drc: fix bad gte unneeded reg assumption

[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 353179a..37706bc 100644 (file)
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -44,9 +44,6 @@
 #define MAXBLOCK 4096
 #define MAX_OUTPUT_BLOCK_SIZE 262144
 
-int cycle_multiplier; // 100 for 1.0
-#define CLOCK_ADJUST(x) (((x) * cycle_multiplier + 50) / 100)
-
 struct regstat
 {
   signed char regmap_entry[HOST_REGS];
@@ -61,7 +58,6 @@ struct regstat
   u_int isconst;
   u_int loadedconst;             // host regs that have constants loaded
   u_int waswritten;              // MIPS regs that were used as store base before
-  uint64_t constmap[HOST_REGS];
 };
 
 struct ll_entry
@@ -92,7 +88,6 @@ struct ll_entry
   static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
   static uint64_t gte_rt[MAXBLOCK];
   static uint64_t gte_unneeded[MAXBLOCK];
-  static int gte_reads_flags; // gte flag read encountered
   static u_int smrv[32]; // speculated MIPS register values
   static u_int smrv_strong; // mask or regs that are likely to have correct values
   static u_int smrv_weak; // same, but somewhat less likely
@@ -110,11 +105,10 @@ struct ll_entry
   uint64_t p32[MAXBLOCK];
   uint64_t pr32[MAXBLOCK];
   signed char regmap_pre[MAXBLOCK][HOST_REGS];
-  signed char regmap[MAXBLOCK][HOST_REGS];
-  signed char regmap_entry[MAXBLOCK][HOST_REGS];
-  uint64_t constmap[MAXBLOCK][HOST_REGS];
-  struct regstat regs[MAXBLOCK];
-  struct regstat branch_regs[MAXBLOCK];
+  static uint64_t current_constmap[HOST_REGS];
+  static uint64_t constmap[MAXBLOCK][HOST_REGS];
+  static struct regstat regs[MAXBLOCK];
+  static struct regstat branch_regs[MAXBLOCK];
   signed char minimum_free_regs[MAXBLOCK];
   u_int needed_reg[MAXBLOCK];
   uint64_t requires_32bit[MAXBLOCK];
@@ -145,6 +139,7 @@ struct ll_entry
   static const u_int using_tlb=0;
 #endif
   int new_dynarec_did_compile;
+  int new_dynarec_hacks;
   u_int stop_after_jal;
   extern u_char restore_candidate[512];
   extern int cycle_count;
@@ -279,6 +274,14 @@ int tracedebug=0;
 
 //#define DEBUG_CYCLE_COUNT 1
 
+int cycle_multiplier; // 100 for 1.0
+
+static int CLOCK_ADJUST(int x)
+{
+  int s=(x>>31)|1;
+  return (x * cycle_multiplier + s * 50) / 100;
+}
+
 static void tlb_hacks()
 {
 #ifndef DISABLE_TLB
@@ -343,6 +346,7 @@ static u_int get_page(u_int vaddr)
   return page;
 }
 
+#ifndef PCSX
 static u_int get_vpage(u_int vaddr)
 {
   u_int vpage=(vaddr^0x80000000)>>12;
@@ -352,6 +356,13 @@ static u_int get_vpage(u_int vaddr)
   if(vpage>2048) vpage=2048+(vpage&2047);
   return vpage;
 }
+#else
+// no virtual mem in PCSX
+static u_int get_vpage(u_int vaddr)
+{
+  return get_page(vaddr);
+}
+#endif
 
