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=f1bdbfc25f16b3ebb42ae0cbaf84a6036b298980;hp=975cce78de20bfa07084a80565f45b6f164f2c93;hb=f776eb14d490a2c4391666ce5beefe058e7e03d5;hpb=dadf55f230286fe187e44370758723b35b8c7efb

diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index 975cce78..f1bdbfc2 100644
--- 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  *
@@ -599,6 +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) {
@@ -719,12 +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))
   {
@@ -777,11 +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;
 }
 
@@ -1602,13 +1593,9 @@ void load_alloc(struct regstat *current,int i)
   //if(rs1[i]!=rt1[i]&&needed_again(rs1[i],i)) clear_const(current,rs1[i]); // Does this help or hurt?
   if(!rs1[i]) current->u&=~1LL; // Allow allocating r0 if it's the source register
   if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
-  if(rt1[i]) {
+  if(rt1[i]&&!((current->u>>rt1[i])&1)) {
     alloc_reg(current,i,rt1[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;
-    }
+    assert(get_reg(current->regmap,rt1[i])>=0);
     if(opcode[i]==0x27||opcode[i]==0x37) // LWU/LD
     {
       current->is32&=~(1LL<<rt1[i]);
@@ -1636,12 +1623,14 @@ void load_alloc(struct regstat *current,int i)
   }
   else
   {
-    // Load to r0 (dummy load)
+    // Load to r0 or unneeded register (dummy load)
     // but we still need a register to calculate the address
     if(opcode[i]==0x22||opcode[i]==0x26)
     {
       alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
     }
+    // If using TLB, need a register for pointer to the mapping table
+    if(using_tlb) alloc_reg(current,i,TLREG);
     alloc_reg_temp(current,i,-1);
     minimum_free_regs[i]=1;
     if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
@@ -2851,6 +2840,7 @@ void load_assemble(int i,struct regstat *i_regs)
     if (opcode[i]==0x21||opcode[i]==0x25) x=2; // LH/LHU
     map=get_reg(i_regs->regmap,TLREG);
     assert(map>=0);
+    reglist&=~(1<<map);
     map=do_tlb_r(addr,tl,map,x,-1,-1,c,constmap[i][s]+offset);
     do_tlb_r_branch(map,c,constmap[i][s]+offset,&jaddr);
   }
@@ -3173,6 +3163,7 @@ void store_assemble(int i,struct regstat *i_regs)
     if (opcode[i]==0x29) x=2; // SH
     map=get_reg(i_regs->regmap,TLREG);
     assert(map>=0);
+    reglist&=~(1<<map);
     map=do_tlb_w(addr,temp,map,x,c,constmap[i][s]+offset);
     do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
   }
@@ -3363,6 +3354,7 @@ void storelr_assemble(int i,struct regstat *i_regs)
   }else{ // using tlb
     int map=get_reg(i_regs->regmap,TLREG);
     assert(map>=0);
+    reglist&=~(1<<map);
     map=do_tlb_w(c||s<0||offset?temp:s,temp,map,0,c,constmap[i][s]+offset);
     if(!c&&!offset&&s>=0) emit_mov(s,temp);
     do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
@@ -3613,6 +3605,7 @@ void c1ls_assemble(int i,struct regstat *i_regs)
   {
     map=get_reg(i_regs->regmap,TLREG);
     assert(map>=0);
+    reglist&=~(1<<map);
     if (opcode[i]==0x31||opcode[i]==0x35) { // LWC1/LDC1
       map=do_tlb_r(offset||c||s<0?ar:s,ar,map,0,-1,-1,c,constmap[i][s]+offset);
     }
@@ -3744,7 +3737,7 @@ 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;
@@ -4114,7 +4107,7 @@ 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) {
@@ -4490,7 +4483,7 @@ void load_all_regs(signed char i_regmap[])
         emit_zeroreg(hr);
       }
       else
-      if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG)
+      if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
       {
         emit_loadreg(i_regmap[hr],hr);
       }
@@ -4509,7 +4502,7 @@ void load_needed_regs(signed char i_regmap[],signed char next_regmap[])
           emit_zeroreg(hr);
         }
         else
-        if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG)
+        if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
         {
           emit_loadreg(i_regmap[hr],hr);
         }
@@ -4529,7 +4522,7 @@ void load_regs_entry(int t)
   }
   // Load 32-bit regs
   for(hr=0;hr<HOST_REGS;hr++) {
-    if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<64) {
+    if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
       if(regs[t].regmap_entry[hr]==0) {
         emit_zeroreg(hr);
       }
@@ -4541,7 +4534,7 @@ void load_regs_entry(int t)
   }
   // Load 64-bit regs
   for(hr=0;hr<HOST_REGS;hr++) {
-    if(regs[t].regmap_entry[hr]>=64) {
+    if(regs[t].regmap_entry[hr]>=64&&regs[t].regmap_entry[hr]<TEMPREG+64) {
       assert(regs[t].regmap_entry[hr]!=64);
       if((regs[t].was32>>(regs[t].regmap_entry[hr]&63))&1) {
         int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
@@ -4621,7 +4614,7 @@ void load_regs_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int ad
     }
     // Load 32-bit regs
     for(hr=0;hr<HOST_REGS;hr++) {
-      if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<64) {
+      if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
         #ifdef DESTRUCTIVE_WRITEBACK
         if(i_regmap[hr]!=regs[t].regmap_entry[hr] || ( !((regs[t].dirty>>hr)&1) && ((i_dirty>>hr)&1) && (((i_is32&~unneeded_reg_upper[t])>>i_regmap[hr])&1) ) || (((i_is32&~regs[t].was32&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)) {
         #else
@@ -4639,7 +4632,7 @@ void load_regs_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int ad
     }
     //Load 64-bit regs
     for(hr=0;hr<HOST_REGS;hr++) {
-      if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=64) {
+      if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=64&&regs[t].regmap_entry[hr]<TEMPREG+64) {
         if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
           assert(regs[t].regmap_entry[hr]!=64);
           if((i_is32>>(regs[t].regmap_entry[hr]&63))&1) {
@@ -4680,19 +4673,19 @@ int match_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
       {
         if(i_regmap[hr]!=regs[t].regmap_entry[hr])
         {
-          if(regs[t].regmap_entry[hr]!=-1)
+          if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
           {
             return 0;
           }
           else 
           if((i_dirty>>hr)&1)
           {
-            if(i_regmap[hr]<64)
+            if(i_regmap[hr]<TEMPREG)
             {
               if(!((unneeded_reg[t]>>i_regmap[hr])&1))
                 return 0;
             }
-            else
+            else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
             {
               if(!((unneeded_reg_upper[t]>>(i_regmap[hr]&63))&1))
                 return 0;
@@ -4866,7 +4859,7 @@ 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);
   }
@@ -4932,7 +4925,7 @@ 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 &&
@@ -7649,7 +7642,7 @@ void clean_registers(int istart,int iend,int wr)
             regs[i].wasdirty|=will_dirty_i&(1<<r);
           }
         }
-        else if((nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
+        else if(regmap_pre[i][r]>=0&&(nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
           // Register moved to a different register
           will_dirty_i&=~(1<<r);
           wont_dirty_i&=~(1<<r);
@@ -7772,10 +7765,9 @@ 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));
@@ -7900,7 +7892,7 @@ int new_recompile_block(int addr)
 #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);
+    printf("SP hack enabled (%08x), @%08x\n", psxRegs.GPR.n.sp, psxRegs.pc);
     sp_in_mirror=1;
   }
   if (Config.HLE && start == 0x80001000) // hlecall
@@ -8018,17 +8010,10 @@ 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;
@@ -8039,22 +8024,31 @@ 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;
@@ -8623,7 +8617,7 @@ int new_recompile_block(int addr)
   current.wasconst=0;
   int ds=0;
   int cc=0;
-  int hr;
+  int hr=-1;
 
