/**

 * $Id: sh4core.in,v 1.5 2007-09-12 09:20:38 nkeynes Exp $

 * 

 * SH4 emulation core, and parent module for all the SH4 peripheral

 * modules.

 *

 * Copyright (c) 2005 Nathan Keynes.

 *

 * This program 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.

 *

 * This program 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.

 */

#define MODULE sh4_module

#include <math.h>

#include "dream.h"

#include "sh4/sh4core.h"

#include "sh4/sh4mmio.h"

#include "sh4/intc.h"

#include "mem.h"

#include "clock.h"

#include "syscall.h"

#define SH4_CALLTRACE 1

#define MAX_INT 0x7FFFFFFF

#define MIN_INT 0x80000000

#define MAX_INTF 2147483647.0

#define MIN_INTF -2147483648.0

#define EXV_EXCEPTION    0x100  /* General exception vector */

#define EXV_TLBMISS      0x400  /* TLB-miss exception vector */

#define EXV_INTERRUPT    0x600  /* External interrupt vector */

/********************** SH4 Module Definition ****************************/

uint32_t sh4_run_slice( uint32_t );

static uint16_t *sh4_icache = NULL;

static uint32_t sh4_icache_addr = 0;

uint32_t sh4_run_slice( uint32_t nanosecs ) 

{

    int i;

    sh4r.slice_cycle = 0;

    if( sh4r.sh4_state != SH4_STATE_RUNNING ) {

	if( sh4r.event_pending < nanosecs ) {

	    sh4r.sh4_state = SH4_STATE_RUNNING;

	    sh4r.slice_cycle = sh4r.event_pending;

	}

    }

    if( sh4_breakpoint_count == 0 ) {

	for( ; sh4r.slice_cycle < nanosecs; sh4r.slice_cycle += sh4_cpu_period ) {

	    if( SH4_EVENT_PENDING() ) {

		if( sh4r.event_types & PENDING_EVENT ) {

		    event_execute();

		}

		/* Eventq execute may (quite likely) deliver an immediate IRQ */

		if( sh4r.event_types & PENDING_IRQ ) {

		    sh4_accept_interrupt();

		}

	    }

	    if( !sh4_execute_instruction() ) {

		break;

	    }

	}

    } else {

	for( ;sh4r.slice_cycle < nanosecs; sh4r.slice_cycle += sh4_cpu_period ) {

	    if( SH4_EVENT_PENDING() ) {

		if( sh4r.event_types & PENDING_EVENT ) {

		    event_execute();

		}

		/* Eventq execute may (quite likely) deliver an immediate IRQ */

		if( sh4r.event_types & PENDING_IRQ ) {

		    sh4_accept_interrupt();

		}

	    }

	    if( !sh4_execute_instruction() )

		break;

#ifdef ENABLE_DEBUG_MODE

	    for( i=0; i<sh4_breakpoint_count; i++ ) {

		if( sh4_breakpoints[i].address == sh4r.pc ) {

		    break;

		}

	    }

	    if( i != sh4_breakpoint_count ) {

		dreamcast_stop();

		if( sh4_breakpoints[i].type == BREAK_ONESHOT )

		    sh4_clear_breakpoint( sh4r.pc, BREAK_ONESHOT );

		break;

	    }

#endif	

	}

    }

    /* If we aborted early, but the cpu is still technically running,

     * we're doing a hard abort - cut the timeslice back to what we

     * actually executed

     */

    if( sh4r.slice_cycle != nanosecs && sh4r.sh4_state == SH4_STATE_RUNNING ) {

	nanosecs = sh4r.slice_cycle;

    }

    if( sh4r.sh4_state != SH4_STATE_STANDBY ) {

	TMU_run_slice( nanosecs );

	SCIF_run_slice( nanosecs );

    }

    return nanosecs;

}

/********************** SH4 emulation core  ****************************/

void sh4_set_pc( int pc )

{

    sh4r.pc = pc;

    sh4r.new_pc = pc+2;

}

#define UNDEF(ir) return sh4_raise_slot_exception(EXC_ILLEGAL, EXC_SLOT_ILLEGAL)

#define UNIMP(ir) do{ ERROR( "Halted on unimplemented instruction at %08x, opcode = %04x", sh4r.pc, ir ); dreamcast_stop(); return FALSE; }while(0)

#if(SH4_CALLTRACE == 1)

#define MAX_CALLSTACK 32

static struct call_stack {

    sh4addr_t call_addr;

    sh4addr_t target_addr;

    sh4addr_t stack_pointer;

} call_stack[MAX_CALLSTACK];

static int call_stack_depth = 0;

int sh4_call_trace_on = 0;

static inline trace_call( sh4addr_t source, sh4addr_t dest ) 

{

    if( call_stack_depth < MAX_CALLSTACK ) {

	call_stack[call_stack_depth].call_addr = source;

	call_stack[call_stack_depth].target_addr = dest;

	call_stack[call_stack_depth].stack_pointer = sh4r.r[15];

    }

    call_stack_depth++;

}

static inline trace_return( sh4addr_t source, sh4addr_t dest )

{

    if( call_stack_depth > 0 ) {

	call_stack_depth--;

    }

}

void fprint_stack_trace( FILE *f )

{

    int i = call_stack_depth -1;

    if( i >= MAX_CALLSTACK )

	i = MAX_CALLSTACK - 1;

    for( ; i >= 0; i-- ) {

	fprintf( f, "%d. Call from %08X => %08X, SP=%08X\n", 

		 (call_stack_depth - i), call_stack[i].call_addr,

		 call_stack[i].target_addr, call_stack[i].stack_pointer );

    }

}

#define TRACE_CALL( source, dest ) trace_call(source, dest)

#define TRACE_RETURN( source, dest ) trace_return(source, dest)

#else

#define TRACE_CALL( dest, rts ) 

