X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fnew_dynarec.c;h=28bee08fa5c8484d7221f8779df8e9f5a8b770ed;hb=df4dc2b13cf71f32751cf842b03ef87991b2c55a;hp=21e7feda9b4051900b1085b782837c6fea6caac3;hpb=9f51b4b9aa4fffa7b9411ca274f5dfb179a6a30a;p=pcsx_rearmed.git

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index 21e7feda..28bee08f 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -23,7 +23,19 @@
 #include <assert.h>
 #include <errno.h>
 #include <sys/mman.h>
+#ifdef __MACH__
+#include <libkern/OSCacheControl.h>
+#endif
+#ifdef _3DS
+#include <3ds_utils.h>
+#endif
+#ifdef VITA
+#include <psp2/kernel/sysmem.h>
+static int sceBlock;
+#endif
 
+#include "new_dynarec_config.h"
+#include "../psxhle.h" //emulator interface
 #include "emu_if.h" //emulator interface
 
 //#define DISASM
@@ -42,19 +54,6 @@
 #include "assem_arm.h"
 #endif
 
-#ifdef __BLACKBERRY_QNX__
-#undef __clear_cache
-#define __clear_cache(start,end) msync(start, (size_t)((void*)end - (void*)start), MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
-#elif defined(__MACH__)
-#include <libkern/OSCacheControl.h>
-#define __clear_cache mach_clear_cache
-static void __clear_cache(void *start, void *end) {
-  size_t len = (char *)end - (char *)start;
-  sys_dcache_flush(start, len);
-  sys_icache_invalidate(start, len);
-}
-#endif
-
 #define MAXBLOCK 4096
 #define MAX_OUTPUT_BLOCK_SIZE 262144
 
@@ -83,22 +82,35 @@ struct ll_entry
   struct ll_entry *next;
 };
 
-  u_int start;
-  u_int *source;
-  char insn[MAXBLOCK][10];
-  u_char itype[MAXBLOCK];
-  u_char opcode[MAXBLOCK];
-  u_char opcode2[MAXBLOCK];
-  u_char bt[MAXBLOCK];
-  u_char rs1[MAXBLOCK];
-  u_char rs2[MAXBLOCK];
-  u_char rt1[MAXBLOCK];
-  u_char rt2[MAXBLOCK];
-  u_char us1[MAXBLOCK];
-  u_char us2[MAXBLOCK];
-  u_char dep1[MAXBLOCK];
-  u_char dep2[MAXBLOCK];
-  u_char lt1[MAXBLOCK];
+struct ht_entry
+{
+  u_int vaddr[2];
+  void *tcaddr[2];
+};
+
+  // used by asm:
+  u_char *out;
+  struct ht_entry hash_table[65536]  __attribute__((aligned(16)));
+  struct ll_entry *jump_in[4096] __attribute__((aligned(16)));
+  struct ll_entry *jump_dirty[4096];
+
+  static struct ll_entry *jump_out[4096];
+  static u_int start;
+  static u_int *source;
+  static char insn[MAXBLOCK][10];
+  static u_char itype[MAXBLOCK];
+  static u_char opcode[MAXBLOCK];
+  static u_char opcode2[MAXBLOCK];
+  static u_char bt[MAXBLOCK];
+  static u_char rs1[MAXBLOCK];
+  static u_char rs2[MAXBLOCK];
+  static u_char rt1[MAXBLOCK];
+  static u_char rt2[MAXBLOCK];
+  static u_char us1[MAXBLOCK];
+  static u_char us2[MAXBLOCK];
+  static u_char dep1[MAXBLOCK];
+  static u_char dep2[MAXBLOCK];
+  static u_char lt1[MAXBLOCK];
   static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
   static uint64_t gte_rt[MAXBLOCK];
   static uint64_t gte_unneeded[MAXBLOCK];
@@ -107,52 +119,47 @@ struct ll_entry
   static u_int smrv_weak; // same, but somewhat less likely
   static u_int smrv_strong_next; // same, but after current insn executes
   static u_int smrv_weak_next;
-  int imm[MAXBLOCK];
-  u_int ba[MAXBLOCK];
-  char likely[MAXBLOCK];
-  char is_ds[MAXBLOCK];
-  char ooo[MAXBLOCK];
-  uint64_t unneeded_reg[MAXBLOCK];
-  uint64_t unneeded_reg_upper[MAXBLOCK];
-  uint64_t branch_unneeded_reg[MAXBLOCK];
-  uint64_t branch_unneeded_reg_upper[MAXBLOCK];
-  uint64_t pr32[MAXBLOCK];
-  signed char regmap_pre[MAXBLOCK][HOST_REGS];
+  static int imm[MAXBLOCK];
+  static u_int ba[MAXBLOCK];
+  static char likely[MAXBLOCK];
+  static char is_ds[MAXBLOCK];
+  static char ooo[MAXBLOCK];
+  static uint64_t unneeded_reg[MAXBLOCK];
+  static uint64_t unneeded_reg_upper[MAXBLOCK];
+  static uint64_t branch_unneeded_reg[MAXBLOCK];
+  static uint64_t branch_unneeded_reg_upper[MAXBLOCK];
+  static signed char regmap_pre[MAXBLOCK][HOST_REGS];
   static uint64_t current_constmap[HOST_REGS];
   static uint64_t constmap[MAXBLOCK][HOST_REGS];
   static struct regstat regs[MAXBLOCK];
   static struct regstat branch_regs[MAXBLOCK];
-  signed char minimum_free_regs[MAXBLOCK];
-  u_int needed_reg[MAXBLOCK];
-  u_int wont_dirty[MAXBLOCK];
-  u_int will_dirty[MAXBLOCK];
-  int ccadj[MAXBLOCK];
-  int slen;
-  u_int instr_addr[MAXBLOCK];
-  u_int link_addr[MAXBLOCK][3];
-  int linkcount;
-  u_int stubs[MAXBLOCK*3][8];
-  int stubcount;
-  u_int literals[1024][2];
-  int literalcount;
-  int is_delayslot;
-  int cop1_usable;
-  u_char *out;
-  struct ll_entry *jump_in[4096] __attribute__((aligned(16)));
-  struct ll_entry *jump_out[4096];
-  struct ll_entry *jump_dirty[4096];
-  u_int hash_table[65536][4]  __attribute__((aligned(16)));
-  char shadow[1048576]  __attribute__((aligned(16)));
-  void *copy;
-  int expirep;
-  int new_dynarec_did_compile;
-  int new_dynarec_hacks;
-  u_int stop_after_jal;
+  static signed char minimum_free_regs[MAXBLOCK];
+  static u_int needed_reg[MAXBLOCK];
+  static u_int wont_dirty[MAXBLOCK];
+  static u_int will_dirty[MAXBLOCK];
+  static int ccadj[MAXBLOCK];
+  static int slen;
+  static void *instr_addr[MAXBLOCK];
+  static u_int link_addr[MAXBLOCK][3];
+  static int linkcount;
+  static u_int stubs[MAXBLOCK*3][8];
+  static int stubcount;
+  static u_int literals[1024][2];
+  static int literalcount;
+  static int is_delayslot;
+  static int cop1_usable;
+  static char shadow[1048576]  __attribute__((aligned(16)));
+  static void *copy;
+  static int expirep;
+  static u_int stop_after_jal;
 #ifndef RAM_FIXED
   static u_int ram_offset;
 #else
   static const u_int ram_offset=0;
 #endif
+
+  int new_dynarec_hacks;
+  int new_dynarec_did_compile;
   extern u_char restore_candidate[512];
   extern int cycle_count;
 
@@ -254,15 +261,80 @@ void jump_intcall();
 void new_dyna_leave();
 
 // Needed by assembler
-void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32);
-void wb_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty);
-void wb_needed_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr);
-void load_all_regs(signed char i_regmap[]);
-void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
-void load_regs_entry(int t);
-void load_all_consts(signed char regmap[],int is32,u_int dirty,int i);
+static void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32);
+static void wb_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty);
+static void wb_needed_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr);
+static void load_all_regs(signed char i_regmap[]);
+static void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
+static void load_regs_entry(int t);
+static void load_all_consts(signed char regmap[],int is32,u_int dirty,int i);
+
+static int verify_dirty(u_int *ptr);
+static int get_final_value(int hr, int i, int *value);
+static void add_stub(int type,int addr,int retaddr,int a,int b,int c,int d,int e);
+static void add_to_linker(int addr,int target,int ext);
+
+static int tracedebug=0;
+
+static void mprotect_w_x(void *start, void *end, int is_x)
+{
+#ifdef NO_WRITE_EXEC
+  #if defined(VITA)
+  // *Open* enables write on all memory that was
+  // allocated by sceKernelAllocMemBlockForVM()?
+  if (is_x)
+    sceKernelCloseVMDomain();
+  else
+    sceKernelOpenVMDomain();
+  #else
+  u_long mstart = (u_long)start & ~4095ul;
+  u_long mend = (u_long)end;
+  if (mprotect((void *)mstart, mend - mstart,
+               PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
+    SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
+  #endif
+#endif
+}
 
-int tracedebug=0;
+static void start_tcache_write(void *start, void *end)
+{
+  mprotect_w_x(start, end, 0);
+}
+
+static void end_tcache_write(void *start, void *end)
+{
+#ifdef __arm__
+  size_t len = (char *)end - (char *)start;
+  #if   defined(__BLACKBERRY_QNX__)
+  msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
+  #elif defined(__MACH__)
+  sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
+  #elif defined(VITA)
+  sceKernelSyncVMDomain(sceBlock, start, len);
+  #elif defined(_3DS)
+  ctr_flush_invalidate_cache();
+  #else
+  __clear_cache(start, end);
+  #endif
+  (void)len;
+#endif
+
+  mprotect_w_x(start, end, 1);
+}
+
+static void *start_block(void)
+{
+  u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
+  if (end > (u_char *)BASE_ADDR + (1<<TARGET_SIZE_2))
+    end = (u_char *)BASE_ADDR + (1<<TARGET_SIZE_2);
+  start_tcache_write(out, end);
+  return out;
+}
+
+static void end_block(void *start)
+{
+  end_tcache_write(start, out);
+}
 
 //#define DEBUG_CYCLE_COUNT 1
 
@@ -292,6 +364,26 @@ static u_int get_vpage(u_int vaddr)
   return get_page(vaddr);
 }
 
+static struct ht_entry *hash_table_get(u_int vaddr)
+{
+  return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
+}
+
+static void hash_table_add(struct ht_entry *ht_bin, u_int vaddr, void *tcaddr)
+{
+  ht_bin->vaddr[1] = ht_bin->vaddr[0];
+  ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
+  ht_bin->vaddr[0] = vaddr;
+  ht_bin->tcaddr[0] = tcaddr;
+}
+
+// some messy ari64's code, seems to rely on unsigned 32bit overflow
+static int doesnt_expire_soon(void *tcaddr)
+{
+  u_int diff = (u_int)((u_char *)tcaddr - out) << (32-TARGET_SIZE_2);
+  return diff > (u_int)(0x60000000 + (MAX_OUTPUT_BLOCK_SIZE << (32-TARGET_SIZE_2)));
+}
+
 // Get address from virtual address
 // This is called from the recompiled JR/JALR instructions
 void *get_addr(u_int vaddr)
@@ -304,11 +396,7 @@ void *get_addr(u_int vaddr)
   while(head!=NULL) {
     if(head->vaddr==vaddr) {
   //printf("TRACE: count=%d next=%d (get_addr match %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
-      int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
-      ht_bin[3]=ht_bin[1];
-      ht_bin[2]=ht_bin[0];
-      ht_bin[1]=(int)head->addr;
-      ht_bin[0]=vaddr;
+      hash_table_add(hash_table_get(vaddr), vaddr, head->addr);
       return head->addr;
     }
     head=head->next;
@@ -318,8 +406,8 @@ void *get_addr(u_int vaddr)
     if(head->vaddr==vaddr) {
       //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
       // Don't restore blocks which are about to expire from the cache
-      if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
-      if(verify_dirty(head->addr)) {
+      if (doesnt_expire_soon(head->addr))
+      if (verify_dirty(head->addr)) {
         //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
         invalid_code[vaddr>>12]=0;
         inv_code_start=inv_code_end=~0;
@@ -327,17 +415,12 @@ void *get_addr(u_int vaddr)
           restore_candidate[vpage>>3]|=1<<(vpage&7);
         }
         else restore_candidate[page>>3]|=1<<(page&7);
-        int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
-        if(ht_bin[0]==vaddr) {
-          ht_bin[1]=(int)head->addr; // Replace existing entry
-        }
+        struct ht_entry *ht_bin = hash_table_get(vaddr);
+        if (ht_bin->vaddr[0] == vaddr)
+          ht_bin->tcaddr[0] = head->addr; // Replace existing entry
         else
-        {
-          ht_bin[3]=ht_bin[1];
-          ht_bin[2]=ht_bin[0];
-          ht_bin[1]=(int)head->addr;
-          ht_bin[0]=vaddr;
-        }
+          hash_table_add(ht_bin, vaddr, head->addr);
+
         return head->addr;
       }
     }
@@ -359,9 +442,9 @@ void *get_addr(u_int vaddr)
 void *get_addr_ht(u_int vaddr)
 {
   //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
-  int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
-  if(ht_bin[0]==vaddr) return (void *)ht_bin[1];
-  if(ht_bin[2]==vaddr) return (void *)ht_bin[3];
+  const struct ht_entry *ht_bin = hash_table_get(vaddr);
+  if (ht_bin->vaddr[0] == vaddr) return ht_bin->tcaddr[0];
+  if (ht_bin->vaddr[1] == vaddr) return ht_bin->tcaddr[1];
   return get_addr(vaddr);
 }
 
@@ -627,6 +710,7 @@ int needed_again(int r, int i)
     }
   }*/
   if(rn<10) return 1;
+  (void)b;
   return 0;
 }
 
@@ -729,39 +813,39 @@ void ll_add_flags(struct ll_entry **head,int vaddr,u_int reg_sv_flags,void *addr
 // but don't return addresses which are about to expire from the cache
 void *check_addr(u_int vaddr)
 {
-  u_int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
-  if(ht_bin[0]==vaddr) {
-    if(((ht_bin[1]-MAX_OUTPUT_BLOCK_SIZE-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
-      if(isclean(ht_bin[1])) return (void *)ht_bin[1];
-  }
-  if(ht_bin[2]==vaddr) {
-    if(((ht_bin[3]-MAX_OUTPUT_BLOCK_SIZE-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
-      if(isclean(ht_bin[3])) return (void *)ht_bin[3];
+  struct ht_entry *ht_bin = hash_table_get(vaddr);
+  size_t i;
+  for (i = 0; i < sizeof(ht_bin->vaddr)/sizeof(ht_bin->vaddr[0]); i++) {
+    if (ht_bin->vaddr[i] == vaddr)
+      if (doesnt_expire_soon((u_char *)ht_bin->tcaddr[i] - MAX_OUTPUT_BLOCK_SIZE))
+        if (isclean(ht_bin->tcaddr[i]))
+          return ht_bin->tcaddr[i];
   }
   u_int page=get_page(vaddr);
   struct ll_entry *head;
   head=jump_in[page];
-  while(head!=NULL) {
-    if(head->vaddr==vaddr) {
-      if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
+  while (head != NULL) {
+    if (head->vaddr == vaddr) {
+      if (doesnt_expire_soon(head->addr)) {
         // Update existing entry with current address
-        if(ht_bin[0]==vaddr) {
-          ht_bin[1]=(int)head->addr;
+        if (ht_bin->vaddr[0] == vaddr) {
+          ht_bin->tcaddr[0] = head->addr;
           return head->addr;
         }
-        if(ht_bin[2]==vaddr) {
-          ht_bin[3]=(int)head->addr;
+        if (ht_bin->vaddr[1] == vaddr) {
+          ht_bin->tcaddr[1] = head->addr;
           return head->addr;
         }
         // Insert into hash table with low priority.
         // Don't evict existing entries, as they are probably
         // addresses that are being accessed frequently.
-        if(ht_bin[0]==-1) {
-          ht_bin[1]=(int)head->addr;
-          ht_bin[0]=vaddr;
-        }else if(ht_bin[2]==-1) {
-          ht_bin[3]=(int)head->addr;
-          ht_bin[2]=vaddr;
+        if (ht_bin->vaddr[0] == -1) {
+          ht_bin->vaddr[0] = vaddr;
+          ht_bin->tcaddr[0] = head->addr;
+        }
+        else if (ht_bin->vaddr[1] == -1) {
+          ht_bin->vaddr[1] = vaddr;
+          ht_bin->tcaddr[1] = head->addr;
         }
         return head->addr;
       }
@@ -774,14 +858,16 @@ void *check_addr(u_int vaddr)
 void remove_hash(int vaddr)
 {
   //printf("remove hash: %x\n",vaddr);
-  int *ht_bin=hash_table[(((vaddr)>>16)^vaddr)&0xFFFF];
-  if(ht_bin[2]==vaddr) {
-    ht_bin[2]=ht_bin[3]=-1;
+  struct ht_entry *ht_bin = hash_table_get(vaddr);
+  if (ht_bin->vaddr[1] == vaddr) {
+    ht_bin->vaddr[1] = -1;
+    ht_bin->tcaddr[1] = NULL;
   }
-  if(ht_bin[0]==vaddr) {
-    ht_bin[0]=ht_bin[2];
-    ht_bin[1]=ht_bin[3];
-    ht_bin[2]=ht_bin[3]=-1;
+  if (ht_bin->vaddr[0] == vaddr) {
+    ht_bin->vaddr[0] = ht_bin->vaddr[1];
+    ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
+    ht_bin->vaddr[1] = -1;
+    ht_bin->tcaddr[1] = NULL;
   }
 }
 
@@ -810,7 +896,7 @@ void ll_clear(struct ll_entry **head)
 {
   struct ll_entry *cur;
   struct ll_entry *next;
-  if(cur=*head) {
+  if((cur=*head)) {
     *head=0;
     while(cur) {
       next=cur->next;
@@ -821,7 +907,7 @@ void ll_clear(struct ll_entry **head)
 }
 
 // Dereference the pointers and remove if it matches
-void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
+static void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
 {
   while(head) {
     int ptr=get_pointer(head->addr);
@@ -830,10 +916,11 @@ void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
        (((ptr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift)))
     {
       inv_debug("EXP: Kill pointer at %x (%x)\n",(int)head->addr,head->vaddr);
-      u_int host_addr=(u_int)kill_pointer(head->addr);
+      void *host_addr=find_extjump_insn(head->addr);
       #ifdef __arm__
-        needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
+        mark_clear_cache(host_addr);
       #endif
+      set_jump_target(host_addr, head->addr);
     }
     head=head->next;
   }
@@ -857,10 +944,11 @@ void invalidate_page(u_int page)
   jump_out[page]=0;
   while(head!=NULL) {
     inv_debug("INVALIDATE: kill pointer to %x (%x)\n",head->vaddr,(int)head->addr);
-    u_int host_addr=(u_int)kill_pointer(head->addr);
+    void *host_addr=find_extjump_insn(head->addr);
     #ifdef __arm__
-      needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
+      mark_clear_cache(host_addr);
     #endif
+    set_jump_target(host_addr, head->addr);
     next=head->next;
     free(head);
     head=next;
@@ -984,7 +1072,7 @@ void invalidate_addr(u_int addr)
 // Anything could have changed, so invalidate everything.
 void invalidate_all_pages()
 {
-  u_int page,n;
+  u_int page;
   for(page=0;page<4096;page++)
     invalidate_page(page);
   for(page=0;page<1048576;page++)
@@ -992,9 +1080,6 @@ void invalidate_all_pages()
       restore_candidate[(page&2047)>>3]|=1<<(page&7);
       restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
     }
-  #ifdef __arm__
-  __clear_cache((void *)BASE_ADDR,(void *)BASE_ADDR+(1<<TARGET_SIZE_2));
-  #endif
   #ifdef USE_MINI_HT
   memset(mini_ht,-1,sizeof(mini_ht));
   #endif
@@ -1007,6 +1092,7 @@ void add_link(u_int vaddr,void *src)
   inv_debug("add_link: %x -> %x (%d)\n",(int)src,vaddr,page);
   int *ptr=(int *)(src+4);
   assert((*ptr&0x0fff0000)==0x059f0000);
+  (void)ptr;
   ll_add(jump_out+page,vaddr,src);
   //int ptr=get_pointer(src);
   //inv_debug("add_link: Pointer is to %x\n",(int)ptr);
@@ -1024,9 +1110,9 @@ void clean_blocks(u_int page)
   while(head!=NULL) {
     if(!invalid_code[head->vaddr>>12]) {
       // Don't restore blocks which are about to expire from the cache
-      if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
+      if (doesnt_expire_soon(head->addr)) {
         u_int start,end;
-        if(verify_dirty((int)head->addr)) {
+        if(verify_dirty(head->addr)) {
           //printf("Possibly Restore %x (%x)\n",head->vaddr, (int)head->addr);
           u_int i;
           u_int inv=0;
@@ -1040,20 +1126,18 @@ void clean_blocks(u_int page)
             inv=1;
           }
           if(!inv) {
-            void * clean_addr=(void *)get_clean_addr((int)head->addr);
-            if((((u_int)clean_addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
+            void *clean_addr = get_clean_addr(head->addr);
+            if (doesnt_expire_soon(clean_addr)) {
               u_int ppage=page;
               inv_debug("INV: Restored %x (%x/%x)\n",head->vaddr, (int)head->addr, (int)clean_addr);
               //printf("page=%x, addr=%x\n",page,head->vaddr);
               //assert(head->vaddr>>12==(page|0x80000));
               ll_add_flags(jump_in+ppage,head->vaddr,head->reg_sv_flags,clean_addr);
-              int *ht_bin=hash_table[((head->vaddr>>16)^head->vaddr)&0xFFFF];
-              if(ht_bin[0]==head->vaddr) {
-                ht_bin[1]=(int)clean_addr; // Replace existing entry
-              }
-              if(ht_bin[2]==head->vaddr) {
-                ht_bin[3]=(int)clean_addr; // Replace existing entry
-              }
+              struct ht_entry *ht_bin = hash_table_get(head->vaddr);
+              if (ht_bin->vaddr[0] == head->vaddr)
+                ht_bin->tcaddr[0] = clean_addr; // Replace existing entry
+              if (ht_bin->vaddr[1] == head->vaddr)
+                ht_bin->tcaddr[1] = clean_addr; // Replace existing entry
             }
           }
         }
@@ -1747,7 +1831,7 @@ static void pagespan_alloc(struct regstat *current,int i)
   //else ...
 }
 
-add_stub(int type,int addr,int retaddr,int a,int b,int c,int d,int e)
+static void add_stub(int type,int addr,int retaddr,int a,int b,int c,int d,int e)
 {
   stubs[stubcount][0]=type;
   stubs[stubcount][1]=addr;
@@ -2326,23 +2410,25 @@ void imm16_assemble(int i,struct regstat *i_regs)
                 emit_mov(sh,th);
               }
             }
-            if(opcode[i]==0x0d) //ORI
-            if(sl<0) {
-              emit_orimm(tl,imm[i],tl);
-            }else{
-              if(!((i_regs->wasconst>>sl)&1))
-                emit_orimm(sl,imm[i],tl);
-              else
-                emit_movimm(constmap[i][sl]|imm[i],tl);
+            if(opcode[i]==0x0d) { // ORI
+              if(sl<0) {
+                emit_orimm(tl,imm[i],tl);
+              }else{
+                if(!((i_regs->wasconst>>sl)&1))
+                  emit_orimm(sl,imm[i],tl);
+                else
+                  emit_movimm(constmap[i][sl]|imm[i],tl);
+              }
             }
-            if(opcode[i]==0x0e) //XORI
-            if(sl<0) {
-              emit_xorimm(tl,imm[i],tl);
-            }else{
-              if(!((i_regs->wasconst>>sl)&1))
-                emit_xorimm(sl,imm[i],tl);
-              else
-                emit_movimm(constmap[i][sl]^imm[i],tl);
+            if(opcode[i]==0x0e) { // XORI
+              if(sl<0) {
+                emit_xorimm(tl,imm[i],tl);
+              }else{
+                if(!((i_regs->wasconst>>sl)&1))
+                  emit_xorimm(sl,imm[i],tl);
+                else
+                  emit_movimm(constmap[i][sl]^imm[i],tl);
+              }
             }
           }
           else {
@@ -2526,7 +2612,7 @@ void load_assemble(int i,struct regstat *i_regs)
   //printf("load_assemble: c=%d\n",c);
   //if(c) printf("load_assemble: const=%x\n",(int)constmap[i][s]+offset);
   // FIXME: Even if the load is a NOP, we should check for pagefaults...
-  if(tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80)
+  if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
     ||rt1[i]==0) {
       // could be FIFO, must perform the read
       // ||dummy read
@@ -2783,7 +2869,7 @@ void load_assemble(int i,struct regstat *i_regs)
     emit_call((int)memdebug);
     //emit_popa();
     restore_regs(0x100f);
-  }/**/
+  }*/
 }
 
 #ifndef loadlr_assemble
@@ -2799,7 +2885,7 @@ void store_assemble(int i,struct regstat *i_regs)
   int s,th,tl,map=-1;
   int addr,temp;
   int offset;
-  int jaddr=0,jaddr2,type;
+  int jaddr=0,type;
   int memtarget=0,c=0;
   int agr=AGEN1+(i&1);
   int faststore_reg_override=0;
@@ -2918,7 +3004,7 @@ void store_assemble(int i,struct regstat *i_regs)
       #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
       emit_callne(invalidate_addr_reg[addr]);
       #else
-      jaddr2=(int)out;
+      int jaddr2=(int)out;
       emit_jne(0);
       add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),addr,0,0,0);
       #endif
@@ -2979,18 +3065,18 @@ void store_assemble(int i,struct regstat *i_regs)
     #ifdef __arm__
     restore_regs(0x100f);
     #endif
-  }/**/
+  }*/
 }
 
 void storelr_assemble(int i,struct regstat *i_regs)
 {
   int s,th,tl;
   int temp;
-  int temp2;
+  int temp2=-1;
   int offset;
-  int jaddr=0,jaddr2;
-  int case1,case2,case3;
-  int done0,done1,done2;
+  int jaddr=0;
+  void *case1, *case2, *case3;
+  void *done0, *done1, *done2;
   int memtarget=0,c=0;
   int agr=AGEN1+(i&1);
   u_int hr,reglist=0;
@@ -3041,10 +3127,10 @@ void storelr_assemble(int i,struct regstat *i_regs)
     emit_xorimm(temp,3,temp);
 #endif
   emit_testimm(temp,2);
-  case2=(int)out;
+  case2=out;
   emit_jne(0);
   emit_testimm(temp,1);
-  case1=(int)out;
+  case1=out;
   emit_jne(0);
   // 0
   if (opcode[i]==0x2A) { // SWL
@@ -3061,10 +3147,10 @@ void storelr_assemble(int i,struct regstat *i_regs)
     emit_writebyte_indexed(tl,3,temp);
     if(rs2[i]) emit_shldimm(th,tl,24,temp2);
   }
-  done0=(int)out;
+  done0=out;
   emit_jmp(0);
   // 1
-  set_jump_target(case1,(int)out);
+  set_jump_target(case1, out);
   if (opcode[i]==0x2A) { // SWL
     // Write 3 msb into three least significant bytes
     if(rs2[i]) emit_rorimm(tl,8,tl);
@@ -3091,12 +3177,12 @@ void storelr_assemble(int i,struct regstat *i_regs)
     // Write two lsb into two most significant bytes
     emit_writehword_indexed(tl,1,temp);
   }
-  done1=(int)out;
+  done1=out;
   emit_jmp(0);
   // 2
-  set_jump_target(case2,(int)out);
+  set_jump_target(case2, out);
   emit_testimm(temp,1);
-  case3=(int)out;
+  case3=out;
   emit_jne(0);
   if (opcode[i]==0x2A) { // SWL
     // Write two msb into two least significant bytes
@@ -3126,10 +3212,10 @@ void storelr_assemble(int i,struct regstat *i_regs)
     emit_writehword_indexed(tl,0,temp);
     if(rs2[i]) emit_rorimm(tl,24,tl);
   }
-  done2=(int)out;
+  done2=out;
   emit_jmp(0);
   // 3
-  set_jump_target(case3,(int)out);
+  set_jump_target(case3, out);
   if (opcode[i]==0x2A) { // SWL
     // Write msb into least significant byte
     if(rs2[i]) emit_rorimm(tl,24,tl);
@@ -3152,24 +3238,24 @@ void storelr_assemble(int i,struct regstat *i_regs)
     // Write entire word
     emit_writeword_indexed(tl,-3,temp);
   }
-  set_jump_target(done0,(int)out);
-  set_jump_target(done1,(int)out);
-  set_jump_target(done2,(int)out);
+  set_jump_target(done0, out);
+  set_jump_target(done1, out);
+  set_jump_target(done2, out);
   if (opcode[i]==0x2C) { // SDL
     emit_testimm(temp,4);
-    done0=(int)out;
+    done0=out;
     emit_jne(0);
     emit_andimm(temp,~3,temp);
     emit_writeword_indexed(temp2,4,temp);
-    set_jump_target(done0,(int)out);
+    set_jump_target(done0, out);
   }
   if (opcode[i]==0x2D) { // SDR
     emit_testimm(temp,4);
-    done0=(int)out;
+    done0=out;
     emit_jeq(0);
     emit_andimm(temp,~3,temp);
     emit_writeword_indexed(temp2,-4,temp);
-    set_jump_target(done0,(int)out);
+    set_jump_target(done0, out);
   }
   if(!c||!memtarget)
     add_stub(STORELR_STUB,jaddr,(int)out,i,(int)i_regs,temp,ccadj[i],reglist);
@@ -3191,7 +3277,7 @@ void storelr_assemble(int i,struct regstat *i_regs)
     #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
     emit_callne(invalidate_addr_reg[temp]);
     #else
-    jaddr2=(int)out;
+    int jaddr2=(int)out;
     emit_jne(0);
     add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),temp,0,0,0);
     #endif
@@ -3208,7 +3294,7 @@ void storelr_assemble(int i,struct regstat *i_regs)
     emit_call((int)memdebug);
     emit_popa();
     //restore_regs(0x100f);
-  /**/
+  */
 }
 
 void c1ls_assemble(int i,struct regstat *i_regs)
@@ -3222,7 +3308,7 @@ void c2ls_assemble(int i,struct regstat *i_regs)
   int ar;
   int offset;
   int memtarget=0,c=0;
-  int jaddr2=0,jaddr3,type;
+  int jaddr2=0,type;
   int agr=AGEN1+(i&1);
   int fastio_reg_override=0;
   u_int hr,reglist=0;
@@ -3303,7 +3389,7 @@ void c2ls_assemble(int i,struct regstat *i_regs)
     #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
     emit_callne(invalidate_addr_reg[ar]);
     #else
-    jaddr3=(int)out;
+    int jaddr3=(int)out;
     emit_jne(0);
     add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),ar,0,0,0);
     #endif
@@ -3364,6 +3450,7 @@ void syscall_assemble(int i,struct regstat *i_regs)
   signed char ccreg=get_reg(i_regs->regmap,CCREG);
   assert(ccreg==HOST_CCREG);
   assert(!is_delayslot);
+  (void)ccreg;
   emit_movimm(start+i*4,EAX); // Get PC
   emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right?  There should probably be an extra cycle...
   emit_jmp((int)jump_syscall_hle); // XXX
@@ -3371,11 +3458,17 @@ void syscall_assemble(int i,struct regstat *i_regs)
 
 void hlecall_assemble(int i,struct regstat *i_regs)
 {
+  extern void psxNULL();
   signed char ccreg=get_reg(i_regs->regmap,CCREG);
   assert(ccreg==HOST_CCREG);
   assert(!is_delayslot);
+  (void)ccreg;
   emit_movimm(start+i*4+4,0); // Get PC
-  emit_movimm((int)psxHLEt[source[i]&7],1);
+  uint32_t hleCode = source[i] & 0x03ffffff;
+  if (hleCode >= (sizeof(psxHLEt) / sizeof(psxHLEt[0])))
+    emit_movimm((int)psxNULL,1);
+  else
+    emit_movimm((int)psxHLEt[hleCode],1);
   emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
   emit_jmp((int)jump_hlecall);
 }
@@ -3385,6 +3478,7 @@ void intcall_assemble(int i,struct regstat *i_regs)
   signed char ccreg=get_reg(i_regs->regmap,CCREG);
   assert(ccreg==HOST_CCREG);
   assert(!is_delayslot);
+  (void)ccreg;
   emit_movimm(start+i*4,0); // Get PC
   emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
   emit_jmp((int)jump_intcall);
@@ -3703,7 +3797,7 @@ void address_generation(int i,struct regstat *i_regs,signed char entry[])
   }
 }
 
-int get_final_value(int hr, int i, int *value)
+static int get_final_value(int hr, int i, int *value)
 {
   int reg=regs[i].regmap[hr];
   while(i<slen-1) {
@@ -4186,13 +4280,37 @@ int match_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
   return 1;
 }
 
+#ifdef DRC_DBG
+static void drc_dbg_emit_do_cmp(int i)
+{
+  extern void do_insn_cmp();
+  extern int cycle;
+  u_int hr,reglist=0;
+
+  for(hr=0;hr<HOST_REGS;hr++)
+    if(regs[i].regmap[hr]>=0) reglist|=1<<hr;
+  save_regs(reglist);
+  emit_movimm(start+i*4,0);
+  emit_writeword(0,(int)&pcaddr);
+  emit_call((int)do_insn_cmp);
+  //emit_readword((int)&cycle,0);
+  //emit_addimm(0,2,0);
+  //emit_writeword(0,(int)&cycle);
+  restore_regs(reglist);
+}
+#else
+#define drc_dbg_emit_do_cmp(x)
+#endif
+
 // Used when a branch jumps into the delay slot of another branch
 void ds_assemble_entry(int i)
 {
   int t=(ba[i]-start)>>2;
-  if(!instr_addr[t]) instr_addr[t]=(u_int)out;
+  if (!instr_addr[t])
+    instr_addr[t] = out;
   assem_debug("Assemble delay slot at %x\n",ba[i]);
   assem_debug("<->\n");
+  drc_dbg_emit_do_cmp(t);
   if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
     wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty,regs[t].was32);
   load_regs(regs[t].regmap_entry,regs[t].regmap,regs[t].was32,rs1[t],rs2[t]);
@@ -4318,7 +4436,7 @@ void do_ccstub(int n)
 {
   literal_pool(256);
   assem_debug("do_ccstub %x\n",start+stubs[n][4]*4);
-  set_jump_target(stubs[n][1],(int)out);
+  set_jump_target(stubs[n][1], out);
   int i=stubs[n][4];
   if(stubs[n][6]==NULLDS) {
     // Delay slot instruction is nullified ("likely" branch)
@@ -4584,7 +4702,7 @@ void do_ccstub(int n)
   emit_jmpreg(EAX);*/
 }
 
-add_to_linker(int addr,int target,int ext)
+static void add_to_linker(int addr,int target,int ext)
 {
   link_addr[linkcount][0]=addr;
   link_addr[linkcount][1]=target;
@@ -4616,12 +4734,12 @@ static void ujump_assemble_write_ra(int i)
       #ifdef REG_PREFETCH
       if(temp>=0)
       {
-        if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
+        if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table_get(return_address),temp);
       }
       #endif
       emit_movimm(return_address,rt); // PC into link register
       #ifdef IMM_PREFETCH
-      emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
+      emit_prefetch(hash_table_get(return_address));
       #endif
     }
   }
@@ -4629,7 +4747,6 @@ static void ujump_assemble_write_ra(int i)
 
 void ujump_assemble(int i,struct regstat *i_regs)
 {
-  signed char *i_regmap=i_regs->regmap;
   int ra_done=0;
   if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
   address_generation(i+1,i_regs,regs[i].regmap_entry);
@@ -4637,9 +4754,10 @@ void ujump_assemble(int i,struct regstat *i_regs)
   int temp=get_reg(branch_regs[i].regmap,PTEMP);
   if(rt1[i]==31&&temp>=0)
   {
+    signed char *i_regmap=i_regs->regmap;
     int return_address=start+i*4+8;
     if(get_reg(branch_regs[i].regmap,31)>0)
-    if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
+    if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table_get(return_address),temp);
   }
   #endif
   if(rt1[i]==31&&(rt1[i]==rs1[i+1]||rt1[i]==rs2[i+1])) {
@@ -4691,20 +4809,19 @@ static void rjump_assemble_write_ra(int i)
   #ifdef REG_PREFETCH
   if(temp>=0)
   {
-    if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
+    if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table_get(return_address),temp);
   }
   #endif
   emit_movimm(return_address,rt); // PC into link register
   #ifdef IMM_PREFETCH
-  emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
+  emit_prefetch(hash_table_get(return_address));
   #endif
 }
 
 void rjump_assemble(int i,struct regstat *i_regs)
 {
-  signed char *i_regmap=i_regs->regmap;
   int temp;
-  int rs,cc,adj;
+  int rs,cc;
   int ra_done=0;
   rs=get_reg(branch_regs[i].regmap,rs1[i]);
   assert(rs>=0);
@@ -4721,8 +4838,9 @@ void rjump_assemble(int i,struct regstat *i_regs)
   if(rt1[i]==31)
   {
     if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
+      signed char *i_regmap=i_regs->regmap;
       int return_address=start+i*4+8;
-      if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
+      if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table_get(return_address),temp);
     }
   }
   #endif
@@ -4749,6 +4867,7 @@ void rjump_assemble(int i,struct regstat *i_regs)
     rjump_assemble_write_ra(i);
   cc=get_reg(branch_regs[i].regmap,CCREG);
   assert(cc==HOST_CCREG);
+  (void)cc;
   #ifdef USE_MINI_HT
   int rh=get_reg(branch_regs[i].regmap,RHASH);
   int ht=get_reg(branch_regs[i].regmap,RHTBL);
@@ -4937,7 +5056,7 @@ void cjump_assemble(int i,struct regstat *i_regs)
       add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
     }
     else {
-      int taken=0,nottaken=0,nottaken1=0;
+      void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
       do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
       if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
       if(!only32)
@@ -4947,32 +5066,32 @@ void cjump_assemble(int i,struct regstat *i_regs)
         {
           if(s2h>=0) emit_cmp(s1h,s2h);
           else emit_test(s1h,s1h);
-          nottaken1=(int)out;
+          nottaken1=out;
           emit_jne(1);
         }
         if(opcode[i]==5) // BNE
         {
           if(s2h>=0) emit_cmp(s1h,s2h);
           else emit_test(s1h,s1h);
-          if(invert) taken=(int)out;
+          if(invert) taken=out;
           else add_to_linker((int)out,ba[i],internal);
           emit_jne(0);
         }
         if(opcode[i]==6) // BLEZ
         {
           emit_test(s1h,s1h);
-          if(invert) taken=(int)out;
+          if(invert) taken=out;
           else add_to_linker((int)out,ba[i],internal);
           emit_js(0);
-          nottaken1=(int)out;
+          nottaken1=out;
           emit_jne(1);
         }
         if(opcode[i]==7) // BGTZ
         {
           emit_test(s1h,s1h);
-          nottaken1=(int)out;
+          nottaken1=out;
           emit_js(1);
-          if(invert) taken=(int)out;
+          if(invert) taken=out;
           else add_to_linker((int)out,ba[i],internal);
           emit_jne(0);
         }
@@ -4985,7 +5104,7 @@ void cjump_assemble(int i,struct regstat *i_regs)
         if(s2l>=0) emit_cmp(s1l,s2l);
         else emit_test(s1l,s1l);
         if(invert){
-          nottaken=(int)out;
+          nottaken=out;
           emit_jne(1);
         }else{
           add_to_linker((int)out,ba[i],internal);
@@ -4997,7 +5116,7 @@ void cjump_assemble(int i,struct regstat *i_regs)
         if(s2l>=0) emit_cmp(s1l,s2l);
         else emit_test(s1l,s1l);
         if(invert){
-          nottaken=(int)out;
+          nottaken=out;
           emit_jeq(1);
         }else{
           add_to_linker((int)out,ba[i],internal);
@@ -5008,7 +5127,7 @@ void cjump_assemble(int i,struct regstat *i_regs)
       {
         emit_cmpimm(s1l,1);
         if(invert){
-          nottaken=(int)out;
+          nottaken=out;
           emit_jge(1);
         }else{
           add_to_linker((int)out,ba[i],internal);
@@ -5019,7 +5138,7 @@ void cjump_assemble(int i,struct regstat *i_regs)
       {
         emit_cmpimm(s1l,1);
         if(invert){
-          nottaken=(int)out;
+          nottaken=out;
           emit_jl(1);
         }else{
           add_to_linker((int)out,ba[i],internal);
@@ -5027,7 +5146,7 @@ void cjump_assemble(int i,struct regstat *i_regs)
         }
       }
       if(invert) {
-        if(taken) set_jump_target(taken,(int)out);
+        if(taken) set_jump_target(taken, out);
         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
         if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
           if(adj) {
@@ -5056,10 +5175,10 @@ void cjump_assemble(int i,struct regstat *i_regs)
             emit_jmp(0);
           }
         }
-        set_jump_target(nottaken,(int)out);
+        set_jump_target(nottaken, out);
       }
 
-      if(nottaken1) set_jump_target(nottaken1,(int)out);
+      if(nottaken1) set_jump_target(nottaken1, out);
       if(adj) {
         if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
       }
@@ -5071,7 +5190,7 @@ void cjump_assemble(int i,struct regstat *i_regs)
     //if(likely[i]) printf("IOL\n");
     //else
     //printf("IOE\n");
-    int taken=0,nottaken=0,nottaken1=0;
+    void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
     if(!unconditional&&!nop) {
       if(!only32)
       {
@@ -5080,30 +5199,30 @@ void cjump_assemble(int i,struct regstat *i_regs)
         {
           if(s2h>=0) emit_cmp(s1h,s2h);
           else emit_test(s1h,s1h);
-          nottaken1=(int)out;
+          nottaken1=out;
           emit_jne(2);
         }
         if((opcode[i]&0x2f)==5) // BNE
         {
           if(s2h>=0) emit_cmp(s1h,s2h);
           else emit_test(s1h,s1h);
-          taken=(int)out;
+          taken=out;
           emit_jne(1);
         }
         if((opcode[i]&0x2f)==6) // BLEZ
         {
           emit_test(s1h,s1h);
-          taken=(int)out;
+          taken=out;
           emit_js(1);
-          nottaken1=(int)out;
+          nottaken1=out;
           emit_jne(2);
         }
         if((opcode[i]&0x2f)==7) // BGTZ
         {
           emit_test(s1h,s1h);
-          nottaken1=(int)out;
+          nottaken1=out;
           emit_js(2);
-          taken=(int)out;
+          taken=out;
           emit_jne(1);
         }
       } // if(!only32)
@@ -5114,26 +5233,26 @@ void cjump_assemble(int i,struct regstat *i_regs)
       {
         if(s2l>=0) emit_cmp(s1l,s2l);
         else emit_test(s1l,s1l);
-        nottaken=(int)out;
+        nottaken=out;
         emit_jne(2);
       }
       if((opcode[i]&0x2f)==5) // BNE
       {
         if(s2l>=0) emit_cmp(s1l,s2l);
         else emit_test(s1l,s1l);
-        nottaken=(int)out;
+        nottaken=out;
         emit_jeq(2);
       }
       if((opcode[i]&0x2f)==6) // BLEZ
       {
         emit_cmpimm(s1l,1);
-        nottaken=(int)out;
+        nottaken=out;
         emit_jge(2);
       }
       if((opcode[i]&0x2f)==7) // BGTZ
       {
         emit_cmpimm(s1l,1);
-        nottaken=(int)out;
+        nottaken=out;
         emit_jl(2);
       }
     } // if(!unconditional)
@@ -5147,7 +5266,7 @@ void cjump_assemble(int i,struct regstat *i_regs)
     ds_unneeded_upper|=1;
     // branch taken
     if(!nop) {
-      if(taken) set_jump_target(taken,(int)out);
+      if(taken) set_jump_target(taken, out);
       assem_debug("1:\n");
       wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
                     ds_unneeded,ds_unneeded_upper);
@@ -5182,8 +5301,8 @@ void cjump_assemble(int i,struct regstat *i_regs)
     // branch not taken
     cop1_usable=prev_cop1_usable;
     if(!unconditional) {
-      if(nottaken1) set_jump_target(nottaken1,(int)out);
-      set_jump_target(nottaken,(int)out);
+      if(nottaken1) set_jump_target(nottaken1, out);
+      set_jump_target(nottaken, out);
       assem_debug("2:\n");
       if(!likely[i]) {
         wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
@@ -5284,7 +5403,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
         return_address=start+i*4+8;
         emit_movimm(return_address,rt); // PC into link register
         #ifdef IMM_PREFETCH
-        if(!nevertaken) emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
+        if(!nevertaken) emit_prefetch(hash_table_get(return_address));
         #endif
       }
     }
@@ -5322,7 +5441,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
       add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
     }
     else {
-      int nottaken=0;
+      void *nottaken = NULL;
       do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
       if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
       if(!only32)
@@ -5332,7 +5451,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
         {
           emit_test(s1h,s1h);
           if(invert){
-            nottaken=(int)out;
+            nottaken=out;
             emit_jns(1);
           }else{
             add_to_linker((int)out,ba[i],internal);
@@ -5343,7 +5462,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
         {
           emit_test(s1h,s1h);
           if(invert){
-            nottaken=(int)out;
+            nottaken=out;
             emit_js(1);
           }else{
             add_to_linker((int)out,ba[i],internal);
@@ -5358,7 +5477,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
         {
           emit_test(s1l,s1l);
           if(invert){
-            nottaken=(int)out;
+            nottaken=out;
             emit_jns(1);
           }else{
             add_to_linker((int)out,ba[i],internal);
@@ -5369,7 +5488,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
         {
           emit_test(s1l,s1l);
           if(invert){
-            nottaken=(int)out;
+            nottaken=out;
             emit_js(1);
           }else{
             add_to_linker((int)out,ba[i],internal);
@@ -5407,7 +5526,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
             emit_jmp(0);
           }
         }
-        set_jump_target(nottaken,(int)out);
+        set_jump_target(nottaken, out);
       }
 
       if(adj) {
@@ -5419,7 +5538,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
   {
     // In-order execution (branch first)
     //printf("IOE\n");
-    int nottaken=0;
+    void *nottaken = NULL;
     if(rt1[i]==31) {
       int rt,return_address;
       rt=get_reg(branch_regs[i].regmap,31);
@@ -5428,7 +5547,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
         return_address=start+i*4+8;
         emit_movimm(return_address,rt); // PC into link register
         #ifdef IMM_PREFETCH
-        emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
+        emit_prefetch(hash_table_get(return_address));
         #endif
       }
     }
@@ -5440,13 +5559,13 @@ void sjump_assemble(int i,struct regstat *i_regs)
         if((opcode2[i]&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
         {
           emit_test(s1h,s1h);
-          nottaken=(int)out;
+          nottaken=out;
           emit_jns(1);
         }
         if((opcode2[i]&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
         {
           emit_test(s1h,s1h);
-          nottaken=(int)out;
+          nottaken=out;
           emit_js(1);
         }
       } // if(!only32)
@@ -5456,13 +5575,13 @@ void sjump_assemble(int i,struct regstat *i_regs)
         if((opcode2[i]&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
         {
           emit_test(s1l,s1l);
-          nottaken=(int)out;
+          nottaken=out;
           emit_jns(1);
         }
         if((opcode2[i]&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
         {
           emit_test(s1l,s1l);
-          nottaken=(int)out;
+          nottaken=out;
           emit_js(1);
         }
       }
@@ -5511,7 +5630,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
     // branch not taken
     cop1_usable=prev_cop1_usable;
     if(!unconditional) {
-      set_jump_target(nottaken,(int)out);
+      set_jump_target(nottaken, out);
       assem_debug("1:\n");
       if(!likely[i]) {
         wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
@@ -5599,7 +5718,7 @@ void fjump_assemble(int i,struct regstat *i_regs)
     do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
     assem_debug("cycle count (adj)\n");
     if(1) {
-      int nottaken=0;
+      void *nottaken = NULL;
       if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
       if(1) {
         assert(fs>=0);
@@ -5607,7 +5726,7 @@ void fjump_assemble(int i,struct regstat *i_regs)
         if(source[i]&0x10000) // BC1T
         {
           if(invert){
-            nottaken=(int)out;
+            nottaken=out;
             emit_jeq(1);
           }else{
             add_to_linker((int)out,ba[i],internal);
@@ -5616,7 +5735,7 @@ void fjump_assemble(int i,struct regstat *i_regs)
         }
         else // BC1F
           if(invert){
-            nottaken=(int)out;
+            nottaken=out;
             emit_jne(1);
           }else{
             add_to_linker((int)out,ba[i],internal);
@@ -5644,7 +5763,7 @@ void fjump_assemble(int i,struct regstat *i_regs)
           add_to_linker((int)out,ba[i],internal);
           emit_jmp(0);
         }
-        set_jump_target(nottaken,(int)out);
+        set_jump_target(nottaken, out);
       }
 
       if(adj) {
@@ -5656,7 +5775,7 @@ void fjump_assemble(int i,struct regstat *i_regs)
   {
     // In-order execution (branch first)
     //printf("IOE\n");
-    int nottaken=0;
+    void *nottaken = NULL;
     if(1) {
       //printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
       if(1) {
@@ -5664,12 +5783,12 @@ void fjump_assemble(int i,struct regstat *i_regs)
         emit_testimm(fs,0x800000);
         if(source[i]&0x10000) // BC1T
         {
-          nottaken=(int)out;
+          nottaken=out;
           emit_jeq(1);
         }
         else // BC1F
         {
-          nottaken=(int)out;
+          nottaken=out;
           emit_jne(1);
         }
       }
@@ -5716,7 +5835,7 @@ void fjump_assemble(int i,struct regstat *i_regs)
 
