M M X _ S U B           //
-//                                                                           //
-//===========================================================================//
-
-// Optimized code for PNG Sub filter decoder
-
-static void /* PRIVATE */
-png_read_filter_row_mmx_sub(png_row_infop row_info, png_bytep row)
-{
-   int bpp;
-   int dummy_value_a;
-   int dummy_value_D;
-
-   bpp = (row_info->pixel_depth + 7) >> 3;   // calc number of bytes per pixel
-   _FullLength = row_info->rowbytes - bpp;   // number of bytes to filter
-
-   __asm__ __volatile__ (
-//pre "movl row, %%edi             \n\t"
-      "movl %%edi, %%esi           \n\t" // lp = row
-//pre "movl bpp, %%eax             \n\t"
-      "addl %%eax, %%edi           \n\t" // rp = row + bpp
-//irr "xorl %%eax, %%eax           \n\t"
-      // get # of bytes to alignment
-      "movl %%edi, _dif            \n\t" // take start of row
-      "addl $0xf, _dif             \n\t" // add 7 + 8 to incr past
-                                         //  alignment boundary
-      "xorl %%ecx, %%ecx           \n\t"
-      "andl $0xfffffff8, _dif      \n\t" // mask to alignment boundary
-      "subl %%edi, _dif            \n\t" // subtract from start ==> value
-      "jz sub_go                   \n\t" //  ecx at alignment
-
-   "sub_lp1:                       \n\t" // fix alignment
-      "movb (%%esi,%%ecx,), %%al   \n\t"
-      "addb %%al, (%%edi,%%ecx,)   \n\t"
-      "incl %%ecx                  \n\t"
-      "cmpl _dif, %%ecx            \n\t"
-      "jb sub_lp1                  \n\t"
-
-   "sub_go:                        \n\t"
-      "movl _FullLength, %%eax     \n\t"
-      "movl %%eax, %%edx           \n\t"
-      "subl %%ecx, %%edx           \n\t" // subtract alignment fix
-      "andl $0x00000007, %%edx     \n\t" // calc bytes over mult of 8
-      "subl %%edx, %%eax           \n\t" // drop over bytes from length
-      "movl %%eax, _MMXLength      \n\t"
-
-      : "=a" (dummy_value_a),   // 0      // output regs (dummy)
-        "=D" (dummy_value_D)    // 1
-
-      : "0" (bpp),              // eax    // input regs
-        "1" (row)               // edi
-
-      : "%ebx", "%ecx", "%edx"            // clobber list
-      , "%esi"
-
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-      , "%mm0", "%mm1", "%mm2", "%mm3"
-      , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-   );
-
-   // now do the math for the rest of the row
-   switch (bpp)
-   {
-      case 3:
-      {
-         _ActiveMask.use  = 0x0000ffffff000000LL;
-         _ShiftBpp.use = 24;       // == 3 * 8
-         _ShiftRem.use  = 40;      // == 64 - 24
-
-         __asm__ __volatile__ (
-// preload  "movl row, %%edi              \n\t"
-            "movq _ActiveMask, %%mm7       \n\t" // load _ActiveMask for 2nd
-                                                //  active byte group
-            "movl %%edi, %%esi            \n\t" // lp = row
-// preload  "movl bpp, %%eax              \n\t"
-            "addl %%eax, %%edi            \n\t" // rp = row + bpp
-            "movq %%mm7, %%mm6            \n\t"
-            "movl _dif, %%edx             \n\t"
-            "psllq _ShiftBpp, %%mm6       \n\t" // move mask in mm6 to cover
-                                                //  3rd active byte group
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%edx,), %%mm1 \n\t"
-
-         "sub_3lp:                        \n\t" // shift data for adding first
-            "psrlq _ShiftRem, %%mm1       \n\t" //  bpp bytes (no need for mask;
-                                                //  shift clears inactive bytes)
-            // add 1st active group
-            "movq (%%edi,%%edx,), %%mm0   \n\t"
-            "paddb %%mm1, %%mm0           \n\t"
-
-            // add 2nd active group
-            "movq %%mm0, %%mm1            \n\t" // mov updated Raws to mm1
-            "psllq _ShiftBpp, %%mm1       \n\t" // shift data to pos. correctly
-            "pand %%mm7, %%mm1            \n\t" // mask to use 2nd active group
-            "paddb %%mm1, %%mm0           \n\t"
-
-            // add 3rd active group
-            "movq %%mm0, %%mm1            \n\t" // mov updated Raws to mm1
-            "psllq _ShiftBpp, %%mm1       \n\t" // shift data to pos. correctly
-            "pand %%mm6, %%mm1            \n\t" // mask to use 3rd active group
-            "addl $8, %%edx               \n\t"
-            "paddb %%mm1, %%mm0           \n\t"
-
-            "cmpl _MMXLength, %%edx       \n\t"
-            "movq %%mm0, -8(%%edi,%%edx,) \n\t" // write updated Raws to array
-            "movq %%mm0, %%mm1            \n\t" // prep 1st add at top of loop
-            "jb sub_3lp                   \n\t"
-
-            : "=a" (dummy_value_a),   // 0      // output regs (dummy)
-              "=D" (dummy_value_D)    // 1
-
-            : "0" (bpp),              // eax    // input regs
-              "1" (row)               // edi
-
-            : "%edx", "%esi"                    // clobber list
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;
-
-      case 1:
-      {
-         __asm__ __volatile__ (
-            "movl _dif, %%edx            \n\t"
-// preload  "movl row, %%edi             \n\t"
-            "cmpl _FullLength, %%edx     \n\t"
-            "jnb sub_1end                \n\t"
-            "movl %%edi, %%esi           \n\t" // lp = row
-            "xorl %%eax, %%eax           \n\t"
-// preload  "movl bpp, %%eax             \n\t"
-            "addl %%eax, %%edi           \n\t" // rp = row + bpp
-
-         "sub_1lp:                       \n\t"
-            "movb (%%esi,%%edx,), %%al   \n\t"
-            "addb %%al, (%%edi,%%edx,)   \n\t"
-            "incl %%edx                  \n\t"
-            "cmpl _FullLength, %%edx     \n\t"
-            "jb sub_1lp                  \n\t"
-
-         "sub_1end:                      \n\t"
-
-            : "=a" (dummy_value_a),   // 0      // output regs (dummy)
-              "=D" (dummy_value_D)    // 1
-
-            : "0" (bpp),              // eax    // input regs
-              "1" (row)               // edi
-
-            : "%edx", "%esi"                    // clobber list
-         );
-      }
-      return;
-
-      case 6:
-      case 4:
-      //case 7:   // GRR BOGUS
-      //case 5:   // GRR BOGUS
-      {
-         _ShiftBpp.use = bpp << 3;
-         _ShiftRem.use = 64 - _ShiftBpp.use;
-
-         __asm__ __volatile__ (
-// preload  "movl row, %%edi              \n\t"
-            "movl _dif, %%edx             \n\t"
-            "movl %%edi, %%esi            \n\t" // lp = row
-// preload  "movl bpp, %%eax              \n\t"
-            "addl %%eax, %%edi            \n\t" // rp = row + bpp
-
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%edx,), %%mm1 \n\t"
-
-         "sub_4lp:                        \n\t" // shift data for adding first
-            "psrlq _ShiftRem, %%mm1       \n\t" //  bpp bytes (no need for mask;
-                                                //  shift clears inactive bytes)
-            "movq (%%edi,%%edx,), %%mm0   \n\t"
-            "paddb %%mm1, %%mm0           \n\t"
-
-            // add 2nd active group
-            "movq %%mm0, %%mm1            \n\t" // mov updated Raws to mm1
-            "psllq _ShiftBpp, %%mm1       \n\t" // shift data to pos. correctly
-            "addl $8, %%edx               \n\t"
-            "paddb %%mm1, %%mm0           \n\t"
-
-            "cmpl _MMXLength, %%edx       \n\t"
-            "movq %%mm0, -8(%%edi,%%edx,) \n\t"
-            "movq %%mm0, %%mm1            \n\t" // prep 1st add at top of loop
-            "jb sub_4lp                   \n\t"
-
-            : "=a" (dummy_value_a),   // 0      // output regs (dummy)
-              "=D" (dummy_value_D)    // 1
-
-            : "0" (bpp),              // eax    // input regs
-              "1" (row)               // edi
-
-            : "%edx", "%esi"                    // clobber list
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1"
-#endif
-         );
-      }
-      break;
-
-      case 2:
-      {
-         _ActiveMask.use = 0x00000000ffff0000LL;
-         _ShiftBpp.use = 16;       // == 2 * 8
-         _ShiftRem.use = 48;       // == 64 - 16
-
-         __asm__ __volatile__ (
-            "movq _ActiveMask, %%mm7      \n\t" // load _ActiveMask for 2nd
-                                                //  active byte group
-            "movl _dif, %%edx             \n\t"
-            "movq %%mm7, %%mm6            \n\t"
-// preload  "movl row, %%edi              \n\t"
-            "psllq _ShiftBpp, %%mm6       \n\t" // move mask in mm6 to cover
-                                                //  3rd active byte group
-            "movl %%edi, %%esi            \n\t" // lp = row
-            "movq %%mm6, %%mm5            \n\t"
-// preload  "movl bpp, %%eax              \n\t"
-            "addl %%eax, %%edi            \n\t" // rp = row + bpp
-            "psllq _ShiftBpp, %%mm5       \n\t" // move mask in mm5 to cover
-                                                //  4th active byte group
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%edx,), %%mm1 \n\t"
-
-         "sub_2lp:                        \n\t" // shift data for adding first
-            "psrlq _ShiftRem, %%mm1       \n\t" //  bpp bytes (no need for mask;
-                                                //  shift clears inactive bytes)
-            // add 1st active group
-            "movq (%%edi,%%edx,), %%mm0   \n\t"
-            "paddb %%mm1, %%mm0           \n\t"
-
-            // add 2nd active group
-            "movq %%mm0, %%mm1            \n\t" // mov updated Raws to mm1
-            "psllq _ShiftBpp, %%mm1       \n\t" // shift data to pos. correctly
-            "pand %%mm7, %%mm1            \n\t" // mask to use 2nd active group
-            "paddb %%mm1, %%mm0           \n\t"
-
-            // add 3rd active group
-            "movq %%mm0, %%mm1            \n\t" // mov updated Raws to mm1
-            "psllq _ShiftBpp, %%mm1       \n\t" // shift data to pos. correctly
-            "pand %%mm6, %%mm1            \n\t" // mask to use 3rd active group
-            "paddb %%mm1, %%mm0           \n\t"
-
-            // add 4th active group
-            "movq %%mm0, %%mm1            \n\t" // mov updated Raws to mm1
-            "psllq _ShiftBpp, %%mm1       \n\t" // shift data to pos. correctly
-            "pand %%mm5, %%mm1            \n\t" // mask to use 4th active group
-            "addl $8, %%edx               \n\t"
-            "paddb %%mm1, %%mm0           \n\t"
-            "cmpl _MMXLength, %%edx       \n\t"
-            "movq %%mm0, -8(%%edi,%%edx,) \n\t" // write updated Raws to array
-            "movq %%mm0, %%mm1            \n\t" // prep 1st add at top of loop
-            "jb sub_2lp                   \n\t"
-
-            : "=a" (dummy_value_a),   // 0      // output regs (dummy)
-              "=D" (dummy_value_D)    // 1
-
-            : "0" (bpp),              // eax    // input regs
-              "1" (row)               // edi
-
-            : "%edx", "%esi"                    // clobber list
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm5", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;
-
-      case 8:
-      {
-         __asm__ __volatile__ (
-// preload  "movl row, %%edi              \n\t"
-            "movl _dif, %%edx             \n\t"
-            "movl %%edi, %%esi            \n\t" // lp = row
-// preload  "movl bpp, %%eax              \n\t"
-            "addl %%eax, %%edi            \n\t" // rp = row + bpp
-            "movl _MMXLength, %%ecx       \n\t"
-
-            // prime the pump:  load the first Raw(x-bpp) data set
-            "movq -8(%%edi,%%edx,), %%mm7 \n\t"
-            "andl $0x0000003f, %%ecx      \n\t" // calc bytes over mult of 64
-
-         "sub_8lp:                        \n\t"
-            "movq (%%edi,%%edx,), %%mm0   \n\t" // load Sub(x) for 1st 8 bytes
-            "paddb %%mm7, %%mm0           \n\t"
-            "movq 8(%%edi,%%edx,), %%mm1  \n\t" // load Sub(x) for 2nd 8 bytes
-            "movq %%mm0, (%%edi,%%edx,)   \n\t" // write Raw(x) for 1st 8 bytes
-
-            // Now mm0 will be used as Raw(x-bpp) for the 2nd group of 8 bytes.
-            // This will be repeated for each group of 8 bytes with the 8th
-            // group being used as the Raw(x-bpp) for the 1st group of the
-            // next loop.
-
-            "paddb %%mm0, %%mm1           \n\t"
-            "movq 16(%%edi,%%edx,), %%mm2 \n\t" // load Sub(x) for 3rd 8 bytes
-            "movq %%mm1, 8(%%edi,%%edx,)  \n\t" // write Raw(x) for 2nd 8 bytes
-            "paddb %%mm1, %%mm2           \n\t"
-            "movq 24(%%edi,%%edx,), %%mm3 \n\t" // load Sub(x) for 4th 8 bytes
-            "movq %%mm2, 16(%%edi,%%edx,) \n\t" // write Raw(x) for 3rd 8 bytes
-            "paddb %%mm2, %%mm3           \n\t"
-            "movq 32(%%edi,%%edx,), %%mm4 \n\t" // load Sub(x) for 5th 8 bytes
-            "movq %%mm3, 24(%%edi,%%edx,) \n\t" // write Raw(x) for 4th 8 bytes
-            "paddb %%mm3, %%mm4           \n\t"
-            "movq 40(%%edi,%%edx,), %%mm5 \n\t" // load Sub(x) for 6th 8 bytes
-            "movq %%mm4, 32(%%edi,%%edx,) \n\t" // write Raw(x) for 5th 8 bytes
-            "paddb %%mm4, %%mm5           \n\t"
-            "movq 48(%%edi,%%edx,), %%mm6 \n\t" // load Sub(x) for 7th 8 bytes
-            "movq %%mm5, 40(%%edi,%%edx,) \n\t" // write Raw(x) for 6th 8 bytes
-            "paddb %%mm5, %%mm6           \n\t"
-            "movq 56(%%edi,%%edx,), %%mm7 \n\t" // load Sub(x) for 8th 8 bytes
-            "movq %%mm6, 48(%%edi,%%edx,) \n\t" // write Raw(x) for 7th 8 bytes
-            "addl $64, %%edx              \n\t"
-            "paddb %%mm6, %%mm7           \n\t"
-            "cmpl %%ecx, %%edx            \n\t"
-            "movq %%mm7, -8(%%edi,%%edx,) \n\t" // write Raw(x) for 8th 8 bytes
-            "jb sub_8lp                   \n\t"
-
-            "cmpl _MMXLength, %%edx       \n\t"
-            "jnb sub_8lt8                 \n\t"
-
-         "sub_8lpA:                       \n\t"
