OPtimize signed+unsigned char arithmetic

prepare output directory also to build only the compiler

CMakeLists.txt

@@ -0,0 +1,67 @@
+cmake_minimum_required(VERSION 3.15)
+
+# Project name and version
+project(oscar64 VERSION 1.0 LANGUAGES CXX)
+
+# Define the executable
+add_executable(oscar64)
+
+
+# Glob all source files in the oscar64/ directory
+file(GLOB OSCAR64_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/oscar64/*.cpp" "${CMAKE_CURRENT_SOURCE_DIR}/oscar64/*.h")
+
+# Add the sources to the target
+target_sources(${PROJECT_NAME} PRIVATE ${OSCAR64_SOURCES})
+
+
+# Add header files
+target_include_directories(oscar64 PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}
+)
+
+# Compiler settings based on configuration
+set_target_properties(oscar64 PROPERTIES
+    CXX_STANDARD 17
+    CXX_STANDARD_REQUIRED YES
+    CXX_EXTENSIONS NO
+)
+
+# Set debug and release configurations
+target_compile_definitions(oscar64 PRIVATE $<$<CONFIG:Debug>:_DEBUG;_CONSOLE> $<$<CONFIG:Release>:NDEBUG;_CONSOLE>)
+target_compile_options(oscar64 PRIVATE $<$<CONFIG:Debug>:/W3 /WX> $<$<CONFIG:Release>:/O2 /W3>)
+
+# Post-build steps (only for Release builds)
+if(CMAKE_BUILD_TYPE STREQUAL "Release")
+    add_custom_command(TARGET oscar64 POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy $<TARGET_FILE:oscar64> ${CMAKE_BINARY_DIR}/bin
+    )
+endif()
+
+if (WIN32)
+    target_link_libraries(oscar64 PRIVATE kernel32 user32 gdi32 winspool comdlg32 advapi32 shell32 ole32 oleaut32 uuid odbc32 odbccp32 version)
+elseif (APPLE)
+    # Add macOS-specific frameworks to replace Windows libraries
+    target_link_libraries(${PROJECT_NAME} 
+        "-framework Cocoa"        # For GUI and application support
+        "-framework CoreFoundation" # For Core Foundation utilities (similar to advapi32 or shell32)
+        "-framework CoreGraphics"  # For graphics and rendering (similar to gdi32)
+        "-framework IOKit"         # For hardware interaction
+        "-framework AppKit"        # For GUI (equivalent to user32)
+    )
+elseif (UNIX)
+    # Add Linux-specific libraries for equivalent functionality
+    target_link_libraries(${PROJECT_NAME} 
+        pthread                    # For threading (similar to kernel32)
+        dl                         # For dynamic library loading (similar to LoadLibrary)
+        m                          # For math library (optional but common on Linux)
+        X11                        # For X Window System (similar to user32 for GUI support)
+    )
+endif()
+
+
+# Output directories
+set_target_properties(oscar64 PROPERTIES
+    RUNTIME_OUTPUT_DIRECTORY_DEBUG ${CMAKE_BINARY_DIR}/Debug
+    RUNTIME_OUTPUT_DIRECTORY_RELEASE ${CMAKE_BINARY_DIR}/Release
+)
+

README.md

@@ -4,7 +4,7 @@ Oscar64 is a C/C++ cross compiler running on a modern system (such as a Windows
 
 The purpose of this compiler is to eliminate the need to write 6502 assembler code to achieve high code density and fast execution speed.  It continues to improve with all the games, demos and tools written by it.  It supports disk overlays and banked cartridges for larger projects.
 
-The C64 executes 418 dhrystone V2.2 iteration per second, when compiled with Oscar64 and -O3 (which shows that the ancient dhrystone benchmark is no match to an optimizing compiler).
+The C64 executes 442 dhrystone V2.2 iteration per second, when compiled with Oscar64 and -O3 (which shows that the ancient dhrystone benchmark is no match to an optimizing compiler).
 
 [Full reference manual](oscar64.md)
 [Additional samples and tutorials](https://github.com/drmortalwombat/OscarTutorials)

autotest/autotest.bat

@@ -1,5 +1,8 @@
 rem @echo off
 
+@call :test rolrortest.cpp
+@if %errorlevel% neq 0 goto :error
+
 @call :test bitfields.cpp
 @if %errorlevel% neq 0 goto :error
 
@@ -15,6 +18,9 @@ rem @echo off
 @call :testh opp_vector.cpp
 @if %errorlevel% neq 0 goto :error
 
+@call :testh opp_static_vector.cpp
+@if %errorlevel% neq 0 goto :error
+
 @call :testh opp_vector_string.cpp
 @if %errorlevel% neq 0 goto :error
 
@@ -33,7 +39,7 @@ rem @echo off
 @call :testh opp_list.cpp
 @if %errorlevel% neq 0 goto :error
 
-@call :testh opp_functional.cpp
+@call :testn opp_functional.cpp
 @if %errorlevel% neq 0 goto :error
 
 @call :testh operatoroverload.cpp
@@ -51,7 +57,7 @@ rem @echo off
 @call :testh constructortest.cpp
 @if %errorlevel% neq 0 goto :error
 
-@call :testh copyconstructor.cpp
+@call :testn copyconstructor.cpp
 @if %errorlevel% neq 0 goto :error
 
 @call :testh copyassign.cpp
@@ -274,6 +280,15 @@ exit /b %errorlevel%
 ..\bin\oscar64 -e -O2 -n -dHEAPCHECK %~1
 @if %errorlevel% neq 0 goto :error
 
+..\bin\oscar64 -e -O2 -xz -Oz -n %~1
+@if %errorlevel% neq 0 goto :error
+
+..\bin\oscar64 -e -O2 -Oo -n %~1
+@if %errorlevel% neq 0 goto :error
+
+..\bin\oscar64 -e -O2 -Ox -n %~1
+@if %errorlevel% neq 0 goto :error
+
 ..\bin\oscar64 -e -O0 -bc %~1
 @if %errorlevel% neq 0 goto :error
 
@@ -331,6 +346,12 @@ exit /b %errorlevel%
 ..\bin\oscar64 -e -O2 -xz -Oz -n %~1
 @if %errorlevel% neq 0 goto :error
 
+..\bin\oscar64 -e -O2 -Oo -n %~1
+@if %errorlevel% neq 0 goto :error
+
+..\bin\oscar64 -e -O2 -Ox -n %~1
+@if %errorlevel% neq 0 goto :error
+
 @exit /b 0
 
 :testb
@@ -370,4 +391,10 @@ exit /b %errorlevel%
 ..\bin\oscar64 -e -O2 -xz -Oz -n %~1
 @if %errorlevel% neq 0 goto :error
 
+..\bin\oscar64 -e -O2 -Oo -n %~1
+@if %errorlevel% neq 0 goto :error
+
+..\bin\oscar64 -e -O2 -Ox -n %~1
+@if %errorlevel% neq 0 goto :error
+
 @exit /b 0

autotest/copyconstructor.cpp

@@ -1,4 +1,5 @@
 #include <assert.h>
+#include <stdio.h>
 
 int t, n;
 
@@ -161,7 +162,7 @@ void test_return_value(void)
 		C2 c(test_ret_v());
 	}
 
-	assert(n == 4 && t == 0);
+	assert(n == 6 && t == 0);
 }
 
 void test_return_reference(void)
@@ -184,7 +185,7 @@ void test_retparam_value(void)
 		test_param_fv(test_ret_v());
 	}
 
-	assert(n == 4 && t == 0);	
+	assert(n == 6 && t == 0);	
 }
 
 void test_retparam_reference(void)
@@ -200,7 +201,6 @@ void test_retparam_reference(void)
 
 int main(void)
 {
-#if 0
 	test_dcopy_init();
 	test_copy_init();
 	test_minit();
@@ -208,9 +208,8 @@ int main(void)
 	test_param_value();
 	test_param_ref();
 	test_return_value();
-#endif
 	test_retparam_value();
-//	test_retparam_reference();
+	test_retparam_reference();
 
 	return 0;
 }

autotest/makefile

@@ -50,5 +50,11 @@ autorefreturn: autorefreturn.cpp
 	$(OSCAR64_CC) -e -Os -n $<
 	$(OSCAR64_CC) -e -O3 -n $<
 
+copyconstructor: copyconstructor.cpp
+	$(OSCAR64_CC) -e -O2 -n $<
+	$(OSCAR64_CC) -e -O0 -n $<
+	$(OSCAR64_CC) -e -Os -n $<
+	$(OSCAR64_CC) -e -O3 -n $<
+
 clean:
 	@$(RM) *.asm *.bcs *.int *.lbl *.map *.prg

autotest/opp_static_vector.cpp

@@ -0,0 +1,78 @@
+#include <opp/static_vector.h>
+#include <opp/algorithm.h>
+#include <assert.h>
+#include <opp/iostream.h>
+
+int main(void)
+{
+	opp::static_vector<int, 20>		a;
+
+	for(int i=0; i<10; i++)
+		a.push_back(i);
+
+	int s = 0;
+	for(int i=0; i<a.size(); i++)
+		s += a[i];
+
+	assert(s == 45);
+
+	for(int i=0; i<5; i++)
+		a.erase(i);
+
+	s = 0;
+	for(int i=0; i<a.size(); i++)
+		s += a[i];
+
+	assert(s == 1 + 3 + 5 + 7 + 9);
+
+	opp::static_vector<int, 100>	v;
+
+	for(int i=0; i<10; i++)
+		v.push_back(i);
+
+	assert(v.size() == 10);
+	v.insert(0, 20);
+	assert(v.size() == 11);
+	v.insert(6, 21);
+	assert(v.size() == 12);
+	v.insert(12, 22);
+
+	int * fi = opp::find(v.begin(), v.end(), 21);
+
+	fi = v.insert(fi, 30);
+	fi = v.insert(fi, 31);
+	fi = v.insert(fi, 32);
+
+	assert(v.size() == 16);
+	assert(v[0] == 20);
+	assert(v[15] == 22);
+	assert(v[8] == 32);
+
+	fi = opp::find(v.begin(), v.end(), 32);
+
+	for(int i=0; i<30; i++)
+	{
+		fi = v.insert(fi, i + 40);
+	}
+
+	assert(v.size() == 46);
+	assert(v[28] == 60);
+
+	v.erase(10, 10);
+
+	for(int i : v) 
+		opp::cout << i << ", ";
+	opp::cout << "\n";
+
+	assert(v.size() == 36);
+	assert(v[18] == 60);
+
+	v.assign(42, 11);
+
+	assert(v.size() == 42);
+	assert(v[0] == 11);
+	assert(v[15] == 11);
+	assert(v[41] == 11);
+
+	return 0;
+}

autotest/opp_vector.cpp

@@ -64,5 +64,15 @@ int main(void)
 		opp::cout << i << ", ";
 	opp::cout << "\n";
 
+	assert(v.size() == 36);
+	assert(v[18] == 60);
+
+	v.assign(42, 11);
+
+	assert(v.size() == 42);
+	assert(v[0] == 11);
+	assert(v[15] == 11);
+	assert(v[41] == 11);
+
 	return 0;
 }

autotest/rolrortest.cpp

@@ -0,0 +1,76 @@
+#include <assert.h>
+
+__noinline unsigned long rollright32(unsigned long a) {
+    unsigned long tmp = a & 1;
+    return ( a >> 1) + (tmp << 31);
+}
+
+__noinline unsigned rollright16(unsigned a) {
+    unsigned tmp = a & 1;
+    return ( a >> 1) + (tmp << 15);
+}
+
+__noinline char rollright8(char a) {
+    char tmp = a & 1;
+    return ( a >> 1) + (tmp << 7); 
+}
+
+__noinline unsigned long rollleft32(unsigned long a) {
+    unsigned long tmp = (a >> 31) & 1;
+    return ( a << 1) + tmp; 
+}
+
+__noinline unsigned rollleft16(unsigned a) {
+    unsigned tmp = (a >> 15) & 1;
+    return ( a << 1) + tmp; 
+}
+
+__noinline char rollleft8(char a) {
+    char tmp = (a >> 7) & 1;
+    return ( a << 1) + tmp; 
+}
+
+int main() {
+    unsigned long lv = 0x12345678ul;
+    unsigned val = 0x1234;
+    char c=0x12;
+
+    unsigned long lvt[33];
+    unsigned valt[17];
+    char ct[9];
+
+    lvt[0] = lv;
+    valt[0] = val;
+    ct[0] = c;  
+
+    assert(rollleft8(rollright8(c)) == c);
+    assert(rollleft16(rollright16(val)) == val);
+    assert(rollleft32(rollright32(lv)) == lv);
+     
+    for(int i=0; i<32; i++)
+        lvt[i + 1] = rollright32(lvt[i]);
+    for(int i=0; i<16; i++)
+        valt[i + 1] = rollright16(valt[i]);
+    for(int i=0; i<8; i++)
+        ct[i + 1] = rollright8(ct[i]);
+
+    for(int i=0; i<=32; i++)
+    {        
+        assert(lvt[32 - i] == lv);
+        lv = rollleft32(lv);
+    }
+
+    for(int i=0; i<=16; i++)
+    {        
+        assert(valt[16 - i] == val);
+        val = rollleft16(val);
+    }
+
+    for(int i=0; i<=8; i++)
+    {        
+        assert(ct[8 - i] == c);
+        c = rollleft8(c);
+    }
+
+    return 0;
+}

autotest/testint32cmp.c

@@ -63,6 +63,33 @@ bool nge(long a, long b)
 
 
 
+
+inline bool ieq(long a, long b)
+{
+	return a == b;
+}
+
+inline bool ilt(long a, long b)
+{
+	return a < b;
+}
+
+inline bool igt(long a, long b)
+{
+	return a > b;
+}
+
+inline bool ile(long a, long b)
+{
+	return a <= b;
+}
+
+inline bool ige(long a, long b)
+{
+	return a >= b;
+}
+
+
 bool beqz(long a)
 {
 	return a == 0;
@@ -188,7 +215,71 @@ bool nge1(long a)
 
 
 
-void cmp(long a, long b)
+bool beqm(long a)
+{
+	return a == -1;
+}
+
+bool bltm(long a)
+{
+	return a < -1;
+}
+
+bool bgtm(long a)
+{
+	return a > -1;
+}
+
+bool blem(long a)
+{
+	return a <= -1;
+}
+
+bool bgem(long a)
+{
+	return a >= -1;
+}
+
+
+
+bool neqm(long a)
+{
+	return a == -1;
+}
+
+#pragma native(neqm)
+
+bool nltm(long a)
+{
+	return a < -1;
+}
+
+#pragma native(nltm)
+
+bool ngtm(long a)
+{
+	return a > -1;
+}
+
+#pragma native(ngtm)
+
+bool nlem(long a)
+{
+	return a <= -1;
+}
+
+#pragma native(nlem)
+
+bool ngem(long a)
+{
+	return a >= -1;
+}
+
+#pragma native(ngem)
+
+
+
+void cmpc(long a, long b)
 {
 	bool	beqf = beq(a, b), bltf = blt(a, b), bgtf = bgt(a, b), blef = ble(a, b), bgef = bge(a, b);
 	bool	neqf = neq(a, b), nltf = nlt(a, b), ngtf = ngt(a, b), nlef = nle(a, b), ngef = nge(a, b);
@@ -203,6 +294,27 @@ void cmp(long a, long b)
 	assert(bgef == ngef);	
 }
 
+void cmpi(long a, long b)
+{
+	bool	ieqf = ieq(a, b), iltf = ilt(a, b), igtf = igt(a, b), ilef = ile(a, b), igef = ige(a, b);
+	bool	neqf = neq(a, b), nltf = nlt(a, b), ngtf = ngt(a, b), nlef = nle(a, b), ngef = nge(a, b);
+	
+	printf("INLINE %ld, %ld : EQ %d LT %d GT %d\r", a, b, ieqf, iltf, igtf);
+	printf("NATIVE %ld, %ld : EQ %d LT %d GT %d\r", a, b, neqf, nltf, ngtf);
+	
+	assert(ieqf == neqf);
+	assert(iltf == nltf);
+	assert(igtf == ngtf);	
+	assert(ilef == nlef);
+	assert(igef == ngef);	
+}
+
+void cmp(long a, long b)
+{
+	cmpc(a, b);
+	cmpi(a, b);
+}
+
 void cmpz(long a)
 {
 	bool	beqf = beqz(a), bltf = bltz(a), bgtf = bgtz(a), blef = blez(a), bgef = bgez(a);
@@ -233,6 +345,21 @@ void cmp1(long a)
 	assert(bgef == ngef);	
 }
 
+void cmpm(long a)
+{
+	bool	beqf = beqm(a), bltf = bltm(a), bgtf = bgtm(a), blef = blem(a), bgef = bgem(a);
+	bool	neqf = neqm(a), nltf = nltm(a), ngtf = ngtm(a), nlef = nlem(a), ngef = ngem(a);
+	
+	printf("BYTE   %ld, 1 : EQ %d LT %d GT %d LE %d GE %d\r", a, beqf, bltf, bgtf, blef, bgef);
+	printf("NATIVE %ld, 1 : EQ %d LT %d GT %d LE %d GE %d\r", a, neqf, nltf, ngtf, nlef, ngef);
+	
+	assert(beqf == neqf);
+	assert(bltf == nltf);
+	assert(bgtf == ngtf);	
+	assert(blef == nlef);
+	assert(bgef == ngef);	
+}
+
 int main(void)
 {
 	cmp( 0,  1);
@@ -327,6 +454,10 @@ int main(void)
 	cmp1(256);
 	cmp1(10000);
 	cmp1(20000);
+	cmp1(1000000l);
+	cmp1(2000000l);
+	cmp1(100000000l);
+	cmp1(200000000l);
 	cmp1(-1);
 	cmp1(-2);
 	cmp1(-3);
@@ -334,6 +465,34 @@ int main(void)
 	cmp1(-256);
 	cmp1(-10000);
 	cmp1(-20000);
+	cmp1(-1000000l);
+	cmp1(-2000000l);
+	cmp1(-100000000l);
+	cmp1(-200000000l);
+
+	cmpm(0);
+	cmpm(1);
+	cmpm(2);
+	cmpm(3);
+	cmpm(255);
+	cmpm(256);
+	cmpm(10000);
+	cmpm(20000);
+	cmpm(1000000l);
+	cmpm(2000000l);
+	cmpm(100000000l);
+	cmpm(200000000l);
+	cmpm(-1);
+	cmpm(-2);
+	cmpm(-3);
+	cmpm(-255);
+	cmpm(-256);
+	cmpm(-10000);
+	cmpm(-20000);
+	cmpm(-1000000l);
+	cmpm(-2000000l);
+	cmpm(-100000000l);
+	cmpm(-200000000l);
 
 	return 0;
 	

include/audio/sidfx.c

@@ -10,7 +10,7 @@ enum SIDFXState
 	SIDFX_WAIT
 };
 
-static struct SIDFXChannel
+__striped static struct SIDFXChannel
 {
 	const SIDFX	* volatile 		com;
 	byte						delay, priority;
@@ -27,6 +27,7 @@ void sidfx_init(void)
 		channels[i].com = nullptr;
 		channels[i].state = SIDFX_IDLE;
 		channels[i].priority = 0;
+		channels[i].delay = 1;
 	}
 }
 
@@ -35,6 +36,11 @@ bool sidfx_idle(byte chn)
 	return channels[chn].state == SIDFX_IDLE;
 }
 
+char sidfx_cnt(byte chn)
+{
+	return channels[chn].cnt;	
+}
+
 void sidfx_play(byte chn, const SIDFX * fx, byte cnt)
 {
 	SIDFXState		ns = channels[chn].state;
@@ -47,6 +53,7 @@ void sidfx_play(byte chn, const SIDFX * fx, byte cnt)
 		return;
 
 	channels[chn].state = SIDFX_IDLE;
+	channels[chn].delay = 1;
 
 	channels[chn].com = fx;
 	channels[chn].cnt = cnt - 1;
@@ -59,30 +66,58 @@ void sidfx_stop(byte chn)
 {
 	channels[chn].com = nullptr;
 	if (channels[chn].state != SIDFX_IDLE)
+	{
 		channels[chn].state = SIDFX_RESET_0;
+		channels[chn].delay = 1;
+	}
 }
 
 inline void sidfx_loop_ch(byte ch)
 {
+	if (channels[ch].state)
+	{
+		const SIDFX	*	com = channels[ch].com;
+
+		channels[ch].delay--;
+		if (channels[ch].delay)
+		{
+			if (com->dfreq)
+			{
+				channels[ch].freq += com->dfreq;
+				sid.voices[ch].freq = channels[ch].freq;
+			}
+			if (com->dpwm)
+			{
+				channels[ch].pwm += com->dpwm;
+				sid.voices[ch].pwm = channels[ch].pwm;
+			}
+		}
+
+		while (!channels[ch].delay)
+		{
 			switch (channels[ch].state)
 			{
 			case SIDFX_IDLE:
+				channels[ch].delay = 1;
 				break;
 			case SIDFX_RESET_0:
 				sid.voices[ch].ctrl = 0;
 				sid.voices[ch].attdec = 0;
 				sid.voices[ch].susrel = 0;
+				if (com)
 					channels[ch].state = SIDFX_READY;
+				else
+					channels[ch].state = SIDFX_IDLE;
+				channels[ch].delay = 1;
 				break;
 			case SIDFX_RESET_1:
 				sid.voices[ch].ctrl = SID_CTRL_TEST;
+				sid.voices[ch].ctrl = 0;
+				sid.voices[ch].attdec = 0;
+				sid.voices[ch].susrel = 0;
 				channels[ch].state = SIDFX_READY;
 				break;
 			case SIDFX_READY:
-			{
-				const SIDFX	*	com = channels[ch].com;
-				if (com)
-				{
 				channels[ch].freq = com->freq;
 				channels[ch].pwm = com->pwm;
 
@@ -92,38 +127,34 @@ inline void sidfx_loop_ch(byte ch)
 				sid.voices[ch].susrel = com->susrel;
 				sid.voices[ch].ctrl = com->ctrl;
 
+				if (com->ctrl & SID_CTRL_GATE)
+				{
 					channels[ch].delay = com->time1;
 					channels[ch].state = SIDFX_PLAY;
 				}
 				else
-					channels[ch].state = SIDFX_IDLE;
+				{
+					channels[ch].delay = com->time0;
+					channels[ch].state = SIDFX_PLAY;
 				}
 				break;
 			case SIDFX_PLAY:
-			{
+				if (com->time0)
-				const SIDFX	*	com = channels[ch].com;
-				if (com->dfreq)
-				{
-					channels[ch].freq += com->dfreq;
-					sid.voices[ch].freq = channels[ch].freq;
-				}
-				if (com->dpwm)
-				{
-					channels[ch].pwm += com->dpwm;
-					sid.voices[ch].pwm = channels[ch].pwm;
-				}
-
-				if (channels[ch].delay)
-					channels[ch].delay--;
-				else if (com->time0)
 				{
 					sid.voices[ch].ctrl = com->ctrl & ~SID_CTRL_GATE;
-					channels[ch].delay = com->time0;
+					channels[ch].delay = com->time0 - 1;
 					channels[ch].state = SIDFX_WAIT;
 				}
 				else if (channels[ch].cnt)
 				{
+					char sr = com->susrel & 0xf0;
 					com++;
+					char ctrl = com->ctrl;
+					if ((com->attdec & 0xef) == 0 && (ctrl & SID_CTRL_GATE) && (com->susrel & 0xf0) > sr)
+					{
+						sid.voices[ch].ctrl = ctrl & ~SID_CTRL_GATE;
+						sid.voices[ch].ctrl = ctrl |  SID_CTRL_GATE;
+					}
 					channels[ch].cnt--;
 					channels[ch].com = com;
 					channels[ch].priority = com->priority;
@@ -131,46 +162,31 @@ inline void sidfx_loop_ch(byte ch)
 				}
 				else
 				{
-					channels[ch].com = nullptr;
+					com = nullptr;
 					channels[ch].state = SIDFX_RESET_0;					
 				}
-			}
 				break;
 			case SIDFX_WAIT:
-			{
+				if (channels[ch].cnt)
-				const SIDFX	*	com = channels[ch].com;			
-				if (com->dfreq)
-				{
-					channels[ch].freq += com->dfreq;
-					sid.voices[ch].freq = channels[ch].freq;
-				}
-				if (com->dpwm)
-				{
-					channels[ch].pwm += com->dpwm;
-					sid.voices[ch].pwm = channels[ch].pwm;
-				}
-
-				if (channels[ch].delay)
-					channels[ch].delay--;
-				else if (channels[ch].cnt)
 				{
 					com++;
 					channels[ch].cnt--;
 					channels[ch].com = com;
 					channels[ch].priority = com->priority;
-					if (com->time0)
+					if (com->ctrl & SID_CTRL_GATE)
 						channels[ch].state = SIDFX_RESET_0;
 					else
 						channels[ch].state = SIDFX_READY;
 				}
 				else
 				{
-					channels[ch].com = nullptr;
+					com = nullptr;
 					channels[ch].state = SIDFX_RESET_0;					
 				}
-			}
 				break;
 			}
+		}
+	}
 }
 
 void sidfx_loop_2(void)

include/audio/sidfx.h

@@ -20,6 +20,8 @@ inline void sidfx_play(byte chn, const SIDFX * fx, byte cnt);
 
 void sidfx_stop(byte chn);
 
