HGE2D引擎+中文chm帮助文档.rar

/* 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 1.2.8 - December 3, 2004 * For conditions of distribution and use, see copyright notice in png.h * Copyright (c) 1998-2004 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... */ /* * 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 fi