path: root/code/qcommon/vm_powerpc.c

/*
===========================================================================
Copyright (C) 2008 Przemyslaw Iskra <sparky@pld-linux.org>

This file is part of Quake III Arena source code.

Quake III Arena source code is free software; you can redistribute it
and/or modify it under the terms of the GNU General Public License as
published by the Free Software Foundation; either version 2 of the License,
or (at your option) any later version.

Quake III Arena source code is distributed in the hope that it will be
useful, but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with Quake III Arena source code; if not, write to the Free Software
Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
===========================================================================
*/

#include <sys/types.h> /* needed by sys/mman.h on OSX */
#include <sys/mman.h>
#include <sys/time.h>
#include <time.h>
#include <stddef.h>

#ifndef MAP_ANONYMOUS
# define MAP_ANONYMOUS MAP_ANON
#endif

#include "vm_local.h"
#include "vm_powerpc_asm.h"

/*
 * VM_TIMES enables showing information about time spent inside
 * and outside generated code
 */
//#define VM_TIMES
#ifdef VM_TIMES
#include <sys/times.h>
static clock_t time_outside_vm = 0;
static clock_t time_total_vm = 0;
#endif

/* exit() won't be called but use it because it is marked with noreturn */
#define DIE( reason ) \
	do { \
		Com_Error(ERR_DROP, "vm_powerpc compiler error: " reason "\n"); \
		exit(1); \
	} while(0)

/*
 * vm_powerpc uses large quantities of memory during compilation,
 * Z_Malloc memory may not be enough for some big qvm files
 */

//#define VM_SYSTEM_MALLOC
#ifdef VM_SYSTEM_MALLOC
static inline void *
PPC_Malloc( size_t size )
{
	void *mem = malloc( size );
	if ( ! mem )
		DIE( "Not enough memory" );

	return mem;
}
# define PPC_Free free
#else
# define PPC_Malloc Z_Malloc
# define PPC_Free Z_Free
#endif

/*
 * optimizations:
 * - hole: bubble optimization (OP_CONST+instruction)
 * - copy: inline OP_BLOCK_COPY for lengths under 16/32 bytes
 * - mask: use rlwinm instruction as dataMask
 */

#ifdef __OPTIMIZE__
# define OPTIMIZE_HOLE 1
# define OPTIMIZE_COPY 1
# define OPTIMIZE_MASK 1
#else
# define OPTIMIZE_HOLE 0
# define OPTIMIZE_COPY 0
# define OPTIMIZE_MASK 0
#endif

/*
 * SUPPORTED TARGETS:
 * - Linux 32 bits
 *   ( http://refspecs.freestandards.org/elf/elfspec_ppc.pdf )
 *   * LR at r0 + 4
 *   * Local variable space not needed
 *     -> store caller safe regs at 16+
 *
 * - Linux 64 bits (not fully conformant)
 *   ( http://www.ibm.com/developerworks/linux/library/l-powasm4.html )
 *   * needs "official procedure descriptors" (only first function has one)
 *   * LR at r0 + 16
 *   * local variable space required, min 64 bytes, starts at 48
 *     -> store caller safe regs at 128+
 *
 * - OS X 32 bits
 *   ( http://developer.apple.com/documentation/DeveloperTools/Conceptual/LowLevelABI/Articles/32bitPowerPC.html )
 *   * LR at r0 + 8
 *   * local variable space required, min 32 bytes (?), starts at 24
 *     -> store caller safe regs at 64+
 *
 * - OS X 64 bits (completely untested)
 *   ( http://developer.apple.com/documentation/DeveloperTools/Conceptual/LowLevelABI/Articles/64bitPowerPC.html )
 *   * LR at r0 + 16
 *   * local variable space required, min 64 bytes (?), starts at 48
 *     -> store caller safe regs at 128+
 */

/* Select Length - first value on 32 bits, second on 64 */
#ifdef __PPC64__
#  define SL( a, b ) (b)
#else
#  define SL( a, b ) (a)
#endif

/* Select ABI - first for ELF, second for OS X */
#ifdef __ELF__
#  define SA( a, b ) (a)
#else
#  define SA( a, b ) (b)
#endif

#define ELF32	SL( SA( 1, 0 ), 0 )
#define ELF64	SL( 0, SA( 1, 0 ) )
#define OSX32	SL( SA( 0, 1 ), 0 )
#define OSX64	SL( 0, SA( 0, 1 ) )

/* native length load/store instructions ( L stands for long ) */
#define iSTLU	SL( iSTWU, iSTDU )
#define iSTL	SL( iSTW, iSTD )
#define iLL	SL( iLWZ, iLD )
#define iLLX	SL( iLWZX, iLDX )

/* register length */
#define GPRLEN	SL( 4, 8 )
#define FPRLEN	(8)
/* shift that many bits to obtain value miltiplied by GPRLEN */
#define GPRLEN_SHIFT	SL( 2, 3 )

/* Link register position */
#define STACK_LR	SL( SA( 4, 8 ), 16 )
/* register save position */
#define STACK_SAVE	SL( SA( 16, 64 ), 128 )
/* temporary space, for float<->int exchange */
#define STACK_TEMP	SL( SA( 8, 24 ), 48 )
/* red zone temporary space, used instead of STACK_TEMP if stack isn't
 * prepared properly */
#define STACK_RTEMP	(-16)

#if ELF64
/*
 * Official Procedure Descriptor
 *  we need to prepare one for generated code if we want to call it
 * as function
 */
typedef struct {
	void *function;
	void *toc;
	void *env;
} opd_t;
#endif


/*
 * opcode information table:
 * - length of immediate value
 * - returned register type
 * - required register(s) type
 */
#define opImm0	0x0000 /* no immediate */
#define opImm1	0x0001 /* 1 byte immadiate value after opcode */
#define opImm4	0x0002 /* 4 bytes immediate value after opcode */

#define opRet0	0x0000 /* returns nothing */
#define opRetI	0x0004 /* returns integer */
#define opRetF	0x0008 /* returns float */
#define opRetIF	(opRetI | opRetF) /* returns integer or float */

#define opArg0	0x0000 /* requires nothing */
#define opArgI	0x0010 /* requires integer(s) */
#define opArgF	0x0020 /* requires float(s) */
#define opArgIF	(opArgI | opArgF) /* requires integer or float */

#define opArg2I	0x0040 /* requires second argument, integer */
#define opArg2F	0x0080 /* requires second argument, float */
#define opArg2IF (opArg2I | opArg2F) /* requires second argument, integer or float */

static const unsigned char vm_opInfo[256] =
{
	[OP_UNDEF]	= opImm0,
	[OP_IGNORE]	= opImm0,
	[OP_BREAK]	= opImm0,
	[OP_ENTER]	= opImm4,
			/* OP_LEAVE has to accept floats, they will be converted to ints */
	[OP_LEAVE]	= opImm4 | opRet0 | opArgIF,
			/* only STORE4 and POP use values from OP_CALL,
			 * no need to convert floats back */
	[OP_CALL]	= opImm0 | opRetI | opArgI,
	[OP_PUSH]	= opImm0 | opRetIF,
	[OP_POP]	= opImm0 | opRet0 | opArgIF,
	[OP_CONST]	= opImm4 | opRetIF,
	[OP_LOCAL]	= opImm4 | opRetI,
	[OP_JUMP]	= opImm0 | opRet0 | opArgI,

	[OP_EQ]		= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_NE]		= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_LTI]	= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_LEI]	= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_GTI]	= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_GEI]	= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_LTU]	= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_LEU]	= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_GTU]	= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_GEU]	= opImm4 | opRet0 | opArgI | opArg2I,
	[OP_EQF]	= opImm4 | opRet0 | opArgF | opArg2F,
	[OP_NEF]	= opImm4 | opRet0 | opArgF | opArg2F,
	[OP_LTF]	= opImm4 | opRet0 | opArgF | opArg2F,
	[OP_LEF]	= opImm4 | opRet0 | opArgF | opArg2F,
	[OP_GTF]	= opImm4 | opRet0 | opArgF | opArg2F,
	[OP_GEF]	= opImm4 | opRet0 | opArgF | opArg2F,

	[OP_LOAD1]	= opImm0 | opRetI | opArgI,
	[OP_LOAD2]	= opImm0 | opRetI | opArgI,
	[OP_LOAD4]	= opImm0 | opRetIF| opArgI,
	[OP_STORE1]	= opImm0 | opRet0 | opArgI | opArg2I,
	[OP_STORE2]	= opImm0 | opRet0 | opArgI | opArg2I,
	[OP_STORE4]	= opImm0 | opRet0 | opArgIF| opArg2I,
	[OP_ARG]	= opImm1 | opRet0 | opArgIF,
	[OP_BLOCK_COPY]	= opImm4 | opRet0 | opArgI | opArg2I,