+char sidfx_cnt(byte chn);
+
 void sidfx_loop(void);
 
 void sidfx_loop_2(void);

include/c64/kernalio.c

@@ -69,7 +69,7 @@ BANKINLINE bool krnio_open(char fnum, char device, char channel)
 {
 	krnio_pstatus[fnum] = KRNIO_OK;
 
-	return __asm
+	return char(__asm
 	{
 		lda	#0
 		sta accu
@@ -94,8 +94,7 @@ BANKINLINE bool krnio_open(char fnum, char device, char channel)
 
 		BANKOUT
 	E2:
-	};
+	});
-
 }
 
 #pragma native(krnio_open)
@@ -131,7 +130,7 @@ BANKINLINE krnioerr krnio_status(void)
 
 BANKINLINE bool krnio_load(char fnum, char device, char channel)
 {
-	return __asm
+	return char(__asm
 	{
 		BANKIN
 		lda	fnum
@@ -149,14 +148,14 @@ BANKINLINE bool krnio_load(char fnum, char device, char channel)
 		rol
 		eor #1
 		sta accu
-	};
+	});
 }
 
 #pragma native(krnio_load)
 
 BANKINLINE bool krnio_save(char device, const char* start, const char* end)
 {
-	return __asm
+	return char(__asm
 	{
 		BANKIN
 		lda	#0
@@ -175,14 +174,14 @@ BANKINLINE bool krnio_save(char device, const char* start, const char* end)
 		rol
 		eor #1
 		sta accu
-	};
+	});
 }
 
 #pragma native(krnio_save)
 
 BANKINLINE bool krnio_chkout(char fnum)
 {
-	return __asm
+	return char(__asm
 	{
 		BANKIN
 		ldx fnum
@@ -193,14 +192,14 @@ BANKINLINE bool krnio_chkout(char fnum)
 		rol
 		eor #1
 		sta accu
-	};
+	});
 }
 
 #pragma native(krnio_chkout)
 
 BANKINLINE bool krnio_chkin(char fnum)
 {
-	return __asm
+	return char(__asm
 	{
 		BANKIN
 		ldx fnum
@@ -211,7 +210,7 @@ BANKINLINE bool krnio_chkin(char fnum)
 		rol
 		eor #1
 		sta accu
-	};
+	});
 }
 
 #pragma native(krnio_chkin)
@@ -230,14 +229,14 @@ BANKINLINE void krnio_clrchn(void)
 
 BANKINLINE bool krnio_chrout(char ch)
 {
-	return __asm
+	return char(__asm
 	{
 		BANKIN
 		lda ch
 		jsr $ffd2		// chrout
 		sta accu
 		BANKOUT
-	};
+	});
 }
 
 #pragma native(krnio_chrout)
@@ -422,7 +421,7 @@ int krnio_gets(char fnum, char * data, int num)
 
 	if (krnio_chkin(fnum))
 	{
-		krnioerr err;
+		krnioerr err = KRNIO_OK;
 		int i = 0;
 		int ch;
 		while (i + 1 < num)

include/c64/rasterirq.c

@@ -5,16 +5,16 @@
 #include <c64/asm6502.h>
 #include <stdlib.h>
 
-volatile char npos = 1, tpos = 0;
 volatile byte rirq_count;
+static byte rirq_pcount;
 
 byte		rasterIRQRows[NUM_IRQS + 1];	
-byte		rasterIRQIndex[NUM_IRQS + 1];
+byte		rasterIRQIndex[NUM_IRQS + 1];	// Sort order of interrupt index, offset by one
 #ifdef ZPAGE_IRQS
 __zeropage
 #endif
-byte		rasterIRQNext[NUM_IRQS + 1];
+byte		rasterIRQNext[NUM_IRQS + 1];	// Rasterline of interrupt, terminated by 0xff
-byte		rasterIRQLow[NUM_IRQS];
+byte		rasterIRQLow[NUM_IRQS];			// Address of interrupt code
 byte		rasterIRQHigh[NUM_IRQS];
 
 #ifdef ZPAGE_IRQS
@@ -22,28 +22,25 @@ __zeropage
 #endif
 byte		nextIRQ;
 
-__asm irq0
+// nextIRQ is the index of the next expected IRQ, or $ff if no IRQ is scheduled
+
+__asm rirq_isr_ram_io
 {	
-    pha
+	stx plrx + 1
-    txa
-    pha
-    tya
-    pha
-kentry:
-	asl $d019
 
 	ldx	nextIRQ
+	bmi	exi
+
+	sta plra + 1
+	sty plry + 1
+
 l1:	
 	lda	rasterIRQNext, x
-	cmp	#$ff
-	beq	e1
-
 	ldy rasterIRQIndex + 1, x
-	tax
+	ldx	rasterIRQLow, y
-	lda	rasterIRQLow, y
+	stx ji + 1
-	sta ji + 1
+	ldx	rasterIRQHigh, y
-	lda	rasterIRQHigh, y
+	stx ji + 2
-	sta ji + 2
 
 ji:	
 	jsr $0000
@@ -51,21 +48,82 @@ ji:
 	inc	nextIRQ
 	ldx nextIRQ
 
-	lda	rasterIRQNext, x
+	ldy	rasterIRQNext, x
-	cmp #$ff
-	beq e2
-	// carry is cleared at this point
 
-	tay
+	asl $d019
-	sbc #2
+
-	cmp $d012
+	cpy #$ff
+	beq e2
+
+	dey
+	sty	$d012
+	dey
+	cpy $d012
+	bcc l1
+
+plry:
+	ldy #0
+plra:
+	lda #0
+plrx:
+	ldx #0
+    rti
+
+exi:
+	asl $d019
+	jmp plrx
+
+    // No more interrupts to service
+e2:
+	inc rirq_count
+
+	ldy	rasterIRQNext
+	dey
+	sty	$d012
+	ldx	#0
+	stx nextIRQ
+	beq	plry
+}
+
+__asm rirq_isr_io
+{	
+    pha
+    txa
+    pha
+    tya
+    pha
+kentry:
+
+	ldx	nextIRQ
+	bmi	exi
+l1:
+	lda	rasterIRQNext, x
+	ldy rasterIRQIndex + 1, x
+	ldx	rasterIRQLow, y
+	stx ji + 1
+	ldx	rasterIRQHigh, y
+	stx ji + 2
+
+ji:	
+	jsr $0000
+
+	inc	nextIRQ
+	ldx nextIRQ
+
+	ldy	rasterIRQNext, x
+
+	asl $d019
+
+	cpy #$ff
+	beq e2
+
+	dey
+	sty	$d012
+	dey
+	cpy $d012
 	bcc l1
 
 exd:
-	dey
-ex:
-	sty	$d012
-
     pla
     tay
     pla
@@ -73,25 +131,22 @@ ex:
     pla
     rti
 
-e2:
+exi:
-
+	asl $d019
-	ldx npos
-	stx tpos
-	inc rirq_count
-	tay
-
-	bit	$d011
-	bpl ex
-
-e1:
-	ldx	#0
-	stx nextIRQ
-	ldy	rasterIRQNext
 	jmp exd
 
+e2:
+	inc rirq_count
+
+	ldy	rasterIRQNext
+	dey
+	sty	$d012
+	ldx	#0
+	stx nextIRQ
+	beq	exd
 }
 
-__asm irq2
+__asm rirq_isr_noio
 {	
     pha
     txa
@@ -105,20 +160,15 @@ kentry:
     lda #$35
     sta $01
 
-	asl $d019
-
 	ldx	nextIRQ
+	bmi	exi
 l1:
 	lda	rasterIRQNext, x
-	cmp	#$ff
-	beq	e1
-
 	ldy rasterIRQIndex + 1, x
-	tax
+	ldx	rasterIRQLow, y
-	lda	rasterIRQLow, y
+	stx ji + 1
-	sta ji + 1
+	ldx	rasterIRQHigh, y
-	lda	rasterIRQHigh, y
+	stx ji + 2
-	sta ji + 2
 
 ji:	
 	jsr $0000
@@ -126,21 +176,20 @@ ji:
 	inc	nextIRQ
 	ldx nextIRQ
 
-	lda	rasterIRQNext, x
+	ldy	rasterIRQNext, x
-	cmp #$ff
+
-	beq e2
+	asl $d019
-	// carry is cleared at this point
+
+	cpy #$ff
+	beq e2
 
-	tay
 	dey
-	sbc #2
+	sty	$d012
-	cmp $d012
+	dey
+	cpy $d012
 	bcc l1
 
-	sty	$d012
+exd:
-
-ex:
-
 	pla
 	sta $01
 
@@ -151,45 +200,35 @@ ex:
     pla
     rti
 
-e2:
+exi:
+	asl $d019
+	jmp exd
 
-	ldx npos
+e2:
-	stx tpos
 	inc rirq_count
 
-	bit	$d011
-	bmi e1
-
-	sta	$d012
-	jmp ex
-
-e1:
-	ldx	#0
-	stx nextIRQ
 	ldy	rasterIRQNext
 	dey
 	sty	$d012
-	jmp	ex
+	ldx	#0
-
+	stx nextIRQ
+	beq	exd
 }
 
-__asm irq1
+__asm rirq_isr_kernal_io
 {	
 	lda $d019
 	bpl ex2
 	
 	ldx	nextIRQ
+	bmi exi
 l1:
 	lda	rasterIRQNext, x
-	cmp	#$ff
-	beq	e1
-
 	ldy rasterIRQIndex + 1, x
-	tax
+	ldx	rasterIRQLow, y
-	lda	rasterIRQLow, y
+	stx ji + 1
-	sta ji + 1
+	ldx	rasterIRQHigh, y
-	lda	rasterIRQHigh, y
+	stx ji + 2
-	sta ji + 2
 
 ji:	
 	jsr $0000
@@ -198,45 +237,36 @@ jx:
 	inc	nextIRQ
 	ldx nextIRQ
 
-	lda	rasterIRQNext, x
+	ldy	rasterIRQNext, x
-	cmp #$ff
-	beq e2
 
-	tay
-
-	sec
-	sbc	#4
-	cmp	$d012
-	bcc l1
-
-	dey
-	sty	$d012
-w1:
-	jmp ex
-
-e2:
-	ldx npos
-	stx tpos
-	inc rirq_count
-
-	bit	$d011
-	bmi e1
-
-	sta $d012
-
-	jmp ex
-
-e1:
-	ldx	#0
-	stx nextIRQ
-	lda	rasterIRQNext, x
-	sec
-	sbc	#1
-	sta $d012
-	
-ex:
 	asl $d019
 
+	cpy #$ff
+	beq e2
+
+	dey
+	dey
+	sty	$d012
+	dey
+	cpy $d012
+	bcc l1
+
+exd:
+	jmp $ea81
+
+exi:
+	asl $d019
+	jmp $ea81
+
+e2:
+	inc rirq_count
+
+	ldy	rasterIRQNext
+	dey
+	dey
+	sty	$d012
+	ldx	#0
+	stx nextIRQ
 	jmp $ea81
 
 ex2:
@@ -245,7 +275,7 @@ ex2:
 	jmp $ea31
 }
 
-__asm irq3
+__asm rirq_isr_kernal_noio
 {	
 	lda $01
 	pha
@@ -256,17 +286,14 @@ __asm irq3
 	bpl ex2
 	
 	ldx	nextIRQ
+	bmi exi
 l1:
 	lda	rasterIRQNext, x
-	cmp	#$ff
-	beq	e1
-
 	ldy rasterIRQIndex + 1, x
-	tax
+	ldx	rasterIRQLow, y
-	lda	rasterIRQLow, y
+	stx ji + 1
-	sta ji + 1
+	ldx	rasterIRQHigh, y
-	lda	rasterIRQHigh, y
+	stx ji + 2
-	sta ji + 2
 
 ji:	
 	jsr $0000
@@ -275,49 +302,40 @@ jx:
 	inc	nextIRQ
 	ldx nextIRQ
 
-	lda	rasterIRQNext, x
+	ldy	rasterIRQNext, x
-	cmp #$ff
+
+	asl $d019
+
+	cpy #$ff
 	beq e2
 
-	tay
+	dey
-
-	sec
-	sbc	#4
-	cmp	$d012
-	bcc l1
-
 	dey
 	sty	$d012
-w1:
+	dey
-	jmp ex
+	cpy $d012
-
+	bcc l1
-e2:
+exd:
-	ldx npos
-	stx tpos
-	inc rirq_count
-
-	bit	$d011
-	bmi e1
-
-	sta $d012
-
-	jmp ex
-
-e1:
-	ldx	#0
-	stx nextIRQ
-	lda	rasterIRQNext, x
-	sec
-	sbc	#1
-	sta $d012
-	
-ex:
-	asl $d019
 	pla
 	sta $01
 
 	jmp $ea81
 
+exi:
+	asl $d019
+	jmp exd
+
+e2:
+	inc rirq_count
+
+	ldy	rasterIRQNext
+	dey
+	dey
+	sty	$d012
+	ldx	#0
+	stx nextIRQ
+	beq	exd
+
 ex2:
 	LDA $DC0D
 	cli
@@ -347,8 +365,8 @@ void rirq_build(RIRQCode * ic, byte size)
 	ic->size = size;
 
 	asm_im(ic->code + 0, ASM_LDY, 0);
-	asm_im(ic->code + 2, ASM_LDA, 0);
+	asm_im(ic->code + 2, ASM_LDX, 0);
-	asm_ab(ic->code + 4, ASM_CPX, 0xd012);
+	asm_ab(ic->code + 4, ASM_CMP, 0xd012);
 	asm_rl(ic->code + 7, ASM_BCS, -5);
 	asm_ab(ic->code + 9, ASM_STY, 0x0000);
 
@@ -362,7 +380,7 @@ void rirq_build(RIRQCode * ic, byte size)
 	}
 	else
 	{
-		asm_ab(ic->code + 12, ASM_STA, 0x0000);
+		asm_ab(ic->code + 12, ASM_STX, 0x0000);
 
 		byte p = 15;
 		for(byte i=2; i<size; i++)
@@ -482,7 +500,7 @@ void rirq_init_kernal(void)
         sei
     }
 
-   	*(void **)0x0314 = irq1;
+   	*(void **)0x0314 = rirq_isr_kernal_io;
 
 	vic.intr_enable = 1;
 	vic.ctrl1 &= 0x7f;
@@ -490,7 +508,7 @@ void rirq_init_kernal(void)
 
 }
 
-void rirq_init_kernal_io(void)
+void rirq_init_kernal_noio(void)
 {
 	rirq_init_tables();
 
@@ -499,7 +517,7 @@ void rirq_init_kernal_io(void)
         sei
     }
 
-   	*(void **)0x0314 = irq3;
+   	*(void **)0x0314 = rirq_isr_kernal_noio;
 
 	vic.intr_enable = 1;
 	vic.ctrl1 &= 0x7f;
@@ -516,8 +534,8 @@ void rirq_init_crt(void)
         sei
     }
 
-   	*(void **)0x0314 = irq0.kentry;
+   	*(void **)0x0314 = rirq_isr_io.kentry;
-   	*(void **)0xfffe = irq0;
+   	*(void **)0xfffe = rirq_isr_io;
 
 	vic.intr_enable = 1;
 	vic.ctrl1 &= 0x7f;
@@ -525,7 +543,7 @@ void rirq_init_crt(void)
 
 }
 
-void rirq_init_crt_io(void)
+void rirq_init_crt_noio(void)
 {
 	rirq_init_tables();
 
@@ -534,8 +552,8 @@ void rirq_init_crt_io(void)
         sei
     }
 
-   	*(void **)0x0314 = irq2.kentry;
+   	*(void **)0x0314 = rirq_isr_noio.kentry;
-   	*(void **)0xfffe = irq2;
+   	*(void **)0xfffe = rirq_isr_noio;
 
 	vic.intr_enable = 1;
 	vic.ctrl1 &= 0x7f;
@@ -552,7 +570,7 @@ void rirq_init_io(void)
         sei
     }
 
-   	*(void **)0xfffe = irq0;
+   	*(void **)0xfffe = rirq_isr_ram_io;
 
 	vic.intr_enable = 1;
 	vic.ctrl1 &= 0x7f;
@@ -569,7 +587,7 @@ void rirq_init_memmap(void)
         sei
     }
 
-   	*(void **)0xfffe = irq2;
+   	*(void **)0xfffe = rirq_isr_noio;
 
 	vic.intr_enable = 1;
 	vic.ctrl1 &= 0x7f;
@@ -587,12 +605,18 @@ void rirq_init(bool kernalIRQ)
 
 void rirq_wait(void)
 {
-	while (tpos != npos) ;
+	char i0 = rirq_pcount;
-	npos++;
+	char i1;
+	do {
+		i1 = rirq_count;
+	} while (i0 == i1);
+	rirq_pcount = i1;
 }
 
-void rirq_sort(void)
+void rirq_sort(bool inirq)
 {
+	// disable raster interrupts while sorting
+	nextIRQ = 0xff;
 #if 1
 	byte maxr = rasterIRQRows[rasterIRQIndex[1]];
 	for(byte i = 2; i<NUM_IRQS + 1; i++)
@@ -640,11 +664,18 @@ void rirq_sort(void)
 	}
 #endif
 
-	npos++;
+	rirq_pcount = rirq_count;
-	nextIRQ = 0;
+	if (inirq)
-	byte	yp = rasterIRQNext[nextIRQ];
+		nextIRQ = NUM_IRQS - 1;
+	else
+	{
+		byte	yp = rasterIRQNext[0];
 		if (yp != 0xff)
+		{
 			vic.raster = yp - 1;
+			nextIRQ = 0;
+		}
+	}
 }
 
 void rirq_start(void)
@@ -658,6 +689,7 @@ void rirq_start(void)
         lda #100
         sta $d012
 
+        asl $d019
         cli
     }
 }

include/c64/rasterirq.h

@@ -154,7 +154,7 @@ void rirq_init_kernal(void);
 
 // Raster IRQ through kernal, with IO range not always enabled
 // calls kernal continuation
-void rirq_init_kernal_io(void);
+void rirq_init_kernal_noio(void);
 
 // Raster IRQ through RAM and ROM vector, with ROM disabled or not and IO range always enabled
 // does not call kernal continuation
@@ -162,7 +162,7 @@ void rirq_init_crt(void);
 
 // Raster IRQ through RAM and ROM vector, with ROM disabled or not and IO range not always enabled
 // does not call kernal continuation
-void rirq_init_crt_io(void);
+void rirq_init_crt_noio(void);
 
 // Raster IRQ through RAM vector, with ROM disabled and IO range always enabled
 // does not call kernal continuation
@@ -179,8 +179,9 @@ void rirq_start(void);
 void rirq_stop(void);
 
 // Sort the raster IRQ, must be performed at the end of the frame after changing 
-// the vertical position of one of the interrupt operatins.
+// the vertical position of one of the interrupt operations.
-void rirq_sort(void);
+// Set the inirq flag to true when calling this from an interrupt
+void rirq_sort(bool inirq = false);
 
 // Wait for the last raster IRQ op to have completed.  Must be called before a
 // sort if the raster IRQ system is active

include/c64/sprites.c

@@ -171,6 +171,12 @@ void vspr_init(char * screen)
 	}
 }
 
+void vspr_shutdown(void)
+{
+	for(int i=0; i<VSPRITES_MAX - 7; i++)
+		rirq_clear(i);
+}
+
 void vspr_screen(char * screen)
 {	
 	vspriteScreen = screen + 0x3f8;	

include/c64/sprites.h

@@ -68,6 +68,8 @@ inline void spr_expand(char sp, bool xexpand, bool yexpand);
 
 void vspr_init(char * screen);
 
+void vspr_shutdown(void);
+
 void vspr_screen(char * screen);
 
 // set one sprite with the given attribute

include/c64/vic.c

@@ -114,4 +114,21 @@ void vic_waitBelow(int line)
 	}
 }
 
+void vic_waitRange(char below, char above)
+{
+	while (vic.ctrl1 & VIC_CTRL1_RST8)
+		;
+
+	if (vic.raster >= above)
+	{
+		while (!(vic.ctrl1 & VIC_CTRL1_RST8))					
+			;
+		while (vic.ctrl1 & VIC_CTRL1_RST8)
+			;
+	}
+
+	while (vic.raster < below)
+		;
+}
+
 #pragma native(vic_waitLine)

include/c64/vic.h

@@ -118,6 +118,9 @@ void vic_waitLine(int line);
 // wait for beam to be below a line
 void vic_waitBelow(int line);
 
+// wait for beam to be in a given range on screen
+void vic_waitRange(char below, char above);
+
 // reference to the VIC chip
 #define vic	(*((struct VIC *)0xd000))
 

include/conio.c

@@ -131,6 +131,18 @@ __asm bsinit
 	lda #147
 	jmp $ffd2	
 }
+#elif defined(__CBMPET__)
+#define bsout	0xffd2
+#define bsin	0xffe4
+__asm bsplot
+{
+    /* no equivalent on PET */
+}
+__asm bsinit
+{
+    /* no equivalent on PET */
+}
+#define bsget	0xffcf
 #else
 #define bsout	0xffd2
 #define bsin	0xffe4
@@ -212,11 +224,10 @@ void putpch(char c)
 						c ^= 0xa0;
 					c ^= 0x20;
 #else
-					if (c >= 97)
 					c ^= 0x20;
 #endif				
 
-					if (giocharmap == IOCHM_PETSCII_2)
+					if (giocharmap == IOCHM_PETSCII_1)
 						c &= 0xdf;
 				}
 			}
@@ -268,11 +279,19 @@ char getpch(void)
 
 char kbhit(void)
 {
+#if defined(__CBMPET__)
+	return __asm
+	{
+		lda $9e
+		sta	accu
+	};
+#else
 	return __asm
 	{
 		lda $c6
 		sta	accu
 	};
+#endif
 }
 
 char getche(void)
@@ -323,10 +342,7 @@ void putch(char c)
 
 void clrscr(void)
 {
-	__asm
+	putrch(147);
-	{
-		jsr	bsinit
-	}
 }
 
 void textcursor(bool show)
@@ -336,6 +352,24 @@ void textcursor(bool show)
 
 void gotoxy(char cx, char cy)
 {
+#if defined(__CBMPET__)
+#define CURS_X 0xc6
+#define CURS_Y 0xd8
+#define SCREEN_PTR 0xc4
+#define SCR_LINELEN 0xd5
+
+	__assume(cy < 25);
+
+	*(volatile char *)CURS_X = cx;
+	*(volatile char *)CURS_Y = cy;
+
+	if (*(volatile char *)SCR_LINELEN > 40)
+		cy <<= 1;
+
+	const unsigned	off = cy * 40;
+
+	* (volatile unsigned *)SCREEN_PTR = off + 0x8000;
+#else
 	__asm
 	{
 		ldx	cy
@@ -343,6 +377,7 @@ void gotoxy(char cx, char cy)
 		clc
 		jsr bsplot
 	}
+#endif
 }
 
 void textcolor(char c)
@@ -350,6 +385,24 @@ void textcolor(char c)
 	*(volatile char *)0x0286 = c;
 }
 
+void bgcolor(char c)
+{
+	*(volatile char *)0xd021 = c;
+}
+
+void bordercolor(char c)
+{
+	*(volatile char *)0xd020 = c;
+}
+
+void revers(char r)
+{
+	if (r) 
+		putrch(18);
+	else
+		putrch(18 + 128);
+}
+
 char wherex(void)
 {
 #if defined(__C128__) || defined(__C128B__) || defined(__C128E__)

include/conio.h

@@ -43,6 +43,26 @@ void dispmode80col(void);
 #define PETSCII_F7				0x88
 #define PETSCII_F8				0x8c
 
+enum ConioColors
+{
+	COLOR_BLACK,
+	COLOR_WHITE,
+	COLOR_RED,
+	COLOR_CYAN,
+	COLOR_PURPLE,
+	COLOR_GREEN,
+	COLOR_BLUE,
+	COLOR_YELLOW,
+
+	COLOR_ORANGE,
+	COLOR_BROWN,
+	COLOR_LT_RED,
+	COLOR_DARK_GREY,
+	COLOR_MED_GREY,
+	COLOR_LT_GREEN,
+	COLOR_LT_BLUE,
+	COLOR_LT_GREY
+};
 // Lowlevel console in/out
 
 // using petscii translation
@@ -74,6 +94,12 @@ void gotoxy(char x, char y);
 
 inline void textcolor(char c);
 
+inline void bgcolor(char c);
+
+inline void bordercolor(char c);
+
+inline void revers(char r);
+
 inline char wherex(void);
 
 inline char wherey(void);

include/crt.c

@@ -667,21 +667,19 @@ L1:		rol	accu
 		sbc	tmp + 1
 		lda	tmp + 6
 		sbc	tmp + 2
+		tax
 		lda	tmp + 7
 		sbc	tmp + 3
 		bcc	W1
+		stx tmp + 6
+		sta tmp + 7
 		lda	tmp + 4
 		sbc	tmp + 0
 		sta tmp + 4
 		lda	tmp + 5
 		sbc	tmp + 1
 		sta tmp + 5
-		lda	tmp + 6
+		sec
-		sbc	tmp + 2
-		sta tmp + 6
-		lda	tmp + 7
-		sbc	tmp + 3
-		sta tmp + 7
 W1:		dey
 		bne	L1
 		rol	accu
@@ -758,6 +756,11 @@ __asm mul32
 		sta	tmp + 6
 		sta	tmp + 7
 
+		lda tmp + 3
+		ora tmp + 2
+		ora tmp + 1
+		beq WB
+
 		lda	tmp + 0
 		jsr WM
 		lda	tmp + 1
@@ -775,6 +778,7 @@ WM:
 		stx accu + 1
 		sta accu
 		rts
+WB:		lda tmp + 0
 W0:
 		sec
 		ror
@@ -3072,7 +3076,7 @@ __asm inp_binop_sub_f32
 
 #pragma	bytecode(BC_BINOP_SUB_F32, inp_binop_sub_f32)
 		
-__asm fmul8
+__asm crt_fmul8
 {
 		sec
 		ror
@@ -3106,7 +3110,7 @@ W1:
 		rts
 }
 
-__asm fmul
+__asm crt_fmul
 {
 		lda	accu
 		ora	accu + 1
@@ -3147,13 +3151,13 @@ W2:
 		beq W5
 		bne W6
 W4:
-		jsr	fmul8
+		jsr	crt_fmul8
 		lda	tmp + 1
 W6:		
-		jsr	fmul8
+		jsr	crt_fmul8
 W5:		
 		lda	tmp + 2
-		jsr	fmul8
+		jsr	crt_fmul8
 		
 		sec
 		lda	tmp + 8
@@ -3210,12 +3214,12 @@ __asm inp_binop_mul_f32
 {
 		jsr	freg.split_exp
 		sty tmpy
-		jsr fmul
+		jsr crt_fmul
 		ldy tmpy
 		jmp	startup.exec
 }
 
-__asm fdiv
+__asm crt_fdiv
 {
 		lda	accu
 		ora	accu + 1
@@ -3337,13 +3341,13 @@ ZERO:
 __asm inp_binop_div_f32
 {
 		jsr	freg.split_exp
-		jsr fdiv
+		jsr crt_fdiv
 		jmp	startup.exec
 }
 
 #pragma	bytecode(BC_BINOP_DIV_F32, inp_binop_div_f32)
 
-__asm fcmp
+__asm crt_fcmp
 {
 		lda	accu + 3
 		eor	tmp + 3
@@ -4058,9 +4062,9 @@ __asm load32
 #pragma runtime(fmergea, freg.merge_aexp)
 #pragma runtime(fadd, faddsub.fadd)
 #pragma runtime(fsub, faddsub.fsub)
-#pragma runtime(fmul, fmul)
+#pragma runtime(fmul, crt_fmul)
-#pragma runtime(fdiv, fdiv)
+#pragma runtime(fdiv, crt_fdiv)
-#pragma runtime(fcmp, fcmp)
+#pragma runtime(fcmp, crt_fcmp)
 #pragma runtime(ffromi, sint16_to_float)
 #pragma runtime(ffromu, uint16_to_float)
 #pragma runtime(ftoi, f32_to_i16)
@@ -4558,7 +4562,7 @@ struct Heap {
 
 #pragma section(heap, 0x0000, HeapStart, HeapEnd)
 
-__asm malloc
+__asm crt_malloc
 {
 		// make room for two additional bytes
 		// to store pointer to end of used memory
@@ -4785,14 +4789,14 @@ hc2:
 __asm inp_malloc
 {
 		sty tmpy
-		jsr malloc
+		jsr crt_malloc
 		ldy tmpy
 		jmp	startup.exec
 }
 
 #pragma bytecode(BC_MALLOC, inp_malloc)
 
-__asm free
+__asm crt_free
 {
 
 		// check nullptr free
@@ -5009,21 +5013,21 @@ nostart:
 __asm inp_free
 {
 		sty tmpy
-		jsr free
+		jsr crt_free
 		ldy tmpy
 		jmp	startup.exec
 }
 
 #pragma bytecode(BC_FREE, inp_free)
 
-__asm breakpoint
+__asm crt_breakpoint
 {
 		rts
 }
 
-#pragma runtime(malloc, malloc)
+#pragma runtime(malloc, crt_malloc)
-#pragma runtime(free, free)
+#pragma runtime(free, crt_free)
-#pragma runtime(breakpoint, breakpoint)
+#pragma runtime(breakpoint, crt_breakpoint)
 
 #if 0
 

include/fixmath.c

@@ -123,8 +123,9 @@ int lmul4f12s(int x, int y)
 
 		lda #0
 		sta	accu + 1
-L2:		
+
 		bcc	W4
+L2:		
 		tay
 		clc
 		lda	accu + 1
@@ -165,7 +166,7 @@ W1:
 		bcc W2
 
 		tay
-		sec
+//		sec 		; we know it is set here
 		lda accu + 1
 		sbc y
 		sta accu + 1

include/gfx/bitmap.c

@@ -525,7 +525,7 @@ void bm_polygon_nc_fill(const Bitmap * bm, const ClipRect * clip, int * px, int
 static inline void buildline(char ly, char lx, int dx, int dy, int stride, bool left, bool up, char pattern, LineOp op)
 {
 	char	ip = 0;
-	bool	delta16 = ((dx | dy) & 0xff80) != 0;
+	bool	delta16 =((dx | dy) & 0xff80) != 0;
 
 	// ylow
 	ip += asm_im(BLIT_CODE + ip, ASM_LDY, ly);
@@ -579,33 +579,51 @@ static inline void buildline(char ly, char lx, int dx, int dy, int stride, bool
 		break;
 	}
 
-	if (dx && dy)
+	if (dy)
+	{
+		bool	delta8 = false;
+
+		if (dx)
 		{
 			// m >= 0
 			ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_DP + delta16);
-		ip += asm_rl(BLIT_CODE + ip, ASM_BMI, delta16 ? 5 + 15 + 13 + 2 : 5 + 15 + 7 + 2);
+			char n = delta16 ? 18 + 13 + 2 : 18 + 7 + 2;
+			if (!up) n++;
+			ip += asm_rl(BLIT_CODE + ip, ASM_BMI, n);
+			delta8 = !delta16;
 		}
 
-	if (dy)
+		if (up)
 		{
-		ip += asm_np(BLIT_CODE + ip, up ? ASM_DEY : ASM_INY);
+			ip += asm_np(BLIT_CODE + ip, ASM_DEY);
-		ip += asm_im(BLIT_CODE + ip, ASM_CPY, up ? 0xff : 0x08);
+			ip += asm_rl(BLIT_CODE + ip, ASM_BPL, delta8 ? 17 : 15);
-		ip += asm_rl(BLIT_CODE + ip, ASM_BNE, 15);
-
 			ip += asm_np(BLIT_CODE + ip, ASM_CLC);
+			ip += asm_im(BLIT_CODE + ip, ASM_LDY, 0x07);
 			ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_SP);
 			ip += asm_im(BLIT_CODE + ip, ASM_ADC, stride & 0xff);
 			ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_SP);
 			ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_SP + 1);
 			ip += asm_im(BLIT_CODE + ip, ASM_ADC, stride >> 8);
 			ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_SP + 1);
-		ip += asm_im(BLIT_CODE + ip, ASM_LDY, up ? 0x07 : 0x00);
+		}
+		else
+		{
+			ip += asm_np(BLIT_CODE + ip, ASM_INY);
+			ip += asm_im(BLIT_CODE + ip, ASM_CPY, 0x08);
+			ip += asm_rl(BLIT_CODE + ip, ASM_BNE, delta8 ? 16 : 14);
+			ip += asm_im(BLIT_CODE + ip, ASM_LDY, 0x00);
+			ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_SP);
+			ip += asm_im(BLIT_CODE + ip, ASM_ADC, (stride - 1) & 0xff);
+			ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_SP);
+			ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_SP + 1);
+			ip += asm_im(BLIT_CODE + ip, ASM_ADC, (stride - 1) >> 8);
+			ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_SP + 1);
 		}
 
-	if (dx && dy)
+		if (dx)
 		{
-		ip += asm_np(BLIT_CODE + ip, ASM_SEC);
 			ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_DP);
+			ip += asm_np(BLIT_CODE + ip, ASM_SEC);
 			ip += asm_im(BLIT_CODE + ip, ASM_SBC, dx & 0xff);
 			ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_DP);
 
@@ -614,47 +632,52 @@ static inline void buildline(char ly, char lx, int dx, int dy, int stride, bool
 				ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_DP + 1);
 				ip += asm_im(BLIT_CODE + ip, ASM_SBC, dx >> 8);
 				ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_DP + 1);
+				ip += asm_rl(BLIT_CODE + ip, ASM_BPL, 13 + 4 + 12);
+
+				ip += asm_np(BLIT_CODE + ip, ASM_CLC);
+				ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_DP);
+				ip += asm_im(BLIT_CODE + ip, ASM_ADC, dy & 0xff);
+				ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_DP);
+				ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_DP + 1);
+				ip += asm_im(BLIT_CODE + ip, ASM_ADC, dy >> 8);
+				ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_DP + 1);
+			}
+			else
+			{
+				// We know regdp to be in the accu at this point
+				ip += asm_rl(BLIT_CODE + ip, ASM_BPL, 5 + 4 + 12);
+				ip += asm_np(BLIT_CODE + ip, ASM_CLC);
+				ip += asm_im(BLIT_CODE + ip, ASM_ADC, dy & 0xff);
+				ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_DP);
+			}
 		}
 
 		// m < 0
-		ip += asm_rl(BLIT_CODE + ip, ASM_BPL, delta16 ? 4 + 13 + 13 : 4 + 13 + 7);
 	}
 
 	if (dx)
 	{
 		ip += asm_zp(BLIT_CODE + ip, left ? ASM_ASL : ASM_LSR, REG_D0);
-		ip += asm_rl(BLIT_CODE + ip, ASM_BCC, 13);
+		ip += asm_rl(BLIT_CODE + ip, ASM_BCC, 12);
 
 		ip += asm_zp(BLIT_CODE + ip, left ? ASM_ROL : ASM_ROR, REG_D0);
-		ip += asm_np(BLIT_CODE + ip, ASM_CLC);
+
 		ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_SP);
-		ip += asm_im(BLIT_CODE + ip, ASM_ADC, left ? 0xf8 : 0x08);
-		ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_SP);
 
 		if (left)
 		{
+			ip += asm_im(BLIT_CODE + ip, ASM_ADC, 0xf8);
 			ip += asm_rl(BLIT_CODE + ip, ASM_BCS, 2);
 			ip += asm_zp(BLIT_CODE + ip, ASM_DEC, REG_SP + 1);
 		}
 		else
 		{
+			ip += asm_im(BLIT_CODE + ip, ASM_ADC, 0x08);
 			ip += asm_rl(BLIT_CODE + ip, ASM_BCC, 2);
 			ip += asm_zp(BLIT_CODE + ip, ASM_INC, REG_SP + 1);
 		}
-	}
 
-	if (dx && dy)
+		ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_SP);
-	{
-		ip += asm_np(BLIT_CODE + ip, ASM_CLC);
-		ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_DP);
-		ip += asm_im(BLIT_CODE + ip, ASM_ADC, dy & 0xff);
-		ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_DP);
-		if (delta16)
-		{
-			ip += asm_zp(BLIT_CODE + ip, ASM_LDA, REG_DP + 1);
-			ip += asm_im(BLIT_CODE + ip, ASM_ADC, dy >> 8);
-			ip += asm_zp(BLIT_CODE + ip, ASM_STA, REG_DP + 1);
-		}
 	}
 
 	// l --

include/math.c

@@ -110,12 +110,12 @@ float atan2(float p, float q)
 	return s;
 }
 
-#define F_EXP_0 	1.0000003
+#define F_EXP_0 	1.0
-#define F_EXP_1 	0.693147059
+#define F_EXP_1 	0.69315668
-#define F_EXP_2 	0.240173099
+#define F_EXP_2 	0.240132068
-#define F_EXP_3 	0.055816392
+#define F_EXP_3 	0.055876024
-#define F_EXP_4 	0.008965036
+#define F_EXP_4 	0.008940801
-#define F_EXP_5 	0.001898429
+#define F_EXP_5 	0.001894414
 
 float exp(float f)
 {
@@ -143,12 +143,12 @@ float exp(float f)
 	return s * x.f;
 }
 
-#define F_LOG_0 	-3.78712618	
+#define F_LOG_0 	-3.78717706
-#define F_LOG_1		10.0957081
+#define F_LOG_1 	10.0960498
-#define F_LOG_2		-13.9747486
+#define F_LOG_2 	-13.975654
-#define F_LOG_3		12.7568806
+#define F_LOG_3 	12.7580616
-#define F_LOG_4		-6.48114552
+#define F_LOG_4 	-6.48190725
-#define F_LOG_5		1.39045416
+#define F_LOG_5 	1.39064767
 
 float log(float f)
 {

include/math.h

@@ -33,10 +33,15 @@ bool isfinite(float f);
 #pragma intrinsic(sin)
 #pragma intrinsic(cos)
 #pragma intrinsic(tan)
+#pragma intrinsic(atan)
+#pragma intrinsic(atan2)
 
 #pragma intrinsic(log)
 #pragma intrinsic(exp)
 
+#pragma intrinsic(pow)
+#pragma intrinsic(sqrt)
+
 #pragma compile("math.c")
 
 

include/opp/boundint.h

@@ -0,0 +1,59 @@
+#ifndef OPP_BOUNDINT_H
+#define OPP_BOUNDINT_H
+
+namespace opp {
+	
+	template<int tmin, int tmax>
+	constexpr auto boundinttype(void)
+	{
+		if constexpr (tmin >= 0 && tmax <= 255)
+			return (char)0;
+		else if constexpr (tmin >= -128 && tmax <= 127)
+			return (signed char)0;
+		else
+			return (int)0;
+	}
+
+	template<int tmin, int tmax> 
+	class boundint
+	{
+	protected:
+		decltype(boundinttype<tmin, tmax>()) 	v;
+	public:
+		boundint(int i) 
+			: v(i) 
+		{
+
+		}
+
+		void operator=(int k)
+		{
+			__assume(k >= tmin && k <= tmax);
+			v = k;
+		}
+
+		void operator+=(int k)
+		{
+			k += v;
+			__assume(k >= tmin && k <= tmax);
+			v = k;
+		}
+
+		void operator-=(int k)
+		{
+			k = v - k;
+			__assume(k >= tmin && k <= tmax);
+			v = k;
+		}
+
+		operator int() const
+		{
+			int k = v;
+			__assume(k >= tmin && k <= tmax);
+			return k;
+		}
+	};
+
+}
+
+#endif

include/opp/static_vector.h

@@ -4,6 +4,7 @@
 #include <new>
 #include <stdlib.h>
 #include <opp/utility.h>
+#include <oscar.h>
 
 namespace opp {
 
@@ -11,8 +12,8 @@ template <class T, int N>
 class static_vector
 {
 protected:
-	char	_space[N * sizeof(T)];
 	enum { m = N }	_size;
+	char	_space[N * sizeof(T)];
 public:
 	typedef T 	element_type;
 
@@ -20,23 +21,28 @@ public:
 
 	static_vector(size_t n) : _size(n) 
 	{
+#ifdef CAPACITYCHECK
+	if (n > N) debugcrash();
+#endif
 		T *	data = (T*)_space;
 		for(size_t i=0; i<n; i++)
-			new (data + i) T;
+			new (data + i) T();
 	}
 
 	static_vector(const static_vector & v)
 		: _size(v._size) 
 	{
+		size_t n = _size;
 		T *	data = (T*)_space, * vdata = (T*)(v._space);
-		for(size_t i=0; i<_size; i++)
+		for(size_t i=0; i<n; i++)
 			new (data + i)T(vdata[i]);
 	}
 
 	~static_vector(void)
 	{
 		T *	data = (T*)_space;
-		for(size_t i=0; i<_size; i++)
+		size_t n = _size;
+		for(size_t i=0; i<n; i++)
 			data[i].~T();
 	}
 
@@ -45,10 +51,12 @@ public:
 		if (this != &v)
 		{
 			T *	data = (T*)_space, * vdata = (T*)(v._space);
-			for(size_t i=0; i<_size; i++)
+			size_t n = _size;
+			for(size_t i=0; i<n; i++)
 				data[i].~T();
 			_size = v._size; 
-			for(size_t i=0; i<_size; i++)
+			n = _size;
+			for(size_t i=0; i<n; i++)
 				new (data + i)T(vdata[i]);
 		}
 		return *this;
@@ -69,6 +77,11 @@ public:
 		return _size == 0;
 	}
 
+	bool full(void) const
+	{
+		return _size == N;
+	}
+
 	size_t capacity(void) const
 	{
 		return N;
@@ -76,6 +89,8 @@ public:
 
 	void resize(size_t n);
 
+	void clear(void);
+
 	T & at(size_t at)
 	{
 		return ((T*)_space)[at];
@@ -166,6 +181,8 @@ public:
 		((T*)_space)[_size].~T();
 	}
 
+	void assign(size_t count, const T & t);
+
 	void insert(size_t at, const T & t);
 
 	void erase(size_t at, size_t n = 1);
@@ -177,34 +194,50 @@ public:
 };
 
 
+template <class T, int N>
+void static_vector<T, N>::clear(void)
+{
+	T *	data = (T*)_space;
+	for(size_t i=0; i<_size; i++)
+		data[i].~T();			
+	_size = 0;
+}
 
 template <class T, int N>
 void static_vector<T, N>::resize(size_t n)
 {
+#ifdef CAPACITYCHECK
+	if (n > N) debugcrash();
+#endif
 	T *	data = (T*)_space;
 	if (n < _size)
 	{
 		for(size_t i=n; i<_size; i++)
 			data[i].~T();			
-		_size = n;
 	}
-	else
+	else if (n > 0)
 	{
 		for(size_t i=_size; i<n; i++)
-			new(data + i)T;
+			new(data + i)T();
-		_size = n;
 	}
+	_size = n;
 }
 
 template <class T, int N>
 void static_vector<T, N>::push_back(const T & t)
 {
+#ifdef CAPACITYCHECK
+	if (_size >= N) debugcrash();
+#endif
 	new ((T*)_space + _size++)T(t);
 }
 
 template <class T, int N>
 void static_vector<T, N>::push_back(T && t)
 {
+#ifdef CAPACITYCHECK
+	if (_size >= N) debugcrash();
+#endif
 	new ((T*)_space + _size++)T(t);
 }
 
@@ -212,14 +245,28 @@ template <class T, int N>
 template <typename ...P>
 void static_vector<T, N>::emplace_back(const P&... p)
 {
+#ifdef CAPACITYCHECK
+	if (_size >= N) debugcrash();
+#endif
 	new ((T*)_space + _size++)T(p...);
 }
 
+template <class T, int N>
+void static_vector<T, N>::assign(size_t count, const T & t)
+{
+	T *	data = (T*)_space;
+	for(size_t i=0; i<_size; i++)
+		data[i].~T();
+	for(size_t i=0; i<count; i++)
+		new (data + i)T(t);
+	_size = count;
+}
+
 template <class T, int N>
 void static_vector<T, N>::insert(size_t at, const T & t)
 {
 	T *	data = (T*)_space;
-	new (data + _size)T;
+	new (data + _size)T();
 	for(size_t i=_size; i>at; i--)
 		data[i] = move(data[i - 1]);
 	data[at] = t;
@@ -240,9 +287,12 @@ void static_vector<T, N>::erase(size_t at, size_t n)
 template <class T, int N>
 T * static_vector<T, N>::insert(T * at, const T & t)
 {
+#ifdef CAPACITYCHECK
+	if (_size >= N) debugcrash();
+#endif
 	T *	data = (T*)_space;
 	T * dp = data + _size;
-	new (dp)T;
+	new (dp)T();
 	while (dp != at)
 	{
 		dp--;

include/opp/vector.h

@@ -21,13 +21,14 @@ public:
 	vector(size_t n) : _data((T*)malloc(n * sizeof(T))), _size(n), _capacity(n) 
 	{
 		for(size_t i=0; i<n; i++)
-			new (_data + i) T;
+			new (_data + i) T();
 	}
 
 	vector(const vector & v)
 		: _data((T*)malloc(v._size * sizeof(T))), _size(v._size), _capacity(v._size) 
 	{
-		for(size_t i=0; i<_size; i++)
+		size_t n = _size;
+		for(size_t i=0; i<n; i++)
 			new (_data + i)T(v._data[i]);
 	}
 
@@ -50,14 +51,16 @@ public:
 	{
 		if (this != &v)
 		{
-			for(size_t i=0; i<_size; i++)
+			size_t n = _size;
+			for(size_t i=0; i<n; i++)
 				_data[i].~T();
 			free(_data);
 
 			_data = (T*)malloc(v._size * sizeof(T));
 			_size = v._size; 
 			_capacity = v._size; 
-			for(size_t i=0; i<_size; i++)
+			n = _size;
+			for(size_t i=0; i<n; i++)
 				new (_data + i)T(v._data[i]);		
 		}
 		return *this;
@@ -94,6 +97,8 @@ public:
 		return _capacity;
 	}
 
+	void clear(void);
+
 	void resize(size_t n);
 
 	void reserve(size_t n);
@@ -190,6 +195,8 @@ public:
 		_data[_size].~T();
 	}
 
+	void assign(size_t count, const T & t);
+
 	void insert(size_t at, const T & t);
 
 	void erase(size_t at, size_t n = 1);
@@ -204,7 +211,7 @@ protected:
 
 
 template <class T>
-void vector<T>::reserve(size_t n)
+__noinline void vector<T>::reserve(size_t n)
 {
 	if (n > _capacity)
 	{
@@ -221,7 +228,15 @@ void vector<T>::reserve(size_t n)
 }
 
 template <class T>
-void vector<T>::resize(size_t n)
+void vector<T>::clear(void)
+{
+	for(size_t i=0; i<_size; i++)
+		_data[i].~T();			
+	_size = 0;
+}
+
+template <class T>
+__noinline void vector<T>::resize(size_t n)
 {
 	if (n < _size)
 	{
@@ -232,7 +247,7 @@ void vector<T>::resize(size_t n)
 	else if (n < _capacity)
 	{
 		for(size_t i=_size; i<n; i++)
-			new(_data + i)T;
+			new(_data + i)T();
 		_size = n;
 	}
 	else
@@ -286,12 +301,27 @@ void vector<T>::emplace_back(const P&... p)
 	new (add_back())T(p...);
 }
 
+template <class T>
+void vector<T>::assign(size_t count, const T & t)
+{
+	for(size_t i=0; i<_size; i++)
+		_data[i].~T();
+	if (count > _capacity)
+	{
+		_size = 0;
+		reserve(count);		
+	}
+	for(size_t i=0; i<count; i++)
+		new (_data + i)T(t);
+	_size = count;
+}
+
 template <class T>
 void vector<T>::insert(size_t at, const T & t)
 {
 	if (_size == _capacity)
 		reserve(_size + 1 + (_size >> 1));
-	new (_data + _size)T;
+	new (_data + _size)T();
 	for(size_t i=_size; i>at; i--)
 		_data[i] = move(_data[i - 1]);
 	_data[at] = t;
@@ -318,7 +348,7 @@ T * vector<T>::insert(T * at, const T & t)
 		at = (T *)(f + unsigned(_data));
 	}
 	T * dp = _data + _size;
-	new (dp)T;
+	new (dp)T();
 	while (dp != at)
 	{
 		dp--;

include/oscar.c

@@ -129,3 +129,10 @@ __native const char * oscar_expand_lzo_buf(char * dp, const char * sp)
 
 	return sp + 1;
 }
+
+void debugcrash(void)
+{
+	__asm volatile {
+		byt $02
+	}
+}

include/oscar.h

@@ -18,6 +18,8 @@ void breakpoint(void);
 
 #pragma intrinsic(breakpoint)
 
+void debugcrash(void);
+
 #pragma compile("oscar.c")
 
 #endif

include/stdio.c

@@ -417,7 +417,7 @@ char * sformat(char * buff, const char * fmt, int * fps, bool print)
 				si.precision = i;		
 			}
 
-			if (c == 'd' || c == p'd')
+			if (c == 'd' || c == p'd' || c == 'i' || c == p'i')
 			{
 				bi = nformi(&si, bp, *fps++, true);
 			}
@@ -556,6 +556,17 @@ int sprintf(char * str, const char * fmt, ...)
 	return d - str;
 }
 
+void vprintf(const char * fmt, va_list vlist)
+{
+	char	buff[50];
+	sformat(buff, fmt, (int *)vlist, true);
+}
+
+int vsprintf(char * str, const char * fmt, va_list vlist)
+{
+	char * d = sformat(str, fmt, (int *)vlist, false);
+	return d - str;
+}
 
 static inline bool isspace(char c)
 {

include/stdio.h

@@ -3,6 +3,7 @@
 
 #include <stdlib.h>
 #include <stddef.h>
+#include <stdarg.h>
 
 void putchar(char c);
 
@@ -18,6 +19,10 @@ void printf(const char * fmt, ...);
 
 int sprintf(char * str, const char * fmt, ...);
 
+void vprintf(const char * fmt, va_list vlist);
+
+int vsprintf(char * str, const char * fmt, va_list vlist);
+
 int scanf(const char * fmt, ...);
 
 int sscanf(const char * str, const char * fmt, ...);

include/stdlib.c

@@ -544,7 +544,7 @@ unsigned heapfree(void)
 }
 
 #if 0
-struct Heap {
+struct Heap {q
 	unsigned int	size;
 	Heap		*	next;
 }	*	freeHeap;
@@ -667,6 +667,97 @@ void * calloc(int num, int size)
 	return p;
 }
 
+void * realloc(void * ptr, unsigned size)
+{
+	if (ptr)
+	{
+#ifdef HEAPCHECK
+		Heap	*	pheap = (Heap *)((char *)ptr - 6);
+		Heap	*	eheap = pheap->next;
+
+		unsigned	psize = (char *)eheap - (char *)pheap;
+
+		void * nptr = malloc(size);
+		memcpy(nptr, ptr, psize - 6);
+		free(ptr);
+		return nptr;
+#else
+		unsigned	nsize = (size + 5) & ~3;
+
+		// Get heap info
+		Heap	*	pheap = (Heap *)((char *)ptr - 2);
+		Heap	*	eheap = pheap->next;
+
+		unsigned	psize = (char *)eheap - (char *)pheap;
+
+		Heap	*	h = HeapNode.next;
+		Heap	*	hp = &HeapNode;
+		while (h && h < pheap)
+		{
+			hp = h;
+			h = h->next;
+		}		
+
+		if (nsize <= psize)
+		{
+			// check if we should free some memory
+			if (nsize + sizeof(HeapNode) < psize)
+			{
+				Heap	*	nheap = (Heap *)((char *)pheap + nsize);
+				pheap->next = nheap;
+
+				if (h == eheap)
+				{
+					nheap->end = h->end;
+					nheap->next = h->next;
+				}
+				else 
+				{
+					nheap->end = eheap;
+					nheap->next = h;
+				}
+
+				hp->next = nheap;
+			}
+
+			return ptr;
+		}
+		else if (h == eheap)
+		{
+			// Free space after this
+
+			// Check if enough space if extending
+			unsigned	xsize = (char *)h->end - (char *)pheap;
+			if (xsize >= nsize)
+			{
+				if (xsize > nsize + sizeof(HeapNode))
+				{
+					Heap	*	nheap = (Heap *)((char *)pheap + nsize);
+					pheap->next = nheap;
+					nheap->end = h->end;
+					nheap->next = h->next;
+					hp->next = nheap;
+				}
+				else
+				{
+					pheap->next = h->end;
+					hp->next = h->next;
+				}
+
+				return ptr;
+			}
+		}
+
+		void * nptr = malloc(size);
+		memcpy(nptr, ptr, psize - 2);
+		free(ptr);
+		return nptr;
+#endif		
+	}
+	else
+		return malloc(size);
+}
+
 static unsigned seed = 31232;
 
 unsigned int rand(void)

include/stdlib.h

@@ -45,6 +45,8 @@ void free(void * ptr);
 
 void * calloc(int num, int size);
 
+void * realloc(void * ptr, unsigned size);
+
 unsigned heapfree(void);
 
 unsigned int rand(void);

include/string.c

@@ -138,6 +138,20 @@ char * cpycat(char * dst, const char * src)
 
 #pragma native(cpycat)
 
+char* strchr( const char* str, int ch )
+{
+	char * p = (char *)str;
+
+	while (*p != (char)ch)
+	{
+		if (!*p)
+			return nullptr;
+		p++;
+	}
+	return p;
+}
+
+
 void * memset(void * dst, int value, int size)
 {
 	__asm

include/string.h

@@ -21,6 +21,8 @@ int memcmp(const void * ptr1, const void * ptr2, int size);
 
 void * memmove(void * dst, const void * src, int size);
 
+char* strchr( const char* str, int ch );
+
 #pragma intrinsic(strcpy)
 
 #pragma intrinsic(memcpy)

make/makefile

@@ -12,8 +12,8 @@ export OSCAR64_CFLAGS =
 export OSCAR64_CXX = $(project_dir)/bin/oscar64
 MKDIR_PARENT = /bin/mkdir -p -m 755
 INSTALL = /usr/bin/install
-INSTALL_PROGRAM = $(INSTALL) --mode=755
+INSTALL_PROGRAM = $(INSTALL) -m 755
-INSTALL_DATA = $(INSTALL) --mode=644
+INSTALL_DATA = $(INSTALL) -m 644
 DESTDIR =
 prefix = /usr/local
 exec_prefix = $(prefix)
@@ -38,7 +38,7 @@ else
 endif
 
 
-all: --prep-build-dir compiler samples tests
+all: compiler samples check
 
 
 $(srcdir)/%.o: $(project_dir)/oscar64/%.cpp
@@ -56,7 +56,7 @@ $(srcdir)/%.d: $(project_dir)/oscar64/%.cpp
 	$(RM) $@.$$$$
 
 
-compiler: $(objects)
+compiler: --prep-build-dir $(objects)
 	@$(MKDIR_PARENT) $(srcdir)
 	@echo "Linking compiler..."
 	$(CXX) $(CPPFLAGS) $(objects) $(linklibs) -o $(project_dir)/bin/oscar64
@@ -85,7 +85,7 @@ samples: compiler
 	@$(MAKE) -C $(project_dir)/samples all
 
 
-tests: compiler
+check: compiler
 	@$(MAKE) -C $(project_dir)/autotest all
 
 install: compiler
@@ -93,16 +93,16 @@ install: compiler
 	@$(MKDIR_PARENT) $(DESTDIR)$(bindir)
 	$(INSTALL_PROGRAM) $(project_dir)/bin/oscar64 $(DESTDIR)$(bindir)
 	@$(MKDIR_PARENT) $(DESTDIR)$(includedir)/oscar64/{audio,c64,c128,cx16,gfx,nes,opp,plus4,vic20}
-	$(INSTALL_DATA) $(project_dir)/include/*.{h,c} $(DESTDIR)$(includedir)/oscar64
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/*.h $(project_dir)/include/*.c) $(DESTDIR)$(includedir)/oscar64
-	$(INSTALL_DATA) $(project_dir)/include/audio/*.{h,c} $(DESTDIR)$(includedir)/oscar64/audio
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/audio/*.h $(project_dir)/include/audio/*.c) $(DESTDIR)$(includedir)/oscar64/audio
-	$(INSTALL_DATA) $(project_dir)/include/c64/*.{h,c} $(DESTDIR)$(includedir)/oscar64/c64
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/c64/*.h $(project_dir)/include/c64/*.c) $(DESTDIR)$(includedir)/oscar64/c64
-	$(INSTALL_DATA) $(project_dir)/include/c128/*.{h,c} $(DESTDIR)$(includedir)/oscar64/c128
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/c128/*.h $(project_dir)/include/c128/*.c) $(DESTDIR)$(includedir)/oscar64/c128
-	$(INSTALL_DATA) $(project_dir)/include/cx16/*.{h,c} $(DESTDIR)$(includedir)/oscar64/cx16
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/cx16/*.h $(project_dir)/include/cx16/*.c) $(DESTDIR)$(includedir)/oscar64/cx16
-	$(INSTALL_DATA) $(project_dir)/include/gfx/*.{h,c} $(DESTDIR)$(includedir)/oscar64/gfx
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/gfx/*.h $(project_dir)/include/gfx/*.c) $(DESTDIR)$(includedir)/oscar64/gfx
-	$(INSTALL_DATA) $(project_dir)/include/nes/*.{h,c} $(DESTDIR)$(includedir)/oscar64/nes
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/nes/*.h $(project_dir)/include/nes/*.c) $(DESTDIR)$(includedir)/oscar64/nes
-	$(INSTALL_DATA) $(project_dir)/include/opp/*.{h,cpp} $(DESTDIR)$(includedir)/oscar64/opp
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/opp/*.h $(project_dir)/include/opp/*.cpp) $(DESTDIR)$(includedir)/oscar64/opp
-	$(INSTALL_DATA) $(project_dir)/include/plus4/*.{h,c} $(DESTDIR)$(includedir)/oscar64/plus4
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/plus4/*.h $(project_dir)/include/plus4/*.c) $(DESTDIR)$(includedir)/oscar64/plus4
-	$(INSTALL_DATA) $(project_dir)/include/vic20/*.{h,c} $(DESTDIR)$(includedir)/oscar64/vic20
+	$(INSTALL_DATA) $(wildcard $(project_dir)/include/vic20/*.h $(project_dir)/include/vic20/*.c) $(DESTDIR)$(includedir)/oscar64/vic20
 
 
 uninstall:

oscar64.md

@@ -15,7 +15,7 @@ When I started this project, I was under the assumption that the 6502 would make
 
 The original plan was thus to compile to an intermediate language that would be interpreted, offering the 16bit support without the added code cost.  This turned out to work great from a code size perspective, but the performance was poor.  So I decided to give compiling to native code a shot.
 
-Most of the 6502 limitations can be overcome by carefull code analysis and optimizations.  
+Most of the 6502 limitations can be overcome by careful code analysis and optimizations.  
 
 * A second data stack relieves the CPU stack. 
 * Static call graph analysis removes the need for a stack completely
@@ -28,7 +28,7 @@ In the end, it turned out that the native code is not only significantly faster
  
 ## Limits and Errors
 
-There are still several open areas, but most targets have been reached.  The current Dhrystone performance is 94 iterations per second with byte code (10993) and 405 iterations with native code (9173 Bytes).  This clearly shows that Dhrystone is not a valid benchmark for optimizing compilers, because it puts the 6502 on par with a 4MHz 8088 or 68k, which it clearly is not.
+There are still several open areas, but most targets have been reached.  The current Dhrystone performance is 94 iterations per second with byte code (10993) and 442 iterations with native code (8952 Bytes).  This clearly shows that Dhrystone is not a valid benchmark for optimizing compilers, because it puts the 6502 on par with a 4MHz 8088 or 68k, which it clearly is not.
 
 ### Language
 
@@ -110,6 +110,7 @@ The compiler is command line driven, and creates an executable .prg file.
 * -v : verbose output for diagnostics
 * -v2 : more verbose output
 * -i : additional include paths
+* -ii : set default include path
 * -o : optional output file name
 * -rt : alternative runtime library, replaces the crt.c (or empty for none)
 * -e : execute the result in the integrated emulator
@@ -128,6 +129,8 @@ The compiler is command line driven, and creates an executable .prg file.
 * -Oa : optimize inline assembler (part of O2/O3)
 * -Oz : enable auto placement of global variables in zero page (part of O3)
 * -Op : optimize constant parameters
+* -Oo : optimize size using "outliner" (extract repeated code sequences into functions)
+* -Ox : optimize pointer arithmetic by blocking shorter arrays to not cross page boundaries
 * -g : create source level debug info and add source line numbers to asm listing
 * -gp : create source level debug info and add source line numbers to asm listing and static profile data
 * -tf : target format, may be prg, crt or bin
@@ -139,6 +142,7 @@ The compiler is command line driven, and creates an executable .prg file.
 * -xz : extended zero page usage, more zero page space, but no return to basic
 * -cid : cartridge type ID, used by vice emulator
 * -pp : compile in C++ mode
+* -strict : use strict ANSI C parsing (no C++ goodies)
 * -psci : use PETSCII encoding for all strings without prefix
 * -rmp : generate error files .error.map, .error.asm when linker fails
 
@@ -173,6 +177,7 @@ A list of source files can be provided.
 * nes_mmc3 : Nintendo entertainment system, MMC3, 512K PROM, 256K CROM
 * atari : Atari 8bit systems, (0x2000..0xbc00)
 * x16 : Commander X16, (0x0800..0x9f00)
+* mega65 : Mega 65, (0x2000..0xc000)
 
 ### C64 Cartridge formats
 
@@ -332,6 +337,17 @@ The .asm file is a great resource when debugging from within e.g. the VICE monit
 
 The compiler has various extensions to simplify developing for the C64.
 
+## Pragmas
+
+Warnings can be turned on or off using the warning pragma.  The scope of the pragma is currently global in most cases, so if it is turned off at some place, it is off everywhere.
+
+	#pragma warning(disable: 2000,2001)
+
+A message can be displayed during compilation with the message pragma
+
+	#pragma message("Hello User")
+	
+
 ## Embedding binary data
 
 The compiler supports the #embed preprocessor directive to import binary data.  It converts a section of an external binary file into a sequence of numbers that can be placed into an initializer of an array.
@@ -482,6 +498,8 @@ Set optimizer options that are active for the functions after it
 * maxinline : inline any function suitable
 * constparams : enable constant parameter folding into called functions
 * noconstparams : disable constant parameter folding into called functions
+* outline : enable outliner
+* nooutline : disable outliner
 * 0 : no optimization
 * 1 : default optimizations
 * 2 : aggressive optimizations
@@ -733,6 +751,11 @@ Regions can also be used to place assets such as character sets at fixed locatio
 
 The #pragma data(), #pragma code() and #pragma bss() control the placement of the generated objects into sections other than the default sections.
 
+A global variable or function can be aligned on a given power of two start with the align pragma.  This is most useful if a page crossing is problematic.  The compiler may also be able to generate more efficient code, if a larger variable is page aligned.
+
+	#pragma align(myvar, 8)
+	#pragma align(myfunc, 256)
+		
 ### Additional BSS sections
 
 Additional bss sections can be defined on request.
@@ -740,7 +763,7 @@ Additional bss sections can be defined on request.
 	#pragma section( mybss, 0, , , bss )
 	#pragma region( mybssregion, 0xc000, 0xd000, , , {mybss} )
 	
-They will not be cleared on startup, so all variables placed in these segments will be unitialized.  A common use for these sections is overlapping runtime variables with compressed load time assests that get expanded to their actual location during startup.
+They will not be cleared on startup, so all variables placed in these segments will be uninitialized.  A common use for these sections is overlapping runtime variables with compressed load time assests that get expanded to their actual location during startup.
 	
 ### Heap and Stack sections
 
@@ -807,7 +830,7 @@ With NROM mappers, the prg and chr code is put into cartridge bank zero.
 
 	#pragma data(data)
 
-The 32KByte of prg code starts at 0x8000 and goes upto 0xff80.  A startup section from 0xff80 to 0xfff9 is taken from the crt.c.
+The 32KByte of prg code starts at 0x8000 and goes up to 0xff80.  A startup section from 0xff80 to 0xfff9 is taken from the crt.c.
 
 A six byte boot section is placed from 0xfffa to 0xffff in all modes, which has to be populated by the game code with the appropriate pointers.
 
@@ -931,6 +954,11 @@ The compiler provides two levels of interrupt safe functions.  The specifier __i
 		return 0
 	}
 	
+### Assembler optimizer
+
+The compiler uses various optimizations on inline assembler that may not have the expected result in all cases.  The optimizer can be disabled for a range of code with e.g. #pragma optimize(noasm) or for an individual __asm statement by using __asm volatile {}.
+The assembler optimizer is enabled with optimization levels O2 and up.
+	
 # Helping the compiler optimizing
 
 The compiler does various optimization when compiling the code, but there are some rules to follow, things to avoid and hints to place to help the compiler to generate optimal code.

oscar64.sln

@@ -1,7 +1,7 @@
 
 Microsoft Visual Studio Solution File, Format Version 12.00
-# Visual Studio Version 16
+# Visual Studio Version 17
-VisualStudioVersion = 16.0.31624.102
+VisualStudioVersion = 17.13.35931.197
 MinimumVisualStudioVersion = 10.0.40219.1
 Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "oscar64", "oscar64\oscar64.vcxproj", "{1DBC623E-6109-41FE-B1BB-9B43FC984F7D}"
 EndProject
@@ -23,10 +23,14 @@ Global
 		{1DBC623E-6109-41FE-B1BB-9B43FC984F7D}.Release|x64.Build.0 = Release|x64
 		{1DBC623E-6109-41FE-B1BB-9B43FC984F7D}.Release|x86.ActiveCfg = Release|Win32
 		{1DBC623E-6109-41FE-B1BB-9B43FC984F7D}.Release|x86.Build.0 = Release|Win32
-		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Debug|x64.ActiveCfg = Debug
+		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Debug|x64.ActiveCfg = Debug|x64
-		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Debug|x86.ActiveCfg = Debug
+		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Debug|x64.Build.0 = Debug|x64
-		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Release|x64.ActiveCfg = Release
+		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Debug|x86.ActiveCfg = Debug|x86
-		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Release|x86.ActiveCfg = Release
+		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Debug|x86.Build.0 = Debug|x86
+		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Release|x64.ActiveCfg = Release|x64
+		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Release|x64.Build.0 = Release|x64
+		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Release|x86.ActiveCfg = Release|x86
+		{567F9C0F-5888-442A-BC9A-D7F9591E5AB4}.Release|x86.Build.0 = Release|x86
 	EndGlobalSection
 	GlobalSection(SolutionProperties) = preSolution
 		HideSolutionNode = FALSE

oscar64/Array.h

@@ -421,11 +421,13 @@ protected:
 			array = a2;
 		}
 
+		for (int i = size; i < to; i++) array[i] = T{};
+
 		size = to;
 	}
 
 	template<typename F>
-	void Parition(const F & f, int l, int r)
+	void Partition(const F & f, int l, int r)
 	{
 		if (r > l + 1)
 		{
@@ -442,8 +444,8 @@ protected:
 			}
 			array[pi] = p;
 
-			Parition(f, l, pi);
+			Partition(f, l, pi);
-			Parition(f, pi + 1, r);
+			Partition(f, pi + 1, r);
 		}
 	}
 public:
@@ -618,10 +620,22 @@ public:
 		return false;
 	}
 
+	void Fill(const T& t)
+	{
+		for (int i = 0; i < size; i++)
+			array[i] = t;
+	}
+
+	void Clear(void)
+	{
+		for (int i = 0; i < size; i++)
+			array[i] = T{};
+	}
+
 	template<typename F>
 	void Sort(const F & f)
 	{
-		Parition(f, 0, size);
+		Partition(f, 0, size);
 	}
 
 	__forceinline T& operator[](int n)

oscar64/Assembler.cpp

@@ -293,7 +293,7 @@ const char* AsmInstructionNames[NUM_ASM_INS_TYPES] = {
 	"INV", "BYT"
 };
 
-int AsmInsModeSize[NUM_ASM_INS_MODES] = {
+int AsmInsModeSize[NUM_ASM_INS_MODES_X] = {
 	1,
 	2,
 	2,
@@ -306,6 +306,8 @@ int AsmInsModeSize[NUM_ASM_INS_MODES] = {
 	2,
 	2,
 	2,
+	0,
+	2
 };
 
 void InitAssembler(void)

oscar64/Assembler.h

@@ -31,7 +31,9 @@ enum AsmInsMode
 
 	NUM_ASM_INS_MODES,
 
-	ASMIM_IMMEDIATE_ADDRESS
+	ASMIM_IMMEDIATE_ADDRESS,
+
+	NUM_ASM_INS_MODES_X,
 };
 
 struct AsmInsData
@@ -45,6 +47,8 @@ extern AsmInsData	DecInsData[256];
 
 extern short AsmInsOpcodes[NUM_ASM_INS_TYPES][NUM_ASM_INS_MODES];
 
+extern int AsmInsModeSize[NUM_ASM_INS_MODES_X];
+
 extern const char* AsmInstructionNames[NUM_ASM_INS_TYPES];
 
 AsmInsType FindAsmInstruction(const char * ins);

oscar64/BitVector.h

@@ -3,6 +3,7 @@
 #include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
+#include <assert.h>
 
 class BitVector
 {

oscar64/ByteCodeGenerator.cpp

@@ -3461,11 +3461,14 @@ ByteCode ByteCodeBasicBlock::RelationalOperator(InterCodeProcedure* proc, const
 	else
 	{
 		if (ins->mSrc[1].mTemp < 0)
+		{
+			if (ins->mSrc[1].mType == IT_INT16 || ins->mSrc[1].mMemory == IM_ABSOLUTE)
 			{
 				ByteCodeInstruction	lins(BC_LOAD_REG_16);
 				lins.mRegister = BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp];
 				lins.mRegisterFinal = ins->mSrc[0].mFinal;
 				mIns.Push(lins);
+
 				if (csigned)
 				{
 					ByteCodeInstruction	cins(BC_BINOP_CMPSI_16);
@@ -3479,7 +3482,36 @@ ByteCode ByteCodeBasicBlock::RelationalOperator(InterCodeProcedure* proc, const
 					mIns.Push(cins);
 				}
 			}
+			else
+			{
+				ByteCodeInstruction	bins(BC_LEA_ABS);
+				bins.mRegister = BC_REG_ACCU;
+				bins.mLinkerObject = ins->mSrc[1].mLinkerObject;
+				bins.mValue = int(ins->mSrc[1].mIntConst);
+				bins.mRelocate = true;
+				mIns.Push(bins);
+
+				if (csigned)
+				{
+					ByteCodeInstruction	cins(BC_BINOP_CMPSR_16);
+					cins.mRegister = BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp];
+					cins.mRegisterFinal = ins->mSrc[0].mFinal;
+					mIns.Push(cins);
+				}
+				else
+				{
+					ByteCodeInstruction	cins(BC_BINOP_CMPUR_16);
+					cins.mRegister = BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp];
+					cins.mRegisterFinal = ins->mSrc[0].mFinal;
+					mIns.Push(cins);
+				}
+
+				code = TransposeBranchCondition(code);
+			}
+		}
 		else if (ins->mSrc[0].mTemp < 0)
+		{
+			if (ins->mSrc[0].mType == IT_INT16 || ins->mSrc[0].mMemory == IM_ABSOLUTE)
 			{
 				ByteCodeInstruction	lins(BC_LOAD_REG_16);
 				lins.mRegister = BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp];
@@ -3552,6 +3584,31 @@ ByteCode ByteCodeBasicBlock::RelationalOperator(InterCodeProcedure* proc, const
 				code = TransposeBranchCondition(code);
 			}
 			else
+			{
+				ByteCodeInstruction	bins(BC_LEA_ABS);
+				bins.mRegister = BC_REG_ACCU;
+				bins.mLinkerObject = ins->mSrc[0].mLinkerObject;
+				bins.mValue = int(ins->mSrc[0].mIntConst);
+				bins.mRelocate = true;
+				mIns.Push(bins);
+
+				if (csigned)
+				{
+					ByteCodeInstruction	cins(BC_BINOP_CMPSR_16);
+					cins.mRegister = BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp];
+					cins.mRegisterFinal = ins->mSrc[1].mFinal;
+					mIns.Push(cins);
+				}
+				else
+				{
+					ByteCodeInstruction	cins(BC_BINOP_CMPUR_16);
+					cins.mRegister = BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp];
+					cins.mRegisterFinal = ins->mSrc[1].mFinal;
+					mIns.Push(cins);
+				}
+			}
+		}
+		else
 		{
 			ByteCodeInstruction	lins(BC_LOAD_REG_16);
 			lins.mRegister = BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp];

oscar64/Compiler.cpp

@@ -174,13 +174,13 @@ void Compiler::RegisterRuntime(const Location & loc, const Ident* ident)
 		if (bcdec->mType == DT_CONST_ASSEMBLER)
 		{
 			if (!bcdec->mLinkerObject)
-				mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec->mValue, nullptr);
+				mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec, nullptr);
 			linkerObject = bcdec->mLinkerObject;
 		}
 		else if (bcdec->mType == DT_LABEL)
 		{
 			if (!bcdec->mBase->mLinkerObject)
-				mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec->mBase->mValue, nullptr);
+				mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec->mBase, nullptr);
 
 			linkerObject = bcdec->mBase->mLinkerObject;
 			offset = int(bcdec->mInteger);
@@ -466,7 +466,7 @@ void Compiler::CompileProcedure(InterCodeProcedure* proc)
 				printf("Generate native code <%s>\n", proc->mIdent->mString);
 
 			ncproc->Compile(proc);
-			mNativeProcedures.Push(ncproc);
+			mNativeCodeGenerator->mProcedures.Push(ncproc);
 		}
 		else
 		{
@@ -508,6 +508,8 @@ bool Compiler::GenerateCode(void)
 			regionZeroPage = mLinker->AddRegion(identZeroPage, 0x00e0, 0x00ff);
 		else if (mCompilerOptions & (COPT_EXTENDED_ZERO_PAGE | COPT_TARGET_NES))
 			regionZeroPage = mLinker->AddRegion(identZeroPage, 0x0080, 0x00ff);
+		else if (mTargetMachine == TMACH_PET_8K || mTargetMachine == TMACH_PET_16K || mTargetMachine == TMACH_PET_32K)
+			regionZeroPage = mLinker->AddRegion(identZeroPage, 0x00ed, 0x00f7);
 		else
 			regionZeroPage = mLinker->AddRegion(identZeroPage, 0x00f7, 0x00ff);
 	}
@@ -521,6 +523,12 @@ bool Compiler::GenerateCode(void)
 		{
 			switch (mTargetMachine)
 			{
+			case TMACH_MEGA65:
+				if (mCompilerOptions & COPT_NATIVE)
+					regionStartup = mLinker->AddRegion(identStartup, 0x2001, 0x2080);
+				else
+					regionStartup = mLinker->AddRegion(identStartup, 0x2001, 0x2100);
+				break;
 			case TMACH_C64:
 			case TMACH_X16:
 				if (mCompilerOptions & COPT_NATIVE)
@@ -623,6 +631,9 @@ bool Compiler::GenerateCode(void)
 		{
 			switch (mTargetMachine)
 			{
+			case TMACH_MEGA65:
+				regionBytecode = mLinker->AddRegion(identBytecode, 0x2100, 0x2200);
+				break;
 			case TMACH_C64:
 			case TMACH_X16:
 				regionBytecode = mLinker->AddRegion(identBytecode, 0x0900, 0x0a00);
@@ -684,6 +695,9 @@ bool Compiler::GenerateCode(void)
 			{
 				switch (mTargetMachine)
 				{
+				case TMACH_MEGA65:
+					regionMain = mLinker->AddRegion(identMain, 0x2300, 0xc000);
+					break;
 				case TMACH_C64:
 					regionMain = mLinker->AddRegion(identMain, 0x0a00, 0xa000);
 					break;
@@ -735,6 +749,14 @@ bool Compiler::GenerateCode(void)
 			{
 				switch (mTargetMachine)
 				{
+				case TMACH_MEGA65:
+					// TODO: Disable M65 cartridges for now.
+					//
+					// if (mCompilerOptions & (COPT_TARGET_CRT8 | COPT_TARGET_CRT16))
+					// 	regionMain = mLinker->AddRegion(identMain, 0x2666, 0xff00);
+					// else
+					regionMain = mLinker->AddRegion(identMain, 0x2080, 0xc000);
+					break;
 				case TMACH_C64:
 
 					if (mCompilerOptions & (COPT_TARGET_CRT8 | COPT_TARGET_CRT16))
@@ -910,7 +932,7 @@ bool Compiler::GenerateCode(void)
 	if (mCompilerOptions & COPT_VERBOSE)
 		printf("Generate intermediate code\n");
 
-	mInterCodeGenerator->TranslateAssembler(mInterCodeModule, dcrtstart->mValue, nullptr);
+	mInterCodeGenerator->TranslateAssembler(mInterCodeModule, dcrtstart, nullptr);
 	for (int i = 0; i < mCompilationUnits->mReferenced.Size(); i++)
 	{
 		Declaration* dec = mCompilationUnits->mReferenced[i];
@@ -993,7 +1015,7 @@ bool Compiler::GenerateCode(void)
 			if (bcdec->mType == DT_CONST_ASSEMBLER)
 			{
 				if (!bcdec->mLinkerObject)
-					mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec->mValue, nullptr);
+					mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec, nullptr);
 				mByteCodeGenerator->mExtByteCodes[i] = bcdec->mLinkerObject;
 			}
 		}
@@ -1036,13 +1058,18 @@ bool Compiler::GenerateCode(void)
 			mCompilationUnits->mSectionStack->mSections.Push(proc->mLinkerObject->mStackSection);
 	}
 
+	mNativeCodeGenerator->OutlineFunctions();
+
 	mNativeCodeGenerator->BuildFunctionProxies();
 
-	for (int i = 0; i < mNativeProcedures.Size(); i++)
+	for (int i = 0; i < mNativeCodeGenerator->mProcedures.Size(); i++)
 	{
 		if (mCompilerOptions & COPT_VERBOSE2)
-			printf("Assemble native code <%s>\n", mNativeProcedures[i]->mInterProc->mIdent->mString);
+		{
-		mNativeProcedures[i]->Assemble();
+			if (mNativeCodeGenerator->mProcedures[i]->mInterProc)
+				printf("Assemble native code <%s>\n", mNativeCodeGenerator->mProcedures[i]->mInterProc->mIdent->mString);
+		}
+		mNativeCodeGenerator->mProcedures[i]->Assemble();
 	}
 
 	LinkerObject* byteCodeObject = nullptr;
@@ -1068,13 +1095,13 @@ bool Compiler::GenerateCode(void)
 					if (bcdec->mType == DT_CONST_ASSEMBLER)
 					{
 						if (!bcdec->mLinkerObject)
-							mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec->mValue, nullptr);
+							mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec, nullptr);
 						linkerObject = bcdec->mLinkerObject;
 					}
 					else if (bcdec->mType == DT_LABEL)
 					{
 						if (!bcdec->mBase->mLinkerObject)
-							mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec->mBase->mValue, nullptr);
+							mInterCodeGenerator->TranslateAssembler(mInterCodeModule, bcdec->mBase, nullptr);
 						linkerObject = bcdec->mBase->mLinkerObject;
 						offset = int(bcdec->mInteger);
 					}

oscar64/Compiler.h

@@ -29,7 +29,6 @@ public:
 	GlobalOptimizer* mGlobalOptimizer;
 
 	GrowingArray<ByteCodeProcedure*>	mByteCodeFunctions;
-	ExpandingArray<NativeCodeProcedure*>	mNativeProcedures;
 
 	TargetMachine	mTargetMachine;
 	uint64			mCompilerOptions;

oscar64/CompilerTypes.h

@@ -14,6 +14,8 @@ static const uint64 COPT_OPTIMIZE_AUTO_ZEROPAGE = 1ULL << 8;
 static const uint64 COPT_OPTIMIZE_CONST_PARAMS = 1ULL << 9;
 static const uint64 COPT_OPTIMIZE_MERGE_CALLS = 1ULL << 10;
 static const uint64 COPT_OPTIMIZE_GLOBAL = 1ULL << 11;
+static const uint64 COPT_OPTIMIZE_OUTLINE = 1ULL << 12;
+static const uint64 COPT_OPTIMIZE_PAGE_CROSSING = 1ULL << 13;
 
 static const uint64 COPT_OPTIMIZE_CODE_SIZE = 1ULL << 16;
 static const uint64 COPT_NATIVE = 1ULL << 17;
@@ -42,6 +44,7 @@ static const uint64 COPT_PETSCII = 1ULL << 53;
 static const uint64 COPT_ERROR_FILES = 1ULL << 54;
 
 static const uint64 COPT_PROFILEINFO = 1ULL << 55;
+static const uint64 COPT_STRICT = 1ULL << 56;
 
 
 
@@ -49,7 +52,7 @@ static const uint64 COPT_DEFAULT = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE |
 
 static const uint64 COPT_OPTIMIZE_DEFAULT = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS;
 
-static const uint64 COPT_OPTIMIZE_SIZE = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_AUTO_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_CODE_SIZE | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL;
+static const uint64 COPT_OPTIMIZE_SIZE = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_AUTO_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_CODE_SIZE | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL;// | COPT_OPTIMIZE_OUTLINE;
 
 static const uint64 COPT_OPTIMIZE_SPEED = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_AUTO_INLINE | COPT_OPTIMIZE_AUTO_UNROLL | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_ASSEMBLER | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL;
 
@@ -76,7 +79,8 @@ enum TargetMachine
 	TMACH_NES_MMC1,
 	TMACH_NES_MMC3,
 	TMACH_ATARI,
-	TMACH_X16
+	TMACH_X16,
+	TMACH_MEGA65
 };
 
 

oscar64/Compression.cpp

@@ -11,7 +11,7 @@ int CompressLZO(uint8* dst, const uint8* source, int size)
 		while (pi < 127 && pos < size)
 		{
 			int	bi = pi, bj = 0;
-			for (int i = 1; i < (pos < 255 ? pos : 255); i++)
+			for (int i = 1; i <= (pos < 255 ? pos : 255); i++)
 			{
 				int j = 0;
 				while (j < 127 && pos + j < size && source[pos - i + j] == source[pos + j])

oscar64/Constexpr.cpp

@@ -651,6 +651,7 @@ Expression* ConstexprInterpreter::EvalConstructor(Expression* exp)
 {
 	mProcType = exp->mLeft->mDecType;
 
+	Expression* pex = exp->mRight;
 	Declaration* cdec = exp->mLeft->mDecType->mParams;
 
 	int	pos = 0;
@@ -659,6 +660,28 @@ Expression* ConstexprInterpreter::EvalConstructor(Expression* exp)
 	mParams[pos].PutPtr(Value(&mResult));
 	pos = 2;
 
+	if (pex->mType == EX_LIST)
+		pex = pex->mRight;
+	else
+		pex = nullptr;
+	cdec = cdec->mNext;
+
+	while (pex && pex->mType == EX_LIST)
+	{
+		if (!AddParam(pos, pex->mLeft, cdec))
+			return exp;
+
+		pex = pex->mRight;
+		if (cdec)
+			cdec = cdec->mNext;
+	}
+
+	if (pex)
+	{
+		if (!AddParam(pos, pex, cdec))
+			return exp;
+	}
+
 	mHeap = new ExpandingArray<Value*>();
 
 	Execute(exp->mLeft->mDecValue->mValue);
@@ -754,6 +777,9 @@ bool ConstexprInterpreter::AddParam(int& pos, Expression* pex, Declaration* dec)
 
 Expression* ConstexprInterpreter::EvalCall(Expression* exp)
 {
+	if (!exp->mLeft->mDecValue || !exp->mLeft->mDecValue->mValue)
+		return exp;
+
 	mProcType = exp->mLeft->mDecType;
 
 	Expression* pex = exp->mRight;
@@ -1020,6 +1046,9 @@ ConstexprInterpreter::Value ConstexprInterpreter::EvalUnary(Expression* exp, con
 			break;
 		case TK_MUL:
 			return vl.GetPtr();
+		case TK_SIZEOF:
+			v.PutInt(vl.mDecType->mSize);
+			break;
 		default:
 			mErrors->Error(exp->mLocation, EERR_INCOMPATIBLE_OPERATOR, "Incompatible operator", TokenNames[exp->mToken]);
 		}
@@ -1168,6 +1197,26 @@ ConstexprInterpreter::Value ConstexprInterpreter::EvalCall(Expression* exp, Cons
 			mResult = Value(exp->mLocation, TheFloatTypeDeclaration);
 			mResult.PutFloat(::exp(mParams[0].GetFloat()));
 		}
+		else if (!strcmp(iname->mString, "sqrt"))
+		{
+			mResult = Value(exp->mLocation, TheFloatTypeDeclaration);
+			mResult.PutFloat(::sqrt(mParams[0].GetFloat()));
+		}
+		else if (!strcmp(iname->mString, "atan"))
+		{
+			mResult = Value(exp->mLocation, TheFloatTypeDeclaration);
+			mResult.PutFloat(::atan(mParams[0].GetFloat()));
+		}
+		else if (!strcmp(iname->mString, "atan2"))
+		{
+			mResult = Value(exp->mLocation, TheFloatTypeDeclaration);
+			mResult.PutFloat(::atan2(mParams[0].GetFloat(), mParams[1].GetFloat()));
+		}
+		else if (!strcmp(iname->mString, "pow"))
+		{
+			mResult = Value(exp->mLocation, TheFloatTypeDeclaration);
+			mResult.PutFloat(::pow(mParams[0].GetFloat(), mParams[1].GetFloat()));
+		}
 		else
 			mErrors->Error(exp->mLeft->mDecValue->mLocation, EERR_OBJECT_NOT_FOUND, "Unknown intrinsic function", iname);
 	}

oscar64/Declaration.cpp

@@ -31,6 +31,17 @@ DeclarationScope::~DeclarationScope(void)
 	delete[] mHash;
 }
 
+DeclarationScope* DeclarationScope::Clone(void) const
+{
+	DeclarationScope* scope = new DeclarationScope(mParent, mLevel, mName);
+	for (int i = 0; i < mHashSize; i++)
+	{
+		if (mHash[i].mIdent)
+			scope->Insert(mHash[i].mIdent, mHash[i].mDec);
+	}
+	return scope;
+}
+
 const Ident* DeclarationScope::Mangle(const Ident* ident) const
 {
 	if (mName && ident)
@@ -145,7 +156,7 @@ void DeclarationScope::End(const Location& loc)
 }
 
 Expression::Expression(const Location& loc, ExpressionType type)
-	:	mLocation(loc), mEndLocation(loc), mType(type), mLeft(nullptr), mRight(nullptr), mConst(false), mDecType(nullptr), mDecValue(nullptr), mToken(TK_NONE)
+	:	mLocation(loc), mEndLocation(loc), mType(type), mLeft(nullptr), mRight(nullptr), mConst(false), mDecType(nullptr), mDecValue(nullptr), mToken(TK_NONE), mFlags(0)
 {
 	static uint32	gUID = 0;
 	mUID = gUID++;
@@ -263,6 +274,9 @@ void Expression::Dump(int ident) const
 	case EX_FOR:
 		printf("FOR");
 		break;
+	case EX_FORBODY:
+		printf("FORBODY");
+		break;
 	case EX_DO:
 		printf("DO");
 		break;
@@ -335,6 +349,9 @@ void Expression::Dump(int ident) const
 	case EX_LABEL:
 		printf("LABEL %s", mDecValue->mIdent->mString);
 		break;
+	case EX_AGGREGATE:
+		printf("AGGREGATE");
+		break;
 	}
 	printf("\n");
 
@@ -488,6 +505,63 @@ Expression* Expression::LogicInvertExpression(void)
 	}
 }
 
+void Expression::ReplaceVariable(Declaration* pvar, Declaration* nvar)
+{
+	if (mLeft) mLeft->ReplaceVariable(pvar, nvar);
+	if (mRight) mRight->ReplaceVariable(pvar, nvar);
+
+	if (mType == EX_VARIABLE && mDecValue == pvar)
+		mDecValue = nvar;
+}
+
+Expression* Expression::ToAlternateThis(Declaration* pthis, Declaration* nthis)
+{
+	Expression* left = mLeft ? mLeft->ToAlternateThis(pthis, nthis) : nullptr;
+	Expression* right = mRight ? mRight->ToAlternateThis(pthis, nthis) : nullptr;
+	Declaration* decType = mDecType, * decValue = mDecValue;
+
+	Declaration* lp = pthis, * np = nthis;
+	while (lp)
+	{
+		if (decType == lp->mBase)
+			decType = np->mBase;
+		else if (decType == lp->mBase->mBase)
+			decType = np->mBase->mBase;
+		if (decValue == lp)
+			decValue = np;
+
+		lp = lp->mNext;
+		np = np->mNext;
+	}
+
+	if (mType == EX_QUALIFY && mLeft->mDecType != left->mDecType)
+	{
+		Declaration* pe = mLeft->mDecType->mParams, * ne = left->mDecType->mParams;
+		while (pe && ne && pe != mDecValue)
+		{
+			pe = pe->mNext;
+			ne = ne->mNext;
+		}
+		decValue = ne;
+		decType = ne->mBase;
+	}
+
+	if (left == mLeft && right == mRight && decType == mDecType && decValue == mDecValue)
+		return this;
+	else
+	{
+		Expression* nexp = new Expression(mLocation, mType);
+		nexp->mLeft = left;
+		nexp->mRight = right;
+		nexp->mDecType = decType;
+		nexp->mDecValue = decValue;
+		nexp->mToken = mToken;
+		nexp->mConst = mConst;
+
+		return nexp;
+	}
+}
+
 Expression* Expression::ConstantDereference(Errors* errors, LinkerSection* dataSection)
 {
 	if (mType == EX_VARIABLE)
@@ -605,6 +679,19 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 
 		return this;
 	}
+	else if (mType == EX_PREFIX && mToken == TK_SIZEOF)
+	{
+		if (mLeft->mDecType->mFlags & DTF_DEFINED)
+		{
+			Expression* ex = new Expression(mLocation, EX_CONSTANT);
+			Declaration* dec = new Declaration(mLocation, DT_CONST_INTEGER);
+			dec->mBase = TheSignedIntTypeDeclaration;
+			dec->mInteger = mLeft->mDecType->mSize;
+			ex->mDecValue = dec;
+			ex->mDecType = dec->mBase;
+			return ex;
+		}
+	}
 	else if (mType == EX_PREFIX && mLeft->mType == EX_CONSTANT)
 	{
 
@@ -688,14 +775,29 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 		ex->mDecType = mDecType;
 		return ex;
 	}
-	else if (mType == EX_PREFIX && mToken == TK_BINARY_AND && mLeft->mType == EX_INDEX && mLeft->mLeft->mType == EX_VARIABLE && (mLeft->mLeft->mDecValue->mFlags & (DTF_STATIC | DTF_GLOBAL)) && mLeft->mRight->mType == EX_CONSTANT)
+	else if (mType == EX_PREFIX && mToken == TK_BINARY_AND && mLeft->mType == EX_INDEX && 
+			 mLeft->mLeft->mType == EX_VARIABLE && mLeft->mLeft->mDecType->mType == DT_TYPE_ARRAY &&
+			 (mLeft->mLeft->mDecValue->mFlags & (DTF_STATIC | DTF_GLOBAL)) && mLeft->mRight->mType == EX_CONSTANT)
 	{
+		Declaration* vdec = mLeft->mLeft->mDecValue;
+
 		Expression* ex = new Expression(mLocation, EX_VARIABLE);
 		Declaration* dec = new Declaration(mLocation, DT_VARIABLE_REF);
-		dec->mFlags = mLeft->mLeft->mDecValue->mFlags;
+		
-		dec->mBase = mLeft->mLeft->mDecValue;
+		if (vdec->mType == DT_VARIABLE_REF)
+		{
+			dec->mFlags = vdec->mFlags;
+			dec->mBase = vdec->mBase;
+			dec->mSize = mLeft->mLeft->mDecType->mBase->mSize - int(mLeft->mRight->mDecValue->mInteger) * dec->mSize;
+			dec->mOffset = int(mLeft->mRight->mDecValue->mInteger) * dec->mSize + vdec->mOffset;
+		}
+		else
+		{
+			dec->mFlags = vdec->mFlags;
+			dec->mBase = vdec;
 			dec->mSize = mLeft->mLeft->mDecType->mBase->mSize - int(mLeft->mRight->mDecValue->mInteger) * dec->mSize;
 			dec->mOffset = int(mLeft->mRight->mDecValue->mInteger) * dec->mSize;
+		}
 		ex->mDecValue = dec;
 		ex->mDecType = mLeft->mLeft->mDecType;
 		return ex;
@@ -819,11 +921,19 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 		{
 			int64	ival = 0, ileft = mLeft->mDecValue->mInteger, iright = mRight->mDecValue->mInteger;
 
-			bool	signop =
+			Declaration* dtype = TheSignedIntTypeDeclaration;
-				(mLeft->mDecValue->mBase->mSize < 2 || (mLeft->mDecValue->mBase->mFlags & DTF_SIGNED)) &&
+			if (mLeft->mDecValue->mBase->mSize > mRight->mDecValue->mBase->mSize)
-				(mRight->mDecValue->mBase->mSize < 2 || (mRight->mDecValue->mBase->mFlags & DTF_SIGNED));
+				dtype = mLeft->mDecValue->mBase;
+			else if (mLeft->mDecValue->mBase->mSize < mRight->mDecValue->mBase->mSize)
+				dtype = mRight->mDecValue->mBase;
+			else if (mLeft->mDecValue->mBase->mSize > 1)
+			{
+				if ((mLeft->mDecValue->mBase->mFlags & DTF_SIGNED) && !(mRight->mDecValue->mBase->mFlags & DTF_SIGNED))
+					dtype = mRight->mDecValue->mBase;
+				else
+					dtype = mLeft->mDecValue->mBase;
+			}
 
-			bool	promote = true;
 			switch (mToken)
 			{
 			case TK_ADD:
@@ -837,27 +947,25 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 				break;
 			case TK_DIV:
 				if (iright == 0)
-					errors->Error(mLocation, EERR_INVALID_VALUE, "Constant division by zero");
+					return this;
-				else if (signop)
+				else if (dtype->mFlags & DTF_SIGNED)
 					ival = ileft / iright;
 				else
 					ival = (uint64)ileft / (uint64)iright;
 				break;
 			case TK_MOD:
 				if (iright == 0)
-					errors->Error(mLocation, EERR_INVALID_VALUE, "Constant division by zero");
+					return this;
-				else if (signop)
+				else if (dtype->mFlags & DTF_SIGNED)
 					ival = ileft % iright;
 				else
 					ival = (uint64)ileft % (uint64)iright;
 				break;
 			case TK_LEFT_SHIFT:
 				ival = ileft << iright;
-				promote = false;
 				break;
 			case TK_RIGHT_SHIFT:
 				ival = ileft >> iright;
-				promote = false;
 				break;
 			case TK_BINARY_AND:
 				ival = ileft & iright;
@@ -874,29 +982,14 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 
 			Expression* ex = new Expression(mLocation, EX_CONSTANT);
 			Declaration* dec = new Declaration(mLocation, DT_CONST_INTEGER);
-			if (promote)
+			dec->mBase = dtype;
-			{
+			dec->mInteger = signextend(ival, dec->mBase);
-				if (mLeft->mDecValue->mBase->mSize <= 2 && mRight->mDecValue->mBase->mSize <= 2)
-					dec->mBase = ival < 32768 ? TheSignedIntTypeDeclaration : TheUnsignedIntTypeDeclaration;
-				else
-					dec->mBase = ival < 2147483648 ? TheSignedLongTypeDeclaration : TheUnsignedLongTypeDeclaration;
-			}
-			else
-			{
-				if (mLeft->mDecValue->mBase->mSize < 2)
-					dec->mBase = TheSignedIntTypeDeclaration;
-				else
-					dec->mBase = mLeft->mDecValue->mBase;
-			}
-
-			dec->mInteger = ival;
 			ex->mDecValue = dec;
 			ex->mDecType = dec->mBase;
 			return ex;
 		}
 		else if ((mLeft->mDecValue->mType == DT_CONST_INTEGER || mLeft->mDecValue->mType == DT_CONST_FLOAT) && (mRight->mDecValue->mType == DT_CONST_INTEGER || mRight->mDecValue->mType == DT_CONST_FLOAT))
 		{
-
 			double	dval;
 			double	dleft = mLeft->mDecValue->mType == DT_CONST_INTEGER ? mLeft->mDecValue->mInteger : mLeft->mDecValue->mNumber;
 			double	dright = mRight->mDecValue->mType == DT_CONST_INTEGER ? mRight->mDecValue->mInteger : mRight->mDecValue->mNumber;
@@ -951,6 +1044,20 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 			ex->mDecType = dec->mBase;
 			return ex;
 		}
+		else if (mLeft->mDecValue->mType == DT_CONST_ADDRESS && mRight->mDecValue->mType == DT_CONST_ADDRESS && mToken == TK_SUB)
+		{
+			if (mLeft->mDecType->mType == DT_TYPE_POINTER && mRight->mDecType->mType == DT_TYPE_POINTER &&
+				mLeft->mDecType->mBase->IsConstSame(mRight->mDecType->mBase))
+			{
+				Expression* ex = new Expression(mLocation, EX_CONSTANT);
+				Declaration* dec = new Declaration(mLocation, DT_CONST_INTEGER);
+				dec->mBase = TheSignedIntTypeDeclaration;
+				dec->mInteger = (mLeft->mDecValue->mInteger - mRight->mDecValue->mInteger) / mLeft->mDecType->mBase->mSize;
+				ex->mDecValue = dec;
+				ex->mDecType = dec->mBase;
+				return ex;
+			}
+		}
 #if 0
 		else if (mLeft->mDecValue->mType == DT_CONST_POINTER && mRight->mDecValue->mType == DT_CONST_INTEGER && (mToken == TK_ADD || mToken == TK_SUB))
 		{
@@ -1031,6 +1138,26 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 
 		}
 	}
+	else if (mType == EX_LOGICAL_AND && mLeft->mType == EX_CONSTANT)
+	{
+		if (mLeft->mDecValue->mType == DT_CONST_INTEGER)
+		{
+			if (mLeft->mDecValue->mInteger == 0)
+				return mLeft;
+			else
+				return mRight->ConstantFold(errors, dataSection);
+		}
+	}
+	else if (mType == EX_LOGICAL_OR && mLeft->mType == EX_CONSTANT)
+	{
+		if (mLeft->mDecValue->mType == DT_CONST_INTEGER)
+		{
+			if (mLeft->mDecValue->mInteger != 0)
+				return mLeft;
+			else
+				return mRight->ConstantFold(errors, dataSection);
+		}
+	}
 	else if (mType == EX_CONDITIONAL && mLeft->mType == EX_CONSTANT)
 	{
 		if (mLeft->mDecValue->mType == DT_CONST_INTEGER)
@@ -1066,7 +1193,7 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 		return ex;
 	}
 	else if (mType == EX_BINARY && mToken == TK_ADD && mLeft->mType == EX_VARIABLE && mLeft->mDecValue->mType == DT_VARIABLE && (mLeft->mDecValue->mFlags & DTF_CONST) && 
-		mLeft->mDecValue && mRight->mDecValue && mLeft->mDecValue->mValue->mType == EX_CONSTANT &&
+		mLeft->mDecValue && mRight->mDecValue && mLeft->mDecValue->mValue && mLeft->mDecValue->mValue->mType == EX_CONSTANT &&
 		mLeft->mDecType->mType == DT_TYPE_POINTER && mRight->mType == EX_CONSTANT && mRight->mDecValue->mType == DT_CONST_INTEGER)
 	{
 		mLeft = mLeft->mDecValue->mValue;
@@ -1132,6 +1259,10 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 		ex->mDecType = mDecType;
 		return ex;
 	}
+	else if (mType == EX_VARIABLE && mDecType->IsSimpleType() && (mDecValue->mFlags & DTF_CONST) && mDecValue->mValue && mDecValue->mValue->mType == EX_INITIALIZATION && mDecValue->mValue->mRight->mType == EX_CONSTANT)
+	{
+		return mDecValue->mValue->mRight;
+	}
 	else if (mType == EX_CALL && mLeft->mType == EX_CONSTANT && (mLeft->mDecValue->mFlags & DTF_INTRINSIC) && mRight && mRight->mType == EX_CONSTANT)
 	{
 		Declaration* decf = mLeft->mDecValue, * decp = mRight->mDecValue;
@@ -1140,6 +1271,7 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 		if (decp->mType == DT_CONST_FLOAT || decp->mType == DT_CONST_INTEGER)
 		{
 			double d = decp->mType == DT_CONST_FLOAT ? decp->mNumber : decp->mInteger;
+			double e = 0.0;
 
 			bool	check = false;
 
@@ -1159,6 +1291,10 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 				d = log(d);
 			else if (!strcmp(iname->mString, "exp"))
 				d = exp(d);
+			else if (!strcmp(iname->mString, "sqrt"))
+				d = sqrt(d);
+			else if (!strcmp(iname->mString, "atan"))
+				d = atan(d);
 			else
 				return this;
 
@@ -1171,6 +1307,36 @@ Expression* Expression::ConstantFold(Errors * errors, LinkerSection * dataSectio
 			return ex;
 		}
 	}
+	else if (mType == EX_CALL && mLeft->mType == EX_CONSTANT && (mLeft->mDecValue->mFlags & DTF_INTRINSIC) && mRight && 
+		mRight->mType == EX_LIST && mRight->mLeft->mType == EX_CONSTANT && mRight->mRight->mType == EX_CONSTANT)
+	{
+		Declaration* decf = mLeft->mDecValue, * decp1 = mRight->mLeft->mDecValue, * decp2 = mRight->mRight->mDecValue;
+		const Ident* iname = decf->mQualIdent;
+
+		if ((decp1->mType == DT_CONST_FLOAT || decp1->mType == DT_CONST_INTEGER) &&
+			(decp2->mType == DT_CONST_FLOAT || decp2->mType == DT_CONST_INTEGER))
+		{
+			double d1 = decp1->mType == DT_CONST_FLOAT ? decp1->mNumber : decp1->mInteger;
+			double d2 = decp2->mType == DT_CONST_FLOAT ? decp2->mNumber : decp2->mInteger;
+
+			bool	check = false;
+
+			if (!strcmp(iname->mString, "pow"))
+				d1 = pow(d1, d2);
+			else if (!strcmp(iname->mString, "atan2"))
+				d1 = atan2(d1, d2);
+			else
+				return this;
+
+			Expression* ex = new Expression(mLocation, EX_CONSTANT);
+			Declaration* dec = new Declaration(mLocation, DT_CONST_FLOAT);
+			dec->mBase = TheFloatTypeDeclaration;
+			dec->mNumber = d1;
+			ex->mDecValue = dec;
+			ex->mDecType = dec->mBase;
+			return ex;
+		}
+	}
 	else if (mType == EX_CALL && mLeft->mType == EX_CONSTANT && (mLeft->mDecValue->mFlags & DTF_CONSTEXPR) && dataSection)
 	{
 		ConstexprInterpreter	cinter(mLocation, errors, dataSection);
@@ -1195,6 +1361,7 @@ Declaration::Declaration(const Location& loc, DecType type)
 	mConst(nullptr), mMutable(nullptr), mVolatile(nullptr),
 	mDefaultConstructor(nullptr), mDestructor(nullptr), mCopyConstructor(nullptr), mCopyAssignment(nullptr), mMoveConstructor(nullptr), mMoveAssignment(nullptr),
 	mVectorConstructor(nullptr), mVectorDestructor(nullptr), mVectorCopyConstructor(nullptr), mVectorCopyAssignment(nullptr),
+	mVectorMoveConstructor(nullptr), mVectorMoveAssignment(nullptr),
 	mVTable(nullptr), mTemplate(nullptr), mForwardParam(nullptr), mForwardCall(nullptr),
 	mVarIndex(-1), mLinkerObject(nullptr), mCallers(nullptr), mCalled(nullptr), mAlignment(1), mFriends(nullptr),
 	mInteger(0), mNumber(0), mMinValue(-0x80000000LL), mMaxValue(0x7fffffffLL), mFastCallBase(0), mFastCallSize(0), mStride(0), mStripe(1),
@@ -1216,7 +1383,7 @@ int Declaration::Stride(void) const
 {
 	if (mStride > 0)
 		return mStride;
-	else if (mBase)
+	else if (mBase && mBase->mType != DT_TYPE_VOID)
 		return mBase->mSize;
 	else
 		return 1;
@@ -1319,6 +1486,51 @@ Declaration* Declaration::ConstCast(Declaration* ntype)
 			pdec->mSize = 2;
 			return pdec;
 		}
+		else if (mType == DT_VARIABLE && mBase->mType == DT_TYPE_ARRAY)
+		{
+			Expression* ex = new Expression(mLocation, EX_VARIABLE);
+			ex->mDecType = mBase;
+			ex->mDecValue = this;
+
+			Declaration* pdec = this->Clone();
+			pdec->mType = DT_CONST_POINTER;
+			pdec->mValue = ex;
+			pdec->mBase = ntype;
+			pdec->mSize = 2;
+
+			return pdec;
+		}
+		else if (mType == DT_VARIABLE_REF)
+		{
+			Declaration* vtype = mBase->mBase;
+			while (vtype && vtype->mType == DT_TYPE_STRUCT)
+			{
+				Declaration* dec = vtype->mParams;
+				while (dec && dec->mOffset != mOffset)
+					dec = dec->mNext;
+				if (dec)
+					vtype = dec->mBase;
+				else
+					vtype = nullptr;
+			}
+
+			if (vtype && vtype->mType == DT_TYPE_ARRAY)
+			{
+				Expression* ex = new Expression(mLocation, EX_VARIABLE);
+				ex->mDecType = mBase->mBase;
+				ex->mDecValue = this;
+
+				Declaration* pdec = this->Clone();
+				pdec->mType = DT_CONST_POINTER;
+				pdec->mValue = ex;
+				pdec->mBase = ntype;
+				pdec->mSize = 2;
+
+				return pdec;
+			}
+			else
+				return this;
+		}
 		else
 			return this;
 	}
@@ -1464,7 +1676,9 @@ const Ident* Declaration::FullIdent(void)
 		Declaration* dec = mBase->mParams;
 		while (dec)
 		{
-			tident = tident->Mangle(dec->mBase->MangleIdent()->mString);
+			const Ident* mident = dec->mBase->MangleIdent();
+			if (mident)
+				tident = tident->Mangle(mident->mString);
 			dec = dec->mNext;
 			if (dec)
 				tident = tident->Mangle(",");
@@ -1600,26 +1814,31 @@ const Ident* Declaration::MangleIdent(void)
 					}
 
 					dec = dec->mNext;
-					if (dec)
+					if (mMangleIdent && dec)
 						mMangleIdent = mMangleIdent->Mangle(",");
 				}
 			}
 			else
 				mMangleIdent = Ident::Unique("void");
 		}
-		else
+		else if (mQualIdent)
 			mMangleIdent = mQualIdent;
+		else
+			mMangleIdent = mIdent;
 
 		if (mTemplate)
 		{
 
 		}
 
+		if (mMangleIdent)
+		{
 			if (mFlags & DTF_CONST)
 				mMangleIdent = mMangleIdent->PreMangle("const ");
 			if (mFlags & DTF_VOLATILE)
 				mMangleIdent = mMangleIdent->PreMangle("volatile ");
 		}
+	}
 
 	return mMangleIdent;
 }
@@ -1679,19 +1898,8 @@ Declaration* Declaration::ExpandTemplate(DeclarationScope* scope)
 	return this;
 }
 
-bool Declaration::ResolveTemplate(Expression* pexp, Declaration* tdec)
+bool Declaration::ResolveTemplateParameterList(Expression* pexp, Declaration* pdec, bool preliminary)
 {
-	Declaration* pdec = tdec->mBase->mParams;
-
-	// Insert partially resolved templates
-	Declaration* ptdec = tdec->mTemplate->mParams;
-	while (ptdec)
-	{
-		if (ptdec->mBase)
-			mScope->Insert(ptdec->mIdent, ptdec->mBase);
-		ptdec = ptdec->mNext;
-	}
-
 	Declaration* phead = nullptr, * ptail = nullptr;
 	int	pcnt = 0;
 
@@ -1727,7 +1935,7 @@ bool Declaration::ResolveTemplate(Expression* pexp, Declaration* tdec)
 			}
 			else
 			{
-				if (!ResolveTemplate(ex->mDecType, pdec->mBase))
+				if (!ResolveTemplate(ex->mDecType, pdec->mBase, false, preliminary))
 					return false;
 
 				pdec = pdec->mNext;
@@ -1741,6 +1949,9 @@ bool Declaration::ResolveTemplate(Expression* pexp, Declaration* tdec)
 	{
 		if (pdec->mType == DT_PACK_ARGUMENT)
 		{
+			if (preliminary)
+				return true;
+
 			Declaration* tpdec = new Declaration(pdec->mLocation, DT_PACK_TYPE);
 			if (pdec->mBase->mType == DT_TYPE_REFERENCE)
 				tpdec->mIdent = pdec->mBase->mBase->mIdent;
@@ -1754,6 +1965,25 @@ bool Declaration::ResolveTemplate(Expression* pexp, Declaration* tdec)
 			return false;
 	}
 
+	return true;
+}
+
+bool Declaration::ResolveTemplate(Expression* pexp, Declaration* tdec)
+{
+	Declaration* pdec = tdec->mBase->mParams;
+
+	// Insert partially resolved templates
+	Declaration* ptdec = tdec->mTemplate->mParams;
+	while (ptdec)
+	{
+		if (ptdec->mBase)
+			mScope->Insert(ptdec->mIdent, ptdec->mBase);
+		ptdec = ptdec->mNext;
+	}
+
+	if (!ResolveTemplateParameterList(pexp, pdec, true) || !ResolveTemplateParameterList(pexp, pdec, false))
+		return false;
+
 	Declaration* ppdec = nullptr;
 	ptdec = tdec->mTemplate->mParams;
 	while (ptdec)
@@ -1796,7 +2026,7 @@ bool Declaration::CanResolveTemplate(Expression* pexp, Declaration* tdec)
 
 		if (pdec)
 		{
-			if (!ResolveTemplate(ex->mDecType, pdec->mBase))
+			if (!ResolveTemplate(ex->mDecType, pdec->mBase, false, false))
 				return false;
 
 			if (pdec->mType != DT_PACK_ARGUMENT)
@@ -1812,19 +2042,19 @@ bool Declaration::CanResolveTemplate(Expression* pexp, Declaration* tdec)
 		return true;
 }
 
-bool Declaration::ResolveTemplate(Declaration* fdec, Declaration* tdec)
+bool Declaration::ResolveTemplate(Declaration* fdec, Declaration* tdec, bool same, bool preliminary)
 {
 	if (tdec->IsSame(fdec))
 		return true;
 	else if (fdec->IsReference())
-		return ResolveTemplate(fdec->mBase, tdec);
+		return ResolveTemplate(fdec->mBase, tdec, same, preliminary);
 	else if (tdec->mType == DT_TYPE_FUNCTION)
 	{
 		if (fdec->mType == DT_TYPE_FUNCTION)
 		{
 			if (fdec->mBase)
 			{
-				if (!tdec->mBase || !ResolveTemplate(fdec->mBase, tdec->mBase))
+				if (!tdec->mBase || !ResolveTemplate(fdec->mBase, tdec->mBase, true, preliminary))
 					return false;
 			}
 			else if (tdec->mBase)
@@ -1863,7 +2093,7 @@ bool Declaration::ResolveTemplate(Declaration* fdec, Declaration* tdec)
 				if (!fpdec)
 					return false;
 
-				if (!ResolveTemplate(fpdec->mBase, tpdec->mBase))
+				if (!ResolveTemplate(fpdec->mBase, tpdec->mBase, true, preliminary))
 					return false;
 
 				fpdec = fpdec->mNext;
@@ -1879,16 +2109,16 @@ bool Declaration::ResolveTemplate(Declaration* fdec, Declaration* tdec)
 	}
 	else if (tdec->mType == DT_TYPE_REFERENCE)
 	{
-		return ResolveTemplate(fdec, tdec->mBase);
+		return ResolveTemplate(fdec, tdec->mBase, true, preliminary);
 	}
 	else if (tdec->mType == DT_TYPE_RVALUEREF)
 	{
-		return ResolveTemplate(fdec, tdec->mBase);
+		return ResolveTemplate(fdec, tdec->mBase, true, preliminary);
 	}
 	else if (tdec->mType == DT_TYPE_POINTER)
 	{
 		if (fdec->mType == DT_TYPE_POINTER || fdec->mType == DT_TYPE_ARRAY)
-			return ResolveTemplate(fdec->mBase, tdec->mBase);
+			return ResolveTemplate(fdec->mBase, tdec->mBase, true, preliminary);
 		else
 			return false;
 	}
@@ -1908,10 +2138,12 @@ bool Declaration::ResolveTemplate(Declaration* fdec, Declaration* tdec)
 			if (pdec->mType == DT_TYPE_STRUCT)
 				pdec = pdec->mScope->Lookup(tdec->mIdent);
 		}
+		else if (preliminary && !same)
+			return true;
 		else
 			pdec = mScope->Insert(tdec->mIdent, fdec);
 
-		if (pdec && !pdec->IsSame(fdec))
+		if (pdec && !(same ? pdec->IsSame(fdec) : pdec->CanAssign(fdec)))
 			return false;
 
 		return true;
@@ -1952,6 +2184,21 @@ bool Declaration::ResolveTemplate(Declaration* fdec, Declaration* tdec)
 		}
 
 		return true;
+	}
+	else if (tdec->mType == DT_TYPE_ARRAY && fdec->mType == DT_TYPE_ARRAY && tdec->mTemplate && tdec->mBase->IsConstSame(fdec->mBase))
+	{
+		Declaration* ifdec = new Declaration(fdec->mLocation, DT_CONST_INTEGER);
+		ifdec->mBase = TheSignedIntTypeDeclaration;
+		ifdec->mSize = 2;
+		ifdec->mInteger = fdec->mSize / fdec->mBase->mSize;
+
+		Declaration * ipdec = mScope->Insert(tdec->mTemplate->mIdent, ifdec);
+		if (ipdec && !ipdec->IsSame(ifdec))
+			return false;
+
+		return true;
+
+
 	}
 	else if (tdec->mType == DT_TYPE_STRUCT && fdec->mType == DT_TYPE_STRUCT && tdec->mTemplate)
 	{
@@ -2017,13 +2264,27 @@ bool Declaration::ResolveTemplate(Declaration* fdec, Declaration* tdec)
 					tpdec = tpdec->mNext;
 				}
 
-				return !fpdec && !tpdec;
+				if (fpdec)
+					return false;
+				else if (!tpdec)
+					return true;
+				else if (tpdec->mBase->mType == DT_PACK_TEMPLATE)
+				{
+					Declaration*tppack = new Declaration(tpdec->mLocation, DT_PACK_TYPE);
+					Declaration* pdec = mScope->Insert(tpdec->mBase->mIdent, tppack);
+
+					return true;
+				}
+				else
+					return false;
 			}
 			
 			ftdec = ftdec->mNext;
 		}
 		return false;
 	}
+	else if (!same && tdec->CanAssign(fdec))
+		return true;
 	else
 		return tdec->IsSame(fdec);
 }
@@ -2036,6 +2297,8 @@ Declaration* Declaration::Clone(void)
 	ndec->mOffset = mOffset;
 	ndec->mStride = mStride;
 	ndec->mStripe = mStripe;
+	ndec->mBits = mBits;
+	ndec->mShift = mShift;
 	ndec->mBase = mBase;
 	ndec->mFlags = mFlags;
 	ndec->mScope = mScope;
@@ -2055,6 +2318,7 @@ Declaration* Declaration::Clone(void)
 	ndec->mMaxValue = mMaxValue;
 	ndec->mCompilerOptions = mCompilerOptions;
 	ndec->mParser = mParser;
+	ndec->mNumVars = mNumVars;
 
 	return ndec;
 }
@@ -2106,6 +2370,8 @@ Declaration* Declaration::ToStriped(int stripe)
 	ndec->mOffset = mOffset * stripe;
 	ndec->mStride = mStride;
 	ndec->mStripe = stripe;
+	ndec->mBits = mBits;
+	ndec->mShift = mShift;
 	ndec->mFlags = mFlags;
 	ndec->mIdent = mIdent;
 	ndec->mQualIdent = mQualIdent;
@@ -2133,6 +2399,24 @@ Declaration* Declaration::ToStriped(int stripe)
 			prev = pnec;
 			p = p->mNext;
 		}
+
+		Declaration* pndec = ndec->BuildPointer(mLocation);
+
+		mScope->Iterate([=](const Ident* ident, Declaration* vdec) {
+			if (vdec->mType == DT_CONST_FUNCTION)
+			{
+				ndec->mScope->Insert(ident, vdec->ToAlternateThis(pndec));
+			}
+		});
+
+		ndec->mDestructor = mDestructor ? mDestructor->ToAlternateThis(pndec) : nullptr;
+		ndec->mDefaultConstructor = mDefaultConstructor ? mDefaultConstructor->ToAlternateThis(pndec) : nullptr;
+		ndec->mCopyConstructor = mCopyConstructor ? mCopyConstructor->ToAlternateThis(pndec) : nullptr;
+		ndec->mMoveConstructor = mMoveConstructor ? mMoveConstructor->ToAlternateThis(pndec) : nullptr;
+		ndec->mVectorConstructor = mVectorConstructor ? mVectorConstructor->ToAlternateThis(pndec, 2) : nullptr;
+		ndec->mVectorDestructor = mVectorDestructor ? mVectorDestructor->ToAlternateThis(pndec, 2) : nullptr;
+		ndec->mVectorCopyConstructor = mVectorCopyConstructor ? mVectorCopyConstructor->ToAlternateThis(pndec, 2) : nullptr;
+		ndec->mVectorMoveConstructor = mVectorMoveConstructor ? mVectorMoveConstructor->ToAlternateThis(pndec, 2) : nullptr;
 	}
 	else if (mType == DT_TYPE_FUNCTION)
 	{
@@ -2163,6 +2447,8 @@ Declaration* Declaration::ToVolatileType(void)
 		ndec->mSize = mSize;
 		ndec->mStride = mStride;
 		ndec->mBase = mBase;
+		ndec->mBits = mBits;
+		ndec->mShift = mShift;
 		ndec->mFlags = mFlags | DTF_VOLATILE;
 		ndec->mScope = mScope;
 		ndec->mParams = mParams;
@@ -2198,6 +2484,7 @@ Declaration* Declaration::ToVolatileType(void)
 		ndec->mVectorConstructor = mVectorConstructor;
 		ndec->mVectorDestructor = mVectorDestructor;
 		ndec->mVectorCopyConstructor = mVectorCopyConstructor;
+		ndec->mVectorMoveConstructor = mVectorMoveConstructor;
 		ndec->mVTable = mVTable;
 
 		mVolatile = ndec;
@@ -2217,7 +2504,12 @@ Declaration* Declaration::ToConstType(void)
 		ndec->mSize = mSize;
 		ndec->mStride = mStride;
 		ndec->mStripe = mStripe;
+		if (mType == DT_TYPE_ARRAY)
+			ndec->mBase = mBase->ToConstType();
+		else
 			ndec->mBase = mBase;
+		ndec->mBits = mBits;
+		ndec->mShift = mShift;
 		ndec->mFlags = mFlags | DTF_CONST;
 		ndec->mScope = mScope;
 		ndec->mParams = mParams;
@@ -2232,6 +2524,7 @@ Declaration* Declaration::ToConstType(void)
 		ndec->mVectorConstructor = mVectorConstructor;
 		ndec->mVectorDestructor = mVectorDestructor;
 		ndec->mVectorCopyConstructor = mVectorCopyConstructor;
+		ndec->mVectorMoveConstructor = mVectorMoveConstructor;
 		ndec->mVTable = mVTable;
 
 		ndec->mMutable = this;
@@ -2241,6 +2534,35 @@ Declaration* Declaration::ToConstType(void)
 	return mConst;
 }
 
+Declaration* Declaration::ToAlternateThis(Declaration* pthis, int nthis)
+{
+	Declaration* ndec = this->Clone();
+
+	if (mType == DT_CONST_FUNCTION)
+	{
+		ndec->mBase = mBase->ToAlternateThis(pthis, nthis);
+		if (mValue)
+			ndec->mValue = mValue->ToAlternateThis(mBase->mParams, ndec->mBase->mParams);
+	}
+	else
+	{
+		Declaration* nparam = ndec->mParams->Clone();
+		Declaration* npp = nparam;
+		Declaration* kpp = ndec->mParams->mNext;
+		while (kpp)
+		{
+			npp->mNext = kpp->Clone();
+			npp = npp->mNext;
+			kpp = npp->mNext;
+		}
+		nparam->mBase = pthis;
+		if (nthis == 2)
+			nparam->mNext->mBase = pthis;
+		ndec->mParams = nparam;
+	}
+	return ndec;
+}
+
 Declaration* Declaration::ToMutableType(void)
 {
 	if (!(mFlags & DTF_CONST))
@@ -2253,7 +2575,9 @@ Declaration* Declaration::ToMutableType(void)
 		ndec->mStride = mStride;
 		ndec->mStripe = mStripe;
 		ndec->mBase = mBase;
-		ndec->mFlags = mFlags | DTF_CONST;
+		ndec->mBits = mBits;
+		ndec->mShift = mShift;
+		ndec->mFlags = mFlags & ~DTF_CONST;
 		ndec->mScope = mScope;
 		ndec->mParams = mParams;
 		ndec->mIdent = mIdent;
@@ -2267,6 +2591,7 @@ Declaration* Declaration::ToMutableType(void)
 		ndec->mVectorConstructor = mVectorConstructor;
 		ndec->mVectorDestructor = mVectorDestructor;
 		ndec->mVectorCopyConstructor = mVectorCopyConstructor;
+		ndec->mVectorMoveConstructor = mVectorMoveConstructor;
 		ndec->mVTable = mVTable;
 
 		ndec->mConst = this;
@@ -2596,6 +2921,9 @@ bool Declaration::IsSame(const Declaration* dec) const
 	if (mType != dec->mType)
 		return false;
 
