/**

  * $Id$

  * 

  * SH4 => x86 translation. This version does no real optimization, it just

  * outputs straight-line x86 code - it mainly exists to provide a baseline

  * to test the optimizing versions against.

  *

  * Copyright (c) 2007 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.

  */

 #include <assert.h>

 #include <math.h>

 #ifndef NDEBUG

 #define DEBUG_JUMPS 1

 #endif

 #include "lxdream.h"

 #include "sh4/sh4core.h"

 #include "sh4/sh4dasm.h"

 #include "sh4/sh4trans.h"

 #include "sh4/sh4stat.h"

 #include "sh4/sh4mmio.h"

 #include "sh4/mmu.h"

 #include "xlat/xltcache.h"

 #include "xlat/x86/x86op.h"

 #include "xlat/xlatdasm.h"

 #include "clock.h"

 #define DEFAULT_BACKPATCH_SIZE 4096

 /* Offset of a reg relative to the sh4r structure */

 #define REG_OFFSET(reg)  (((char *)&sh4r.reg) - ((char *)&sh4r) - 128)

 #define R_T      REG_OFFSET(t)

 #define R_Q      REG_OFFSET(q)

 #define R_S      REG_OFFSET(s)

 #define R_M      REG_OFFSET(m)

 #define R_SR     REG_OFFSET(sr)

 #define R_GBR    REG_OFFSET(gbr)

 #define R_SSR    REG_OFFSET(ssr)

 #define R_SPC    REG_OFFSET(spc)

 #define R_VBR    REG_OFFSET(vbr)

 #define R_MACH   REG_OFFSET(mac)+4

 #define R_MACL   REG_OFFSET(mac)

 #define R_PC     REG_OFFSET(pc)

 #define R_NEW_PC REG_OFFSET(new_pc)

 #define R_PR     REG_OFFSET(pr)

 #define R_SGR    REG_OFFSET(sgr)

 #define R_FPUL   REG_OFFSET(fpul)

 #define R_FPSCR  REG_OFFSET(fpscr)

 #define R_DBR    REG_OFFSET(dbr)

 #define R_R(rn)  REG_OFFSET(r[rn])

 #define R_FR(f)  REG_OFFSET(fr[0][(f)^1])

 #define R_XF(f)  REG_OFFSET(fr[1][(f)^1])

 #define R_DR(f)  REG_OFFSET(fr[(f)&1][(f)&0x0E])

 #define R_DRL(f) REG_OFFSET(fr[(f)&1][(f)|0x01])

 #define R_DRH(f) REG_OFFSET(fr[(f)&1][(f)&0x0E])

 #define DELAY_NONE 0

 #define DELAY_PC 1

 #define DELAY_PC_PR 2

 #define SH4_MODE_UNKNOWN -1

 struct backpatch_record {

     uint32_t fixup_offset;

     uint32_t fixup_icount;

     int32_t exc_code;

 };

 /** 

  * Struct to manage internal translation state. This state is not saved -

  * it is only valid between calls to sh4_translate_begin_block() and

  * sh4_translate_end_block()

  */

 struct sh4_x86_state {

     int in_delay_slot;

     uint8_t *code;

     gboolean fpuen_checked; /* true if we've already checked fpu enabled. */

     gboolean branch_taken; /* true if we branched unconditionally */

     gboolean double_prec; /* true if FPU is in double-precision mode */

     gboolean double_size; /* true if FPU is in double-size mode */

     gboolean sse3_enabled; /* true if host supports SSE3 instructions */

     uint32_t block_start_pc;

     uint32_t stack_posn;   /* Trace stack height for alignment purposes */

     uint32_t sh4_mode;     /* Mirror of sh4r.xlat_sh4_mode */

     int tstate;

     /* mode settings */

     gboolean tlb_on; /* True if tlb translation is active */

     struct mem_region_fn **priv_address_space;

     struct mem_region_fn **user_address_space;

     /* Instrumentation */

     xlat_block_begin_callback_t begin_callback;

     xlat_block_end_callback_t end_callback;

     gboolean fastmem;

     /* Allocated memory for the (block-wide) back-patch list */

     struct backpatch_record *backpatch_list;

     uint32_t backpatch_posn;

     uint32_t backpatch_size;

 };

 static struct sh4_x86_state sh4_x86;

 static uint8_t sh4_entry_stub[128];

 typedef FASTCALL void (*entry_point_t)(void *);

 entry_point_t sh4_translate_enter;

 static uint32_t max_int = 0x7FFFFFFF;

 static uint32_t min_int = 0x80000000;

 static uint32_t save_fcw; /* save value for fpu control word */

 static uint32_t trunc_fcw = 0x0F7F; /* fcw value for truncation mode */

 static void sh4_x86_translate_unlink_block( void *use_list );

 static struct xlat_target_fns x86_target_fns = {

 	sh4_x86_translate_unlink_block

 };	

 gboolean is_sse3_supported()

 {

     uint32_t features;

     __asm__ __volatile__(

         "mov $0x01, %%eax\n\t"

         "cpuid\n\t" : "=c" (features) : : "eax", "edx", "ebx");

     return (features & 1) ? TRUE : FALSE;

 }

 void sh4_translate_set_address_space( struct mem_region_fn **priv, struct mem_region_fn **user )

 {

     sh4_x86.priv_address_space = priv;

     sh4_x86.user_address_space = user;

 }

 void sh4_translate_write_entry_stub(void)

 {

 	mem_unprotect(sh4_entry_stub, sizeof(sh4_entry_stub));

 	xlat_output = sh4_entry_stub;

 	PUSH_r32(REG_EBP);

 	MOVP_immptr_rptr( ((uint8_t *)&sh4r) + 128, REG_EBP );

 	PUSH_r32(REG_EBX);

 	PUSH_r32(REG_SAVE1);

 	PUSH_r32(REG_SAVE2);

 #if SIZEOF_VOID_P == 8

     PUSH_r32(REG_SAVE3);

     PUSH_r32(REG_SAVE4);

     CALL_r32( REG_ARG1 );

     POP_r32(REG_SAVE4);

     POP_r32(REG_SAVE3);

 #else

     SUBL_imms_r32( 8, REG_ESP ); 

 	CALL_r32( REG_ARG1 );

 	ADDL_imms_r32( 8, REG_ESP );

 #endif

 	POP_r32(REG_SAVE2);	

 	POP_r32(REG_SAVE1);

 	POP_r32(REG_EBX);

 	POP_r32(REG_EBP);

 	RET();

 	sh4_translate_enter = (entry_point_t)sh4_entry_stub;

 }

 void sh4_translate_init(void)

 {

     sh4_x86.backpatch_list = malloc(DEFAULT_BACKPATCH_SIZE);

     sh4_x86.backpatch_size = DEFAULT_BACKPATCH_SIZE / sizeof(struct backpatch_record);

     sh4_x86.begin_callback = NULL;

     sh4_x86.end_callback = NULL;

     sh4_x86.fastmem = TRUE;

     sh4_x86.sse3_enabled = is_sse3_supported();

     xlat_set_target_fns(&x86_target_fns);

     sh4_translate_set_address_space( sh4_address_space, sh4_user_address_space );

     sh4_translate_write_entry_stub();

 }

 void sh4_translate_set_callbacks( xlat_block_begin_callback_t begin, xlat_block_end_callback_t end )

 {

     sh4_x86.begin_callback = begin;

     sh4_x86.end_callback = end;

 }

 void sh4_translate_set_fastmem( gboolean flag )

 {

     sh4_x86.fastmem = flag;

 }

 static void sh4_x86_add_backpatch( uint8_t *fixup_addr, uint32_t fixup_pc, uint32_t exc_code )

 {

     int reloc_size = 4;

     if( exc_code == -2 ) {

         reloc_size = sizeof(void *);

     }

     if( sh4_x86.backpatch_posn == sh4_x86.backpatch_size ) {

 	sh4_x86.backpatch_size <<= 1;

 	sh4_x86.backpatch_list = realloc( sh4_x86.backpatch_list, 

 					  sh4_x86.backpatch_size * sizeof(struct backpatch_record));

 	assert( sh4_x86.backpatch_list != NULL );

     }

     if( sh4_x86.in_delay_slot ) {

 	fixup_pc -= 2;

     }

     sh4_x86.backpatch_list[sh4_x86.backpatch_posn].fixup_offset = 

 	(((uint8_t *)fixup_addr) - ((uint8_t *)xlat_current_block->code)) - reloc_size;

     sh4_x86.backpatch_list[sh4_x86.backpatch_posn].fixup_icount = (fixup_pc - sh4_x86.block_start_pc)>>1;

     sh4_x86.backpatch_list[sh4_x86.backpatch_posn].exc_code = exc_code;

     sh4_x86.backpatch_posn++;

 }

 #define TSTATE_NONE -1

 #define TSTATE_O    X86_COND_O

 #define TSTATE_C    X86_COND_C

 #define TSTATE_E    X86_COND_E

 #define TSTATE_NE   X86_COND_NE

 #define TSTATE_G    X86_COND_G

 #define TSTATE_GE   X86_COND_GE

 #define TSTATE_A    X86_COND_A

 #define TSTATE_AE   X86_COND_AE

 #define MARK_JMP8(x) uint8_t *_mark_jmp_##x = (xlat_output-1)

 #define JMP_TARGET(x) *_mark_jmp_##x += (xlat_output - _mark_jmp_##x)

 /* Convenience instructions */

 #define LDC_t()          CMPB_imms_rbpdisp(1,R_T); CMC()

 #define SETE_t()         SETCCB_cc_rbpdisp(X86_COND_E,R_T)

 #define SETA_t()         SETCCB_cc_rbpdisp(X86_COND_A,R_T)

 #define SETAE_t()        SETCCB_cc_rbpdisp(X86_COND_AE,R_T)

 #define SETG_t()         SETCCB_cc_rbpdisp(X86_COND_G,R_T)

 #define SETGE_t()        SETCCB_cc_rbpdisp(X86_COND_GE,R_T)

 #define SETC_t()         SETCCB_cc_rbpdisp(X86_COND_C,R_T)

 #define SETO_t()         SETCCB_cc_rbpdisp(X86_COND_O,R_T)

 #define SETNE_t()        SETCCB_cc_rbpdisp(X86_COND_NE,R_T)

 #define SETC_r8(r1)      SETCCB_cc_r8(X86_COND_C, r1)

 #define JAE_label(label) JCC_cc_rel8(X86_COND_AE,-1); MARK_JMP8(label)

 #define JBE_label(label) JCC_cc_rel8(X86_COND_BE,-1); MARK_JMP8(label)

 #define JE_label(label)  JCC_cc_rel8(X86_COND_E,-1); MARK_JMP8(label)

 #define JGE_label(label) JCC_cc_rel8(X86_COND_GE,-1); MARK_JMP8(label)

 #define JNA_label(label) JCC_cc_rel8(X86_COND_NA,-1); MARK_JMP8(label)

 #define JNE_label(label) JCC_cc_rel8(X86_COND_NE,-1); MARK_JMP8(label)

 #define JNO_label(label) JCC_cc_rel8(X86_COND_NO,-1); MARK_JMP8(label)

 #define JP_label(label)  JCC_cc_rel8(X86_COND_P,-1); MARK_JMP8(label)

 #define JS_label(label)  JCC_cc_rel8(X86_COND_S,-1); MARK_JMP8(label)

 #define JMP_label(label) JMP_rel8(-1); MARK_JMP8(label)

 #define JNE_exc(exc)     JCC_cc_rel32(X86_COND_NE,0); sh4_x86_add_backpatch(xlat_output, pc, exc)

 #define LOAD_t() if( sh4_x86.tstate == TSTATE_NONE ) { \

 	CMPL_imms_rbpdisp( 1, R_T ); sh4_x86.tstate = TSTATE_E; }     

 /** Branch if T is set (either in the current cflags, or in sh4r.t) */

 #define JT_label(label) LOAD_t() \

     JCC_cc_rel8(sh4_x86.tstate,-1); MARK_JMP8(label)

 /** Branch if T is clear (either in the current cflags or in sh4r.t) */

 #define JF_label(label) LOAD_t() \

     JCC_cc_rel8(sh4_x86.tstate^1, -1); MARK_JMP8(label)

 #define load_reg(x86reg,sh4reg)     MOVL_rbpdisp_r32( REG_OFFSET(r[sh4reg]), x86reg )

 #define store_reg(x86reg,sh4reg)    MOVL_r32_rbpdisp( x86reg, REG_OFFSET(r[sh4reg]) )

 /**

  * Load an FR register (single-precision floating point) into an integer x86

  * register (eg for register-to-register moves)

  */

 #define load_fr(reg,frm)  MOVL_rbpdisp_r32( REG_OFFSET(fr[0][(frm)^1]), reg )

 #define load_xf(reg,frm)  MOVL_rbpdisp_r32( REG_OFFSET(fr[1][(frm)^1]), reg )

 /**

  * Load the low half of a DR register (DR or XD) into an integer x86 register 

  */

 #define load_dr0(reg,frm) MOVL_rbpdisp_r32( REG_OFFSET(fr[frm&1][frm|0x01]), reg )

 #define load_dr1(reg,frm) MOVL_rbpdisp_r32( REG_OFFSET(fr[frm&1][frm&0x0E]), reg )

 /**

  * Store an FR register (single-precision floating point) from an integer x86+

  * register (eg for register-to-register moves)

  */

 #define store_fr(reg,frm) MOVL_r32_rbpdisp( reg, REG_OFFSET(fr[0][(frm)^1]) )

 #define store_xf(reg,frm) MOVL_r32_rbpdisp( reg, REG_OFFSET(fr[1][(frm)^1]) )

 #define store_dr0(reg,frm) MOVL_r32_rbpdisp( reg, REG_OFFSET(fr[frm&1][frm|0x01]) )

 #define store_dr1(reg,frm) MOVL_r32_rbpdisp( reg, REG_OFFSET(fr[frm&1][frm&0x0E]) )

 #define push_fpul()  FLDF_rbpdisp(R_FPUL)

 #define pop_fpul()   FSTPF_rbpdisp(R_FPUL)

 #define push_fr(frm) FLDF_rbpdisp( REG_OFFSET(fr[0][(frm)^1]) )

 #define pop_fr(frm)  FSTPF_rbpdisp( REG_OFFSET(fr[0][(frm)^1]) )

 #define push_xf(frm) FLDF_rbpdisp( REG_OFFSET(fr[1][(frm)^1]) )

 #define pop_xf(frm)  FSTPF_rbpdisp( REG_OFFSET(fr[1][(frm)^1]) )

 #define push_dr(frm) FLDD_rbpdisp( REG_OFFSET(fr[0][(frm)&0x0E]) )

 #define pop_dr(frm)  FSTPD_rbpdisp( REG_OFFSET(fr[0][(frm)&0x0E]) )

 #define push_xdr(frm) FLDD_rbpdisp( REG_OFFSET(fr[1][(frm)&0x0E]) )

 #define pop_xdr(frm)  FSTPD_rbpdisp( REG_OFFSET(fr[1][(frm)&0x0E]) )

 #ifdef ENABLE_SH4STATS

 #define COUNT_INST(id) MOVL_imm32_r32( id, REG_EAX ); CALL1_ptr_r32(sh4_stats_add, REG_EAX); sh4_x86.tstate = TSTATE_NONE

 #else

 #define COUNT_INST(id)

 #endif

 /* Exception checks - Note that all exception checks will clobber EAX */

 #define check_priv( ) \

     if( (sh4_x86.sh4_mode & SR_MD) == 0 ) { \

         if( sh4_x86.in_delay_slot ) { \

             exit_block_exc(EXC_SLOT_ILLEGAL, (pc-2), 4 ); \

         } else { \

             exit_block_exc(EXC_ILLEGAL, pc, 2); \

         } \

         sh4_x86.branch_taken = TRUE; \

         sh4_x86.in_delay_slot = DELAY_NONE; \

         return 2; \

     }

 #define check_fpuen( ) \

     if( !sh4_x86.fpuen_checked ) {\

 	sh4_x86.fpuen_checked = TRUE;\

 	MOVL_rbpdisp_r32( R_SR, REG_EAX );\

 	ANDL_imms_r32( SR_FD, REG_EAX );\

 	if( sh4_x86.in_delay_slot ) {\

 	    JNE_exc(EXC_SLOT_FPU_DISABLED);\

 	} else {\

 	    JNE_exc(EXC_FPU_DISABLED);\

 	}\

 	sh4_x86.tstate = TSTATE_NONE; \

     }

 #define check_ralign16( x86reg ) \

     TESTL_imms_r32( 0x00000001, x86reg ); \

     JNE_exc(EXC_DATA_ADDR_READ)

 #define check_walign16( x86reg ) \

     TESTL_imms_r32( 0x00000001, x86reg ); \

     JNE_exc(EXC_DATA_ADDR_WRITE);

 #define check_ralign32( x86reg ) \

     TESTL_imms_r32( 0x00000003, x86reg ); \

     JNE_exc(EXC_DATA_ADDR_READ)

 #define check_walign32( x86reg ) \

     TESTL_imms_r32( 0x00000003, x86reg ); \

     JNE_exc(EXC_DATA_ADDR_WRITE);

 #define check_ralign64( x86reg ) \

     TESTL_imms_r32( 0x00000007, x86reg ); \

     JNE_exc(EXC_DATA_ADDR_READ)

 #define check_walign64( x86reg ) \

     TESTL_imms_r32( 0x00000007, x86reg ); \

     JNE_exc(EXC_DATA_ADDR_WRITE);

 #define address_space() ((sh4_x86.sh4_mode&SR_MD) ? (uintptr_t)sh4_x86.priv_address_space : (uintptr_t)sh4_x86.user_address_space)

 #define UNDEF(ir)

 /* Note: For SR.MD == 1 && MMUCR.AT == 0, there are no memory exceptions, so 

  * don't waste the cycles expecting them. Otherwise we need to save the exception pointer.

  */

 #ifdef HAVE_FRAME_ADDRESS

 static void call_read_func(int addr_reg, int value_reg, int offset, int pc)

 {

     decode_address(address_space(), addr_reg, REG_CALLPTR);

     if( !sh4_x86.tlb_on && (sh4_x86.sh4_mode & SR_MD) ) { 

         CALL1_r32disp_r32(REG_CALLPTR, offset, addr_reg);

     } else {

         if( addr_reg != REG_ARG1 ) {

             MOVL_r32_r32( addr_reg, REG_ARG1 );

         }

         MOVP_immptr_rptr( 0, REG_ARG2 );

         sh4_x86_add_backpatch( xlat_output, pc, -2 );

         CALL2_r32disp_r32_r32(REG_CALLPTR, offset, REG_ARG1, REG_ARG2);

     }

     if( value_reg != REG_RESULT1 ) { 

         MOVL_r32_r32( REG_RESULT1, value_reg );

     }

 }

 static void call_write_func(int addr_reg, int value_reg, int offset, int pc)

 {

     decode_address(address_space(), addr_reg, REG_CALLPTR);

     if( !sh4_x86.tlb_on && (sh4_x86.sh4_mode & SR_MD) ) { 

         CALL2_r32disp_r32_r32(REG_CALLPTR, offset, addr_reg, value_reg);

     } else {

         if( value_reg != REG_ARG2 ) {

             MOVL_r32_r32( value_reg, REG_ARG2 );

 	}        

         if( addr_reg != REG_ARG1 ) {

             MOVL_r32_r32( addr_reg, REG_ARG1 );

         }

 #if MAX_REG_ARG > 2        

         MOVP_immptr_rptr( 0, REG_ARG3 );

         sh4_x86_add_backpatch( xlat_output, pc, -2 );

         CALL3_r32disp_r32_r32_r32(REG_CALLPTR, offset, REG_ARG1, REG_ARG2, REG_ARG3);

 #else

         MOVL_imm32_rspdisp( 0, 0 );

         sh4_x86_add_backpatch( xlat_output, pc, -2 );

         CALL3_r32disp_r32_r32_r32(REG_CALLPTR, offset, REG_ARG1, REG_ARG2, 0);

 #endif

     }

 }

 #else

 static void call_read_func(int addr_reg, int value_reg, int offset, int pc)

 {

     decode_address(address_space(), addr_reg, REG_CALLPTR);

     CALL1_r32disp_r32(REG_CALLPTR, offset, addr_reg);

     if( value_reg != REG_RESULT1 ) {

         MOVL_r32_r32( REG_RESULT1, value_reg );

     }

 }     

 static void call_write_func(int addr_reg, int value_reg, int offset, int pc)

 {

     decode_address(address_space(), addr_reg, REG_CALLPTR);

     CALL2_r32disp_r32_r32(REG_CALLPTR, offset, addr_reg, value_reg);

 }

 #endif

 #define MEM_REGION_PTR(name) offsetof( struct mem_region_fn, name )

 #define MEM_READ_BYTE( addr_reg, value_reg ) call_read_func(addr_reg, value_reg, MEM_REGION_PTR(read_byte), pc)

 #define MEM_READ_BYTE_FOR_WRITE( addr_reg, value_reg ) call_read_func( addr_reg, value_reg, MEM_REGION_PTR(read_byte_for_write), pc) 

 #define MEM_READ_WORD( addr_reg, value_reg ) call_read_func(addr_reg, value_reg, MEM_REGION_PTR(read_word), pc)

 #define MEM_READ_LONG( addr_reg, value_reg ) call_read_func(addr_reg, value_reg, MEM_REGION_PTR(read_long), pc)

 #define MEM_WRITE_BYTE( addr_reg, value_reg ) call_write_func(addr_reg, value_reg, MEM_REGION_PTR(write_byte), pc)

 #define MEM_WRITE_WORD( addr_reg, value_reg ) call_write_func(addr_reg, value_reg, MEM_REGION_PTR(write_word), pc)

 #define MEM_WRITE_LONG( addr_reg, value_reg ) call_write_func(addr_reg, value_reg, MEM_REGION_PTR(write_long), pc)

 #define MEM_PREFETCH( addr_reg ) call_read_func(addr_reg, REG_RESULT1, MEM_REGION_PTR(prefetch), pc)

 #define SLOTILLEGAL() exit_block_exc(EXC_SLOT_ILLEGAL, pc-2, 4); sh4_x86.in_delay_slot = DELAY_NONE; return 2;

 /** Offset of xlat_sh4_mode field relative to the code pointer */ 

 #define XLAT_SH4_MODE_CODE_OFFSET  (int32_t)(offsetof(struct xlat_cache_block, xlat_sh4_mode) - offsetof(struct xlat_cache_block,code) )

 #define XLAT_CHAIN_CODE_OFFSET (int32_t)(offsetof(struct xlat_cache_block, chain) - offsetof(struct xlat_cache_block,code) )

 #define XLAT_ACTIVE_CODE_OFFSET (int32_t)(offsetof(struct xlat_cache_block, active) - offsetof(struct xlat_cache_block,code) )

 void sh4_translate_begin_block( sh4addr_t pc ) 

 {

 	sh4_x86.code = xlat_output;

     sh4_x86.in_delay_slot = FALSE;

     sh4_x86.fpuen_checked = FALSE;

     sh4_x86.branch_taken = FALSE;

     sh4_x86.backpatch_posn = 0;

     sh4_x86.block_start_pc = pc;

     sh4_x86.tlb_on = IS_TLB_ENABLED();

     sh4_x86.tstate = TSTATE_NONE;

     sh4_x86.double_prec = sh4r.fpscr & FPSCR_PR;

     sh4_x86.double_size = sh4r.fpscr & FPSCR_SZ;

     sh4_x86.sh4_mode = sh4r.xlat_sh4_mode;

     if( sh4_x86.begin_callback ) {

         CALL_ptr( sh4_x86.begin_callback );

     }

     if( sh4_profile_blocks ) {

     	MOVP_immptr_rptr( sh4_x86.code + XLAT_ACTIVE_CODE_OFFSET, REG_EAX );

     	ADDL_imms_r32disp( 1, REG_EAX, 0 );

     }  

 }

 uint32_t sh4_translate_end_block_size()

 {

 	uint32_t epilogue_size = EPILOGUE_SIZE;

 	if( sh4_x86.end_callback ) {

 	    epilogue_size += (CALL1_PTR_MIN_SIZE - 1);

 	}

     if( sh4_x86.backpatch_posn <= 3 ) {

         epilogue_size += (sh4_x86.backpatch_posn*(12+CALL1_PTR_MIN_SIZE));

     } else {

         epilogue_size += (3*(12+CALL1_PTR_MIN_SIZE)) + (sh4_x86.backpatch_posn-3)*(15+CALL1_PTR_MIN_SIZE);

     }

     return epilogue_size;

 }

 /**

  * Embed a breakpoint into the generated code

  */

 void sh4_translate_emit_breakpoint( sh4vma_t pc )

 {

     MOVL_imm32_r32( pc, REG_EAX );

     CALL1_ptr_r32( sh4_translate_breakpoint_hit, REG_EAX );

     sh4_x86.tstate = TSTATE_NONE;

 }

 #define UNTRANSLATABLE(pc) !IS_IN_ICACHE(pc)

 /**

  * Test if the loaded target code pointer in %eax is valid, and if so jump

  * directly into it, bypassing the normal exit.

  */

 static void jump_next_block()

 {

 	uint8_t *ptr = xlat_output;

 	TESTP_rptr_rptr(REG_EAX, REG_EAX);

 	JE_label(nocode);

 	if( sh4_x86.sh4_mode == SH4_MODE_UNKNOWN ) {

 	    /* sr/fpscr was changed, possibly updated xlat_sh4_mode, so reload it */

 	    MOVL_rbpdisp_r32( REG_OFFSET(xlat_sh4_mode), REG_ECX );

 	    CMPL_r32_r32disp( REG_ECX, REG_EAX, XLAT_SH4_MODE_CODE_OFFSET );

 	} else {

 	    CMPL_imms_r32disp( sh4_x86.sh4_mode, REG_EAX, XLAT_SH4_MODE_CODE_OFFSET );

 	}

 	JNE_label(wrongmode);

 	if( sh4_x86.end_callback ) {

 	    /* Note this does leave the stack out of alignment, but doesn't matter

 	     * for what we're currently using it for.

 	     */

 	    PUSH_r32(REG_EAX);

 	    MOVP_immptr_rptr(sh4_x86.end_callback, REG_ECX);

 	    JMP_rptr(REG_ECX);

 	} else {

 	    JMP_rptr(REG_EAX);

 	}

 	JMP_TARGET(wrongmode);

 	MOVP_rptrdisp_rptr( REG_EAX, XLAT_CHAIN_CODE_OFFSET, REG_EAX );

 	int rel = ptr - xlat_output;

     JMP_prerel(rel);

 	JMP_TARGET(nocode); 

 }

 /**

  * 

  */

 void FASTCALL sh4_translate_link_block( uint32_t pc )

 {

     uint8_t *target = (uint8_t *)xlat_get_code_by_vma(pc);

     while( target != NULL && sh4r.xlat_sh4_mode != XLAT_BLOCK_MODE(target) ) {

         target = XLAT_BLOCK_CHAIN(target);

 	}

     if( target == NULL ) {

         target = sh4_translate_basic_block( pc );

     }

     uint8_t *backpatch = ((uint8_t *)__builtin_return_address(0)) - (CALL1_PTR_MIN_SIZE);

     *backpatch = 0xE9;

     *(uint32_t *)(backpatch+1) = (uint32_t)(target-backpatch)-5;

     *(void **)(backpatch+5) = XLAT_BLOCK_FOR_CODE(target)->use_list;

     XLAT_BLOCK_FOR_CODE(target)->use_list = backpatch; 

     uint8_t * volatile *retptr = ((uint8_t * volatile *)__builtin_frame_address(0))+1;

     assert( *retptr == ((uint8_t *)__builtin_return_address(0)) );

 	*retptr = backpatch;

 }

 static void emit_translate_and_backpatch()

 {

     /* NB: this is either 7 bytes (i386) or 12 bytes (x86-64) */

     CALL1_ptr_r32(sh4_translate_link_block, REG_ARG1);

     /* When patched, the jmp instruction will be 5 bytes (either platform) -

      * we need to reserve sizeof(void*) bytes for the use-list

 	 * pointer

 	 */ 

     if( sizeof(void*) == 8 ) {

         NOP();

     } else {

         NOP2();

     }

 }

 /**

  * If we're jumping to a fixed address (or at least fixed relative to the

  * current PC, then we can do a direct branch. REG_ARG1 should contain

  * the PC at this point.

  */

 static void jump_next_block_fixed_pc( sh4addr_t pc )

 {

 	if( IS_IN_ICACHE(pc) ) {

 	    if( sh4_x86.sh4_mode != SH4_MODE_UNKNOWN && sh4_x86.end_callback == NULL ) {

 	        /* Fixed address, in cache, and fixed SH4 mode - generate a call to the

 	         * fetch-and-backpatch routine, which will replace the call with a branch */

            emit_translate_and_backpatch();	         

            return;

 		} else {

             MOVP_moffptr_rax( xlat_get_lut_entry(GET_ICACHE_PHYS(pc)) );

             ANDP_imms_rptr( -4, REG_EAX );

         }

 	} else if( sh4_x86.tlb_on ) {

         CALL1_ptr_r32(xlat_get_code_by_vma, REG_ARG1);

     } else {

         CALL1_ptr_r32(xlat_get_code, REG_ARG1);

     }

     jump_next_block();

 }

 static void sh4_x86_translate_unlink_block( void *use_list )

 {

 	uint8_t *tmp = xlat_output; /* In case something is active, which should never happen */

 	void *next = use_list;

 	while( next != NULL ) {

     	xlat_output = (uint8_t *)next;

  	    next = *(void **)(xlat_output+5);

  		emit_translate_and_backpatch();

  	}

  	xlat_output = tmp;

 }

 static void exit_block()

 {

 	if( sh4_x86.end_callback ) {

 	    MOVP_immptr_rptr(sh4_x86.end_callback, REG_ECX);

 	    JMP_rptr(REG_ECX);

 	} else {

 	    RET();

 	}

 }

 /**

  * Exit the block with sh4r.pc already written

  */

 void exit_block_pcset( sh4addr_t pc )

 {

     MOVL_imm32_r32( ((pc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period, REG_ECX );

     ADDL_rbpdisp_r32( REG_OFFSET(slice_cycle), REG_ECX );

     MOVL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );

     CMPL_r32_rbpdisp( REG_ECX, REG_OFFSET(event_pending) );

     JBE_label(exitloop);

     MOVL_rbpdisp_r32( R_PC, REG_ARG1 );

     if( sh4_x86.tlb_on ) {

         CALL1_ptr_r32(xlat_get_code_by_vma,REG_ARG1);

     } else {

         CALL1_ptr_r32(xlat_get_code,REG_ARG1);

     }

     jump_next_block();

     JMP_TARGET(exitloop);

     exit_block();

 }

 /**

  * Exit the block with sh4r.new_pc written with the target pc

  */

 void exit_block_newpcset( sh4addr_t pc )

 {

     MOVL_imm32_r32( ((pc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period, REG_ECX );

     ADDL_rbpdisp_r32( REG_OFFSET(slice_cycle), REG_ECX );

     MOVL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );

     MOVL_rbpdisp_r32( R_NEW_PC, REG_ARG1 );

     MOVL_r32_rbpdisp( REG_ARG1, R_PC );

     CMPL_r32_rbpdisp( REG_ECX, REG_OFFSET(event_pending) );

     JBE_label(exitloop);

     if( sh4_x86.tlb_on ) {

         CALL1_ptr_r32(xlat_get_code_by_vma,REG_ARG1);

     } else {

         CALL1_ptr_r32(xlat_get_code,REG_ARG1);

     }

 	jump_next_block();

     JMP_TARGET(exitloop);

     exit_block();

 }

 /**

  * Exit the block to an absolute PC

  */

 void exit_block_abs( sh4addr_t pc, sh4addr_t endpc )

 {

     MOVL_imm32_r32( ((endpc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period, REG_ECX );

     ADDL_rbpdisp_r32( REG_OFFSET(slice_cycle), REG_ECX );

     MOVL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );

     MOVL_imm32_r32( pc, REG_ARG1 );

     MOVL_r32_rbpdisp( REG_ARG1, R_PC );

     CMPL_r32_rbpdisp( REG_ECX, REG_OFFSET(event_pending) );

     JBE_label(exitloop);

     jump_next_block_fixed_pc(pc);    

     JMP_TARGET(exitloop);

     exit_block();

 }

 /**

  * Exit the block to a relative PC

  */

 void exit_block_rel( sh4addr_t pc, sh4addr_t endpc )

 {

     MOVL_imm32_r32( ((endpc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period, REG_ECX );

     ADDL_rbpdisp_r32( REG_OFFSET(slice_cycle), REG_ECX );

     MOVL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );

 	if( pc == sh4_x86.block_start_pc && sh4_x86.sh4_mode == sh4r.xlat_sh4_mode ) {

 	    /* Special case for tight loops - the PC doesn't change, and

 	     * we already know the target address. Just check events pending before

 	     * looping.

 	     */

         CMPL_r32_rbpdisp( REG_ECX, REG_OFFSET(event_pending) );

         uint32_t backdisp = ((uintptr_t)(sh4_x86.code - xlat_output));

         JCC_cc_prerel(X86_COND_A, backdisp);

 	} else {

         MOVL_imm32_r32( pc - sh4_x86.block_start_pc, REG_ARG1 );

         ADDL_rbpdisp_r32( R_PC, REG_ARG1 );

         MOVL_r32_rbpdisp( REG_ARG1, R_PC );

         CMPL_r32_rbpdisp( REG_ECX, REG_OFFSET(event_pending) );

         JBE_label(exitloop2);

         jump_next_block_fixed_pc(pc);

         JMP_TARGET(exitloop2);

     }

     exit_block();

 }

 /**

  * Exit unconditionally with a general exception

  */

 void exit_block_exc( int code, sh4addr_t pc, int inst_adjust )

 {

     MOVL_imm32_r32( pc - sh4_x86.block_start_pc, REG_ECX );

     ADDL_r32_rbpdisp( REG_ECX, R_PC );

     MOVL_imm32_r32( ((pc - sh4_x86.block_start_pc + inst_adjust)>>1)*sh4_cpu_period, REG_ECX );

     ADDL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );

     MOVL_imm32_r32( code, REG_ARG1 );

     CALL1_ptr_r32( sh4_raise_exception, REG_ARG1 );

     exit_block();

 }    

 /**

  * Embed a call to sh4_execute_instruction for situations that we

  * can't translate (just page-crossing delay slots at the moment).

  * Caller is responsible for setting new_pc before calling this function.

  *

  * Performs:

  *   Set PC = endpc

  *   Set sh4r.in_delay_slot = sh4_x86.in_delay_slot

  *   Update slice_cycle for endpc+2 (single step doesn't update slice_cycle)

  *   Call sh4_execute_instruction

  *   Call xlat_get_code_by_vma / xlat_get_code as for normal exit

  */

 void exit_block_emu( sh4vma_t endpc )

 {

     MOVL_imm32_r32( endpc - sh4_x86.block_start_pc, REG_ECX );   // 5

     ADDL_r32_rbpdisp( REG_ECX, R_PC );

     MOVL_imm32_r32( (((endpc - sh4_x86.block_start_pc)>>1)+1)*sh4_cpu_period, REG_ECX ); // 5

     ADDL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );     // 6

     MOVL_imm32_r32( sh4_x86.in_delay_slot ? 1 : 0, REG_ECX );

     MOVL_r32_rbpdisp( REG_ECX, REG_OFFSET(in_delay_slot) );

     CALL_ptr( sh4_execute_instruction );

     exit_block();

 } 

 /**

  * Write the block trailer (exception handling block)

  */

 void sh4_translate_end_block( sh4addr_t pc ) {

     if( sh4_x86.branch_taken == FALSE ) {

         // Didn't exit unconditionally already, so write the termination here

         exit_block_rel( pc, pc );

     }

     if( sh4_x86.backpatch_posn != 0 ) {

         unsigned int i;

         // Exception raised - cleanup and exit

         uint8_t *end_ptr = xlat_output;

         MOVL_r32_r32( REG_EDX, REG_ECX );

         ADDL_r32_r32( REG_EDX, REG_ECX );

         ADDL_r32_rbpdisp( REG_ECX, R_SPC );

         MOVL_moffptr_eax( &sh4_cpu_period );

         INC_r32( REG_EDX );  /* Add 1 for the aborting instruction itself */ 

         MULL_r32( REG_EDX );

         ADDL_r32_rbpdisp( REG_EAX, REG_OFFSET(slice_cycle) );

         exit_block();

         for( i=0; i< sh4_x86.backpatch_posn; i++ ) {

             uint32_t *fixup_addr = (uint32_t *)&xlat_current_block->code[sh4_x86.backpatch_list[i].fixup_offset];

             if( sh4_x86.backpatch_list[i].exc_code < 0 ) {

                 if( sh4_x86.backpatch_list[i].exc_code == -2 ) {

                     *((uintptr_t *)fixup_addr) = (uintptr_t)xlat_output; 

                 } else {

                     *fixup_addr += xlat_output - (uint8_t *)&xlat_current_block->code[sh4_x86.backpatch_list[i].fixup_offset] - 4;

                 }

                 MOVL_imm32_r32( sh4_x86.backpatch_list[i].fixup_icount, REG_EDX );

                 int rel = end_ptr - xlat_output;

                 JMP_prerel(rel);

             } else {

                 *fixup_addr += xlat_output - (uint8_t *)&xlat_current_block->code[sh4_x86.backpatch_list[i].fixup_offset] - 4;

                 MOVL_imm32_r32( sh4_x86.backpatch_list[i].exc_code, REG_ARG1 );

                 CALL1_ptr_r32( sh4_raise_exception, REG_ARG1 );

                 MOVL_imm32_r32( sh4_x86.backpatch_list[i].fixup_icount, REG_EDX );

                 int rel = end_ptr - xlat_output;

                 JMP_prerel(rel);

             }

         }

     }

 }

 /**

  * Translate a single instruction. Delayed branches are handled specially

  * by translating both branch and delayed instruction as a single unit (as

  * 

  * The instruction MUST be in the icache (assert check)

  *

  * @return true if the instruction marks the end of a basic block

  * (eg a branch or 

  */

 uint32_t sh4_translate_instruction( sh4vma_t pc )

 {

     uint32_t ir;

     /* Read instruction from icache */

     assert( IS_IN_ICACHE(pc) );

     ir = *(uint16_t *)GET_ICACHE_PTR(pc);

     if( !sh4_x86.in_delay_slot ) {

 	sh4_translate_add_recovery( (pc - sh4_x86.block_start_pc)>>1 );

     }

     /* check for breakpoints at this pc */

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

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

             sh4_translate_emit_breakpoint(pc);

             break;

         }

     }

 %%

 /* ALU operations */

 ADD Rm, Rn {:

     COUNT_INST(I_ADD);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     ADDL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 ADD #imm, Rn {:  

     COUNT_INST(I_ADDI);

     ADDL_imms_rbpdisp( imm, REG_OFFSET(r[Rn]) );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 ADDC Rm, Rn {:

     COUNT_INST(I_ADDC);

     if( sh4_x86.tstate != TSTATE_C ) {

         LDC_t();

     }

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     ADCL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     SETC_t();

     sh4_x86.tstate = TSTATE_C;

 :}

 ADDV Rm, Rn {:

     COUNT_INST(I_ADDV);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     ADDL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     SETO_t();

     sh4_x86.tstate = TSTATE_O;

 :}

 AND Rm, Rn {:

     COUNT_INST(I_AND);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     ANDL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 AND #imm, R0 {:  

     COUNT_INST(I_ANDI);

     load_reg( REG_EAX, 0 );

     ANDL_imms_r32(imm, REG_EAX); 

     store_reg( REG_EAX, 0 );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 AND.B #imm, @(R0, GBR) {: 

     COUNT_INST(I_ANDB);

     load_reg( REG_EAX, 0 );

     ADDL_rbpdisp_r32( R_GBR, REG_EAX );

     MOVL_r32_r32(REG_EAX, REG_SAVE1);

     MEM_READ_BYTE_FOR_WRITE( REG_EAX, REG_EDX );

     MOVL_r32_r32(REG_SAVE1, REG_EAX);

     ANDL_imms_r32(imm, REG_EDX );

     MEM_WRITE_BYTE( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 CMP/EQ Rm, Rn {:  

     COUNT_INST(I_CMPEQ);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     CMPL_r32_r32( REG_EAX, REG_ECX );

     SETE_t();

     sh4_x86.tstate = TSTATE_E;

 :}

 CMP/EQ #imm, R0 {:  

     COUNT_INST(I_CMPEQI);

     load_reg( REG_EAX, 0 );

     CMPL_imms_r32(imm, REG_EAX);

     SETE_t();

     sh4_x86.tstate = TSTATE_E;

 :}

 CMP/GE Rm, Rn {:  

     COUNT_INST(I_CMPGE);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     CMPL_r32_r32( REG_EAX, REG_ECX );

     SETGE_t();

     sh4_x86.tstate = TSTATE_GE;

 :}

 CMP/GT Rm, Rn {: 

     COUNT_INST(I_CMPGT);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     CMPL_r32_r32( REG_EAX, REG_ECX );

     SETG_t();

     sh4_x86.tstate = TSTATE_G;

 :}

 CMP/HI Rm, Rn {:  

     COUNT_INST(I_CMPHI);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     CMPL_r32_r32( REG_EAX, REG_ECX );

     SETA_t();

     sh4_x86.tstate = TSTATE_A;

 :}

 CMP/HS Rm, Rn {: 

     COUNT_INST(I_CMPHS);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     CMPL_r32_r32( REG_EAX, REG_ECX );

     SETAE_t();

     sh4_x86.tstate = TSTATE_AE;

  :}

 CMP/PL Rn {: 

     COUNT_INST(I_CMPPL);

     load_reg( REG_EAX, Rn );

     CMPL_imms_r32( 0, REG_EAX );

     SETG_t();

     sh4_x86.tstate = TSTATE_G;

 :}

 CMP/PZ Rn {:  

     COUNT_INST(I_CMPPZ);

     load_reg( REG_EAX, Rn );

     CMPL_imms_r32( 0, REG_EAX );

     SETGE_t();

     sh4_x86.tstate = TSTATE_GE;

 :}

 CMP/STR Rm, Rn {:  

     COUNT_INST(I_CMPSTR);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     XORL_r32_r32( REG_ECX, REG_EAX );

     TESTB_r8_r8( REG_AL, REG_AL );

     JE_label(target1);

     TESTB_r8_r8( REG_AH, REG_AH );

     JE_label(target2);

     SHRL_imm_r32( 16, REG_EAX );

     TESTB_r8_r8( REG_AL, REG_AL );

     JE_label(target3);

     TESTB_r8_r8( REG_AH, REG_AH );

     JMP_TARGET(target1);

     JMP_TARGET(target2);

     JMP_TARGET(target3);

     SETE_t();

     sh4_x86.tstate = TSTATE_E;

 :}

 DIV0S Rm, Rn {:

     COUNT_INST(I_DIV0S);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     SHRL_imm_r32( 31, REG_EAX );

     SHRL_imm_r32( 31, REG_ECX );

     MOVL_r32_rbpdisp( REG_EAX, R_M );

     MOVL_r32_rbpdisp( REG_ECX, R_Q );

     CMPL_r32_r32( REG_EAX, REG_ECX );

     SETNE_t();

     sh4_x86.tstate = TSTATE_NE;

 :}

 DIV0U {:  

     COUNT_INST(I_DIV0U);

     XORL_r32_r32( REG_EAX, REG_EAX );

     MOVL_r32_rbpdisp( REG_EAX, R_Q );

     MOVL_r32_rbpdisp( REG_EAX, R_M );

     MOVL_r32_rbpdisp( REG_EAX, R_T );

     sh4_x86.tstate = TSTATE_C; // works for DIV1

 :}

 DIV1 Rm, Rn {:

     COUNT_INST(I_DIV1);

     MOVL_rbpdisp_r32( R_M, REG_ECX );

     load_reg( REG_EAX, Rn );

     if( sh4_x86.tstate != TSTATE_C ) {

 	LDC_t();

     }

     RCLL_imm_r32( 1, REG_EAX );

     SETC_r8( REG_DL ); // Q'

     CMPL_rbpdisp_r32( R_Q, REG_ECX );

     JE_label(mqequal);

     ADDL_rbpdisp_r32( REG_OFFSET(r[Rm]), REG_EAX );

     JMP_label(end);

     JMP_TARGET(mqequal);

     SUBL_rbpdisp_r32( REG_OFFSET(r[Rm]), REG_EAX );

     JMP_TARGET(end);

     store_reg( REG_EAX, Rn ); // Done with Rn now

     SETC_r8(REG_AL); // tmp1

     XORB_r8_r8( REG_DL, REG_AL ); // Q' = Q ^ tmp1

     XORB_r8_r8( REG_AL, REG_CL ); // Q'' = Q' ^ M

     MOVL_r32_rbpdisp( REG_ECX, R_Q );

     XORL_imms_r32( 1, REG_AL );   // T = !Q'

     MOVZXL_r8_r32( REG_AL, REG_EAX );

     MOVL_r32_rbpdisp( REG_EAX, R_T );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 DMULS.L Rm, Rn {:  

     COUNT_INST(I_DMULS);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     IMULL_r32(REG_ECX);

     MOVL_r32_rbpdisp( REG_EDX, R_MACH );

     MOVL_r32_rbpdisp( REG_EAX, R_MACL );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 DMULU.L Rm, Rn {:  

     COUNT_INST(I_DMULU);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     MULL_r32(REG_ECX);

     MOVL_r32_rbpdisp( REG_EDX, R_MACH );

     MOVL_r32_rbpdisp( REG_EAX, R_MACL );    

     sh4_x86.tstate = TSTATE_NONE;

 :}

 DT Rn {:  

     COUNT_INST(I_DT);

     load_reg( REG_EAX, Rn );

     ADDL_imms_r32( -1, REG_EAX );

     store_reg( REG_EAX, Rn );

     SETE_t();

     sh4_x86.tstate = TSTATE_E;

 :}

 EXTS.B Rm, Rn {:  

     COUNT_INST(I_EXTSB);

     load_reg( REG_EAX, Rm );

     MOVSXL_r8_r32( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

 :}

 EXTS.W Rm, Rn {:  

     COUNT_INST(I_EXTSW);

     load_reg( REG_EAX, Rm );

     MOVSXL_r16_r32( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

 :}

 EXTU.B Rm, Rn {:  

     COUNT_INST(I_EXTUB);

     load_reg( REG_EAX, Rm );

     MOVZXL_r8_r32( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

 :}

 EXTU.W Rm, Rn {:  

     COUNT_INST(I_EXTUW);

     load_reg( REG_EAX, Rm );

     MOVZXL_r16_r32( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

 :}

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

     COUNT_INST(I_MACL);

     if( Rm == Rn ) {

 	load_reg( REG_EAX, Rm );

 	check_ralign32( REG_EAX );

 	MEM_READ_LONG( REG_EAX, REG_EAX );

 	MOVL_r32_r32(REG_EAX, REG_SAVE1);

 	load_reg( REG_EAX, Rm );

 	LEAL_r32disp_r32( REG_EAX, 4, REG_EAX );

 	MEM_READ_LONG( REG_EAX, REG_EAX );

         ADDL_imms_rbpdisp( 8, REG_OFFSET(r[Rn]) );

     } else {

 	load_reg( REG_EAX, Rm );

 	check_ralign32( REG_EAX );

 	MEM_READ_LONG( REG_EAX, REG_EAX );

 	MOVL_r32_r32(REG_EAX, REG_SAVE1);

 	load_reg( REG_EAX, Rn );

 	check_ralign32( REG_EAX );

 	MEM_READ_LONG( REG_EAX, REG_EAX );

 	ADDL_imms_rbpdisp( 4, REG_OFFSET(r[Rn]) );

 	ADDL_imms_rbpdisp( 4, REG_OFFSET(r[Rm]) );

     }

     IMULL_r32( REG_SAVE1 );

     ADDL_r32_rbpdisp( REG_EAX, R_MACL );

     ADCL_r32_rbpdisp( REG_EDX, R_MACH );

     MOVL_rbpdisp_r32( R_S, REG_ECX );

     TESTL_r32_r32(REG_ECX, REG_ECX);

     JE_label( nosat );

     CALL_ptr( signsat48 );

     JMP_TARGET( nosat );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MACW);

     if( Rm == Rn ) {

 	load_reg( REG_EAX, Rm );

 	check_ralign16( REG_EAX );

 	MEM_READ_WORD( REG_EAX, REG_EAX );

         MOVL_r32_r32( REG_EAX, REG_SAVE1 );

 	load_reg( REG_EAX, Rm );

 	LEAL_r32disp_r32( REG_EAX, 2, REG_EAX );

 	MEM_READ_WORD( REG_EAX, REG_EAX );

 	ADDL_imms_rbpdisp( 4, REG_OFFSET(r[Rn]) );

 	// Note translate twice in case of page boundaries. Maybe worth

 	// adding a page-boundary check to skip the second translation

     } else {

 	load_reg( REG_EAX, Rn );

 	check_ralign16( REG_EAX );

 	MEM_READ_WORD( REG_EAX, REG_EAX );

         MOVL_r32_r32( REG_EAX, REG_SAVE1 );

 	load_reg( REG_EAX, Rm );

 	check_ralign16( REG_EAX );

 	MEM_READ_WORD( REG_EAX, REG_EAX );

 	ADDL_imms_rbpdisp( 2, REG_OFFSET(r[Rn]) );

 	ADDL_imms_rbpdisp( 2, REG_OFFSET(r[Rm]) );

     }

     IMULL_r32( REG_SAVE1 );

     MOVL_rbpdisp_r32( R_S, REG_ECX );

     TESTL_r32_r32( REG_ECX, REG_ECX );

     JE_label( nosat );

     ADDL_r32_rbpdisp( REG_EAX, R_MACL );  // 6

     JNO_label( end );            // 2

     MOVL_imm32_r32( 1, REG_EDX );         // 5

     MOVL_r32_rbpdisp( REG_EDX, R_MACH );   // 6

     JS_label( positive );        // 2

     MOVL_imm32_r32( 0x80000000, REG_EAX );// 5

     MOVL_r32_rbpdisp( REG_EAX, R_MACL );   // 6

     JMP_label(end2);           // 2

     JMP_TARGET(positive);

     MOVL_imm32_r32( 0x7FFFFFFF, REG_EAX );// 5

     MOVL_r32_rbpdisp( REG_EAX, R_MACL );   // 6

     JMP_label(end3);            // 2

     JMP_TARGET(nosat);

     ADDL_r32_rbpdisp( REG_EAX, R_MACL );  // 6

     ADCL_r32_rbpdisp( REG_EDX, R_MACH );  // 6

     JMP_TARGET(end);

     JMP_TARGET(end2);

     JMP_TARGET(end3);

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOVT Rn {:  

     COUNT_INST(I_MOVT);

     MOVL_rbpdisp_r32( R_T, REG_EAX );

     store_reg( REG_EAX, Rn );

 :}

 MUL.L Rm, Rn {:  

     COUNT_INST(I_MULL);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     MULL_r32( REG_ECX );

     MOVL_r32_rbpdisp( REG_EAX, R_MACL );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MULS.W Rm, Rn {:

     COUNT_INST(I_MULSW);

     MOVSXL_rbpdisp16_r32( R_R(Rm), REG_EAX );

     MOVSXL_rbpdisp16_r32( R_R(Rn), REG_ECX );

     MULL_r32( REG_ECX );

     MOVL_r32_rbpdisp( REG_EAX, R_MACL );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MULU.W Rm, Rn {:  

     COUNT_INST(I_MULUW);

     MOVZXL_rbpdisp16_r32( R_R(Rm), REG_EAX );

     MOVZXL_rbpdisp16_r32( R_R(Rn), REG_ECX );

     MULL_r32( REG_ECX );

     MOVL_r32_rbpdisp( REG_EAX, R_MACL );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 NEG Rm, Rn {:

     COUNT_INST(I_NEG);

     load_reg( REG_EAX, Rm );

     NEGL_r32( REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 NEGC Rm, Rn {:  

     COUNT_INST(I_NEGC);

     load_reg( REG_EAX, Rm );

     XORL_r32_r32( REG_ECX, REG_ECX );

     LDC_t();

     SBBL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     SETC_t();

     sh4_x86.tstate = TSTATE_C;

 :}

 NOT Rm, Rn {:  

     COUNT_INST(I_NOT);

     load_reg( REG_EAX, Rm );

     NOTL_r32( REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 OR Rm, Rn {:  

     COUNT_INST(I_OR);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     ORL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 OR #imm, R0 {:

     COUNT_INST(I_ORI);

     load_reg( REG_EAX, 0 );

     ORL_imms_r32(imm, REG_EAX);

     store_reg( REG_EAX, 0 );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 OR.B #imm, @(R0, GBR) {:  

     COUNT_INST(I_ORB);

     load_reg( REG_EAX, 0 );

     ADDL_rbpdisp_r32( R_GBR, REG_EAX );

     MOVL_r32_r32( REG_EAX, REG_SAVE1 );

     MEM_READ_BYTE_FOR_WRITE( REG_EAX, REG_EDX );

     MOVL_r32_r32( REG_SAVE1, REG_EAX );

     ORL_imms_r32(imm, REG_EDX );

     MEM_WRITE_BYTE( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 ROTCL Rn {:

     COUNT_INST(I_ROTCL);

     load_reg( REG_EAX, Rn );

     if( sh4_x86.tstate != TSTATE_C ) {

 	LDC_t();

     }

     RCLL_imm_r32( 1, REG_EAX );

     store_reg( REG_EAX, Rn );

     SETC_t();

     sh4_x86.tstate = TSTATE_C;

 :}

 ROTCR Rn {:  

     COUNT_INST(I_ROTCR);

     load_reg( REG_EAX, Rn );

     if( sh4_x86.tstate != TSTATE_C ) {

 	LDC_t();

     }

     RCRL_imm_r32( 1, REG_EAX );

     store_reg( REG_EAX, Rn );

     SETC_t();

     sh4_x86.tstate = TSTATE_C;

 :}

 ROTL Rn {:  

     COUNT_INST(I_ROTL);

     load_reg( REG_EAX, Rn );

     ROLL_imm_r32( 1, REG_EAX );

     store_reg( REG_EAX, Rn );

     SETC_t();

     sh4_x86.tstate = TSTATE_C;

 :}

 ROTR Rn {:  

     COUNT_INST(I_ROTR);

     load_reg( REG_EAX, Rn );

     RORL_imm_r32( 1, REG_EAX );

     store_reg( REG_EAX, Rn );

     SETC_t();

     sh4_x86.tstate = TSTATE_C;

 :}

 SHAD Rm, Rn {:

     COUNT_INST(I_SHAD);

     /* Annoyingly enough, not directly convertible */

     load_reg( REG_EAX, Rn );

     load_reg( REG_ECX, Rm );

     CMPL_imms_r32( 0, REG_ECX );

     JGE_label(doshl);

     NEGL_r32( REG_ECX );      // 2

     ANDB_imms_r8( 0x1F, REG_CL ); // 3

     JE_label(emptysar);     // 2

     SARL_cl_r32( REG_EAX );       // 2

     JMP_label(end);          // 2

     JMP_TARGET(emptysar);

     SARL_imm_r32(31, REG_EAX );  // 3

     JMP_label(end2);

     JMP_TARGET(doshl);

     ANDB_imms_r8( 0x1F, REG_CL ); // 3

     SHLL_cl_r32( REG_EAX );       // 2

     JMP_TARGET(end);

     JMP_TARGET(end2);

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SHLD Rm, Rn {:  

     COUNT_INST(I_SHLD);

     load_reg( REG_EAX, Rn );

     load_reg( REG_ECX, Rm );

     CMPL_imms_r32( 0, REG_ECX );

     JGE_label(doshl);

     NEGL_r32( REG_ECX );      // 2

     ANDB_imms_r8( 0x1F, REG_CL ); // 3

     JE_label(emptyshr );

     SHRL_cl_r32( REG_EAX );       // 2

     JMP_label(end);          // 2

     JMP_TARGET(emptyshr);

     XORL_r32_r32( REG_EAX, REG_EAX );

     JMP_label(end2);

     JMP_TARGET(doshl);

     ANDB_imms_r8( 0x1F, REG_CL ); // 3

     SHLL_cl_r32( REG_EAX );       // 2

     JMP_TARGET(end);

     JMP_TARGET(end2);

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SHAL Rn {: 

     COUNT_INST(I_SHAL);

     load_reg( REG_EAX, Rn );

     SHLL_imm_r32( 1, REG_EAX );

     SETC_t();

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_C;

 :}

 SHAR Rn {:  

     COUNT_INST(I_SHAR);

     load_reg( REG_EAX, Rn );

     SARL_imm_r32( 1, REG_EAX );

     SETC_t();

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_C;

 :}

 SHLL Rn {:  

     COUNT_INST(I_SHLL);

     load_reg( REG_EAX, Rn );

     SHLL_imm_r32( 1, REG_EAX );

     SETC_t();

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_C;

 :}

 SHLL2 Rn {:

     COUNT_INST(I_SHLL);

     load_reg( REG_EAX, Rn );

     SHLL_imm_r32( 2, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SHLL8 Rn {:  

     COUNT_INST(I_SHLL);

     load_reg( REG_EAX, Rn );

     SHLL_imm_r32( 8, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SHLL16 Rn {:  

     COUNT_INST(I_SHLL);

     load_reg( REG_EAX, Rn );

     SHLL_imm_r32( 16, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SHLR Rn {:  

     COUNT_INST(I_SHLR);

     load_reg( REG_EAX, Rn );

     SHRL_imm_r32( 1, REG_EAX );

     SETC_t();

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_C;

 :}

 SHLR2 Rn {:  

     COUNT_INST(I_SHLR);

     load_reg( REG_EAX, Rn );

     SHRL_imm_r32( 2, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SHLR8 Rn {:  

     COUNT_INST(I_SHLR);

     load_reg( REG_EAX, Rn );

     SHRL_imm_r32( 8, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SHLR16 Rn {:  

     COUNT_INST(I_SHLR);

     load_reg( REG_EAX, Rn );

     SHRL_imm_r32( 16, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SUB Rm, Rn {:  

     COUNT_INST(I_SUB);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     SUBL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SUBC Rm, Rn {:  

     COUNT_INST(I_SUBC);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     if( sh4_x86.tstate != TSTATE_C ) {

 	LDC_t();

     }

     SBBL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     SETC_t();

     sh4_x86.tstate = TSTATE_C;

 :}

 SUBV Rm, Rn {:  

     COUNT_INST(I_SUBV);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     SUBL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     SETO_t();

     sh4_x86.tstate = TSTATE_O;

 :}

 SWAP.B Rm, Rn {:  

     COUNT_INST(I_SWAPB);

     load_reg( REG_EAX, Rm );

     XCHGB_r8_r8( REG_AL, REG_AH ); // NB: does not touch EFLAGS

     store_reg( REG_EAX, Rn );

 :}

 SWAP.W Rm, Rn {:  

     COUNT_INST(I_SWAPB);

     load_reg( REG_EAX, Rm );

     MOVL_r32_r32( REG_EAX, REG_ECX );

     SHLL_imm_r32( 16, REG_ECX );

     SHRL_imm_r32( 16, REG_EAX );

     ORL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 TAS.B @Rn {:  

     COUNT_INST(I_TASB);

     load_reg( REG_EAX, Rn );

     MOVL_r32_r32( REG_EAX, REG_SAVE1 );

     MEM_READ_BYTE_FOR_WRITE( REG_EAX, REG_EDX );

     TESTB_r8_r8( REG_DL, REG_DL );

     SETE_t();

     ORB_imms_r8( 0x80, REG_DL );

     MOVL_r32_r32( REG_SAVE1, REG_EAX );

     MEM_WRITE_BYTE( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 TST Rm, Rn {:  

     COUNT_INST(I_TST);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     TESTL_r32_r32( REG_EAX, REG_ECX );

     SETE_t();

     sh4_x86.tstate = TSTATE_E;

 :}

 TST #imm, R0 {:  

     COUNT_INST(I_TSTI);

     load_reg( REG_EAX, 0 );

     TESTL_imms_r32( imm, REG_EAX );

     SETE_t();

     sh4_x86.tstate = TSTATE_E;

 :}

 TST.B #imm, @(R0, GBR) {:  

     COUNT_INST(I_TSTB);

     load_reg( REG_EAX, 0);

     ADDL_rbpdisp_r32( R_GBR, REG_EAX );

     MEM_READ_BYTE( REG_EAX, REG_EAX );

     TESTB_imms_r8( imm, REG_AL );

     SETE_t();

     sh4_x86.tstate = TSTATE_E;

 :}

 XOR Rm, Rn {:  

     COUNT_INST(I_XOR);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     XORL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 XOR #imm, R0 {:  

     COUNT_INST(I_XORI);

     load_reg( REG_EAX, 0 );

     XORL_imms_r32( imm, REG_EAX );

     store_reg( REG_EAX, 0 );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 XOR.B #imm, @(R0, GBR) {:  

     COUNT_INST(I_XORB);

     load_reg( REG_EAX, 0 );

     ADDL_rbpdisp_r32( R_GBR, REG_EAX ); 

     MOVL_r32_r32( REG_EAX, REG_SAVE1 );

     MEM_READ_BYTE_FOR_WRITE(REG_EAX, REG_EDX);

     MOVL_r32_r32( REG_SAVE1, REG_EAX );

     XORL_imms_r32( imm, REG_EDX );

     MEM_WRITE_BYTE( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 XTRCT Rm, Rn {:

     COUNT_INST(I_XTRCT);

     load_reg( REG_EAX, Rm );

     load_reg( REG_ECX, Rn );

     SHLL_imm_r32( 16, REG_EAX );

     SHRL_imm_r32( 16, REG_ECX );

     ORL_r32_r32( REG_EAX, REG_ECX );

     store_reg( REG_ECX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 /* Data move instructions */

 MOV Rm, Rn {:  

     COUNT_INST(I_MOV);

     load_reg( REG_EAX, Rm );

     store_reg( REG_EAX, Rn );

 :}

 MOV #imm, Rn {:  

     COUNT_INST(I_MOVI);

     MOVL_imm32_r32( imm, REG_EAX );

     store_reg( REG_EAX, Rn );

 :}

 MOV.B Rm, @Rn {:  

     COUNT_INST(I_MOVB);

     load_reg( REG_EAX, Rn );

     load_reg( REG_EDX, Rm );

