(c) in 1995-97 by Volker Barthelmann This document is under construction! Not for public distribution! This document describes some of the internals of vbcc and tries to explain what must be done to write a code generator for vbcc. However if someone wants to write one, I suggest to contact me first, so that it can be integrated into the source tree. You should only have to rewrite/change the files machine.c and machine.h. It is probably best to choose an appropriate one from the generic* ones and only change something or add something and leave the rest untouched. From now on integer means any of {char, short, int, long} or their unsigned couterparts. Arithmetic means integer or float or double. Elementary type means arithmetic or pointer. If You intend to write a code generator for a machine with different kinds of pointers You might have some problems. THE INTERMEDIATE CODE vbcc will generate intermediate code for every function and pass this code to the code generator which has to convert it into the desired output. In the future there may be a code generator generator which reads a machine description file and generates a code generator from that, but it is not clear whether this could simplify much without taking penalties in the generated code. Anyway this would be a layer on top of the current interface to the code generator so that the interface described in this document would still be valid and accessable. The intermediate code is represented as a doubly linked list of quadruples (I am calling them ICs from now on) consisting mainly of an operator, two source operands and a target. They are represented like this: struct IC{ struct IC *prev; struct IC *next; int code; int typf; [...] struct obj q1; struct obj q2; struct obj z; [...] }; The only members relevant to the code generator are 'prev', 'next', 'code', 'typf', 'q1', 'q2' and 'z'. 'prev' and 'next' are pointers to the previous and next IC. The first IC has 'prev'==0 and the last one has 'next'==0. 'typf' is the type of the operands of this IC. This can be one of: #define CHAR 1 #define SHORT 2 #define INT 3 #define LONG 4 #define FLOAT 5 #define DOUBLE 6 #define VOID 7 #define POINTER 8 #define ARRAY 9 #define STRUCT 10 #define UNION 11 #define ENUM 12 /* not relevant for code generator */ #define FUNKT 13 and can be additionally or'ed by #define UNSIGNED 16 #define CONST 64 #define VOLATILE 128 #define UNCOMPLETE 256 However only UNSIGNED is of real importance for the code generator. 'typf'&15 yields the type without any qualifiers. 'q1', 'q2' and 'z' are the source1 (quelle1 in German), source2 and target (ziel). If a result has to be computed, it always will be stored in the object 'z' and the objects 'q1' and 'q2' may not be destroyed during this operation. The objects are described by this structure. struct obj{ int flags; int reg; struct Var *v; struct AddressingMode *am; union atyps{ zchar vchar; zchar vuchar; zshort vshort; zushort vushort; zint vint; zuint vuint; zlong vlong; zulong vulong; zfloat vfloat; zdouble vdouble; zpointer vpointer; }val; }; 'flags' describes what kind the object is. It can be a combination of #define VAR 1 The object is a variable. The pointer to its struct Var is in 'v'. 'val.vlong' vontains an offset that has to be added to it. In certain cases (e.g. flags==VAR|REG) where an offset is useless, it can always be taken as 0. A struct Var looks like: struct Var{ int storage_class; [...] char *identifier; [...] zlong offset; [...] }; The relevant entries are: 'identifier': The name of the variable. Usually only of interest for variables with external-linkage. 'storage_class': One of: #define AUTO 1 #define REGISTER 2 #define STATIC 3 #define EXTERN 4 #define TYPEDEF 5 /* not relevant */ If the variable is not assigned to a register (i.e. bit REG is not set in the flags of the struct obj) then the variable can be addressed in the following ways (with examples of 68k-code): 'storage_class' == AUTO or 'storage_class' == REGISTER: 'offset' contains the offset inside the local-variables section. The code generator must decide how it's going to handle the activation record. If 'offset' < 0 then the variable is a function argument on the stack. In this case the offset in the parameter-area is - ('offset' + 'maxalign'). The code generator may have to calculate the actual offset to a stack- or frame-pointer from the value in 'offset'. 'offset'+'val.vlong'(sp) Note that 'storage_class' REGISTER is equivalent to AUTO - whether the variable is actually assigned a register is specified by the bit REG in the 'flags' of the 'struct obj'. 'storage_class' == EXTERN The variable can be addressed through its name in 'identifier'. 'val.vlong'+_'identifier' 'storage_class' == STATIC The variable can be addressed through a numbered label. The label number is stored in 'offset'. 'val.vlong'+l'offset' #define KONST 2 The object is a constant. Its value is in the corresponding (to 'typf') member of 'val'. #define DREFOBJ 32 The content of the location in memory the object points to is used. #define REG 64 The object is a register. 'reg' contains its number. #define VARADR 128 The address of the object is to be used. Only together with static variables (i.e. 'storage_class' STATIC or EXTERN). Certain combination of these flags are possible, e.g. VAR|REG, VAR|DREFOBJ, VAR|REG|DREFOBJ. 'flags' can be 0. Also some other bits which are not relevant to the code generator may be set. Constants will usually be in 'q2' if possible. One of the sources always is not constant and the target is always an lvalue. Unless otherwise specified all operands of an IC are of the type 'typf' (which may be further restricted by 'code'). However not all objects must be used. This depends on 'code' and is listed below. In most cases (i.e. when not explicitly stated) 'typf' is an elementary type (i.e. arithmetic or pointer). 'am' can be used to store information on special addressing modes. This has to be handled by the by the code generator. However 'am' has to be 0 or has to point to a struct AddressingMode that was allocated using malloc() when the code generator returns. 'val' stores either the value of the object if it is a constant or an offset if it is a variable. 'code' describes the operation and can be one of: #define ASSIGN 2 Copy 'q1' to 'z'. 'q2.val.vlong' contains the size of the objects (this is necessary if it is an array or a struct). 'typf' does not have to be an elementary type! The only case where 'typf' == ARRAY should be in automatic initializations. It is also possible that ('typf'&15) == CHAR but the size is != 1. This is created for an inline memcpy/strcpy where the type is not known. #define OR 16 #define XOR 17 #define AND 18 Bitwise boolean operations. q1,q2->z. All operands are integers. #define LSHIFT 25 #define RSHIFT 26 Bit shifting. q1,q2->z. 'q2' is the number of shifts. All operands are integers. AFAIK the behaviour is implementation defined if 'q2' is negative. #define ADD 27 #define SUB 28 #define MULT 29 #define DIV 30 Standard arithmetic operations. q1,q2->z. All operands are of arithmetic types (integers or floating point). #define MOD 31 Modulo (%). q1,q2->z. All operands are integers. #define KOMPLEMENT 33 Bitwise complement. q1->z. All operands are integers. #define MINUS 38 Unary minus. q1->z. All operands are of arithmetic types (integers or floating point). #define ADDRESS 40 Get the address of an object. q1->z. 'z' is always a pointer and 'q1' is always an automatic variable. #define CALL 42 Call the function 'q1'. Currently 'q1' is a function rather than a pointer to a function. This may change in the future. 'q2.val.vlong' contains the number of bytes pushed on the stack as function arguments. Those may have to be popped from the stack after the function returns depending on the calling mechanism. #define CONVCHAR 50 #define CONVSHORT 51 #define CONVINT 52 #define CONVLONG 53 #define CONVFLOAT 54 #define CONVDOUBLE 55 #define CONVPOINTER 57 #define CONVUCHAR 58 #define CONVUSHORT 59 #define CONVUINT 60 #define CONVULONG 61 Convert one type to another. q1->z. 'z' is always of the type 'typf'. 'q1' is a short in CONVSHORT and an unsigned long in CONVULONG etc. Conversions floating point<->pointers do not occur. #define ALLOCREG 65 From now on the register 'q1.reg' is in use. No code has to be generated for this, but it is probably necessary to keep track of the registers in use to know which registers are available for the code generator at a time and which registers the function trashes. #define FREEREG 66 From now on the register 'q1.reg' is free. Also it means that the value currently stored in 'q1.reg' is not used any more and therefore provides a little bit of data flow information. Note however that if a FREEREG follows a branch the value of the register may be used at the target of the branch. #define COMPARE 77 Compare and set condition codes. q1,q2(->z). Compare the operands and set the condition code, so that BEQ, BNE, BLT, BGE, BLE or BGT works. If 'z.flags' == 0 then the condition codes will be evaluated immediately after the COMPARE, i.e. the next instruction (except possible FREEREGs) will be a conditional branch. However if a target supports several condition code registers and sets the global variable 'multiple_ccs' to 1 vbcc might use those registers and perform certain optimizations. Then 'z' may be non-empty and the condition codes have to be stored in 'z' (this will usually be a condition code register). #define TEST 68 Test 'q1' to 0 and set condition codes. q1. This is equal to COMPARE 'q1',(corresponding constant 0) but only the condition code for BEQ and BNE has to be set. #define LABEL 69 Generate a label. 