From f56a2cb8c610ab09f98d98e366e595b7c2e2d9e1 Mon Sep 17 00:00:00 2001
From: tudnai <tamas.rudnai@gmail.com>
Date: Wed, 29 Apr 2020 15:35:54 -0700
Subject: [PATCH] Speed bump (2x) - Precise MHz measurements

---
 A2Mac.xcodeproj/project.pbxproj              | 12 ++--
 A2Mac/Base.lproj/Main.storyboard             |  4 +-
 A2Mac/ViewController.swift                   | 40 +++++++++----
 src/cpu/6502.c                               | 39 +++++++------
 src/cpu/6502.h                               |  4 +-
 src/cpu/instructions/6502_instr_load_store.h | 60 +++++---------------
 src/cpu/instructions/6502_instr_stack.h      |  8 ++-
 src/dev/mem/mmio.h                           | 12 ++--
 src/util/common.h                            |  2 +
 9 files changed, 91 insertions(+), 90 deletions(-)

diff --git a/A2Mac.xcodeproj/project.pbxproj b/A2Mac.xcodeproj/project.pbxproj
index 0b81307..cac7417 100644
--- a/A2Mac.xcodeproj/project.pbxproj
+++ b/A2Mac.xcodeproj/project.pbxproj
@@ -1216,9 +1216,9 @@
 					"@executable_path/../Frameworks",
 				);
 				OTHER_CFLAGS = (
-					"-DDISASSEMBLER",
-					"-DINTERRUPT_CHECK_PER_STEP",
-					"-DCLK_ABSOLUTE_PRECISE",
+					"-D_NO_DISASSEMBLER",
+					"-D_NO_INTERRUPT_CHECK_PER_STEP",
+					"-D_NO_CLK_ABSOLUTE_PRECISE",
 				);
 				OTHER_SWIFT_FLAGS = "-DHIRES -DMETAL_NO -DHIRESDRAW -D_NO_HIRESLOW -DHIRESLOWCOLOR";
 				PRODUCT_BUNDLE_IDENTIFIER = com.gamealloy.A2Mac;
@@ -1252,9 +1252,9 @@
 					"@executable_path/../Frameworks",
 				);
 				OTHER_CFLAGS = (
-					"-DDISASSEMBLER",
-					"-DINTERRUPT_CHECK_PER_STEP",
-					"-DCLK_ABSOLUTE_PRECISE",
+					"-D_NO_DISASSEMBLER",
+					"-D_NO_INTERRUPT_CHECK_PER_STEP",
+					"-D_NO_CLK_ABSOLUTE_PRECISE",
 				);
 				OTHER_SWIFT_FLAGS = "-DHIRES -DMETAL_NO -DHIRESDRAW -D_NO_HIRESLOW -DHIRESLOWCOLOR";
 				PRODUCT_BUNDLE_IDENTIFIER = com.gamealloy.A2Mac;
diff --git a/A2Mac/Base.lproj/Main.storyboard b/A2Mac/Base.lproj/Main.storyboard
index 6b555a0..3ce6940 100644
--- a/A2Mac/Base.lproj/Main.storyboard
+++ b/A2Mac/Base.lproj/Main.storyboard
@@ -970,8 +970,8 @@
                                                 </connections>
                                             </button>
                                             <button verticalHuggingPriority="750" translatesAutoresizingMaskIntoConstraints="NO" id="umb-DW-vGc">
-                                                <rect key="frame" x="-1" y="-1" width="47" height="18"/>
-                                                <buttonCell key="cell" type="radio" title="666" bezelStyle="regularSquare" imagePosition="left" alignment="left" refusesFirstResponder="YES" inset="2" id="8VU-4B-7Pe">
+                                                <rect key="frame" x="-1" y="-1" width="51" height="18"/>
+                                                <buttonCell key="cell" type="radio" title="MAX" bezelStyle="regularSquare" imagePosition="left" alignment="left" refusesFirstResponder="YES" inset="2" id="8VU-4B-7Pe">
                                                     <behavior key="behavior" changeContents="YES" doesNotDimImage="YES" lightByContents="YES"/>
                                                     <font key="font" metaFont="system"/>
                                                 </buttonCell>
diff --git a/A2Mac/ViewController.swift b/A2Mac/ViewController.swift
index 86f8396..d6dea29 100644
--- a/A2Mac/ViewController.swift
+++ b/A2Mac/ViewController.swift
@@ -368,13 +368,24 @@ class ViewController: NSViewController  {
     var was = 0;
     
     var currentVideoMode = videoMode
+    var lastFrameTime = CACurrentMediaTime() as Double
+    var frameCounter : UInt = 0
+    var clkCounter : Double = 0
     
     func Update() {
+        clk_6502_per_frm_max = 0
+        
+        clkCounter += Double(clkfrm)
 
-        #if SPEEDTEST
-        #else
-        m6502_Run()
-        #endif
+        frameCounter += 1
+        
+        if ( frameCounter % UInt(fps) == 0 ) {
+            let currentTime = CACurrentMediaTime() as Double
+            let elpasedTime = currentTime - lastFrameTime
+            lastFrameTime = currentTime
+            mhz = Double( clkCounter ) / (elpasedTime * M);
+            clkCounter = 0
+        }
         
 //        render()
 //        hires.compute()
@@ -468,12 +479,15 @@ class ViewController: NSViewController  {
             self.display.stringValue = txt
 //            self.display.stringValue = "testing\nit\nout"
 
-            if ( (mhz < 10) && (mhz != floor(mhz)) ) {
+            if ( (mhz < 1.5) && (mhz != floor(mhz)) ) {
                 self.speedometer.stringValue = String(format: "%0.3lf MHz", mhz);
             }
             else {
-                self.speedometer.stringValue = String(format: "%.0lf MHz", mhz);
+                self.speedometer.stringValue = String(format: "%0.1lf MHz", mhz);
             }
+//            else {
+//                self.speedometer.stringValue = String(format: "%.0lf MHz", mhz);
+//            }
             
             #if HIRES
 //            if ( self.was < 300 ) {
@@ -492,6 +506,13 @@ class ViewController: NSViewController  {
 //            }
             #endif
         }
+        
+        #if SPEEDTEST
+        #else
+        m6502_Run()
+        #endif
+        
+        
     }
 
     
@@ -588,15 +609,14 @@ class ViewController: NSViewController  {
 
     
     func setCPUClockSpeed( freq : Double ) {
-        mhz = freq
-        MHz_6502 = UInt64(mhz * M)
+        MHz_6502 = UInt64(freq * M)
         clk_6502_per_frm = MHz_6502 / UInt64(fps)
         clk_6502_per_frm_set = clk_6502_per_frm
     }
 
     @IBAction func speedSelected(_ sender: NSButton) {
-        if ( sender.title == "Max" ) {
-            clk_6502_per_frm = UINT64_MAX
+        if ( sender.title == "MAX" ) {
+            setCPUClockSpeed(freq: 9999)
         }
         else if let freq = Double( sender.title ) {
             setCPUClockSpeed(freq: freq)
diff --git a/src/cpu/6502.c b/src/cpu/6502.c
index 4355562..440c4df 100644
--- a/src/cpu/6502.c
+++ b/src/cpu/6502.c
@@ -35,9 +35,11 @@ unsigned long long int inst_cnt = 0;
 
 //const unsigned int fps = 30;
 const unsigned long long default_MHz_6502 = 1.023 * M; // 2 * M; // 4 * M; // 8 * M; // 16 * M; // 128 * M; // 256 * M; // 512 * M;
+const unsigned long long startup_MHz_6502 = 25 * M;
 unsigned long long MHz_6502 = default_MHz_6502;
-unsigned long long clk_6502_per_frm =  25 * M; //default_MHz_6502 / fps;
+unsigned long long clk_6502_per_frm =  startup_MHz_6502 / fps;
 unsigned long long clk_6502_per_frm_set = default_MHz_6502 / fps;
+unsigned long long clk_6502_per_frm_max = default_MHz_6502 / fps;
 
 
 unsigned long long tick_per_sec = G;
@@ -547,35 +549,35 @@ INLINE int m6502_Step() {
         case 0x7E: ROR( dest_abs_X() ); return 7;                      // ROR abs,X
 //        case 0x7F:
 //        case 0x80:
-        case 0x81: STA( dest_X_ind() ) ; return 6;                     // STA X,ind
+        case 0x81: STA( addr_X_ind() ) ; return 6;                     // STA X,ind
 //        case 0x82:
 //        case 0x83:
-        case 0x84: STY( dest_zp() ); return 3;                         // STY zpg
-        case 0x85: STA( dest_zp() ); return 3;                         // STA zpg
-        case 0x86: STX( dest_zp() ); return 3;                         // STX zpg
+        case 0x84: STY( addr_zp() ); return 3;                         // STY zpg
+        case 0x85: STA( addr_zp() ); return 3;                         // STA zpg
+        case 0x86: STX( addr_zp() ); return 3;                         // STX zpg
 //        case 0x87:
         case 0x88: DEY(); return 2;                                    // DEY
 //        case 0x89: NOP(); imm(); return 4;                             // NOP imm
         case 0x8A: TXA(); return 2;                                    // TXA
 //        case 0x8B:
-        case 0x8C: STY( dest_abs() ); return 4;                        // STY abs
-        case 0x8D: STA( dest_abs() ); return 4;                        // STA abs
-        case 0x8E: STX( dest_abs() ); return 4;                        // STX abs
+        case 0x8C: STY( addr_abs() ); return 4;                        // STY abs
+        case 0x8D: STA( addr_abs() ); return 4;                        // STA abs
+        case 0x8E: STX( addr_abs() ); return 4;                        // STX abs
 //        case 0x8F:
         case 0x90: BCC( rel_addr() ); return 2;                        // BCC rel
-        case 0x91: STA( dest_ind_Y() ); return 6;                      // STA ind,Y
+        case 0x91: STA( addr_ind_Y() ); return 6;                      // STA ind,Y
 //        case 0x92:
 //        case 0x93:
-        case 0x94: STY( dest_zp_X() ); return 4;                       // STY zpg,X
-        case 0x95: STA( dest_zp_X() ); return 4;                       // STA zpg,X
-        case 0x96: STX( dest_zp_Y() ); return 4;                       // STX zpg,Y
+        case 0x94: STY( addr_zp_X() ); return 4;                       // STY zpg,X
+        case 0x95: STA( addr_zp_X() ); return 4;                       // STA zpg,X
+        case 0x96: STX( addr_zp_Y() ); return 4;                       // STX zpg,Y
 //        case 0x97:
         case 0x98: TYA(); return 2;                                    // TYA
-        case 0x99: STA( dest_abs_Y() ); return 5;                      // STA abs,Y
+        case 0x99: STA( addr_abs_Y() ); return 5;                      // STA abs,Y
         case 0x9A: TXS(); return 2;                                    // TXS
 //        case 0x9B:
 //        case 0x9C:
-        case 0x9D: STA( dest_abs_X() ); return 5;                      // STA abs,X
+        case 0x9D: STA( addr_abs_X() ); return 5;                      // STA abs,X
 //        case 0x9E:
 //        case 0x9F:
         case 0xA0: LDY( imm() ); return 2;                             // LDY imm
@@ -694,19 +696,24 @@ double mips = 0;
 double mhz = 0;
 unsigned long long epoch = 0;
 
+
+unsigned int clkfrm = 0;
+
 void m6502_Run() {
     static unsigned int clk = 0;
-    static unsigned int clkfrm = 0;
     
     // init time
 //#ifdef CLK_WAIT
 //    unsigned long long elpased = (unsigned long long)-1LL;
 //#endif
 
+    // we will also use this to pause the simulation if not finished by the end of the frame
+    clk_6502_per_frm_max = clk_6502_per_frm;
+    
 #ifdef SPEEDTEST
     for ( inst_cnt = 0; inst_cnt < iterations ; inst_cnt++ )
 #elif defined( CLK_WAIT )
-        for ( clkfrm = 0; clkfrm < clk_6502_per_frm ; clkfrm += clk )
+        for ( clkfrm = 0; clkfrm < clk_6502_per_frm_max ; clkfrm += clk )
 #else
 //    for ( ; m6502.pc ; )
     for ( ; ; )
diff --git a/src/cpu/6502.h b/src/cpu/6502.h
index 4a58139..cad0a43 100644
--- a/src/cpu/6502.h
+++ b/src/cpu/6502.h
@@ -16,6 +16,8 @@
 extern unsigned long long MHz_6502;
 extern unsigned long long clk_6502_per_frm;
 extern unsigned long long clk_6502_per_frm_set;
+extern unsigned long long clk_6502_per_frm_max;
+extern unsigned int clkfrm;
 
 
 typedef enum {
@@ -135,7 +137,7 @@ extern void hires_Update(void);
 extern double mips;
 extern double mhz;
 