#define TRACE_RETURN( source, dest )

#endif

#define RAISE( x, v ) do{			\

    if( sh4r.vbr == 0 ) { \

        ERROR( "%08X: VBR not initialized while raising exception %03X, halting", sh4r.pc, x ); \

        dreamcast_stop(); return FALSE;	\

    } else { \

        sh4r.spc = sh4r.pc;	\

        sh4r.ssr = sh4_read_sr(); \

        sh4r.sgr = sh4r.r[15]; \

        MMIO_WRITE(MMU,EXPEVT,x); \

        sh4r.pc = sh4r.vbr + v; \

        sh4r.new_pc = sh4r.pc + 2; \

        sh4_write_sr( sh4r.ssr |SR_MD|SR_BL|SR_RB ); \

	if( sh4r.in_delay_slot ) { \

	    sh4r.in_delay_slot = 0; \

	    sh4r.spc -= 2; \

	} \

    } \

    return TRUE; } while(0)

#define MEM_READ_BYTE( addr ) sh4_read_byte(addr)

#define MEM_READ_WORD( addr ) sh4_read_word(addr)

#define MEM_READ_LONG( addr ) sh4_read_long(addr)

#define MEM_WRITE_BYTE( addr, val ) sh4_write_byte(addr, val)

#define MEM_WRITE_WORD( addr, val ) sh4_write_word(addr, val)

#define MEM_WRITE_LONG( addr, val ) sh4_write_long(addr, val)

#define FP_WIDTH (IS_FPU_DOUBLESIZE() ? 8 : 4)

#define MEM_FP_READ( addr, reg ) sh4_read_float( addr, reg );

#define MEM_FP_WRITE( addr, reg ) sh4_write_float( addr, reg );

#define CHECKPRIV() if( !IS_SH4_PRIVMODE() ) return sh4_raise_slot_exception( EXC_ILLEGAL, EXC_SLOT_ILLEGAL )

#define CHECKRALIGN16(addr) if( (addr)&0x01 ) return sh4_raise_exception( EXC_DATA_ADDR_READ )

#define CHECKRALIGN32(addr) if( (addr)&0x03 ) return sh4_raise_exception( EXC_DATA_ADDR_READ )

#define CHECKWALIGN16(addr) if( (addr)&0x01 ) return sh4_raise_exception( EXC_DATA_ADDR_WRITE )

#define CHECKWALIGN32(addr) if( (addr)&0x03 ) return sh4_raise_exception( EXC_DATA_ADDR_WRITE )

#define CHECKFPUEN() if( !IS_FPU_ENABLED() ) { if( ir == 0xFFFD ) { UNDEF(ir); } else { return sh4_raise_slot_exception( EXC_FPU_DISABLED, EXC_SLOT_FPU_DISABLED ); } }

#define CHECKDEST(p) if( (p) == 0 ) { ERROR( "%08X: Branch/jump to NULL, CPU halted", sh4r.pc ); dreamcast_stop(); return FALSE; }

#define CHECKSLOTILLEGAL() if(sh4r.in_delay_slot) return sh4_raise_exception(EXC_SLOT_ILLEGAL)

static void sh4_switch_banks( )

{

    uint32_t tmp[8];

    memcpy( tmp, sh4r.r, sizeof(uint32_t)*8 );

    memcpy( sh4r.r, sh4r.r_bank, sizeof(uint32_t)*8 );

    memcpy( sh4r.r_bank, tmp, sizeof(uint32_t)*8 );

}

void sh4_write_sr( uint32_t newval )

{

    if( (newval ^ sh4r.sr) & SR_RB )

        sh4_switch_banks();

    sh4r.sr = newval;

    sh4r.t = (newval&SR_T) ? 1 : 0;

    sh4r.s = (newval&SR_S) ? 1 : 0;

    sh4r.m = (newval&SR_M) ? 1 : 0;

    sh4r.q = (newval&SR_Q) ? 1 : 0;

    intc_mask_changed();

}

static void sh4_write_float( uint32_t addr, int reg )

{

    if( IS_FPU_DOUBLESIZE() ) {

	if( reg & 1 ) {

	    sh4_write_long( addr, *((uint32_t *)&XF((reg)&0x0E)) );

	    sh4_write_long( addr+4, *((uint32_t *)&XF(reg)) );

	} else {

	    sh4_write_long( addr, *((uint32_t *)&FR(reg)) ); 

	    sh4_write_long( addr+4, *((uint32_t *)&FR((reg)|0x01)) );

	}

    } else {

	sh4_write_long( addr, *((uint32_t *)&FR((reg))) );

    }

}

static void sh4_read_float( uint32_t addr, int reg )

{

    if( IS_FPU_DOUBLESIZE() ) {

	if( reg & 1 ) {

	    *((uint32_t *)&XF((reg) & 0x0E)) = sh4_read_long(addr);

	    *((uint32_t *)&XF(reg)) = sh4_read_long(addr+4);

	} else {

	    *((uint32_t *)&FR(reg)) = sh4_read_long(addr);

	    *((uint32_t *)&FR((reg) | 0x01)) = sh4_read_long(addr+4);

	}

    } else {

	*((uint32_t *)&FR(reg)) = sh4_read_long(addr);

    }

}

uint32_t sh4_read_sr( void )

{

    /* synchronize sh4r.sr with the various bitflags */

    sh4r.sr &= SR_MQSTMASK;

    if( sh4r.t ) sh4r.sr |= SR_T;

    if( sh4r.s ) sh4r.sr |= SR_S;

    if( sh4r.m ) sh4r.sr |= SR_M;

    if( sh4r.q ) sh4r.sr |= SR_Q;

    return sh4r.sr;

}

/**

 * Raise a general CPU exception for the specified exception code.