-            "movq (%%edi,%%edx,), %%mm0   \n\t"
-            "addl $8, %%edx               \n\t"
-            "paddb %%mm7, %%mm0           \n\t"
-            "cmpl _MMXLength, %%edx       \n\t"
-            "movq %%mm0, -8(%%edi,%%edx,) \n\t" // -8 to offset early addl edx
-            "movq %%mm0, %%mm7            \n\t" // move calculated Raw(x) data
-                                                //  to mm1 to be new Raw(x-bpp)
-                                                //  for next loop
-            "jb sub_8lpA                  \n\t"
-
-         "sub_8lt8:                       \n\t"
-
-            : "=a" (dummy_value_a),   // 0      // output regs (dummy)
-              "=D" (dummy_value_D)    // 1
-
-            : "0" (bpp),              // eax    // input regs
-              "1" (row)               // edi
-
-            : "%ecx", "%edx", "%esi"            // clobber list
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-         );
-      }
-      break;
-
-      default:                // bpp greater than 8 bytes   GRR BOGUS
-      {
-         __asm__ __volatile__ (
-            "movl _dif, %%edx             \n\t"
-// preload  "movl row, %%edi              \n\t"
-            "movl %%edi, %%esi            \n\t" // lp = row
-// preload  "movl bpp, %%eax              \n\t"
-            "addl %%eax, %%edi            \n\t" // rp = row + bpp
-
-         "sub_Alp:                        \n\t"
-            "movq (%%edi,%%edx,), %%mm0   \n\t"
-            "movq (%%esi,%%edx,), %%mm1   \n\t"
-            "addl $8, %%edx               \n\t"
-            "paddb %%mm1, %%mm0           \n\t"
-            "cmpl _MMXLength, %%edx       \n\t"
-            "movq %%mm0, -8(%%edi,%%edx,) \n\t" // mov does not affect flags;
-                                                //  -8 to offset addl edx
-            "jb sub_Alp                   \n\t"
-
-            : "=a" (dummy_value_a),   // 0      // output regs (dummy)
-              "=D" (dummy_value_D)    // 1
-
-            : "0" (bpp),              // eax    // input regs
-              "1" (row)               // edi
-
-            : "%edx", "%esi"                    // clobber list
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-            , "%mm0", "%mm1"
-#endif
-         );
-      }
-      break;
-
-   } // end switch (bpp)
-
-   __asm__ __volatile__ (
-      "movl _MMXLength, %%edx       \n\t"
-//pre "movl row, %%edi              \n\t"
-      "cmpl _FullLength, %%edx      \n\t"
-      "jnb sub_end                  \n\t"
-
-      "movl %%edi, %%esi            \n\t" // lp = row
-//pre "movl bpp, %%eax              \n\t"
-      "addl %%eax, %%edi            \n\t" // rp = row + bpp
-      "xorl %%eax, %%eax            \n\t"
-
-   "sub_lp2:                        \n\t"
-      "movb (%%esi,%%edx,), %%al    \n\t"
-      "addb %%al, (%%edi,%%edx,)    \n\t"
-      "incl %%edx                   \n\t"
-      "cmpl _FullLength, %%edx      \n\t"
-      "jb sub_lp2                   \n\t"
-
-   "sub_end:                        \n\t"
-      "EMMS                         \n\t" // end MMX instructions
-
-      : "=a" (dummy_value_a),   // 0      // output regs (dummy)
-        "=D" (dummy_value_D)    // 1
-
-      : "0" (bpp),              // eax    // input regs
-        "1" (row)               // edi
-
-      : "%edx", "%esi"                    // clobber list
-   );
-
-} // end of png_read_filter_row_mmx_sub()
-#endif
-
-
-
-
-//===========================================================================//
-//                                                                           //
-//            P N G _ R E A D _ F I L T E R _ R O W _ M M X _ U P            //
-//                                                                           //
-//===========================================================================//
-
-// Optimized code for PNG Up filter decoder
-
-static void /* PRIVATE */
-png_read_filter_row_mmx_up(png_row_infop row_info, png_bytep row,
-                           png_bytep prev_row)
-{
-   png_uint_32 len;
-   int dummy_value_d;   // fix 'forbidden register 3 (dx) was spilled' error
-   int dummy_value_S;
-   int dummy_value_D;
-
-   len = row_info->rowbytes;              // number of bytes to filter
-
-   __asm__ __volatile__ (
-//pre "movl row, %%edi              \n\t"
-      // get # of bytes to alignment
-#ifdef __PIC__
-      "pushl %%ebx                  \n\t"
-#endif
-      "movl %%edi, %%ecx            \n\t"
-      "xorl %%ebx, %%ebx            \n\t"
-      "addl $0x7, %%ecx             \n\t"
-      "xorl %%eax, %%eax            \n\t"
-      "andl $0xfffffff8, %%ecx      \n\t"
-//pre "movl prev_row, %%esi         \n\t"
-      "subl %%edi, %%ecx            \n\t"
