drc: merge Ari64's patch: 12_cycle_count

[pcsx_rearmed.git] / libpcsxcore / new_dynarec / new_dynarec.c

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index 34673b9..7727384 100644 (file)
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -1,6 +1,6 @@
 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
  *   Mupen64plus - new_dynarec.c                                           *
- *   Copyright (C) 2009-2010 Ari64                                         *
+ *   Copyright (C) 2009-2011 Ari64                                         *
  *                                                                         *
  *   This program is free software; you can redistribute it and/or modify  *
  *   it under the terms of the GNU General Public License as published by  *
@@ -126,6 +126,7 @@ struct ll_entry
 #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;
@@ -598,6 +599,7 @@ void clear_const(struct regstat *cur,signed char reg)
 int is_const(struct regstat *cur,signed char reg)
 {
   int hr;
+  if(reg<0) return 0;
   if(!reg) return 1;
   for (hr=0;hr<HOST_REGS;hr++) {
     if((cur->regmap[hr]&63)==reg) {
@@ -718,12 +720,6 @@ int needed_again(int r, int i)
   int j;
   int b=-1;
   int rn=10;
-  int hr;
-  u_char hsn[MAXREG+1];
-  int preferred_reg;
-  
-  memset(hsn,10,sizeof(hsn));
-  lsn(hsn,i,&preferred_reg);
   
   if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000))
   {
@@ -776,11 +772,7 @@ int needed_again(int r, int i)
       }
     }
   }*/
-  for(hr=0;hr<HOST_REGS;hr++) {
-    if(hr!=EXCLUDE_REG) {
-      if(rn<hsn[hr]) return 1;
-    }
-  }
+  if(rn<10) return 1;
   return 0;
 }
 
@@ -2826,6 +2818,13 @@ 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
@@ -2865,6 +2864,9 @@ 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);
         }
@@ -2890,6 +2892,9 @@ 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);
@@ -2915,13 +2920,17 @@ 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);
@@ -2947,6 +2956,9 @@ 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);
         }
@@ -2972,6 +2984,9 @@ 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);
@@ -2998,13 +3013,17 @@ 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);
@@ -3017,6 +3036,10 @@ 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);
@@ -3025,7 +3048,7 @@ 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);
@@ -3105,9 +3128,15 @@ 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);
@@ -3115,6 +3144,7 @@ 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
       {
@@ -3146,6 +3176,9 @@ void store_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
       //gen_tlb_addr_w(temp,map);
       //emit_writebyte_indexed(tl,(int)rdram-0x80000000,temp);
@@ -3161,6 +3194,9 @@ void store_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_writehword_indexed_tlb(tl,x,temp,map,temp);
@@ -3174,23 +3210,32 @@ void store_assemble(int i,struct regstat *i_regs)
     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;
@@ -3690,7 +3735,7 @@ void c2ls_assemble(int i,struct regstat *i_regs)
   int ar;
   int offset;
   int memtarget=0,c=0;
-  int jaddr,jaddr2=0,jaddr3,type;
+  int jaddr2=0,jaddr3,type;
   int agr=AGEN1+(i&1);
   u_int hr,reglist=0;
   u_int copr=(source[i]>>16)&0x1f;
@@ -4060,7 +4105,7 @@ static void loop_preload(signed char pre[],signed char entry[])
 void address_generation(int i,struct regstat *i_regs,signed char entry[])
 {
   if(itype[i]==LOAD||itype[i]==LOADLR||itype[i]==STORE||itype[i]==STORELR||itype[i]==C1LS||itype[i]==C2LS) {
-    int ra;
+    int ra=-1;
     int agr=AGEN1+(i&1);
     int mgr=MGEN1+(i&1);
     if(itype[i]==LOAD) {
@@ -4812,7 +4857,7 @@ void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
   }
   else
   {
-    emit_cmpimm(HOST_CCREG,-2*(count+2));
+    emit_cmpimm(HOST_CCREG,-CLOCK_DIVIDER*(count+2));
     jaddr=(int)out;
     emit_jns(0);
   }
@@ -4878,7 +4923,7 @@ void do_ccstub(int n)
           emit_loadreg(rs2[i],s2l);
       #endif
       int hr=0;
-      int addr,alt,ntaddr;
+      int addr=-1,alt=-1,ntaddr=-1;
       while(hr<HOST_REGS)
       {
         if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
@@ -7718,10 +7763,9 @@ void disassemble_inst(int i)
 void new_dynarec_clear_full()
 {
   int n;
-  for(n=0x80000;n<0x80800;n++)
-    invalid_code[n]=1;
-  for(n=0;n<65536;n++)
-    hash_table[n][0]=hash_table[n][2]=-1;
+  out=(u_char *)BASE_ADDR;
+  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));
@@ -7734,6 +7778,7 @@ void new_dynarec_clear_full()
 #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
@@ -7843,6 +7888,11 @@ 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, psxRegs.pc);
+    sp_in_mirror=1;
+  }
   if (Config.HLE && start == 0x80001000) // hlecall
   {
     // XXX: is this enough? Maybe check hleSoftCall?
@@ -7958,17 +8008,10 @@ int new_recompile_block(int addr)
           case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
           case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
           case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
-          case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
-          case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
-          case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
           case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
           case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
           case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
           case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
-          case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
-          case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
-          case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
-          case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
           case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
           case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
           case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
@@ -7979,22 +8022,31 @@ int new_recompile_block(int addr)
           case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
           case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
           case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
-          case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
-          case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
-          case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
-          case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
           case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
           case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
           case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
           case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
           case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
           case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
+#ifndef FORCE32
+          case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
+          case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
+          case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
+          case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
+          case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
+          case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
+          case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
+          case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
+          case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
+          case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
+          case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
           case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
           case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
           case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
           case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
           case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
           case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
+#endif
         }
         break;
       case 0x01: strcpy(insn[i],"regimm"); type=NI;
@@ -8563,7 +8615,7 @@ int new_recompile_block(int addr)
   current.wasconst=0;
   int ds=0;
   int cc=0;
-  int hr;
+  int hr=-1;
 