 // Get address from virtual address
 // This is called from the recompiled JR/JALR instructions
@@ -592,11 +603,11 @@ void set_const(struct regstat *cur,signed char reg,uint64_t value)
   for (hr=0;hr<HOST_REGS;hr++) {
     if(cur->regmap[hr]==reg) {
       cur->isconst|=1<<hr;
-      cur->constmap[hr]=value;
+      current_constmap[hr]=value;
     }
     else if((cur->regmap[hr]^64)==reg) {
       cur->isconst|=1<<hr;
-      cur->constmap[hr]=value>>32;
+      current_constmap[hr]=value>>32;
     }
   }
 }
@@ -630,7 +641,7 @@ uint64_t get_const(struct regstat *cur,signed char reg)
   if(!reg) return 0;
   for (hr=0;hr<HOST_REGS;hr++) {
     if(cur->regmap[hr]==reg) {
-      return cur->constmap[hr];
+      return current_constmap[hr];
     }
   }
   printf("Unknown constant in r%d\n",reg);
@@ -1226,7 +1237,7 @@ 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; }
-  u_int page=get_page(addr);
+  u_int page=get_vpage(addr);
   if(page<2048) { // RAM
     struct ll_entry *head;
     u_int addr_min=~0, addr_max=0;
@@ -1265,11 +1276,9 @@ void invalidate_addr(u_int addr)
       return;
     }
     else {
-      inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);//rhits);
-    }
-    //rhits=0;
-    if(page!=0) // FIXME: don't know what's up with page 0 (Klonoa)
+      inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);
       return;
+    }
   }
 #endif
   invalidate_block(addr>>12);
@@ -3300,7 +3309,7 @@ void store_assemble(int i,struct regstat *i_regs)
     jaddr=0;
   }
 #endif
-  if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))) {
+  if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
     if(!c||memtarget) {
       #ifdef DESTRUCTIVE_SHIFT
       // The x86 shift operation is 'destructive'; it overwrites the
@@ -3574,7 +3583,7 @@ void storelr_assemble(int i,struct regstat *i_regs)
   }
   if(!c||!memtarget)
     add_stub(STORELR_STUB,jaddr,(int)out,i,(int)i_regs,temp,ccadj[i],reglist);
-  if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))) {
+  if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
     #ifdef RAM_OFFSET
     int map=get_reg(i_regs->regmap,ROREG);
     if(map<0) map=HOST_TEMPREG;
@@ -3752,7 +3761,7 @@ void c1ls_assemble(int i,struct regstat *i_regs)
     emit_writedword_indexed_tlb(th,tl,0,offset||c||s<0?temp:s,map,temp);
     type=STORED_STUB;
   }
-  if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))) {
+  if(!using_tlb&&!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
     if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
       #ifndef DESTRUCTIVE_SHIFT
       temp=offset||c||s<0?ar:s;
@@ -3871,7 +3880,8 @@ void c2ls_assemble(int i,struct regstat *i_regs)
   }
   if(jaddr2)
     add_stub(type,jaddr2,(int)out,i,ar,(int)i_regs,ccadj[i],reglist);
-  if (!(i_regs->waswritten&(1<<rs1[i]))&&opcode[i]==0x3a) { // SWC2
+  if(opcode[i]==0x3a) // SWC2
+  if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
 #if defined(HOST_IMM8)
     int ir=get_reg(i_regs->regmap,INVCP);
     assert(ir>=0);
@@ -6770,16 +6780,20 @@ void unneeded_registers(int istart,int iend,int r)
 {
   int i;
   uint64_t u,uu,gte_u,b,bu,gte_bu;
-  uint64_t temp_u,temp_uu,temp_gte_u;
+  uint64_t temp_u,temp_uu,temp_gte_u=0;
   uint64_t tdep;
+  uint64_t gte_u_unknown=0;
+  if(new_dynarec_hacks&NDHACK_GTE_UNNEEDED)
+    gte_u_unknown=~0ll;
   if(iend==slen-1) {
     u=1;uu=1;
+    gte_u=gte_u_unknown;
   }else{
     u=unneeded_reg[iend+1];
     uu=unneeded_reg_upper[iend+1];
     u=1;uu=1;
+    gte_u=gte_unneeded[iend+1];
   }
-  gte_u=temp_gte_u=0;
 