-#define fps 30
+#define fps 30U
 
 extern void rom_loadFile( const char * bundlePath, const char * filename );
 extern void woz_loadFile( const char * bundlePath, const char * filename );
diff --git a/src/cpu/instructions/6502_instr_load_store.h b/src/cpu/instructions/6502_instr_load_store.h
index 53acbbc..9eb8a0c 100644
--- a/src/cpu/instructions/6502_instr_load_store.h
+++ b/src/cpu/instructions/6502_instr_load_store.h
@@ -90,51 +90,17 @@ char * charConv =
  
  (not a real instruction, only a helper function)
 **/
-INLINE void STR( uint8_t * dst, uint8_t src ) {
-    dbgPrintf("STR [%04X], %02X ", (int)(dst - RAM), src );
- 
-    uint16_t addr = dst - RAM;
+INLINE void STR( uint16_t addr, uint8_t src ) {
+    dbgPrintf("STR [%04X], %02X ", addr, src );
 
     // I/O or ROM or RAM EXP
-    if ( addr >= 0xC000 ) {
-        // ROM or RAM EXP
-        if ( addr >= 0xD000 ) {
-            uint8_t * pgaddr;
-            // DO NOT MAKE IT NICER! faster this way!
-            if ( (pgaddr = RAM_PG_WR_TBL[ addr >> 8 ]) ) {
-                dst = pgaddr + (addr & 0xFF);
-            }
-            else {
-                // NULL page table, so no need to do anything
-                return;
-            }
-        }
-        // I/O
-        else {
-    //        printf("mmio write:[%04X] = %02X\n", addr, src);
-    //        dst = Apple2_Dummy_Page;
-            
-            ioWrite( addr, src );
-            return;
-        }
+    if ( ( addr >= 0xC000 ) && ( addr < 0xC100 ) ) {
+        return ioWrite( addr, src );
     }
-//    else if ( ( addr >= 0x400 ) && ( addr < 0x800 ) ) {
-//        // Peripheral-card scratchpad RAM locations should be available on both memory
-//        if ( ( ( addr >= 0x478 ) && ( addr <= 0x47F ) )
-//          || ( ( addr >= 0x4F8 ) && ( addr <= 0x4FF ) )
-//          || ( ( addr >= 0x578 ) && ( addr <= 0x57F ) )
-//          || ( ( addr >= 0x5F8 ) && ( addr <= 0x5FF ) )
-//          || ( ( addr >= 0x678 ) && ( addr <= 0x67F ) )
-//          || ( ( addr >= 0x6F8 ) && ( addr <= 0x6FF ) )
-//          || ( ( addr >= 0x778 ) && ( addr <= 0x77F ) )
-//          || ( ( addr >= 0x7F8 ) && ( addr <= 0x7FF ) )
-//        ) {
-//            AUX_VID_RAM[ addr - 0x400 ] = src;
-//        }
-//    }
     
-    *dst = src;
-
+    // DO NOT MAKE IT NICER! faster this way!
+    *(RAM_PG_WR_TBL[ addr >> 8 ] + (addr & 0xFF)) = src;
+//    RAM[addr] = src;
 }
 
 /**
@@ -153,10 +119,10 @@ INLINE void STR( uint8_t * dst, uint8_t src ) {
  (indirect,X)  STA (oper,X)  81    2     6
  (indirect),Y  STA (oper),Y  91    2     6
 **/
-INLINE void STA( uint8_t * dst ) {
+INLINE void STA( uint16_t addr ) {
     dbgPrintf("STA ");
     disPrintf(disassembly.inst, "STA");
-    STR(dst, m6502.A);
+    STR(addr, m6502.A);
 }
 
 /**
@@ -171,10 +137,10 @@ INLINE void STA( uint8_t * dst ) {
  zeropage,Y    STX oper,Y    96    2     4
  absolute      STX oper      8E    3     4
  **/
-INLINE void STX( uint8_t * dst ) {
+INLINE void STX( uint16_t addr ) {
     dbgPrintf("STX ");
     disPrintf(disassembly.inst, "STX");
-    STR(dst, m6502.X);
+    STR(addr, m6502.X);
 }
 
 /**
@@ -189,10 +155,10 @@ INLINE void STX( uint8_t * dst ) {
  zeropage,X    STY oper,X    94    2     4
  absolute      STY oper      8C    3     4
  **/
-INLINE void STY( uint8_t * dst ) {
+INLINE void STY( uint16_t addr ) {
     dbgPrintf("STY ");
     disPrintf(disassembly.inst, "STY");
-    STR(dst, m6502.Y);
+    STR(addr, m6502.Y);
 }
 
