X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fnew_dynarec.c;h=518d07165cfcba25e037049132cf7a963a682162;hb=4bdc30ab36281e9f9934efb026e264def03cef46;hp=955642308768cfcbffb07f69d750444f1e82779f;hpb=53358c1d5b032cc7186b71e3cc14f0ad6c2d5468;p=pcsx_rearmed.git

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index 95564230..518d0716 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -50,6 +50,7 @@
 
 //#define DISASM
 //#define ASSEM_PRINT
+//#define STAT_PRINT
 
 #ifdef ASSEM_PRINT
 #define assem_debug printf
@@ -191,11 +192,11 @@ static struct decoded_insn
 } dops[MAXBLOCK];
 
   // 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 u_char *out;
+  static struct ll_entry *jump_dirty[4096];
   static struct ll_entry *jump_out[4096];
   static u_int start;
   static u_int *source;
@@ -235,6 +236,19 @@ static struct decoded_insn
   static int expirep;
   static u_int stop_after_jal;
   static u_int f1_hack;
+#ifdef STAT_PRINT
+  static int stat_bc_direct;
+  static int stat_bc_pre;
+  static int stat_bc_restore;
+  static int stat_jump_in_lookups;
+  static int stat_restore_tries;
+  static int stat_restore_compares;
+  static int stat_inv_addr_calls;
+  static int stat_inv_hits;
+  #define stat_inc(s) s++
+#else
+  #define stat_inc(s)
+#endif
 
   int new_dynarec_hacks;
   int new_dynarec_hacks_pergame;
@@ -250,7 +264,6 @@ static struct decoded_insn
   extern int branch_target;
   extern uintptr_t ram_offset;
   extern uintptr_t mini_ht[32][2];
-  extern u_char restore_candidate[512];
 
   /* registers that may be allocated */
   /* 1-31 gpr */
@@ -301,7 +314,7 @@ static struct decoded_insn
 //#define FCOMP 21  // Floating point compare (sets FSREG)
 #define SYSCALL 22// SYSCALL,BREAK
 #define OTHER 23  // Other
-#define SPAN 24   // Branch/delay slot spans 2 pages
+//#define SPAN 24   // Branch/delay slot spans 2 pages
 #define NI 25     // Not implemented
 #define HLECALL 26// PCSX fake opcodes for HLE
 #define COP2 27   // Coprocessor 2 move
@@ -322,11 +335,8 @@ int new_recompile_block(u_int addr);
 void *get_addr_ht(u_int vaddr);
 void invalidate_block(u_int block);
 void invalidate_addr(u_int addr);
-void remove_hash(int vaddr);
 void dyna_linker();
-void dyna_linker_ds();
 void verify_code();
-void verify_code_ds();
 void cc_interrupt();
 void fp_exception();
 void fp_exception_ds();
@@ -336,10 +346,13 @@ void jump_break   (u_int u0, u_int u1, u_int pc);
 void jump_break_ds(u_int u0, u_int u1, u_int pc);
 void jump_to_new_pc();
 void call_gteStall();
-void clean_blocks(u_int page);
 void add_jump_out(u_int vaddr, void *src);
 void new_dyna_leave();
 
+static void *get_clean_addr(void *addr);
+static void get_bounds(void *addr, u_char **start, u_char **end);
+static void ll_add_flags(struct ll_entry **head,int vaddr,u_int reg_sv_flags,void *addr);
+
 // Needed by assembler
 static void wb_register(signed char r, const signed char regmap[], uint64_t dirty);
 static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty);
@@ -531,6 +544,21 @@ static void hash_table_add(struct ht_entry *ht_bin, u_int vaddr, void *tcaddr)
   ht_bin->tcaddr[0] = tcaddr;
 }
 
+static void mark_valid_code(u_int vaddr, u_int len)
+{
+  u_int i, j;
+  vaddr &= 0x1fffffff;
+  for (i = vaddr & ~0xfff; i < vaddr + len; i += 0x1000) {
+    // ram mirrors, but should not hurt bios
+    for (j = 0; j < 0x800000; j += 0x200000) {
+      invalid_code[(i|j) >> 12] =
+      invalid_code[(i|j|0x80000000u) >> 12] =
+      invalid_code[(i|j|0xa0000000u) >> 12] = 0;
+    }
+  }
+  inv_code_start = inv_code_end = ~0;
+}
+
 // some messy ari64's code, seems to rely on unsigned 32bit overflow
 static int doesnt_expire_soon(void *tcaddr)
 {
@@ -538,51 +566,94 @@ static int doesnt_expire_soon(void *tcaddr)
   return diff > (u_int)(0x60000000 + (MAX_OUTPUT_BLOCK_SIZE << (32-TARGET_SIZE_2)));
 }
 
+void *ndrc_try_restore_block(u_int vaddr)
+{
+  u_char *source_start = NULL, *source_end = NULL;
+  void *found_stub = NULL, *found_clean = NULL;
+  u_int len, page = get_page(vaddr);
+  const struct ll_entry *head;
+  int ep_count = 0;
+
+  stat_inc(stat_restore_tries);
+  for (head = jump_dirty[page]; head != NULL; head = head->next)
+  {
+    if (head->vaddr != vaddr)
+      continue;
+    // don't restore blocks which are about to expire from the cache
+    if (!doesnt_expire_soon(head->addr))
+      continue;
+    stat_inc(stat_restore_compares);
+    if (!verify_dirty(head->addr))
+      continue;
+
+    found_stub = head->addr;
+    break;
+  }
+  if (!found_stub)
+    return NULL;
+
+  found_clean = get_clean_addr(found_stub);
+  get_bounds(found_stub, &source_start, &source_end);
+  assert(source_start < source_end);
+  len = source_end - source_start;
+  mark_valid_code(vaddr, len);
+
+  // restore all entry points
+  for (head = jump_dirty[page]; head != NULL; head = head->next)
+  {
+    if (head->vaddr < vaddr || head->vaddr >= vaddr + len)
+      continue;
+
+    u_char *start = NULL, *end = NULL;
+    get_bounds(head->addr, &start, &end);
+    if (start != source_start || end != source_end)
+      continue;
+
+    void *clean_addr = get_clean_addr(head->addr);
+    ll_add_flags(jump_in + page, head->vaddr, head->reg_sv_flags, clean_addr);
+
+    int in_ht = 0;
+    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
+      in_ht = 1;
+    }
+    if (ht_bin->vaddr[1] == head->vaddr) {
+      ht_bin->tcaddr[1] = clean_addr; // Replace existing entry
+      in_ht = 1;
+    }
+    if (!in_ht)
+      hash_table_add(ht_bin, head->vaddr, clean_addr);
+    ep_count++;
+  }
+  inv_debug("INV: Restored %08x %p (%d)\n", vaddr, found_stub, ep_count);
+  stat_inc(stat_bc_restore);
+  return found_clean;
+}
+
 // Get address from virtual address
 // This is called from the recompiled JR/JALR instructions
 void noinline *get_addr(u_int vaddr)
 {
-  u_int page=get_page(vaddr);
-  u_int vpage=get_vpage(vaddr);
+  u_int page = get_page(vaddr);
   struct ll_entry *head;
-  //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
-  head=jump_in[page];
-  while(head!=NULL) {
-    if(head->vaddr==vaddr) {
-  //printf("TRACE: count=%d next=%d (get_addr match %x: %p)\n",Count,next_interupt,vaddr,head->addr);
+  void *code;
+
+  stat_inc(stat_jump_in_lookups);
+  for (head = jump_in[page]; head != NULL; head = head->next) {
+    if (head->vaddr == vaddr) {
       hash_table_add(hash_table_get(vaddr), vaddr, head->addr);
       return head->addr;
     }
-    head=head->next;
   }
-  head=jump_dirty[vpage];
-  while(head!=NULL) {
-    if(head->vaddr==vaddr) {
-      //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %p)\n",Count,next_interupt,vaddr,head->addr);
-      // Don't restore blocks which are about to expire from the cache
-      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;
-        if(vpage<2048) {
-          restore_candidate[vpage>>3]|=1<<(vpage&7);
-        }
-        else restore_candidate[page>>3]|=1<<(page&7);
-        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
-          hash_table_add(ht_bin, vaddr, head->addr);
+  code = ndrc_try_restore_block(vaddr);
+  if (code)
+    return code;
+
+  int r = new_recompile_block(vaddr);
+  if (r == 0)
+    return get_addr(vaddr);
 
-        return head->addr;
-      }
-    }
-    head=head->next;
-  }
-  //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
-  int r=new_recompile_block(vaddr);
-  if(r==0) return get_addr(vaddr);
   // generate an address error
   Status|=2;
   Cause=(vaddr<<31)|(4<<2);
@@ -991,7 +1062,6 @@ static const struct {
   FUNCNAME(jump_syscall),
   FUNCNAME(jump_syscall_ds),
   FUNCNAME(call_gteStall),
-  FUNCNAME(clean_blocks),
   FUNCNAME(new_dyna_leave),
   FUNCNAME(pcsx_mtc0),
   FUNCNAME(pcsx_mtc0_ds),
@@ -1135,7 +1205,7 @@ static void *check_addr(u_int vaddr)
   return 0;
 }
 
-void remove_hash(int vaddr)
+static void remove_hash(int vaddr)
 {
   //printf("remove hash: %x\n",vaddr);
   struct ht_entry *ht_bin = hash_table_get(vaddr);
@@ -1214,6 +1284,7 @@ static void invalidate_page(u_int page)
   struct ll_entry *head;
   struct ll_entry *next;
   head=jump_in[page];
+  if (head) stat_inc(stat_inv_hits);
   jump_in[page]=0;
   while(head!=NULL) {
     inv_debug("INVALIDATE: %x\n",head->vaddr);
@@ -1293,6 +1364,7 @@ void invalidate_addr(u_int addr)
   //static int rhits;
   // this check is done by the caller
   //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
+  stat_inc(stat_inv_addr_calls);
   u_int page=get_vpage(addr);
   if(page<2048) { // RAM
     struct ll_entry *head;
@@ -1352,11 +1424,6 @@ void invalidate_all_pages(void)
   u_int page;
   for(page=0;page<4096;page++)
     invalidate_page(page);
-  for(page=0;page<1048576;page++)
-    if(!invalid_code[page]) {
-      restore_candidate[(page&2047)>>3]|=1<<(page&7);
-      restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
-    }
   #ifdef USE_MINI_HT
   memset(mini_ht,-1,sizeof(mini_ht));
   #endif
@@ -1386,55 +1453,6 @@ void add_jump_out(u_int vaddr,void *src)
   //inv_debug("add_jump_out:  to %p\n",get_pointer(src));
 }
 
-// If a code block was found to be unmodified (bit was set in
-// restore_candidate) and it remains unmodified (bit is clear
-// in invalid_code) then move the entries for that 4K page from
-// the dirty list to the clean list.
-void clean_blocks(u_int page)
-{
-  struct ll_entry *head;
-  inv_debug("INV: clean_blocks page=%d\n",page);
-  head=jump_dirty[page];
-  while(head!=NULL) {
-    if(!invalid_code[head->vaddr>>12]) {
-      // Don't restore blocks which are about to expire from the cache
-      if (doesnt_expire_soon(head->addr)) {
-        if(verify_dirty(head->addr)) {
-          u_char *start, *end;
-          //printf("Possibly Restore %x (%p)\n",head->vaddr, head->addr);
-          u_int i;
-          u_int inv=0;
-          get_bounds(head->addr, &start, &end);
-          if (start - rdram < RAM_SIZE) {
-            for (i = (start-rdram+0x80000000)>>12; i <= (end-1-rdram+0x80000000)>>12; i++) {
-              inv|=invalid_code[i];
-            }
-          }
-          else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
-            inv=1;
-          }
-          if(!inv) {
-            void *clean_addr = get_clean_addr(head->addr);
-            if (doesnt_expire_soon(clean_addr)) {
-              u_int ppage=page;
-              inv_debug("INV: Restored %x (%p/%p)\n",head->vaddr, head->addr, 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);
-              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
-            }
-          }
-        }
-      }
-    }
-    head=head->next;
-  }
-}
-
 /* Register allocation */
 