   for (i=iend;i>=istart;i--)
   {
@@ -6794,7 +6808,7 @@ void unneeded_registers(int istart,int iend,int r)
         // Branch out of this block, flush all regs
         u=1;
         uu=1;
-        gte_u=0;
+        gte_u=gte_u_unknown;
         /* Hexagon hack 
         if(itype[i]==UJUMP&&rt1[i]==31)
         {
@@ -6840,7 +6854,7 @@ void unneeded_registers(int istart,int iend,int r)
           {
             u=1;
             uu=1;
-            gte_u=0;
+            gte_u=gte_u_unknown;
           }
         }
       }
@@ -6883,7 +6897,7 @@ void unneeded_registers(int istart,int iend,int r)
             {
               temp_u=1;
               temp_uu=1;
-              temp_gte_u=0;
+              temp_gte_u=gte_u_unknown;
             }
           }
           tdep=(~temp_uu>>rt1[i])&1;
@@ -6905,7 +6919,7 @@ void unneeded_registers(int istart,int iend,int r)
           }else{
             unneeded_reg[(ba[i]-start)>>2]=1;
             unneeded_reg_upper[(ba[i]-start)>>2]=1;
-            gte_unneeded[(ba[i]-start)>>2]=0;
+            gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
           }
         } /*else*/ if(1) {
           if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
@@ -7010,6 +7024,8 @@ void unneeded_registers(int istart,int iend,int r)
     uu&=~(1LL<<us1[i]);
     uu&=~(1LL<<us2[i]);
     gte_u&=~gte_rs[i];
+    if(gte_rs[i]&&rt1[i]&&(unneeded_reg[i+1]&(1ll<<rt1[i])))
+      gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
     // Source-target dependencies
     uu&=~(tdep<<dep1[i]);
     uu&=~(tdep<<dep2[i]);
@@ -7944,7 +7960,6 @@ void new_dynarec_clear_full()
   literalcount=0;
   stop_after_jal=0;
   inv_code_start=inv_code_end=~0;
-  gte_reads_flags=0;
   // TLB
 #ifndef DISABLE_TLB
   using_tlb=0;
@@ -8654,12 +8669,7 @@ int new_recompile_block(int addr)
         {
           case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
           case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
-          case 0x02: gte_rs[i]=1ll<<(gr+32); // CFC2
-            if(gr==31&&!gte_reads_flags) {
-              assem_debug("gte flag read encountered @%08x\n",addr + i*4);
-              gte_reads_flags=1;
-            }
-            break;
+          case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
           case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
         }
         break;
@@ -8687,6 +8697,12 @@ int new_recompile_block(int addr)
         gte_rs[i]=gte_reg_reads[source[i]&0x3f];
         gte_rt[i]=gte_reg_writes[source[i]&0x3f];
         gte_rt[i]|=1ll<<63; // every op changes flags
+        if((source[i]&0x3f)==GTE_MVMVA) {
+          int v = (source[i] >> 15) & 3;
+          gte_rs[i]&=~0xe3fll;
+          if(v==3) gte_rs[i]|=0xe00ll;
+          else gte_rs[i]|=3ll<<(v*2);
+        }
         break;
       case FLOAT:
       case FCONV:
@@ -9505,7 +9521,7 @@ int new_recompile_block(int addr)
             branch_regs[i-1].is32|=1LL<<31;
           }
           memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
-          memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
+          memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
           break;
         case RJUMP:
           memcpy(&branch_regs[i-1],&current,sizeof(current));
@@ -9530,7 +9546,7 @@ int new_recompile_block(int addr)
           }
           #endif
           memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
-          memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
+          memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
           break;
         case CJUMP:
           if((opcode[i-1]&0x3E)==4) // BEQ/BNE
@@ -9566,7 +9582,7 @@ int new_recompile_block(int addr)
             branch_regs[i-1].isconst=0;
             branch_regs[i-1].wasconst=0;
             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
-            memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
+            memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
           }
           else
           if((opcode[i-1]&0x3E)==6) // BLEZ/BGTZ
@@ -9599,7 +9615,7 @@ int new_recompile_block(int addr)
             branch_regs[i-1].isconst=0;
             branch_regs[i-1].wasconst=0;
             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
-            memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
+            memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
           }
           else
           // Alloc the delay slot in case the branch is taken
@@ -9665,7 +9681,7 @@ int new_recompile_block(int addr)
             branch_regs[i-1].isconst=0;
             branch_regs[i-1].wasconst=0;
             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
-            memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
+            memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
           }
           else
           // Alloc the delay slot in case the branch is taken
@@ -9779,7 +9795,7 @@ int new_recompile_block(int addr)
     {
       cc=0;
     }
-#ifdef PCSX
+#if defined(PCSX) && !defined(DRC_DBG)
     else if(itype[i]==C2OP&&gte_cycletab[source[i]&0x3f]>2)
     {
       // GTE runs in parallel until accessed, divide by 2 for a rough guess
@@ -9804,7 +9820,7 @@ int new_recompile_block(int addr)
       regs[i].is32=current.is32;
       regs[i].dirty=current.dirty;
       regs[i].isconst=current.isconst;
-      memcpy(constmap[i],current.constmap,sizeof(current.constmap));
+      memcpy(constmap[i],current_constmap,sizeof(current_constmap));
     }
     for(hr=0;hr<HOST_REGS;hr++) {
       if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {