Decide whether SuperFX and SA-1 chips are enabled only once per frame. This saves a few million instructions per second. (Inspired by Snes9x-Euphoria)

3 files changed, 459 insertions, 7 deletions

diff --git a/source/cpuexec.cpp b/source/cpuexec.cpp
index c4836ec..a0c7ebc 100644
--- a/source/cpuexec.cpp
+++ b/source/cpuexec.cpp
@@ -111,8 +111,116 @@
 
 extern void S9xProcessSound (unsigned int);
 
+void S9xMainLoop_SA1_SFX     (void);
+void S9xMainLoop_SA1_NoSFX   (void);
+void S9xMainLoop_NoSA1_SFX   (void);
+void S9xMainLoop_NoSA1_NoSFX (void);
+
+/*
+ * This is a CATSFC modification inspired by a Snes9x-Euphoria modification.
+ * The emulator selects a main loop based on the chips used in an entire
+ * frame. This avoids the constant SA1.Executing and Settings.SuperFX checks.
+ *
+ * The original version of S9xMainLoop is S9xMainLoop_SA1_SFX below. Remember
+ * to propagate modifications to the SA1_NoSFX, NoSA1_SFX and NoSA1_NoSFX
+ * versions.
+ */
 void S9xMainLoop (void)
 {
+	if (Settings.SA1)
+	{
+		if (Settings.SuperFX)             S9xMainLoop_SA1_SFX   ();
+		else /* if (!Settings.SuperFX) */ S9xMainLoop_SA1_NoSFX ();
+	}
+	else /* if (!Settings.SA1) */
+	{
+		if (Settings.SuperFX)             S9xMainLoop_NoSA1_SFX   ();
+		else /* if (!Settings.SuperFX) */ S9xMainLoop_NoSA1_NoSFX ();
+	}
+}
+
+void S9xMainLoop_SA1_SFX (void)
+{
+    for (;;)
+	{
+    	APU_EXECUTE ();
+    
+    	if (CPU.Flags)
+    	{
+    	    if (CPU.Flags & NMI_FLAG)
+    	    {
+        		if (--CPU.NMICycleCount == 0) {
+        		    CPU.Flags &= ~NMI_FLAG;
+        		    if (CPU.WaitingForInterrupt) {
+            			CPU.WaitingForInterrupt = FALSE;
+            			CPU.PC++;
+        		    }
+        		    S9xOpcode_NMI ();
+        		}
+    	    }
+
+    	    CHECK_SOUND ();
+
+    	    if (CPU.Flags & IRQ_PENDING_FLAG)
+    	    {
+        		if (CPU.IRQCycleCount == 0)
+        		{
+        		    if (CPU.WaitingForInterrupt) {
+            			CPU.WaitingForInterrupt = FALSE;
+            			CPU.PC++;
+        		    }
+        		    if (CPU.IRQActive && !Settings.DisableIRQ) {
+            			if (!CheckFlag (IRQ))
+            			    S9xOpcode_IRQ ();
+        		    }
+        		    else
+            			CPU.Flags &= ~IRQ_PENDING_FLAG;
+        		}
+        		else
+                {
+                    if(--CPU.IRQCycleCount==0 && CheckFlag (IRQ))
+                        CPU.IRQCycleCount=1;
+                }
+    	    }
+
+    	    if (CPU.Flags & SCAN_KEYS_FLAG)
+	        	break;
+    	}
+
+#ifdef CPU_SHUTDOWN
+    	CPU.PCAtOpcodeStart = CPU.PC;
+#endif
+    	CPU.Cycles += CPU.MemSpeed;
+
+    	(*ICPU.S9xOpcodes [*CPU.PC++].S9xOpcode) ();
+	
+    	if (SA1.Executing)
+    	    S9xSA1MainLoop ();
+    	DO_HBLANK_CHECK_SFX();
+    }
+
+    ICPU.Registers.PC = CPU.PC - CPU.PCBase;
+    S9xPackStatus ();
+    IAPU.Registers.PC = IAPU.PC - IAPU.RAM;
+    S9xAPUPackStatus ();
+    if (CPU.Flags & SCAN_KEYS_FLAG)
+    {
+	    S9xSyncSpeed ();
+        CPU.Flags &= ~SCAN_KEYS_FLAG;
+    }
+
+#ifdef DETECT_NASTY_FX_INTERLEAVE
+    if (CPU.BRKTriggered && Settings.SuperFX && !CPU.TriedInterleavedMode2)
+    {
+    	CPU.TriedInterleavedMode2 = TRUE;
+    	CPU.BRKTriggered = FALSE;
+    	S9xDeinterleaveMode2 ();
+    }
+#endif
+}
+
+void S9xMainLoop_SA1_NoSFX (void)
+{
     for (;;)
 	{
     	APU_EXECUTE ();
@@ -168,7 +276,76 @@ void S9xMainLoop (void)
 	
     	if (SA1.Executing)
     	    S9xSA1MainLoop ();
-    	DO_HBLANK_CHECK();
+    	DO_HBLANK_CHECK_NoSFX();
+    }
+
+    ICPU.Registers.PC = CPU.PC - CPU.PCBase;
+    S9xPackStatus ();
+    IAPU.Registers.PC = IAPU.PC - IAPU.RAM;
+    S9xAPUPackStatus ();
+    if (CPU.Flags & SCAN_KEYS_FLAG)
+    {
+	    S9xSyncSpeed ();
+        CPU.Flags &= ~SCAN_KEYS_FLAG;
+    }
+}
+
+void S9xMainLoop_NoSA1_SFX (void)
+{
+    for (;;)
+	{
+    	APU_EXECUTE ();
+    
+    	if (CPU.Flags)
+    	{
+    	    if (CPU.Flags & NMI_FLAG)
+    	    {
+        		if (--CPU.NMICycleCount == 0) {
+        		    CPU.Flags &= ~NMI_FLAG;
+        		    if (CPU.WaitingForInterrupt) {
+            			CPU.WaitingForInterrupt = FALSE;
+            			CPU.PC++;
+        		    }
+        		    S9xOpcode_NMI ();
+        		}
+    	    }
+
+    	    CHECK_SOUND ();
+
+    	    if (CPU.Flags & IRQ_PENDING_FLAG)
+    	    {
+        		if (CPU.IRQCycleCount == 0)
+        		{
+        		    if (CPU.WaitingForInterrupt) {
+            			CPU.WaitingForInterrupt = FALSE;
+            			CPU.PC++;
+        		    }
+        		    if (CPU.IRQActive && !Settings.DisableIRQ) {
+            			if (!CheckFlag (IRQ))
+            			    S9xOpcode_IRQ ();
+        		    }
+        		    else
+            			CPU.Flags &= ~IRQ_PENDING_FLAG;
+        		}
+        		else
+                {
+                    if(--CPU.IRQCycleCount==0 && CheckFlag (IRQ))
+                        CPU.IRQCycleCount=1;
+                }
+    	    }
+
+    	    if (CPU.Flags & SCAN_KEYS_FLAG)
+	        	break;
+    	}
+
+#ifdef CPU_SHUTDOWN
+    	CPU.PCAtOpcodeStart = CPU.PC;
+#endif
+    	CPU.Cycles += CPU.MemSpeed;
+
+    	(*ICPU.S9xOpcodes [*CPU.PC++].S9xOpcode) ();
+	
+    	DO_HBLANK_CHECK_SFX();
     }
 
     ICPU.Registers.PC = CPU.PC - CPU.PCBase;
@@ -191,6 +368,75 @@ void S9xMainLoop (void)
 #endif
 }
 