	[OP_SEX8]	= opImm0 | opRetI | opArgI,
	[OP_SEX16]	= opImm0 | opRetI | opArgI,
	[OP_NEGI]	= opImm0 | opRetI | opArgI,
	[OP_ADD]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_SUB]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_DIVI]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_DIVU]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_MODI]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_MODU]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_MULI]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_MULU]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_BAND]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_BOR]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_BXOR]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_BCOM]	= opImm0 | opRetI | opArgI,
	[OP_LSH]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_RSHI]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_RSHU]	= opImm0 | opRetI | opArgI | opArg2I,
	[OP_NEGF]	= opImm0 | opRetF | opArgF,
	[OP_ADDF]	= opImm0 | opRetF | opArgF | opArg2F,
	[OP_SUBF]	= opImm0 | opRetF | opArgF | opArg2F,
	[OP_DIVF]	= opImm0 | opRetF | opArgF | opArg2F,
	[OP_MULF]	= opImm0 | opRetF | opArgF | opArg2F,
	[OP_CVIF]	= opImm0 | opRetF | opArgI,
	[OP_CVFI]	= opImm0 | opRetI | opArgF,
};

/*
 * source instruction data
 */
typedef struct source_instruction_s source_instruction_t;
struct source_instruction_s {
	// opcode
	unsigned long int op;

	// number of instruction
	unsigned long int i_count;

	// immediate value (if any)
	union {
		unsigned int i;
		signed int si;
		signed short ss[2];
		unsigned short us[2];
		unsigned char b;
	} arg;

	// required and returned registers
	unsigned char regA1;
	unsigned char regA2;
	unsigned char regR;
	unsigned char regPos;

	// next instruction
	source_instruction_t *next;
};



/*
 * read-only data needed by the generated code
 */
typedef struct VM_Data {
	// length of this struct + data
	size_t dataLength;
	// compiled code size (in bytes)
	// it only is code size, without the data
	size_t codeLength;

	// function pointers, no use to waste registers for them
	long int (* AsmCall)( int, int );
	void (* BlockCopy )( unsigned int, unsigned int, unsigned int );

	// instruction pointers, rarely used so don't waste register
	ppc_instruction_t *iPointers;

	// data mask for load and store, not used if optimized
	unsigned int dataMask;

	// fixed number used to convert from integer to float
	unsigned int floatBase; // 0x59800004

#if ELF64
	// official procedure descriptor
	opd_t opd;
#endif

	// additional constants, for floating point OP_CONST
	// this data has dynamic length, thus '0' here
	unsigned int data[0];
} vm_data_t;

#ifdef offsetof
# define VM_Data_Offset( field )	offsetof( vm_data_t, field )
#else
# define OFFSET( structName, field ) \
	( (void *)&(((structName *)NULL)->field) - NULL )
# define VM_Data_Offset( field )	OFFSET( vm_data_t, field )
#endif


/*
 * functions used by generated code
 */
static long int
VM_AsmCall( int callSyscallInvNum, int callProgramStack )
{
	vm_t *savedVM = currentVM;
	long int i, ret;
#ifdef VM_TIMES
	struct tms start_time, stop_time;
	clock_t saved_time = time_outside_vm;
	times( &start_time );
#endif

	// save the stack to allow recursive VM entry
	currentVM->programStack = callProgramStack - 4;

	// we need to convert ints to longs on 64bit powerpcs
	if ( sizeof( intptr_t ) == sizeof( int ) ) {
		intptr_t *argPosition = (intptr_t *)((byte *)currentVM->dataBase + callProgramStack + 4);

		// generated code does not invert syscall number
		argPosition[ 0 ] = -1 - callSyscallInvNum;

		ret = currentVM->systemCall( argPosition );
	} else {
		intptr_t args[11];

		// generated code does not invert syscall number
		args[0] = -1 - callSyscallInvNum;

		int *argPosition = (int *)((byte *)currentVM->dataBase + callProgramStack + 4);
		for( i = 1; i < 11; i++ )
			args[ i ] = argPosition[ i ];

		ret = currentVM->systemCall( args );
	}

	currentVM = savedVM;

#ifdef VM_TIMES
	times( &stop_time );
	time_outside_vm = saved_time + ( stop_time.tms_utime - start_time.tms_utime );
#endif

	return ret;
}

static void
VM_BlockCopy( unsigned int dest, unsigned int src, unsigned int count )
{
	unsigned dataMask = currentVM->dataMask;

	if ( (dest & dataMask) != dest
		|| (src & dataMask) != src
		|| ((dest+count) & dataMask) != dest + count
		|| ((src+count) & dataMask) != src + count)
	{
		DIE( "OP_BLOCK_COPY out of range!");
	}

	memcpy( currentVM->dataBase+dest, currentVM->dataBase+src, count );
}


/*
 * code-block descriptors
 */
typedef struct dest_instruction dest_instruction_t;
typedef struct symbolic_jump symbolic_jump_t;

struct symbolic_jump {
	// number of source instruction it has to jump to
	unsigned long int jump_to;

	// jump condition true/false, (4*cr7+(eq|gt..))
	long int bo, bi;

	// extensions / modifiers (branch-link)
	unsigned long ext;

	// dest_instruction refering to this jump
	dest_instruction_t *parent;

	// next jump
	symbolic_jump_t *nextJump;
};

struct dest_instruction {
	// position in the output chain
	unsigned long int count;

	// source instruction number
	unsigned long int i_count;

	// exact (for instructins), or maximum (for jump) length
	unsigned short length;

	dest_instruction_t *next;

	// if the instruction is a jump than jump will be non NULL
	symbolic_jump_t *jump;

	// if jump is NULL than all the instructions will be here
	ppc_instruction_t code[0];
};

// first and last instruction,
// di_first is a dummy instruction
static dest_instruction_t *di_first = NULL, *di_last = NULL;
// number of instructions
static unsigned long int di_count = 0;
// pointers needed to compute local jumps, those aren't pointers to
// actual instructions, just used to check how long the jump is going
// to be and whether it is positive or negative
static dest_instruction_t **di_pointers = NULL;

// output instructions which does not come from source code
// use false i_count value
#define FALSE_ICOUNT 0xffffffff


/*
 * append specified instructions at the end of instruction chain
 */
static void
PPC_Append(
		dest_instruction_t *di_now,
		unsigned long int i_count
  	  )
{
	di_now->count = di_count++;
	di_now->i_count = i_count;
	di_now->next = NULL;

	di_last->next = di_now;
	di_last = di_now;

	if ( i_count != FALSE_ICOUNT ) {
		if ( ! di_pointers[ i_count ] )
			di_pointers[ i_count ] = di_now;
	}
}

/*
 * make space for instructions and append
 */
static void
PPC_AppendInstructions(
		unsigned long int i_count,
		size_t num_instructions,
		const ppc_instruction_t *is
	)
{
	if ( num_instructions < 0 )
		num_instructions = 0;
	size_t iBytes = sizeof( ppc_instruction_t ) * num_instructions;
	dest_instruction_t *di_now = PPC_Malloc( sizeof( dest_instruction_t ) + iBytes );

	di_now->length = num_instructions;
	di_now->jump = NULL;

	if ( iBytes > 0 )
		memcpy( &(di_now->code[0]), is, iBytes );

	PPC_Append( di_now, i_count );
}

/*
 * create symbolic jump and append
 */
static symbolic_jump_t *sj_first = NULL, *sj_last = NULL;
static void
PPC_PrepareJump(
		unsigned long int i_count,
		unsigned long int dest,
		long int bo,
		long int bi,
		unsigned long int ext
	)
{
	dest_instruction_t *di_now = PPC_Malloc( sizeof( dest_instruction_t ) );
	symbolic_jump_t *sj = PPC_Malloc( sizeof( symbolic_jump_t ) );

	sj->jump_to = dest;
	sj->bo = bo;
	sj->bi = bi;
	sj->ext = ext;
	sj->parent = di_now;
	sj->nextJump = NULL;

	sj_last->nextJump = sj;
	sj_last = sj;

	di_now->length = (bo == branchAlways ? 1 : 2);
	di_now->jump = sj;

	PPC_Append( di_now, i_count );
}

/*
 * simplyfy instruction emission
 */
#define emitStart( i_cnt ) \
	unsigned long int i_count = i_cnt; \
	size_t num_instructions = 0; \
	long int force_emit = 0; \
	ppc_instruction_t instructions[50];

#define pushIn( inst ) \
	(instructions[ num_instructions++ ] = inst)
#define in( inst, args... ) pushIn( IN( inst, args ) )