 * (NOT for TRAPA or TLB exceptions)

 */

gboolean sh4_raise_exception( int code )

{

    RAISE( code, EXV_EXCEPTION );

}

gboolean sh4_raise_slot_exception( int normal_code, int slot_code ) {

    if( sh4r.in_delay_slot ) {

	return sh4_raise_exception(slot_code);

    } else {

	return sh4_raise_exception(normal_code);

    }

}

gboolean sh4_raise_tlb_exception( int code )

{

    RAISE( code, EXV_TLBMISS );

}

void sh4_accept_interrupt( void )

{

    uint32_t code = intc_accept_interrupt();

    sh4r.ssr = sh4_read_sr();

    sh4r.spc = sh4r.pc;

    sh4r.sgr = sh4r.r[15];

    sh4_write_sr( sh4r.ssr|SR_BL|SR_MD|SR_RB );

    MMIO_WRITE( MMU, INTEVT, code );

    sh4r.pc = sh4r.vbr + 0x600;

    sh4r.new_pc = sh4r.pc + 2;

    //    WARN( "Accepting interrupt %03X, from %08X => %08X", code, sh4r.spc, sh4r.pc );

}

gboolean sh4_execute_instruction( void )

{

    uint32_t pc;

    unsigned short ir;

    uint32_t tmp;

    float ftmp;

    double dtmp;

#define R0 sh4r.r[0]

    pc = sh4r.pc;

    if( pc > 0xFFFFFF00 ) {

	/* SYSCALL Magic */

	syscall_invoke( pc );

	sh4r.in_delay_slot = 0;

	pc = sh4r.pc = sh4r.pr;

	sh4r.new_pc = sh4r.pc + 2;

    }

    CHECKRALIGN16(pc);

    /* Read instruction */

    uint32_t pageaddr = pc >> 12;

    if( sh4_icache != NULL && pageaddr == sh4_icache_addr ) {

	ir = sh4_icache[(pc&0xFFF)>>1];

    } else {

	sh4_icache = (uint16_t *)mem_get_page(pc);

	if( ((uint32_t)sh4_icache) < MAX_IO_REGIONS ) {

	    /* If someone's actually been so daft as to try to execute out of an IO

	     * region, fallback on the full-blown memory read

	     */

	    sh4_icache = NULL;

	    ir = MEM_READ_WORD(pc);

	} else {

	    sh4_icache_addr = pageaddr;

	    ir = sh4_icache[(pc&0xFFF)>>1];

	}

    }

%%

AND Rm, Rn {: sh4r.r[Rn] &= sh4r.r[Rm]; :}

AND #imm, R0 {: R0 &= imm; :}

AND.B #imm, @(R0, GBR) {: MEM_WRITE_BYTE( R0 + sh4r.gbr, imm & MEM_READ_BYTE(R0 + sh4r.gbr) ); :}

NOT Rm, Rn {: sh4r.r[Rn] = ~sh4r.r[Rm]; :}

OR Rm, Rn {: sh4r.r[Rn] |= sh4r.r[Rm]; :}

OR #imm, R0  {: R0 |= imm; :}

OR.B #imm, @(R0, GBR) {: MEM_WRITE_BYTE( R0 + sh4r.gbr, imm | MEM_READ_BYTE(R0 + sh4r.gbr) ); :}

TAS.B @Rn {:

    tmp = MEM_READ_BYTE( sh4r.r[Rn] );

    sh4r.t = ( tmp == 0 ? 1 : 0 );

    MEM_WRITE_BYTE( sh4r.r[Rn], tmp | 0x80 );

:}

TST Rm, Rn {: sh4r.t = (sh4r.r[Rn]&sh4r.r[Rm] ? 0 : 1); :}

TST #imm, R0 {: sh4r.t = (R0 & imm ? 0 : 1); :}

TST.B #imm, @(R0, GBR) {: sh4r.t = ( MEM_READ_BYTE(R0 + sh4r.gbr) & imm ? 0 : 1 ); :}

XOR Rm, Rn {: sh4r.r[Rn] ^= sh4r.r[Rm]; :}

XOR #imm, R0 {: R0 ^= imm; :}

XOR.B #imm, @(R0, GBR) {: MEM_WRITE_BYTE( R0 + sh4r.gbr, imm ^ MEM_READ_BYTE(R0 + sh4r.gbr) ); :}

XTRCT Rm, Rn {: sh4r.r[Rn] = (sh4r.r[Rn]>>16) | (sh4r.r[Rm]<<16); :}

ROTL Rn {:

    sh4r.t = sh4r.r[Rn] >> 31;

    sh4r.r[Rn] <<= 1;

    sh4r.r[Rn] |= sh4r.t;

:}

ROTR Rn {:

    sh4r.t = sh4r.r[Rn] & 0x00000001;

    sh4r.r[Rn] >>= 1;

    sh4r.r[Rn] |= (sh4r.t << 31);

:}

ROTCL Rn {:

    tmp = sh4r.r[Rn] >> 31;

    sh4r.r[Rn] <<= 1;

    sh4r.r[Rn] |= sh4r.t;

    sh4r.t = tmp;

:}

ROTCR Rn {:

    tmp = sh4r.r[Rn] & 0x00000001;

    sh4r.r[Rn] >>= 1;

    sh4r.r[Rn] |= (sh4r.t << 31 );

    sh4r.t = tmp;

:}

SHAD Rm, Rn {:

    tmp = sh4r.r[Rm];

    if( (tmp & 0x80000000) == 0 ) sh4r.r[Rn] <<= (tmp&0x1f);

    else if( (tmp & 0x1F) == 0 )  

        sh4r.r[Rn] = ((int32_t)sh4r.r[Rn]) >> 31;

    else 

	sh4r.r[Rn] = ((int32_t)sh4r.r[Rn]) >> (((~sh4r.r[Rm]) & 0x1F)+1);

:}

SHLD Rm, Rn {:

    tmp = sh4r.r[Rm];

    if( (tmp & 0x80000000) == 0 ) sh4r.r[Rn] <<= (tmp&0x1f);

    else if( (tmp & 0x1F) == 0 ) sh4r.r[Rn] = 0;

    else sh4r.r[Rn] >>= (((~tmp) & 0x1F)+1);

:}

SHAL Rn {:

    sh4r.t = sh4r.r[Rn] >> 31;

    sh4r.r[Rn] <<= 1;

:}

SHAR Rn {:

    sh4r.t = sh4r.r[Rn] & 0x00000001;

    sh4r.r[Rn] = ((int32_t)sh4r.r[Rn]) >> 1;

:}

SHLL Rn {: sh4r.t = sh4r.r[Rn] >> 31; sh4r.r[Rn] <<= 1; :}

SHLR Rn {: sh4r.t = sh4r.r[Rn] & 0x00000001; sh4r.r[Rn] >>= 1; :}

SHLL2 Rn {: sh4r.r[Rn] <<= 2; :}

SHLR2 Rn {: sh4r.r[Rn] >>= 2; :}

SHLL8 Rn {: sh4r.r[Rn] <<= 8; :}

SHLR8 Rn {: sh4r.r[Rn] >>= 8; :}

SHLL16 Rn {: sh4r.r[Rn] <<= 16; :}

SHLR16 Rn {: sh4r.r[Rn] >>= 16; :}

EXTU.B Rm, Rn {: sh4r.r[Rn] = sh4r.r[Rm]&0x000000FF; :}

EXTU.W Rm, Rn {: sh4r.r[Rn] = sh4r.r[Rm]&0x0000FFFF; :}

EXTS.B Rm, Rn {: sh4r.r[Rn] = SIGNEXT8( sh4r.r[Rm]&0x000000FF ); :}

EXTS.W Rm, Rn {: sh4r.r[Rn] = SIGNEXT16( sh4r.r[Rm]&0x0000FFFF ); :}

SWAP.B Rm, Rn {: sh4r.r[Rn] = (sh4r.r[Rm]&0xFFFF0000) | ((sh4r.r[Rm]&0x0000FF00)>>8) | ((sh4r.r[Rm]&0x000000FF)<<8); :}

SWAP.W Rm, Rn {: sh4r.r[Rn] = (sh4r.r[Rm]>>16) | (sh4r.r[Rm]<<16); :}

CLRT {: sh4r.t = 0; :}

SETT {: sh4r.t = 1; :}

CLRMAC {: sh4r.mac = 0; :}

LDTLB {: /* TODO */ :}

CLRS {: sh4r.s = 0; :}

SETS {: sh4r.s = 1; :}

MOVT Rn {: sh4r.r[Rn] = sh4r.t; :}

NOP {: /* NOP */ :}

PREF @Rn {:

     tmp = sh4r.r[Rn];

     if( (tmp & 0xFC000000) == 0xE0000000 ) {

	 sh4_flush_store_queue(tmp);

     }

:}

OCBI @Rn {: :}

OCBP @Rn {: :}

OCBWB @Rn {: :}

MOVCA.L R0, @Rn {:

    tmp = sh4r.r[Rn];

    CHECKWALIGN32(tmp);

    MEM_WRITE_LONG( tmp, R0 );

:}

MOV.B Rm, @(R0, Rn) {: MEM_WRITE_BYTE( R0 + sh4r.r[Rn], sh4r.r[Rm] ); :}

MOV.W Rm, @(R0, Rn) {: 

    CHECKWALIGN16( R0 + sh4r.r[Rn] );

    MEM_WRITE_WORD( R0 + sh4r.r[Rn], sh4r.r[Rm] );

:}

MOV.L Rm, @(R0, Rn) {:

    CHECKWALIGN32( R0 + sh4r.r[Rn] );

    MEM_WRITE_LONG( R0 + sh4r.r[Rn], sh4r.r[Rm] );

:}

MOV.B @(R0, Rm), Rn {: sh4r.r[Rn] = MEM_READ_BYTE( R0 + sh4r.r[Rm] ); :}

MOV.W @(R0, Rm), Rn {: CHECKRALIGN16( R0 + sh4r.r[Rm] );

                    sh4r.r[Rn] = MEM_READ_WORD( R0 + sh4r.r[Rm] );

:}

MOV.L @(R0, Rm), Rn {: CHECKRALIGN32( R0 + sh4r.r[Rm] );

                    sh4r.r[Rn] = MEM_READ_LONG( R0 + sh4r.r[Rm] );

:}

MOV.L Rm, @(disp, Rn) {:

    tmp = sh4r.r[Rn] + disp;

    CHECKWALIGN32( tmp );

    MEM_WRITE_LONG( tmp, sh4r.r[Rm] );

:}

MOV.B Rm, @Rn {: MEM_WRITE_BYTE( sh4r.r[Rn], sh4r.r[Rm] ); :}

MOV.W Rm, @Rn {: CHECKWALIGN16( sh4r.r[Rn] ); MEM_WRITE_WORD( sh4r.r[Rn], sh4r.r[Rm] ); :}

MOV.L Rm, @Rn {: CHECKWALIGN32( sh4r.r[Rn] ); MEM_WRITE_LONG( sh4r.r[Rn], sh4r.r[Rm] ); :}

MOV.B Rm, @-Rn {: sh4r.r[Rn] --; MEM_WRITE_BYTE( sh4r.r[Rn], sh4r.r[Rm] ); :}

MOV.W Rm, @-Rn {: sh4r.r[Rn] -= 2; CHECKWALIGN16( sh4r.r[Rn] ); MEM_WRITE_WORD( sh4r.r[Rn], sh4r.r[Rm] ); :}

MOV.L Rm, @-Rn {: sh4r.r[Rn] -= 4; CHECKWALIGN32( sh4r.r[Rn] ); MEM_WRITE_LONG( sh4r.r[Rn], sh4r.r[Rm] ); :}

MOV.L @(disp, Rm), Rn {:

    tmp = sh4r.r[Rm] + disp;

    CHECKRALIGN32( tmp );

    sh4r.r[Rn] = MEM_READ_LONG( tmp );

:}

MOV.B @Rm, Rn {: sh4r.r[Rn] = MEM_READ_BYTE( sh4r.r[Rm] ); :}

MOV.W @Rm, Rn {: CHECKRALIGN16( sh4r.r[Rm] ); sh4r.r[Rn] = MEM_READ_WORD( sh4r.r[Rm] ); :}

MOV.L @Rm, Rn {: CHECKRALIGN32( sh4r.r[Rm] ); sh4r.r[Rn] = MEM_READ_LONG( sh4r.r[Rm] ); :}

MOV Rm, Rn {: sh4r.r[Rn] = sh4r.r[Rm]; :}

MOV.B @Rm+, Rn {: sh4r.r[Rn] = MEM_READ_BYTE( sh4r.r[Rm] ); sh4r.r[Rm] ++; :}

MOV.W @Rm+, Rn {: CHECKRALIGN16( sh4r.r[Rm] ); sh4r.r[Rn] = MEM_READ_WORD( sh4r.r[Rm] ); sh4r.r[Rm] += 2; :}

MOV.L @Rm+, Rn {: CHECKRALIGN32( sh4r.r[Rm] ); sh4r.r[Rn] = MEM_READ_LONG( sh4r.r[Rm] ); sh4r.r[Rm] += 4; :}

MOV.L @(disp, PC), Rn {:

    CHECKSLOTILLEGAL();

    tmp = (pc&0xFFFFFFFC) + disp + 4;

    sh4r.r[Rn] = MEM_READ_LONG( tmp );

:}

MOV.B R0, @(disp, GBR) {: MEM_WRITE_BYTE( sh4r.gbr + disp, R0 ); :}

MOV.W R0, @(disp, GBR) {:

    tmp = sh4r.gbr + disp;

    CHECKWALIGN16( tmp );

    MEM_WRITE_WORD( tmp, R0 );

:}

MOV.L R0, @(disp, GBR) {:

    tmp = sh4r.gbr + disp;

    CHECKWALIGN32( tmp );

    MEM_WRITE_LONG( tmp, R0 );

:}

MOV.B @(disp, GBR), R0 {: R0 = MEM_READ_BYTE( sh4r.gbr + disp ); :}

MOV.W @(disp, GBR), R0 {: 

    tmp = sh4r.gbr + disp;

    CHECKRALIGN16( tmp );

    R0 = MEM_READ_WORD( tmp );

:}

MOV.L @(disp, GBR), R0 {:

    tmp = sh4r.gbr + disp;

    CHECKRALIGN32( tmp );

    R0 = MEM_READ_LONG( tmp );

:}

MOV.B R0, @(disp, Rn) {: MEM_WRITE_BYTE( sh4r.r[Rn] + disp, R0 ); :}

MOV.W R0, @(disp, Rn) {: 

    tmp = sh4r.r[Rn] + disp;

    CHECKWALIGN16( tmp );

    MEM_WRITE_WORD( tmp, R0 );

:}

MOV.B @(disp, Rm), R0 {: R0 = MEM_READ_BYTE( sh4r.r[Rm] + disp ); :}

MOV.W @(disp, Rm), R0 {: 

    tmp = sh4r.r[Rm] + disp;

    CHECKRALIGN16( tmp );

    R0 = MEM_READ_WORD( tmp );

:}

MOV.W @(disp, PC), Rn {:

    CHECKSLOTILLEGAL();

    tmp = pc + 4 + disp;

    sh4r.r[Rn] = MEM_READ_WORD( tmp );

:}

MOVA @(disp, PC), R0 {:

    CHECKSLOTILLEGAL();

    R0 = (pc&0xFFFFFFFC) + disp + 4;

:}

MOV #imm, Rn {:  sh4r.r[Rn] = imm; :}

CMP/EQ #imm, R0 {: sh4r.t = ( R0 == imm ? 1 : 0 ); :}

CMP/EQ Rm, Rn {: sh4r.t = ( sh4r.r[Rm] == sh4r.r[Rn] ? 1 : 0 ); :}

CMP/GE Rm, Rn {: sh4r.t = ( ((int32_t)sh4r.r[Rn]) >= ((int32_t)sh4r.r[Rm]) ? 1 : 0 ); :}

CMP/GT Rm, Rn {: sh4r.t = ( ((int32_t)sh4r.r[Rn]) > ((int32_t)sh4r.r[Rm]) ? 1 : 0 ); :}

CMP/HI Rm, Rn {: sh4r.t = ( sh4r.r[Rn] > sh4r.r[Rm] ? 1 : 0 ); :}

CMP/HS Rm, Rn {: sh4r.t = ( sh4r.r[Rn] >= sh4r.r[Rm] ? 1 : 0 ); :}

CMP/PL Rn {: sh4r.t = ( ((int32_t)sh4r.r[Rn]) > 0 ? 1 : 0 ); :}

CMP/PZ Rn {: sh4r.t = ( ((int32_t)sh4r.r[Rn]) >= 0 ? 1 : 0 ); :}

CMP/STR Rm, Rn {: 

    /* set T = 1 if any byte in RM & RN is the same */

    tmp = sh4r.r[Rm] ^ sh4r.r[Rn];

    sh4r.t = ((tmp&0x000000FF)==0 || (tmp&0x0000FF00)==0 ||

             (tmp&0x00FF0000)==0 || (tmp&0xFF000000)==0)?1:0;

:}

ADD Rm, Rn {: sh4r.r[Rn] += sh4r.r[Rm]; :}

ADD #imm, Rn {: sh4r.r[Rn] += imm; :}

ADDC Rm, Rn {:

    tmp = sh4r.r[Rn];

    sh4r.r[Rn] += sh4r.r[Rm] + sh4r.t;

    sh4r.t = ( sh4r.r[Rn] < tmp || (sh4r.r[Rn] == tmp && sh4r.t != 0) ? 1 : 0 );

:}

ADDV Rm, Rn {:

    tmp = sh4r.r[Rn] + sh4r.r[Rm];

    sh4r.t = ( (sh4r.r[Rn]>>31) == (sh4r.r[Rm]>>31) && ((sh4r.r[Rn]>>31) != (tmp>>31)) );

    sh4r.r[Rn] = tmp;

:}

DIV0U {: sh4r.m = sh4r.q = sh4r.t = 0; :}

DIV0S Rm, Rn {: 

    sh4r.q = sh4r.r[Rn]>>31;

    sh4r.m = sh4r.r[Rm]>>31;

    sh4r.t = sh4r.q ^ sh4r.m;

:}

DIV1 Rm, Rn {:

    /* This is just from the sh4p manual with some

     * simplifications (someone want to check it's correct? :)

     * Why they couldn't just provide a real DIV instruction...

     */

    uint32_t tmp0, tmp1, tmp2, dir;

    dir = sh4r.q ^ sh4r.m;

    sh4r.q = (sh4r.r[Rn] >> 31);

    tmp2 = sh4r.r[Rm];

    sh4r.r[Rn] = (sh4r.r[Rn] << 1) | sh4r.t;

    tmp0 = sh4r.r[Rn];

    if( dir ) {

         sh4r.r[Rn] += tmp2;

         tmp1 = (sh4r.r[Rn]<tmp0 ? 1 : 0 );

    } else {

         sh4r.r[Rn] -= tmp2;

         tmp1 = (sh4r.r[Rn]>tmp0 ? 1 : 0 );

    }

    sh4r.q ^= sh4r.m ^ tmp1;

    sh4r.t = ( sh4r.q == sh4r.m ? 1 : 0 );

:}

DMULS.L Rm, Rn {: sh4r.mac = SIGNEXT32(sh4r.r[Rm]) * SIGNEXT32(sh4r.r[Rn]); :}

DMULU.L Rm, Rn {: sh4r.mac = ((uint64_t)sh4r.r[Rm]) * ((uint64_t)sh4r.r[Rn]); :}

DT Rn {:

    sh4r.r[Rn] --;

    sh4r.t = ( sh4r.r[Rn] == 0 ? 1 : 0 );

:}

MAC.W @Rm+, @Rn+ {:

    CHECKRALIGN16( sh4r.r[Rn] );

    CHECKRALIGN16( sh4r.r[Rm] );

    int32_t stmp = SIGNEXT16(MEM_READ_WORD(sh4r.r[Rn]));

    sh4r.r[Rn] += 2;

    stmp = stmp * SIGNEXT16(MEM_READ_WORD(sh4r.r[Rm]));

    sh4r.r[Rm] += 2;

    if( sh4r.s ) {

	int64_t tmpl = (int64_t)((int32_t)sh4r.mac) + (int64_t)stmp;

	if( tmpl > (int64_t)0x000000007FFFFFFFLL ) {

	    sh4r.mac = 0x000000017FFFFFFFLL;

	} else if( tmpl < (int64_t)0xFFFFFFFF80000000LL ) {

	    sh4r.mac = 0x0000000180000000LL;

	} else {

	    sh4r.mac = (sh4r.mac & 0xFFFFFFFF00000000LL) |

		((uint32_t)(sh4r.mac + stmp));

	}

    } else {

	sh4r.mac += SIGNEXT32(stmp);

    }

:}

MAC.L @Rm+, @Rn+ {:

    CHECKRALIGN32( sh4r.r[Rm] );

    CHECKRALIGN32( sh4r.r[Rn] );

    int64_t tmpl = SIGNEXT32(MEM_READ_LONG(sh4r.r[Rn]));

    sh4r.r[Rn] += 4;

    tmpl = tmpl * SIGNEXT32(MEM_READ_LONG(sh4r.r[Rm])) + sh4r.mac;

    sh4r.r[Rm] += 4;

    if( sh4r.s ) {

        /* 48-bit Saturation. Yuch */

        if( tmpl < (int64_t)0xFFFF800000000000LL )

            tmpl = 0xFFFF800000000000LL;

        else if( tmpl > (int64_t)0x00007FFFFFFFFFFFLL )

            tmpl = 0x00007FFFFFFFFFFFLL;

    }

    sh4r.mac = tmpl;

:}

MUL.L Rm, Rn {: sh4r.mac = (sh4r.mac&0xFFFFFFFF00000000LL) |

                        (sh4r.r[Rm] * sh4r.r[Rn]); :}

MULU.W Rm, Rn {:

    sh4r.mac = (sh4r.mac&0xFFFFFFFF00000000LL) |

               (uint32_t)((sh4r.r[Rm]&0xFFFF) * (sh4r.r[Rn]&0xFFFF));

:}

MULS.W Rm, Rn {:

    sh4r.mac = (sh4r.mac&0xFFFFFFFF00000000LL) |

               (uint32_t)(SIGNEXT32(sh4r.r[Rm]&0xFFFF) * SIGNEXT32(sh4r.r[Rn]&0xFFFF));

:}

NEGC Rm, Rn {:

    tmp = 0 - sh4r.r[Rm];

    sh4r.r[Rn] = tmp - sh4r.t;

    sh4r.t = ( 0<tmp || tmp<sh4r.r[Rn] ? 1 : 0 );

:}

NEG Rm, Rn {: sh4r.r[Rn] = 0 - sh4r.r[Rm]; :}

SUB Rm, Rn {: sh4r.r[Rn] -= sh4r.r[Rm]; :}

SUBC Rm, Rn {: 

    tmp = sh4r.r[Rn];

    sh4r.r[Rn] = sh4r.r[Rn] - sh4r.r[Rm] - sh4r.t;

    sh4r.t = (sh4r.r[Rn] > tmp || (sh4r.r[Rn] == tmp && sh4r.t == 1));

:}

BRAF Rn {:

     CHECKSLOTILLEGAL();

     CHECKDEST( pc + 4 + sh4r.r[Rn] );

     sh4r.in_delay_slot = 1;

     sh4r.pc = sh4r.new_pc;

     sh4r.new_pc = pc + 4 + sh4r.r[Rn];

     return TRUE;

:}

BSRF Rn {:

     CHECKSLOTILLEGAL();

     CHECKDEST( pc + 4 + sh4r.r[Rn] );

     sh4r.in_delay_slot = 1;

     sh4r.pr = sh4r.pc + 4;

     sh4r.pc = sh4r.new_pc;

     sh4r.new_pc = pc + 4 + sh4r.r[Rn];

     TRACE_CALL( pc, sh4r.new_pc );

     return TRUE;

:}

BT disp {:

    CHECKSLOTILLEGAL();

    if( sh4r.t ) {

        CHECKDEST( sh4r.pc + disp + 4 )

        sh4r.pc += disp + 4;

        sh4r.new_pc = sh4r.pc + 2;

        return TRUE;

    }

:}

BF disp {:

    CHECKSLOTILLEGAL();

    if( !sh4r.t ) {

        CHECKDEST( sh4r.pc + disp + 4 )

        sh4r.pc += disp + 4;

        sh4r.new_pc = sh4r.pc + 2;

        return TRUE;

    }

:}

BT/S disp {:

    CHECKSLOTILLEGAL();

    if( sh4r.t ) {

        CHECKDEST( sh4r.pc + disp + 4 )

        sh4r.in_delay_slot = 1;

        sh4r.pc = sh4r.new_pc;

        sh4r.new_pc = pc + disp + 4;

        sh4r.in_delay_slot = 1;

        return TRUE;

    }

:}

BF/S disp {:

    CHECKSLOTILLEGAL();

    if( !sh4r.t ) {

        CHECKDEST( sh4r.pc + disp + 4 )

        sh4r.in_delay_slot = 1;

        sh4r.pc = sh4r.new_pc;

        sh4r.new_pc = pc + disp + 4;

        return TRUE;

    }

:}

BRA disp {:

    CHECKSLOTILLEGAL();

    CHECKDEST( sh4r.pc + disp + 4 );

    sh4r.in_delay_slot = 1;

    sh4r.pc = sh4r.new_pc;

    sh4r.new_pc = pc + 4 + disp;

    return TRUE;

:}

BSR disp {:

    CHECKDEST( sh4r.pc + disp + 4 );

    CHECKSLOTILLEGAL();

    sh4r.in_delay_slot = 1;

    sh4r.pr = pc + 4;

    sh4r.pc = sh4r.new_pc;

    sh4r.new_pc = pc + 4 + disp;

    TRACE_CALL( pc, sh4r.new_pc );

    return TRUE;

:}

TRAPA #imm {:

    CHECKSLOTILLEGAL();

    MMIO_WRITE( MMU, TRA, imm<<2 );

    sh4r.pc += 2;

    sh4_raise_exception( EXC_TRAP );

:}

RTS {: 

    CHECKSLOTILLEGAL();

    CHECKDEST( sh4r.pr );

    sh4r.in_delay_slot = 1;

    sh4r.pc = sh4r.new_pc;

    sh4r.new_pc = sh4r.pr;

    TRACE_RETURN( pc, sh4r.new_pc );

    return TRUE;

:}

SLEEP {:

    if( MMIO_READ( CPG, STBCR ) & 0x80 ) {

	sh4r.sh4_state = SH4_STATE_STANDBY;

    } else {

	sh4r.sh4_state = SH4_STATE_SLEEP;

    }

    return FALSE; /* Halt CPU */

:}

RTE {:

    CHECKPRIV();

    CHECKDEST( sh4r.spc );

    CHECKSLOTILLEGAL();

    sh4r.in_delay_slot = 1;

    sh4r.pc = sh4r.new_pc;

    sh4r.new_pc = sh4r.spc;

    sh4_write_sr( sh4r.ssr );

    return TRUE;

:}

JMP @Rn {:

    CHECKDEST( sh4r.r[Rn] );

    CHECKSLOTILLEGAL();

    sh4r.in_delay_slot = 1;

    sh4r.pc = sh4r.new_pc;

    sh4r.new_pc = sh4r.r[Rn];

    return TRUE;

:}

JSR @Rn {:

    CHECKDEST( sh4r.r[Rn] );

    CHECKSLOTILLEGAL();

    sh4r.in_delay_slot = 1;

    sh4r.pc = sh4r.new_pc;

    sh4r.new_pc = sh4r.r[Rn];

    sh4r.pr = pc + 4;

    TRACE_CALL( pc, sh4r.new_pc );

    return TRUE;

:}

STS MACH, Rn {: sh4r.r[Rn] = (sh4r.mac>>32); :}

STS.L MACH, @-Rn {:

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], (sh4r.mac>>32) );

:}

STC.L SR, @-Rn {:

    CHECKPRIV();

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4_read_sr() );

:}

LDS.L @Rm+, MACH {:

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.mac = (sh4r.mac & 0x00000000FFFFFFFF) |

               (((uint64_t)MEM_READ_LONG(sh4r.r[Rm]))<<32);

    sh4r.r[Rm] += 4;

:}

LDC.L @Rm+, SR {:

    CHECKSLOTILLEGAL();

    CHECKPRIV();

