drc: fix $ra evicting on JAL

[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 d837138..573d0cd 100644 (file)
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -690,6 +690,10 @@ void lsn(u_char hsn[], int i, int *preferred_reg)
     hsn[RHASH]=1;
     hsn[RHTBL]=1;
   }
+  // due to the way JAL is currently done we need DS not to evict $ra
+  if(i>0&&itype[i-1]==UJUMP&&rt1[i-1]==31) {
+    hsn[31]=0;
+  }
   // Coprocessor load/store needs FTEMP, even if not declared
   if(itype[i]==C1LS||itype[i]==C2LS) {
     hsn[FTEMP]=0;
@@ -2766,6 +2770,7 @@ void load_assemble(int i,struct regstat *i_regs)
   int offset;
   int jaddr=0;
   int memtarget=0,c=0;
+  int fastload_reg_override=0;
   u_int hr,reglist=0;
   th=get_reg(i_regs->regmap,rt1[i]|64);
   tl=get_reg(i_regs->regmap,rt1[i]);
@@ -2820,6 +2825,7 @@ void load_assemble(int i,struct regstat *i_regs)
         if(sp_in_mirror&&rs1[i]==29) {
           emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
           emit_cmpimm(HOST_TEMPREG,RAM_SIZE);
+          fastload_reg_override=HOST_TEMPREG;
         }
         else
         #endif
@@ -2864,9 +2870,8 @@ void load_assemble(int i,struct regstat *i_regs)
 #else
           if(!c) a=addr;
 #endif
-#ifdef PCSX
-          if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+          if(fastload_reg_override) a=fastload_reg_override;
+
           emit_movsbl_indexed_tlb(x,a,map,tl);
         }
       }
@@ -2892,9 +2897,7 @@ void load_assemble(int i,struct regstat *i_regs)
 #else
           if(!c) a=addr;
 #endif
-#ifdef PCSX
-          if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+          if(fastload_reg_override) a=fastload_reg_override;
           //#ifdef
           //emit_movswl_indexed_tlb(x,tl,map,tl);
           //else
@@ -2920,9 +2923,7 @@ void load_assemble(int i,struct regstat *i_regs)
     if(!c||memtarget) {
       if(!dummy) {
         int a=addr;
-#ifdef PCSX
-        if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+        if(fastload_reg_override) a=fastload_reg_override;
         //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
         #ifdef HOST_IMM_ADDR32
         if(c)
@@ -2956,9 +2957,8 @@ void load_assemble(int i,struct regstat *i_regs)
 #else
           if(!c) a=addr;
 #endif
-#ifdef PCSX
-          if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+          if(fastload_reg_override) a=fastload_reg_override;
+
           emit_movzbl_indexed_tlb(x,a,map,tl);
         }
       }
@@ -2984,9 +2984,7 @@ void load_assemble(int i,struct regstat *i_regs)
 #else
           if(!c) a=addr;
 #endif
-#ifdef PCSX
-          if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+          if(fastload_reg_override) a=fastload_reg_override;
           //#ifdef
           //emit_movzwl_indexed_tlb(x,tl,map,tl);
           //#else
@@ -3013,9 +3011,7 @@ void load_assemble(int i,struct regstat *i_regs)
     if(!c||memtarget) {
       if(!dummy) {
         int a=addr;
-#ifdef PCSX
-        if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+        if(fastload_reg_override) a=fastload_reg_override;
         //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
         #ifdef HOST_IMM_ADDR32
         if(c)
@@ -3036,9 +3032,7 @@ void load_assemble(int i,struct regstat *i_regs)
     if(!c||memtarget) {
       if(!dummy) {
         int a=addr;
-#ifdef PCSX
-        if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+        if(fastload_reg_override) a=fastload_reg_override;
         //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);
@@ -3103,6 +3097,7 @@ void store_assemble(int i,struct regstat *i_regs)
   int jaddr=0,jaddr2,type;
   int memtarget=0,c=0;
   int agr=AGEN1+(i&1);
+  int faststore_reg_override=0;
   u_int hr,reglist=0;
   th=get_reg(i_regs->regmap,rs2[i]|64);
   tl=get_reg(i_regs->regmap,rs2[i]);
@@ -3131,6 +3126,7 @@ void store_assemble(int i,struct regstat *i_regs)
       if(sp_in_mirror&&rs1[i]==29) {
         emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
         emit_cmpimm(HOST_TEMPREG,RAM_SIZE);
+        faststore_reg_override=HOST_TEMPREG;
       }
       else
       #endif
@@ -3177,9 +3173,7 @@ void store_assemble(int i,struct regstat *i_regs)
 #else
       if(!c) a=addr;
 #endif
-#ifdef PCSX
-      if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+      if(faststore_reg_override) a=faststore_reg_override;
       //gen_tlb_addr_w(temp,map);
       //emit_writebyte_indexed(tl,(int)rdram-0x80000000,temp);
       emit_writebyte_indexed_tlb(tl,x,a,map,a);
@@ -3195,9 +3189,7 @@ void store_assemble(int i,struct regstat *i_regs)
 #else
       if(!c) a=addr;
 #endif
-#ifdef PCSX
-      if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+      if(faststore_reg_override) a=faststore_reg_override;
       //#ifdef
       //emit_writehword_indexed_tlb(tl,x,temp,map,temp);
       //#else
@@ -3212,9 +3204,7 @@ void store_assemble(int i,struct regstat *i_regs)
   if (opcode[i]==0x2B) { // SW
     if(!c||memtarget) {
       int a=addr;
-#ifdef PCSX
-      if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
-#endif
+      if(faststore_reg_override) a=faststore_reg_override;
       //emit_writeword_indexed(tl,(int)rdram-0x80000000,addr);
       emit_writeword_indexed_tlb(tl,0,a,map,temp);
     }
@@ -3223,9 +3213,7 @@ void store_assemble(int i,struct regstat *i_regs)
   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(faststore_reg_override) a=faststore_reg_override;
       if(rs2[i]) {
         assert(th>=0);
         //emit_writeword_indexed(th,(int)rdram-0x80000000,addr);
@@ -8251,7 +8239,9 @@ int new_recompile_block(int addr)
       case 0x24: strcpy(insn[i],"LBU"); type=LOAD; break;
       case 0x25: strcpy(insn[i],"LHU"); type=LOAD; break;
       case 0x26: strcpy(insn[i],"LWR"); type=LOADLR; break;
+#ifndef FORCE32
       case 0x27: strcpy(insn[i],"LWU"); type=LOAD; break;
+#endif
       case 0x28: strcpy(insn[i],"SB"); type=STORE; break;
       case 0x29: strcpy(insn[i],"SH"); type=STORE; break;
       case 0x2A: strcpy(insn[i],"SWL"); type=STORELR; break;
@@ -8895,8 +8885,6 @@ 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);
@@ -8907,6 +8895,8 @@ 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);
@@ -10986,10 +10976,14 @@ int new_recompile_block(int addr)
 #ifdef PCSX
   if (start == 0x80030000) {
     // nasty hack for fastbios thing
+    // override block entry to this code
     instr_addr0_override=(u_int)out;
     emit_movimm(start,0);
-    emit_readword((int)&pcaddr,1);
+    // abuse io address var as a flag that we
+    // have already returned here once
+    emit_readword((int)&address,1);
     emit_writeword(0,(int)&pcaddr);
+    emit_writeword(0,(int)&address);
     emit_cmp(0,1);
     emit_jne((int)new_dyna_leave);
   }