drc: implement ra accesses in ujump DS

[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 edd1413..34673b9 100644 (file)
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -1186,6 +1186,9 @@ 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]) {
@@ -5108,34 +5111,19 @@ 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);
@@ -5156,6 +5144,14 @@ 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);
@@ -8239,18 +8235,6 @@ 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 */
@@ -8483,19 +8467,43 @@ 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;
       }
@@ -8821,8 +8829,8 @@ int new_recompile_block(int addr)
           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
@@ -10755,9 +10763,9 @@ 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);
@@ -11032,6 +11040,12 @@ 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 */