#define emitEnd() \
	do{ \
		if ( num_instructions || force_emit ) \
			PPC_AppendInstructions( i_count, num_instructions, instructions );\
		num_instructions = 0; \
	} while(0)

#define emitJump( dest, bo, bi, ext ) \
	do { \
		emitEnd(); \
		PPC_PrepareJump( i_count, dest, bo, bi, ext ); \
	} while(0)


/*
 * definitions for creating .data section,
 * used in cases where constant float is needed
 */
#define LOCAL_DATA_CHUNK 50
typedef struct local_data_s local_data_t;
struct local_data_s {
	// number of data in this structure
	long int count;

	// data placeholder
	unsigned int data[ LOCAL_DATA_CHUNK ];

	// next chunk, if this one wasn't enough
	local_data_t *next;
};

// first data chunk
static local_data_t *data_first = NULL;
// total number of data
static long int data_acc = 0;

/*
 * append the data and return its offset
 */
static size_t
PPC_PushData( unsigned int datum )
{
	local_data_t *d_now = data_first;
	long int accumulated = 0;

	// check whether we have this one already
	do {
		long int i;
		for ( i = 0; i < d_now->count; i++ ) {
			if ( d_now->data[ i ] == datum ) {
				accumulated += i;
				return VM_Data_Offset( data[ accumulated ] );
			}
		}
		if ( !d_now->next )
			break;

		accumulated += d_now->count;
		d_now = d_now->next;
	} while (1);

	// not found, need to append
	accumulated += d_now->count;

	// last chunk is full, create new one
	if ( d_now->count >= LOCAL_DATA_CHUNK ) {
		d_now->next = PPC_Malloc( sizeof( local_data_t ) );
		d_now = d_now->next;
		d_now->count = 0;
		d_now->next = NULL;
	}

	d_now->data[ d_now->count ] = datum;
	d_now->count += 1;

	data_acc = accumulated + 1;

	return VM_Data_Offset( data[ accumulated ] );
}

/*
 * find leading zeros in dataMask to implement it with
 * "rotate and mask" instruction
 */
static long int fastMaskHi = 0, fastMaskLo = 31;
static void
PPC_MakeFastMask( int mask )
{
#if defined( __GNUC__ ) && ( __GNUC__ >= 4 || ( __GNUC__ == 3 && __GNUC_MINOR__ >= 4 ) )
	/* count leading zeros */
	fastMaskHi = __builtin_clz( mask );

	/* count trailing zeros */
	fastMaskLo = 31 - __builtin_ctz( mask );
#else
	fastMaskHi = 0;
	while ( ( mask & ( 0x80000000 >> fastMaskHi ) ) == 0 )
		fastMaskHi++;

	fastMaskLo = 31;
	while ( ( mask & ( 0x80000000 >> fastMaskLo ) ) == 0 )
		fastMaskLo--;
#endif
}


/*
 * register definitions
 */

/* registers which are global for generated code */

// pointer to VM_Data (constant)
#define rVMDATA r14
// vm->dataBase (constant)
#define rDATABASE r15
// programStack (variable)
#define rPSTACK r16

/*
 * function local registers,
 *
 * normally only volatile registers are used, but if there aren't enough
 * or function has to preserve some value while calling annother one
 * then caller safe registers are used as well
 */
static const long int gpr_list[] = {
	/* caller safe registers, normally only one is used */
	r24, r23, r22, r21,
	r20, r19, r18, r17,
	/* volatile registers (preferred),
	 * normally no more than 5 is used */
	r3, r4, r5, r6,
	r7, r8, r9, r10,
};
static const long int gpr_vstart = 8; /* position of first volatile register */
static const long int gpr_total = sizeof( gpr_list ) / sizeof( gpr_list[0] );

static const long int fpr_list[] = {
	/* static registers, normally none is used */
	f20, f21, f19, f18,
	f17, f16, f15, f14,
	/* volatile registers (preferred),
	 * normally no more than 7 is used */
	f0, f1, f2, f3,
	f4, f5, f6, f7,
	f8, f9, f10, f11,
	f12, f13,
};
static const long int fpr_vstart = 8;
static const long int fpr_total = sizeof( fpr_list ) / sizeof( fpr_list[0] );

/*
 * prepare some dummy structures and emit init code
 */
static void
PPC_CompileInit( void )
{
	di_first = di_last = PPC_Malloc( sizeof( dest_instruction_t ) );
	di_first->count = 0;
	di_first->next = NULL;
	di_first->jump = NULL;

	sj_first = sj_last = PPC_Malloc( sizeof( symbolic_jump_t ) );
	sj_first->nextJump = NULL;

	data_first = PPC_Malloc( sizeof( local_data_t ) );
	data_first->count = 0;
	data_first->next = NULL;

	/*
	 * init function:
	 * saves old values of global registers and sets our values
	 * function prototype is:
	 *  int begin( void *data, int programStack, void *vm->dataBase )
	 */

	/* first instruction must not be placed on instruction list */
	emitStart( FALSE_ICOUNT );

	long int stack = STACK_SAVE + 4 * GPRLEN;

	in( iMFLR, r0 );
	in( iSTLU, r1, -stack, r1 );
	in( iSTL, rVMDATA, STACK_SAVE + 0 * GPRLEN, r1 );
	in( iSTL, rPSTACK, STACK_SAVE + 1 * GPRLEN, r1 );
	in( iSTL, rDATABASE, STACK_SAVE + 2 * GPRLEN, r1 );
	in( iSTL, r0, stack + STACK_LR, r1 );
	in( iMR, rVMDATA, r3 );
	in( iMR, rPSTACK, r4 );
	in( iMR, rDATABASE, r5 );
	in( iBL, +4*8 ); // LINK JUMP: first generated instruction | XXX jump !
	in( iLL, rVMDATA, STACK_SAVE + 0 * GPRLEN, r1 );
	in( iLL, rPSTACK, STACK_SAVE + 1 * GPRLEN, r1 );
	in( iLL, rDATABASE, STACK_SAVE + 2 * GPRLEN, r1 );
	in( iLL, r0, stack + STACK_LR, r1 );
	in( iMTLR, r0 );
	in( iADDI, r1, r1, stack );
	in( iBLR );

	emitEnd();
}

// rFIRST is the copy of the top value on the opstack
#define rFIRST		(gpr_list[ gpr_pos - 1])
// second value on the opstack
#define rSECOND		(gpr_list[ gpr_pos - 2 ])
// temporary registers, not on the opstack
#define rTEMP(x)	(gpr_list[ gpr_pos + x ])
#define rTMP		rTEMP(0)

#define fFIRST		(fpr_list[ fpr_pos - 1 ])
#define fSECOND		(fpr_list[ fpr_pos - 2 ])
#define fTEMP(x)	(fpr_list[ fpr_pos + x ])
#define fTMP		fTEMP(0)

// register types
#define rTYPE_STATIC	0x01
#define rTYPE_FLOAT	0x02

// what type should this opcode return
#define RET_INT		( !(i_now->regR & rTYPE_FLOAT) )
#define RET_FLOAT	( i_now->regR & rTYPE_FLOAT )
// what type should it accept
#define ARG_INT		( ! i_now->regA1 )
#define ARG_FLOAT	( i_now->regA1 )
#define ARG2_INT	( ! i_now->regA2 )
#define ARG2_FLOAT	( i_now->regA2 )

/*
 * emit OP_CONST, called if nothing has used the const value directly
 */
static void
PPC_EmitConst( source_instruction_t * const i_const )
{
	emitStart( i_const->i_count );

	if ( !(i_const->regR & rTYPE_FLOAT) ) {
		// gpr_pos needed for "rFIRST" to work
		long int gpr_pos = i_const->regPos;

		if ( i_const->arg.si >= -0x8000 && i_const->arg.si < 0x8000 ) {
			in( iLI, rFIRST, i_const->arg.si );
		} else if ( i_const->arg.i < 0x10000 ) {
			in( iLI, rFIRST, 0 );
			in( iORI, rFIRST, rFIRST, i_const->arg.i );
		} else {
			in( iLIS, rFIRST, i_const->arg.ss[ 0 ] );
			if ( i_const->arg.us[ 1 ] != 0 )
				in( iORI, rFIRST, rFIRST, i_const->arg.us[ 1 ] );
		}

	} else {
		// fpr_pos needed for "fFIRST" to work
		long int fpr_pos = i_const->regPos;

		// there's no good way to generate the data,
		// just read it from data section
		in( iLFS, fFIRST, PPC_PushData( i_const->arg.i ), rVMDATA );
	}

	emitEnd();
}
#define MAYBE_EMIT_CONST() if ( i_const ) PPC_EmitConst( i_const )