'typf' specifies the number of the label. #define BEQ 70 #define BNE 71 #define BLT 72 #define BGE 73 #define BLE 74 #define BGT 75 Branch on condition codes. (q1) 'typf' specifies the label where program execution shall continue, if the condition code is true (otherwise continue with next statement). The condition codes mean equal, not equal, less than, greater or equal, less or equal and greater than. If 'q1' is empty then the codes set by the last COMPARE or TEST must be evaluated. Otherwise 'q1' contains the condition codes. On some machines the type of operands of a comparison (e.g unsigned or signed) is encoded in the branch instructions rather than in the comparison instructions. In this case the code generator has to keep track of the type of the last comparison. #define BRA 76 Branch always. 'typf' specifies the label where program execution continues. #define PUSH 78 Push q1 on the stack. q1. 'q2.val.vlong' contains the size of the object and 'q1' does not have to be an elementary type (see ASSIGN). This is only used for passing function arguments. #define ADDI2P 81 Add an integer to a pointer. q1,q2->z. 'q1' and 'z' are always pointers and 'q2' is an integer of type 'typf'. 'z' has to be 'q1' increased by 'q2' bytes. #define SUBIFP 82 Subtract an Integer from a pointer. q1,q2->z. 'q1' and 'z' are always pointers and 'q2' is an integer of type 'typf'. 'z' has to be 'q1' decreased by 'q2' bytes. #define SUBPFP 83 Subtract a pointer from a pointer. q1,q2->z. 'q1' and 'q2' is a pointer and 'z' is an integer of type 'typf'. 'z' has to be 'q1' - 'q2' in bytes. #define GETRETURN 93 Get the return value of the last function call. ->z. If the return value is in a register this will be in 'q1.reg'. Otherwise 'q1.reg' will be 0. This follows immediately after a CALL instruction (except possible FREEREGs). #define SETRETURN 94 Set the return value of the current function. q1. If the return value is in a register this will be in 'z.reg'. Otherwise 'z.reg' will be 0. This is immediately followed by a function exit (i.e. it is the last IC or followed by an unconditional branch to a label which is the last IC - always ignoring FREEREGs). #define MOVEFROMREG 95 Move a register to memory. q1->z. 'q1' is always a register and 'z' an array of size 'regsize[q1.reg]'. #define MOVETOREG 96 Load a register from memory. q1->z. 'z' is always a register and 'q1' an array of size 'regsize[z.reg]'. #define NOP 97 Do nothing. MACHINE.H Only a few changes should be necessary to machine.h. TARGET DATA TYPES First you need a typedef every elementary type of the target machine. These are: zchar type char on the target machine zuchar type unsigned char on the target machine zshort ... zushort zint zuint zlong zulong zfloat zdouble zpointer a byte pointer on the target machine If the host machine has an elementary type which can represent a type of the target machine you can use that. However if there is none (e.g. if you want to build a cross-compiler for a 64bit machine on a 32bit host) you have to define it as a struct (if you use an array it has to be embedded in a struct because vbcc assigns those types and passes them as function arguments etc.) and provide some arithmetic functions. In the future there will perhaps be an automated installation procedure which checks if the host machine offers appropriate data types and installs emulation routines otherwise if possible. However at the moment this must be done manually for every host. If you could typedef every type to a host type you can copy the bunch of cryptic #defines from the beginning of an example machine.h and skip the next part about target arithmetic. TARGET ARITHMETIC Now you have to provide a lot of functions performing operations using the target machine's arithmetic. You can look them up in any machine.h. E.g. zladd() takes two zlongs and returns their sum as zlong. zuladd() does the same with zulongs, zdadd() with doubles. No functions for smaller types are needed because vbcc will calculate with the wider types and convert the results down if needed. If you could typedef any target type to an elementary type of the host machine you can simply provide a macro, e.g. #define zladd(x,y) ((x)+(y)). Note that you _must_ provide prototypes for all functions which are not macros. Now you must provide conversion functions which convert between types of the target machine. E.g. zl2zc takes a zlong and returns it converted to a zchar. Again look at an example machine.h to see which ones are needed. A few functions for converting between target and host types are also necessary, e.g. l2zl takes a long and returns it converted to a zlong. At last we need functions for comparing target data types. E.g. zlleq(a,b) must