 // Note: registers are allocated clean (unmodified state)
@@ -2070,11 +2088,6 @@ static void delayslot_alloc(struct regstat *current,int i)
     case RJUMP:
     case SYSCALL:
     case HLECALL:
-    case SPAN:
-      assem_debug("jump in the delay slot.  this shouldn't happen.\n");//abort();
-      SysPrintf("Disabled speculative precompilation\n");
-      stop_after_jal=1;
-      break;
     case IMM16:
       imm16_alloc(current,i);
       break;
@@ -2121,42 +2134,6 @@ static void delayslot_alloc(struct regstat *current,int i)
   }
 }
 
-// Special case where a branch and delay slot span two pages in virtual memory
-static void pagespan_alloc(struct regstat *current,int i)
-{
-  current->isconst=0;
-  current->wasconst=0;
-  regs[i].wasconst=0;
-  minimum_free_regs[i]=HOST_REGS;
-  alloc_all(current,i);
-  alloc_cc(current,i);
-  dirty_reg(current,CCREG);
-  if(dops[i].opcode==3) // JAL
-  {
-    alloc_reg(current,i,31);
-    dirty_reg(current,31);
-  }
-  if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
-  {
-    alloc_reg(current,i,dops[i].rs1);
-    if (dops[i].rt1!=0) {
-      alloc_reg(current,i,dops[i].rt1);
-      dirty_reg(current,dops[i].rt1);
-    }
-  }
-  if((dops[i].opcode&0x2E)==4) // BEQ/BNE/BEQL/BNEL
-  {
-    if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
-    if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
-  }
-  else
-  if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
-  {
-    if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
-  }
-  //else ...
-}
-
 static void add_stub(enum stub_type type, void *addr, void *retaddr,
   u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
 {
@@ -4111,7 +4088,6 @@ static void ujump_assemble(int i, const struct regstat *i_regs);
 static void rjump_assemble(int i, const struct regstat *i_regs);
 static void cjump_assemble(int i, const struct regstat *i_regs);
 static void sjump_assemble(int i, const struct regstat *i_regs);
-static void pagespan_assemble(int i, const struct regstat *i_regs);
 
 static int assemble(int i, const struct regstat *i_regs, int ccadj_)
 {
@@ -4191,9 +4167,6 @@ static int assemble(int i, const struct regstat *i_regs, int ccadj_)
       sjump_assemble(i, i_regs);
       ds = 1;
       break;
-    case SPAN:
-      pagespan_assemble(i, i_regs);
-      break;
     case NOP:
     case OTHER:
     case NI:
@@ -4213,7 +4186,6 @@ static void ds_assemble(int i, const struct regstat *i_regs)
     case SYSCALL:
     case HLECALL:
     case INTCALL:
-    case SPAN:
     case UJUMP:
     case RJUMP:
     case CJUMP:
@@ -4825,7 +4797,6 @@ static void ds_assemble_entry(int i)
     case SYSCALL:
     case HLECALL:
     case INTCALL:
-    case SPAN:
     case UJUMP:
     case RJUMP:
     case CJUMP:
@@ -4851,11 +4822,6 @@ static void emit_extjump(void *addr, u_int target)
   emit_extjump2(addr, target, dyna_linker);
 }
 
-static void emit_extjump_ds(void *addr, u_int target)
-{
-  emit_extjump2(addr, target, dyna_linker_ds);
-}
-
 // Load 2 immediates optimizing for small code size
 static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
 {
@@ -5899,271 +5865,6 @@ static void sjump_assemble(int i, const struct regstat *i_regs)
   }
 }
 
-static void pagespan_assemble(int i, const struct regstat *i_regs)
-{
-  int s1l=get_reg(i_regs->regmap,dops[i].rs1);
-  int s2l=get_reg(i_regs->regmap,dops[i].rs2);
-  void *taken = NULL;
-  void *nottaken = NULL;
-  int unconditional=0;
-  if(dops[i].rs1==0)
-  {
-    s1l=s2l;
-    s2l=-1;
-  }
-  else if(dops[i].rs2==0)
-  {
-    s2l=-1;
-  }
-  int hr=0;
-  int addr=-1,alt=-1,ntaddr=-1;
-  if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
-  else {
-    while(hr<HOST_REGS)
-    {
-      if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
-         i_regs->regmap[hr]!=dops[i].rs1 &&
-         i_regs->regmap[hr]!=dops[i].rs2 )
-      {
-        addr=hr++;break;
-      }
-      hr++;
-    }
-  }
-  while(hr<HOST_REGS)
-  {
-    if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
-       i_regs->regmap[hr]!=dops[i].rs1 &&
-       i_regs->regmap[hr]!=dops[i].rs2 )
-    {
-      alt=hr++;break;
-    }
-    hr++;
-  }
-  if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
-  {
-    while(hr<HOST_REGS)
-    {
-      if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
-         i_regs->regmap[hr]!=dops[i].rs1 &&
-         i_regs->regmap[hr]!=dops[i].rs2 )
-      {
-        ntaddr=hr;break;
-      }
-      hr++;
-    }
-  }
-  assert(hr<HOST_REGS);
-  if((dops[i].opcode&0x2e)==4||dops[i].opcode==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
-    load_reg(regs[i].regmap_entry,regs[i].regmap,CCREG);
-  }
-  emit_addimm(HOST_CCREG, ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
-  if(dops[i].opcode==2) // J
-  {
-    unconditional=1;
-  }
-  if(dops[i].opcode==3) // JAL
-  {
-    // TODO: mini_ht
-    int rt=get_reg(i_regs->regmap,31);
-    emit_movimm(start+i*4+8,rt);
-    unconditional=1;
-  }
-  if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
-  {
-    emit_mov(s1l,addr);
-    if(dops[i].opcode2==9) // JALR
-    {
-      int rt=get_reg(i_regs->regmap,dops[i].rt1);
-      emit_movimm(start+i*4+8,rt);
-    }
-  }
-  if((dops[i].opcode&0x3f)==4) // BEQ
-  {
-    if(dops[i].rs1==dops[i].rs2)
-    {
-      unconditional=1;
-    }
-    else
-    #ifdef HAVE_CMOV_IMM
-    if(1) {
-      if(s2l>=0) emit_cmp(s1l,s2l);
-      else emit_test(s1l,s1l);
-      emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
-    }
-    else
-    #endif
-    {
-      assert(s1l>=0);
-      emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
-      if(s2l>=0) emit_cmp(s1l,s2l);
-      else emit_test(s1l,s1l);
-      emit_cmovne_reg(alt,addr);
-    }
-  }
-  if((dops[i].opcode&0x3f)==5) // BNE
-  {
-    #ifdef HAVE_CMOV_IMM
-    if(s2l>=0) emit_cmp(s1l,s2l);
-    else emit_test(s1l,s1l);
-    emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
-    #else
-    assert(s1l>=0);
-    emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
-    if(s2l>=0) emit_cmp(s1l,s2l);
-    else emit_test(s1l,s1l);
-    emit_cmovne_reg(alt,addr);
-    #endif
-  }
-  if((dops[i].opcode&0x3f)==0x14) // BEQL
-  {
-    if(s2l>=0) emit_cmp(s1l,s2l);
-    else emit_test(s1l,s1l);
-    if(nottaken) set_jump_target(nottaken, out);
-    nottaken=out;
-    emit_jne(0);
-  }
-  if((dops[i].opcode&0x3f)==0x15) // BNEL
-  {
-    if(s2l>=0) emit_cmp(s1l,s2l);
-    else emit_test(s1l,s1l);
-    nottaken=out;
-    emit_jeq(0);
-    if(taken) set_jump_target(taken, out);
-  }
-  if((dops[i].opcode&0x3f)==6) // BLEZ
-  {
-    emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
-    emit_cmpimm(s1l,1);
-    emit_cmovl_reg(alt,addr);
-  }
-  if((dops[i].opcode&0x3f)==7) // BGTZ
-  {
-    emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
-    emit_cmpimm(s1l,1);
-    emit_cmovl_reg(ntaddr,addr);
-  }
-  if((dops[i].opcode&0x3f)==0x16) // BLEZL
-  {
-    assert((dops[i].opcode&0x3f)!=0x16);
-  }
-  if((dops[i].opcode&0x3f)==0x17) // BGTZL
-  {
-    assert((dops[i].opcode&0x3f)!=0x17);
-  }
-  assert(dops[i].opcode!=1); // BLTZ/BGEZ
-
-  //FIXME: Check CSREG
-  if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
-    if((source[i]&0x30000)==0) // BC1F
-    {
-      emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
-      emit_testimm(s1l,0x800000);
-      emit_cmovne_reg(alt,addr);
-    }
-    if((source[i]&0x30000)==0x10000) // BC1T
-    {
-      emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
-      emit_testimm(s1l,0x800000);
-      emit_cmovne_reg(alt,addr);
-    }
-    if((source[i]&0x30000)==0x20000) // BC1FL
-    {
-      emit_testimm(s1l,0x800000);
-      nottaken=out;
-      emit_jne(0);
-    }
-    if((source[i]&0x30000)==0x30000) // BC1TL
-    {
-      emit_testimm(s1l,0x800000);
-      nottaken=out;
-      emit_jeq(0);
-    }
-  }
-
-  assert(i_regs->regmap[HOST_CCREG]==CCREG);
-  wb_dirtys(regs[i].regmap,regs[i].dirty);
-  if(unconditional)
-  {
-    emit_movimm(ba[i],HOST_BTREG);
-  }
-  else if(addr!=HOST_BTREG)
-  {
-    emit_mov(addr,HOST_BTREG);
-  }
-  void *branch_addr=out;
-  emit_jmp(0);
-  int target_addr=start+i*4+5;
-  void *stub=out;
-  void *compiled_target_addr=check_addr(target_addr);
-  emit_extjump_ds(branch_addr, target_addr);
-  if(compiled_target_addr) {
-    set_jump_target(branch_addr, compiled_target_addr);
-    add_jump_out(target_addr,stub);
-  }
-  else set_jump_target(branch_addr, stub);
-}
-
-// Assemble the delay slot for the above
-static void pagespan_ds()
-{
-  assem_debug("initial delay slot:\n");
-  u_int vaddr=start+1;
-  u_int page=get_page(vaddr);
-  u_int vpage=get_vpage(vaddr);
-  ll_add(jump_dirty+vpage,vaddr,(void *)out);
-  do_dirty_stub_ds(slen*4);
-  ll_add(jump_in+page,vaddr,(void *)out);
-  assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
-  if(regs[0].regmap[HOST_CCREG]!=CCREG)
-    wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty);
-  if(regs[0].regmap[HOST_BTREG]!=BTREG)
-    emit_writeword(HOST_BTREG,&branch_target);
-  load_regs(regs[0].regmap_entry,regs[0].regmap,dops[0].rs1,dops[0].rs2);
-  address_generation(0,&regs[0],regs[0].regmap_entry);
-  if (ram_offset && (dops[0].is_load || dops[0].is_store))
-    load_reg(regs[0].regmap_entry,regs[0].regmap,ROREG);
-  if (dops[0].is_store)
-    load_reg(regs[0].regmap_entry,regs[0].regmap,INVCP);
-  is_delayslot=0;
-  switch (dops[0].itype) {
-    case SYSCALL:
-    case HLECALL:
-    case INTCALL:
-    case SPAN:
-    case UJUMP:
-    case RJUMP:
-    case CJUMP:
-    case SJUMP:
-      SysPrintf("Jump in the delay slot.  This is probably a bug.\n");
-      break;
-    default:
-      assemble(0, &regs[0], 0);
-  }
-  int btaddr=get_reg(regs[0].regmap,BTREG);
-  if(btaddr<0) {
-    btaddr=get_reg_temp(regs[0].regmap);
-    emit_readword(&branch_target,btaddr);
-  }
-  assert(btaddr!=HOST_CCREG);
-  if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
-#ifdef HOST_IMM8
-  host_tempreg_acquire();
-  emit_movimm(start+4,HOST_TEMPREG);
-  emit_cmp(btaddr,HOST_TEMPREG);
-  host_tempreg_release();
-#else
-  emit_cmpimm(btaddr,start+4);
-#endif
-  void *branch = out;
-  emit_jeq(0);
-  store_regs_bt(regs[0].regmap,regs[0].dirty,-1);
-  do_jump_vaddr(btaddr);
-  set_jump_target(branch, out);
-  store_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
-  load_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
-}
-
 static void check_regmap(signed char *regmap)
 {
 #ifndef NDEBUG
@@ -6218,8 +5919,6 @@ void disassemble_inst(int i)
         else
           printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
         break;
-      case SPAN:
-        printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,ba[i]);break;
       case IMM16:
         if(dops[i].opcode==0xf) //LUI
           printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,imm[i]&0xffff);
@@ -6347,7 +6046,6 @@ void new_dynarec_clear_full(void)
   memset(invalid_code,1,sizeof(invalid_code));
   memset(hash_table,0xff,sizeof(hash_table));
   memset(mini_ht,-1,sizeof(mini_ht));
-  memset(restore_candidate,0,sizeof(restore_candidate));
   memset(shadow,0,sizeof(shadow));
   copy=shadow;
   expirep=16384; // Expiry pointer, +2 blocks
@@ -6421,6 +6119,8 @@ void new_dynarec_init(void)
   ram_offset=(uintptr_t)rdram-0x80000000;
   if (ram_offset!=0)
     SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
+  SysPrintf("Mapped (RAM/scrp/ROM/LUTs/TC):\n");
+  SysPrintf("%p/%p/%p/%p/%p\n", psxM, psxH, psxR, mem_rtab, out);
 }
 
 void new_dynarec_cleanup(void)
@@ -6442,6 +6142,7 @@ void new_dynarec_cleanup(void)
   #ifdef ROM_COPY
   if (munmap (ROM_COPY, 67108864) < 0) {SysPrintf("munmap() failed\n");}
   #endif
+  new_dynarec_print_stats();
 }
 
 static u_int *get_source_start(u_int addr, u_int *limit)
@@ -6572,6 +6273,20 @@ void new_dynarec_load_blocks(const void *save, int size)
   memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
 }
 
+void new_dynarec_print_stats(void)
+{
+#ifdef STAT_PRINT
+  printf("cc %3d,%3d,%3d lu%3d,%3d c%3d inv%3d,%3d tc_offs %zu\n",
+    stat_bc_pre, stat_bc_direct, stat_bc_restore,
+    stat_jump_in_lookups, stat_restore_tries, stat_restore_compares,
+    stat_inv_addr_calls, stat_inv_hits,
+    out - ndrc->translation_cache);
+  stat_bc_direct = stat_bc_pre = stat_bc_restore =
+  stat_jump_in_lookups = stat_restore_tries = stat_restore_compares =
+  stat_inv_addr_calls = stat_inv_hits = 0;
+#endif
+}
+
 static int apply_hacks(void)
 {
   int i;
@@ -7070,8 +6785,24 @@ static noinline void pass1_disassemble(u_int pagelimit)
 
     /* Is this the end of the block? */
     if (i > 0 && dops[i-1].is_ujump) {
-      if(dops[i-1].rt1==0) { // Continue past subroutine call (JAL)
-        done=2;
+      if (dops[i-1].rt1 == 0) { // not jal
+        int found_bbranch = 0, t = (ba[i-1] - start) / 4;
+        if ((u_int)(t - i) < 64 && start + (t+64)*4 < pagelimit) {
+          // scan for a branch back to i+1
+          for (j = t; j < t + 64; j++) {
+            int tmpop = source[j] >> 26;
+            if (tmpop == 1 || ((tmpop & ~3) == 4)) {
+              int t2 = j + 1 + (int)(signed short)source[j];
+              if (t2 == i + 1) {
+                //printf("blk expand %08x<-%08x\n", start + (i+1)*4, start + j*4);
+                found_bbranch = 1;
+                break;
+              }
+            }
+          }
+        }
+        if (!found_bbranch)
+          done = 2;
       }
       else {
         if(stop_after_jal) done=1;
@@ -7104,13 +6835,11 @@ static noinline void pass1_disassemble(u_int pagelimit)
       SysPrintf("Disabled speculative precompilation\n");
     }
   }
-  slen=i;
-  if (dops[i-1].is_jump) {
-    if(start+i*4==pagelimit) {
-      dops[i-1].itype=SPAN;
-    }
-  }
-  assert(slen>0);
+  while (i > 0 && dops[i-1].is_jump)
+    i--;
+  assert(i > 0);
+  assert(!dops[i-1].is_jump);
+  slen = i;
 }
 
 // Basic liveness analysis for MIPS registers
@@ -7613,9 +7342,6 @@ static noinline void pass3_register_alloc(u_int addr)
         case INTCALL:
           syscall_alloc(&current,i);
           break;
-        case SPAN:
-          pagespan_alloc(&current,i);
-          break;
       }
 
       // Create entry (branch target) regmap
@@ -8046,7 +7772,7 @@ static noinline void pass4_cull_unused_regs(void)
       }
     }
     // Cycle count is needed at branches.  Assume it is needed at the target too.
-    if(i==0||dops[i].bt||dops[i].itype==CJUMP||dops[i].itype==SPAN) {
+    if(i==0||dops[i].bt||dops[i].itype==CJUMP) {
       if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
       if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
     }
@@ -9033,8 +8759,8 @@ int new_recompile_block(u_int addr)
       state_rflags |= 1 << i;
   }
 
-  start = (u_int)addr&~3;
-  //assert(((u_int)addr&1)==0); // start-in-delay-slot flag
+  assert(!(addr & 3));
+  start = addr & ~3;
   new_dynarec_did_compile=1;
   if (Config.HLE && start == 0x80001000) // hlecall
   {
@@ -9138,21 +8864,13 @@ int new_recompile_block(u_int addr)
     }
   }
 
-  if(dops[slen-1].itype==SPAN) {
-    dops[slen-1].bt=1; // Mark as a branch target so instruction can restart after exception
-  }
-
   /* Pass 8 - Assembly */
   linkcount=0;stubcount=0;
   is_delayslot=0;
   u_int dirty_pre=0;
   void *beginning=start_block();
-  int ds = 0;
-  if((u_int)addr&1) {
-    ds=1;
-    pagespan_ds();
-  }
   void *instr_addr0_override = NULL;
+  int ds = 0;
 
   if (start == 0x80030000) {
     // nasty hack for the fastbios thing
@@ -9265,7 +8983,7 @@ int new_recompile_block(u_int addr)
   // If the block did not end with an unconditional branch,
   // add a jump to the next instruction.
   else if (i > 1) {
-    if (!dops[i-2].is_ujump && dops[i-1].itype != SPAN) {
+    if (!dops[i-2].is_ujump) {
       assert(!dops[i-1].is_jump);
       assert(i==slen);
       if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
@@ -9416,17 +9134,7 @@ int new_recompile_block(u_int addr)
     out = ndrc->translation_cache;
 
   // Trap writes to any of the pages we compiled
-  for(i=start>>12;i<=(start+slen*4)>>12;i++) {
-    invalid_code[i]=0;
-  }
-  inv_code_start=inv_code_end=~0;
-
-  // for PCSX we need to mark all mirrors too
-  if(get_page(start)<(RAM_SIZE>>12))
-    for(i=start>>12;i<=(start+slen*4)>>12;i++)
-      invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
-      invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
-      invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
+  mark_valid_code(start, slen*4);
 
   /* Pass 10 - Free memory by expiring oldest blocks */
 
@@ -9435,6 +9143,7 @@ int new_recompile_block(u_int addr)
 #ifdef ASSEM_PRINT
   fflush(stdout);
 #endif
+  stat_inc(stat_bc_direct);
   return 0;
 }