/*
 * emit empty instruction, just sets the needed pointers
 */
static inline void
PPC_EmitNull( source_instruction_t * const i_null )
{
	PPC_AppendInstructions( i_null->i_count, 0, NULL );
}
#define EMIT_FALSE_CONST() PPC_EmitNull( i_const )


/*
 * analize function for register usage and whether it needs stack (r1) prepared
 */
static void
VM_AnalyzeFunction(
		source_instruction_t * const i_first,
		long int *prepareStack,
		long int *gpr_start_pos,
		long int *fpr_start_pos
		)
{
	source_instruction_t *i_now = i_first;

	source_instruction_t *value_provider[20] = { NULL };
	unsigned long int opstack_depth = 0;

	/*
	 * first step:
	 *  remember what codes returned some value and mark the value type
	 *  when we get to know what it should be
	 */
	while ( (i_now = i_now->next) ) {
		unsigned long int op = i_now->op;
		unsigned long int opi = vm_opInfo[ op ];

		if ( opi & opArgIF ) {
			assert( opstack_depth > 0 );

			opstack_depth--;
			source_instruction_t *vp = value_provider[ opstack_depth ];
			unsigned long int vpopi = vm_opInfo[ vp->op ];

			if ( (opi & opArgI) && (vpopi & opRetI) ) {
				// instruction accepts integer, provider returns integer
				//vp->regR |= rTYPE_INT;
				//i_now->regA1 = rTYPE_INT;
			} else if ( (opi & opArgF) && (vpopi & opRetF) ) {
				// instruction accepts float, provider returns float
				vp->regR |= rTYPE_FLOAT; // use OR here - could be marked as static
				i_now->regA1 = rTYPE_FLOAT;
			} else {
				// instruction arg type does not agree with
				// provider return type
				DIE( "unrecognized instruction combination" );
			}

		}
		if ( opi & opArg2IF ) {
			assert( opstack_depth > 0 );

			opstack_depth--;
			source_instruction_t *vp = value_provider[ opstack_depth ];
			unsigned long int vpopi = vm_opInfo[ vp->op ];

			if ( (opi & opArg2I) && (vpopi & opRetI) ) {
				// instruction accepts integer, provider returns integer
				//vp->regR |= rTYPE_INT;
				//i_now->regA2 = rTYPE_INT;
			} else if ( (opi & opArg2F) && (vpopi & opRetF) ) {
				// instruction accepts float, provider returns float
				vp->regR |= rTYPE_FLOAT; // use OR here - could be marked as static
				i_now->regA2 = rTYPE_FLOAT;
			} else {
				// instruction arg type does not agree with
				// provider return type
				DIE( "unrecognized instruction combination" );
			}
		}


		if (
			( op == OP_CALL )
				||
			( op == OP_BLOCK_COPY && ( i_now->arg.i > SL( 16, 32 ) || !OPTIMIZE_COPY ) )
		) {
			long int i;
			*prepareStack = 1;
			// force caller safe registers so we won't have to save them
			for ( i = 0; i < opstack_depth; i++ ) {
				source_instruction_t *vp = value_provider[ i ];
				vp->regR |= rTYPE_STATIC;
			}
		}


		if ( opi & opRetIF ) {
			value_provider[ opstack_depth ] = i_now;
			opstack_depth++;
		}
	}

	/*
	 * second step:
	 *  now that we know register types; compute exactly how many registers
	 *  of each type we need
	 */

	i_now = i_first;
	long int needed_reg[4] = {0,0,0,0}, max_reg[4] = {0,0,0,0};
	opstack_depth = 0;
	while ( (i_now = i_now->next) ) {
		unsigned long int op = i_now->op;
		unsigned long int opi = vm_opInfo[ op ];

		if ( opi & opArgIF ) {
			assert( opstack_depth > 0 );
			opstack_depth--;
			source_instruction_t *vp = value_provider[ opstack_depth ];

			needed_reg[ ( vp->regR & 2 ) ] -= 1;
			if ( vp->regR & 1 ) // static
				needed_reg[ ( vp->regR & 3 ) ] -= 1;
		}
		if ( opi & opArg2IF ) {
			assert( opstack_depth > 0 );
			opstack_depth--;
			source_instruction_t *vp = value_provider[ opstack_depth ];

			needed_reg[ ( vp->regR & 2 ) ] -= 1;
			if ( vp->regR & 1 ) // static
				needed_reg[ ( vp->regR & 3 ) ] -= 1;
		}

		if ( opi & opRetIF ) {
			long int i;
			value_provider[ opstack_depth ] = i_now;
			opstack_depth++;

			i = i_now->regR & 2;
			needed_reg[ i ] += 1;
			if ( max_reg[ i ] < needed_reg[ i ] )
				max_reg[ i ] = needed_reg[ i ];

			i = i_now->regR & 3;
			if ( i & 1 ) {
				needed_reg[ i ] += 1;
				if ( max_reg[ i ] < needed_reg[ i ] )
					max_reg[ i ] = needed_reg[ i ];
			}
		}
	}

	long int gpr_start = gpr_vstart;
	const long int gpr_volatile = gpr_total - gpr_vstart;
	if ( max_reg[ 1 ] > 0 || max_reg[ 0 ] > gpr_volatile ) {
		// max_reg[ 0 ] - all gprs needed
		// max_reg[ 1 ] - static gprs needed
		long int max = max_reg[ 0 ] - gpr_volatile;
		if ( max_reg[ 1 ] > max )
			max = max_reg[ 1 ];
		if ( max > gpr_vstart ) {
			/* error */
			DIE( "Need more GPRs" );
		}

		gpr_start -= max;

		// need stack to save caller safe registers
		*prepareStack = 1;
	}
	*gpr_start_pos = gpr_start;

	long int fpr_start = fpr_vstart;
	const long int fpr_volatile = fpr_total - fpr_vstart;
	if ( max_reg[ 3 ] > 0 || max_reg[ 2 ] > fpr_volatile ) {
		// max_reg[ 2 ] - all fprs needed
		// max_reg[ 3 ] - static fprs needed
		long int max = max_reg[ 2 ] - fpr_volatile;
		if ( max_reg[ 3 ] > max )
			max = max_reg[ 3 ];
		if ( max > fpr_vstart ) {
			/* error */
			DIE( "Need more FPRs" );
		}

		fpr_start -= max;

		// need stack to save caller safe registers
		*prepareStack = 1;
	}
	*fpr_start_pos = fpr_start;
}

/*
 * translate opcodes to ppc instructions,
 * it works on functions, not on whole code at once
 */
static void
VM_CompileFunction( source_instruction_t * const i_first )
{
	long int prepareStack = 0;
	long int gpr_start_pos, fpr_start_pos;

	VM_AnalyzeFunction( i_first, &prepareStack, &gpr_start_pos, &fpr_start_pos );

	long int gpr_pos = gpr_start_pos, fpr_pos = fpr_start_pos;

	// OP_CONST combines well with many opcodes so we treat it in a special way
	source_instruction_t *i_const = NULL;
	source_instruction_t *i_now = i_first;

	// how big the stack has to be
	long int save_space = STACK_SAVE;
	{
		if ( gpr_start_pos < gpr_vstart )
			save_space += (gpr_vstart - gpr_start_pos) * GPRLEN;
		save_space = ( save_space + 15 ) & ~0x0f;

		if ( fpr_start_pos < fpr_vstart )
			save_space += (fpr_vstart - fpr_start_pos) * FPRLEN;
		save_space = ( save_space + 15 ) & ~0x0f;
	}

	long int stack_temp = prepareStack ? STACK_TEMP : STACK_RTEMP;

	while ( (i_now = i_now->next) ) {
		emitStart( i_now->i_count );

		switch ( i_now->op )
		{
			default:
			case OP_UNDEF:
			case OP_IGNORE:
				MAYBE_EMIT_CONST();
				in( iNOP );
				break;

			case OP_BREAK:
				MAYBE_EMIT_CONST();
				// force SEGV
				in( iLWZ, r0, 0, r0 );
				break;

			case OP_ENTER:
				if ( i_const )
					DIE( "Weird opcode order" );

				// don't prepare stack if not needed
				if ( prepareStack ) {
					long int i, save_pos = STACK_SAVE;

					in( iMFLR, r0 );
					in( iSTLU, r1, -save_space, r1 );
					in( iSTL, r0, save_space + STACK_LR, r1 );

