pcnt: measure gte too

[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 4cb43c8..e2c63ed 100644 (file)
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -5403,6 +5403,12 @@ void rjump_assemble(int i,struct regstat *i_regs)
   //assert(adj==0);
   emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
   add_stub(CC_STUB,(int)out,jump_vaddr_reg[rs],0,i,-1,TAKEN,0);
+#ifdef PCSX
+  if(itype[i+1]==COP0&&(source[i+1]&0x3f)==0x10)
+    // special case for RFE
+    emit_jmp(0);
+  else
+#endif
   emit_jns(0);
   //load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,-1);
   #ifdef USE_MINI_HT
@@ -8645,8 +8651,9 @@ int new_recompile_block(int addr)
         rs2[i]=0;
         rt1[i]=0;
         rt2[i]=0;
-        gte_rt[i]=1ll<<63; // every op changes flags
-        // TODO: other regs?
+        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
         break;
       case FLOAT:
       case FCONV:
@@ -9729,6 +9736,11 @@ int new_recompile_block(int addr)
       cc=0;
     }
 #ifdef PCSX
+    else if(itype[i]==C2OP&&gte_cycletab[source[i]&0x3f]>2)
+    {
+      // GTE runs in parallel until accessed, divide by 2 for a rough guess
+      cc+=gte_cycletab[source[i]&0x3f]/2;
+    }
     else if(/*itype[i]==LOAD||*/itype[i]==STORE||itype[i]==C1LS) // load causes weird timing issues
     {
       cc+=2; // 2 cycle penalty (after CLOCK_DIVIDER)