vbcc - C compiler (c) in 1995-96 by Volker Barthelmann INTRODUCTION vbcc is a free portable and retargetable ANSI C compiler. It is clearly split into a target independant and a target dependant part and supports emulating datatypes of the target machine on any other machine so that it is possible to e.g. make a crosscompiler for a 64bit machine on a 32bit machine. This document only deals with the target dependant parts of the Amiga68k version. LEGAL vbcc is (c) in 1995-96 by Volker Barthelmann. All code is written by me and may be freely redistributed as long as no modifications are made and nothing is charged for it. Non-commercial usage of vbcc is allowed without any restrictions. Commercial usage needs my written consent. Sending me money, gifts, postcards etc. would of course be very nice and may encourage further development of vbcc, but is not legally or morally necessary to use vbcc. ADDITIONAL OPTIONS FOR THIS VERSION -cpu=n Generate code for cpu n (e.g. -cpu=68020), default: 68000 -fpu=n Generate code for fpu n (e.g. -fpu=68881), default: 0 -sd Use small data model (see below). -sc Use small code model (see below). -d2scratch Use d2 as scratch register (for testing with certain C libraries). -noa4 Don't use register a4 (for testing with certain C libraries). -prof Insert code for profiling (not really usable yet). -const-in-data By default constant data will be placed in the code section (and therefore is accessable with faster pc-relative addressing modes). Using this option it will be placed in the data section. This could e.g. be useful if you want to use small data and small code, but your code gets too big with all the constant data. -use-framepointer By default automatic variables are addressed through a7 instead of a5. This generates slightly better code, because the function entry and exit overhead is reduced and a5 can be used as register variable etc. However this may be a bit confusing when debugging and you can force vbcc to use a5 as a fixed framepointer. -no-addressing-modes The intermediate code does not contain any of the 68k addressing modes, so if they are to be used an extra pass over the intermediate code is necessary to recognize certain patterns that can be expressed using a 68k addressing mode. By default vbcc tries to use some 68k addressing modes. Currently (ax)+ and subsets of (displ,ax,dy*skal) are used. However not all cases where those addressing modes could be used are recognized. With this option you can prevent vbcc from searching for possible addressing modes. -no-delayed-popping By default arguments of function calls are not always popped from the stack immediately after the call, so that the arguments of several calls may be popped at once. With this option vbcc can be forced to pop them after every function call. This may simplify debugging and very slightly reduce the stack size needed by the compiled program. SOME INTERNALS The current version generates assembler output for use with the PhxAss assembler (c) by Frank Wille. Most peephole optimizations are done by the assembler so vbcc only does some that the assembler cannot make. The generated executables will probably only work with OS2.0 or higher. The registers d0, d1, a0, a1, fp0 and fp1 are used as scratch registers (i.e. they can be destroyed in function calls), all other registers are preserved. All elementary types up to 4 bytes are returned in register d0 like common on the Amiga (although I think pointers should better be returned in a0). If compiled for an fpu, floating point values are returned in fp0. All other types are returned by passing the function the address of the result as a hidden argument - so when you call such a function without a proper declaration in scope you can expect a crash. vbcc uses d0-d7 and a0-a6 for temporary results and register variables (a4 is used as small data pointer if -sd is used). a5 can be used as frame pointer for automatic variables (but this is not necessary - they can be accessed through a7, too). At the moment all local variables are addressed via (dist,ax), so a function may have only ~32k of local variables if code for <=68000 is generated. The elementary data types are represented like: type size in bits alignment in bytes char 8 1 short 16 2 int 32 2 long 32 2 all pointers 32 2 float(fpu) 32 2 see below double(fpu) 64 2 see below Although it would be better to have all 32bit+ types aligned to 4 bytes I chose 2 bytes to be compatible with the Amiga system structures which unfortunately have longwords aligned to 4n+2-addresses. The amiga68k code generator at the moment only works on systems that store floats and doubles in a similar way (IEEE) like the Amiga. SMALL DATA vbcc can access static data in two ways. By default all such data will be accessed with full 32bit addresses (large data model). However there is a second way. You can set up an address register (a4) to point into your data segment and then address data with a 16bit offset through this register. The advantages of the small data model are that your program will usually be smaller (because the 16bit offsets use less space and no relocation information is needed) and faster. The disadvantages are that one address register cannot be used by the compiler and that you can use it only if all your static data occupies less than 64kb. Also you may not mix object modules and libraries that have been compiled with different data models (you can call functions compiled with large data model from object files compiled with small data model, but not vice versa and only functions can be called that way - other data cannot be accessed) and you probably have to use PhxLnk then. SMALL CODE If you use the small code model calls to external functions (i.e. from libraries or other object files) are done with 16bit offsets over the program counter rather than with absolute 32bit addresses. The advantage is slightly smaller and faster code. The disadvantages are that all the code (including library functions) must be small enough and that you may have to use PhxLnk. However you can link objects/libraries together if they have been compiled with different code models. CPUs At the moment the values of -cpu=n have those effects: n>=68020: - 32bit multiplication/division/modulo is done with the mul?.l, div?.l and div?l.l instructions - tst.l ax is used - extb.l dx is used - 16/32bit offsets are used in certain addressing modes - link.l is used - addressing modes with scaling are used - (dx) is used if no address register is available (not yet) n>=68040 - 8bit constants are not copied in data registers FPUs At the moment the values of -fpu=n have those effects: n>68000: - floating point calculations are done using the fpu n=68040: n=68060: - instructions that have to be emulated on these fpus will not be used; at the moment this only includes the fintrz instruction in case of the 040 MATH Integer math hopefully works without problems on all cpus. Long multiply on cpus <68020 uses inline routines. This may increase code size a bit, but it should be significantly faster, because function call overhead is not necessary and the compiler can use the registers which already contain the sources and/or need the result. However if anyone really wants an option for using library routines for multiply, this can easily be implemented. Long division and modulo is handled by calls to library functions. At the moment standard library calls with parameter passing via stack are used. This is rather slow, but division takes quite some time anyway. (mult/div/mod with constant powers of two are replaced by corresponding bitwise operations (mod only if the other operand is unsigned), but sums of powers of two not yet). If no FPU is specified floating point math is done using the C= math libraries. 32 bit IEEE format is used for float and 64 bit IEEE for double. Float return values are passed in d0 and double is passed via pointers. This is rather slow, so if you need good floating point performance specify an FPU. Floating point math is done with the FPU if one is specified (see above). Floating point values are kept in registers then and therefore may have extended precision sometimes, which is not ANSI compliant (but will usually cause no harm). When floating point values are stored in memory they use the same IEEE formats as without FPU. Float or double return values are passed in fp0. Note that you must not link object files together if they were not compiled with the same -fpu settings and that you have to link with the proper math library (see vclib.doc). KNOWN PROBLEMS - Converting floating point values to unsigned integers is not correct if the value is >LONG_MAX and code for an FPU is generated. - The extended precision of the FPU registers can cause problems if a program depends on the exact precision. Most programs will not have trouble with that, but programs which do exact comparisons with floating point types (e.g. to try to calculate the number of significant bits) may not work as expected (especially if the optimizer was turned on). - Alignment of arrays of char is incompatible with the Amiga include files. AFAIK this will only conflict with the cia-includes. - The Amiga68k code-generator is not fully portable, but depends on the host machine having similar data types like the Amiga. Also it currently reads floating point variables via an int * which is not strictly conforming. Volker Barthelmann volker@vb.franken.de Kennedy-Ring 39 91301 Forchheim Germany