-      "jz up_go                     \n\t"
-
-   "up_lp1:                         \n\t" // fix alignment
-      "movb (%%edi,%%ebx,), %%al    \n\t"
-      "addb (%%esi,%%ebx,), %%al    \n\t"
-      "incl %%ebx                   \n\t"
-      "cmpl %%ecx, %%ebx            \n\t"
-      "movb %%al, -1(%%edi,%%ebx,)  \n\t" // mov does not affect flags; -1 to
-      "jb up_lp1                    \n\t" //  offset incl ebx
-
-   "up_go:                          \n\t"
-//pre "movl len, %%edx              \n\t"
-      "movl %%edx, %%ecx            \n\t"
-      "subl %%ebx, %%edx            \n\t" // subtract alignment fix
-      "andl $0x0000003f, %%edx      \n\t" // calc bytes over mult of 64
-      "subl %%edx, %%ecx            \n\t" // drop over bytes from length
-
-      // unrolled loop - use all MMX registers and interleave to reduce
-      // number of branch instructions (loops) and reduce partial stalls
-   "up_loop:                        \n\t"
-      "movq (%%esi,%%ebx,), %%mm1   \n\t"
-      "movq (%%edi,%%ebx,), %%mm0   \n\t"
-      "movq 8(%%esi,%%ebx,), %%mm3  \n\t"
-      "paddb %%mm1, %%mm0           \n\t"
-      "movq 8(%%edi,%%ebx,), %%mm2  \n\t"
-      "movq %%mm0, (%%edi,%%ebx,)   \n\t"
-      "paddb %%mm3, %%mm2           \n\t"
-      "movq 16(%%esi,%%ebx,), %%mm5 \n\t"
-      "movq %%mm2, 8(%%edi,%%ebx,)  \n\t"
-      "movq 16(%%edi,%%ebx,), %%mm4 \n\t"
-      "movq 24(%%esi,%%ebx,), %%mm7 \n\t"
-      "paddb %%mm5, %%mm4           \n\t"
-      "movq 24(%%edi,%%ebx,), %%mm6 \n\t"
-      "movq %%mm4, 16(%%edi,%%ebx,) \n\t"
-      "paddb %%mm7, %%mm6           \n\t"
-      "movq 32(%%esi,%%ebx,), %%mm1 \n\t"
-      "movq %%mm6, 24(%%edi,%%ebx,) \n\t"
-      "movq 32(%%edi,%%ebx,), %%mm0 \n\t"
-      "movq 40(%%esi,%%ebx,), %%mm3 \n\t"
-      "paddb %%mm1, %%mm0           \n\t"
-      "movq 40(%%edi,%%ebx,), %%mm2 \n\t"
-      "movq %%mm0, 32(%%edi,%%ebx,) \n\t"
-      "paddb %%mm3, %%mm2           \n\t"
-      "movq 48(%%esi,%%ebx,), %%mm5 \n\t"
-      "movq %%mm2, 40(%%edi,%%ebx,) \n\t"
-      "movq 48(%%edi,%%ebx,), %%mm4 \n\t"
-      "movq 56(%%esi,%%ebx,), %%mm7 \n\t"
-      "paddb %%mm5, %%mm4           \n\t"
-      "movq 56(%%edi,%%ebx,), %%mm6 \n\t"
-      "movq %%mm4, 48(%%edi,%%ebx,) \n\t"
-      "addl $64, %%ebx              \n\t"
-      "paddb %%mm7, %%mm6           \n\t"
-      "cmpl %%ecx, %%ebx            \n\t"
-      "movq %%mm6, -8(%%edi,%%ebx,) \n\t" // (+56)movq does not affect flags;
-      "jb up_loop                   \n\t" //  -8 to offset addl ebx
-
-      "cmpl $0, %%edx               \n\t" // test for bytes over mult of 64
-      "jz up_end                    \n\t"
-
-      "cmpl $8, %%edx               \n\t" // test for less than 8 bytes
-      "jb up_lt8                    \n\t" //  [added by lcreeve@netins.net]
-
-      "addl %%edx, %%ecx            \n\t"
-      "andl $0x00000007, %%edx      \n\t" // calc bytes over mult of 8
-      "subl %%edx, %%ecx            \n\t" // drop over bytes from length
-      "jz up_lt8                    \n\t"
-
-   "up_lpA:                         \n\t" // use MMX regs to update 8 bytes sim.
-      "movq (%%esi,%%ebx,), %%mm1   \n\t"
-      "movq (%%edi,%%ebx,), %%mm0   \n\t"
-      "addl $8, %%ebx               \n\t"
-      "paddb %%mm1, %%mm0           \n\t"
-      "cmpl %%ecx, %%ebx            \n\t"
-      "movq %%mm0, -8(%%edi,%%ebx,) \n\t" // movq does not affect flags; -8 to
-      "jb up_lpA                    \n\t" //  offset add ebx
-      "cmpl $0, %%edx               \n\t" // test for bytes over mult of 8
-      "jz up_end                    \n\t"
-
-   "up_lt8:                         \n\t"
-      "xorl %%eax, %%eax            \n\t"
-      "addl %%edx, %%ecx            \n\t" // move over byte count into counter
-
-   "up_lp2:                         \n\t" // use x86 regs for remaining bytes
-      "movb (%%edi,%%ebx,), %%al    \n\t"
-      "addb (%%esi,%%ebx,), %%al    \n\t"
-      "incl %%ebx                   \n\t"
-      "cmpl %%ecx, %%ebx            \n\t"
-      "movb %%al, -1(%%edi,%%ebx,)  \n\t" // mov does not affect flags; -1 to
-      "jb up_lp2                    \n\t" //  offset inc ebx
-
-   "up_end:                         \n\t"
-      "EMMS                         \n\t" // conversion of filtered row complete
-#ifdef __PIC__
-      "popl %%ebx                   \n\t"
-#endif
-
-      : "=d" (dummy_value_d),   // 0      // output regs (dummy)
-        "=S" (dummy_value_S),   // 1
-        "=D" (dummy_value_D)    // 2
-
-      : "0" (len),              // edx    // input regs
-        "1" (prev_row),         // esi
-        "2" (row)               // edi
-
-      : "%eax", "%ebx", "%ecx"            // clobber list (no input regs!)
