DSP LLE: Remove byteswaps and other bloat from the hot path (do them at load/dma time). Minor speedup, perhaps.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@2939 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Plugins/Plugin_DSP_LLE/Src/DSPInterpreter.cpp

@@ -32,7 +32,7 @@ namespace DSPInterpreter {
 void unknown(const UDSPInstruction& opc)
 {
 	//_assert_msg_(MASTER_LOG, !g_dsp.exception_in_progress_hack, "assert while exception");
-	ERROR_LOG(DSPLLE, "LLE: Unrecognized opcode 0x%04x, pc 0x%04x", opc.hex, g_dsp.err_pc);
+	ERROR_LOG(DSPLLE, "LLE: Unrecognized opcode 0x%04x, pc 0x%04x", opc.hex, g_dsp.pc);
 }
 
 // test register and updates SR accordingly
@@ -158,7 +158,7 @@ void rti(const UDSPInstruction& opc)
 void halt(const UDSPInstruction& opc)
 {
 	g_dsp.cr |= 0x4;
-	g_dsp.pc = g_dsp.err_pc;
+	g_dsp.pc--;
 }
 
 
@@ -1688,7 +1688,7 @@ void msubc(const UDSPInstruction& opc)
 void srs(const UDSPInstruction& opc)
 {
 	u8 reg   = ((opc.hex >> 8) & 0x7) + 0x18;
-	u16 addr = (s8)opc.hex;
+	u16 addr = (u16)(s16)(s8)opc.hex;
 	dsp_dmem_write(addr, g_dsp.r[reg]);
 }
   
@@ -1701,7 +1701,7 @@ void srs(const UDSPInstruction& opc)
 void lrs(const UDSPInstruction& opc)
 {
 	u8 reg   = ((opc.hex >> 8) & 0x7) + 0x18;
-	u16 addr = (s8) opc.hex;
+	u16 addr = (u16)(s16)(s8)opc.hex;
 	g_dsp.r[reg] = dsp_dmem_read(addr);
 }
 

Source/Plugins/Plugin_DSP_LLE/Src/DSPInterpreter.h

@@ -25,6 +25,8 @@
 
 namespace DSPInterpreter {
 
+typedef void (*DSPInterpreterFunc)(const UDSPInstruction& opc);
+
 void unknown(const UDSPInstruction& opc);
 void call(const UDSPInstruction& opc);
 void callr(const UDSPInstruction& opc);

Source/Plugins/Plugin_DSP_LLE/Src/Globals.h

@@ -36,6 +36,5 @@ u32 Memory_Read_U32(u32 _uAddress);
 	void ProfilerStart();
 #endif
 
-
 #endif
 

Source/Plugins/Plugin_DSP_LLE/Src/Tools.cpp

@@ -102,7 +102,7 @@ bool DumpCWCode(u32 _Address, u32 _Length)
 	{
 		for (size_t i = _Address; i < _Address + _Length; i++)
 		{
-			u16 val = Common::swap16(g_dsp.iram[i]);
+			u16 val = g_dsp.iram[i];
 			fprintf(pFile, "    cw 0x%04x \n", val);
 		}
 

Source/Plugins/Plugin_DSP_LLE/Src/disassemble.cpp

@@ -341,7 +341,7 @@ char* gd_dis_opcode(gd_globals_t* gdg)
 	}
 
 	pc &= 0x0fff;
-	op1 = Common::swap16(gdg->binbuf[pc]);
+	op1 = gdg->binbuf[pc];
 
 	// find opcode
 	for (j = 0; j < opcodes_size; j++)

Source/Plugins/Plugin_DSP_LLE/Src/gdsp_condition_codes.cpp

@@ -68,6 +68,8 @@ void Update_SR_Register16(s16 _Value)
 	}
 }
 
+// If this always returns 1, Hermes' demo sounds better.
+// However, most AX games are negatively affected.
 int GetMultiplyModifier()
 {
 	if (g_dsp.r[DSP_REG_SR] & (1 << 13))

Source/Plugins/Plugin_DSP_LLE/Src/gdsp_interface.cpp

@@ -90,7 +90,7 @@ void gdsp_mbox_write_l(u8 mbx, u16 val)
 
 	if (mbx == GDSP_MBOX_DSP)
 	{
-		DEBUG_LOG(DSPLLE, " - DSP writes mail to mbx %i: 0x%08x (pc=0x%04x)", mbx, gdsp_mbox_peek(GDSP_MBOX_DSP), g_dsp.err_pc);
+		DEBUG_LOG(DSPLLE, " - DSP writes mail to mbx %i: 0x%08x (pc=0x%04x)", mbx, gdsp_mbox_peek(GDSP_MBOX_DSP), g_dsp.pc);
 	}
 }
 
@@ -110,7 +110,7 @@ u16 gdsp_mbox_read_l(u8 mbx)
 	val = gdsp_mbox[mbx][1];
 	gdsp_mbox[mbx][0] &= ~0x8000;
 
-	DEBUG_LOG(DSPLLE, "- DSP reads mail from mbx %i: %08x (pc=0x%04x)", mbx, gdsp_mbox_peek(mbx), g_dsp.err_pc);
+	DEBUG_LOG(DSPLLE, "- DSP reads mail from mbx %i: %08x (pc=0x%04x)", mbx, gdsp_mbox_peek(mbx), g_dsp.pc);
 
 	if (g_dspInitialize.bOnThread)
 		g_CriticalSection.Leave();
@@ -127,7 +127,6 @@ void gdsp_ifx_write(u16 addr, u16 val)
 	    case 0xfb: // DIRQ
 		    if (val & 0x1)
 			    g_dsp.irq_request();
-
 		    break;
 
 	    case 0xfc: // DMBH
@@ -211,7 +210,7 @@ void gdsp_idma_in(u16 dsp_addr, u32 addr, u32 size)
 	for (u32 i = 0; i < size; i += 2)
 	{
 		// TODO : this may be different on Wii.
-		*(u16*)&dst[dsp_addr + i] = *(u16*)&g_dsp.cpu_ram[(addr + i) & 0x0fffffff];
+		*(u16*)&dst[dsp_addr + i] = Common::swap16(*(u16*)&g_dsp.cpu_ram[(addr + i) & 0x0fffffff]);
 	}
 	WriteProtectMemory(g_dsp.iram, DSP_IRAM_BYTE_SIZE, false);
 
@@ -242,7 +241,7 @@ void gdsp_ddma_in(u16 dsp_addr, u32 addr, u32 size)
 
 	for (u32 i = 0; i < size; i += 2)
 	{
-		*(u16*)&dst[dsp_addr + i] = *(u16*)&g_dsp.cpu_ram[(addr + i) & 0x7FFFFFFF];
+		*(u16*)&dst[dsp_addr + i] = Common::swap16(*(u16*)&g_dsp.cpu_ram[(addr + i) & 0x7FFFFFFF]);
 	}
 
 	INFO_LOG(DSPLLE, "*** ddma_in RAM (0x%08x) -> DRAM_DSP (0x%04x) : size (0x%08x)\n", addr, dsp_addr / 2, size);
@@ -261,7 +260,7 @@ void gdsp_ddma_out(u16 dsp_addr, u32 addr, u32 size)
 
 	for (u32 i = 0; i < size; i += 2)
 	{
-		*(u16*)&g_dsp.cpu_ram[(addr + i) & 0x7FFFFFFF] = *(u16*)&src[dsp_addr + i];
+		*(u16*)&g_dsp.cpu_ram[(addr + i) & 0x7FFFFFFF] = Common::swap16(*(u16*)&src[dsp_addr + i]);
 	}
 
 	INFO_LOG(DSPLLE, "*** ddma_out DRAM_DSP (0x%04x) -> RAM (0x%08x) : size (0x%08x)\n", dsp_addr / 2, addr, size);

Source/Plugins/Plugin_DSP_LLE/Src/gdsp_interpreter.cpp

@@ -89,7 +89,7 @@ void gdsp_init()
 	// Just zero out DRAM.
 	for (int i = 0; i < DSP_DRAM_SIZE; i++)
 	{
-		g_dsp.dram[i] = 0x0021;
+		g_dsp.dram[i] = 0x2100;
 	}
 
 	// copied from a real console after the custom UCode has been loaded
@@ -145,6 +145,11 @@ bool gdsp_load_irom(const char *fname)
 			return false;
 		}
 		fclose(pFile);
+	
+		// Byteswap the rom.
+		for (int i = 0; i < DSP_IROM_SIZE; i++)
+			g_dsp.irom[i] = Common::swap16(g_dsp.irom[i]);
+
 		return true;
 	}
 	// Always keep IROM write protected.
@@ -166,6 +171,9 @@ bool gdsp_load_coef(const char *fname)
 			return false;
 		}
 		fclose(pFile);
+		// Byteswap the rom.
+		for (int i = 0; i < DSP_IROM_SIZE; i++)
+			g_dsp.coef[i] = Common::swap16(g_dsp.coef[i]);
 		return true;
 	}
 	// Always keep COEF write protected. We unprotect only when DMA-ing
@@ -213,9 +221,9 @@ void gdsp_step()
 {
 	g_dsp.step_counter++;
 
+#if PROFILE
 	g_dsp.err_pc = g_dsp.pc;
 
-#if PROFILE
 	ProfilerAddDelta(g_dsp.err_pc, 1);
 	if (g_dsp.step_counter == 1)
 	{
@@ -226,7 +234,6 @@ void gdsp_step()
 	{
 		ProfilerDump(g_dsp.step_counter);
 	}
-
 #endif
 
 	u16 opc = dsp_fetch_code();
@@ -334,8 +341,6 @@ void gdsp_run_cycles(int cycles)
 	// idle loop and if so we waste some time here. Might be beneficial to slice even further.
 	while (cycles > 0)
 	{
-		if (cr_halt)
-			return;
 		gdsp_step();
 		cycles--;
 		// We don't bother directly supporting pause - if the main emu pauses,

Source/Plugins/Plugin_DSP_LLE/Src/gdsp_interpreter.h

@@ -75,7 +75,9 @@ struct SDSP
 {
 	u16 r[32];
 	u16 pc;
+#if PROFILE
 	u16 err_pc;
+#endif
 	u16* iram;
 	u16* dram;
 	u16* irom;

Source/Plugins/Plugin_DSP_LLE/Src/gdsp_memory.cpp

@@ -30,19 +30,14 @@
 #include "gdsp_memory.h"
 #include "gdsp_interface.h"
 
-u16 dsp_swap16(u16 x)
-{
-	return (x >> 8) | (x << 8);
-}
-
 u16 dsp_imem_read(u16 addr)
 {
 	switch (addr >> 12)
 	{
 	case 0:
-		return dsp_swap16(g_dsp.iram[addr & DSP_IRAM_MASK]);
+		return g_dsp.iram[addr & DSP_IRAM_MASK];
 	case 8:
-		return dsp_swap16(g_dsp.irom[addr & DSP_IROM_MASK]);
+		return g_dsp.irom[addr & DSP_IROM_MASK];
 	default:
 		ERROR_LOG(DSPLLE, "%04x DSP ERROR: Executing from invalid (%04x) memory", g_dsp.pc, addr);
 		return 0;
@@ -54,10 +49,10 @@ u16 dsp_dmem_read(u16 addr)
 	switch (addr >> 12)
 	{
 	case 0x0: // 0xxx DRAM
-		return dsp_swap16(g_dsp.dram[addr & DSP_DRAM_MASK]);
+		return g_dsp.dram[addr & DSP_DRAM_MASK];
 		
 	case 0x1: // 1xxx COEF
-		return dsp_swap16(g_dsp.coef[addr & DSP_COEF_MASK]);
+		return g_dsp.coef[addr & DSP_COEF_MASK];
 
 		// FIXME: unknown addresses used by zelda 
  /*	case 0x2:
@@ -74,17 +69,16 @@ u16 dsp_dmem_read(u16 addr)
 	}
 }
 
-
 void dsp_dmem_write(u16 addr, u16 val)
 {
 	switch (addr >> 12)
 	{
 	case 0x0: // 0xxx DRAM
-		  g_dsp.dram[addr & DSP_DRAM_MASK] = dsp_swap16(val);
+		g_dsp.dram[addr & DSP_DRAM_MASK] = val;
 		break;
 
 	case 0x1: // 1xxx COEF
-		  ERROR_LOG(DSPLLE, "someone writes to COEF");
+		ERROR_LOG(DSPLLE, "someone writes to COEF (pc = %02x)", g_dsp.pc);
 		break;
 
 	case 0xf: // Fxxx HW regs
@@ -96,15 +90,3 @@ void dsp_dmem_write(u16 addr, u16 val)
 		break;
 	}
 }
-
-u16 dsp_fetch_code()
-{
-	u16 opc = dsp_imem_read(g_dsp.pc);
-	g_dsp.pc++;
-	return opc;
-}
-
-u16 dsp_peek_code()
-{
-	return dsp_imem_read(g_dsp.pc);
-}

Source/Plugins/Plugin_DSP_LLE/Src/gdsp_memory.h

@@ -27,10 +27,22 @@
 
 #include "Globals.h"
 
-u16  dsp_fetch_code();
+#include "gdsp_interpreter.h"
-u16  dsp_peek_code();
+
 u16  dsp_imem_read(u16 addr);
 void dsp_dmem_write(u16 addr, u16 val);
 u16  dsp_dmem_read(u16 addr);
 
+inline u16 dsp_fetch_code()
+{
+	u16 opc = dsp_imem_read(g_dsp.pc);
+	g_dsp.pc++;
+	return opc;
+}
+
+inline u16 dsp_peek_code()
+{
+	return dsp_imem_read(g_dsp.pc);
+}
+
 #endif

Source/Plugins/Plugin_DSP_LLE/Src/main.cpp

@@ -49,9 +49,6 @@ SoundStream *soundStream = NULL;
 #define GDSP_MBOX_CPU   0
 #define GDSP_MBOX_DSP   1
 
-extern u32 m_addressPBs;
-bool AXTask(u32& _uMail);
-
 bool bCanWork = false;
 bool bIsRunning = false;
 
@@ -312,18 +309,6 @@ void DSP_WriteMailboxLow(bool _CPUMailbox, u16 _uLowMail)
 	{
 		gdsp_mbox_write_l(GDSP_MBOX_CPU, _uLowMail);
 
-		u32 uAddress = gdsp_mbox_peek(GDSP_MBOX_CPU);
-		u16 errpc = g_dsp.err_pc;
-
-		DEBUG_LOG(DSPLLE, "CPU writes mail to mbx 0: 0x%08x (pc=0x%04x)\n", uAddress, errpc);
-
-		// I couldn't find any better way to detect the AX mails so this had to
-		// do. Please feel free to change it.
-		if ((errpc == 0x0054 || errpc == 0x0055) && m_addressPBs == 0)
-		{
-			DEBUG_LOG(DSPLLE, "AXTask ======== 0x%08x (pc=0x%04x)", uAddress, errpc);
-			AXTask(uAddress);
-		}
 	}
 	else
 	{