From 0d57b7f01acb209d6472024dbca2b0fec29069df Mon Sep 17 00:00:00 2001
From: hrydgard <hrydgard@gmail.com>
Date: Fri, 1 May 2009 19:07:29 +0000
Subject: [PATCH] DSP: I hereby name R08-R11 WR0-WR3, standing for Wrap control
 Registers 0-3. Kill "CR". Document decrements a little bit.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@3126 8ced0084-cf51-0410-be5f-012b33b47a6e
---
 Source/Core/DSPCore/Src/DSPInterpreter.cpp    | 21 ++----
 Source/Core/DSPCore/Src/DSPTables.cpp         | 23 ++++---
 Source/Core/DSPCore/Src/gdsp_ext_op.cpp       | 29 ++++----
 Source/Core/DSPCore/Src/gdsp_ext_op.h         |  1 -
 Source/Core/DSPCore/Src/gdsp_opcodes_helper.h | 68 ++++++++-----------
 Source/Core/DSPCore/Src/gdsp_registers.h      | 29 ++++----
 Source/DSPSpy/tests/dr_test.ds                | 20 +++---
 Source/DSPSpy/tests/ir_test.ds                | 18 ++---
 8 files changed, 95 insertions(+), 114 deletions(-)

diff --git a/Source/Core/DSPCore/Src/DSPInterpreter.cpp b/Source/Core/DSPCore/Src/DSPInterpreter.cpp
index 774da08cbb..9c1c2640c1 100644
--- a/Source/Core/DSPCore/Src/DSPInterpreter.cpp
+++ b/Source/Core/DSPCore/Src/DSPInterpreter.cpp
@@ -457,6 +457,7 @@ void lri(const UDSPInstruction& opc)
 	u8 reg  = opc.hex & DSP_REG_MASK;
 	u16 imm = dsp_fetch_code();
 	dsp_op_write_reg(reg, imm);
+	dsp_conditional_extend_accum(reg);
 }
 
 // LRIS $(0x18+D), #I
@@ -468,6 +469,7 @@ void lris(const UDSPInstruction& opc)
 	u8 reg  = ((opc.hex >> 8) & 0x7) + DSP_REG_AXL0;
 	u16 imm = (s8)opc.hex;
 	dsp_op_write_reg(reg, imm);
+	dsp_conditional_extend_accum(reg);
 }
 
 // LR $D, @M
@@ -481,6 +483,7 @@ void lr(const UDSPInstruction& opc)
 	u16 addr = dsp_fetch_code();
 	u16 val  = dsp_dmem_read(addr);
 	dsp_op_write_reg(reg, val);
+	dsp_conditional_extend_accum(reg);
 }
 
 // SR @M, $S
@@ -1392,12 +1395,7 @@ void dar(const UDSPInstruction& opc)
 {
 	int reg = opc.hex & 0x3;
 
-	int temp = g_dsp.r[reg] + g_dsp.r[DSP_REG_R08];
-
-	if (temp <= 0x7ff)  // ???
-		g_dsp.r[reg] = temp;
-	else
-		g_dsp.r[reg]--;
+	g_dsp.r[reg]--;   // TODO: Wrap properly.
 }
 
 
@@ -1410,12 +1408,7 @@ void iar(const UDSPInstruction& opc)
 {
 	int reg = opc.hex & 0x3;
 
-	int temp = g_dsp.r[reg] + g_dsp.r[DSP_REG_R08];
-
-	if (temp <= 0x7ff)  // ???
-		g_dsp.r[reg] = temp;
-	else
-		g_dsp.r[reg]++;
+	g_dsp.r[reg]++;   // TODO: Wrap properly according to the corresponding WR register.
 }
 
 //-------------------------------------------------------------
@@ -1476,7 +1469,7 @@ void srbith(const UDSPInstruction& opc)
 		g_dsp.r[DSP_REG_SR] |= SR_TOP_BIT_UNK;
 		break;
 
