minor timing tweaks

src/ARM.cpp

@@ -647,7 +647,7 @@ void ARMv5::Execute()
             R[15] += 2;
             CurInstr = NextInstr[0];
             NextInstr[0] = NextInstr[1];
-            if (R[15] & 0x2) { NextInstr[1] >>= 16; CodeCycles = 0; }
+            if (R[15] & 0x2) { NextInstr[1] >>= 16; CodeCycles = 1; }
             else             NextInstr[1] = CodeRead32(R[15], false);
 
             // actually execute

src/ARM.h

@@ -317,15 +317,16 @@ public:
     void AddCycles_C() override
     {
         // code only. always nonseq 32-bit for ARM9.
-        s32 numC = (R[15] & 0x2) ? 0 : CodeCycles;
+        s32 numC = CodeCycles;
         Cycles += numC;
     }
 
     void AddCycles_CI(s32 numI) override
     {
         // code+internal
-        s32 numC = (R[15] & 0x2) ? 0 : CodeCycles;
-        Cycles += numC + numI;
+        s32 numC = CodeCycles;
+        numI += 1;
+        Cycles += std::max(numC, numI);
     }
 
     void AddCycles_CDI_LDR() override;