					/* save registers */
					for ( i = gpr_start_pos; i < gpr_vstart; i++ ) {
						in( iSTL, gpr_list[ i ], save_pos, r1 );
						save_pos += GPRLEN;
					}
					save_pos = ( save_pos + 15 ) & ~0x0f;

					for ( i = fpr_start_pos; i < fpr_vstart; i++ ) {
						in( iSTFD, fpr_list[ i ], save_pos, r1 );
						save_pos += FPRLEN;
					}
					prepareStack = 2;
				}

				in( iADDI, rPSTACK, rPSTACK, - i_now->arg.si );
				break;

			case OP_LEAVE:
				if ( i_const ) {
					EMIT_FALSE_CONST();

					if ( i_const->regR & rTYPE_FLOAT)
						DIE( "constant float in OP_LEAVE" );

					if ( i_const->arg.si >= -0x8000 && i_const->arg.si < 0x8000 ) {
						in( iLI, r3, i_const->arg.si );
					} else if ( i_const->arg.i < 0x10000 ) {
						in( iLI, r3, 0 );
						in( iORI, r3, r3, i_const->arg.i );
					} else {
						in( iLIS, r3, i_const->arg.ss[ 0 ] );
						if ( i_const->arg.us[ 1 ] != 0 )
							in( iORI, r3, r3, i_const->arg.us[ 1 ] );
					}
					gpr_pos--;
				} else {
					MAYBE_EMIT_CONST();

					/* place return value in r3 */
					if ( ARG_INT ) {
						if ( rFIRST != r3 )
							in( iMR, r3, rFIRST );
						gpr_pos--;
					} else {
						in( iSTFS, fFIRST, stack_temp, r1 );
						in( iLWZ, r3, stack_temp, r1 );
						fpr_pos--;
					}
				}

				// don't undo stack if not prepared
				if ( prepareStack >= 2 ) {
					long int i, save_pos = STACK_SAVE;

					in( iLL, r0, save_space + STACK_LR, r1 );


					/* restore registers */
					for ( i = gpr_start_pos; i < gpr_vstart; i++ ) {
						in( iLL, gpr_list[ i ], save_pos, r1 );
						save_pos += GPRLEN;
					}
					save_pos = ( save_pos + 15 ) & ~0x0f;
					for ( i = fpr_start_pos; i < fpr_vstart; i++ ) {
						in( iLFD, fpr_list[ i ], save_pos, r1 );
						save_pos += FPRLEN;
					}

					in( iMTLR, r0 );
					in( iADDI, r1, r1, save_space );
				}
				in( iADDI, rPSTACK, rPSTACK, i_now->arg.si);
				in( iBLR );
				assert( gpr_pos == gpr_start_pos );
				assert( fpr_pos == fpr_start_pos );
				break;

			case OP_CALL:
				if ( i_const ) {
					EMIT_FALSE_CONST();

					if ( i_const->arg.si >= 0 ) {
						emitJump(
							i_const->arg.i,
							branchAlways, 0, branchExtLink
						);
					} else {
						/* syscall */
						in( iLL, r0, VM_Data_Offset( AsmCall ), rVMDATA );

						in( iLI, r3, i_const->arg.si ); // negative value
						in( iMR, r4, rPSTACK ); // push PSTACK on argument list

						in( iMTCTR, r0 );
						in( iBCTRL );
					}
					if ( rFIRST != r3 )
						in( iMR, rFIRST, r3 );
				} else {
					MAYBE_EMIT_CONST();

					in( iCMPWI, cr7, rFIRST, 0 );
					in( iBLTm, cr7, +4*5 /* syscall */ ); // XXX jump !
					/* instruction call */

					// get instruction address
					in( iLL, r0, VM_Data_Offset( iPointers ), rVMDATA );
					in( iRLWINM, rFIRST, rFIRST, GPRLEN_SHIFT, 0, 31-GPRLEN_SHIFT ); // mul * GPRLEN
					in( iLLX, r0, rFIRST, r0 ); // load pointer

					in( iB, +4*(3 + (rFIRST != r3 ? 1 : 0) ) ); // XXX jump !

					/* syscall */
					in( iLL, r0, VM_Data_Offset( AsmCall ), rVMDATA ); // get asmCall pointer
					/* rFIRST can be r3 or some static register */
					if ( rFIRST != r3 )
						in( iMR, r3, rFIRST ); // push OPSTACK top value on argument list
					in( iMR, r4, rPSTACK ); // push PSTACK on argument list

					/* common code */
					in( iMTCTR, r0 );
					in( iBCTRL );

					if ( rFIRST != r3 )
						in( iMR, rFIRST, r3 ); // push return value on the top of the opstack
				}
				break;

			case OP_PUSH:
				MAYBE_EMIT_CONST();
				if ( RET_INT )
					gpr_pos++;
				else
					fpr_pos++;
				/* no instructions here */
				force_emit = 1;
				break;

			case OP_POP:
				MAYBE_EMIT_CONST();
				if ( ARG_INT )
					gpr_pos--;
				else
					fpr_pos--;
				/* no instructions here */
				force_emit = 1;
				break;

			case OP_CONST:
				MAYBE_EMIT_CONST();
				/* nothing here */
				break;

			case OP_LOCAL:
				MAYBE_EMIT_CONST();
				{
					signed long int hi, lo;
					hi = i_now->arg.ss[ 0 ];
					lo = i_now->arg.ss[ 1 ];
					if ( lo < 0 )
						hi += 1;

					gpr_pos++;
					if ( hi == 0 ) {
						in( iADDI, rFIRST, rPSTACK, lo );
					} else {
						in( iADDIS, rFIRST, rPSTACK, hi );
						if ( lo != 0 )
							in( iADDI, rFIRST, rFIRST, lo );
					}
				}
				break;

			case OP_JUMP:
				if ( i_const ) {
					EMIT_FALSE_CONST();

					emitJump(
						i_const->arg.i,
						branchAlways, 0, 0
					);
				} else {
					MAYBE_EMIT_CONST();

					in( iLL, r0, VM_Data_Offset( iPointers ), rVMDATA );
					in( iRLWINM, rFIRST, rFIRST, GPRLEN_SHIFT, 0, 31-GPRLEN_SHIFT ); // mul * GPRLEN
					in( iLLX, r0, rFIRST, r0 ); // load pointer
					in( iMTCTR, r0 );
					in( iBCTR );
				}
				gpr_pos--;
				break;

			case OP_EQ:
			case OP_NE:
				if ( i_const && i_const->arg.si >= -0x8000 && i_const->arg.si < 0x10000 ) {
					EMIT_FALSE_CONST();
					if ( i_const->arg.si >= 0x8000 )
						in( iCMPLWI, cr7, rSECOND, i_const->arg.i );
					else
						in( iCMPWI, cr7, rSECOND, i_const->arg.si );
				} else {
					MAYBE_EMIT_CONST();
					in( iCMPW, cr7, rSECOND, rFIRST );
				}
				emitJump(
					i_now->arg.i,
					(i_now->op == OP_EQ ? branchTrue : branchFalse),
					4*cr7+eq, 0
				);
				gpr_pos -= 2;
				break;

			case OP_LTI:
			case OP_GEI:
				if ( i_const && i_const->arg.si >= -0x8000 && i_const->arg.si < 0x8000 ) {
					EMIT_FALSE_CONST();
					in( iCMPWI, cr7, rSECOND, i_const->arg.si );
				} else {
					MAYBE_EMIT_CONST();
					in( iCMPW, cr7, rSECOND, rFIRST );
				}
				emitJump(
					i_now->arg.i,
					( i_now->op == OP_LTI ? branchTrue : branchFalse ),
					4*cr7+lt, 0
				);
				gpr_pos -= 2;
				break;

			case OP_GTI:
			case OP_LEI:
				if ( i_const && i_const->arg.si >= -0x8000 && i_const->arg.si < 0x8000 ) {
					EMIT_FALSE_CONST();
					in( iCMPWI, cr7, rSECOND, i_const->arg.si );
				} else {
					MAYBE_EMIT_CONST();
					in( iCMPW, cr7, rSECOND, rFIRST );
				}
				emitJump(
					i_now->arg.i,
					( i_now->op == OP_GTI ? branchTrue : branchFalse ),
					4*cr7+gt, 0
				);
				gpr_pos -= 2;
				break;

			case OP_LTU:
			case OP_GEU:
				if ( i_const && i_const->arg.i < 0x10000 ) {
					EMIT_FALSE_CONST();
					in( iCMPLWI, cr7, rSECOND, i_const->arg.i );
				} else {
					MAYBE_EMIT_CONST();
					in( iCMPLW, cr7, rSECOND, rFIRST );
				}
				emitJump(
					i_now->arg.i,
					( i_now->op == OP_LTU ? branchTrue : branchFalse ),
					4*cr7+lt, 0
				);
				gpr_pos -= 2;
				break;