 #endif // __6502_INSTR_LOAD_STORE_H__
diff --git a/src/cpu/instructions/6502_instr_stack.h b/src/cpu/instructions/6502_instr_stack.h
index 58bfb81..51fafc1 100644
--- a/src/cpu/instructions/6502_instr_stack.h
+++ b/src/cpu/instructions/6502_instr_stack.h
@@ -13,13 +13,17 @@ static const uint16_t stack_base_addr = 0x100;
 
 
 INLINE void PUSH( uint8_t src ) {
-    RAM[ stack_base_addr | m6502.SP-- ] = src;
+//    RAM[ stack_base_addr | m6502.SP-- ] = src;
+    // DO NOT MAKE IT NICER! faster this way!
+    *( RAM_PG_WR_TBL[ stack_base_addr >> 8 ] + m6502.SP-- ) = src;
 }
 
 INLINE uint8_t POP() {
-    return RAM[ stack_base_addr | ++m6502.SP ];
+//    return RAM[ stack_base_addr | ++m6502.SP ];
+    return *( RAM_PG_WR_TBL[ stack_base_addr >> 8 ] + ++m6502.SP );
 }
 
+
 INLINE void PUSH_addr( uint16_t addr ) {
     PUSH( (uint8_t)(addr >> 8) );
     PUSH( (uint8_t)addr );
diff --git a/src/dev/mem/mmio.h b/src/dev/mem/mmio.h
index 47bfd4d..1a17c63 100644
--- a/src/dev/mem/mmio.h
+++ b/src/dev/mem/mmio.h
@@ -437,7 +437,7 @@ void auxMemorySelect() {
 
 
 INLINE uint8_t ioRead( uint16_t addr ) {
-    if (outdev) fprintf(outdev, "ioRead:%04X\n", addr);
+//    if (outdev) fprintf(outdev, "ioRead:%04X\n", addr);
     
     uint8_t currentMagnet = 0;
     
@@ -677,7 +677,7 @@ INLINE uint8_t ioRead( uint16_t addr ) {
     return RAM[addr];
 }
 
-
+// TODO:
 void setIO ( uint16_t ioaddr, uint8_t val ) {
     RAM[ioaddr] = val;
 }
@@ -710,7 +710,7 @@ void kbdInput ( uint8_t code ) {
 
 
 INLINE void ioWrite( uint16_t addr, uint8_t val ) {
-    if (outdev) fprintf(outdev, "ioWrite:%04X (A:%02X)\n", addr, m6502.A);
+//    if (outdev) fprintf(outdev, "ioWrite:%04X (A:%02X)\n", addr, m6502.A);
     switch (addr) {
         case io_KBDSTRB:
             RAM[io_KBD] &= 0x7F;
@@ -834,14 +834,14 @@ INLINE void ioWrite( uint16_t addr, uint8_t val ) {
  **/
 INLINE uint8_t memread8( uint16_t addr ) {
     return * ( RAM_PG_RD_TBL[addr >> 8] + (addr & 0xFF) );
-//    return RAM[ addr ];
+//    return   RAM[addr];
 }
 /**
  Naive implementation of RAM read from address
  **/
 INLINE uint16_t memread16( uint16_t addr ) {
     return * (uint16_t*) ( RAM_PG_RD_TBL[addr >> 8] + (addr & 0xFF) );
-//    return * (uint16_t*) (& RAM[ addr ]);
+//    return * (uint16_t*) (RAM + addr);
 }
 
 INLINE uint8_t memread( uint16_t addr ) {
@@ -890,7 +890,7 @@ static  void memwrite( uint16_t addr, uint8_t byte ) {
  increase pc by one
  **/
 INLINE uint8_t fetch() {
-    disHexB( disassembly.pOpcode, RAM[m6502.PC] );
+    disHexB( disassembly.pOpcode, memread(m6502.PC) );
 #ifdef CLK_ABSOLUTE_PRECISE
     if ( (m6502.PC & 0xFF) >= 0xFF ) {
         m6502.clktime++;
diff --git a/src/util/common.h b/src/util/common.h
index fcb81b1..e7efebe 100644
--- a/src/util/common.h
+++ b/src/util/common.h
@@ -59,6 +59,8 @@ union {
 
 
 #define BITTEST(n,x) ((bits_t)(n)).b##x
+//#define BITTEST(n,x) ( (n) >> (x) )
+//#define BITTEST(n,x) ( (n) & (1 << (x)) )
 
 
 extern FILE * outdev;