-
-#if 0  /* MMX regs (%mm0, etc.) not supported by gcc 2.7.2.3 or egcs 1.1 */
-      , "%mm0", "%mm1", "%mm2", "%mm3"
-      , "%mm4", "%mm5", "%mm6", "%mm7"
-#endif
-   );
-
-} // end of png_read_filter_row_mmx_up()
-
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-
-
-
-
-/*===========================================================================*/
-/*                                                                           */
-/*                   P N G _ R E A D _ F I L T E R _ R O W                   */
-/*                                                                           */
-/*===========================================================================*/
-
-
-/* Optimized png_read_filter_row routines */
-
-void /* PRIVATE */
-png_read_filter_row(png_structp png_ptr, png_row_infop row_info, png_bytep
-   row, png_bytep prev_row, int filter)
-{
-#ifdef PNG_DEBUG
-   char filnm[10];
-#endif
-
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED)
-/* GRR:  these are superseded by png_ptr->asm_flags: */
-#define UseMMX_sub    1   // GRR:  converted 20000730
-#define UseMMX_up     1   // GRR:  converted 20000729
-#define UseMMX_avg    1   // GRR:  converted 20000828 (+ 16-bit bugfix 20000916)
-#define UseMMX_paeth  1   // GRR:  converted 20000828
-
-   if (_mmx_supported == 2) {
-       /* this should have happened in png_init_mmx_flags() already */
-       png_warning(png_ptr, "asm_flags may not have been initialized");
-       png_mmx_support();
-   }
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-
-#ifdef PNG_DEBUG
-   png_debug(1, "in png_read_filter_row (pnggccrd.c)\n");
-   switch (filter)
-   {
-      case 0: sprintf(filnm, "none");
-         break;
-      case 1: sprintf(filnm, "sub-%s",
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-        (png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_SUB)? "MMX" : 
-#endif
-#endif
-"x86");
-         break;
-      case 2: sprintf(filnm, "up-%s",
-#ifdef PNG_ASSEMBLER_CODE_SUPPORTED
-#if !defined(PNG_1_0_X)
-        (png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_UP)? "MMX" :
-#endif
-#endif
- "x86");
-         break;
-      case 3: sprintf(filnm, "avg-%s",
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-        (png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_AVG)? "MMX" :
-#endif
-#endif
- "x86");
-         break;
-      case 4: sprintf(filnm, "Paeth-%s",
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-        (png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_PAETH)? "MMX":
-#endif
-#endif
-"x86");
-         break;
-      default: sprintf(filnm, "unknw");
-         break;
-   }
-   png_debug2(0, "row_number=%5ld, %5s, ", png_ptr->row_number, filnm);
-   png_debug1(0, "row=0x%08lx, ", (unsigned long)row);
-   png_debug2(0, "pixdepth=%2d, bytes=%d, ", (int)row_info->pixel_depth,
-      (int)((row_info->pixel_depth + 7) >> 3));
-   png_debug1(0,"rowbytes=%8ld\n", row_info->rowbytes);
-#endif /* PNG_DEBUG */
-
-   switch (filter)
-   {
-      case PNG_FILTER_VALUE_NONE:
-         break;
-
-      case PNG_FILTER_VALUE_SUB:
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-         if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_SUB) &&
-             (row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
-             (row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
-#else
-         if (_mmx_supported)
-#endif
-         {
-            png_read_filter_row_mmx_sub(row_info, row);
-         }
-         else
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-         {
-            png_uint_32 i;
-            png_uint_32 istop = row_info->rowbytes;
-            png_uint_32 bpp = (row_info->pixel_depth + 7) >> 3;
-            png_bytep rp = row + bpp;
-            png_bytep lp = row;
-
-            for (i = bpp; i < istop; i++)
-            {
-               *rp = (png_byte)(((int)(*rp) + (int)(*lp++)) & 0xff);
-               rp++;
-            }
-         }  /* end !UseMMX_sub */
-         break;
-
-      case PNG_FILTER_VALUE_UP:
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED)
-#if !defined(PNG_1_0_X)
-         if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_UP) &&
-             (row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
-             (row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
-#else
-         if (_mmx_supported)
-#endif
-         {
-            png_read_filter_row_mmx_up(row_info, row, prev_row);
-         }
-          else
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-         {
-            png_uint_32 i;
-            png_uint_32 istop = row_info->rowbytes;
-            png_bytep rp = row;
-            png_bytep pp = prev_row;
-
-            for (i = 0; i < istop; ++i)
-            {
-               *rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff);
-               rp++;
-            }
-         }  /* end !UseMMX_up */
-         break;
-
