load register only if needed

- do thumb bl long merge in the first step
- preparations for better branch jitting

src/ARMJIT.cpp

@@ -159,6 +159,7 @@ CompiledBlock CompileBlock(ARM* cpu)
     u32 r15 = cpu->R[15];
 	cpu->FillPipeline();
     u32 nextInstr[2] = {cpu->NextInstr[0], cpu->NextInstr[1]};
+	u32 nextInstrAddr[2] = {blockAddr, r15};
     do
     {
         r15 += thumb ? 2 : 4;
@@ -167,6 +168,10 @@ CompiledBlock CompileBlock(ARM* cpu)
         instrs[i].Instr = nextInstr[0];
         instrs[i].NextInstr[0] = nextInstr[0] = nextInstr[1];
 	
+		instrs[i].Addr = nextInstrAddr[0];
+		nextInstrAddr[0] = nextInstrAddr[1];
+		nextInstrAddr[1] = r15;
+
         if (cpu->Num == 0)
         {
             ARMv5* cpuv5 = (ARMv5*)cpu;
@@ -193,8 +198,19 @@ CompiledBlock CompileBlock(ARM* cpu)
         instrs[i].NextInstr[1] = nextInstr[1];
         instrs[i].Info = ARMInstrInfo::Decode(thumb, cpu->Num, instrs[i].Instr);
 
+		if (thumb && instrs[i].Info.Kind == ARMInstrInfo::tk_BL_LONG_2 && i > 0
+			&& instrs[i - 1].Info.Kind == ARMInstrInfo::tk_BL_LONG_1)
+		{
+			instrs[i - 1].Info.Kind = ARMInstrInfo::tk_BL_LONG;
+			instrs[i - 1].Instr = (instrs[i - 1].Instr & 0xFFFF) | (instrs[i].Instr << 16);
+			instrs[i - 1].Info.DstRegs = 0xC000;
+			instrs[i - 1].Info.SrcRegs = 0;
+			instrs[i - 1].Info.EndBlock = true;
+			i--;
+		}
         i++;
 
+
 		bool canCompile = compiler->CanCompile(thumb, instrs[i - 1].Info.Kind);
 		if (instrs[i - 1].Info.ReadFlags != 0 || !canCompile)
 			floodFillSetFlags(instrs, i - 2, canCompile ? instrs[i - 1].Info.ReadFlags : 0xF);

src/ARMJIT.h

@@ -31,6 +31,7 @@ struct FetchedInstr
 	u8 SetFlags;
     u32 Instr;
     u32 NextInstr[2];
+	u32 Addr;
 
     u8 CodeCycles;
 

src/ARMJIT_RegisterCache.h

@@ -38,7 +38,7 @@ public:
         Mapping[reg] = (Reg)-1;
     }
 
-    void LoadRegister(int reg)
+    void LoadRegister(int reg, bool loadValue)
     {
         assert(Mapping[reg] == -1);
         for (int i = 0; i < NativeRegsAvailable; i++)
@@ -50,6 +50,7 @@ public:
                 NativeRegsUsed |= 1 << (int)nativeReg;
                 LoadedRegs |= 1 << reg;
 
+                if (loadValue)
                     Compiler->LoadReg(reg, nativeReg);
 
                 return;
@@ -66,7 +67,7 @@ public:
             UnloadRegister(reg);
     }
 
-	void Prepare(int i)
+	void Prepare(bool thumb, int i)
     {
         u16 futureNeeded = 0;
         int ranking[16];
@@ -111,8 +112,11 @@ public:
                 loadedSet.m_val = LoadedRegs;
             }
 
+            BitSet16 needValueLoaded(needToBeLoaded);
+            if (thumb || Instr.Cond() >= 0xE)
+                needValueLoaded = BitSet16(Instr.Info.SrcRegs);
             for (int reg : needToBeLoaded)
-                LoadRegister(reg);
+                LoadRegister(reg, needValueLoaded[reg]);
         }
         DirtyRegs |= Instr.Info.DstRegs & ~(1 << 15);
     }

src/ARMJIT_x64/ARMJIT_Branch.cpp

@@ -271,15 +271,17 @@ void Compiler::T_Comp_BL_LONG_2()
     Comp_JumpTo(RSCRATCH);
 }
 
-void Compiler::T_Comp_BL_Merged(FetchedInstr part1)
+void Compiler::T_Comp_BL_Merged()
 {
-    assert(part1.Info.Kind == ARMInstrInfo::tk_BL_LONG_1);
     Comp_AddCycles_C();
 
-    u32 target = (R15 - 2) + ((s32)((part1.Instr & 0x7FF) << 21) >> 9);
+    R15 += 2;
-    target += (CurInstr.Instr & 0x7FF) << 1;
 
-    if (Num == 1 || CurInstr.Instr & (1 << 12))
+    u32 upperPart = CurInstr.Instr >> 16;
+    u32 target = (R15 - 2) + ((s32)((CurInstr.Instr & 0x7FF) << 21) >> 9);
+    target += (upperPart & 0x7FF) << 1;
+
+    if (Num == 1 || upperPart & (1 << 12))
         target |= 1;
 
     MOV(32, MapReg(14), Imm32((R15 - 2) | 1));

src/ARMJIT_x64/ARMJIT_Compiler.cpp

@@ -338,7 +338,8 @@ const Compiler::CompileFunc T_Comp[ARMInstrInfo::tk_Count] = {
     // Branch
     F(T_Comp_BCOND), F(T_Comp_BranchXchangeReg), F(T_Comp_BranchXchangeReg), F(T_Comp_B), F(T_Comp_BL_LONG_1), F(T_Comp_BL_LONG_2), 
     // Unk, SVC
-    NULL, NULL
+    NULL, NULL,
+    F(T_Comp_BL_Merged)
 };
 #undef F
 
@@ -361,21 +362,18 @@ CompiledBlock Compiler::CompileBlock(ARM* cpu, FetchedInstr instrs[], int instrs
     ConstantCycles = 0;
     Thumb = cpu->CPSR & 0x20;
     Num = cpu->Num;
-    R15 = cpu->R[15];
     CodeRegion = cpu->CodeRegion;
     CurCPU = cpu;
 
     CompiledBlock res = (CompiledBlock)GetWritableCodePtr();
 
     if (!(Num == 0 
-        ? IsMapped<0>(R15 - (Thumb ? 2 : 4)) 
+        ? IsMapped<0>(instrs[0].Addr - (Thumb ? 2 : 4)) 
-        : IsMapped<1>(R15 - (Thumb ? 2 : 4))))
+        : IsMapped<1>(instrs[0].Addr - (Thumb ? 2 : 4))))
     {
         printf("Trying to compile a block in unmapped memory\n");
     }
 
-    bool mergedThumbBL = false;
-
     ABI_PushRegistersAndAdjustStack(BitSet32(ABI_ALL_CALLEE_SAVED & ABI_ALL_GPRS & ~BitSet32({RSP})), 8);
 
     MOV(64, R(RCPU), ImmPtr(cpu));
@@ -387,8 +385,8 @@ CompiledBlock Compiler::CompileBlock(ARM* cpu, FetchedInstr instrs[], int instrs
 
     for (int i = 0; i < instrsCount; i++)
     {
-        R15 += Thumb ? 2 : 4;
         CurInstr = instrs[i];
+        R15 = CurInstr.Addr + (Thumb ? 4 : 8);
 
         CompileFunc comp = Thumb
             ? T_Comp[CurInstr.Info.Kind]
@@ -406,16 +404,11 @@ CompiledBlock Compiler::CompileBlock(ARM* cpu, FetchedInstr instrs[], int instrs
         }
         
         if (comp != NULL)
-            RegCache.Prepare(i);
+            RegCache.Prepare(Thumb, i);
         else
             RegCache.Flush();
 
         if (Thumb)
-        {
-            if (i < instrsCount - 1 && CurInstr.Info.Kind == ARMInstrInfo::tk_BL_LONG_1
-                && instrs[i + 1].Info.Kind == ARMInstrInfo::tk_BL_LONG_2)
-                mergedThumbBL = true;
-            else
         {
             u32 icode = (CurInstr.Instr >> 6) & 0x3FF;
             if (comp == NULL)
@@ -424,12 +417,9 @@ CompiledBlock Compiler::CompileBlock(ARM* cpu, FetchedInstr instrs[], int instrs
 
                 ABI_CallFunction(ARMInterpreter::THUMBInstrTable[icode]);
             }
-                else if (mergedThumbBL)
-                    T_Comp_BL_Merged(instrs[i - 1]);
             else
                 (this->*comp)();
         }
-        }
         else
         {
             u32 cond = CurInstr.Cond();

src/ARMJIT_x64/ARMJIT_Compiler.h

@@ -90,7 +90,7 @@ public:
     void T_Comp_BranchXchangeReg();
     void T_Comp_BL_LONG_1();
     void T_Comp_BL_LONG_2();
-    void T_Comp_BL_Merged(FetchedInstr prefix);
+    void T_Comp_BL_Merged();
 
     void Comp_MemAccess(Gen::OpArg rd, bool signExtend, bool store, int size);
     s32 Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc, bool decrement, bool usermode);

src/ARM_InstrInfo.h

@@ -212,6 +212,9 @@ enum
     tk_UNK,
     tk_SVC,
 
+    // not a real instruction
+    tk_BL_LONG,
+
     tk_Count
 };