     MEM_WRITE_BYTE( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.B Rm, @-Rn {:  

     COUNT_INST(I_MOVB);

     load_reg( REG_EAX, Rn );

     LEAL_r32disp_r32( REG_EAX, -1, REG_EAX );

     load_reg( REG_EDX, Rm );

     MEM_WRITE_BYTE( REG_EAX, REG_EDX );

     ADDL_imms_rbpdisp( -1, REG_OFFSET(r[Rn]) );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVB);

     load_reg( REG_EAX, 0 );

     ADDL_rbpdisp_r32( REG_OFFSET(r[Rn]), REG_EAX );

     load_reg( REG_EDX, Rm );

     MEM_WRITE_BYTE( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVB);

     MOVL_rbpdisp_r32( R_GBR, REG_EAX );

     ADDL_imms_r32( disp, REG_EAX );

     load_reg( REG_EDX, 0 );

     MEM_WRITE_BYTE( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVB);

     load_reg( REG_EAX, Rn );

     ADDL_imms_r32( disp, REG_EAX );

     load_reg( REG_EDX, 0 );

     MEM_WRITE_BYTE( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.B @Rm, Rn {:  

     COUNT_INST(I_MOVB);

     load_reg( REG_EAX, Rm );

     MEM_READ_BYTE( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.B @Rm+, Rn {:  

     COUNT_INST(I_MOVB);

     load_reg( REG_EAX, Rm );

     MEM_READ_BYTE( REG_EAX, REG_EAX );

     if( Rm != Rn ) {

     	ADDL_imms_rbpdisp( 1, REG_OFFSET(r[Rm]) );

     }

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVB);

     load_reg( REG_EAX, 0 );

     ADDL_rbpdisp_r32( REG_OFFSET(r[Rm]), REG_EAX );

     MEM_READ_BYTE( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVB);

     MOVL_rbpdisp_r32( R_GBR, REG_EAX );

     ADDL_imms_r32( disp, REG_EAX );

     MEM_READ_BYTE( REG_EAX, REG_EAX );

     store_reg( REG_EAX, 0 );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVB);

     load_reg( REG_EAX, Rm );

     ADDL_imms_r32( disp, REG_EAX );

     MEM_READ_BYTE( REG_EAX, REG_EAX );

     store_reg( REG_EAX, 0 );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.L Rm, @Rn {:

     COUNT_INST(I_MOVL);

     load_reg( REG_EAX, Rn );

     check_walign32(REG_EAX);

     MOVL_r32_r32( REG_EAX, REG_ECX );

     ANDL_imms_r32( 0xFC000000, REG_ECX );

     CMPL_imms_r32( 0xE0000000, REG_ECX );

     JNE_label( notsq );

     ANDL_imms_r32( 0x3C, REG_EAX );

     load_reg( REG_EDX, Rm );

     MOVL_r32_sib( REG_EDX, 0, REG_EBP, REG_EAX, REG_OFFSET(store_queue) );

     JMP_label(end);

     JMP_TARGET(notsq);

     load_reg( REG_EDX, Rm );

     MEM_WRITE_LONG( REG_EAX, REG_EDX );

     JMP_TARGET(end);

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.L Rm, @-Rn {:  

     COUNT_INST(I_MOVL);

     load_reg( REG_EAX, Rn );

     ADDL_imms_r32( -4, REG_EAX );

     check_walign32( REG_EAX );

     load_reg( REG_EDX, Rm );

     MEM_WRITE_LONG( REG_EAX, REG_EDX );

     ADDL_imms_rbpdisp( -4, REG_OFFSET(r[Rn]) );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVL);

     load_reg( REG_EAX, 0 );

     ADDL_rbpdisp_r32( REG_OFFSET(r[Rn]), REG_EAX );

     check_walign32( REG_EAX );

     load_reg( REG_EDX, Rm );

     MEM_WRITE_LONG( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVL);

     MOVL_rbpdisp_r32( R_GBR, REG_EAX );

     ADDL_imms_r32( disp, REG_EAX );

     check_walign32( REG_EAX );

     load_reg( REG_EDX, 0 );

     MEM_WRITE_LONG( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVL);

     load_reg( REG_EAX, Rn );

     ADDL_imms_r32( disp, REG_EAX );

     check_walign32( REG_EAX );

     MOVL_r32_r32( REG_EAX, REG_ECX );

     ANDL_imms_r32( 0xFC000000, REG_ECX );

     CMPL_imms_r32( 0xE0000000, REG_ECX );

     JNE_label( notsq );

     ANDL_imms_r32( 0x3C, REG_EAX );

     load_reg( REG_EDX, Rm );

     MOVL_r32_sib( REG_EDX, 0, REG_EBP, REG_EAX, REG_OFFSET(store_queue) );

     JMP_label(end);

     JMP_TARGET(notsq);

     load_reg( REG_EDX, Rm );

     MEM_WRITE_LONG( REG_EAX, REG_EDX );

     JMP_TARGET(end);

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.L @Rm, Rn {:  

     COUNT_INST(I_MOVL);

     load_reg( REG_EAX, Rm );

     check_ralign32( REG_EAX );

     MEM_READ_LONG( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.L @Rm+, Rn {:  

     COUNT_INST(I_MOVL);

     load_reg( REG_EAX, Rm );

     check_ralign32( REG_EAX );

     MEM_READ_LONG( REG_EAX, REG_EAX );

     if( Rm != Rn ) {

     	ADDL_imms_rbpdisp( 4, REG_OFFSET(r[Rm]) );

     }

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVL);

     load_reg( REG_EAX, 0 );

     ADDL_rbpdisp_r32( REG_OFFSET(r[Rm]), REG_EAX );

     check_ralign32( REG_EAX );

     MEM_READ_LONG( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVL);

     MOVL_rbpdisp_r32( R_GBR, REG_EAX );

     ADDL_imms_r32( disp, REG_EAX );

     check_ralign32( REG_EAX );

     MEM_READ_LONG( REG_EAX, REG_EAX );

     store_reg( REG_EAX, 0 );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVLPC);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	uint32_t target = (pc & 0xFFFFFFFC) + disp + 4;

 	if( sh4_x86.fastmem && IS_IN_ICACHE(target) ) {

 	    // If the target address is in the same page as the code, it's

 	    // pretty safe to just ref it directly and circumvent the whole

 	    // memory subsystem. (this is a big performance win)

 	    // FIXME: There's a corner-case that's not handled here when

 	    // the current code-page is in the ITLB but not in the UTLB.

 	    // (should generate a TLB miss although need to test SH4 

 	    // behaviour to confirm) Unlikely to be anyone depending on this

 	    // behaviour though.

 	    sh4ptr_t ptr = GET_ICACHE_PTR(target);

 	    MOVL_moffptr_eax( ptr );

 	} else {

 	    // Note: we use sh4r.pc for the calc as we could be running at a

 	    // different virtual address than the translation was done with,

 	    // but we can safely assume that the low bits are the same.

 	    MOVL_imm32_r32( (pc-sh4_x86.block_start_pc) + disp + 4 - (pc&0x03), REG_EAX );

 	    ADDL_rbpdisp_r32( R_PC, REG_EAX );

 	    MEM_READ_LONG( REG_EAX, REG_EAX );

 	    sh4_x86.tstate = TSTATE_NONE;

 	}

 	store_reg( REG_EAX, Rn );

     }

 :}

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

     COUNT_INST(I_MOVL);

     load_reg( REG_EAX, Rm );

     ADDL_imms_r32( disp, REG_EAX );

     check_ralign32( REG_EAX );

     MEM_READ_LONG( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.W Rm, @Rn {:  

     COUNT_INST(I_MOVW);

     load_reg( REG_EAX, Rn );

     check_walign16( REG_EAX );

     load_reg( REG_EDX, Rm );

     MEM_WRITE_WORD( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.W Rm, @-Rn {:  

     COUNT_INST(I_MOVW);

     load_reg( REG_EAX, Rn );

     check_walign16( REG_EAX );

     LEAL_r32disp_r32( REG_EAX, -2, REG_EAX );

     load_reg( REG_EDX, Rm );

     MEM_WRITE_WORD( REG_EAX, REG_EDX );

     ADDL_imms_rbpdisp( -2, REG_OFFSET(r[Rn]) );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVW);

     load_reg( REG_EAX, 0 );

     ADDL_rbpdisp_r32( REG_OFFSET(r[Rn]), REG_EAX );

     check_walign16( REG_EAX );

     load_reg( REG_EDX, Rm );

     MEM_WRITE_WORD( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVW);

     MOVL_rbpdisp_r32( R_GBR, REG_EAX );

     ADDL_imms_r32( disp, REG_EAX );

     check_walign16( REG_EAX );

     load_reg( REG_EDX, 0 );

     MEM_WRITE_WORD( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVW);

     load_reg( REG_EAX, Rn );

     ADDL_imms_r32( disp, REG_EAX );

     check_walign16( REG_EAX );

     load_reg( REG_EDX, 0 );

     MEM_WRITE_WORD( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.W @Rm, Rn {:  

     COUNT_INST(I_MOVW);

     load_reg( REG_EAX, Rm );

     check_ralign16( REG_EAX );

     MEM_READ_WORD( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOV.W @Rm+, Rn {:  

     COUNT_INST(I_MOVW);

     load_reg( REG_EAX, Rm );

     check_ralign16( REG_EAX );

     MEM_READ_WORD( REG_EAX, REG_EAX );

     if( Rm != Rn ) {

         ADDL_imms_rbpdisp( 2, REG_OFFSET(r[Rm]) );

     }

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVW);

     load_reg( REG_EAX, 0 );

     ADDL_rbpdisp_r32( REG_OFFSET(r[Rm]), REG_EAX );

     check_ralign16( REG_EAX );

     MEM_READ_WORD( REG_EAX, REG_EAX );

     store_reg( REG_EAX, Rn );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVW);

     MOVL_rbpdisp_r32( R_GBR, REG_EAX );

     ADDL_imms_r32( disp, REG_EAX );

     check_ralign16( REG_EAX );

     MEM_READ_WORD( REG_EAX, REG_EAX );

     store_reg( REG_EAX, 0 );

     sh4_x86.tstate = TSTATE_NONE;

 :}

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

     COUNT_INST(I_MOVW);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	// See comments for MOV.L @(disp, PC), Rn

 	uint32_t target = pc + disp + 4;

 	if( sh4_x86.fastmem && IS_IN_ICACHE(target) ) {

 	    sh4ptr_t ptr = GET_ICACHE_PTR(target);

 	    MOVL_moffptr_eax( ptr );

 	    MOVSXL_r16_r32( REG_EAX, REG_EAX );

 	} else {

 	    MOVL_imm32_r32( (pc - sh4_x86.block_start_pc) + disp + 4, REG_EAX );

 	    ADDL_rbpdisp_r32( R_PC, REG_EAX );

 	    MEM_READ_WORD( REG_EAX, REG_EAX );

 	    sh4_x86.tstate = TSTATE_NONE;

 	}

 	store_reg( REG_EAX, Rn );

     }

 :}

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

     COUNT_INST(I_MOVW);

     load_reg( REG_EAX, Rm );

     ADDL_imms_r32( disp, REG_EAX );

     check_ralign16( REG_EAX );

     MEM_READ_WORD( REG_EAX, REG_EAX );

     store_reg( REG_EAX, 0 );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 MOVA @(disp, PC), R0 {:  

     COUNT_INST(I_MOVA);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	MOVL_imm32_r32( (pc - sh4_x86.block_start_pc) + disp + 4 - (pc&0x03), REG_ECX );

 	ADDL_rbpdisp_r32( R_PC, REG_ECX );

 	store_reg( REG_ECX, 0 );

 	sh4_x86.tstate = TSTATE_NONE;

     }

 :}

 MOVCA.L R0, @Rn {:  

     COUNT_INST(I_MOVCA);

     load_reg( REG_EAX, Rn );

     check_walign32( REG_EAX );

     load_reg( REG_EDX, 0 );

     MEM_WRITE_LONG( REG_EAX, REG_EDX );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 /* Control transfer instructions */

 BF disp {:

     COUNT_INST(I_BF);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	sh4vma_t target = disp + pc + 4;

 	JT_label( nottaken );

 	exit_block_rel(target, pc+2 );

 	JMP_TARGET(nottaken);

 	return 2;

     }

 :}

 BF/S disp {:

     COUNT_INST(I_BFS);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	sh4_x86.in_delay_slot = DELAY_PC;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    MOVL_imm32_r32( pc + 4 - sh4_x86.block_start_pc, REG_EAX );

 	    JT_label(nottaken);

 	    ADDL_imms_r32( disp, REG_EAX );

 	    JMP_TARGET(nottaken);