-      case PNG_FILTER_VALUE_AVG:
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-         if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_AVG) &&
-             (row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
-             (row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
-#else
-         if (_mmx_supported)
-#endif
-         {
-            png_read_filter_row_mmx_avg(row_info, row, prev_row);
-         }
-         else
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-         {
-            png_uint_32 i;
-            png_bytep rp = row;
-            png_bytep pp = prev_row;
-            png_bytep lp = row;
-            png_uint_32 bpp = (row_info->pixel_depth + 7) >> 3;
-            png_uint_32 istop = row_info->rowbytes - bpp;
-
-            for (i = 0; i < bpp; i++)
-            {
-               *rp = (png_byte)(((int)(*rp) +
-                  ((int)(*pp++) >> 1)) & 0xff);
-               rp++;
-            }
-
-            for (i = 0; i < istop; i++)
-            {
-               *rp = (png_byte)(((int)(*rp) +
-                  ((int)(*pp++ + *lp++) >> 1)) & 0xff);
-               rp++;
-            }
-         }  /* end !UseMMX_avg */
-         break;
-
-      case PNG_FILTER_VALUE_PAETH:
-#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_THREAD_UNSAFE_OK)
-#if !defined(PNG_1_0_X)
-         if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_PAETH) &&
-             (row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
-             (row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
-#else
-         if (_mmx_supported)
-#endif
-         {
-            png_read_filter_row_mmx_paeth(row_info, row, prev_row);
-         }
-         else
-#endif /* PNG_ASSEMBLER_CODE_SUPPORTED */
-         {
-            png_uint_32 i;
-            png_bytep rp = row;
-            png_bytep pp = prev_row;
-            png_bytep lp = row;
-            png_bytep cp = prev_row;
-            png_uint_32 bpp = (row_info->pixel_depth + 7) >> 3;
-            png_uint_32 istop = row_info->rowbytes - bpp;
-
-            for (i = 0; i < bpp; i++)
-            {
-               *rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff);
-               rp++;
-            }
-
-            for (i = 0; i < istop; i++)   /* use leftover rp,pp */
-            {
-               int a, b, c, pa, pb, pc, p;
-
-               a = *lp++;
-               b = *pp++;
-               c = *cp++;
-
-               p = b - c;
-               pc = a - c;
-
-#ifdef PNG_USE_ABS
-               pa = abs(p);
-               pb = abs(pc);
-               pc = abs(p + pc);
-#else
-               pa = p < 0 ? -p : p;
-               pb = pc < 0 ? -pc : pc;
-               pc = (p + pc) < 0 ? -(p + pc) : p + pc;
-#endif
-
-               /*
-                  if (pa <= pb && pa <= pc)
-                     p = a;
-                  else if (pb <= pc)
-                     p = b;
-                  else
-                     p = c;
-                */
-
-               p = (pa <= pb && pa <= pc) ? a : (pb <= pc) ? b : c;
-
-               *rp = (png_byte)(((int)(*rp) + p) & 0xff);
-               rp++;
-            }
-         }  /* end !UseMMX_paeth */
-         break;
-
-      default:
-         png_warning(png_ptr, "Ignoring bad row-filter type");
-         *row=0;
-         break;
-   }
-}
-
-#endif /* PNG_HAVE_ASSEMBLER_READ_FILTER_ROW */
-
-
-/*===========================================================================*/
-/*                                                                           */
-/*                      P N G _ M M X _ S U P P O R T                        */
-/*                                                                           */
-/*===========================================================================*/
-
-/* GRR NOTES:  (1) the following code assumes 386 or better (pushfl/popfl)
- *             (2) all instructions compile with gcc 2.7.2.3 and later
- *             (3) the function is moved down here to prevent gcc from
- *                  inlining it in multiple places and then barfing be-
- *                  cause the ".NOT_SUPPORTED" label is multiply defined
- *             [is there a way to signal that a *single* function should
- *              not be inlined?  is there a way to modify the label for
- *              each inlined instance, e.g., by appending _1, _2, etc.?
- *              maybe if don't use leading "." in label name? (nope...sigh)]
- */
+int PNGAPI
+png_dummy_mmx_support(void) __attribute__((noinline));
 
 int PNGAPI
-png_mmx_support(void)
+png_dummy_mmx_support(void)
 {
-#if defined(PNG_MMX_CODE_SUPPORTED)
+   int result;
+#if defined(PNG_MMX_CODE_SUPPORTED)  // superfluous, but what the heck
     __asm__ __volatile__ (
+#if defined(__x86_64__)
+        "pushq %%rbx          \n\t"  // rbx gets clobbered by CPUID instruction
+        "pushq %%rcx          \n\t"  // so does rcx...