+	if ((mFlags & (DTF_SIGNED | DTF_CONST | DTF_VOLATILE)) != (dec->mFlags & (DTF_SIGNED | DTF_CONST | DTF_VOLATILE)))
+		return false;
+
 	if (!(mFlags & dec->mFlags & DTF_DEFINED) && mType == DT_TYPE_STRUCT)
 		return mIdent == dec->mIdent;
 
@@ -2604,9 +2932,6 @@ bool Declaration::IsSame(const Declaration* dec) const
 	if (mStripe != dec->mStripe)
 		return false;
 
-	if ((mFlags & (DTF_SIGNED | DTF_CONST | DTF_VOLATILE)) != (dec->mFlags & (DTF_SIGNED | DTF_CONST | DTF_VOLATILE)))
-		return false;
-
 	if (mType == DT_CONST_INTEGER)
 		return mInteger == dec->mInteger;
 	else if (mType == DT_TYPE_INTEGER)
@@ -2653,7 +2978,7 @@ bool Declaration::IsSame(const Declaration* dec) const
 
 		return true;
 	}
-	else if (mType == DT_TYPE_TEMPLATE)
+	else if (mType == DT_TYPE_TEMPLATE || mType == DT_PACK_TEMPLATE)
 	{
 		return mIdent == dec->mIdent;
 	}
@@ -2786,7 +3111,7 @@ bool Declaration::CanAssign(const Declaration* fromType) const
 		if (mScope == fromType->mScope || (mIdent == fromType->mIdent && mSize == fromType->mSize))
 			return true;
 		if (fromType->mBase)
-			return this->CanAssign(fromType->mBase);
+			return this->CanAssign(fromType->mBase->mBase);
 		return false;
 	}
 	else if (mType == DT_TYPE_ARRAY && fromType->mType == DT_TYPE_ARRAY)
@@ -2823,6 +3148,10 @@ bool Declaration::CanAssign(const Declaration* fromType) const
 		{
 			return true;
 		}
+		else if (mBase->mType == DT_TYPE_FUNCTION && fromType->mType == DT_TYPE_ASSEMBLER)
+		{
+			return (mBase->mBase->mType == DT_TYPE_VOID && !mBase->mBase->mParams);
+		}
 	}
 
 	return false;
@@ -2849,6 +3178,52 @@ bool Declaration::IsIndexed(void) const
 	return mType == DT_TYPE_ARRAY || mType == DT_TYPE_POINTER;
 }
 
+bool Declaration::ContainsArray(void) const
+{
+	if (mType == DT_TYPE_ARRAY)
+		return true;
+	else if (mType == DT_TYPE_STRUCT)
+	{
+		Declaration* p = mParams;
+		while (p)
+		{
+			if (p->mType == DT_ELEMENT && p->mBase->ContainsArray())
+				return true;
+			p = p->mNext;
+		}
+	}
+	return false;
+}
+
+bool Declaration::IsShortIntStruct(void) const
+{
+	if (mType == DT_TYPE_STRUCT && mSize <= 4 && !mDestructor && !mCopyConstructor)
+	{
+		Declaration* em = mParams;
+		while (em)
+		{
+			if (em->mType == DT_ELEMENT && em->mBase->IsIntegerType() && em->mBits == 0)
+				em = em->mNext;
+			else
+				return false;
+		}
+
+//		if (strcmp(mIdent->mString, "connection")) return false;
+		return true;
+	}
+
+	return false;
+}
+
+bool Declaration::HasConstructor(void) const
+{
+	return mType == DT_TYPE_STRUCT && mScope && mScope->Lookup(mIdent->PreMangle("+"));
+}
+
+bool Declaration::IsComplexStruct(void) const
+{
+	return mType == DT_TYPE_STRUCT && !IsShortIntStruct();
+}
 
 bool Declaration::IsSimpleType(void) const
 {
@@ -2868,14 +3243,17 @@ Declaration* TheVoidFunctionTypeDeclaration, * TheConstVoidValueDeclaration;
 Declaration* TheCharPointerTypeDeclaration, * TheConstCharPointerTypeDeclaration;
 Expression* TheVoidExpression;
 Declaration* TheNullptrConstDeclaration, * TheZeroIntegerConstDeclaration, * TheZeroFloatConstDeclaration, * TheNullPointerTypeDeclaration;
+Declaration* TheTrueConstDeclaration, * TheFalseConstDeclaration;
 
 void InitDeclarations(void)
 {
 	static Location	noloc;
 	TheVoidTypeDeclaration = new Declaration(noloc, DT_TYPE_VOID);
+	TheVoidTypeDeclaration->mSize = 0;
 	TheVoidTypeDeclaration->mFlags = DTF_DEFINED;
 
 	TheConstVoidTypeDeclaration = new Declaration(noloc, DT_TYPE_VOID);
+	TheConstVoidTypeDeclaration->mSize = 0;
 	TheConstVoidTypeDeclaration->mFlags = DTF_DEFINED | DTF_CONST;
 
 	TheVoidPointerTypeDeclaration = new Declaration(noloc, DT_TYPE_POINTER);
@@ -2967,4 +3345,13 @@ void InitDeclarations(void)
 	TheZeroFloatConstDeclaration->mBase = TheFloatTypeDeclaration;
 	TheZeroFloatConstDeclaration->mSize = 4;
 
+	TheTrueConstDeclaration = new Declaration(noloc, DT_CONST_INTEGER);
+	TheTrueConstDeclaration->mBase = TheBoolTypeDeclaration;
+	TheTrueConstDeclaration->mSize = 1;
+	TheTrueConstDeclaration->mInteger = 1;
+
+	TheFalseConstDeclaration = new Declaration(noloc, DT_CONST_INTEGER);
+	TheFalseConstDeclaration->mBase = TheBoolTypeDeclaration;
+	TheFalseConstDeclaration->mSize = 1;
+	TheFalseConstDeclaration->mInteger = 0;
 }

oscar64/Declaration.h

@@ -127,6 +127,7 @@ static const uint64 DTF_FPARAM_UNUSED	= (1ULL << 49);
 static const uint64 DTF_DEPRECATED		= (1ULL << 50);
 static const uint64 DTF_FUNC_NO_RETURN	= (1ULL << 51);
 static const uint64 DTF_PLACED			= (1ULL << 52);
+static const uint64 DTF_NO_PAGE_CROSS	= (1ULL << 53);
 
 
 
@@ -166,6 +167,8 @@ public:
 	ScopeLevel		mLevel;
 	const Ident	*	mName;
 
+	DeclarationScope* Clone(void) const;
+
 	DeclarationScope* mParent;
 protected:
 	struct Entry
@@ -215,6 +218,7 @@ enum ExpressionType
 	EX_IF,
 	EX_ELSE,
 	EX_FOR,
+	EX_FORBODY,
 	EX_DO,
 	EX_SCOPE,
 	EX_BREAK,
@@ -238,9 +242,16 @@ enum ExpressionType
 	EX_PACK,
 	EX_PACK_TYPE,
 	EX_LABEL,
-	EX_GOTO
+	EX_GOTO,
+	EX_AGGREGATE
 };
 
+static const uint32 ANAFL_LHS = (1U << 0);
+static const uint32 ANAFL_RHS = (1U << 1);
+static const uint32 ANAFL_ASSIGN = (1U << 2);
+static const uint32 ANAFL_ALIAS = (1U << 3);
+
+
 class Expression
 {
 public:
@@ -257,12 +268,16 @@ public:
 	AsmInsType				mAsmInsType;
 	AsmInsMode				mAsmInsMode;
 	bool					mConst;
+	uint32					mFlags;
 
 	Expression* LogicInvertExpression(void);
 	Expression* ConstantFold(Errors * errors, LinkerSection* dataSection, Linker * linker = nullptr);
 	Expression* ConstantDereference(Errors* errors, LinkerSection* dataSection);
 	bool HasSideEffects(void) const;
 	Expression* ListAppend(Expression* lexp);
+	Expression* ToAlternateThis(Declaration* pthis, Declaration* nthis);
+
+	void ReplaceVariable(Declaration* pvar, Declaration* nvar);
 
 	bool IsSame(const Expression* exp) const;
 	bool IsRValue(void) const;
@@ -270,6 +285,7 @@ public:
 	bool IsConstRef(void) const;
 	bool IsVolatile(void) const;
 
+
 	void Dump(int ident) const;
 };
 
@@ -286,7 +302,7 @@ public:
 	Token				mToken;
 	Declaration		*	mBase, * mParams, * mParamPack, * mNext, * mPrev, * mConst, * mMutable, * mVolatile;
 	Declaration		*	mDefaultConstructor, * mDestructor, * mCopyConstructor, * mCopyAssignment, * mMoveConstructor, * mMoveAssignment;
-	Declaration		*	mVectorConstructor, * mVectorDestructor, * mVectorCopyConstructor, * mVectorCopyAssignment;
+	Declaration		*	mVectorConstructor, * mVectorDestructor, * mVectorCopyConstructor, * mVectorCopyAssignment, * mVectorMoveConstructor, * mVectorMoveAssignment;
 	Declaration		*	mVTable, * mClass, * mTemplate;
 	Declaration		*	mForwardParam, * mForwardCall;
 
@@ -328,12 +344,17 @@ public:
 	bool IsSimpleType(void) const;
 	bool IsReference(void) const;
 	bool IsIndexed(void) const;
+	bool ContainsArray(void) const;
+	bool IsShortIntStruct(void) const;
+	bool IsComplexStruct(void) const;
+	bool HasConstructor(void) const;
 
 	void SetDefined(void);
 
 	Declaration* ToConstType(void);
 	Declaration* ToMutableType(void);
 	Declaration* ToVolatileType(void);
+	Declaration* ToAlternateThis(Declaration* pthis, int nthis = 1);
 
 	Declaration* ToStriped(int stripe);
 	Declaration* ToStriped(Errors* errors);
@@ -357,8 +378,9 @@ public:
 	DecType ValueType(void) const;
 
 	bool CanResolveTemplate(Expression* pexp, Declaration* tdec);
-	bool ResolveTemplate(Declaration* fdec, Declaration * tdec);
+	bool ResolveTemplate(Declaration* fdec, Declaration * tdec, bool same, bool preliminary);
 	bool ResolveTemplate(Expression* pexp, Declaration* tdec);
+	bool ResolveTemplateParameterList(Expression* pexp, Declaration* pdec, bool preliminary);
 
 	Declaration* ExpandTemplate(DeclarationScope* scope);
 
@@ -376,5 +398,6 @@ extern Declaration* TheBoolTypeDeclaration, * TheFloatTypeDeclaration, * TheVoid
 extern Declaration* TheVoidFunctionTypeDeclaration, * TheConstVoidValueDeclaration;
 extern Declaration* TheCharPointerTypeDeclaration, * TheConstCharPointerTypeDeclaration;
 extern Declaration* TheNullptrConstDeclaration, * TheZeroIntegerConstDeclaration, * TheZeroFloatConstDeclaration, * TheNullPointerTypeDeclaration;
+extern Declaration* TheTrueConstDeclaration, * TheFalseConstDeclaration;
 extern Expression* TheVoidExpression;
 

oscar64/Errors.cpp

@@ -4,11 +4,18 @@
 #include <stdlib.h>
 
 Errors::Errors(void)
-	: mErrorCount(0)
+	: mErrorCount(0), mMinLevel(EINFO_GENERIC), mDisabled(EERR_GENERIC)
 {
 
 }
 
+ErrorID	Errors::SetMinLevel(ErrorID id)
+{
+	ErrorID	pid = mMinLevel;
+	mMinLevel = id;
+	return pid;
+}
+
 void Errors::Error(const Location& loc, ErrorID eid, const char* msg, const Ident* info1, const Ident* info2)
 {
 	if (!info1)
@@ -22,6 +29,8 @@ void Errors::Error(const Location& loc, ErrorID eid, const char* msg, const Iden
 
 void Errors::Error(const Location& loc, ErrorID eid, const char* msg, const char* info1, const char * info2) 
 {
+	if (eid >= mMinLevel && !(eid < EERR_GENERIC && mDisabled[eid]))
+	{
 		const char* level = "info";
 		if (eid >= EERR_GENERIC)
 		{
@@ -57,6 +66,7 @@ void Errors::Error(const Location& loc, ErrorID eid, const char* msg, const char
 
 		if (mErrorCount > 10 || eid >= EFATAL_GENERIC)
 			exit(20);
+	}
 }
 
 

oscar64/Errors.h

@@ -1,5 +1,8 @@
 #pragma once
 
+#include "NumberSet.h"
+
+
 class Location
 {
 public:
@@ -10,7 +13,11 @@ public:
 	Location() : mFileName(nullptr), mLine(0), mColumn(0), mFrom(nullptr) {}
 	Location(const Location& loc, const Location* from)
 		: mFileName(loc.mFileName), mLine(loc.mLine), mColumn(loc.mColumn), mFrom(from)
-	{}
+	{
+		static volatile int k;
+		if (from)
+			k = from->mLine;
+	}
 };
 
 class Ident;
@@ -44,6 +51,8 @@ enum ErrorID
 	EWARN_DEFAULT_COPY_DEPRECATED,
 	EWARN_INSUFFICIENT_MEMORY,
 	EWARN_FUNCTION_NOT_INLINED,
+	EWARN_INVALID_VOID_POINTER_ARITHMETIC,
+	EWARN_DIVISION_BY_ZERO,
 
 	EERR_GENERIC = 3000,
 	EERR_FILE_NOT_FOUND,
@@ -103,7 +112,7 @@ enum ErrorID
 	ERRR_INSTANTIATE_ABSTRACT_CLASS,
 	ERRR_INVALID_GOTO,
 	EERR_INVALID_INITIALIZER,
-
+	ERRR_INVALID_VOID_POINTER_ARITHMETIC,
 
 	EERR_INVALID_CONSTEXPR,
 	EERR_DOUBLE_FREE,
@@ -118,9 +127,13 @@ enum ErrorID
 	EERR_INVALID_CLASS_INITIALIZER,
 	EERR_CALL_OF_DELETED_FUNCTION,
 
+	EERR_STATIC_ASSERT,
+
 	EFATAL_GENERIC = 4000,
 	EFATAL_OUT_OF_MEMORY,
 	EFATAL_MACRO_EXPANSION_DEPTH,
+
+	ERROR_MAX = 5000
 };
 
 class Errors
@@ -128,7 +141,12 @@ class Errors
 public:
 	Errors(void);
 
+	ErrorID	SetMinLevel(ErrorID id);
+
 	int		mErrorCount;
+	ErrorID	mMinLevel;
+
+	NumberSet	mDisabled;
 
 	void Error(const Location& loc, ErrorID eid, const char* msg, const Ident* info1, const Ident* info2 = nullptr);
 	void Error(const Location& loc, ErrorID eid, const char* msg, const char* info1 = nullptr, const char* info2 = nullptr);

oscar64/GlobalAnalyzer.cpp

@@ -393,7 +393,7 @@ void GlobalAnalyzer::CheckFastcall(Declaration* procDec, bool head)
 
 		if (procDec->mValue && procDec->mValue->mType == EX_DISPATCH)
 		{
-			Declaration* maxf = nullptr;
+			Declaration* maxf = nullptr, * reff = nullptr;
 
 			bool	stackCall = false;
 
@@ -402,7 +402,10 @@ void GlobalAnalyzer::CheckFastcall(Declaration* procDec, bool head)
 				Declaration* cf = procDec->mCalled[i];
 
 				if (cf->mBase->mFlags & DTF_STACKCALL)
+				{
 					stackCall = true;
+					reff = cf;
+				}
 
 				if (!maxf)
 					maxf = cf;
@@ -410,6 +413,9 @@ void GlobalAnalyzer::CheckFastcall(Declaration* procDec, bool head)
 					maxf = cf;
 			}
 
+			if (!reff)
+				reff = maxf;
+
 			for (int i = 0; i < procDec->mCalled.Size(); i++)
 			{
 				Declaration* cf = procDec->mCalled[i];
@@ -420,9 +426,9 @@ void GlobalAnalyzer::CheckFastcall(Declaration* procDec, bool head)
 					cf->mBase->mFlags |= DTF_STACKCALL;
 				}
 
-				if (cf != maxf)
+				if (cf != reff)
 				{
-					Declaration* fp = cf->mBase->mParams, * mp = maxf->mBase->mParams;
+					Declaration* fp = cf->mBase->mParams, * mp = reff->mBase->mParams;
 					while (fp)
 					{
 						fp->mVarIndex = mp->mVarIndex;
@@ -431,10 +437,11 @@ void GlobalAnalyzer::CheckFastcall(Declaration* procDec, bool head)
 					}
 
 					assert(!mp);
-
-					cf->mBase->mFastCallBase = cf->mFastCallBase = maxf->mBase->mFastCallBase;
 					cf->mBase->mFastCallSize = cf->mFastCallSize = maxf->mBase->mFastCallSize;
 				}
+
+				if (cf != maxf)
+					cf->mBase->mFastCallBase = cf->mFastCallBase = maxf->mBase->mFastCallBase;
 			}
 
 			procDec->mFastCallBase = procDec->mBase->mFastCallBase;
@@ -474,7 +481,7 @@ void GlobalAnalyzer::CheckFastcall(Declaration* procDec, bool head)
 					fplimit += 256;
 			}
 
-			if (procDec->mBase->mBase->mType == DT_TYPE_STRUCT)
+			if (procDec->mBase->mBase->IsComplexStruct())
 			{
 				if (nbase < numfpzero && nbase + 2 > numfpzero)
 					nbase = numfpzero;
@@ -658,7 +665,7 @@ void GlobalAnalyzer::AnalyzeProcedure(Expression* cexp, Expression* exp, Declara
 			if (mCompilerOptions & COPT_OPTIMIZE_CONST_EXPRESSIONS)
 				dec->mFlags |= DTF_FUNC_CONSTEXPR;
 			dec->mFlags |= DTF_FUNC_PURE;
-			Analyze(exp, dec, false, false);
+			Analyze(exp, dec, 0);
 
 			Declaration* pdec = dec->mBase->mParams;
 			int vi = 0;
@@ -678,7 +685,7 @@ void GlobalAnalyzer::AnalyzeProcedure(Expression* cexp, Expression* exp, Declara
 		}
 		else
 		{
-			mErrors->Error(dec->mLocation, EERR_UNDEFINED_OBJECT, "Calling undefined function", dec->mQualIdent);
+			mErrors->Error(dec->mLocation, EERR_UNDEFINED_OBJECT, "Calling undefined function", dec->FullIdent());
 			if (cexp)
 				mErrors->Error(cexp->mLocation, EINFO_CALLED_FROM, "Called from here");
 
@@ -754,7 +761,7 @@ void GlobalAnalyzer::AnalyzeGlobalVariable(Declaration* dec)
 
 		if (dec->mValue)
 		{
-			Analyze(dec->mValue, dec, false, false);
+			Analyze(dec->mValue, dec, 0);
 		}
 	}
 }
@@ -764,17 +771,19 @@ void GlobalAnalyzer::AnalyzeInit(Declaration* mdec)
 	while (mdec)
 	{
 		if (mdec->mValue)
-			RegisterProc(Analyze(mdec->mValue, mdec, false, false));
+			RegisterProc(Analyze(mdec->mValue, mdec, 0));
 		else if (mdec->mParams)
 			AnalyzeInit(mdec->mParams);
 		mdec = mdec->mNext;
 	}
 }
 
-Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, bool lhs, bool aliasing)
+Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, uint32 flags)
 {
 	Declaration* ldec, * rdec;
 
+	exp->mFlags = flags;
+
 	switch (exp->mType)
 	{
 	case EX_ERROR:
@@ -792,7 +801,7 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 		}
 		else if (exp->mDecValue->mType == DT_CONST_POINTER)
 		{
-			ldec = Analyze(exp->mDecValue->mValue, procDec, true, true);
+			ldec = Analyze(exp->mDecValue->mValue, procDec, ANAFL_LHS | ANAFL_ALIAS);
 			if (ldec->mType == DT_VARIABLE)
 				ldec->mFlags |= DTF_VAR_ALIASING;
 			RegisterProc(ldec);
@@ -816,7 +825,7 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 		if (mCompilerOptions & COPT_DEBUGINFO)
 			exp->mDecValue->mReferences.Push(exp);
 
-		if (aliasing)
+		if (flags & ANAFL_ALIAS)
 		{
 			Declaration* dec = exp->mDecValue;
 			while (dec->mType == DT_VARIABLE_REF)
@@ -833,14 +842,14 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 			if (!(type->mFlags & DTF_CONST))
 				procDec->mFlags &= ~DTF_FUNC_CONSTEXPR;
 
-			if (lhs)
+			if (flags & ANAFL_LHS)
 				procDec->mFlags &= ~DTF_FUNC_PURE;
 
 			AnalyzeGlobalVariable(exp->mDecValue);
 		}
 		else
 		{
-			if (lhs)
+			if (flags & ANAFL_LHS)
 				exp->mDecValue->mFlags |= DTF_VAR_ADDRESS;
 
 			if (!(exp->mDecValue->mFlags & DTF_ANALYZED))
@@ -854,8 +863,8 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 	case EX_ASSIGNMENT:
 		procDec->mComplexity += 5 * exp->mLeft->mDecType->mSize;
 
-		ldec = Analyze(exp->mLeft, procDec, true, false);
+		ldec = Analyze(exp->mLeft, procDec, ANAFL_LHS | ANAFL_RHS);
-		rdec = Analyze(exp->mRight, procDec, false, false);
+		rdec = Analyze(exp->mRight, procDec, ANAFL_RHS);
 		if (exp->mLeft->mType == EX_VARIABLE && exp->mRight->mType == EX_CALL && exp->mLeft->mDecType->mType == DT_TYPE_STRUCT)
 			exp->mLeft->mDecValue->mFlags |= DTF_VAR_ALIASING;
 		RegisterProc(rdec);
@@ -864,33 +873,33 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 	case EX_BINARY:
 		procDec->mComplexity += 10 * exp->mDecType->mSize;
 
-		ldec = Analyze(exp->mLeft, procDec, lhs, false);
+		ldec = Analyze(exp->mLeft, procDec, flags & ~ANAFL_ALIAS);
-		rdec = Analyze(exp->mRight, procDec, lhs, false);
+		rdec = Analyze(exp->mRight, procDec, flags & ~ANAFL_ALIAS);
 		return ldec;
 
 	case EX_RELATIONAL:
 		procDec->mComplexity += 10 * exp->mLeft->mDecType->mSize;
 
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, ANAFL_RHS);
-		rdec = Analyze(exp->mRight, procDec, false, false);
+		rdec = Analyze(exp->mRight, procDec, ANAFL_RHS);
 		return TheBoolTypeDeclaration;
 
 	case EX_PREINCDEC:
 		procDec->mComplexity += 10 * exp->mLeft->mDecType->mSize;
 
-		return Analyze(exp->mLeft, procDec, true, false);
+		return Analyze(exp->mLeft, procDec, ANAFL_LHS | ANAFL_RHS);
 	case EX_PREFIX:
 		if (exp->mToken == TK_BINARY_AND)
 		{
-			ldec = Analyze(exp->mLeft, procDec, true, true);
+			ldec = Analyze(exp->mLeft, procDec, ANAFL_LHS | ANAFL_ALIAS);
 			if (ldec->mType == DT_VARIABLE)
 				ldec->mFlags |= DTF_VAR_ALIASING;
 		} 
 		else if (exp->mToken == TK_MUL)
 		{
-			ldec = Analyze(exp->mLeft, procDec, false, false);
+			ldec = Analyze(exp->mLeft, procDec, 0);
 			procDec->mFlags &= ~DTF_FUNC_CONSTEXPR;
-			if (lhs)
+			if (flags & ANAFL_LHS)
 				procDec->mFlags &= ~DTF_FUNC_PURE;
 
 			return exp->mDecType;
@@ -899,10 +908,14 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 		{
 			return TheUnsignedCharTypeDeclaration;
 		}
+		else if (exp->mToken == TK_SIZEOF)
+		{
+			return TheUnsignedIntTypeDeclaration;
+		}
 		else
 		{
 			procDec->mComplexity += 10 * exp->mLeft->mDecType->mSize;
-			return Analyze(exp->mLeft, procDec, false, false);
+			return Analyze(exp->mLeft, procDec, 0);
 		}
 		break;
 	case EX_POSTFIX:
@@ -911,31 +924,31 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 	case EX_POSTINCDEC:
 		procDec->mComplexity += 10 * exp->mLeft->mDecType->mSize;
 
-		return Analyze(exp->mLeft, procDec, true, false);
+		return Analyze(exp->mLeft, procDec, ANAFL_LHS | ANAFL_RHS);
 	case EX_INDEX:
 		procDec->mComplexity += 10 * exp->mRight->mDecType->mSize;
 
