FishMarket 1.0

home *** CD-ROM | disk | FTP | other *** search

/ FishMarket 1.0 / FishMarket v1.0.iso / fishies / 451-475 / disk_466 / dice / compiler.doc < prev next >

Wrap

Text File | 1992-05-06 | 14KB | 364 lines

compiler/compiler compiler/compiler compiler/COMPILER compiler/COMPILER DICE SYSTEM Matthew Dillon 891 Regal Rd. Berkeley, Ca. 94708 USA dillon@overload.Berkeley.CA.US uunet.uu.net!overload!dillon BIX: mdillon COMPILER FEATURES LIST AND ORGANIZATION NOTE: net distributed releases will not contain commodore includes or amiga*.lib (1) Organization of includes and libraries, default: DINCLUDE: location of top level includes such as stdio.h DINCLUDE:Amiga/ location of amiga includes (e.g. amiga/exec/types.h) (amiga includes not included) NOTE !!! Amiga includes should be placed in the subdirectory dinclude:Amiga, that is, dinclude:Amiga/Exec/*.h, dinclude:Amiga/graphics/*.h, etc... DCC references startup and libraries to DLIB: DLIB:amigas.lib A small-data model version of AMIGA.LIB DLIB:c.lib The main C library for which you have source code DLIB:m.lib The math library DLIB:auto.lib The auto library-open library DLIB:c.o The startup module DLIB:x.o The terminator module (2) Organization of includes and libraries, -2.0 option (usually set in the DCCOPTS enviroment variable) DINCLUDE:AMIGA20/ instead of DINCLUDE:AMIGA DLIB:amigas20.lib instead of DLIB:amigas.lib (3) DCC generated link lines If you give DICE the following line: 1> DCC foo.o -o foo It will run DLink with the following options: DLink dlib:c.o foo.o dlib:c.lib dlib:amigas.lib dlib:auto.lib dlib:x.o -o foo DLIB:C.O DICE startup code FOO.O your object module(s) DLIB:C.LIB DICE C.LIB DLIB:AMIGAS.LIB small data model version of commodore amiga.lib DLIB:AUTO.LIB DICE AUTO.LIB DLIB:X.O DICE section terminator code Some explanation is in order. AUTO.LIB catches any references to common library base variables when no storage has been declared and imports code to automatically open said libraries on startup and close them on exit. DICE uses this feature to import code to open and close "dos.library" when DOSBase is referenced, and also for the floating point libraries. X.O is used to supply an RTS at the end of the special autoinit and autoexit sections. Any module between C.O and X.O may declare these special sections to handle run-time relocations and other DICE features, and do so such that when the base of the section is JSR'd all the various module's code is sequenced through until the RTS in X.O is hit for the sections in question. (4) SMALL DATA AMIGA.LIB DICE uses a small-data model version of Commodore's AMIGA.LIB to allow it to generate residentable executables (-r option to DCC). The small data amiga.lib, called AMIGAS.LIB (note the 's') is generated from the large modem amiga.lib with the LIBTOS program. LIBTOS, including source, is included. The source is available to allow developers using DICE to add new base variables to LIBTOS's search list when new operating systems come out without having to wait for my update. ---------------------------------------------------------------------- COMPILER FEATURES LIST (0) ANSI. Pretty much ansi. Please report non-ansism problems. Check out the doc/KnownBugs list before reporting a problem. (1) autos are placed in registers based on weighted usage. A0,A1,D0,D1 will be used as register variables whenever possible, which usually reduces the number of registers that must be saved/restored in critical low level routines. The 'register' keyword is essentially ignored (use 'volatile' to force a variable to NOT be in a register). (2) LINK/UNLK is removed from low level routines that do not reference A5 and do not make any subroutine calls. (3) MOVEM is converted to MOVE or deleted entirely when possible. (4) Various obvious optimizations: use of BSET, BCLR, BTST as short cuts to logic operations, CMP/TST, etc... (5) Switch() statements are optimized as follows: * if the range of values is less than double the number of cases a lookup table will be used * if a small number of cases exists a sub[q]/beq sequence will be used * for a sufficiently large set of cases that cannot be made into a jump table, a binary subdivision method is used. (6) Branch optimization. Bxx to BRAs are optimized (up to 20 hops) by DAS, and massive optimization is done to multi-level conditionals by the main compiler pass. (7) Workbench support. The standard startup module supports the workbench. Specifically, the low level _main() in c.o does this. When a program is run from the workbench, a different entry point is called: wbmain(wbmsg) If you do not supply a wbmain() then the one from the c.lib library is used which immediately return(-1)s, exiting the program. The standard return from wbmain() or using exit() automatically ReplyMsg()s the workbench message for you. (8) _main() shortcut for extremely low level programs. You may overide c.lib's _main with your own. If you do so you may not use any c.lib memory allocation, stdio, or fd (open,close...) related routines. The startup module (c.o) will still run autoinits, autoexits, setup resident programs, and initialize your BSS space. _main(len, arg) (as passed to the program on startup) long len; char *arg; { } If running from the workbench, you must GetMsg() and process the workbench startup message yourself, as well return the message when you exit. _exit exists in c.o and cannot be overridden. If you use the _main override you may NOT call exit() unless you also override it. Calling exit() will generate references to stdio library code and bring it in from the library, making your executable larger unnecessarily. i.e. if you use _main() you should either fall through or use _exit() to exit. (9) AutoInit support. You may qualify a subroutine with the __autoinit keyword that will cause it to be automatically called by the startup code rather than having to explicitly call it from main. This feature is more useful to bring code in indirectly. For example, to automatically OpenLibrary() (& CloseLibrary()) floating point libraries if their base variables are referenced. Refer to the auto.lib source code for examples. __autoexit works the same way. Note that __autoinit routines are called BEFORE _main, just after autoinit libraries are openned, and __autoexit routines are called in __exit just before autoinit libraries are closed. (10) DCC runs fast. Make everything resident and watch it fly. Most of this speed comes from loading entire files into memory at once. This might present a memory problem, but in general DICE's memory usage is comparable to Lattice and Aztec. (11) CODE SIZE is comparable with Lattice and Aztec, even better in many cases. The minimum program sizes are as follows: _main() { } 448 no stdio/fd main() { } 2736 some stdio main() { printf..} 5492 most of stdio (12) EXTENSIONS , see EXTENSIONS.DOC. DICE provides many extensions to C to support various AMIGAisms and ROM based code, including special treatment of the 'const' storage qualifier. (13) The preprocessor is able to handle arbitrarily sized macros. There are no limits to symbol or macro length. The preprocessor is able to handle up to 64 levels of includes / macro recursions with arbitrary use of the string-ize (#) and token pasting (##) operators. The preprocessor implements the __BASE_FILE__ macro as well as all standard ANSI macros (__FILE__, __LINE__, __DATE__, __TIME__, ...) (14) DICE supports the creation of residentable executables in nearly all cases (exception: using the __geta4 procedure qualifier precludes being able to make an executable resident, also non const __chip data). Not only is residentability supported, but it is supported without requiring the programmer to #include all sorts of prototype/pragma files. (15) DICE supports the small-data model, large data model, small and large code models, and special data models (word absolute, absolute known address) for supporting ROMable code. The linker automatically generates JMP tables when the small-code model is used in large executables. (16) REGISTERED user's DICE supports __chip, __near, and __far keywords, as well as others. (17) Simplified library and startup module design. There are no multiple versions of a single library for different memory models and the frontend, DCC, makes all library and startup module interaction except m.lib invisible. The same c.o handles both resident and non-resident startup. c.lib and m.lib may be used with programs that declare more than 64KBytes of BSS (though not with programs that declare more than 64KBytes of __near initialized storage). There is no large-data-model version of c.lib and m.lib, although you can easily recompile the libraries into a large-data model version if you wish. I suggest that instead of doing that you stick with the small-data model and use the __geta4 type qualifier to procedures which are called out of context. (18) Optimization of args to prototyped routines -------------------------------------------------------------------------- GOALS In writing DICE my goals are as follows: - reasonably fast compilation. Modular executables for ease of use, reliability, and testability. - concentrate more on reliability and less on optimizations. Remember that this is only the third major distribution of DICE, I expect some bugs to show up. I expect to become more reliable than the two other commercial C compilers in the Amiga market within a year. - but do not forget optimizations... put in relatively easy to implement optimizations that do not destroy the reliability of DICE. DICE does no common sub-expression or loop unrolling optimizations, but does do relatively smart register allocation and multi-level history to propogate conditional expressions. - provide comprehensive support of the Amiga, especially for new versions of the OS that come out. - I have always torn my hair out at not being able to easily fix bugs in the support libraries for commercial compilers. DICE includes full source for its support libraries (c.lib, m.lib, auto.lib) and a means to remake the library. DICE also includes full source to some of its own utilities, namely LIBTOS and the DCC frontend which are the most likely candidates for programmer hair loss. ------------------------------------------------------------------------ COMMENTS Registerized parameters are not implemented. While such a feature would make an executable somewhat smaller and a little faster, a few percent of gain is not worth the complexity of the implementation (and thus the possibility of introduced bugs). The difference in execution speed is generally so miniscule that it isn't noticable, and where it is noticable it is usually due to a badly implemented algorithm in the first place. So, cute, but not as much of a boost as you might think. Inline library calls are not implemented either. Inline library calls are actually useful and when properly implemented increase efficiency to low level calls such as AddHead() and GetMsg(). Other calls such as Move() and Draw() do not increase noticably in efficiency. #pragma's must be used to define inline calls and this makes the compiler enviroment less portable. Lattice made a big mistake *requiring* the use of inline library calls with residentable programs. My solution was to write a program to convert AMIGA.LIB into a small-data version of same called AMIGAS.LIB. In general, I refuse to implement a thousand nearly useless features that will only introduce more bugs into the compiler. ------------------------------------------------------------------------ COMPATIBILITY ANSI C DICE is getting closer to being ANSI C compliant, there are only a few things missing. I spent a great deal of time ensuring that STDIO routines run relatively fast. I decided to write nearly all of the support library in C instead of falling back to optimized assembly to keep the system uniform and portable. One does not notice much of a difference between the C strcpy() and an assembly strcpy() relative to the run time of their program. AMIGA As said in feature (7), if you wish your program to be runnable from the workbench you must supply a wbmain() entry point. If you do not then running the program from the workbench will result in its immediate termination (e.g. nothing happens). On return from wbmain() or exit() the workbench message is automatically ReplyMsg()d by the exit code. DICE separates workbench startup and puts it in the hands of the user to simplify the user's code. Since there are two entry points, main() for CLI run programs and wbmain() for WORKBENCH-run programs, the programmer can more easily modularize his code. The Registered release of DICE supports the __near, __far, and __chip keywords. UNBUFFERED CONSOLE IO You may set a console to RAW mode using the setbuf() and setvbuf() calls. You may set a console back to COOKED mode using the same calls. ------------------------------------------------------------------------ Your first program 1> cd examples 1> dcc hello.c -o ram:hello 1> ram:hello hello world 1> Your first residentable program 1> cd examples 1> dcc hello.c -o ram:hello -r 1> ram:hello hello world 1> resident ram:hello 1> hello ; instantanious execution hello world 1> * NOTE, a common problem is that users have removed various floating point libraries from LIBS: and this will prevent DC1 from running. Make sure you have the full complement of floating point libraries in LIBS: