X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=pcsx_rearmed.git;a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fnew_dynarec.c;h=1fe2c0c4114c8f4ff1077623c022e15376c74cbf;hp=fb6ace4725e1651034ad153ee5105d94ce445649;hb=2adcd6fad4594a18025b4f00c49e43a23f8f8992;hpb=dd3a91a17c7b45c56cde69ff39ea35185f476898

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index fb6ace47..1fe2c0c4 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -84,6 +84,7 @@ struct ll_entry
   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];
@@ -94,10 +95,9 @@ struct ll_entry
   signed char regmap[MAXBLOCK][HOST_REGS];
   signed char regmap_entry[MAXBLOCK][HOST_REGS];
   uint64_t constmap[MAXBLOCK][HOST_REGS];
-  uint64_t known_value[HOST_REGS];
-  u_int known_reg;
   struct regstat regs[MAXBLOCK];
   struct regstat branch_regs[MAXBLOCK];
+  signed char minimum_free_regs[MAXBLOCK];
   u_int needed_reg[MAXBLOCK];
   uint64_t requires_32bit[MAXBLOCK];
   u_int wont_dirty[MAXBLOCK];
@@ -121,7 +121,12 @@ struct ll_entry
   char shadow[1048576]  __attribute__((aligned(16)));
   void *copy;
   int expirep;
+#ifndef PCSX
   u_int using_tlb;
+#else
+  static const u_int using_tlb=0;
+#endif
+  static u_int sp_in_mirror;
   u_int stop_after_jal;
   extern u_char restore_candidate[512];
   extern int cycle_count;
@@ -134,19 +139,21 @@ struct ll_entry
 #define CSREG 35 // Coprocessor status
 #define CCREG 36 // Cycle count
 #define INVCP 37 // Pointer to invalid_code
-#define TEMPREG 38
-#define FTEMP 38 // FPU/LDL/LDR temporary register
-#define PTEMP 39 // Prefetch temporary register
-#define TLREG 40 // TLB mapping offset
-#define RHASH 41 // Return address hash
-#define RHTBL 42 // Return address hash table address
-#define RTEMP 43 // JR/JALR address register
-#define MAXREG 43
-#define AGEN1 44 // Address generation temporary register
-#define AGEN2 45 // Address generation temporary register
-#define MGEN1 46 // Maptable address generation temporary register
-#define MGEN2 47 // Maptable address generation temporary register
-#define BTREG 48 // Branch target temporary register
+#define MMREG 38 // Pointer to memory_map
+#define ROREG 39 // ram offset (if rdram!=0x80000000)
+#define TEMPREG 40
+#define FTEMP 40 // FPU temporary register
+#define PTEMP 41 // Prefetch temporary register
+#define TLREG 42 // TLB mapping offset
+#define RHASH 43 // Return address hash
+#define RHTBL 44 // Return address hash table address
+#define RTEMP 45 // JR/JALR address register
+#define MAXREG 45
+#define AGEN1 46 // Address generation temporary register
+#define AGEN2 47 // Address generation temporary register
+#define MGEN1 48 // Maptable address generation temporary register
+#define MGEN2 49 // Maptable address generation temporary register
+#define BTREG 50 // Branch target temporary register
 
   /* instruction types */
 #define NOP 0     // No operation
@@ -1180,6 +1187,9 @@ void invalidate_block(u_int block)
   
   // Don't trap writes
   invalid_code[block]=1;
+#ifdef PCSX
+  invalid_code[((u_int)0x80000000>>12)|page]=1;
+#endif
 #ifndef DISABLE_TLB
   // If there is a valid TLB entry for this page, remove write protect
   if(tlb_LUT_w[block]) {
@@ -1388,7 +1398,10 @@ void shift_alloc(struct regstat *current,int i)
       if(rs1[i]) alloc_reg(current,i,rs1[i]);
       if(rs2[i]) alloc_reg(current,i,rs2[i]);
       alloc_reg(current,i,rt1[i]);
-      if(rt1[i]==rs2[i]) alloc_reg_temp(current,i,-1);
+      if(rt1[i]==rs2[i]) {
+        alloc_reg_temp(current,i,-1);
+        minimum_free_regs[i]=1;
+      }
       current->is32|=1LL<<rt1[i];
     } else { // DSLLV/DSRLV/DSRAV
       if(rs1[i]) alloc_reg64(current,i,rs1[i]);
@@ -1396,7 +1409,10 @@ void shift_alloc(struct regstat *current,int i)
       alloc_reg64(current,i,rt1[i]);
       current->is32&=~(1LL<<rt1[i]);
       if(opcode2[i]==0x16||opcode2[i]==0x17) // DSRLV and DSRAV need a temporary register
+      {
         alloc_reg_temp(current,i,-1);
+        minimum_free_regs[i]=1;
+      }
     }
     clear_const(current,rs1[i]);
     clear_const(current,rs2[i]);
@@ -1591,6 +1607,7 @@ void load_alloc(struct regstat *current,int i)
     if(get_reg(current->regmap,rt1[i])<0) {
       // dummy load, but we still need a register to calculate the address
       alloc_reg_temp(current,i,-1);
+      minimum_free_regs[i]=1;
     }
     if(opcode[i]==0x27||opcode[i]==0x37) // LWU/LD
     {
@@ -1603,6 +1620,7 @@ void load_alloc(struct regstat *current,int i)
       alloc_reg64(current,i,rt1[i]);
       alloc_all(current,i);
       alloc_reg64(current,i,FTEMP);
+      minimum_free_regs[i]=HOST_REGS;
     }
     else current->is32|=1LL<<rt1[i];
     dirty_reg(current,rt1[i]);
@@ -1613,6 +1631,7 @@ void load_alloc(struct regstat *current,int i)
     {
       alloc_reg(current,i,FTEMP);
       alloc_reg_temp(current,i,-1);
+      minimum_free_regs[i]=1;
     }
   }
   else
@@ -1624,10 +1643,12 @@ void load_alloc(struct regstat *current,int i)
       alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
     }
     alloc_reg_temp(current,i,-1);
+    minimum_free_regs[i]=1;
     if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
     {
       alloc_all(current,i);
       alloc_reg64(current,i,FTEMP);
+      minimum_free_regs[i]=HOST_REGS;
     }
   }
 }
@@ -1653,6 +1674,7 @@ void store_alloc(struct regstat *current,int i)
   }
   // We need a temporary register for address generation
   alloc_reg_temp(current,i,-1);
+  minimum_free_regs[i]=1;
 }
 
 void c1ls_alloc(struct regstat *current,int i)
@@ -1690,6 +1712,7 @@ void c2ls_alloc(struct regstat *current,int i)
   #endif
   // We need a temporary register for address generation
   alloc_reg_temp(current,i,-1);
+  minimum_free_regs[i]=1;
 }
 
 #ifndef multdiv_alloc
@@ -1735,6 +1758,7 @@ void multdiv_alloc(struct regstat *current,int i)
       current->is32&=~(1LL<<LOREG);
       dirty_reg(current,HIREG);
       dirty_reg(current,LOREG);
+      minimum_free_regs[i]=HOST_REGS;
     }
   }
   else
@@ -1783,6 +1807,7 @@ void cop0_alloc(struct regstat *current,int i)
     assert(opcode2[i]==0x10);
     alloc_all(current,i);
   }
+  minimum_free_regs[i]=HOST_REGS;
 }
 
 void cop1_alloc(struct regstat *current,int i)
@@ -1819,16 +1844,19 @@ void cop1_alloc(struct regstat *current,int i)
       alloc_reg_temp(current,i,-1);
     }
   }
+  minimum_free_regs[i]=1;
 }
 void fconv_alloc(struct regstat *current,int i)
 {
   alloc_reg(current,i,CSREG); // Load status
   alloc_reg_temp(current,i,-1);
+  minimum_free_regs[i]=1;
 }
 void float_alloc(struct regstat *current,int i)
 {
   alloc_reg(current,i,CSREG); // Load status
   alloc_reg_temp(current,i,-1);
+  minimum_free_regs[i]=1;
 }
 void c2op_alloc(struct regstat *current,int i)
 {
@@ -1840,6 +1868,7 @@ void fcomp_alloc(struct regstat *current,int i)
   alloc_reg(current,i,FSREG); // Load flags
   dirty_reg(current,FSREG); // Flag will be modified
   alloc_reg_temp(current,i,-1);
+  minimum_free_regs[i]=1;
 }
 
 void syscall_alloc(struct regstat *current,int i)
@@ -1847,6 +1876,7 @@ void syscall_alloc(struct regstat *current,int i)
   alloc_cc(current,i);
   dirty_reg(current,CCREG);
   alloc_all(current,i);
+  minimum_free_regs[i]=HOST_REGS;
   current->isconst=0;
 }
 
@@ -1925,6 +1955,7 @@ 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);
@@ -2757,8 +2788,10 @@ void load_assemble(int i,struct regstat *i_regs)
   if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
   if(s>=0) {
     c=(i_regs->wasconst>>s)&1;
-    memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
-    if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
+    if (c) {
+      memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
+      if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
+    }
   }
   //printf("load_assemble: c=%d\n",c);
   //if(c) printf("load_assemble: const=%x\n",(int)constmap[i][s]+offset);
@@ -2794,6 +2827,13 @@ void load_assemble(int i,struct regstat *i_regs)
       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
       #endif
       {
+        #ifdef PCSX
+        if(sp_in_mirror&&rs1[i]==29) {
+          emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
+          emit_cmpimm(HOST_TEMPREG,RAM_SIZE);
+        }
+        else
+        #endif
         emit_cmpimm(addr,RAM_SIZE);
         jaddr=(int)out;
         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
@@ -2833,6 +2873,9 @@ void load_assemble(int i,struct regstat *i_regs)
           else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
 #else
           if(!c) a=addr;
+#endif
+#ifdef PCSX
+          if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
 #endif
           emit_movsbl_indexed_tlb(x,a,map,tl);
         }
@@ -2858,6 +2901,9 @@ void load_assemble(int i,struct regstat *i_regs)
           else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
 #else
           if(!c) a=addr;
+#endif
+#ifdef PCSX
+          if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
 #endif
           //#ifdef
           //emit_movswl_indexed_tlb(x,tl,map,tl);
@@ -2883,13 +2929,17 @@ void load_assemble(int i,struct regstat *i_regs)
   if (opcode[i]==0x23) { // LW
     if(!c||memtarget) {
       if(!dummy) {
+        int a=addr;
+#ifdef PCSX
+        if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
         //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
         #ifdef HOST_IMM_ADDR32
         if(c)
           emit_readword_tlb(constmap[i][s]+offset,map,tl);
         else
         #endif
-        emit_readword_indexed_tlb(0,addr,map,tl);
+        emit_readword_indexed_tlb(0,a,map,tl);
       }
       if(jaddr)
         add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
@@ -2915,6 +2965,9 @@ void load_assemble(int i,struct regstat *i_regs)
           else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
 #else
           if(!c) a=addr;
+#endif
+#ifdef PCSX
+          if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
 #endif
           emit_movzbl_indexed_tlb(x,a,map,tl);
         }
@@ -2940,6 +2993,9 @@ void load_assemble(int i,struct regstat *i_regs)
           else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
 #else
           if(!c) a=addr;
+#endif
+#ifdef PCSX
+          if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
 #endif
           //#ifdef
           //emit_movzwl_indexed_tlb(x,tl,map,tl);
@@ -2966,13 +3022,17 @@ void load_assemble(int i,struct regstat *i_regs)
     assert(th>=0);
     if(!c||memtarget) {
       if(!dummy) {
+        int a=addr;
+#ifdef PCSX
+        if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
         //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
         #ifdef HOST_IMM_ADDR32
         if(c)
           emit_readword_tlb(constmap[i][s]+offset,map,tl);
         else
         #endif
-        emit_readword_indexed_tlb(0,addr,map,tl);
+        emit_readword_indexed_tlb(0,a,map,tl);
       }
       if(jaddr)
         add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
@@ -2985,6 +3045,10 @@ void load_assemble(int i,struct regstat *i_regs)
   if (opcode[i]==0x37) { // LD
     if(!c||memtarget) {
       if(!dummy) {
+        int a=addr;
+#ifdef PCSX
+        if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
         //gen_tlb_addr_r(tl,map);
         //if(th>=0) emit_readword_indexed((int)rdram-0x80000000,addr,th);
         //emit_readword_indexed((int)rdram-0x7FFFFFFC,addr,tl);
@@ -2993,7 +3057,7 @@ void load_assemble(int i,struct regstat *i_regs)
           emit_readdword_tlb(constmap[i][s]+offset,map,th,tl);
         else
         #endif
-        emit_readdword_indexed_tlb(0,addr,map,th,tl);
+        emit_readdword_indexed_tlb(0,a,map,th,tl);
       }
       if(jaddr)
         add_stub(LOADD_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
@@ -3058,8 +3122,10 @@ void store_assemble(int i,struct regstat *i_regs)
   offset=imm[i];
   if(s>=0) {
     c=(i_regs->wasconst>>s)&1;
-    memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
-    if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
+    if(c) {
+      memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
+      if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
+    }
   }
   assert(tl>=0);
   assert(temp>=0);
@@ -3071,9 +3137,15 @@ void store_assemble(int i,struct regstat *i_regs)
   else addr=s;
   if(!using_tlb) {
     if(!c) {
+      #ifdef PCSX
+      if(sp_in_mirror&&rs1[i]==29) {
+        emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
+        emit_cmpimm(HOST_TEMPREG,RAM_SIZE);
+      }
+      else
+      #endif
       #ifdef R29_HACK
       // Strmnnrmn's speed hack
-      memtarget=1;
       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
       #endif
       emit_cmpimm(addr,RAM_SIZE);
@@ -3081,6 +3153,7 @@ void store_assemble(int i,struct regstat *i_regs)
       if(s==addr) emit_mov(s,temp);
       #endif
       #ifdef R29_HACK
+      memtarget=1;
       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
       #endif
       {
@@ -3106,71 +3179,76 @@ void store_assemble(int i,struct regstat *i_regs)
 
   if (opcode[i]==0x28) { // SB
     if(!c||memtarget) {
-      int x=0;
+      int x=0,a=temp;
 #ifdef BIG_ENDIAN_MIPS
       if(!c) emit_xorimm(addr,3,temp);
       else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
 #else
-      if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
-      else if (addr!=temp) emit_mov(addr,temp);
+      if(!c) a=addr;
+#endif
+#ifdef PCSX
+      if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
 #endif
       //gen_tlb_addr_w(temp,map);
       //emit_writebyte_indexed(tl,(int)rdram-0x80000000,temp);
-      emit_writebyte_indexed_tlb(tl,x,temp,map,temp);
+      emit_writebyte_indexed_tlb(tl,x,a,map,a);
     }
     type=STOREB_STUB;
   }
   if (opcode[i]==0x29) { // SH
     if(!c||memtarget) {
-      int x=0;
+      int x=0,a=temp;
 #ifdef BIG_ENDIAN_MIPS
       if(!c) emit_xorimm(addr,2,temp);
       else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
 #else
-      if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
-      else if (addr!=temp) emit_mov(addr,temp);
+      if(!c) a=addr;
+#endif
+#ifdef PCSX
+      if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
 #endif
       //#ifdef
       //emit_writehword_indexed_tlb(tl,x,temp,map,temp);
       //#else
       if(map>=0) {
-        gen_tlb_addr_w(temp,map);
-        emit_writehword_indexed(tl,x,temp);
+        gen_tlb_addr_w(a,map);
+        emit_writehword_indexed(tl,x,a);
       }else
-        emit_writehword_indexed(tl,(int)rdram-0x80000000+x,temp);
+        emit_writehword_indexed(tl,(int)rdram-0x80000000+x,a);
     }
     type=STOREH_STUB;
   }
   if (opcode[i]==0x2B) { // SW
-    if(!c||memtarget)
+    if(!c||memtarget) {
+      int a=addr;
+#ifdef PCSX
+      if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
       //emit_writeword_indexed(tl,(int)rdram-0x80000000,addr);
-      emit_writeword_indexed_tlb(tl,0,addr,map,temp);
+      emit_writeword_indexed_tlb(tl,0,a,map,temp);
+    }
     type=STOREW_STUB;
   }
   if (opcode[i]==0x3F) { // SD
     if(!c||memtarget) {
+      int a=addr;
+#ifdef PCSX
+      if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
+#endif
       if(rs2[i]) {
         assert(th>=0);
         //emit_writeword_indexed(th,(int)rdram-0x80000000,addr);
         //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,addr);
-        emit_writedword_indexed_tlb(th,tl,0,addr,map,temp);
+        emit_writedword_indexed_tlb(th,tl,0,a,map,temp);
       }else{
         // Store zero
         //emit_writeword_indexed(tl,(int)rdram-0x80000000,temp);
         //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,temp);
-        emit_writedword_indexed_tlb(tl,tl,0,addr,map,temp);
+        emit_writedword_indexed_tlb(tl,tl,0,a,map,temp);
       }
     }
     type=STORED_STUB;
   }
-  if(!using_tlb&&(!c||memtarget))
-    // addr could be a temp, make sure it survives STORE*_STUB
-    reglist|=1<<addr;
-  if(jaddr) {
-    add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
-  } else if(!memtarget) {
-    inline_writestub(type,i,constmap[i][s]+offset,i_regs->regmap,rs2[i],ccadj[i],reglist);
-  }
   if(!using_tlb) {
     if(!c||memtarget) {
       #ifdef DESTRUCTIVE_SHIFT
@@ -3185,11 +3263,20 @@ void store_assemble(int i,struct regstat *i_regs)
       #else
       emit_cmpmem_indexedsr12_imm((int)invalid_code,addr,1);
       #endif
+      #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
+      emit_callne(invalidate_addr_reg[addr]);
+      #else
       jaddr2=(int)out;
       emit_jne(0);
       add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),addr,0,0,0);
+      #endif
     }
   }
+  if(jaddr) {
+    add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
+  } else if(c&&!memtarget) {
+    inline_writestub(type,i,constmap[i][s]+offset,i_regs->regmap,rs2[i],ccadj[i],reglist);
+  }
   //if(opcode[i]==0x2B || opcode[i]==0x3F)
   //if(opcode[i]==0x2B || opcode[i]==0x28)
   //if(opcode[i]==0x2B || opcode[i]==0x29)
@@ -3230,7 +3317,7 @@ void storelr_assemble(int i,struct regstat *i_regs)
   int jaddr=0,jaddr2;
   int case1,case2,case3;
   int done0,done1,done2;
-  int memtarget,c=0;
+  int memtarget=0,c=0;
   int agr=AGEN1+(i&1);
   u_int hr,reglist=0;
   th=get_reg(i_regs->regmap,rs2[i]|64);
@@ -3241,8 +3328,10 @@ void storelr_assemble(int i,struct regstat *i_regs)
   offset=imm[i];
   if(s>=0) {
     c=(i_regs->isconst>>s)&1;
-    memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
-    if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
+    if(c) {
+      memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
+      if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
+    }
   }
   assert(tl>=0);
   for(hr=0;hr<HOST_REGS;hr++) {
@@ -3614,9 +3703,13 @@ void c1ls_assemble(int i,struct regstat *i_regs)
       #else
       emit_cmpmem_indexedsr12_imm((int)invalid_code,temp,1);
       #endif
+      #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
+      emit_callne(invalidate_addr_reg[temp]);
+      #else
       jaddr3=(int)out;
       emit_jne(0);
       add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),temp,0,0,0);
+      #endif
     }
   }
   if(jaddr2) add_stub(type,jaddr2,(int)out,i,offset||c||s<0?ar:s,(int)i_regs,ccadj[i],reglist);
@@ -3722,9 +3815,13 @@ void c2ls_assemble(int i,struct regstat *i_regs)
 #else
     emit_cmpmem_indexedsr12_imm((int)invalid_code,ar,1);
 #endif
+    #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
+    emit_callne(invalidate_addr_reg[ar]);
+    #else
     jaddr3=(int)out;
     emit_jne(0);
     add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),ar,0,0,0);
+    #endif
   }
   if (opcode[i]==0x32) { // LWC2
     cop2_put_dreg(copr,tl,HOST_TEMPREG);
@@ -5068,34 +5165,19 @@ void ujump_assemble(int i,struct regstat *i_regs)
     if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
   }
   #endif
-  ds_assemble(i+1,i_regs);
-  uint64_t bc_unneeded=branch_regs[i].u;
-  uint64_t bc_unneeded_upper=branch_regs[i].uu;
-  bc_unneeded|=1|(1LL<<rt1[i]);
-  bc_unneeded_upper|=1|(1LL<<rt1[i]);
-  wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
-                bc_unneeded,bc_unneeded_upper);
-  load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
   if(rt1[i]==31) {
     int rt;
     unsigned int return_address;
-    assert(rt1[i+1]!=31);
-    assert(rt2[i+1]!=31);
     rt=get_reg(branch_regs[i].regmap,31);
     assem_debug("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]);
     //assert(rt>=0);
     return_address=start+i*4+8;
     if(rt>=0) {
       #ifdef USE_MINI_HT
-      if(internal_branch(branch_regs[i].is32,return_address)) {
-        int temp=rt+1;
-        if(temp==EXCLUDE_REG||temp>=HOST_REGS||
-           branch_regs[i].regmap[temp]>=0)
-        {
-          temp=get_reg(branch_regs[i].regmap,-1);
-        }
+      if(internal_branch(branch_regs[i].is32,return_address)&&rt1[i+1]!=31) {
+        int temp=-1; // note: must be ds-safe
         #ifdef HOST_TEMPREG
-        if(temp<0) temp=HOST_TEMPREG;
+        temp=HOST_TEMPREG;
         #endif
         if(temp>=0) do_miniht_insert(return_address,rt,temp);
         else emit_movimm(return_address,rt);
@@ -5116,6 +5198,14 @@ void ujump_assemble(int i,struct regstat *i_regs)
       }
     }
   }
+  ds_assemble(i+1,i_regs);
+  uint64_t bc_unneeded=branch_regs[i].u;
+  uint64_t bc_unneeded_upper=branch_regs[i].uu;
+  bc_unneeded|=1|(1LL<<rt1[i]);
+  bc_unneeded_upper|=1|(1LL<<rt1[i]);
+  wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
+                bc_unneeded,bc_unneeded_upper);
+  load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
   int cc,adj;
   cc=get_reg(branch_regs[i].regmap,CCREG);
   assert(cc==HOST_CCREG);
@@ -5290,26 +5380,15 @@ void cjump_assemble(int i,struct regstat *i_regs)
   int prev_cop1_usable=cop1_usable;
   int unconditional=0,nop=0;
   int only32=0;
-  int ooo=1;
   int invert=0;
   int internal=internal_branch(branch_regs[i].is32,ba[i]);
   if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
-  if(likely[i]) ooo=0;
   if(!match) invert=1;
   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
   if(i>(ba[i]-start)>>2) invert=1;
   #endif
-    
-  if(ooo)
-    if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))||
-       (rs2[i]&&(rs2[i]==rt1[i+1]||rs2[i]==rt2[i+1])))
-  {
-    // Write-after-read dependency prevents out of order execution
-    // First test branch condition, then execute delay slot, then branch
-    ooo=0;
-  }
-
-  if(ooo) {
+  
+  if(ooo[i]) {
     s1l=get_reg(branch_regs[i].regmap,rs1[i]);
     s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
     s2l=get_reg(branch_regs[i].regmap,rs2[i]);
@@ -5345,7 +5424,7 @@ void cjump_assemble(int i,struct regstat *i_regs)
     only32=(regs[i].was32>>rs1[i])&(regs[i].was32>>rs2[i])&1;
   }
 
-  if(ooo) {
+  if(ooo[i]) {
     // Out of order execution (delay slot first)
     //printf("OOOE\n");
     address_generation(i+1,i_regs,regs[i].regmap_entry);
@@ -5684,11 +5763,9 @@ void sjump_assemble(int i,struct regstat *i_regs)
   int prev_cop1_usable=cop1_usable;
   int unconditional=0,nevertaken=0;
   int only32=0;
-  int ooo=1;
   int invert=0;
   int internal=internal_branch(branch_regs[i].is32,ba[i]);
   if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
-  if(likely[i]) ooo=0;
   if(!match) invert=1;
   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
   if(i>(ba[i]-start)>>2) invert=1;
@@ -5697,19 +5774,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
   //if(opcode2[i]>=0x10) return; // FIXME (BxxZAL)
   //assert(opcode2[i]<0x10||rs1[i]==0); // FIXME (BxxZAL)
 
-  if(ooo) {
-    if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))
-    {
-      // Write-after-read dependency prevents out of order execution
-      // First test branch condition, then execute delay slot, then branch
-      ooo=0;
-    }
-    if(rt1[i]==31&&(rs1[i+1]==31||rs2[i+1]==31||rt1[i+1]==31||rt2[i+1]==31))
-      // BxxZAL $ra is available to delay insn, so do it in order
-      ooo=0;
-  }
-
-  if(ooo) {
+  if(ooo[i]) {
     s1l=get_reg(branch_regs[i].regmap,rs1[i]);
     s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
   }
@@ -5731,7 +5796,7 @@ void sjump_assemble(int i,struct regstat *i_regs)
     only32=(regs[i].was32>>rs1[i])&1;
   }
 
-  if(ooo) {
+  if(ooo[i]) {
     // Out of order execution (delay slot first)
     //printf("OOOE\n");
     address_generation(i+1,i_regs,regs[i].regmap_entry);
@@ -6024,25 +6089,15 @@ void fjump_assemble(int i,struct regstat *i_regs)
   assem_debug("fmatch=%d\n",match);
   int fs,cs;
   int eaddr;
-  int ooo=1;
   int invert=0;
   int internal=internal_branch(branch_regs[i].is32,ba[i]);
   if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
-  if(likely[i]) ooo=0;
   if(!match) invert=1;
   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
   if(i>(ba[i]-start)>>2) invert=1;
   #endif
 
-  if(ooo)
-    if(itype[i+1]==FCOMP)
-  {
-    // Write-after-read dependency prevents out of order execution
-    // First test branch condition, then execute delay slot, then branch
-    ooo=0;
-  }
-
-  if(ooo) {
+  if(ooo[i]) {
     fs=get_reg(branch_regs[i].regmap,FSREG);
     address_generation(i+1,i_regs,regs[i].regmap_entry); // Is this okay?
   }
@@ -6061,7 +6116,7 @@ void fjump_assemble(int i,struct regstat *i_regs)
     cop1_usable=1;
   }
 
-  if(ooo) {
+  if(ooo[i]) {
     // Out of order execution (delay slot first)
     //printf("OOOE\n");
     ds_assemble(i+1,i_regs);
@@ -7712,20 +7767,10 @@ void disassemble_inst(int i)
     }
 }
 
-void new_dynarec_init()
+// clear the state completely, instead of just marking
+// things invalid like invalidate_all_pages() does
+void new_dynarec_clear_full()
 {
-  printf("Init new dynarec\n");
-  out=(u_char *)BASE_ADDR;
-  if (mmap (out, 1<<TARGET_SIZE_2,
-            PROT_READ | PROT_WRITE | PROT_EXEC,
-            MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
-            -1, 0) <= 0) {printf("mmap() failed\n");}
-#ifdef MUPEN64
-  rdword=&readmem_dword;
-  fake_pc.f.r.rs=&readmem_dword;
-  fake_pc.f.r.rt=&readmem_dword;
-  fake_pc.f.r.rd=&readmem_dword;
-#endif
   int n;
   for(n=0x80000;n<0x80800;n++)
     invalid_code[n]=1;
@@ -7733,23 +7778,48 @@ void new_dynarec_init()
     hash_table[n][0]=hash_table[n][2]=-1;
   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
   pending_exception=0;
   literalcount=0;
-#ifdef HOST_IMM8
-  // Copy this into local area so we don't have to put it in every literal pool
-  invc_ptr=invalid_code;
-#endif
   stop_after_jal=0;
   // TLB
+#ifndef DISABLE_TLB
   using_tlb=0;
+#endif
+  sp_in_mirror=0;
   for(n=0;n<524288;n++) // 0 .. 0x7FFFFFFF
     memory_map[n]=-1;
   for(n=524288;n<526336;n++) // 0x80000000 .. 0x807FFFFF
     memory_map[n]=((u_int)rdram-0x80000000)>>2;
   for(n=526336;n<1048576;n++) // 0x80800000 .. 0xFFFFFFFF
     memory_map[n]=-1;
+  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);
+}
+
+void new_dynarec_init()
+{
+  printf("Init new dynarec\n");
+  out=(u_char *)BASE_ADDR;
+  if (mmap (out, 1<<TARGET_SIZE_2,
+            PROT_READ | PROT_WRITE | PROT_EXEC,
+            MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
+            -1, 0) <= 0) {printf("mmap() failed\n");}
+#ifdef MUPEN64
+  rdword=&readmem_dword;
+  fake_pc.f.r.rs=&readmem_dword;
+  fake_pc.f.r.rt=&readmem_dword;
+  fake_pc.f.r.rd=&readmem_dword;
+#endif
+  int n;
+  new_dynarec_clear_full();
+#ifdef HOST_IMM8
+  // Copy this into local area so we don't have to put it in every literal pool
+  invc_ptr=invalid_code;
+#endif
 #ifdef MUPEN64
   for(n=0;n<0x8000;n++) { // 0 .. 0x7FFFFFFF
     writemem[n] = write_nomem_new;
@@ -7828,6 +7898,11 @@ int new_recompile_block(int addr)
   start = (u_int)addr&~3;
   //assert(((u_int)addr&1)==0);
 #ifdef PCSX
+  if(!sp_in_mirror&&(signed int)(psxRegs.GPR.n.sp&0xffe00000)>0x80200000&&
+     0x10000<=psxRegs.GPR.n.sp&&(psxRegs.GPR.n.sp&~0xe0e00000)<RAM_SIZE) {
+    printf("SP hack enabled (%08x), @%08x\n", psxRegs.GPR.n.sp);
+    sp_in_mirror=1;
+  }
   if (Config.HLE && start == 0x80001000) // hlecall
   {
     // XXX: is this enough? Maybe check hleSoftCall?
@@ -7919,7 +7994,8 @@ int new_recompile_block(int addr)
   /* Pass 1 disassembly */
 
   for(i=0;!done;i++) {
-    bt[i]=0;likely[i]=0;op2=0;
+    bt[i]=0;likely[i]=0;ooo[i]=0;op2=0;
+    minimum_free_regs[i]=0;
     opcode[i]=op=source[i]>>26;
     switch(op)
     {
@@ -8219,18 +8295,6 @@ int new_recompile_block(int addr)
         printf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
         break;
     }
-#ifdef PCSX
-    /* detect branch in delay slot early */
-    if(type==RJUMP||type==UJUMP||type==CJUMP||type==SJUMP||type==FJUMP) {
-      opcode[i+1]=source[i+1]>>26;
-      opcode2[i+1]=source[i+1]&0x3f;
-      if((0<opcode[i+1]&&opcode[i+1]<8)||(opcode[i+1]==0&&(opcode2[i+1]==8||opcode2[i+1]==9))) {
-        printf("branch in delay slot @%08x (%08x)\n", addr + i*4+4, addr);
-        // don't handle first branch and call interpreter if it's hit
-        type=INTCALL;
-      }
-    }
-#endif
     itype[i]=type;
     opcode2[i]=op2;
     /* Get registers/immediates */
@@ -8463,19 +8527,43 @@ int new_recompile_block(int addr)
     else if(type==CJUMP||type==SJUMP||type==FJUMP)
       ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
     else ba[i]=-1;
-    /* Is this the end of the block? */
-    if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)) {
 #ifdef PCSX
-      // check for link register access in delay slot
-      int rt1_=rt1[i-1];
-      if(rt1_!=0&&(rs1[i]==rt1_||rs2[i]==rt1_||rt1[i]==rt1_||rt2[i]==rt1_)) {
-        printf("link access in delay slot @%08x (%08x)\n", addr + i*4, addr);
+    if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)) {
+      int do_in_intrp=0;
+      // branch in delay slot?
+      if(type==RJUMP||type==UJUMP||type==CJUMP||type==SJUMP||type==FJUMP) {
+        // don't handle first branch and call interpreter if it's hit
+        printf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
+        do_in_intrp=1;
+      }
+      // basic load delay detection
+      else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&rt1[i]!=0) {
+        int t=(ba[i-1]-start)/4;
+        if(0 <= t && t < i &&(rt1[i]==rs1[t]||rt1[i]==rs2[t])&&itype[t]!=CJUMP&&itype[t]!=SJUMP) {
+          // jump target wants DS result - potential load delay effect
+          printf("load delay @%08x (%08x)\n", addr + i*4, addr);
+          do_in_intrp=1;
+          bt[t+1]=1; // expected return from interpreter
+        }
+        else if(i>=2&&rt1[i-2]==2&&rt1[i]==2&&rs1[i]!=2&&rs2[i]!=2&&rs1[i-1]!=2&&rs2[i-1]!=2&&
+              !(i>=3&&(itype[i-3]==RJUMP||itype[i-3]==UJUMP||itype[i-3]==CJUMP||itype[i-3]==SJUMP))) {
+          // v0 overwrite like this is a sign of trouble, bail out
+          printf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
+          do_in_intrp=1;
+        }
+      }
+      if(do_in_intrp) {
+        rs1[i-1]=CCREG;
+        rs2[i-1]=rt1[i-1]=rt2[i-1]=0;
         ba[i-1]=-1;
         itype[i-1]=INTCALL;
         done=2;
+        i--; // don't compile the DS
       }
-      else
+    }
 #endif
+    /* Is this the end of the block? */
+    if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)) {
       if(rt1[i-1]==0) { // Continue past subroutine call (JAL)
         done=2;
       }
@@ -8798,17 +8886,18 @@ int new_recompile_block(int addr)
           clear_const(&current,rt1[i]);
           alloc_cc(&current,i);
           dirty_reg(&current,CCREG);
+          ooo[i]=1;
+          delayslot_alloc(&current,i+1);
           if (rt1[i]==31) {
             alloc_reg(&current,i,31);
             dirty_reg(&current,31);
-            assert(rs1[i+1]!=31&&rs2[i+1]!=31);
-            assert(rt1[i+1]!=rt1[i]);
+            //assert(rs1[i+1]!=31&&rs2[i+1]!=31);
+            //assert(rt1[i+1]!=rt1[i]);
             #ifdef REG_PREFETCH
             alloc_reg(&current,i,PTEMP);
             #endif
             //current.is32|=1LL<<rt1[i];
           }
-          delayslot_alloc(&current,i+1);
           //current.isconst=0; // DEBUG
           ds=1;
           //printf("i=%d, isconst=%x\n",i,current.isconst);
@@ -8852,6 +8941,7 @@ int new_recompile_block(int addr)
             alloc_reg(&current,i,RTEMP);
           }
           //current.isconst=0; // DEBUG
+          ooo[i]=1;
           ds=1;
           break;
         case CJUMP:
@@ -8875,9 +8965,6 @@ int new_recompile_block(int addr)
                (rs2[i]&&(rs2[i]==rt1[i+1]||rs2[i]==rt2[i+1]))) {
               // The delay slot overwrites one of our conditions.
               // Allocate the branch condition registers instead.
-              // Note that such a sequence of instructions could
-              // be considered a bug since the branch can not be
-              // re-executed if an exception occurs.
               current.isconst=0;
               current.wasconst=0;
               regs[i].wasconst=0;
@@ -8889,7 +8976,11 @@ int new_recompile_block(int addr)
                 if(rs2[i]) alloc_reg64(&current,i,rs2[i]);
               }
             }
-            else delayslot_alloc(&current,i+1);
+            else
+            {
+              ooo[i]=1;
+              delayslot_alloc(&current,i+1);
+            }
           }
           else
           if((opcode[i]&0x3E)==6) // BLEZ/BGTZ
@@ -8904,9 +8995,6 @@ int new_recompile_block(int addr)
             if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) {
               // The delay slot overwrites one of our conditions.
               // Allocate the branch condition registers instead.
-              // Note that such a sequence of instructions could
-              // be considered a bug since the branch can not be
-              // re-executed if an exception occurs.
               current.isconst=0;
               current.wasconst=0;
               regs[i].wasconst=0;
@@ -8916,7 +9004,11 @@ int new_recompile_block(int addr)
                 if(rs1[i]) alloc_reg64(&current,i,rs1[i]);
               }
             }
-            else delayslot_alloc(&current,i+1);
+            else
+            {
+              ooo[i]=1;
+              delayslot_alloc(&current,i+1);
+            }
           }
           else
           // Don't alloc the delay slot yet because we might not execute it
@@ -8976,12 +9068,9 @@ int new_recompile_block(int addr)
               //#endif
               //current.is32|=1LL<<rt1[i];
             }
-            if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) {
-              // The delay slot overwrites the branch condition.
+            if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) // The delay slot overwrites the branch condition.
+               ||(rt1[i]==31&&(rs1[i+1]==31||rs2[i+1]==31||rt1[i+1]==31||rt2[i+1]==31))) { // DS touches $ra
               // Allocate the branch condition registers instead.
-              // Note that such a sequence of instructions could
-              // be considered a bug since the branch can not be
-              // re-executed if an exception occurs.
               current.isconst=0;
               current.wasconst=0;
               regs[i].wasconst=0;
@@ -8991,7 +9080,11 @@ int new_recompile_block(int addr)
                 if(rs1[i]) alloc_reg64(&current,i,rs1[i]);
               }
             }
-            else delayslot_alloc(&current,i+1);
+            else
+            {
+              ooo[i]=1;
+              delayslot_alloc(&current,i+1);
+            }
           }
           else
           // Don't alloc the delay slot yet because we might not execute it
@@ -9029,15 +9122,13 @@ int new_recompile_block(int addr)
             if(itype[i+1]==FCOMP) {
               // The delay slot overwrites the branch condition.
               // Allocate the branch condition registers instead.
-              // Note that such a sequence of instructions could
-              // be considered a bug since the branch can not be
-              // re-executed if an exception occurs.
               alloc_cc(&current,i);
               dirty_reg(&current,CCREG);
               alloc_reg(&current,i,CSREG);
               alloc_reg(&current,i,FSREG);
             }
             else {
+              ooo[i]=1;
               delayslot_alloc(&current,i+1);
               alloc_reg(&current,i+1,CSREG);
             }
@@ -9477,6 +9568,16 @@ int new_recompile_block(int addr)
     {
       cc=0;
     }
+#ifdef PCSX
+    else if(/*itype[i]==LOAD||*/itype[i]==STORE||itype[i]==C1LS) // load causes weird timing issues
+    {
+      cc+=2; // 2 cycle penalty (after CLOCK_DIVIDER)
+    }
+    else if(itype[i]==C2LS)
+    {
+      cc+=4;
+    }
+#endif
     else
     {
       cc++;
@@ -9828,22 +9929,26 @@ int new_recompile_block(int addr)
               f_regmap[hr]=branch_regs[i].regmap[hr];
             }
           }
-          if(itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS
-          ||itype[i+1]==SHIFT||itype[i+1]==COP1||itype[i+1]==FLOAT
-          ||itype[i+1]==FCOMP||itype[i+1]==FCONV
-          ||itype[i+1]==COP2||itype[i+1]==C2LS||itype[i+1]==C2OP)
-          {
-            // Test both in case the delay slot is ooo,
-            // could be done better...
-            if(count_free_regs(branch_regs[i].regmap)<2
-             ||count_free_regs(regs[i].regmap)<2) 
+          if(ooo[i]) {
+            if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1]) 
+              f_regmap[hr]=branch_regs[i].regmap[hr];
+          }else{
+            if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1]) 
               f_regmap[hr]=branch_regs[i].regmap[hr];
           }
           // Avoid dirty->clean transition
-          // #ifdef DESTRUCTIVE_WRITEBACK here?
+          #ifdef DESTRUCTIVE_WRITEBACK
           if(t>0) if(get_reg(regmap_pre[t],f_regmap[hr])>=0) if((regs[t].wasdirty>>get_reg(regmap_pre[t],f_regmap[hr]))&1) f_regmap[hr]=-1;
+          #endif
+          // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
+          // case above, however it's always a good idea.  We can't hoist the
+          // load if the register was already allocated, so there's no point
+          // wasting time analyzing most of these cases.  It only "succeeds"
+          // when the mapping was different and the load can be replaced with
+          // a mov, which is of negligible benefit.  So such cases are
+          // skipped below.
           if(f_regmap[hr]>0) {
-            if(regs[t].regmap_entry[hr]<0) {
+            if(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0) {
               int r=f_regmap[hr];
               for(j=t;j<=i;j++)
               {
@@ -9855,6 +9960,7 @@ int new_recompile_block(int addr)
                   // register is lower numbered than the lower-half
                   // register.  Not sure if it's worth fixing...
                   if(get_reg(regs[j].regmap,r&63)<0) break;
+                  if(get_reg(regs[j].regmap_entry,r&63)<0) break;
                   if(regs[j].is32&(1LL<<(r&63))) break;
                 }
                 if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
@@ -9868,17 +9974,10 @@ int new_recompile_block(int addr)
                     }
                     k=i;
                     while(k>1&&regs[k-1].regmap[hr]==-1) {
-                      if(itype[k-1]==STORE||itype[k-1]==STORELR
-                      ||itype[k-1]==C1LS||itype[k-1]==SHIFT||itype[k-1]==COP1
-                      ||itype[k-1]==FLOAT||itype[k-1]==FCONV||itype[k-1]==FCOMP
-                      ||itype[k-1]==COP2||itype[k-1]==C2LS||itype[k-1]==C2OP) {
-                        if(count_free_regs(regs[k-1].regmap)<2) {
-                          //printf("no free regs for store %x\n",start+(k-1)*4);
-                          break;
-                        }
+                      if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
+                        //printf("no free regs for store %x\n",start+(k-1)*4);
+                        break;
                       }
-                      else
-                      if(itype[k-1]!=NOP&&itype[k-1]!=MOV&&itype[k-1]!=ALU&&itype[k-1]!=SHIFTIMM&&itype[k-1]!=IMM16&&itype[k-1]!=LOAD) break;
                       if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
                         //printf("no-match due to different register\n");
                         break;
@@ -9955,13 +10054,31 @@ int new_recompile_block(int addr)
                     }
                   }
                   for(k=t;k<j;k++) {
+                    // Alloc register clean at beginning of loop,
+                    // but may dirty it in pass 6
                     regs[k].regmap_entry[hr]=f_regmap[hr];
                     regs[k].regmap[hr]=f_regmap[hr];
-                    regmap_pre[k+1][hr]=f_regmap[hr];
-                    regs[k+1].wasdirty&=~(1<<hr);
                     regs[k].dirty&=~(1<<hr);
                     regs[k].wasconst&=~(1<<hr);
                     regs[k].isconst&=~(1<<hr);
+                    if(itype[k]==UJUMP||itype[k]==RJUMP||itype[k]==CJUMP||itype[k]==SJUMP||itype[k]==FJUMP) {
+                      branch_regs[k].regmap_entry[hr]=f_regmap[hr];
+                      branch_regs[k].regmap[hr]=f_regmap[hr];
+                      branch_regs[k].dirty&=~(1<<hr);
+                      branch_regs[k].wasconst&=~(1<<hr);
+                      branch_regs[k].isconst&=~(1<<hr);
+                      if(itype[k]!=RJUMP&&itype[k]!=UJUMP&&(source[k]>>16)!=0x1000) {
+                        regmap_pre[k+2][hr]=f_regmap[hr];
+                        regs[k+2].wasdirty&=~(1<<hr);
+                        assert((branch_regs[k].is32&(1LL<<f_regmap[hr]))==
+                          (regs[k+2].was32&(1LL<<f_regmap[hr])));
+                      }
+                    }
+                    else
+                    {
+                      regmap_pre[k+1][hr]=f_regmap[hr];
+                      regs[k+1].wasdirty&=~(1<<hr);
+                    }
                   }
                   if(regs[j].regmap[hr]==f_regmap[hr])
                     regs[j].regmap_entry[hr]=f_regmap[hr];
@@ -9978,16 +10095,29 @@ int new_recompile_block(int addr)
                   //printf("32/64 mismatch %x %d\n",start+j*4,hr);
                   break;
                 }
-                if(itype[j]==STORE||itype[j]==STORELR||itype[j]==C1LS
-                ||itype[j]==SHIFT||itype[j]==COP1||itype[j]==FLOAT
-                ||itype[j]==FCOMP||itype[j]==FCONV
-                ||itype[j]==COP2||itype[j]==C2LS||itype[j]==C2OP) {
-                  if(count_free_regs(regs[j].regmap)<2) {
-                    //printf("No free regs for store %x\n",start+j*4);
+                if(itype[j]==UJUMP||itype[j]==RJUMP||(source[j]>>16)==0x1000)
+                {
+                  // Stop on unconditional branch
+                  break;
+                }
+                if(itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP)
+                {
+                  if(ooo[j]) {
+                    if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1]) 
+                      break;
+                  }else{
+                    if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1]) 
+                      break;
+                  }
+                  if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
+                    //printf("no-match due to different register (branch)\n");
                     break;
                   }
                 }
-                else if(itype[j]!=NOP&&itype[j]!=MOV&&itype[j]!=ALU&&itype[j]!=SHIFTIMM&&itype[j]!=IMM16&&itype[j]!=LOAD) break;
+                if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
+                  //printf("No free regs for store %x\n",start+j*4);
+                  break;
+                }
                 if(f_regmap[hr]>=64) {
                   if(regs[j].is32&(1LL<<(f_regmap[hr]&63))) {
                     break;
@@ -10032,17 +10162,10 @@ int new_recompile_block(int addr)
       if(bt[i]) {
         for(j=i;j<slen-1;j++) {
           if(regs[j].regmap[HOST_CCREG]!=-1) break;
-          if(itype[j]==STORE||itype[j]==STORELR||itype[j]==C1LS
-          ||itype[j]==SHIFT||itype[j]==COP1||itype[j]==FLOAT
-          ||itype[j]==FCOMP||itype[j]==FCONV
-          ||itype[j]==COP2||itype[j]==C2LS||itype[j]==C2OP) {
-            if(count_free_regs(regs[j].regmap)<2) {
-              //printf("no free regs for store %x\n",start+j*4);
-              break;
-            }
+          if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
+            //printf("no free regs for store %x\n",start+j*4);
+            break;
           }
-          else
-          if(itype[j]!=NOP&&itype[j]!=MOV&&itype[j]!=ALU&&itype[j]!=SHIFTIMM&&itype[j]!=IMM16&&itype[j]!=LOAD) break;
         }
         if(regs[j].regmap[HOST_CCREG]==CCREG) {
           int k=i;
@@ -10066,17 +10189,10 @@ int new_recompile_block(int addr)
           int k;
           k=i;
           while(regs[k-1].regmap[HOST_CCREG]==-1) {
-            if(itype[k-1]==STORE||itype[k-1]==STORELR||itype[k-1]==C1LS
-            ||itype[k-1]==SHIFT||itype[k-1]==COP1||itype[k-1]==FLOAT
-            ||itype[k-1]==FCONV||itype[k-1]==FCOMP
-            ||itype[k-1]==COP2||itype[k-1]==C2LS||itype[k-1]==C2OP) {
-              if(count_free_regs(regs[k-1].regmap)<2) {
-                //printf("no free regs for store %x\n",start+(k-1)*4);
-                break;
-              }
+            if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
+              //printf("no free regs for store %x\n",start+(k-1)*4);
+              break;
             }
-            else
-            if(itype[k-1]!=NOP&&itype[k-1]!=MOV&&itype[k-1]!=ALU&&itype[k-1]!=SHIFTIMM&&itype[k-1]!=IMM16&&itype[k-1]!=LOAD) break;
             k--;
           }
           if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
@@ -10100,8 +10216,7 @@ int new_recompile_block(int addr)
       if(itype[i]!=STORE&&itype[i]!=STORELR&&itype[i]!=C1LS&&itype[i]!=SHIFT&&
          itype[i]!=NOP&&itype[i]!=MOV&&itype[i]!=ALU&&itype[i]!=SHIFTIMM&&
          itype[i]!=IMM16&&itype[i]!=LOAD&&itype[i]!=COP1&&itype[i]!=FLOAT&&
-         itype[i]!=FCONV&&itype[i]!=FCOMP&&
-         itype[i]!=COP2&&itype[i]!=C2LS&&itype[i]!=C2OP)
+         itype[i]!=FCONV&&itype[i]!=FCOMP)
       {
         memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
       }
@@ -10708,9 +10823,9 @@ int new_recompile_block(int addr)
       if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
       {
         // Load the delay slot registers if necessary
-        if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i])
+        if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i]&&(rs1[i+1]!=rt1[i]||rt1[i]==0))
           load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i+1],rs1[i+1]);
-        if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i])
+        if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i]&&(rs2[i+1]!=rt1[i]||rt1[i]==0))
           load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs2[i+1],rs2[i+1]);
         if(itype[i+1]==STORE||itype[i+1]==STORELR||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a)
           load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,INVCP,INVCP);
@@ -10985,6 +11100,12 @@ int new_recompile_block(int addr)
     }
 #endif
   }
+#ifdef PCSX
+  // PCSX maps all RAM mirror invalid_code tests to 0x80000000..0x80000000+RAM_SIZE
+  if(get_page(start)<(RAM_SIZE>>12))
+    for(i=start>>12;i<=(start+slen*4)>>12;i++)
+      invalid_code[((u_int)0x80000000>>12)|i]=0;
+#endif
   
   /* Pass 10 - Free memory by expiring oldest blocks */