jit: make everything configurable

12 files changed, 192 insertions, 46 deletions

diff --git a/src/ARM.cpp b/src/ARM.cpp
index a77fbc4..6cc80c0 100644
--- a/src/ARM.cpp
+++ b/src/ARM.cpp
@@ -489,7 +489,7 @@ void ARMv5::Execute()
 
     while (NDS::ARM9Timestamp < NDS::ARM9Target)
     {
-        /*if (CPSR & 0x20) // THUMB
+        if (CPSR & 0x20) // THUMB
         {
             // prefetch
             R[15] += 2;
@@ -522,14 +522,8 @@ void ARMv5::Execute()
             }
             else
                 AddCycles_C();
-        }*/
-
-        /*if (!ARMJIT::IsMapped(0, R[15] - ((CPSR&0x20)?2:4)))
-            printf("aaarg ungempappter raum %x\n", R[15]);*/
-
-        ARMJIT::CompiledBlock block = ARMJIT::LookUpBlock(0, R[15] - ((CPSR&0x20)?2:4));
-        Cycles += (block ? block : ARMJIT::CompileBlock(this))();
-        
+        }
+ 
         // TODO optimize this shit!!!
         if (Halted)
         {
@@ -554,6 +548,58 @@ void ARMv5::Execute()
         Halted = 0;
 }
 
+void ARMv5::ExecuteJIT()
+{
+    if (Halted)
+    {
+        if (Halted == 2)
+        {
+            Halted = 0;
+        }
+        else if (NDS::HaltInterrupted(0))
+        {
+            Halted = 0;
+            if (NDS::IME[0] & 0x1)
+                TriggerIRQ();
+        }
+        else
+        {
+            NDS::ARM9Timestamp = NDS::ARM9Target;
+            return;
+        }
+    }
+
+    while (NDS::ARM9Timestamp < NDS::ARM9Target)
+    {
+        u32 instrAddr = R[15] - ((CPSR&0x20)?2:4);
+        if (!ARMJIT::IsMapped(0, instrAddr))
+        {
+            NDS::ARM9Timestamp = NDS::ARM9Target;
+            printf("ARMv5 PC in non executable region %08X\n", R[15]);
+            return;
+        }
+
+        ARMJIT::CompiledBlock block = ARMJIT::LookUpBlock(0, instrAddr);
+        Cycles += (block ? block : ARMJIT::CompileBlock(this))();
+
+        if (Halted)
+        {
+            if (Halted == 1 && NDS::ARM9Timestamp < NDS::ARM9Target)
+            {
+                NDS::ARM9Timestamp = NDS::ARM9Target;
+            }
+            break;
+        }
+        if (IRQ) TriggerIRQ();
+
+        NDS::ARM9Timestamp += Cycles;
+        Cycles = 0;
+    }
+
+    if (Halted == 2)
+        Halted = 0;
+}
+
 void ARMv4::Execute()
 {
     if (Halted)
@@ -577,7 +623,7 @@ void ARMv4::Execute()
 
     while (NDS::ARM7Timestamp < NDS::ARM7Target)
     {
-        /*if (CPSR & 0x20) // THUMB
+        if (CPSR & 0x20) // THUMB
         {
             // prefetch
             R[15] += 2;
@@ -605,13 +651,7 @@ void ARMv4::Execute()
             }
             else
                 AddCycles_C();
-        }*/
-
-        /*if (!ARMJIT::IsMapped(1, R[15] - ((CPSR&0x20)?2:4)))
-            printf("aaarg ungempappter raum %x\n", R[15]);*/
-
-        ARMJIT::CompiledBlock block = ARMJIT::LookUpBlock(1, R[15] - ((CPSR&0x20)?2:4));
-        Cycles += (block ? block : ARMJIT::CompileBlock(this))();
+        }
 