+        "pushq %%rdx          \n\t"  // ...and rdx (but rcx & rdx safe on Linux)
+        "pushfq               \n\t"  // save Eflag to stack
+        "popq %%rax           \n\t"  // get Eflag from stack into rax
+        "movq %%rax, %%rcx    \n\t"  // make another copy of Eflag in rcx
+        "xorl $0x200000, %%eax \n\t" // toggle ID bit in Eflag (i.e., bit 21)
+        "pushq %%rax          \n\t"  // save modified Eflag back to stack
+        "popfq                \n\t"  // restore modified value to Eflag reg
+        "pushfq               \n\t"  // save Eflag to stack
+        "popq %%rax           \n\t"  // get Eflag from stack
+        "pushq %%rcx          \n\t"  // save original Eflag to stack
+        "popfq                \n\t"  // restore original Eflag
+#else
         "pushl %%ebx          \n\t"  // ebx gets clobbered by CPUID instruction
         "pushl %%ecx          \n\t"  // so does ecx...
         "pushl %%edx          \n\t"  // ...and edx (but ecx & edx safe on Linux)
-//      ".byte  0x66          \n\t"  // convert 16-bit pushf to 32-bit pushfd
-//      "pushf                \n\t"  // 16-bit pushf
         "pushfl               \n\t"  // save Eflag to stack
         "popl %%eax           \n\t"  // get Eflag from stack into eax
         "movl %%eax, %%ecx    \n\t"  // make another copy of Eflag in ecx
         "xorl $0x200000, %%eax \n\t" // toggle ID bit in Eflag (i.e., bit 21)
         "pushl %%eax          \n\t"  // save modified Eflag back to stack
-//      ".byte  0x66          \n\t"  // convert 16-bit popf to 32-bit popfd
-//      "popf                 \n\t"  // 16-bit popf
         "popfl                \n\t"  // restore modified value to Eflag reg
         "pushfl               \n\t"  // save Eflag to stack
         "popl %%eax           \n\t"  // get Eflag from stack
         "pushl %%ecx          \n\t"  // save original Eflag to stack
         "popfl                \n\t"  // restore original Eflag
+#endif
         "xorl %%ecx, %%eax    \n\t"  // compare new Eflag with original Eflag
-        "jz .NOT_SUPPORTED    \n\t"  // if same, CPUID instr. is not supported
+        "jz 0f                \n\t"  // if same, CPUID instr. is not supported
 
         "xorl %%eax, %%eax    \n\t"  // set eax to zero
 //      ".byte  0x0f, 0xa2    \n\t"  // CPUID instruction (two-byte opcode)
         "cpuid                \n\t"  // get the CPU identification info
         "cmpl $1, %%eax       \n\t"  // make sure eax return non-zero value
-        "jl .NOT_SUPPORTED    \n\t"  // if eax is zero, MMX is not supported
+        "jl 0f                \n\t"  // if eax is zero, MMX is not supported
 
         "xorl %%eax, %%eax    \n\t"  // set eax to zero and...
         "incl %%eax           \n\t"  // ...increment eax to 1.  This pair is
@@ -5359,37 +61,41 @@ png_mmx_support(void)
         "cpuid                \n\t"  // get the CPU identification info again
         "andl $0x800000, %%edx \n\t" // mask out all bits but MMX bit (23)
         "cmpl $0, %%edx       \n\t"  // 0 = MMX not supported
-        "jz .NOT_SUPPORTED    \n\t"  // non-zero = yes, MMX IS supported
+        "jz 0f                \n\t"  // non-zero = yes, MMX IS supported
 
         "movl $1, %%eax       \n\t"  // set return value to 1
-        "jmp  .RETURN         \n\t"  // DONE:  have MMX support
+        "jmp  1f              \n\t"  // DONE:  have MMX support
 
-    ".NOT_SUPPORTED:          \n\t"  // target label for jump instructions
+    "0:                       \n\t"  // .NOT_SUPPORTED: target label for jump instructions
         "movl $0, %%eax       \n\t"  // set return value to 0
-    ".RETURN:          \n\t"  // target label for jump instructions
-        "movl %%eax, _mmx_supported \n\t" // save in global static variable, too
+    "1:                       \n\t"  // .RETURN: target label for jump instructions
+#if defined(__x86_64__)
+        "popq %%rdx           \n\t"  // restore rdx
+        "popq %%rcx           \n\t"  // restore rcx
+        "popq %%rbx           \n\t"  // restore rbx
+#else
         "popl %%edx           \n\t"  // restore edx
         "popl %%ecx           \n\t"  // restore ecx
         "popl %%ebx           \n\t"  // restore ebx
+#endif
 
 //      "ret                  \n\t"  // DONE:  no MMX support
                                      // (fall through to standard C "ret")
 
-        :                            // output list (none)
+        : "=a" (result)              // output list
 
         :                            // any variables used on input (none)
 
-        : "%eax"                     // clobber list
+                                     // no clobber list
 //      , "%ebx", "%ecx", "%edx"     // GRR:  we handle these manually
 //      , "memory"   // if write to a variable gcc thought was in a reg
 //      , "cc"       // "condition codes" (flag bits)
     );
-#else     
+    _mmx_supported = result;
+#else
     _mmx_supported = 0;
 #endif /* PNG_MMX_CODE_SUPPORTED */
 
     return _mmx_supported;
 }
-
-
-#endif /* PNG_USE_PNGGCCRD */
+#endif