     // branch not taken
     if(1) { // <- FIXME (don't need this)
-      set_jump_target(nottaken,(int)out);
+      set_jump_target(nottaken, out);
       assem_debug("1:\n");
       if(!likely[i]) {
         wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
@@ -5754,9 +5873,8 @@ static void pagespan_assemble(int i,struct regstat *i_regs)
   int s1h=get_reg(i_regs->regmap,rs1[i]|64);
   int s2l=get_reg(i_regs->regmap,rs2[i]);
   int s2h=get_reg(i_regs->regmap,rs2[i]|64);
-  void *nt_branch=NULL;
-  int taken=0;
-  int nottaken=0;
+  void *taken = NULL;
+  void *nottaken = NULL;
   int unconditional=0;
   if(rs1[i]==0)
   {
@@ -5771,7 +5889,7 @@ static void pagespan_assemble(int i,struct regstat *i_regs)
     s1h=s2h=-1;
   }
   int hr=0;
-  int addr,alt,ntaddr;
+  int addr=-1,alt=-1,ntaddr=-1;
   if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
   else {
     while(hr<HOST_REGS)
@@ -5889,13 +6007,13 @@ static void pagespan_assemble(int i,struct regstat *i_regs)
     if(s1h>=0) {
       if(s2h>=0) emit_cmp(s1h,s2h);
       else emit_test(s1h,s1h);
-      nottaken=(int)out;
+      nottaken=out;
       emit_jne(0);
     }
     if(s2l>=0) emit_cmp(s1l,s2l);
     else emit_test(s1l,s1l);
-    if(nottaken) set_jump_target(nottaken,(int)out);
-    nottaken=(int)out;
+    if(nottaken) set_jump_target(nottaken, out);
+    nottaken=out;
     emit_jne(0);
   }
   if((opcode[i]&0x3f)==0x15) // BNEL
@@ -5903,14 +6021,14 @@ static void pagespan_assemble(int i,struct regstat *i_regs)
     if(s1h>=0) {
       if(s2h>=0) emit_cmp(s1h,s2h);
       else emit_test(s1h,s1h);
-      taken=(int)out;
+      taken=out;
       emit_jne(0);
     }
     if(s2l>=0) emit_cmp(s1l,s2l);
     else emit_test(s1l,s1l);
-    nottaken=(int)out;
+    nottaken=out;
     emit_jeq(0);
-    if(taken) set_jump_target(taken,(int)out);
+    if(taken) set_jump_target(taken, out);
   }
   if((opcode[i]&0x3f)==6) // BLEZ
   {
@@ -5963,13 +6081,13 @@ static void pagespan_assemble(int i,struct regstat *i_regs)
     if((source[i]&0x30000)==0x20000) // BC1FL
     {
       emit_testimm(s1l,0x800000);
-      nottaken=(int)out;
+      nottaken=out;
       emit_jne(0);
     }
     if((source[i]&0x30000)==0x30000) // BC1TL
     {
       emit_testimm(s1l,0x800000);
-      nottaken=(int)out;
+      nottaken=out;
       emit_jeq(0);
     }
   }
@@ -5991,13 +6109,13 @@ static void pagespan_assemble(int i,struct regstat *i_regs)
   void *compiled_target_addr=check_addr(target_addr);
   emit_extjump_ds((int)branch_addr,target_addr);
   if(compiled_target_addr) {
-    set_jump_target((int)branch_addr,(int)compiled_target_addr);
+    set_jump_target(branch_addr, compiled_target_addr);
     add_link(target_addr,stub);
   }
-  else set_jump_target((int)branch_addr,(int)stub);
+  else set_jump_target(branch_addr, stub);
   if(likely[i]) {
     // Not-taken path
-    set_jump_target((int)nottaken,(int)out);
+    set_jump_target(nottaken, out);
     wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
     void *branch_addr=out;
     emit_jmp(0);
@@ -6006,10 +6124,10 @@ static void pagespan_assemble(int i,struct regstat *i_regs)
     void *compiled_target_addr=check_addr(target_addr);
     emit_extjump_ds((int)branch_addr,target_addr);
     if(compiled_target_addr) {
-      set_jump_target((int)branch_addr,(int)compiled_target_addr);
+      set_jump_target(branch_addr, compiled_target_addr);
       add_link(target_addr,stub);
     }
-    else set_jump_target((int)branch_addr,(int)stub);
+    else set_jump_target(branch_addr, stub);
   }
 }
 
@@ -6097,11 +6215,11 @@ static void pagespan_ds()
 #else
   emit_cmpimm(btaddr,start+4);
 #endif
-  int branch=(int)out;
+  void *branch = out;
   emit_jeq(0);
   store_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,-1);
   emit_jmp(jump_vaddr_reg[btaddr]);
-  set_jump_target(branch,(int)out);
+  set_jump_target(branch, out);
   store_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,start+4);
   load_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,start+4);
 }
@@ -6764,7 +6882,7 @@ void clean_registers(int istart,int iend,int wr)
               if(r!=EXCLUDE_REG) {
                 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
                   regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
-                }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);/*assert(!((wont_dirty_i>>r)&1));*/}
+                }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
               }
             }
           }
@@ -6776,7 +6894,7 @@ void clean_registers(int istart,int iend,int wr)
               if(r!=EXCLUDE_REG) {
                 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
                   regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
-                }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);/*assert(!((wont_dirty_i>>r)&1));*/}
+                }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
               }
             }
           }
@@ -6819,7 +6937,7 @@ void clean_registers(int istart,int iend,int wr)
             wont_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
           } else {
             wont_dirty_i|=1<<r;
-            /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);/*assert(!((will_dirty>>r)&1));*/
+            /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
           }
         }
       }
@@ -6925,13 +7043,14 @@ static void disassemble_inst(int i) {}
 static int new_dynarec_test(void)
 {
   int (*testfunc)(void) = (void *)out;
+  void *beginning;
   int ret;
+
+  beginning = start_block();
   emit_movimm(DRC_TEST_VAL,0); // test
   emit_jmpreg(14);
   literal_pool(0);
-#ifdef __arm__
-  __clear_cache((void *)testfunc, out);
-#endif
+  end_block(beginning);
   SysPrintf("testing if we can run recompiled code..\n");
   ret = testfunc();
   if (ret == DRC_TEST_VAL)
@@ -6968,20 +7087,43 @@ void new_dynarec_clear_full()
 void new_dynarec_init()
 {
   SysPrintf("Init new dynarec\n");
-  out=(u_char *)BASE_ADDR;
-#if BASE_ADDR_FIXED
-  if (mmap (out, 1<<TARGET_SIZE_2,
+
+  // allocate/prepare a buffer for translation cache
+  // see assem_arm.h for some explanation
+#if   defined(BASE_ADDR_FIXED)
+  if (mmap (translation_cache, 1 << TARGET_SIZE_2,
+            PROT_READ | PROT_WRITE | PROT_EXEC,
+            MAP_PRIVATE | MAP_ANONYMOUS,
+            -1, 0) != translation_cache) {
+    SysPrintf("mmap() failed: %s\n", strerror(errno));
+    SysPrintf("disable BASE_ADDR_FIXED and recompile\n");
+    abort();
+  }
+#elif defined(BASE_ADDR_DYNAMIC)
+  #ifdef VITA
+  sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
+  if (sceBlock < 0)
+    SysPrintf("sceKernelAllocMemBlockForVM failed\n");
+  int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&translation_cache);
+  if (ret < 0)
+    SysPrintf("sceKernelGetMemBlockBase failed\n");
+  #else
+  translation_cache = mmap (NULL, 1 << TARGET_SIZE_2,
             PROT_READ | PROT_WRITE | PROT_EXEC,
-            MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
-            -1, 0) <= 0) {
+            MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+  if (translation_cache == MAP_FAILED) {
     SysPrintf("mmap() failed: %s\n", strerror(errno));
+    abort();
   }
+  #endif
 #else
+  #ifndef NO_WRITE_EXEC
   // not all systems allow execute in data segment by default
-  if (mprotect(out, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
+  if (mprotect((void *)BASE_ADDR, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
     SysPrintf("mprotect() failed: %s\n", strerror(errno));
+  #endif
 #endif
-  int n;
+  out=(u_char *)BASE_ADDR;
   cycle_multiplier=200;
   new_dynarec_clear_full();
 #ifdef HOST_IMM8
@@ -7000,9 +7142,15 @@ void new_dynarec_init()
 void new_dynarec_cleanup()
 {
   int n;
-  #if BASE_ADDR_FIXED
-  if (munmap ((void *)BASE_ADDR, 1<<TARGET_SIZE_2) < 0) {SysPrintf("munmap() failed\n");}
+#if defined(BASE_ADDR_FIXED) || defined(BASE_ADDR_DYNAMIC)
+  #ifdef VITA
+  sceKernelFreeMemBlock(sceBlock);
+  sceBlock = -1;
+  #else
+  if (munmap ((void *)BASE_ADDR, 1<<TARGET_SIZE_2) < 0)
+    SysPrintf("munmap() failed\n");
   #endif
+#endif
   for(n=0;n<4096;n++) ll_clear(jump_in+n);
   for(n=0;n<4096;n++) ll_clear(jump_out+n);
   for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
@@ -7030,6 +7178,7 @@ static u_int *get_source_start(u_int addr, u_int *limit)
     *limit = (addr & 0x80600000) + 0x00200000;
     return (u_int *)((u_int)rdram + (addr&0x1fffff));
   }
+  return NULL;
 }
 
 static u_int scan_for_ret(u_int addr)
@@ -7047,6 +7196,7 @@ static u_int scan_for_ret(u_int addr)
     if (*mem == 0x03e00008) // jr $ra
       return addr + 8;
   }
+  return addr;
 }
 
 struct savestate_block {
@@ -7160,16 +7310,15 @@ int new_recompile_block(int addr)
   if (Config.HLE && start == 0x80001000) // hlecall
   {
     // XXX: is this enough? Maybe check hleSoftCall?
-    u_int beginning=(u_int)out;
+    void *beginning=start_block();
     u_int page=get_page(start);
+
     invalid_code[start>>12]=0;
     emit_movimm(start,0);
     emit_writeword(0,(int)&pcaddr);
     emit_jmp((int)new_dyna_leave);
     literal_pool(0);
-#ifdef __arm__
-    __clear_cache((void *)beginning,out);
-#endif
+    end_block(beginning);
     ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
     return 0;
   }
@@ -8352,7 +8501,7 @@ int new_recompile_block(int addr)
       // Create entry (branch target) regmap
       for(hr=0;hr<HOST_REGS;hr++)
       {
-        int r,or,er;
+        int r,or;
         r=current.regmap[hr];
         if(r>=0) {
           if(r!=regmap_pre[i][hr]) {
@@ -9871,17 +10020,17 @@ int new_recompile_block(int addr)
   cop1_usable=0;
   uint64_t is32_pre=0;
   u_int dirty_pre=0;
-  u_int beginning=(u_int)out;
+  void *beginning=start_block();
   if((u_int)addr&1) {
     ds=1;
     pagespan_ds();
   }
-  u_int instr_addr0_override=0;
+  void *instr_addr0_override = NULL;
 
   if (start == 0x80030000) {
     // nasty hack for fastbios thing
     // override block entry to this code
-    instr_addr0_override=(u_int)out;
+    instr_addr0_override = out;
     emit_movimm(start,0);
     // abuse io address var as a flag that we
     // have already returned here once
@@ -9898,7 +10047,7 @@ int new_recompile_block(int addr)
     if(ds) {
       ds=0; // Skip delay slot
       if(bt[i]) assem_debug("OOPS - branch into delay slot\n");
-      instr_addr[i]=0;
+      instr_addr[i] = NULL;
     } else {
       speculate_register_values(i);
       #ifndef DESTRUCTIVE_WRITEBACK
@@ -9923,8 +10072,10 @@ int new_recompile_block(int addr)
         loop_preload(regmap_pre[i],regs[i].regmap_entry);
       }
       // branch target entry point
-      instr_addr[i]=(u_int)out;
+      instr_addr[i] = out;
       assem_debug("<->\n");
+      drc_dbg_emit_do_cmp(i);
+
       // load regs
       if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
         wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty,regs[i].was32);
@@ -10104,10 +10255,10 @@ int new_recompile_block(int addr)
       void *addr=check_addr(link_addr[i][1]);
       emit_extjump(link_addr[i][0],link_addr[i][1]);
       if(addr) {
-        set_jump_target(link_addr[i][0],(int)addr);
+        set_jump_target(link_addr[i][0], addr);
         add_link(link_addr[i][1],stub);
       }
-      else set_jump_target(link_addr[i][0],(int)stub);
+      else set_jump_target(link_addr[i][0], stub);
     }
     else
     {
@@ -10135,22 +10286,20 @@ int new_recompile_block(int addr)
         u_int vpage=get_vpage(vaddr);
         literal_pool(256);
         {
-          assem_debug("%8x (%d) <- %8x\n",instr_addr[i],i,start+i*4);
+          assem_debug("%p (%d) <- %8x\n",instr_addr[i],i,start+i*4);
           assem_debug("jump_in: %x\n",start+i*4);
-          ll_add(jump_dirty+vpage,vaddr,(void *)out);
-          int entry_point=do_dirty_stub(i);
-          ll_add_flags(jump_in+page,vaddr,state_rflags,(void *)entry_point);
+          ll_add(jump_dirty+vpage,vaddr,out);
+          void *entry_point = do_dirty_stub(i);
+          ll_add_flags(jump_in+page,vaddr,state_rflags,entry_point);
           // If there was an existing entry in the hash table,
           // replace it with the new address.
           // Don't add new entries.  We'll insert the
           // ones that actually get used in check_addr().
-          int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
-          if(ht_bin[0]==vaddr) {
-            ht_bin[1]=entry_point;
-          }
-          if(ht_bin[2]==vaddr) {
-            ht_bin[3]=entry_point;
-          }
+          struct ht_entry *ht_bin = hash_table_get(vaddr);
+          if (ht_bin->vaddr[0] == vaddr)
+            ht_bin->tcaddr[0] = entry_point;
+          if (ht_bin->vaddr[1] == vaddr)
+            ht_bin->tcaddr[1] = entry_point;
         }
       }
     }
@@ -10161,14 +10310,12 @@ int new_recompile_block(int addr)
   // Align code
   if(((u_int)out)&7) emit_addnop(13);
   #endif
-  assert((u_int)out-beginning<MAX_OUTPUT_BLOCK_SIZE);
+  assert((u_int)out-(u_int)beginning<MAX_OUTPUT_BLOCK_SIZE);
   //printf("shadow buffer: %x-%x\n",(int)copy,(int)copy+slen*4);
   memcpy(copy,source,slen*4);
   copy+=slen*4;
 
-  #ifdef __arm__
-  __clear_cache((void *)beginning,out);
-  #endif
+  end_block(beginning);
 
   // If we're within 256K of the end of the buffer,
   // start over from the beginning. (Is 256K enough?)
@@ -10193,7 +10340,7 @@ int new_recompile_block(int addr)
   while(expirep!=end)
   {
     int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
-    int base=(int)BASE_ADDR+((expirep>>13)<<shift); // Base address of this block
+    uintptr_t base=(uintptr_t)BASE_ADDR+((expirep>>13)<<shift); // Base address of this block
     inv_debug("EXP: Phase %d\n",expirep);
     switch((expirep>>11)&3)
     {
@@ -10212,18 +10359,20 @@ int new_recompile_block(int addr)
       case 2:
         // Clear hash table
         for(i=0;i<32;i++) {
-          int *ht_bin=hash_table[((expirep&2047)<<5)+i];
-          if((ht_bin[3]>>shift)==(base>>shift) ||
-             ((ht_bin[3]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
-            inv_debug("EXP: Remove hash %x -> %x\n",ht_bin[2],ht_bin[3]);
-            ht_bin[2]=ht_bin[3]=-1;
-          }
-          if((ht_bin[1]>>shift)==(base>>shift) ||
-             ((ht_bin[1]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
-            inv_debug("EXP: Remove hash %x -> %x\n",ht_bin[0],ht_bin[1]);
-            ht_bin[0]=ht_bin[2];
-            ht_bin[1]=ht_bin[3];
-            ht_bin[2]=ht_bin[3]=-1;
+          struct ht_entry *ht_bin = &hash_table[((expirep&2047)<<5)+i];
+          if (((uintptr_t)ht_bin->tcaddr[1]>>shift) == (base>>shift) ||
+             (((uintptr_t)ht_bin->tcaddr[1]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
+            inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[1],ht_bin->tcaddr[1]);
+            ht_bin->vaddr[1] = -1;
+            ht_bin->tcaddr[1] = NULL;
+          }
+          if (((uintptr_t)ht_bin->tcaddr[0]>>shift) == (base>>shift) ||
+             (((uintptr_t)ht_bin->tcaddr[0]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
+            inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[0],ht_bin->tcaddr[0]);
+            ht_bin->vaddr[0] = ht_bin->vaddr[1];
+            ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
+            ht_bin->vaddr[1] = -1;
+            ht_bin->tcaddr[1] = NULL;
           }
         }
         break;