 	    ADDL_rbpdisp_r32( R_PC, REG_EAX );

 	    MOVL_r32_rbpdisp( REG_EAX, R_NEW_PC );

 	    exit_block_emu(pc+2);

 	    sh4_x86.branch_taken = TRUE;

 	    return 2;

 	} else {

 	    LOAD_t();

 	    sh4vma_t target = disp + pc + 4;

 	    JCC_cc_rel32(sh4_x86.tstate,0);

 	    uint32_t *patch = ((uint32_t *)xlat_output)-1;

 	    int save_tstate = sh4_x86.tstate;

 	    sh4_translate_instruction(pc+2);

             sh4_x86.in_delay_slot = DELAY_PC; /* Cleared by sh4_translate_instruction */

 	    exit_block_rel( target, pc+4 );

 	    // not taken

 	    *patch = (xlat_output - ((uint8_t *)patch)) - 4;

 	    sh4_x86.tstate = save_tstate;

 	    sh4_translate_instruction(pc+2);

 	    return 4;

 	}

     }

 :}

 BRA disp {:  

     COUNT_INST(I_BRA);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	sh4_x86.in_delay_slot = DELAY_PC;

 	sh4_x86.branch_taken = TRUE;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    MOVL_rbpdisp_r32( R_PC, REG_EAX );

 	    ADDL_imms_r32( pc + disp + 4 - sh4_x86.block_start_pc, REG_EAX );

 	    MOVL_r32_rbpdisp( REG_EAX, R_NEW_PC );

 	    exit_block_emu(pc+2);

 	    return 2;

 	} else {

 	    sh4_translate_instruction( pc + 2 );

 	    exit_block_rel( disp + pc + 4, pc+4 );

 	    return 4;

 	}

     }

 :}

 BRAF Rn {:  

     COUNT_INST(I_BRAF);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	MOVL_rbpdisp_r32( R_PC, REG_EAX );

 	ADDL_imms_r32( pc + 4 - sh4_x86.block_start_pc, REG_EAX );

 	ADDL_rbpdisp_r32( REG_OFFSET(r[Rn]), REG_EAX );

 	MOVL_r32_rbpdisp( REG_EAX, R_NEW_PC );

 	sh4_x86.in_delay_slot = DELAY_PC;

 	sh4_x86.tstate = TSTATE_NONE;

 	sh4_x86.branch_taken = TRUE;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    exit_block_emu(pc+2);

 	    return 2;

 	} else {

 	    sh4_translate_instruction( pc + 2 );

 	    exit_block_newpcset(pc+4);

 	    return 4;

 	}

     }

 :}

 BSR disp {:  

     COUNT_INST(I_BSR);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	MOVL_rbpdisp_r32( R_PC, REG_EAX );

 	ADDL_imms_r32( pc + 4 - sh4_x86.block_start_pc, REG_EAX );

 	MOVL_r32_rbpdisp( REG_EAX, R_PR );

 	sh4_x86.in_delay_slot = DELAY_PC;

 	sh4_x86.branch_taken = TRUE;

 	sh4_x86.tstate = TSTATE_NONE;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    ADDL_imms_r32( disp, REG_EAX );

 	    MOVL_r32_rbpdisp( REG_EAX, R_NEW_PC );

 	    exit_block_emu(pc+2);

 	    return 2;

 	} else {

 	    sh4_translate_instruction( pc + 2 );

 	    exit_block_rel( disp + pc + 4, pc+4 );

 	    return 4;

 	}

     }

 :}

 BSRF Rn {:  

     COUNT_INST(I_BSRF);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	MOVL_rbpdisp_r32( R_PC, REG_EAX );

 	ADDL_imms_r32( pc + 4 - sh4_x86.block_start_pc, REG_EAX );

 	MOVL_r32_rbpdisp( REG_EAX, R_PR );

 	ADDL_rbpdisp_r32( REG_OFFSET(r[Rn]), REG_EAX );

 	MOVL_r32_rbpdisp( REG_EAX, R_NEW_PC );

 	sh4_x86.in_delay_slot = DELAY_PC;

 	sh4_x86.tstate = TSTATE_NONE;

 	sh4_x86.branch_taken = TRUE;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    exit_block_emu(pc+2);

 	    return 2;

 	} else {

 	    sh4_translate_instruction( pc + 2 );

 	    exit_block_newpcset(pc+4);

 	    return 4;

 	}

     }

 :}

 BT disp {:

     COUNT_INST(I_BT);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	sh4vma_t target = disp + pc + 4;

 	JF_label( nottaken );

 	exit_block_rel(target, pc+2 );

 	JMP_TARGET(nottaken);

 	return 2;

     }

 :}

 BT/S disp {:

     COUNT_INST(I_BTS);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	sh4_x86.in_delay_slot = DELAY_PC;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    MOVL_imm32_r32( pc + 4 - sh4_x86.block_start_pc, REG_EAX );

 	    JF_label(nottaken);

 	    ADDL_imms_r32( disp, REG_EAX );

 	    JMP_TARGET(nottaken);

 	    ADDL_rbpdisp_r32( R_PC, REG_EAX );

 	    MOVL_r32_rbpdisp( REG_EAX, R_NEW_PC );

 	    exit_block_emu(pc+2);

 	    sh4_x86.branch_taken = TRUE;

 	    return 2;

 	} else {

 		LOAD_t();

 	    JCC_cc_rel32(sh4_x86.tstate^1,0);

 	    uint32_t *patch = ((uint32_t *)xlat_output)-1;

 	    int save_tstate = sh4_x86.tstate;

 	    sh4_translate_instruction(pc+2);

             sh4_x86.in_delay_slot = DELAY_PC; /* Cleared by sh4_translate_instruction */

 	    exit_block_rel( disp + pc + 4, pc+4 );

 	    // not taken

 	    *patch = (xlat_output - ((uint8_t *)patch)) - 4;

 	    sh4_x86.tstate = save_tstate;

 	    sh4_translate_instruction(pc+2);

 	    return 4;

 	}

     }

 :}

 JMP @Rn {:  

     COUNT_INST(I_JMP);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	load_reg( REG_ECX, Rn );

 	MOVL_r32_rbpdisp( REG_ECX, R_NEW_PC );

 	sh4_x86.in_delay_slot = DELAY_PC;

 	sh4_x86.branch_taken = TRUE;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    exit_block_emu(pc+2);

 	    return 2;

 	} else {

 	    sh4_translate_instruction(pc+2);

 	    exit_block_newpcset(pc+4);

 	    return 4;

 	}

     }

 :}

 JSR @Rn {:  

     COUNT_INST(I_JSR);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	MOVL_rbpdisp_r32( R_PC, REG_EAX );

 	ADDL_imms_r32( pc + 4 - sh4_x86.block_start_pc, REG_EAX );

 	MOVL_r32_rbpdisp( REG_EAX, R_PR );

 	load_reg( REG_ECX, Rn );

 	MOVL_r32_rbpdisp( REG_ECX, R_NEW_PC );

 	sh4_x86.in_delay_slot = DELAY_PC;

 	sh4_x86.branch_taken = TRUE;

 	sh4_x86.tstate = TSTATE_NONE;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    exit_block_emu(pc+2);

 	    return 2;

 	} else {

 	    sh4_translate_instruction(pc+2);

 	    exit_block_newpcset(pc+4);

 	    return 4;

 	}

     }

 :}

 RTE {:  

     COUNT_INST(I_RTE);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	check_priv();

 	MOVL_rbpdisp_r32( R_SPC, REG_ECX );

 	MOVL_r32_rbpdisp( REG_ECX, R_NEW_PC );

 	MOVL_rbpdisp_r32( R_SSR, REG_EAX );

 	CALL1_ptr_r32( sh4_write_sr, REG_EAX );

 	sh4_x86.in_delay_slot = DELAY_PC;

 	sh4_x86.fpuen_checked = FALSE;

 	sh4_x86.tstate = TSTATE_NONE;

 	sh4_x86.branch_taken = TRUE;

     sh4_x86.sh4_mode = SH4_MODE_UNKNOWN;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    exit_block_emu(pc+2);

 	    return 2;

 	} else {

 	    sh4_translate_instruction(pc+2);

 	    exit_block_newpcset(pc+4);

 	    return 4;

 	}

     }

 :}

 RTS {:  

     COUNT_INST(I_RTS);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	MOVL_rbpdisp_r32( R_PR, REG_ECX );

 	MOVL_r32_rbpdisp( REG_ECX, R_NEW_PC );

 	sh4_x86.in_delay_slot = DELAY_PC;

 	sh4_x86.branch_taken = TRUE;

 	if( UNTRANSLATABLE(pc+2) ) {

 	    exit_block_emu(pc+2);

 	    return 2;

 	} else {

 	    sh4_translate_instruction(pc+2);

 	    exit_block_newpcset(pc+4);

 	    return 4;

 	}

     }

 :}

 TRAPA #imm {:  

     COUNT_INST(I_TRAPA);

     if( sh4_x86.in_delay_slot ) {

 	SLOTILLEGAL();

     } else {

 	MOVL_imm32_r32( pc+2 - sh4_x86.block_start_pc, REG_ECX );   // 5

 	ADDL_r32_rbpdisp( REG_ECX, R_PC );

 	MOVL_imm32_r32( imm, REG_EAX );

 	CALL1_ptr_r32( sh4_raise_trap, REG_EAX );

 	sh4_x86.tstate = TSTATE_NONE;

 	exit_block_pcset(pc+2);

 	sh4_x86.branch_taken = TRUE;

 	return 2;

     }

 :}

 UNDEF {:  

     COUNT_INST(I_UNDEF);

     if( sh4_x86.in_delay_slot ) {

 	exit_block_exc(EXC_SLOT_ILLEGAL, pc-2, 4);    

     } else {

 	exit_block_exc(EXC_ILLEGAL, pc, 2);    

 	return 2;

     }

 :}

 CLRMAC {:  

     COUNT_INST(I_CLRMAC);

     XORL_r32_r32(REG_EAX, REG_EAX);

     MOVL_r32_rbpdisp( REG_EAX, R_MACL );

     MOVL_r32_rbpdisp( REG_EAX, R_MACH );

     sh4_x86.tstate = TSTATE_NONE;

 :}

 CLRS {:

     COUNT_INST(I_CLRS);

     CLC();

     SETCCB_cc_rbpdisp(X86_COND_C, R_S);

     sh4_x86.tstate = TSTATE_NONE;

 :}

 CLRT {:  

     COUNT_INST(I_CLRT);

     CLC();

     SETC_t();

     sh4_x86.tstate = TSTATE_C;

 :}

 SETS {:  

     COUNT_INST(I_SETS);

     STC();

     SETCCB_cc_rbpdisp(X86_COND_C, R_S);

     sh4_x86.tstate = TSTATE_NONE;

 :}

 SETT {:  

     COUNT_INST(I_SETT);

     STC();

     SETC_t();

     sh4_x86.tstate = TSTATE_C;

 :}

 /* Floating point moves */

 FMOV FRm, FRn {:  

     COUNT_INST(I_FMOV1);

     check_fpuen();

     if( sh4_x86.double_size ) {

         load_dr0( REG_EAX, FRm );

         load_dr1( REG_ECX, FRm );

         store_dr0( REG_EAX, FRn );

         store_dr1( REG_ECX, FRn );

     } else {

         load_fr( REG_EAX, FRm ); // SZ=0 branch

         store_fr( REG_EAX, FRn );

     }

 :}

 FMOV FRm, @Rn {: 

     COUNT_INST(I_FMOV2);

     check_fpuen();

     load_reg( REG_EAX, Rn );

     if( sh4_x86.double_size ) {

         check_walign64( REG_EAX );

         load_dr0( REG_EDX, FRm );

         MEM_WRITE_LONG( REG_EAX, REG_EDX );

         load_reg( REG_EAX, Rn );

         LEAL_r32disp_r32( REG_EAX, 4, REG_EAX );

         load_dr1( REG_EDX, FRm );

         MEM_WRITE_LONG( REG_EAX, REG_EDX );

     } else {

         check_walign32( REG_EAX );

         load_fr( REG_EDX, FRm );

         MEM_WRITE_LONG( REG_EAX, REG_EDX );

     }

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FMOV @Rm, FRn {:  

     COUNT_INST(I_FMOV5);

     check_fpuen();

     load_reg( REG_EAX, Rm );

     if( sh4_x86.double_size ) {

         check_ralign64( REG_EAX );

         MEM_READ_LONG( REG_EAX, REG_EAX );

         store_dr0( REG_EAX, FRn );

         load_reg( REG_EAX, Rm );

         LEAL_r32disp_r32( REG_EAX, 4, REG_EAX );

         MEM_READ_LONG( REG_EAX, REG_EAX );

         store_dr1( REG_EAX, FRn );

     } else {

         check_ralign32( REG_EAX );

         MEM_READ_LONG( REG_EAX, REG_EAX );

         store_fr( REG_EAX, FRn );

     }

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FMOV FRm, @-Rn {:  

     COUNT_INST(I_FMOV3);

     check_fpuen();

     load_reg( REG_EAX, Rn );

     if( sh4_x86.double_size ) {

         check_walign64( REG_EAX );

         LEAL_r32disp_r32( REG_EAX, -8, REG_EAX );

         load_dr0( REG_EDX, FRm );

         MEM_WRITE_LONG( REG_EAX, REG_EDX );

         load_reg( REG_EAX, Rn );

         LEAL_r32disp_r32( REG_EAX, -4, REG_EAX );

         load_dr1( REG_EDX, FRm );

         MEM_WRITE_LONG( REG_EAX, REG_EDX );

         ADDL_imms_rbpdisp(-8,REG_OFFSET(r[Rn]));

     } else {

         check_walign32( REG_EAX );

         LEAL_r32disp_r32( REG_EAX, -4, REG_EAX );

         load_fr( REG_EDX, FRm );

         MEM_WRITE_LONG( REG_EAX, REG_EDX );

         ADDL_imms_rbpdisp(-4,REG_OFFSET(r[Rn]));

     }

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FMOV @Rm+, FRn {:

     COUNT_INST(I_FMOV6);

     check_fpuen();

     load_reg( REG_EAX, Rm );

     if( sh4_x86.double_size ) {

         check_ralign64( REG_EAX );

         MEM_READ_LONG( REG_EAX, REG_EAX );

         store_dr0( REG_EAX, FRn );

         load_reg( REG_EAX, Rm );

         LEAL_r32disp_r32( REG_EAX, 4, REG_EAX );

         MEM_READ_LONG( REG_EAX, REG_EAX );

         store_dr1( REG_EAX, FRn );

         ADDL_imms_rbpdisp( 8, REG_OFFSET(r[Rm]) );

     } else {

         check_ralign32( REG_EAX );

         MEM_READ_LONG( REG_EAX, REG_EAX );

         store_fr( REG_EAX, FRn );

         ADDL_imms_rbpdisp( 4, REG_OFFSET(r[Rm]) );

     }

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FMOV FRm, @(R0, Rn) {:  

     COUNT_INST(I_FMOV4);

     check_fpuen();

     load_reg( REG_EAX, Rn );

     ADDL_rbpdisp_r32( REG_OFFSET(r[0]), REG_EAX );

     if( sh4_x86.double_size ) {

         check_walign64( REG_EAX );

         load_dr0( REG_EDX, FRm );

         MEM_WRITE_LONG( REG_EAX, REG_EDX );

         load_reg( REG_EAX, Rn );

         ADDL_rbpdisp_r32( REG_OFFSET(r[0]), REG_EAX );

         LEAL_r32disp_r32( REG_EAX, 4, REG_EAX );

         load_dr1( REG_EDX, FRm );

         MEM_WRITE_LONG( REG_EAX, REG_EDX );

     } else {

         check_walign32( REG_EAX );

         load_fr( REG_EDX, FRm );

         MEM_WRITE_LONG( REG_EAX, REG_EDX ); // 12

     }

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FMOV @(R0, Rm), FRn {:  

     COUNT_INST(I_FMOV7);

     check_fpuen();

     load_reg( REG_EAX, Rm );

     ADDL_rbpdisp_r32( REG_OFFSET(r[0]), REG_EAX );

     if( sh4_x86.double_size ) {

         check_ralign64( REG_EAX );

         MEM_READ_LONG( REG_EAX, REG_EAX );

         store_dr0( REG_EAX, FRn );

         load_reg( REG_EAX, Rm );

         ADDL_rbpdisp_r32( REG_OFFSET(r[0]), REG_EAX );

         LEAL_r32disp_r32( REG_EAX, 4, REG_EAX );

         MEM_READ_LONG( REG_EAX, REG_EAX );

         store_dr1( REG_EAX, FRn );

     } else {

         check_ralign32( REG_EAX );

         MEM_READ_LONG( REG_EAX, REG_EAX );

         store_fr( REG_EAX, FRn );

     }

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FLDI0 FRn {:  /* IFF PR=0 */

     COUNT_INST(I_FLDI0);

     check_fpuen();

     if( sh4_x86.double_prec == 0 ) {

         XORL_r32_r32( REG_EAX, REG_EAX );

         store_fr( REG_EAX, FRn );

     }

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FLDI1 FRn {:  /* IFF PR=0 */

     COUNT_INST(I_FLDI1);

     check_fpuen();

     if( sh4_x86.double_prec == 0 ) {

         MOVL_imm32_r32( 0x3F800000, REG_EAX );

         store_fr( REG_EAX, FRn );

     }

 :}

 FLOAT FPUL, FRn {:  

     COUNT_INST(I_FLOAT);

     check_fpuen();

     FILD_rbpdisp(R_FPUL);

     if( sh4_x86.double_prec ) {

         pop_dr( FRn );

     } else {

         pop_fr( FRn );

     }

 :}

 FTRC FRm, FPUL {:  

     COUNT_INST(I_FTRC);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr( FRm );

     } else {

         push_fr( FRm );

     }

     MOVP_immptr_rptr( &min_int, REG_ECX );

     FILD_r32disp( REG_ECX, 0 );

     FCOMIP_st(1);              

     JAE_label( sat );     

     JP_label( sat2 );       

     MOVP_immptr_rptr( &max_int, REG_ECX );

     FILD_r32disp( REG_ECX, 0 );

     FCOMIP_st(1);

     JNA_label( sat3 );

     MOVP_immptr_rptr( &save_fcw, REG_EAX );

     FNSTCW_r32disp( REG_EAX, 0 );

     MOVP_immptr_rptr( &trunc_fcw, REG_EDX );

     FLDCW_r32disp( REG_EDX, 0 );

     FISTP_rbpdisp(R_FPUL);             

     FLDCW_r32disp( REG_EAX, 0 );

     JMP_label(end);             

     JMP_TARGET(sat);

     JMP_TARGET(sat2);

     JMP_TARGET(sat3);

     MOVL_r32disp_r32( REG_ECX, 0, REG_ECX ); // 2

     MOVL_r32_rbpdisp( REG_ECX, R_FPUL );

     FPOP_st();

     JMP_TARGET(end);

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FLDS FRm, FPUL {:  

     COUNT_INST(I_FLDS);

     check_fpuen();

     load_fr( REG_EAX, FRm );

     MOVL_r32_rbpdisp( REG_EAX, R_FPUL );

 :}

 FSTS FPUL, FRn {:  

     COUNT_INST(I_FSTS);

     check_fpuen();

     MOVL_rbpdisp_r32( R_FPUL, REG_EAX );

     store_fr( REG_EAX, FRn );

 :}

 FCNVDS FRm, FPUL {:  

     COUNT_INST(I_FCNVDS);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr( FRm );

         pop_fpul();

     }

 :}

 FCNVSD FPUL, FRn {:  

     COUNT_INST(I_FCNVSD);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_fpul();

         pop_dr( FRn );

     }

 :}

 /* Floating point instructions */

 FABS FRn {:  

     COUNT_INST(I_FABS);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr(FRn);

         FABS_st0();

         pop_dr(FRn);

     } else {

         push_fr(FRn);

         FABS_st0();

         pop_fr(FRn);

     }

 :}

 FADD FRm, FRn {:  

     COUNT_INST(I_FADD);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr(FRm);

         push_dr(FRn);

         FADDP_st(1);

         pop_dr(FRn);

     } else {

         push_fr(FRm);

         push_fr(FRn);

         FADDP_st(1);

         pop_fr(FRn);

     }

 :}

 FDIV FRm, FRn {:  

     COUNT_INST(I_FDIV);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr(FRn);

         push_dr(FRm);

         FDIVP_st(1);

         pop_dr(FRn);

     } else {

         push_fr(FRn);

         push_fr(FRm);

         FDIVP_st(1);

         pop_fr(FRn);

     }

 :}

 FMAC FR0, FRm, FRn {:  

     COUNT_INST(I_FMAC);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr( 0 );

         push_dr( FRm );

         FMULP_st(1);

         push_dr( FRn );

         FADDP_st(1);

         pop_dr( FRn );

     } else {

         push_fr( 0 );

         push_fr( FRm );

         FMULP_st(1);

         push_fr( FRn );

         FADDP_st(1);

         pop_fr( FRn );

     }

 :}

 FMUL FRm, FRn {:  

     COUNT_INST(I_FMUL);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr(FRm);

         push_dr(FRn);

         FMULP_st(1);

         pop_dr(FRn);

     } else {

         push_fr(FRm);

         push_fr(FRn);

         FMULP_st(1);

         pop_fr(FRn);

     }

 :}

 FNEG FRn {:  

     COUNT_INST(I_FNEG);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr(FRn);

         FCHS_st0();

         pop_dr(FRn);

     } else {

         push_fr(FRn);

         FCHS_st0();

         pop_fr(FRn);

     }

 :}

 FSRRA FRn {:  

     COUNT_INST(I_FSRRA);

     check_fpuen();

     if( sh4_x86.double_prec == 0 ) {

         FLD1_st0();

         push_fr(FRn);

         FSQRT_st0();

         FDIVP_st(1);

         pop_fr(FRn);

     }

 :}

 FSQRT FRn {:  

     COUNT_INST(I_FSQRT);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr(FRn);

         FSQRT_st0();

         pop_dr(FRn);

     } else {

         push_fr(FRn);

         FSQRT_st0();

         pop_fr(FRn);

     }

 :}

 FSUB FRm, FRn {:  

     COUNT_INST(I_FSUB);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr(FRn);

         push_dr(FRm);

         FSUBP_st(1);

         pop_dr(FRn);

     } else {

         push_fr(FRn);

         push_fr(FRm);

         FSUBP_st(1);

         pop_fr(FRn);

     }

 :}

 FCMP/EQ FRm, FRn {:  

     COUNT_INST(I_FCMPEQ);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr(FRm);

         push_dr(FRn);

     } else {

         push_fr(FRm);

         push_fr(FRn);

     }

     XORL_r32_r32(REG_EAX, REG_EAX);

     XORL_r32_r32(REG_EDX, REG_EDX);

     FCOMIP_st(1);

     SETCCB_cc_r8(X86_COND_NP, REG_DL);

     CMOVCCL_cc_r32_r32(X86_COND_E, REG_EDX, REG_EAX);

     MOVL_r32_rbpdisp(REG_EAX, R_T);

     FPOP_st();

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FCMP/GT FRm, FRn {:  

     COUNT_INST(I_FCMPGT);

     check_fpuen();

     if( sh4_x86.double_prec ) {

         push_dr(FRm);

         push_dr(FRn);

     } else {

         push_fr(FRm);

         push_fr(FRn);

     }

     FCOMIP_st(1);

     SETA_t();

     FPOP_st();

     sh4_x86.tstate = TSTATE_A;

 :}

 FSCA FPUL, FRn {:  

     COUNT_INST(I_FSCA);

     check_fpuen();

     if( sh4_x86.double_prec == 0 ) {

         LEAP_rbpdisp_rptr( REG_OFFSET(fr[0][FRn&0x0E]), REG_EDX );

         MOVL_rbpdisp_r32( R_FPUL, REG_EAX );

         CALL2_ptr_r32_r32( sh4_fsca, REG_EAX, REG_EDX );

     }

     sh4_x86.tstate = TSTATE_NONE;

 :}

 FIPR FVm, FVn {:  

     COUNT_INST(I_FIPR);

     check_fpuen();

     if( sh4_x86.double_prec == 0 ) {

         if( sh4_x86.sse3_enabled ) {

             MOVAPS_rbpdisp_xmm( REG_OFFSET(fr[0][FVm<<2]), 4 );

             MULPS_rbpdisp_xmm( REG_OFFSET(fr[0][FVn<<2]), 4 );

             HADDPS_xmm_xmm( 4, 4 ); 

             HADDPS_xmm_xmm( 4, 4 );

             MOVSS_xmm_rbpdisp( 4, REG_OFFSET(fr[0][(FVn<<2)+2]) );

         } else {

             push_fr( FVm<<2 );

             push_fr( FVn<<2 );

             FMULP_st(1);

             push_fr( (FVm<<2)+1);

             push_fr( (FVn<<2)+1);

             FMULP_st(1);

             FADDP_st(1);

             push_fr( (FVm<<2)+2);

             push_fr( (FVn<<2)+2);

             FMULP_st(1);

             FADDP_st(1);

             push_fr( (FVm<<2)+3);