+void S9xMainLoop_NoSA1_NoSFX (void)
+{
+    for (;;)
+	{
+    	APU_EXECUTE ();
+    
+    	if (CPU.Flags)
+    	{
+    	    if (CPU.Flags & NMI_FLAG)
+    	    {
+        		if (--CPU.NMICycleCount == 0) {
+        		    CPU.Flags &= ~NMI_FLAG;
+        		    if (CPU.WaitingForInterrupt) {
+            			CPU.WaitingForInterrupt = FALSE;
+            			CPU.PC++;
+        		    }
+        		    S9xOpcode_NMI ();
+        		}
+    	    }
+
+    	    CHECK_SOUND ();
+
+    	    if (CPU.Flags & IRQ_PENDING_FLAG)
+    	    {
+        		if (CPU.IRQCycleCount == 0)
+        		{
+        		    if (CPU.WaitingForInterrupt) {
+            			CPU.WaitingForInterrupt = FALSE;
+            			CPU.PC++;
+        		    }
+        		    if (CPU.IRQActive && !Settings.DisableIRQ) {
+            			if (!CheckFlag (IRQ))
+            			    S9xOpcode_IRQ ();
+        		    }
+        		    else
+            			CPU.Flags &= ~IRQ_PENDING_FLAG;
+        		}
+        		else
+                {
+                    if(--CPU.IRQCycleCount==0 && CheckFlag (IRQ))
+                        CPU.IRQCycleCount=1;
+                }
+    	    }
+
+    	    if (CPU.Flags & SCAN_KEYS_FLAG)
+	        	break;
+    	}
+
+#ifdef CPU_SHUTDOWN
+    	CPU.PCAtOpcodeStart = CPU.PC;
+#endif
+    	CPU.Cycles += CPU.MemSpeed;
+
+    	(*ICPU.S9xOpcodes [*CPU.PC++].S9xOpcode) ();
+	
+    	DO_HBLANK_CHECK_NoSFX();
+    }
+
+    ICPU.Registers.PC = CPU.PC - CPU.PCBase;
+    S9xPackStatus ();
+    IAPU.Registers.PC = IAPU.PC - IAPU.RAM;
+    S9xAPUPackStatus ();
+    if (CPU.Flags & SCAN_KEYS_FLAG)
+    {
+	    S9xSyncSpeed ();
+        CPU.Flags &= ~SCAN_KEYS_FLAG;
+    }
+}
+
 void S9xSetIRQ (uint32 source)
 {
     CPU.IRQActive |= source;
@@ -211,7 +457,15 @@ void S9xClearIRQ (uint32 source)
     CLEAR_IRQ_SOURCE (source);
 }
 
-void S9xDoHBlankProcessing ()
+/*
+ * This is a CATSFC modification inspired by a Snes9x-Euphoria modification.
+ * The emulator selects an HBlank processor based on the chips used in an
+ * entire frame. This avoids the constant Settings.SuperFX checks.
+ *
+ * The original version of S9xDoHBlankProcessing is S9xDoHBlankProcessing_SFX
+ * below. Remember to propagate modifications to the NoSFX version.
+ */
+void S9xDoHBlankProcessing_SFX ()
 {
 #ifdef CPU_SHUTDOWN
     CPU.WaitCounter++;
@@ -401,4 +655,193 @@ void S9xDoHBlankProcessing ()
 
     S9xReschedule ();
 }
+void S9xDoHBlankProcessing_NoSFX ()
+{
+#ifdef CPU_SHUTDOWN
+    CPU.WaitCounter++;
+#endif
+    switch (CPU.WhichEvent)
+    {
+		case HBLANK_START_EVENT:
+#ifdef SYNC_JOYPAD_AT_HBLANK
+		// Re-get the controls every hblank. A resolution algorithm in
+		// ppu.cpp will determine with greater accuracy whether a key was
+		// pressed or released during the frame.
+		uint32 i;
+		for (i = 0; i < 5; i++)
+		{
+			IPPU.JoypadsAtHBlanks [i][CPU.V_Counter] = S9xReadJoypad (i);
+		}
+#endif
+		if (IPPU.HDMA && CPU.V_Counter <= PPU.ScreenHeight)
+			IPPU.HDMA = S9xDoHDMA (IPPU.HDMA);
+
+		break;
+
+		case HBLANK_END_EVENT:
+
+#ifndef STORM
+		if (Settings.SoundSync)
+			S9xGenerateSound ();
+#endif
+
+		CPU.Cycles -= Settings.H_Max;
+		if (IAPU.APUExecuting)
+		{
+			APU.Cycles -= Settings.H_Max;
+#ifdef MK_APU
+			S9xCatchupCount();
+#endif
+		}
+		else
+			APU.Cycles = 0;
+
+		CPU.NextEvent = -1;
+		ICPU.Scanline++;
+
+		if (++CPU.V_Counter >= (Settings.PAL ? SNES_MAX_PAL_VCOUNTER : SNES_MAX_NTSC_VCOUNTER))
+		{
+			CPU.V_Counter = 0;
+			Memory.FillRAM[0x213F]^=0x80;
+			PPU.RangeTimeOver = 0;
+			CPU.NMIActive = FALSE;
+			ICPU.Frame++;
+			PPU.HVBeamCounterLatched = 0;
+			CPU.Flags |= SCAN_KEYS_FLAG;
+			S9xStartHDMA ();
+		}
+
+		S9xProcessSound (0);
+
+		if (PPU.VTimerEnabled && !PPU.HTimerEnabled && CPU.V_Counter == PPU.IRQVBeamPos)
+		{
+			S9xSetIRQ (PPU_V_BEAM_IRQ_SOURCE);
+		}
+
+		if (CPU.V_Counter == PPU.ScreenHeight + FIRST_VISIBLE_LINE)
+		{
+			// Start of V-blank
+			S9xEndScreenRefresh ();
+			IPPU.HDMA = 0;
+		    // Bits 7 and 6 of $4212 are computed when read in S9xGetPPU.
+		    missing.dma_this_frame = 0;
+		    IPPU.MaxBrightness = PPU.Brightness;
+		    PPU.ForcedBlanking = (Memory.FillRAM [0x2100] >> 7) & 1;
+
+			if(!PPU.ForcedBlanking)
+			{
+				PPU.OAMAddr = PPU.SavedOAMAddr;
+
+				uint8 tmp = 0;
+				if(PPU.OAMPriorityRotation)
+					tmp = (PPU.OAMAddr&0xFE)>>1;
+				if((PPU.OAMFlip&1) || PPU.FirstSprite!=tmp)
+				{
+					PPU.FirstSprite=tmp;
+					IPPU.OBJChanged=TRUE;
+				}
+
+				PPU.OAMFlip = 0;
+			}
+
+			Memory.FillRAM[0x4210] = 0x80 |Model->_5A22;
+		    if (Memory.FillRAM[0x4200] & 0x80)
+		    {
+				CPU.NMIActive = TRUE;
+				CPU.Flags |= NMI_FLAG;
+				CPU.NMICycleCount = CPU.NMITriggerPoint;
+		    }
+
+#ifdef OLD_SNAPSHOT_CODE
+		    if (CPU.Flags & SAVE_SNAPSHOT_FLAG)
+			{
+				CPU.Flags &= ~SAVE_SNAPSHOT_FLAG;
+				Registers.PC = CPU.PC - CPU.PCBase;
+				S9xPackStatus ();
+				S9xAPUPackStatus ();
+				Snapshot (NULL);
+			}
+#endif
+		}
+
+		if (CPU.V_Counter == PPU.ScreenHeight + 3)
+			S9xUpdateJoypads ();
+
+		if (CPU.V_Counter == FIRST_VISIBLE_LINE)
+		{
+			Memory.FillRAM[0x4210] = Model->_5A22;
+			CPU.Flags &= ~NMI_FLAG;
+			S9xStartScreenRefresh ();
+		}
+		if (CPU.V_Counter >= FIRST_VISIBLE_LINE &&
+			CPU.V_Counter < PPU.ScreenHeight + FIRST_VISIBLE_LINE)
+		{
+			RenderLine (CPU.V_Counter - FIRST_VISIBLE_LINE);
+		}
+	// Use TimerErrorCounter to skip update of SPC700 timers once
+	// every 128 updates. Needed because this section of code is called
+	// once every emulated 63.5 microseconds, which coresponds to
+	// 15.750KHz, but the SPC700 timers need to be updated at multiples
+	// of 8KHz, hence the error correction.
+//	IAPU.TimerErrorCounter++;
+//	if (IAPU.TimerErrorCounter >= )
+//	    IAPU.TimerErrorCounter = 0;
+//	else
+	{
+		if (APU.TimerEnabled [2])
+		{
+			APU.Timer [2] += 4;
+			while (APU.Timer [2] >= APU.TimerTarget [2])
+			{
+				IAPU.RAM [0xff] = (IAPU.RAM [0xff] + 1) & 0xf;
+				APU.Timer [2] -= APU.TimerTarget [2];
+#ifdef SPC700_SHUTDOWN		
+				IAPU.WaitCounter++;
+				IAPU.APUExecuting = TRUE;
+#endif
+			}
+		}
+		if (CPU.V_Counter & 1)
+		{
+			if (APU.TimerEnabled [0])
+			{
+				APU.Timer [0]++;
+				if (APU.Timer [0] >= APU.TimerTarget [0])
+				{
+					IAPU.RAM [0xfd] = (IAPU.RAM [0xfd] + 1) & 0xf;
+					APU.Timer [0] = 0;
+#ifdef SPC700_SHUTDOWN		
+					IAPU.WaitCounter++;
+					IAPU.APUExecuting = TRUE;
+#endif
+				}
+			}
+			if (APU.TimerEnabled [1])
+			{
+				APU.Timer [1]++;
+				if (APU.Timer [1] >= APU.TimerTarget [1])
+				{
+					IAPU.RAM [0xfe] = (IAPU.RAM [0xfe] + 1) & 0xf;
+					APU.Timer [1] = 0;
+#ifdef SPC700_SHUTDOWN		
+					IAPU.WaitCounter++;
+					IAPU.APUExecuting = TRUE;
+#endif		    
+				}
+			}
+	    }
+	}
+		break;
+
+	    case HTIMER_BEFORE_EVENT:
+	    case HTIMER_AFTER_EVENT:
+		if (PPU.HTimerEnabled && (!PPU.VTimerEnabled || CPU.V_Counter == PPU.IRQVBeamPos))
+		{
+		    S9xSetIRQ (PPU_H_BEAM_IRQ_SOURCE);
+		}
+		break;
+	}
+
+    S9xReschedule ();
+}
 