-	// 40-bit precision? clamping? no idea :(
+	// Automatic 40-bit sign extension when loading ACx.M.
 	// 40 seems to be the default.
 	// Confirmed these by using DSPSpy and copying the value of SR to R00 after setting.
 	case 0xe:  // SET16  (really, clear SR's 0x4000) 
@@ -1583,7 +1576,7 @@ void mulmvz(const UDSPInstruction& opc)
 	dsp_set_long_acc(rreg, acc);
 
 	// math prod
-	prod = (s64)g_dsp.r[DSP_REG_AXL0 + sreg] * (s64)g_dsp.r[DSP_REG_AXH0 + sreg] * GetMultiplyModifier();
+	prod = (s64)(s16)g_dsp.r[DSP_REG_AXL0 + sreg] * (s64)(s16)g_dsp.r[DSP_REG_AXH0 + sreg] * GetMultiplyModifier();
 	dsp_set_long_prod(prod);
 
 	Update_SR_Register64(prod);
diff --git a/Source/Core/DSPCore/Src/DSPTables.cpp b/Source/Core/DSPCore/Src/DSPTables.cpp
index 40aac6792d..845d414180 100644
--- a/Source/Core/DSPCore/Src/DSPTables.cpp
+++ b/Source/Core/DSPCore/Src/DSPTables.cpp
@@ -309,7 +309,6 @@ const DSPOPCTemplate cw =
 const DSPOPCTemplate opcodes_ext[] =
 {	
 	// FIXME: guessing this is cr need checking
-	{"CR",      0x0000, 0x00fc, DSPInterpreter::Ext::cr, nop, 1, 1, {{P_REG, 1, 0, 0, 0x0003}}, NULL, NULL,},
 	{"DR",      0x0004, 0x00fc, DSPInterpreter::Ext::dr, nop, 1, 1, {{P_REG, 1, 0, 0, 0x0003}}, NULL, NULL,},
 	{"IR",      0x0008, 0x00fc, DSPInterpreter::Ext::ir, nop, 1, 1, {{P_REG, 1, 0, 0, 0x0003}}, NULL, NULL,},
 	{"NR",      0x000c, 0x00fc, DSPInterpreter::Ext::nr, nop, 1, 1, {{P_REG, 1, 0, 0, 0x0003}}, NULL, NULL,},
@@ -401,14 +400,14 @@ const pdlabel_t regnames[] =
 	{0x01, "AR1",       "Addr Reg 01",},
 	{0x02, "AR2",       "Addr Reg 02",},
 	{0x03, "AR3",       "Addr Reg 03",},
-	{0x04, "IX0",       "Index Reg 0(04)",},
-	{0x05, "IX1",       "Index Reg 1(05)",},
-	{0x06, "IX2",       "Index Reg 2(06)",},
-	{0x07, "IX3",       "Indec Reg 3(07)",},
-	{0x08, "R08",       "Register 08",},
-	{0x09, "R09",       "Register 09",},
-	{0x0a, "R10",       "Register 10",},
-	{0x0b, "R11",       "Register 11",},
+	{0x04, "IX0",       "Index Reg 0",},
+	{0x05, "IX1",       "Index Reg 1",},
+	{0x06, "IX2",       "Index Reg 2",},
+	{0x07, "IX3",       "Index Reg 3",},
+	{0x08, "WR0",       "Wrapping Register 0",},
+	{0x09, "WR1",       "Wrapping Register 1",},
+	{0x0a, "WR2",       "Wrapping Register 2",},
+	{0x0b, "WR3",       "Wrapping Register 3",},
 	{0x0c, "ST0",       "Call stack",},
 	{0x0d, "ST1",       "Data stack",},
 	{0x0e, "ST2",       "Loop addr stack",},
@@ -435,6 +434,12 @@ const pdlabel_t regnames[] =
 	{0x21, "ACC1",		"Accu Full 1",},
 	{0x22, "AX0",		"Extra Accu 0",},
 	{0x23, "AX1",		"Extra Accu 1",},
+
+	// Old names for the wrapping registers, for compatibility.
+	{0x08, "R08",       "Wrapping Register 0(08)",},
+	{0x09, "R09",       "Wrapping Register 1(09)",},
+	{0x0a, "R10",       "Wrapping Register 2(10)",},
+	{0x0b, "R11",       "Wrapping Register 3(11)",},
 };
 