         // TODO optimize this shit!!!
         if (Halted)
@@ -636,3 +676,56 @@ void ARMv4::Execute()
     if (Halted == 2)
         Halted = 0;
 }
+
+void ARMv4::ExecuteJIT()
+{
+    if (Halted)
+    {
+        if (Halted == 2)
+        {
+            Halted = 0;
+        }
+        else if (NDS::HaltInterrupted(1))
+        {
+            Halted = 0;
+            if (NDS::IME[1] & 0x1)
+                TriggerIRQ();
+        }
+        else
+        {
+            NDS::ARM7Timestamp = NDS::ARM7Target;
+            return;
+        }
+    }
+
+    while (NDS::ARM7Timestamp < NDS::ARM7Target)
+    {
+        u32 instrAddr = R[15] - ((CPSR&0x20)?2:4);
+        if (!ARMJIT::IsMapped(1, instrAddr))
+        {
+            NDS::ARM7Timestamp = NDS::ARM7Target;
+            printf("ARMv4 PC in non executable region %08X\n", R[15]);
+            return;
+        }
+        ARMJIT::CompiledBlock block = ARMJIT::LookUpBlock(1, instrAddr);
+        Cycles += (block ? block : ARMJIT::CompileBlock(this))();
+
+        // TODO optimize this shit!!!
+        if (Halted)
+        {
+            if (Halted == 1 && NDS::ARM7Timestamp < NDS::ARM7Target)
+            {
+                NDS::ARM7Timestamp = NDS::ARM7Target;
+            }
+            break;
+        }
+
+        if (IRQ) TriggerIRQ();
+
+        NDS::ARM7Timestamp += Cycles;
+        Cycles = 0;
+    }
+
+    if (Halted == 2)
+        Halted = 0;
+}
\ No newline at end of file
diff --git a/src/ARM.h b/src/ARM.h
index b9f5d89..0544301 100644
--- a/src/ARM.h
+++ b/src/ARM.h
@@ -52,6 +52,7 @@ public:
     }
 
     virtual void Execute() = 0;
+    virtual void ExecuteJIT() = 0;
 
     bool CheckCondition(u32 code)
     {
@@ -151,6 +152,7 @@ public:
     void DataAbort();
 
     void Execute();
+    void ExecuteJIT();
 
     // all code accesses are forced nonseq 32bit
     u32 CodeRead32(u32 addr, bool branch);
@@ -269,6 +271,7 @@ public:
     void JumpTo(u32 addr, bool restorecpsr = false);
 
     void Execute();
+    void ExecuteJIT();
 
     u16 CodeRead16(u32 addr)
     {
diff --git a/src/ARMJIT.cpp b/src/ARMJIT.cpp
index 47b425f..e8e6be0 100644
--- a/src/ARMJIT.cpp
+++ b/src/ARMJIT.cpp
@@ -2,6 +2,8 @@
 
 #include <string.h>
 
+#include "Config.h"
+
 #include "ARMJIT_x64/ARMJIT_Compiler.h"
 
 namespace ARMJIT
@@ -125,18 +127,21 @@ CompiledBlock CompileBlock(ARM* cpu)
 {
     bool thumb = cpu->CPSR & 0x20;
 
-    FetchedInstr instrs[12];
+	if (Config::JIT_MaxBlockSize < 1)
+		Config::JIT_MaxBlockSize = 1;
+	if (Config::JIT_MaxBlockSize > 32)
+		Config::JIT_MaxBlockSize = 32;
+
+    FetchedInstr instrs[Config::JIT_MaxBlockSize];
     int i = 0;
-	u32 r15Initial = cpu->R[15];
+	u32 blockAddr = cpu->R[15] - (thumb ? 2 : 4);
     u32 r15 = cpu->R[15];
     u32 nextInstr[2] = {cpu->NextInstr[0], cpu->NextInstr[1]};
-    //printf("block %x %d\n", r15, thumb);
     do
     {
         r15 += thumb ? 2 : 4;
 