			case OP_GTU:
			case OP_LEU:
				if ( i_const && i_const->arg.i < 0x10000 ) {
					EMIT_FALSE_CONST();
					in( iCMPLWI, cr7, rSECOND, i_const->arg.i );
				} else {
					MAYBE_EMIT_CONST();
					in( iCMPLW, cr7, rSECOND, rFIRST );
				}
				emitJump(
					i_now->arg.i,
					( i_now->op == OP_GTU ? branchTrue : branchFalse ),
					4*cr7+gt, 0
				);
				gpr_pos -= 2;
				break;

			case OP_EQF:
			case OP_NEF:
				MAYBE_EMIT_CONST();
				in( iFCMPU, cr7, fSECOND, fFIRST );
				emitJump(
					i_now->arg.i,
					( i_now->op == OP_EQF ? branchTrue : branchFalse ),
					4*cr7+eq, 0
				);
				fpr_pos -= 2;
				break;

			case OP_LTF:
			case OP_GEF:
				MAYBE_EMIT_CONST();
				in( iFCMPU, cr7, fSECOND, fFIRST );
				emitJump(
					i_now->arg.i,
					( i_now->op == OP_LTF ? branchTrue : branchFalse ),
					4*cr7+lt, 0
				);
				fpr_pos -= 2;
				break;

			case OP_GTF:
			case OP_LEF:
				MAYBE_EMIT_CONST();
				in( iFCMPU, cr7, fSECOND, fFIRST );
				emitJump(
					i_now->arg.i,
					( i_now->op == OP_GTF ? branchTrue : branchFalse ),
					4*cr7+gt, 0
				);
				fpr_pos -= 2;
				break;

			case OP_LOAD1:
				MAYBE_EMIT_CONST();
#if OPTIMIZE_MASK
				in( iRLWINM, rFIRST, rFIRST, 0, fastMaskHi, fastMaskLo );
#else
				in( iLWZ, r0, VM_Data_Offset( dataMask ), rVMDATA );
				in( iAND, rFIRST, rFIRST, r0 );
#endif
				in( iLBZX, rFIRST, rFIRST, rDATABASE );
				break;

			case OP_LOAD2:
				MAYBE_EMIT_CONST();
#if OPTIMIZE_MASK
				in( iRLWINM, rFIRST, rFIRST, 0, fastMaskHi, fastMaskLo );
#else
				in( iLWZ, r0, VM_Data_Offset( dataMask ), rVMDATA );
				in( iAND, rFIRST, rFIRST, r0 );
#endif
				in( iLHZX, rFIRST, rFIRST, rDATABASE );
				break;

			case OP_LOAD4:
				MAYBE_EMIT_CONST();
#if OPTIMIZE_MASK
				in( iRLWINM, rFIRST, rFIRST, 0, fastMaskHi, fastMaskLo );
#else
				in( iLWZ, r0, VM_Data_Offset( dataMask ), rVMDATA );
				in( iAND, rFIRST, rFIRST, r0 );
#endif
				if ( RET_INT ) {
					in( iLWZX, rFIRST, rFIRST, rDATABASE );
				} else {
					fpr_pos++;
					in( iLFSX, fFIRST, rFIRST, rDATABASE );
					gpr_pos--;
				}
				break;

			case OP_STORE1:
				MAYBE_EMIT_CONST();
#if OPTIMIZE_MASK
				in( iRLWINM, rSECOND, rSECOND, 0, fastMaskHi, fastMaskLo );
#else
				in( iLWZ, r0, VM_Data_Offset( dataMask ), rVMDATA );
				in( iAND, rSECOND, rSECOND, r0 );
#endif
				in( iSTBX, rFIRST, rSECOND, rDATABASE );
				gpr_pos -= 2;
				break;

			case OP_STORE2:
				MAYBE_EMIT_CONST();
#if OPTIMIZE_MASK
				in( iRLWINM, rSECOND, rSECOND, 0, fastMaskHi, fastMaskLo );
#else
				in( iLWZ, r0, VM_Data_Offset( dataMask ), rVMDATA );
				in( iAND, rSECOND, rSECOND, r0 );
#endif
				in( iSTHX, rFIRST, rSECOND, rDATABASE );
				gpr_pos -= 2;
				break;

			case OP_STORE4:
				MAYBE_EMIT_CONST();
				if ( ARG_INT ) {
#if OPTIMIZE_MASK
					in( iRLWINM, rSECOND, rSECOND, 0, fastMaskHi, fastMaskLo );
#else
					in( iLWZ, r0, VM_Data_Offset( dataMask ), rVMDATA );
					in( iAND, rSECOND, rSECOND, r0 );
#endif

					in( iSTWX, rFIRST, rSECOND, rDATABASE );
					gpr_pos--;
				} else {
#if OPTIMIZE_MASK
					in( iRLWINM, rFIRST, rFIRST, 0, fastMaskHi, fastMaskLo );
#else
					in( iLWZ, r0, VM_Data_Offset( dataMask ), rVMDATA );
					in( iAND, rFIRST, rFIRST, r0 );
#endif

					in( iSTFSX, fFIRST, rFIRST, rDATABASE );
					fpr_pos--;
				}
				gpr_pos--;
				break;

			case OP_ARG:
				MAYBE_EMIT_CONST();
				in( iADDI, r0, rPSTACK, i_now->arg.b );
				if ( ARG_INT ) {
					in( iSTWX, rFIRST, rDATABASE, r0 );
					gpr_pos--;
				} else {
					in( iSTFSX, fFIRST, rDATABASE, r0 );
					fpr_pos--;
				}
				break;

			case OP_BLOCK_COPY:
				MAYBE_EMIT_CONST();
#if OPTIMIZE_COPY
				if ( i_now->arg.i <= SL( 16, 32 ) ) {
					/* block is very short so copy it in-place */

					unsigned int len = i_now->arg.i;
					unsigned int copied = 0, left = len;

					in( iADD, rFIRST, rFIRST, rDATABASE );
					in( iADD, rSECOND, rSECOND, rDATABASE );

					if ( len >= GPRLEN ) {
						long int i, words = len / GPRLEN;
						in( iLL, r0, 0, rFIRST );
						for ( i = 1; i < words; i++ )
							in( iLL, rTEMP( i - 1 ), GPRLEN * i, rFIRST );

						in( iSTL, r0, 0, rSECOND );
						for ( i = 1; i < words; i++ )
							in( iSTL, rTEMP( i - 1 ), GPRLEN * i, rSECOND );

						copied += words * GPRLEN;
						left -= copied;
					}

					if ( SL( 0, left >= 4 ) ) {
						in( iLWZ, r0, copied+0, rFIRST );
						in( iSTW, r0, copied+0, rSECOND );
						copied += 4;
						left -= 4;
					}
					if ( left >= 4 ) {
						DIE("Bug in OP_BLOCK_COPY");
					}
					if ( left == 3 ) {
						in( iLHZ, r0,	copied+0, rFIRST );
						in( iLBZ, rTMP,	copied+2, rFIRST );
						in( iSTH, r0,	copied+0, rSECOND );
						in( iSTB, rTMP,	copied+2, rSECOND );
					} else if ( left == 2 ) {
						in( iLHZ, r0, copied+0, rFIRST );
						in( iSTH, r0, copied+0, rSECOND );
					} else if ( left == 1 ) {
						in( iLBZ, r0, copied+0, rFIRST );
						in( iSTB, r0, copied+0, rSECOND );
					}
				} else
#endif
				{
					unsigned long int r5_ori = 0;
					if ( i_now->arg.si >= -0x8000 && i_now->arg.si < 0x8000 ) {
						in( iLI, r5, i_now->arg.si );
					} else if ( i_now->arg.i < 0x10000 ) {
						in( iLI, r5, 0 );
						r5_ori = i_now->arg.i;
					} else {
						in( iLIS, r5, i_now->arg.ss[ 0 ] );
						r5_ori = i_now->arg.us[ 1 ];
					}

					in( iLL, r0, VM_Data_Offset( BlockCopy ), rVMDATA ); // get blockCopy pointer

					if ( r5_ori )
						in( iORI, r5, r5, r5_ori );

					in( iMTCTR, r0 );

					if ( rFIRST != r4 )
						in( iMR, r4, rFIRST );
					if ( rSECOND != r3 )
						in( iMR, r3, rSECOND );

					in( iBCTRL );
				}

				gpr_pos -= 2;
				break;

			case OP_SEX8:
				MAYBE_EMIT_CONST();
				in( iEXTSB, rFIRST, rFIRST );
				break;

			case OP_SEX16:
				MAYBE_EMIT_CONST();
				in( iEXTSH, rFIRST, rFIRST );
				break;

			case OP_NEGI:
				MAYBE_EMIT_CONST();
				in( iNEG, rFIRST, rFIRST );
				break;

			case OP_ADD:
				if ( i_const ) {
					EMIT_FALSE_CONST();

					signed short int hi, lo;
					hi = i_const->arg.ss[ 0 ];
					lo = i_const->arg.ss[ 1 ];
					if ( lo < 0 )
						hi += 1;

					if ( hi != 0 )
						in( iADDIS, rSECOND, rSECOND, hi );
					if ( lo != 0 )
						in( iADDI, rSECOND, rSECOND, lo );