 u8 opSize[OPTABLE_SIZE];
diff --git a/Source/Core/DSPCore/Src/gdsp_ext_op.cpp b/Source/Core/DSPCore/Src/gdsp_ext_op.cpp
index e8fdee0e4b..554d5a391b 100644
--- a/Source/Core/DSPCore/Src/gdsp_ext_op.cpp
+++ b/Source/Core/DSPCore/Src/gdsp_ext_op.cpp
@@ -29,22 +29,21 @@
 // attached to opcodes that allow extending (8 lower bits of opcode not used by
 // opcode). Extended opcodes do not modify program counter $pc register.
 
+// Most of the suffixes increment or decrement one or more addressing registers
+// (the first four, ARx). The increment/decrement is either 1, or the corresponding 
+// "index" register (the second four, IXx). The addressing registers will wrap
+// in odd ways, dictated by the corresponding wrapping register, WP0-3.
+
+// The following should be applied as a decrement:
+// ar[i] = (ar[i] & wp[i]) == 0 ? ar[i] | wp[i] : ar[i] - 1;
+// I have not found the corresponding algorithms for increments yet.
+// It's gotta be fairly simple though. See R3123, R3125 in Google Code.
 
 namespace DSPInterpreter
 {
 namespace Ext 
 {
 
-// CR $arR
-// xxxx xxxx 0000 00rr
-// Clearing addressing register $arR.
-// This is not in any doc and as such just a guess
-void cr(const UDSPInstruction& opc) {
-	u8 reg = opc.hex & 0x3;	
-	
-	g_dsp.r[reg] = 0;
-}
-
 // DR $arR
 // xxxx xxxx 0000 01rr
 // Decrement addressing register $arR.
@@ -208,7 +207,6 @@ void dsp_op_ext_r_epi(const UDSPInstruction& opc)
 {
 	u8 op  = (opc.hex >> 2) & 0x3;
 	u8 reg = opc.hex & 0x3;
-
 	switch (op) {
 	case 0x00: // 
 		g_dsp.r[reg] = 0;
@@ -375,7 +373,6 @@ void dsp_op_ext_ld(const UDSPInstruction& opc)
 //
 //
 // ================================================================================
-
 void dsp_op_ext_ops_pro(const UDSPInstruction& opc)
 {
 	if ((opc.hex & 0xFF) == 0){return;}
@@ -428,7 +425,10 @@ void dsp_op_ext_ops_pro(const UDSPInstruction& opc)
 
 void dsp_op_ext_ops_epi(const UDSPInstruction& opc)
 {
-	if ((opc.hex & 0xFF) == 0){return;}
+	if ((opc.hex & 0xFF) == 0)
+	{
+		return;
+	}
 
 	switch ((opc.hex >> 4) & 0xf)
 	{
@@ -436,7 +436,6 @@ void dsp_op_ext_ops_epi(const UDSPInstruction& opc)
 	  case 0x09:
 	  case 0x0a:
 	  case 0x0b:
-
 		  if (opc.hex & 0x2)
 		  {
 			  dsp_op_ext_sl_epi(opc.hex);
@@ -450,5 +449,3 @@ void dsp_op_ext_ops_epi(const UDSPInstruction& opc)
 		  return;
 	}
 }
-
-
diff --git a/Source/Core/DSPCore/Src/gdsp_ext_op.h b/Source/Core/DSPCore/Src/gdsp_ext_op.h
index 6601059fac..7040980bd3 100644
--- a/Source/Core/DSPCore/Src/gdsp_ext_op.h
+++ b/Source/Core/DSPCore/Src/gdsp_ext_op.h
@@ -62,7 +62,6 @@ void ldn(const UDSPInstruction& opc);
 void ldm(const UDSPInstruction& opc);
 void ldnm(const UDSPInstruction& opc);
 void mv(const UDSPInstruction& opc);
-void cr(const UDSPInstruction& opc);
 void dr(const UDSPInstruction& opc);
 void ir(const UDSPInstruction& opc);
 void nr(const UDSPInstruction& opc);
diff --git a/Source/Core/DSPCore/Src/gdsp_opcodes_helper.h b/Source/Core/DSPCore/Src/gdsp_opcodes_helper.h
index 02eae01d85..e1bbc236ed 100644
--- a/Source/Core/DSPCore/Src/gdsp_opcodes_helper.h
+++ b/Source/Core/DSPCore/Src/gdsp_opcodes_helper.h
@@ -83,50 +83,19 @@ inline u16 dsp_op_read_reg(u8 reg)
 inline void dsp_op_write_reg(u8 reg, u16 val)
 {
 	switch (reg & 0x1f) {
-
-
+		/*
+	case DSP_REG_ACH0:
+	case DSP_REG_ACH1:
+		// sign extend from the bottom 8 bits.
+		g_dsp.r[reg] = (u16)(s16)(s8)(u8)val;
+		break;
+*/
 	case 0x0c:
 	case 0x0d:
 	case 0x0e:
 	case 0x0f:
 		dsp_reg_store_stack(reg - 0x0c, val);
 		break;
-#if 0 // FIXME
-	case 0x1e: // AC0.M
-	case 0x1f: // AC1.M
-		// in "s16 mode", LRI $AC0.M, xxx  will set AC0.L and AC0.H to 0, 
-		// while it won't in "s40 mode".
-		if (g_dsp.r[DSP_REG_SR] & SR_16_BIT) {
-				g_dsp.r[reg - 0x2] = 0; // L
-				g_dsp.r[reg - 0xe] = 0; // H
-		}
-		
-		g_dsp.r[reg] = val;
-		break;
-
-
-	case 0x1c: // AC0.L
-	case 0x1d: // AC1.L
-		
-		if (g_dsp.r[DSP_REG_SR] & SR_16_BIT) {
-				g_dsp.r[reg + 0x2] = 0; // M
-				g_dsp.r[reg - 0xc] = 0; // H
-		}
-		
-		g_dsp.r[reg] = val;
-		break;
-
-	case 0x10: // AC0.H
-	case 0x11: // AC1.H
-		if (g_dsp.r[DSP_REG_SR] & SR_16_BIT) {
-				g_dsp.r[reg + 0xc] = 0; // L
-				g_dsp.r[reg + 0xe] = 0; // M
-		}
-		
-		g_dsp.r[reg] = val;
-		break;
-#endif
-
 
 	default:
 		g_dsp.r[reg] = val;
@@ -134,6 +103,25 @@ inline void dsp_op_write_reg(u8 reg, u16 val)
 	}
 }
 
+inline void dsp_conditional_extend_accum(u8 reg) 
+{
+#if 0
+	switch (reg) 
+	{
+	case DSP_REG_ACM0:
+	case DSP_REG_ACM1:
+		if (g_dsp.r[DSP_REG_SR] & SR_40_MODE_BIT)
+		{
+			// Sign extend into whole accum.
+			u16 val = g_dsp.r[reg];
+			g_dsp.r[reg - DSP_REG_ACM0 + DSP_REG_ACH0] = (val & 0x8000) ? 0xFFFF : 0x0000;
+			g_dsp.r[reg - DSP_REG_ACM0 + DSP_REG_ACL0] = 0;
+		}
+	}
+#endif
+}
+
+
 
 // ---------------------------------------------------------------------------------------
 //
@@ -253,13 +241,13 @@ inline s64 dsp_get_long_acx(int _reg)
 inline s16 dsp_get_ax_l(int _reg)
 {
 	_assert_(_reg < 2);
-	return g_dsp.r[0x18 + _reg];
+	return (s16)g_dsp.r[0x18 + _reg];
 }
 
 inline s16 dsp_get_ax_h(int _reg)
 {
 	_assert_(_reg < 2);
-	return g_dsp.r[0x1a + _reg];
+	return (s16)g_dsp.r[0x1a + _reg];
 }
 
 #endif
diff --git a/Source/Core/DSPCore/Src/gdsp_registers.h b/Source/Core/DSPCore/Src/gdsp_registers.h
index 08838c2a38..473f5993b8 100644
--- a/Source/Core/DSPCore/Src/gdsp_registers.h
+++ b/Source/Core/DSPCore/Src/gdsp_registers.h
@@ -31,19 +31,19 @@
 #define DSP_REG_AR0         0x00    // address registers
 #define DSP_REG_AR1         0x01
 #define DSP_REG_AR2         0x02
-#define DSP_REG_AR3         0x03    // used as jump function selector
+#define DSP_REG_AR3         0x03
 
-#define DSP_REG_IX0         0x04    // LEFT_VOLUME accel
-#define DSP_REG_IX1         0x05    // RIGHT_VOLUME accel
-#define DSP_REG_IX2         0x06    // ADDRH_SMP accel
-#define DSP_REG_IX3         0x07    // ADDRL_SMP accel
+#define DSP_REG_IX0         0x04    // indexing registers (actually, mostly used as increments)
+#define DSP_REG_IX1         0x05
+#define DSP_REG_IX2         0x06
+#define DSP_REG_IX3         0x07
 
-#define DSP_REG_R08         0x08    // fixed to 48000 value
-#define DSP_REG_R09         0x09    // problems using this
-#define DSP_REG_R0A         0x0a    // ADDREH_SMP accel
-#define DSP_REG_R0B         0x0b    // ADDREL_SMP accel
+#define DSP_REG_WR0         0x08    // address wrapping registers. should be initialized to 0xFFFF if not used.
+#define DSP_REG_WR1         0x09
+#define DSP_REG_WR2         0x0a
+#define DSP_REG_WR3         0x0b
 
-#define DSP_REG_ST0         0x0c
+#define DSP_REG_ST0         0x0c    // stacks.
 #define DSP_REG_ST1         0x0d
 #define DSP_REG_ST2         0x0e
 #define DSP_REG_ST3         0x0f
@@ -51,15 +51,15 @@
 #define DSP_REG_CONFIG      0x12
 #define DSP_REG_SR          0x13
 
-#define DSP_REG_PRODL       0x14
+#define DSP_REG_PRODL       0x14    // product.
 #define DSP_REG_PRODM       0x15
 #define DSP_REG_PRODH       0x16
 #define DSP_REG_PRODM2      0x17 
 
 #define DSP_REG_AXL0        0x18
 #define DSP_REG_AXL1        0x19
-#define DSP_REG_AXH0        0x1a    // SMP_R accel
-#define DSP_REG_AXH1        0x1b    // SMP_L accel
+#define DSP_REG_AXH0        0x1a
+#define DSP_REG_AXH1        0x1b
 
 #define DSP_REG_ACC0        0x1c    // accumulator (global)
 #define DSP_REG_ACC1        0x1d
@@ -85,6 +85,7 @@
 #define DSP_REG_CMBH        0xfffe  // CPU Mailbox H
 #define DSP_REG_CMBL        0xffff  // CPU Mailbox L
 
+
 #define DMA_TO_DSP          0
 #define DMA_TO_CPU          1
 
@@ -92,7 +93,6 @@
 #define DSP_STACK_C 0
 #define DSP_STACK_D 1
 
-
 // SR bits
 #define SR_CARRY       0x0001
 #define SR_2           0x0002   // overflow???
@@ -107,7 +107,6 @@
 #define SR_40_MODE_BIT 0x4000   // 0 = "16", 1 = "40"  (SET16, SET40)  Controls sign extension when loading mid accums.
 #define SR_TOP_BIT_UNK 0x8000   // 1 = normal. 0 = x2  (CLR15, SET15) ????????
 
-
 void dsp_reg_store_stack(u8 stack_reg, u16 val);
 u16 dsp_reg_load_stack(u8 stack_reg);
 
diff --git a/Source/DSPSpy/tests/dr_test.ds b/Source/DSPSpy/tests/dr_test.ds
index 03f9dabbd8..2eb3d31ec0 100644
--- a/Source/DSPSpy/tests/dr_test.ds
+++ b/Source/DSPSpy/tests/dr_test.ds
@@ -2,24 +2,24 @@
 incdir  "tests"
 include "dsp_base.inc"
 
-; Tests done using AR1 = 0x0010, IX1 = 0. LP1 means R09.
-; LP1 = 0
+; Tests done using AR1 = 0x0010, IX1 = 0. WR1 (wrap 1) means R09.
+; WR1 = 0
 ;   10, 10, 10, 10,
-; LP1 = 1
+; WR1 = 1
 ;   10, 11, 10, 11,
-; LP1 = 2
+; WR1 = 2
 ;   10, 12, 11, 13, 12, 11, 13, 12, 11, 13, 12, 11
-; LP1 = 3
+; WR1 = 3
 ;   10, 13, 12, 11, 10, 13, 12...
-; LP1 = 4
+; WR1 = 4
 ;   10, 14, 13, 17, 16, 15, 14, 13, 17, 16, 15, 14, 13, ... 
-; LP1 = 5
+; WR1 = 5
 ;   10, 15, 14, 13, 12, 17, 16, 15, 14, 13, 12, 17, 
-; LP1 = 6
+; WR1 = 6
 ;   10, 16, 15, 14, 13, 12, 11, 17, 16, 15, 14, 13
-; LP1 = 7
+; WR1 = 7
 ;   10, 17, 16, 15, .. normal
-; LP1 = 8
+; WR1 = 8
 ;   10, 18, 17, 1f, 1e, 1d, 1c, 1b, 1a, 19, 18, 17, 1f, 1e, .....
 
 
diff --git a/Source/DSPSpy/tests/ir_test.ds b/Source/DSPSpy/tests/ir_test.ds
index a52fe0d63b..e7ee5e332a 100644
--- a/Source/DSPSpy/tests/ir_test.ds
+++ b/Source/DSPSpy/tests/ir_test.ds
@@ -8,23 +8,23 @@ include "dsp_base.inc"
 ; These were proven FALSE though by the following:
 
 ; Tests done using AR1 = 0x0010, IX1 = 0
-; LP1 = 0
+; WR1 = 0
 ;   10, 11, 11, 11, 11......
-; LP1 = 1
+; WR1 = 1
 ;   10, 11, 10, 11, 10......
-; LP1 = 2
+; WR1 = 2
 ;   10, 11, 12, 13, 11, 12, 13, 11, 12, 13 ......
-; LP1 = 3
+; WR1 = 3
 ;   10, 11, 12, 13, 10, 11, 12, 13, 10, 11, 12, 13.......
-; LP1 = 4
+; WR1 = 4
 ;   10, 11, 12, 13, 14, 15, 16, 17, 13, 14, 15, 16, 17, 13, 14, 15 ......
-; LP1 = 5
+; WR1 = 5
 ;   10, 11, 12, 13, 14, 15, 16, 17, 12, 13, 14, 15 ...
-; LP1 = 6
+; WR1 = 6
 ;   10, 11, 12, 13, 14, 15, 16, 17, 11, 12, 13, 14...
-; LP1 = 7
+; WR1 = 7
 ;   10, 11, 12, 13, 14, 15, 16, 17, 10, 11, ....
-; LP1 = 8
+; WR1 = 8
 ;   10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 1a, 1b, 1c, 1d, 1e, 1f, 17, 18, 19, 1a, 1b.....