diff --git a/source/cpuexec.h b/source/cpuexec.h
index a094220..dab9099 100644
--- a/source/cpuexec.h
+++ b/source/cpuexec.h
@@ -93,9 +93,13 @@
 #include "memmap.h"
 #include "65c816.h"
 
-#define DO_HBLANK_CHECK() \
+#define DO_HBLANK_CHECK_SFX() \
     if (CPU.Cycles >= CPU.NextEvent) \
-	S9xDoHBlankProcessing ();
+	S9xDoHBlankProcessing_SFX ();
+
+#define DO_HBLANK_CHECK_NoSFX() \
+    if (CPU.Cycles >= CPU.NextEvent) \
+	S9xDoHBlankProcessing_NoSFX ();
 
 struct SOpcodes {
 #ifdef __WIN32__
@@ -126,7 +130,8 @@ START_EXTERN_C
 void S9xMainLoop (void);
 void S9xReset (void);
 void S9xSoftReset (void);
-void S9xDoHBlankProcessing ();
+void S9xDoHBlankProcessing_SFX ();
+void S9xDoHBlankProcessing_NoSFX ();
 void S9xClearIRQ (uint32);
 void S9xSetIRQ (uint32);
 
diff --git a/source/dma.cpp b/source/dma.cpp
index 2addceb..96f705f 100644
--- a/source/dma.cpp
+++ b/source/dma.cpp
@@ -855,8 +855,12 @@ void S9xDoDMA (uint8 Channel)
     IAPU.APUExecuting = Settings.APUEnabled;
     APU_EXECUTE ();
 #endif
-    while (CPU.Cycles > CPU.NextEvent)
-		S9xDoHBlankProcessing ();
+	if (Settings.SuperFX)
+		while (CPU.Cycles > CPU.NextEvent)
+			S9xDoHBlankProcessing_SFX ();
+	else /* if (!Settings.SuperFX) */
+		while (CPU.Cycles > CPU.NextEvent)
+			S9xDoHBlankProcessing_NoSFX ();
 
 	if(Settings.SPC7110&&spc7110_dma)
 	{