					// if both are zero no instruction will be written
					if ( hi == 0 && lo == 0 )
						force_emit = 1;
				} else {
					MAYBE_EMIT_CONST();
					in( iADD, rSECOND, rSECOND, rFIRST );
				}
				gpr_pos--;
				break;

			case OP_SUB:
				MAYBE_EMIT_CONST();
				in( iSUB, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_DIVI:
				MAYBE_EMIT_CONST();
				in( iDIVW, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_DIVU:
				MAYBE_EMIT_CONST();
				in( iDIVWU, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_MODI:
				MAYBE_EMIT_CONST();
				in( iDIVW, r0, rSECOND, rFIRST );
				in( iMULLW, r0, r0, rFIRST );
				in( iSUB, rSECOND, rSECOND, r0 );
				gpr_pos--;
				break;

			case OP_MODU:
				MAYBE_EMIT_CONST();
				in( iDIVWU, r0, rSECOND, rFIRST );
				in( iMULLW, r0, r0, rFIRST );
				in( iSUB, rSECOND, rSECOND, r0 );
				gpr_pos--;
				break;

			case OP_MULI:
			case OP_MULU:
				MAYBE_EMIT_CONST();
				in( iMULLW, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_BAND:
				MAYBE_EMIT_CONST();
				in( iAND, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_BOR:
				MAYBE_EMIT_CONST();
				in( iOR, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_BXOR:
				MAYBE_EMIT_CONST();
				in( iXOR, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_BCOM:
				MAYBE_EMIT_CONST();
				in( iNOT, rFIRST, rFIRST );
				break;

			case OP_LSH:
				MAYBE_EMIT_CONST();
				in( iSLW, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_RSHI:
				MAYBE_EMIT_CONST();
				in( iSRAW, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_RSHU:
				MAYBE_EMIT_CONST();
				in( iSRW, rSECOND, rSECOND, rFIRST );
				gpr_pos--;
				break;

			case OP_NEGF:
				MAYBE_EMIT_CONST();
				in( iFNEG, fFIRST, fFIRST );
				break;

			case OP_ADDF:
				MAYBE_EMIT_CONST();
				in( iFADDS, fSECOND, fSECOND, fFIRST );
				fpr_pos--;
				break;

			case OP_SUBF:
				MAYBE_EMIT_CONST();
				in( iFSUBS, fSECOND, fSECOND, fFIRST );
				fpr_pos--;
				break;

			case OP_DIVF:
				MAYBE_EMIT_CONST();
				in( iFDIVS, fSECOND, fSECOND, fFIRST );
				fpr_pos--;
				break;

			case OP_MULF:
				MAYBE_EMIT_CONST();
				in( iFMULS, fSECOND, fSECOND, fFIRST );
				fpr_pos--;
				break;

			case OP_CVIF:
				MAYBE_EMIT_CONST();
				fpr_pos++;
				in( iXORIS, rFIRST, rFIRST, 0x8000 );
				in( iLIS, r0, 0x4330 );
				in( iSTW, rFIRST, stack_temp + 4, r1 );
				in( iSTW, r0, stack_temp, r1 );
				in( iLFS, fTMP, VM_Data_Offset( floatBase ), rVMDATA );
				in( iLFD, fFIRST, stack_temp, r1 );
				in( iFSUB, fFIRST, fFIRST, fTMP );
				in( iFRSP, fFIRST, fFIRST );
				gpr_pos--;
				break;

			case OP_CVFI:
				MAYBE_EMIT_CONST();
				gpr_pos++;
				in( iFCTIWZ, fFIRST, fFIRST );
				in( iSTFD, fFIRST, stack_temp, r1 );
				in( iLWZ, rFIRST, stack_temp + 4, r1 );
				fpr_pos--;
				break;
		}

		i_const = NULL;

		if ( i_now->op != OP_CONST ) {
			// emit the instructions if it isn't OP_CONST
			emitEnd();
		} else {
			// mark in what register the value should be saved
			if ( RET_INT )
				i_now->regPos = ++gpr_pos;
			else
				i_now->regPos = ++fpr_pos;

#if OPTIMIZE_HOLE
			i_const = i_now;
#else
			PPC_EmitConst( i_now );
#endif
		}
	}
	if ( i_const )
		DIE( "left (unused) OP_CONST" );

	{
		// free opcode information, don't free first dummy one
		source_instruction_t *i_next = i_first->next;
		while ( i_next ) {
			i_now = i_next;
			i_next = i_now->next;
			PPC_Free( i_now );
		}
	}
}


/*
 * check which jumps are short enough to use signed 16bit immediate branch
 */
static void
PPC_ShrinkJumps( void )
{
	symbolic_jump_t *sj_now = sj_first;
	while ( (sj_now = sj_now->nextJump) ) {
		if ( sj_now->bo == branchAlways )
			// non-conditional branch has 26bit immediate
			sj_now->parent->length = 1;

		else {
			dest_instruction_t *di = di_pointers[ sj_now->jump_to ];
			dest_instruction_t *ji = sj_now->parent;
			long int jump_length = 0;
			if ( ! di )
				DIE( "No instruction to jump to" );
			if ( ji->count > di->count ) {
				do {
					jump_length += di->length;
				} while ( ( di = di->next ) != ji );
			} else {
				jump_length = 1;
				while ( ( ji = ji->next ) != di )
					jump_length += ji->length;
			}
			if ( jump_length < 0x2000 )
				// jump is short, use normal instruction
				sj_now->parent->length = 1;
		}
	}
}

/*
 * puts all the data in one place, it consists of many different tasks
 */
static void
PPC_ComputeCode( vm_t *vm )
{
	dest_instruction_t *di_now = di_first;

	unsigned long int codeInstructions = 0;
	// count total instruciton number
	while ( (di_now = di_now->next ) )
		codeInstructions += di_now->length;

	size_t codeLength = sizeof( vm_data_t )
		+ sizeof( unsigned int ) * data_acc
		+ sizeof( ppc_instruction_t ) * codeInstructions;

	// get the memory for the generated code, smarter ppcs need the
	// mem to be marked as executable (whill change later)
	unsigned char *dataAndCode = mmap( NULL, codeLength,
		PROT_READ|PROT_WRITE, MAP_SHARED|MAP_ANONYMOUS, -1, 0 );

	if ( ! dataAndCode )
		DIE( "Not enough memory" );

	ppc_instruction_t *codeNow, *codeBegin;
	codeNow = codeBegin = (ppc_instruction_t *)( dataAndCode + VM_Data_Offset( data[ data_acc ] ) );

	ppc_instruction_t nop = IN( iNOP );

	// copy instructions to the destination
	// fills the jump instructions with nops
	// saves pointers of all instructions
	di_now = di_first;
	while ( (di_now = di_now->next ) ) {
		unsigned long int i_count = di_now->i_count;
		if ( i_count != FALSE_ICOUNT ) {
			if ( ! di_pointers[ i_count ] )
				di_pointers[ i_count ] = (void *) codeNow;
		}

		if ( di_now->jump == NULL ) {
			memcpy( codeNow, &(di_now->code[0]), di_now->length * sizeof( ppc_instruction_t ) );
			codeNow += di_now->length;
		} else {
			long int i;
			symbolic_jump_t *sj;
			for ( i = 0; i < di_now->length; i++ )
				codeNow[ i ] = nop;
			codeNow += di_now->length;

			sj = di_now->jump;
			// save position of jumping instruction
			sj->parent = (void *)(codeNow - 1);
		}
	}

	// compute the jumps and write corresponding instructions
	symbolic_jump_t *sj_now = sj_first;
	while ( (sj_now = sj_now->nextJump ) ) {
		ppc_instruction_t *jumpFrom = (void *) sj_now->parent;
		ppc_instruction_t *jumpTo = (void *) di_pointers[ sj_now->jump_to ];
		signed long int jumpLength = jumpTo - jumpFrom;

		// if jump is short, just write it
		if ( jumpLength >= - 8192 && jumpLength < 8192 ) {
			powerpc_iname_t branchConditional = sj_now->ext & branchExtLink ? iBCL : iBC;
			*jumpFrom = IN( branchConditional, sj_now->bo, sj_now->bi, jumpLength * 4 );
			continue;
		}