    CHECKWALIGN32( sh4r.r[Rm] );

    sh4_write_sr( MEM_READ_LONG(sh4r.r[Rm]) );

    sh4r.r[Rm] +=4;

:}

LDS Rm, MACH {:

    sh4r.mac = (sh4r.mac & 0x00000000FFFFFFFF) |

               (((uint64_t)sh4r.r[Rm])<<32);

:}

LDC Rm, SR {:

    CHECKSLOTILLEGAL();

    CHECKPRIV();

    sh4_write_sr( sh4r.r[Rm] );

:}

LDC Rm, SGR {:

    CHECKPRIV();

    sh4r.sgr = sh4r.r[Rm];

:}

LDC.L @Rm+, SGR {:

    CHECKPRIV();

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.sgr = MEM_READ_LONG(sh4r.r[Rm]);

    sh4r.r[Rm] +=4;

:}

STS MACL, Rn {: sh4r.r[Rn] = (uint32_t)sh4r.mac; :}

STS.L MACL, @-Rn {:

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], (uint32_t)sh4r.mac );

:}

STC.L GBR, @-Rn {:

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.gbr );

:}

LDS.L @Rm+, MACL {:

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.mac = (sh4r.mac & 0xFFFFFFFF00000000LL) |

               (uint64_t)((uint32_t)MEM_READ_LONG(sh4r.r[Rm]));

    sh4r.r[Rm] += 4;

:}

LDC.L @Rm+, GBR {:

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.gbr = MEM_READ_LONG(sh4r.r[Rm]);

    sh4r.r[Rm] +=4;

:}

LDS Rm, MACL {:

    sh4r.mac = (sh4r.mac & 0xFFFFFFFF00000000LL) |

               (uint64_t)((uint32_t)(sh4r.r[Rm]));

:}

LDC Rm, GBR {: sh4r.gbr = sh4r.r[Rm]; :}

STS PR, Rn {: sh4r.r[Rn] = sh4r.pr; :}

STS.L PR, @-Rn {:

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.pr );

:}

STC.L VBR, @-Rn {:

    CHECKPRIV();

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.vbr );

:}

LDS.L @Rm+, PR {:

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.pr = MEM_READ_LONG( sh4r.r[Rm] );

    sh4r.r[Rm] += 4;

:}

LDC.L @Rm+, VBR {:

    CHECKPRIV();

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.vbr = MEM_READ_LONG(sh4r.r[Rm]);

    sh4r.r[Rm] +=4;

:}

LDS Rm, PR {: sh4r.pr = sh4r.r[Rm]; :}

LDC Rm, VBR {:

    CHECKPRIV();

    sh4r.vbr = sh4r.r[Rm];

:}

STC SGR, Rn {:

    CHECKPRIV();

    sh4r.r[Rn] = sh4r.sgr;

:}

STC.L SGR, @-Rn {:

    CHECKPRIV();

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.sgr );

:}

STC.L SSR, @-Rn {:

    CHECKPRIV();

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.ssr );

:}

LDC.L @Rm+, SSR {:

    CHECKPRIV();

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.ssr = MEM_READ_LONG(sh4r.r[Rm]);

    sh4r.r[Rm] +=4;

:}

LDC Rm, SSR {:

    CHECKPRIV();

    sh4r.ssr = sh4r.r[Rm];

:}

STC.L SPC, @-Rn {:

    CHECKPRIV();

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.spc );

:}

LDC.L @Rm+, SPC {:

    CHECKPRIV();

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.spc = MEM_READ_LONG(sh4r.r[Rm]);

    sh4r.r[Rm] +=4;

:}

LDC Rm, SPC {:

    CHECKPRIV();

    sh4r.spc = sh4r.r[Rm];

:}

STS FPUL, Rn {: sh4r.r[Rn] = sh4r.fpul; :}

STS.L FPUL, @-Rn {:

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.fpul );

:}

LDS.L @Rm+, FPUL {:

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.fpul = MEM_READ_LONG(sh4r.r[Rm]);

    sh4r.r[Rm] +=4;

:}

LDS Rm, FPUL {: sh4r.fpul = sh4r.r[Rm]; :}

STS FPSCR, Rn {: sh4r.r[Rn] = sh4r.fpscr; :}

STS.L FPSCR, @-Rn {:

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.fpscr );

:}

LDS.L @Rm+, FPSCR {:

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.fpscr = MEM_READ_LONG(sh4r.r[Rm]);

    sh4r.r[Rm] +=4;

    sh4r.fr_bank = &sh4r.fr[(sh4r.fpscr&FPSCR_FR)>>21][0];

:}

LDS Rm, FPSCR {: 

    sh4r.fpscr = sh4r.r[Rm]; 

    sh4r.fr_bank = &sh4r.fr[(sh4r.fpscr&FPSCR_FR)>>21][0];

:}

STC DBR, Rn {: CHECKPRIV(); sh4r.r[Rn] = sh4r.dbr; :}

STC.L DBR, @-Rn {:

    CHECKPRIV();

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.dbr );

:}

LDC.L @Rm+, DBR {:

    CHECKPRIV();

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.dbr = MEM_READ_LONG(sh4r.r[Rm]);

    sh4r.r[Rm] +=4;

:}

LDC Rm, DBR {:

    CHECKPRIV();

    sh4r.dbr = sh4r.r[Rm];

:}

STC.L Rm_BANK, @-Rn {:

    CHECKPRIV();

    sh4r.r[Rn] -= 4;

    CHECKWALIGN32( sh4r.r[Rn] );

    MEM_WRITE_LONG( sh4r.r[Rn], sh4r.r_bank[Rm_BANK] );

:}

LDC.L @Rm+, Rn_BANK {:

    CHECKPRIV();

    CHECKRALIGN32( sh4r.r[Rm] );

    sh4r.r_bank[Rn_BANK] = MEM_READ_LONG( sh4r.r[Rm] );