         instrs[i].Instr = nextInstr[0];
-        //printf("%x %x\n", instrs[i].Instr, r15);
         instrs[i].NextInstr[0] = nextInstr[0] = nextInstr[1];
 
         if (cpu->Num == 0)
@@ -166,16 +171,16 @@ CompiledBlock CompileBlock(ARM* cpu)
         instrs[i].Info = ARMInstrInfo::Decode(thumb, cpu->Num, instrs[i].Instr);
 
         i++;
-    } while(!instrs[i - 1].Info.Branches() && i < 10);
+    } while(!instrs[i - 1].Info.Branches() && i < Config::JIT_MaxBlockSize);
 
     CompiledBlock block = compiler->CompileBlock(cpu, instrs, i);
 
-    InsertBlock(cpu->Num, r15Initial - (thumb ? 2 : 4), block);
+    InsertBlock(cpu->Num, blockAddr, block);
 
 	return block;
 }
 
-void ResetBlocks()
+void InvalidateBlockCache()
 {
 	memset(cache.MainRAM, 0, sizeof(cache.MainRAM));
 	memset(cache.SWRAM, 0, sizeof(cache.SWRAM));
@@ -185,6 +190,8 @@ void ResetBlocks()
 	memset(cache.ARM7_BIOS, 0, sizeof(cache.ARM7_BIOS));
 	memset(cache.ARM7_WRAM, 0, sizeof(cache.ARM7_WRAM));
 	memset(cache.ARM7_WVRAM, 0, sizeof(cache.ARM7_WVRAM));
+
+	compiler->Reset();
 }
 
 }
\ No newline at end of file
diff --git a/src/ARMJIT.h b/src/ARMJIT.h
index 45bb4ed..004256c 100644
--- a/src/ARMJIT.h
+++ b/src/ARMJIT.h
@@ -111,7 +111,7 @@ void DeInit();
 
 CompiledBlock CompileBlock(ARM* cpu);
 
-void ResetBlocks();
+void InvalidateBlockCache();
 
 }
 
diff --git a/src/ARMJIT_x64/ARMJIT_Compiler.cpp b/src/ARMJIT_x64/ARMJIT_Compiler.cpp
index 2b7ccd2..fe23859 100644
--- a/src/ARMJIT_x64/ARMJIT_Compiler.cpp
+++ b/src/ARMJIT_x64/ARMJIT_Compiler.cpp
@@ -336,13 +336,15 @@ const Compiler::CompileFunc T_Comp[ARMInstrInfo::tk_Count] = {
 };
 #undef F
 
+void Compiler::Reset()
+{
+    SetCodePtr((u8*)ResetStart);
+}
+
 CompiledBlock Compiler::CompileBlock(ARM* cpu, FetchedInstr instrs[], int instrsCount)
 {
     if (IsAlmostFull())
-    {
-        ResetBlocks();
-        SetCodePtr((u8*)ResetStart);
-    }
+        InvalidateBlockCache();
 
     CompiledBlock res = (CompiledBlock)GetWritableCodePtr();
 
@@ -355,7 +357,7 @@ CompiledBlock Compiler::CompileBlock(ARM* cpu, FetchedInstr instrs[], int instrs
 
     bool mergedThumbBL = false;
 
-    ABI_PushRegistersAndAdjustStack(BitSet32(ABI_ALL_CALLEE_SAVED & ABI_ALL_GPRS & ~RSP), 8);
+    ABI_PushRegistersAndAdjustStack(BitSet32(ABI_ALL_CALLEE_SAVED & ABI_ALL_GPRS & ~BitSet32({RSP})), 8);
 
     MOV(64, R(RCPU), ImmPtr(cpu));
 
@@ -469,7 +471,7 @@ CompiledBlock Compiler::CompileBlock(ARM* cpu, FetchedInstr instrs[], int instrs
 
     MOV(32, R(RAX), Imm32(ConstantCycles));
 