		// jump isn't short so write it as two instructions
		//
		// the letter one is a non-conditional branch instruction which
		// accepts immediate values big enough (26 bits)
		*jumpFrom = IN( (sj_now->ext & branchExtLink ? iBL : iB), jumpLength * 4 );
		if ( sj_now->bo == branchAlways )
			continue;

		// there should have been additional space prepared for this case
		if ( jumpFrom[ -1 ] != nop )
			DIE( "additional space for long jump not prepared" );

		// invert instruction condition
		long int bo = 0;
		switch ( sj_now->bo ) {
			case branchTrue:
				bo = branchFalse;
				break;
			case branchFalse:
				bo = branchTrue;
				break;
			default:
				DIE( "unrecognized branch type" );
				break;
		}

		// the former instruction is an inverted conditional branch which
		// jumps over the non-conditional one
		jumpFrom[ -1 ] = IN( iBC, bo, sj_now->bi, +2*4 );
	}

	vm->codeBase = dataAndCode;
	vm->codeLength = codeLength;

	vm_data_t *data = (vm_data_t *)dataAndCode;

#if ELF64
	// prepare Official Procedure Descriptor for the generated code
	// and retrieve real function pointer for helper functions

	opd_t *ac = (void *)VM_AsmCall, *bc = (void *)VM_BlockCopy;
	data->opd.function = codeBegin;
	// trick it into using the same TOC
	// this way we won't have to switch TOC before calling AsmCall or BlockCopy
	data->opd.toc = ac->toc;
	data->opd.env = ac->env;

	data->AsmCall = ac->function;
	data->BlockCopy = bc->function;
#else
	data->AsmCall = VM_AsmCall;
	data->BlockCopy = VM_BlockCopy;
#endif

	data->dataMask = vm->dataMask;
	data->iPointers = (ppc_instruction_t *)vm->instructionPointers;
	data->dataLength = VM_Data_Offset( data[ data_acc ] );
	data->codeLength = ( codeNow - codeBegin ) * sizeof( ppc_instruction_t );
	data->floatBase = 0x59800004;


	/* write dynamic data (float constants) */
	{
		local_data_t *d_next, *d_now = data_first;
		long int accumulated = 0;

		do {
			long int i;
			for ( i = 0; i < d_now->count; i++ )
				data->data[ accumulated + i ] = d_now->data[ i ];

			accumulated += d_now->count;
			d_next = d_now->next;
			PPC_Free( d_now );

			if ( !d_next )
				break;
			d_now = d_next;
		} while (1);
		data_first = NULL;
	}

	/* free most of the compilation memory */
	{
		di_now = di_first->next;
		PPC_Free( di_first );
		PPC_Free( sj_first );

		while ( di_now ) {
			di_first = di_now->next;
			if ( di_now->jump )
				PPC_Free( di_now->jump );
			PPC_Free( di_now );
			di_now = di_first;
		}
	}

	return;
}

static void
VM_Destroy_Compiled( vm_t *self )
{
	if ( self->codeBase ) {
		if ( munmap( self->codeBase, self->codeLength ) )
			Com_Printf( S_COLOR_RED "Memory unmap failed, possible memory leak\n" );
	}
	self->codeBase = NULL;
}

void
VM_Compile( vm_t *vm, vmHeader_t *header )
{
	long int pc = 0;
	unsigned long int i_count;
	char* code;
	struct timeval tvstart = {0, 0};
	source_instruction_t *i_first /* dummy */, *i_last = NULL, *i_now;

	vm->compiled = qfalse;

	gettimeofday(&tvstart, NULL);

	PPC_MakeFastMask( vm->dataMask );

	i_first = PPC_Malloc( sizeof( source_instruction_t ) );
	i_first->next = NULL;

	// realloc instructionPointers with correct size
	// use Z_Malloc so vm.c will be able to free the memory
	if ( sizeof( void * ) != sizeof( int ) ) {
		Z_Free( vm->instructionPointers );
		vm->instructionPointers = Z_Malloc( header->instructionCount * sizeof( void * ) );
	}
	di_pointers = (void *)vm->instructionPointers;
	memset( di_pointers, 0, header->instructionCount * sizeof( void * ) );


	PPC_CompileInit();

	/*
	 * read the input program
	 * divide it into functions and send each function to compiler
	 */
	code = (char *)header + header->codeOffset;
	for ( i_count = 0; i_count < header->instructionCount; ++i_count )
	{
		unsigned char op = code[ pc++ ];

		if ( op == OP_ENTER ) {
			if ( i_first->next )
				VM_CompileFunction( i_first );
			i_first->next = NULL;
			i_last = i_first;
		}

		i_now = PPC_Malloc( sizeof( source_instruction_t ) );
		i_now->op = op;
		i_now->i_count = i_count;
		i_now->arg.i = 0;
		i_now->regA1 = 0;
		i_now->regA2 = 0;
		i_now->regR = 0;
		i_now->regPos = 0;
		i_now->next = NULL;

		if ( vm_opInfo[op] & opImm4 ) {
			union {
				unsigned char b[4];
				unsigned int i;
			} c = { { code[ pc + 3 ], code[ pc + 2 ], code[ pc + 1 ], code[ pc + 0 ] }, };

			i_now->arg.i = c.i;
			pc += 4;
		} else if ( vm_opInfo[op] & opImm1 ) {
			i_now->arg.b = code[ pc++ ];
		}

		i_last->next = i_now;
		i_last = i_now;
	}
	VM_CompileFunction( i_first );
	PPC_Free( i_first );

	PPC_ShrinkJumps();
	memset( di_pointers, 0, header->instructionCount * sizeof( void * ) );
	PPC_ComputeCode( vm );

	/* check for uninitialized pointers */
#ifdef DEBUG_VM
	long int i;
	for ( i = 0; i < header->instructionCount; i++ )
		if ( di_pointers[ i ] == 0 )
			Com_Printf( S_COLOR_RED "Pointer %ld not initialized !\n", i );
#endif

	/* mark memory as executable and not writeable */
	if ( mprotect( vm->codeBase, vm->codeLength, PROT_READ|PROT_EXEC ) ) {

		// it has failed, make sure memory is unmapped before throwing the error
		VM_Destroy_Compiled( vm );
		DIE( "mprotect failed" );
	}

	vm->destroy = VM_Destroy_Compiled;
	vm->compiled = qtrue;

	{
		struct timeval tvdone = {0, 0};
		struct timeval dur = {0, 0};
		Com_Printf( "VM file %s compiled to %i bytes of code (%p - %p)\n",
			vm->name, vm->codeLength, vm->codeBase, vm->codeBase+vm->codeLength );

		gettimeofday(&tvdone, NULL);
		timersub(&tvdone, &tvstart, &dur);
		Com_Printf( "compilation took %lu.%06lu seconds\n",
			(long unsigned int)dur.tv_sec, (long unsigned int)dur.tv_usec );
	}
}

int
VM_CallCompiled( vm_t *vm, int *args )
{
	int retVal;
	int *argPointer;

	vm_data_t *vm_dataAndCode = (void *)( vm->codeBase );
	int programStack = vm->programStack;
	int stackOnEntry = programStack;

	byte *image = vm->dataBase;

	currentVM = vm;

	vm->currentlyInterpreting = qtrue;

	programStack -= 48;
	argPointer = (int *)&image[ programStack + 8 ];
	memcpy( argPointer, args, 4 * 9 );
	argPointer[ -1 ] = 0;
	argPointer[ -2 ] = -1;

#ifdef VM_TIMES
	struct tms start_time, stop_time;
	clock_t time_diff;

	times( &start_time );
	time_outside_vm = 0;
#endif

	/* call generated code */
	{
		int ( *entry )( void *, int, void * );
#ifdef __PPC64__
		entry = (void *)&(vm_dataAndCode->opd);
#else
		entry = (void *)(vm->codeBase + vm_dataAndCode->dataLength);
#endif
		retVal = entry( vm->codeBase, programStack, vm->dataBase );
	}

#ifdef VM_TIMES
	times( &stop_time );
	time_diff = stop_time.tms_utime - start_time.tms_utime;
	time_total_vm += time_diff - time_outside_vm;
	if ( time_diff > 100 ) {
		printf( "App clock: %ld, vm total: %ld, vm this: %ld, vm real: %ld, vm out: %ld\n"
			"Inside VM %f%% of app time\n",
			stop_time.tms_utime,
			time_total_vm,
			time_diff,
			time_diff - time_outside_vm,
			time_outside_vm,
			(double)100 * time_total_vm / stop_time.tms_utime );
	}
#endif

	vm->programStack = stackOnEntry;
	vm->currentlyInterpreting = qfalse;

	return retVal;
}