 #ifndef FORCE32
   provisional_32bit();
@@ -8826,6 +8878,8 @@ 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);
@@ -8836,8 +8890,6 @@ int new_recompile_block(int addr)
             #endif
             //current.is32|=1LL<<rt1[i];
           }
-          ooo[i]=1;
-          delayslot_alloc(&current,i+1);
           //current.isconst=0; // DEBUG
           ds=1;
           //printf("i=%d, isconst=%x\n",i,current.isconst);
@@ -9580,8 +9632,8 @@ int new_recompile_block(int addr)
         }
       }
       // Don't need stuff which is overwritten
-      if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
-      if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
+      //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
+      //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
       // Merge in delay slot
       for(hr=0;hr<HOST_REGS;hr++)
       {
@@ -9700,7 +9752,10 @@ int new_recompile_block(int addr)
             if(likely[i]) {
               regs[i].regmap[hr]=-1;
               regs[i].isconst&=~(1<<hr);
-              if(i<slen-2) regmap_pre[i+2][hr]=-1;
+              if(i<slen-2) {
+                regmap_pre[i+2][hr]=-1;
+                regs[i+2].wasconst&=~(1<<hr);
+              }
             }
           }
         }
@@ -9755,6 +9810,7 @@ int new_recompile_block(int addr)
               {
                 if(!likely[i]&&i<slen-2) {
                   regmap_pre[i+2][hr]=-1;
+                  regs[i+2].wasconst&=~(1<<hr);
                 }
               }
             }
@@ -9800,6 +9856,7 @@ int new_recompile_block(int addr)
                 }
                 regmap_pre[i+1][hr]=-1;
                 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
+                regs[i+1].wasconst&=~(1<<hr);
               }
               regs[i].regmap[hr]=-1;
               regs[i].isconst&=~(1<<hr);
@@ -10476,6 +10533,19 @@ int new_recompile_block(int addr)
     }
     //requires_32bit[i]=is32[i]&~unneeded_reg_upper[i]; // DEBUG
   }
+#else
+  for (i=slen-1;i>=0;i--)
+  {
+    if(itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
+    {
+      // Conditional branch
+      if((source[i]>>16)!=0x1000&&i<slen-2) {
+        // Mark this address as a branch target since it may be called
+        // upon return from interrupt
+        bt[i+2]=1;
+      }
+    }
+  }
 #endif
 
   if(itype[slen-1]==SPAN) {