Main Page | Class Hierarchy | Class List | Directories | File List | Class Members | File Members | Related Pages

AggressiveOptimize.h

00001 
00003 // Version 1.40 
00004 // October 22nd, 2002 - .NET (VC7, _MSC_VER=1300) support!
00005 // Version 1.30
00006 // Nov 24th, 2000
00007 // Version 1.20
00008 // Jun 9th, 2000
00009 // Version 1.10
00010 // Jan 23rd, 2000
00011 // Version 1.00
00012 // May 20th, 1999
00013 // Todd C. Wilson, Fresh Ground Software
00014 // (todd@nopcode.com)
00015 // This header file will kick in settings for Visual C++ 5 and 6 that will (usually)
00016 // result in smaller exe's.
00017 // The "trick" is to tell the compiler to not pad out the function calls; this is done
00018 // by not using the /O1 or /O2 option - if you do, you implicitly use /Gy, which pads
00019 // out each and every function call. In one single 500k dll, I managed to cut out 120k
00020 // by this alone!
00021 // The other two "tricks" are telling the Linker to merge all data-type segments together
00022 // in the exe file. The relocation, read-only (constants) data, and code section (.text)
00023 // sections can almost always be merged. Each section merged can save 4k in exe space,
00024 // since each section is padded out to 4k chunks. This is very noticeable with smaller
00025 // exes, since you could have only 700 bytes of data, 300 bytes of code, 94 bytes of
00026 // strings - padded out, this could be 12k of runtime, for 1094 bytes of stuff! For larger
00027 // programs, this is less overall, but can save at least 4k.
00028 // Note that if you're using MFC static or some other 3rd party libs, you may get poor
00029 // results with merging the readonly (.rdata) section - the exe may grow larger.
00030 // To use this feature, define _MERGE_DATA_ in your project or before this header is used.
00031 // With Visual C++ 5, the program uses a file alignment of 512 bytes, which results
00032 // in a small exe. Under VC6, the program instead uses 4k, which is the same as the
00033 // section size. The reason (from what I understand) is that 4k is the chunk size of
00034 // the virtual memory manager, and that WinAlign (an end-user tuning tool for Win98)
00035 // will re-align the programs on this boundary. The problem with this is that all of
00036 // Microsoft's system exes and dlls are *NOT* tuned like this, and using 4k causes serious
00037 // exe bloat. This is very noticeable for smaller programs.
00038 // The "trick" for this is to use the undocumented FILEALIGN linker parm to change the
00039 // padding from 4k to 1/2k, which results in a much smaller exe - anywhere from 20%-75%
00040 // depending on the size. Note that this is the same as using /OPT:NOWIN98, which *is*
00041 // a previously documented switch, but was left out of the docs for some reason in VC6 and
00042 // all of the current MSDN's - see KB:Q235956 for more information.
00043 // Microsoft does say that using the 4k alignment will "speed up process loading",
00044 // but I've been unable to notice a difference, even on my P180, with a very large (4meg) exe.
00045 // Please note, however, that this will probably not change the size of the COMPRESSED
00046 // file (either in a .zip file or in an install archive), since this 4k is all zeroes and
00047 // gets compressed away.
00048 // Also, the /ALIGN:4096 switch will "magically" do the same thing, even though this is the
00049 // default setting for this switch. Apparently this sets the same values as the above two
00050 // switches do. We do not use this in this header, since it smacks of a bug and not a feature.
00051 // Thanks to Michael Geary <Mike@Geary.com> for some additional tips!
00052 //
00053 // Notes about using this header in .NET
00054 // First off, VC7 does not allow a lot of the linker command options in pragma's. There is no
00055 // honest or good reason why Microsoft decided to make this change, it just doesn't.
00056 // So that is why there are a lot of <1300 #if's in the header.
00057 // If you want to take full advantage of the VC7 linker options, you will need to do it on a 
00058 // PER PROJECT BASIS; you can no longer use a global header file like this to make it better.
00059 // Items I strongly suggest putting in all your VC7 project linker options command line settings:
00060 //          /ignore:4078    /RELEASE
00061 // Compiler options:
00062 //          /GL (Whole Program Optimization)
00063 // If you're making an .EXE and not a .DLL, consider adding in:
00064 //          /GA (Optimize for Windows Application)
00065 // Some items to consider using in your VC7 projects (not VC6):
00066 // Link-time Code Generation - whole code optimization. Put this in your exe/dll project link settings.
00067 //          /LTCG:NOSTATUS
00068 // The classic no-padding and no-bloat compiler C/C++ switch:
00069 //          /opt:nowin98
00070 //
00071 // (C++ command line options:  /GL /opt:nowin98  and /GA for .exe files)
00072 // (Link command line options: /ignore:4078 /RELEASE /LTCG:NOSTATUS)
00073 //
00074 // Now, notes on using these options in VC7 vs VC6.
00075 // VC6 consistently, for me, produces smaller code from C++ the exact same sources,
00076 // with or without this header. On average, VC6 produces 5% smaller binaries compared
00077 // to VC7 compiling the exact same project, *without* this header. With this header, VC6
00078 // will make a 13k file, while VC7 will make a 64k one. VC7 is just bloaty, pure and
00079 // simple - all that managed/unmanaged C++ runtimes, and the CLR stuff must be getting
00080 // in the way of code generation. However, template support is better, so there.
00081 // Both VC6 and VC7 show the same end kind of end result savings - larger binary output
00082 // will shave about 2% off, where as smaller projects (support DLL's, cpl's,
00083 // activex controls, ATL libs, etc) get the best result, since the padding is usually
00084 // more than the actual usable code. But again, VC7 does not compile down as small as VC6.
00085 //
00086 // The argument can be made that doing this is a waste of time, since the "zero bytes"
00087 // will be compressed out in a zip file or install archive. Not really - it doesn't matter
00088 // if the data is a string of zeroes or ones or 85858585 - it will still take room (20 bytes
00089 // in a zip file, 29 bytes if only *4* of them 4k bytes are not the same) and time to
00090 // compress that data and decompress it. Also, 20k of zeros is NOT 20k on disk - it's the
00091 // size of the cluster slop- for Fat32 systems, 20k can be 32k, NTFS could make it 24k if you're
00092 // just 1 byte over (round up). Most end users do not have the dual P4 Xeon systems with
00093 // two gigs of RDram and a Raid 0+1 of Western Digital 120meg Special Editions that all
00094 // worthy developers have (all six of us), so they will need any space and LOADING TIME
00095 // savings they will need; taking an extra 32k or more out of your end user's 64megs of
00096 // ram on Windows 98 is Not a Good Thing.
00097 //
00098 // Now, as a ADDED BONUS at NO EXTRA COST TO YOU! Under VC6, using the /merge:.text=.data
00099 // pragma will cause the output file to be un-disassembleable! (is that a word?) At least,
00100 // with the normal tools - WinDisam, DumpBin, and the like will not work. Try it - use the
00101 // header, compile release, and then use DUMPBIN /DISASM filename.exe - no code!
00102 // Thanks to Gëzim Pani <gpani@siu.edu> for discovering this gem - for a full writeup on
00103 // this issue and the ramifactions of it, visit www.nopcode.com for the Aggressive Optimize
00104 // article.
00105 
00106 #ifndef _AGGRESSIVEOPTIMIZE_H_
00107 #define _AGGRESSIVEOPTIMIZE_H_
00108 
00109 #pragma warning(disable:4711)
00110 
00111 #ifdef NDEBUG
00112 // /Og (global optimizations), /Os (favor small code), /Oy (no frame pointers)
00113 #pragma optimize("gsy",on)
00114 
00115 #if (_MSC_VER<1300)
00116     #pragma comment(linker,"/RELEASE")
00117 #endif
00118 
00119 // Note that merging the .rdata section will result in LARGER exe's if you using
00120 // MFC (esp. static link). If this is desirable, define _MERGE_RDATA_ in your project.
00121 #ifdef _MERGE_RDATA_
00122 #pragma comment(linker,"/merge:.rdata=.data")
00123 #endif // _MERGE_RDATA_
00124 
00125 #pragma comment(linker,"/merge:.text=.data")
00126 #if (_MSC_VER<1300)
00127     // In VC7, this causes problems with the relocation and data tables, so best to not merge them
00128     #pragma comment(linker,"/merge:.reloc=.data")
00129 #endif
00130 
00131 // Merging sections with different attributes causes a linker warning, so
00132 // turn off the warning. From Michael Geary. Undocumented, as usual!
00133 #if (_MSC_VER<1300)
00134     // In VC7, you will need to put this in your project settings
00135     #pragma comment(linker,"/ignore:4078")
00136 #endif
00137 
00138 // With Visual C++ 5, you already get the 512-byte alignment, so you will only need
00139 // it for VC6, and maybe later.
00140 #if _MSC_VER >= 1000
00141 
00142 // Option #1: use /filealign
00143 // Totally undocumented! And if you set it lower than 512 bytes, the program crashes.
00144 // Either leave at 0x200 or 0x1000
00145 //#pragma comment(linker,"/FILEALIGN:0x200")
00146 
00147 // Option #2: use /opt:nowin98
00148 // See KB:Q235956 or the READMEVC.htm in your VC directory for info on this one.
00149 // This is our currently preferred option, since it is fully documented and unlikely
00150 // to break in service packs and updates.
00151 #if (_MSC_VER<1300)
00152     // In VC7, you will need to put this in your project settings
00153     #pragma comment(linker,"/opt:nowin98")
00154 #else
00155 
00156 // Option #3: use /align:4096
00157 // A side effect of using the default align value is that it turns on the above switch.
00158 // Does nothing under Vc7 that /opt:nowin98 doesn't already give you
00159 // #pragma comment(linker,"/ALIGN:512")
00160 #endif
00161 
00162 #endif // _MSC_VER >= 1000
00163 
00164 #endif // NDEBUG
00165 
00166 #endif //  _AGGRESSIVEOPTIMIZE_H_

Generated on Wed Oct 26 14:46:48 2005 for Lgi by  doxygen 1.4.1