return true if zlong a <= zlong b and false otherwise. zleqto(a,b) must return true if zlong a == zlong b and false otherwise. ADDRESSING-MODES If the code generator supports extended addressing modes You have to think of a way to represent them and define the structure AddressingMode so that all Modes can be stored in it. The machine independant part of vbcc will not use these modes, so Your code generator has to find a way to combine several statements to make use of these modes. BASIC PARAMETERS #define MAXR to the number of available registers. #define MAXGF to the number of command line flags that can be used to configure the behaviour of the code generator. This must be at least one even if you do not use any flags. #define USEQ2ASZ as 0 or 1; if it is set to 0, no ICs where 'q2' == 'z' will be generated. This is because those ICs might be hard to implement efficiently on certain CPUs. #define MINADDI2P to the smallest integer type (i.e. CHAR, SHORT or INT) that can be added to a pointer. Smaller types will be automatically converted to MINADDI2P when they are to be added to a pointer. This may be subsumed by shortcut() in the future. #define BIGENDIAN as 1 if integers are represented in big endian, i.e. the most significant byte is at the lowest memory address, the least significant byte at the highest. #define LITTLEENDIAN as 1 if integers are represented in little endian, i.e. the least significant byte is at the lowest memory address, the most significant byte at the highest. #define SWITCHSUBS as 1 if switch-statements should be compiled into a series of SUB/TEST/BEQ instructions rather than COMPARE/BEQ. This may be useful if the target has a more efficient SUB-instruction (e.g. 68k). #define INLINEMEMCPY to the largest size in bytes allowed for inline memcpy. Calls to memcpy/strcpy with a known size smaller than INLINEMEMCPY may be replaced by a single ASSIGN IC by vbcc. This may be replaced by a variable of type zlong in the future. MACHINE.C This is the main part of the code generator. The following changes have to be made: COMMANDLINE OPTIONS You can use code generator specific commandline options. The number of flags is specified as MAXGF in machine.h. Insert the names for the flags as char *g_flags_name[MAXGF]. If an option was specified (g_flags[i]&USEDFLAG) is not zero. In int g_flags[MAXGF] You can also choose how the options are to be used: 0 The option can only be specified. E.g. if g_flags_name[2]=="myflag", the commandline may contain "-myflag" and (g_flags[2]&USEDFLAG)!=0. VALFLAG The option must be specified with an integer constant, e.g. "-myflag=1234". This value can be found in g_flags_val[2].l then. STRINGFLAG The option must be specified with a string, e.g. "-myflag=Hallo". The pointer to the string can be found in g_flags_val[2].p then. DATA TYPES The array zlong align[16] must contain the necessary alignments for every type in bytes. Some of the entries in this array are not actually used, but align[type&15] must yield the correct alignment for every type. align[CHAR] must be 1. zlong maxalign; must be set to an alignment in bytes that is ok for every type. The array zlong sizetab[16] must contain the sizes of every type in bytes. The array zlong t_min[32] must contain the smallest number for every integer type (including unsigned ones). The array zulong t_max[32] must contain the greatest number for every integer type (including unsigned ones). As zlong and zulong may be no elementary types on the host machine those arrays have to be initialized dynamically. Also note that if you want the code generator to be portable those values may not be representable as constants by the host architecture and have to be calculated using the functions for arithmetic on the target's data types. E.g. the smallest representable value of a 32bit twos-complement data type is not guaranteed to be valid on every ANSI C implementation. Also note that you may not use simple operators on the target data types but you have to use the functions or convert them to an elementary type of the host machine before (if you know that it is representable as such). REGISTER SET The valid registers are numbered from 1..MAXR. The array char *regnames[MAXR+1] must contain the names for every register. zlong regsize[MAXR+1] must contain the size of each register in bytes. This is used to create storage if registers have to be saved. int regscratch[MAXR+1] must contain information whether a register is a scratchregister i.e. may be destroyed during a function call (1 or 0). vbcc will generate code to save/restore all scratch-registers which are assigned a value when calling a function. However if the code generator uses additional scratch-registers it has to take care to save/restore them. Also the code generator must save/restore used non-scratch-registers on function entry/exit. int regsa[MAXR+1] must contain information whether a register is in use or not at the beginning of a function (1 or 0). The compiler will not use any of those registers for register variables or temporaries. You _must_ set regsratch[i] = 0 if regsa[i] == 1. If you want it to be save across function calls the code generator has to take care of this. You should order the registers so that the ones that should be used first have the smallest number and it is recommended that registers which are used to pass return values be assigned the lowest numbers. Also You may reserve certain registers to the code generator. This may be reasonable if many ICs cannot be converted without using additional registers. FUNCTIONS The following functions have to be implemented by the code generator: - int init_cg(void); This function is called after the commandline arguments are parsed. It can set up certain internal data, etc. The arrays regarding the data types and the register set, can be set up at this point rather than with a static initialization, however the arrays regarding the commandline options have to be static initialized. The results of the commandline options are available at this point. If something goes wrong, 0 has to be returned, otherwise 1. - void cleanup_cg(FILE *f) This function is called before the compiler exits. f is the output file which _must_ be checked against 0 before using. - int freturn(struct Typ *t); This function has to return the number of the register return values of type t are passed in. If the type is not passed in a register, 0 must be returned. - int regok(int r, int t, int mode); Check, whether the type t can be stored in register r; return 0, if not. If t==POINTER and mode==0 the register only has to be able to store the pointer, but if mode!=0 it has to be able to dereference the pointer. If t==0 return whether the register can be used to store condition codes. This is only relevant if multiple_ccs is set to 1. - int dangerous_IC(struct IC *p) Check if this IC can raise exceptions or is otherwise dangerous. Movement of ICs which are dangerous is restricted to preserve the semantics of the program. Typical dangerous ICs are divisions or pointer dereferencing. On certain targets floating point or even signed integer arithmetic can raise exceptions, too. - int must_convert(np p, int t) Check if type in p does not have to be converted to type t. E.g. on many machines certain types have identical representations (integers of the same size or pointers and integers of the same size). - int shortcut(int code, int t) In C no operations are done with chars and shorts because of integral promotion. However sometimes vbcc might see that an operation could be performed with the short types yielding the same result. Before generating such an instruction with short types vbcc will ask the code generator by calling shortcut() to find out whether it should do so. Return true iff it is a win to perform the operation code with type t rather than promoting the operands and using int or so. - void gen_code(FILE *f, struct IC *p, struct Var *v, zlong offset); This function has to generate the output for a function to stream f. v is a pointer to the function which contains the name of the function. p is a pointer to the first IC, that has to be converted. offset is the space needed for local variables in bytes. This function has to take care that only scratchregisters are destroyed by this function. The array regused contains information about the registers that have been used by vbcc in this function. However if the code generator uses additional registers it has to take care of them, too. The regs[] and regused[] arrays may be overwritten by gen_code() as well as parts of the list of ICs. However the list of ICs must still be a valid list of ICs after gen_code() returned. - void gen_ds(FILE *f, zlong size, struct Typ *t); Has to print output that generates size bytes of type t initialized with proper 0. t is a pointer to a struct Typ which contains the precise type of the variable. On machines where every type can be initialized to 0 by setting all bits to zero the type does not matter. - void gen_align(FILE *f, zlong align); Has to print output that ensures the following data to be aligned to align bytes. - void gen_var_head(FILE *f, struct Var *v); Has to print the head of a static or external variable v. This includes the label and necessary informations for external linkage etc. Typically variables will be generated by a call to gen_align followed by gen_var_head and a series of calls to gen_dc and/or gen_ds. - void gen_dc(FILE *f, int t, struct const_list *p); Well..I'm too lazy at the moment to describe that... Volker Barthelmann volker@vb.franken.de