 #ifndef FORCE32
   provisional_32bit();
@@ -8886,6 +8880,8 @@ 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);
@@ -8896,8 +8892,6 @@ 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);
@@ -9640,8 +9634,8 @@ 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++)
       {
@@ -9760,7 +9754,10 @@ 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);
+              }
             }
           }
         }
@@ -9815,6 +9812,7 @@ int new_recompile_block(int addr)
               {
                 if(!likely[i]&&i<slen-2) {
                   regmap_pre[i+2][hr]=-1;
+                  regs[i+2].wasconst&=~(1<<hr);
                 }
               }
             }
@@ -9860,6 +9858,7 @@ 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);
@@ -10536,6 +10535,19 @@ 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) {
@@ -10801,8 +10813,13 @@ int new_recompile_block(int addr)
         wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,is32_pre,
               unneeded_reg[i],unneeded_reg_upper[i]);
       }
-      is32_pre=regs[i].is32;
-      dirty_pre=regs[i].dirty;
+      if((itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)&&!likely[i]) {
+        is32_pre=branch_regs[i].is32;
+        dirty_pre=branch_regs[i].dirty;
+      }else{
+        is32_pre=regs[i].is32;
+        dirty_pre=regs[i].dirty;
+      }
       #endif
       // write back
       if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))