-    ABI_PopRegistersAndAdjustStack(BitSet32(ABI_ALL_CALLEE_SAVED & ABI_ALL_GPRS & ~RSP), 8);
+    ABI_PopRegistersAndAdjustStack(BitSet32(ABI_ALL_CALLEE_SAVED & ABI_ALL_GPRS & ~BitSet32({RSP})), 8);
     RET();
 
     return res;
diff --git a/src/ARMJIT_x64/ARMJIT_Compiler.h b/src/ARMJIT_x64/ARMJIT_Compiler.h
index e04f96a..cd58012 100644
--- a/src/ARMJIT_x64/ARMJIT_Compiler.h
+++ b/src/ARMJIT_x64/ARMJIT_Compiler.h
@@ -22,6 +22,8 @@ class Compiler : public Gen::X64CodeBlock
 public:
     Compiler();
 
+    void Reset();
+
     CompiledBlock CompileBlock(ARM* cpu, FetchedInstr instrs[], int instrsCount);
 
     void LoadReg(int reg, Gen::X64Reg nativeReg);
diff --git a/src/Config.cpp b/src/Config.cpp
index f558ef6..37b701c 100644
--- a/src/Config.cpp
+++ b/src/Config.cpp
@@ -34,6 +34,9 @@ int Threaded3D;
 int GL_ScaleFactor;
 int GL_Antialias;
 
+bool JIT_Enable = false;
+int JIT_MaxBlockSize = 12;
+
 ConfigEntry ConfigFile[] =
 {
     {"3DRenderer", 0, &_3DRenderer, 1, NULL, 0},
@@ -42,6 +45,9 @@ ConfigEntry ConfigFile[] =
     {"GL_ScaleFactor", 0, &GL_ScaleFactor, 1, NULL, 0},
     {"GL_Antialias", 0, &GL_Antialias, 0, NULL, 0},
 
+    {"JIT_Enable", 0, &JIT_Enable, 0, NULL, 0},
+    {"JIT_MaxBlockSize", 0, &JIT_MaxBlockSize, 10, NULL, 0},
+
     {"", -1, NULL, 0, NULL, 0}
 };
 
diff --git a/src/Config.h b/src/Config.h
index 84fd57b..18a7910 100644
--- a/src/Config.h
+++ b/src/Config.h
@@ -46,6 +46,9 @@ extern int Threaded3D;
 extern int GL_ScaleFactor;
 extern int GL_Antialias;
 
+extern bool JIT_Enable;
+extern int JIT_MaxBlockSize;
+
 }
 
 #endif // CONFIG_H
diff --git a/src/NDS.cpp b/src/NDS.cpp
index baa5e0d..4b50d9c 100644
--- a/src/NDS.cpp
+++ b/src/NDS.cpp
@@ -524,7 +524,7 @@ void Reset()
     KeyCnt = 0;
     RCnt = 0;
 
-    ARMJIT::ResetBlocks();
+    ARMJIT::InvalidateBlockCache();
 
     NDSCart::Reset();
     GBACart::Reset();
@@ -741,6 +741,11 @@ bool DoSavestate(Savestate* file)
         GPU::SetPowerCnt(PowerControl9);
     }
 
+    if (!file->Saving)
+    {
+        ARMJIT::InvalidateBlockCache();
+    }
+
     return true;
 }
 
@@ -826,6 +831,7 @@ void RunSystem(u64 timestamp)
     }
 }
 
+template <bool EnableJIT>
 u32 RunFrame()
 {
     FrameStartTimestamp = SysTimestamp;
@@ -858,7 +864,10 @@ u32 RunFrame()
         }
         else
         {
-            ARM9->Execute();
+            if (EnableJIT)
+                ARM9->ExecuteJIT();
+            else
+                ARM9->Execute();
         }
 
         RunTimers(0);
@@ -880,7 +889,10 @@ u32 RunFrame()
             }
             else
             {
-                ARM7->Execute();
+                if (EnableJIT)
+                    ARM7->ExecuteJIT();
+                else
+                    ARM7->Execute();
             }
 
             RunTimers(1);
@@ -910,6 +922,14 @@ u32 RunFrame()
     return GPU::TotalScanlines;
 }
 
+u32 RunFrame()
+{
+    if (Config::JIT_Enable)
+        return RunFrame<true>();
+    else
+        return RunFrame<false>();
+}
+
 void Reschedule(u64 target)
 {
     if (CurCPU == 0)
diff --git a/src/libui_sdl/DlgEmuSettings.cpp b/src/libui_sdl/DlgEmuSettings.cpp
index 0ccaed7..116d2da 100644
--- a/src/libui_sdl/DlgEmuSettings.cpp
+++ b/src/libui_sdl/DlgEmuSettings.cpp
@@ -57,10 +57,20 @@ void OnOk(uiButton* btn, void* blarg)
 {
     Config::DirectBoot = uiCheckboxChecked(cbDirectBoot);
 
+    Config::JIT_Enable = uiCheckboxChecked(cbJITEnabled);
+    long blockSize = strtol(uiEntryText(enJITMaxBlockSize), NULL, 10);
+    if (blockSize < 1)
+        blockSize = 1;
+    if (blockSize > 32)
+        blockSize = 32;
+    Config::JIT_MaxBlockSize = blockSize;
+
     Config::Save();
 
     uiControlDestroy(uiControl(win));
     opened = false;
+
+    ApplyNewSettings(4);
 }
 
 void OnJITStateChanged(uiCheckbox* cb, void* blarg)
@@ -143,6 +153,12 @@ void Open()
 
     uiCheckboxSetChecked(cbDirectBoot, Config::DirectBoot);
 
+    uiCheckboxSetChecked(cbJITEnabled, Config::JIT_Enable);
+    {
+        char maxBlockSizeStr[10];
+        sprintf(maxBlockSizeStr, "%d", Config::JIT_MaxBlockSize);
+        uiEntrySetText(enJITMaxBlockSize, maxBlockSizeStr);
+    }
     OnJITStateChanged(cbJITEnabled, NULL);
 
     uiControlShow(uiControl(win));
diff --git a/src/libui_sdl/PlatformConfig.cpp b/src/libui_sdl/PlatformConfig.cpp
index f78b195..b6d1e8d 100644
--- a/src/libui_sdl/PlatformConfig.cpp
+++ b/src/libui_sdl/PlatformConfig.cpp
@@ -64,6 +64,7 @@ char MicWavPath[512];
 
 char LastROMFolder[512];
 
+bool EnableJIT;
 
 ConfigEntry PlatformConfigFile[] =
 {
diff --git a/src/libui_sdl/main.cpp b/src/libui_sdl/main.cpp
index 8e8bf9e..d6809c3 100644
--- a/src/libui_sdl/main.cpp
+++ b/src/libui_sdl/main.cpp
@@ -48,6 +48,7 @@
 #include "../Wifi.h"
 #include "../Platform.h"
 #include "../Config.h"
+#include "../ARMJIT.h"
 
 #include "../Savestate.h"
 
@@ -2408,19 +2409,11 @@ void ApplyNewSettings(int type)
         GPU3D::InitRenderer(Screen_UseGL);
         if (Screen_UseGL) uiGLMakeContextCurrent(NULL);
     }
-    /*else if (type == 4) // vsync
+    else if (type == 4)
     {
-        if (Screen_UseGL)
-        {
-            uiGLMakeContextCurrent(GLContext);
-            uiGLSetVSync(Config::ScreenVSync);
-            uiGLMakeContextCurrent(NULL);
-        }
-        else
-        {
-            // TODO eventually: VSync for non-GL screen?
-        }
-    }*/
+        if (Config::JIT_Enable)
+            ARMJIT::InvalidateBlockCache();
+    }
 
     EmuRunning = prevstatus;
 }