-		ldec = Analyze(exp->mLeft, procDec, lhs, false);
+		ldec = Analyze(exp->mLeft, procDec, flags & ~ANAFL_ALIAS);
 		if (ldec->mType == DT_VARIABLE || ldec->mType == DT_ARGUMENT)
 		{
 			ldec = ldec->mBase;
 			if (ldec->mType == DT_TYPE_POINTER)
 			{
-				if (lhs)
+				if (flags & ANAFL_LHS)
 					procDec->mFlags &= ~DTF_FUNC_PURE;
 				procDec->mFlags &= ~DTF_FUNC_CONSTEXPR;
 			}
 		}
-		rdec = Analyze(exp->mRight, procDec, false, false);
+		rdec = Analyze(exp->mRight, procDec, 0);
 		if (ldec->mBase)
 			return ldec->mBase;
 		break;
 	case EX_QUALIFY:
-		Analyze(exp->mLeft, procDec, lhs, aliasing);
+		Analyze(exp->mLeft, procDec, flags);
 		return exp->mDecValue->mBase;
 	case EX_DISPATCH:
 		procDec->mFlags |= DTF_PREVENT_INLINE;
-		Analyze(exp->mLeft, procDec, lhs, false);
+		Analyze(exp->mLeft, procDec, flags & ~ANAFL_ALIAS);
 //		RegisterCall(procDec, exp->mLeft->mDecType);
 		break;
 	case EX_VCALL:
@@ -946,7 +959,7 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 	case EX_INLINE:
 		procDec->mComplexity += 10;
 
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, 0);
 		if ((ldec->mFlags & DTF_INTRINSIC) && !ldec->mValue)
 		{
 
@@ -1033,7 +1046,7 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 				if (pex->mType == EX_CALL && IsStackParam(pex->mDecType) && !(pdec && (pdec->mBase->mType == DT_TYPE_REFERENCE || pdec->mBase->mType == DT_TYPE_RVALUEREF)))
 					ldec->mBase->mFlags |= DTF_STACKCALL;
 
-				RegisterProc(Analyze(pex, procDec, pdec && pdec->mBase->IsReference(), false));
+				RegisterProc(Analyze(pex, procDec, (pdec && pdec->mBase->IsReference()) ? ANAFL_LHS : 0));
 
 				if (pdec)
 					pdec = pdec->mNext;
@@ -1047,12 +1060,12 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 		break;
 	case EX_LIST:
 	case EX_COMMA:
-		RegisterProc(Analyze(exp->mLeft, procDec, false, false));
+		RegisterProc(Analyze(exp->mLeft, procDec, 0));
-		return Analyze(exp->mRight, procDec, false, false);
+		return Analyze(exp->mRight, procDec, 0);
 	case EX_RETURN:
 		if (exp->mLeft)
 		{
-			RegisterProc(Analyze(exp->mLeft, procDec, procDec->mBase->mBase->IsReference(), false));
+			RegisterProc(Analyze(exp->mLeft, procDec, procDec->mBase->mBase->IsReference() ? ANAFL_LHS : 0));
 			if (procDec->mBase->mBase && procDec->mBase->mBase->mType == DT_TYPE_STRUCT && procDec->mBase->mBase->mCopyConstructor)
 			{
 				if (procDec->mBase->mBase->mMoveConstructor)
@@ -1073,47 +1086,47 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 			if (exp->mType == EX_SEQUENCE)
 			{
 				if (exp->mLeft)
-					ldec = Analyze(exp->mLeft, procDec, false, false);
+					ldec = Analyze(exp->mLeft, procDec, 0);
 				exp = exp->mRight;
 			}
 			else
-				return Analyze(exp, procDec, false, false);
+				return Analyze(exp, procDec, 0);
 
 		} while (exp);
 		break;
 
 	case EX_SCOPE:
-		Analyze(exp->mLeft, procDec, false, false);
+		Analyze(exp->mLeft, procDec, 0);
 		break;
 
 	case EX_CONSTRUCT:
 		if (exp->mLeft->mLeft)
-			Analyze(exp->mLeft->mLeft, procDec, false, false);
+			Analyze(exp->mLeft->mLeft, procDec, 0);
 		if (exp->mLeft->mRight)
-			Analyze(exp->mLeft->mRight, procDec, false, false);
+			Analyze(exp->mLeft->mRight, procDec, 0);
 		if (exp->mRight)
-			return Analyze(exp->mRight, procDec, false, false);
+			return Analyze(exp->mRight, procDec, 0);
 		break;
 
 	case EX_CLEANUP:
-		Analyze(exp->mRight, procDec, false, false);
+		Analyze(exp->mRight, procDec, 0);
-		return Analyze(exp->mLeft, procDec, lhs, false);
+		return Analyze(exp->mLeft, procDec, flags & ~ANAFL_ALIAS);
 		
 	case EX_WHILE:
 		procDec->mFlags &= ~DTF_FUNC_CONSTEXPR;
 
 		procDec->mComplexity += 20;
 
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, 0);
-		rdec = Analyze(exp->mRight, procDec, false, false);
+		rdec = Analyze(exp->mRight, procDec, 0);
 		break;
 	case EX_IF:
 		procDec->mComplexity += 20;
 
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, 0);
-		rdec = Analyze(exp->mRight->mLeft, procDec, false, false);
+		rdec = Analyze(exp->mRight->mLeft, procDec, 0);
 		if (exp->mRight->mRight)
-			rdec = Analyze(exp->mRight->mRight, procDec, false, false);
+			rdec = Analyze(exp->mRight->mRight, procDec, 0);
 		break;
 	case EX_ELSE:
 		break;
@@ -1123,18 +1136,23 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 		procDec->mComplexity += 30;
 
 		if (exp->mLeft->mRight)
-			ldec = Analyze(exp->mLeft->mRight, procDec, false, false);
+			ldec = Analyze(exp->mLeft->mRight, procDec, 0);
 		if (exp->mLeft->mLeft->mLeft)
-			ldec = Analyze(exp->mLeft->mLeft->mLeft, procDec, false, false);
+			ldec = Analyze(exp->mLeft->mLeft->mLeft, procDec, 0);
-		rdec = Analyze(exp->mRight, procDec, false, false);
+		rdec = Analyze(exp->mRight, procDec, 0);
 		if (exp->mLeft->mLeft->mRight)
-			ldec = Analyze(exp->mLeft->mLeft->mRight, procDec, false, false);
+			ldec = Analyze(exp->mLeft->mLeft->mRight, procDec, 0);
+		break;
+	case EX_FORBODY:
+		ldec = Analyze(exp->mLeft, procDec, 0);
+		if (exp->mRight)
+			Analyze(exp->mRight, procDec, 0);
 		break;
 	case EX_DO:
 		procDec->mComplexity += 20;
 
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, 0);
-		rdec = Analyze(exp->mRight, procDec, false, false);
+		rdec = Analyze(exp->mRight, procDec, 0);
 		break;
 	case EX_BREAK:
 	case EX_CONTINUE:
@@ -1143,18 +1161,18 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 	case EX_TYPE:
 		break;
 	case EX_TYPECAST:
-		return Analyze(exp->mLeft, procDec, false, false);
+		return Analyze(exp->mLeft, procDec, 0);
 		break;
 	case EX_LOGICAL_AND:
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, 0);
-		rdec = Analyze(exp->mRight, procDec, false, false);
+		rdec = Analyze(exp->mRight, procDec, 0);
 		break;
 	case EX_LOGICAL_OR:
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, 0);
-		rdec = Analyze(exp->mRight, procDec, false, false);
+		rdec = Analyze(exp->mRight, procDec, 0);
 		break;
 	case EX_LOGICAL_NOT:
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, 0);
 		break;
 	case EX_ASSEMBLER:
 		procDec->mFlags |= DTF_FUNC_ASSEMBLER;
@@ -1165,14 +1183,14 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 	case EX_UNDEFINED:
 		break;
 	case EX_SWITCH:
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, 0);
 		exp = exp->mRight;
 		while (exp)
 		{
 			procDec->mComplexity += 10;
 
 			if (exp->mLeft->mRight)
-				rdec = Analyze(exp->mLeft->mRight, procDec, false, false);
+				rdec = Analyze(exp->mLeft->mRight, procDec, 0);
 			exp = exp->mRight;
 		}
 		break;
@@ -1183,9 +1201,9 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec, boo
 	case EX_CONDITIONAL:
 		procDec->mComplexity += exp->mDecType->mSize * 10;
 
-		ldec = Analyze(exp->mLeft, procDec, false, false);
+		ldec = Analyze(exp->mLeft, procDec, 0);
-		RegisterProc(Analyze(exp->mRight->mLeft, procDec, lhs, aliasing));
+		RegisterProc(Analyze(exp->mRight->mLeft, procDec, flags));
-		RegisterProc(Analyze(exp->mRight->mRight, procDec, lhs, aliasing));
+		RegisterProc(Analyze(exp->mRight->mRight, procDec, flags));
 		break;
 	}
 

oscar64/GlobalAnalyzer.h

@@ -35,7 +35,7 @@ protected:
 	int CallerInvokes(Declaration* called);
 	int CallerInvokes(Declaration* caller, Declaration* called);
 
-	Declaration* Analyze(Expression* exp, Declaration* procDec, bool lhs, bool aliasing);
+	Declaration* Analyze(Expression* exp, Declaration* procDec, uint32 flags);
 
 	bool IsStackParam(const Declaration* pdec) const;
 	bool MarkCycle(Declaration* rootDec, Declaration* procDec);

oscar64/GlobalOptimizer.cpp

@@ -60,7 +60,7 @@ void GlobalOptimizer::Reset(void)
 	mFunctions.SetSize(0);
 	mGlobalVariables.SetSize(0);
 	mCalledFunctions.SetSize(0);
-	mCalledFunctions.SetSize(0);
+	mCallingFunctions.SetSize(0);
 }
 
 void GlobalOptimizer::PropagateParamCommas(Expression*& fexp, Expression*& exp)
@@ -243,6 +243,19 @@ bool GlobalOptimizer::ReplaceParamConst(Expression* exp, Declaration* param)
 	bool	changed = false;
 	if (exp)
 	{
+		if (ReplaceParamConst(exp->mLeft, param))
+			changed = true;
+		if (ReplaceParamConst(exp->mRight, param))
+			changed = true;
+
+		if (changed)
+		{
+			if (exp->mLeft)
+				exp->mLeft = exp->mLeft->ConstantFold(mErrors, nullptr);
+			if (exp->mRight)
+				exp->mRight = exp->mRight->ConstantFold(mErrors, nullptr);
+		}
+
 		if (exp->mType == EX_VARIABLE && exp->mDecValue == param)
 		{
 			exp->mType = EX_CONSTANT;
@@ -251,10 +264,6 @@ bool GlobalOptimizer::ReplaceParamConst(Expression* exp, Declaration* param)
 			changed = true;
 		}
 
-		if (ReplaceParamConst(exp->mLeft, param))
-			changed = true;
-		if (ReplaceParamConst(exp->mRight, param))
-			changed = true;
 	}
 	return changed;
 }
@@ -266,15 +275,23 @@ bool GlobalOptimizer::ReplaceGlobalConst(Expression* exp)
 	{
 		if (exp->mType == EX_VARIABLE && (exp->mDecValue->mFlags & (DTF_GLOBAL | DTF_STATIC)) && !(exp->mDecValue->mOptFlags & (OPTF_VAR_MODIFIED | OPTF_VAR_ADDRESS)) && exp->mDecValue->mValue)
 		{
+
 			Expression* cexp = exp->mDecValue->mValue;
-			if (cexp->mType == EX_CONSTANT && 
+			if (cexp->mType == EX_CONSTANT)
-				(cexp->mDecValue->mType == DT_CONST_ADDRESS || cexp->mDecValue->mType == DT_CONST_INTEGER || 
+			{
-				 cexp->mDecValue->mType == DT_CONST_POINTER || cexp->mDecValue->mType == DT_CONST_FLOAT))
+
+				if (cexp->mDecValue->mType == DT_CONST_ADDRESS || cexp->mDecValue->mType == DT_CONST_INTEGER ||
+					cexp->mDecValue->mType == DT_CONST_POINTER || cexp->mDecValue->mType == DT_CONST_FLOAT)
 				{
 					exp->mType = EX_CONSTANT;
 					exp->mDecValue = cexp->mDecValue->ConstCast(exp->mDecType);
 					changed = true;
 				}
+				else if (exp->mDecValue->mBase->mType == DT_TYPE_STRUCT)
+				{
+					exp->mDecValue->mBase = exp->mDecValue->mBase->ToConstType();
+				}
+			}
 		}
 
 		if (ReplaceGlobalConst(exp->mLeft))
@@ -324,7 +341,7 @@ bool GlobalOptimizer::Optimize(void)
 
 			if (!(func->mOptFlags & OPTF_FUNC_VARIABLE) && !(func->mBase->mFlags & DTF_VIRTUAL))
 			{
-				if (!(func->mOptFlags & OPTF_VAR_USED) && func->mBase->mBase && (func->mBase->mBase->IsSimpleType() || func->mBase->mBase->IsReference()))
+				if (!(func->mOptFlags & (OPTF_VAR_USED | OPTF_VAR_ADDRESS)) && func->mBase->mBase && (func->mBase->mBase->IsSimpleType() || func->mBase->mBase->IsReference()))
 				{
 #if DUMP_OPTS
 					printf("Remove return value\n");
@@ -382,6 +399,7 @@ bool GlobalOptimizer::Optimize(void)
 						{
 							if (ReplaceParamConst(func->mValue, pdec))
 							{
+								func->mValue = func->mValue->ConstantFold(mErrors, nullptr);
 #if DUMP_OPTS
 								printf("Const parameter %s\n", pdec->mIdent ? pdec->mIdent->mString : "_");
 #endif
@@ -444,7 +462,7 @@ void GlobalOptimizer::AnalyzeProcedure(Expression* exp, Declaration* procDec)
 			Analyze(exp, procDec, false);
 		}
 		else
-			mErrors->Error(procDec->mLocation, EERR_UNDEFINED_OBJECT, "Calling undefined function", procDec->mQualIdent);
+			mErrors->Error(procDec->mLocation, EERR_UNDEFINED_OBJECT, "Calling undefined function", procDec->FullIdent());
 
 		procDec->mOptFlags &= ~OPTF_ANALYZING;
 	}
@@ -596,9 +614,6 @@ void GlobalOptimizer::RegisterProc(Declaration* to)
 	}
 }
 
-static const uint32 ANAFL_LHS	=	(1U << 0);
-static const uint32 ANAFL_RHS	= (1U << 1);
-
 Declaration* GlobalOptimizer::Analyze(Expression* exp, Declaration* procDec, uint32 flags)
 {
 	Declaration* ldec, * rdec;
@@ -644,6 +659,8 @@ Declaration* GlobalOptimizer::Analyze(Expression* exp, Declaration* procDec, uin
 			exp->mDecValue->mOptFlags |= OPTF_VAR_USED;
 		if (flags & ANAFL_LHS)
 			exp->mDecValue->mOptFlags |= OPTF_VAR_ADDRESS;
+		if (exp->mDecValue->mBase->IsReference() && (flags & ANAFL_ASSIGN))
+			exp->mDecValue->mOptFlags |= OPTF_VAR_USED;
 
 		if (exp->mDecValue->mType == DT_ARGUMENT && (flags & ANAFL_LHS))
 		{
@@ -655,7 +672,12 @@ Declaration* GlobalOptimizer::Analyze(Expression* exp, Declaration* procDec, uin
 	case EX_INITIALIZATION:
 	case EX_ASSIGNMENT:
 		if (exp->mToken == TK_ASSIGN)
+		{
+			if (exp->mType == EX_ASSIGNMENT)
+				ldec = Analyze(exp->mLeft, procDec, ANAFL_LHS | ANAFL_ASSIGN | flags);
+			else
 				ldec = Analyze(exp->mLeft, procDec, ANAFL_LHS | flags);
+		}
 		else
 			ldec = Analyze(exp->mLeft, procDec, ANAFL_LHS | ANAFL_RHS | flags);
 
@@ -803,6 +825,22 @@ Declaration* GlobalOptimizer::Analyze(Expression* exp, Declaration* procDec, uin
 								pdec->mOptFlags |= OPTF_VAR_CONST;
 							}
 						}
+						else if (pex->mType == EX_VARIABLE && pdec->mBase->mType == DT_TYPE_POINTER && pex->mDecType->mType == DT_TYPE_ARRAY && (pex->mDecValue->mFlags & (DTF_GLOBAL | DTF_STATIC)) && pdec->mBase->CanAssign(pex->mDecType))
+						{
+							if (pdec->mOptFlags & OPTF_VAR_CONST)
+							{
+								if (pdec->mValue->mType != EX_VARIABLE || pdec->mValue->mDecValue != pex->mDecValue)
+								{
+									pdec->mOptFlags |= OPTF_VAR_NOCONST;
+									pdec->mOptFlags &= ~OPTF_VAR_CONST;
+								}
+							}
+							else
+							{
+								pdec->mValue = pex;
+								pdec->mOptFlags |= OPTF_VAR_CONST;
+							}
+						}
 						else
 						{
 							pdec->mOptFlags |= OPTF_VAR_NOCONST;
@@ -948,13 +986,20 @@ Declaration* GlobalOptimizer::Analyze(Expression* exp, Declaration* procDec, uin
 		if (exp->mLeft->mLeft->mRight)
 			ldec = Analyze(exp->mLeft->mLeft->mRight, procDec, 0);
 		break;
+	case EX_FORBODY:
+		ldec = Analyze(exp->mLeft, procDec, 0);
+		if (exp->mRight)
+			rdec = Analyze(exp->mRight, procDec, 0);
+		break;
 	case EX_DO:
 		ldec = Analyze(exp->mRight, procDec, 0);
 		rdec = Analyze(exp->mLeft, procDec, ANAFL_RHS);
 		break;
 	case EX_BREAK:
 	case EX_CONTINUE:
+		break;
 	case EX_ASSUME:
+		return Analyze(exp->mLeft, procDec, ANAFL_RHS);
 		break;
 	case EX_TYPE:
 		break;

oscar64/Ident.cpp

@@ -47,7 +47,7 @@ const Ident* Ident::Unique(const char* str)
 
 const Ident* Ident::PreMangle(const char* str) const
 {
-	char	buffer[200];
+	char	buffer[1000];
 	strcpy_s(buffer, str);
 	strcat_s(buffer, mString);
 	return Unique(buffer);
@@ -55,14 +55,14 @@ const Ident* Ident::PreMangle(const char* str) const
 
 const Ident* Ident::Unique(const char* str, int id)
 {
-	char	buffer[200];
+	char	buffer[1000];
 	sprintf_s(buffer, "%s#%d", str, id);
 	return Unique(buffer);
 }
 
 const Ident* Ident::Mangle(const char* str) const
 {
-	char	buffer[200];
+	char	buffer[1000];
 	strcpy_s(buffer, mString);
 	strcat_s(buffer, str);
 	return Unique(buffer);

oscar64/InterCode.h

@@ -169,6 +169,7 @@ public:
 	}			mMinState, mMaxState;
 
 	bool Same(const IntegerValueRange& range) const;
+	bool Weaker(const IntegerValueRange& range) const;
 	bool Merge(const IntegerValueRange& range, bool head, bool initial);
 	void Expand(const IntegerValueRange& range);
 	void Union(const IntegerValueRange& range);
@@ -198,7 +199,7 @@ public:
 
 
 
-typedef GrowingArray<IntegerValueRange>		GrowingIntegerValueRangeArray;
+typedef ExpandingArray<IntegerValueRange>		GrowingIntegerValueRangeArray;
 
 class ValueSet
 {
@@ -380,7 +381,7 @@ public:
 	InterCodeBasicBlock			*	mTrueJump, * mFalseJump, * mLoopPrefix, * mDominator;
 	GrowingInstructionArray			mInstructions;
 
-	bool							mVisited, mInPath, mLoopHead, mChecked, mConditionBlockTrue, mUnreachable, mLoopPath, mValueRangeValid;
+	bool							mVisited, mInPath, mLoopHead, mChecked, mConditionBlockTrue, mUnreachable, mLoopPath, mValueRangeValid, mPatched, mLoopDebug;
 	mutable int						mMark;
 
 	NumberSet						mLocalUsedTemps, mLocalModifiedTemps;
@@ -423,14 +424,15 @@ public:
 	InterCodeBasicBlock* Clone(void);
 
 	void Append(InterInstruction * code);
-	void AppendBeforeBranch(InterInstruction* code);
+	void AppendBeforeBranch(InterInstruction* code, bool loopindex = false);
 	const InterInstruction* FindByDst(int dst) const;
 	void Close(InterCodeBasicBlock* trueJump, InterCodeBasicBlock* falseJump);
 
 	void CollectEntries(void);
 	void CollectEntryBlocks(InterCodeBasicBlock* from);
-	void GenerateTraces(bool expand, bool compact);
+	void GenerateTraces(int expand, bool compact);
 	void BuildDominatorTree(InterCodeBasicBlock * from);
+	bool StripLoopHead(void);
 
 	bool MergeSameConditionTraces(void);
 
@@ -448,6 +450,7 @@ public:
 	void CollectConstTemps(GrowingInstructionPtrArray& ctemps, NumberSet& assignedTemps);
 	bool PropagateConstTemps(const GrowingInstructionPtrArray& ctemps);
 	bool ForwardConstTemps(const GrowingInstructionPtrArray& ctemps);
+	bool PropagateConstCompareResults(void);
 
 	bool EarlyBranchElimination(const GrowingInstructionPtrArray& ctemps);
 
@@ -480,6 +483,8 @@ public:
 
 	bool CheckSingleBlockLimitedLoop(InterCodeBasicBlock*& pblock, int64 & nloop);
 
+	bool TempIsUnsigned(int temp);
+
 	void RestartLocalIntegerRangeSets(int num, const GrowingVariableArray& localVars, const GrowingVariableArray& paramVars);
 	void BuildLocalIntegerRangeSets(int num, const GrowingVariableArray& localVars, const GrowingVariableArray& paramVars);
 	void UpdateLocalIntegerRangeSets(const GrowingVariableArray& localVars, const GrowingVariableArray& paramVars);
@@ -515,6 +520,8 @@ public:
 	void PerformMachineSpecificValueUsageCheck(const GrowingInstructionPtrArray& tvalue, FastNumberSet& tvalid, const GrowingVariableArray& staticVars, const GrowingInterCodeProcedurePtrArray& staticProcs, FastNumberSet& fsingle);
 	bool EliminateDeadBranches(void);
 
+	bool EliminateIntegerSumAliasTemps(const GrowingInstructionPtrArray& tvalue);
+
 	bool MergeIndexedLoadStore(const GrowingInstructionPtrArray& tvalue);
 	bool SimplifyIntegerNumeric(const GrowingInstructionPtrArray& tvalue, int& spareTemps);
 	bool SimplifyPointerOffsets(void);
@@ -623,29 +630,42 @@ public:
 	void SingleBlockLoopUnrolling(void);
 	bool SingleBlockLoopPointerSplit(int& spareTemps);
 	bool SingleBlockLoopPointerToByte(int& spareTemps);
+	bool SingleBlockLoopSinking(int& spareTemps);
 	bool CollectLoopBody(InterCodeBasicBlock* head, ExpandingArray<InterCodeBasicBlock*> & body);
 	bool CollectLoopBodyRecursive(InterCodeBasicBlock* head, ExpandingArray<InterCodeBasicBlock*>& body);
 	void CollectLoopPath(const ExpandingArray<InterCodeBasicBlock*>& body, ExpandingArray<InterCodeBasicBlock*>& path);
 	void InnerLoopOptimization(const NumberSet& aliasedParams);
 	void ConstLoopOptimization(void);
+	bool EmptyLoopOptimization(void);
 	void EliminateDoubleLoopCounter(void);
 	void PushMoveOutOfLoop(void);
 	bool MoveConditionOutOfLoop(void);
 	void SingleLoopCountZeroCheck(void);
+	void InnerLoopCountZeroCheck(void);
 	bool PostDecLoopOptimization(void);
+	bool Flatten2DLoop(void);
 
-	void PropagateMemoryAliasingInfo(const GrowingInstructionPtrArray& tvalue);
+	void PropagateMemoryAliasingInfo(const GrowingInstructionPtrArray& tvalue, bool loops);
 	void RemoveUnusedMallocs(void);
 
 	bool PullStoreUpToConstAddress(void);
 
-	bool CollectSingleHeadLoopBody(InterCodeBasicBlock* head, InterCodeBasicBlock* tail, GrowingArray<InterCodeBasicBlock*>& body);
+	bool CollectSingleHeadLoopBody(InterCodeBasicBlock* head, InterCodeBasicBlock* tail, ExpandingArray<InterCodeBasicBlock*>& body);
+
+	bool CollectGenericLoop(ExpandingArray<InterCodeBasicBlock*>& lblocks);
+	bool CollectSingleEntryGenericLoop(ExpandingArray<InterCodeBasicBlock*>& lblocks);
+	void CollectReachable(ExpandingArray<InterCodeBasicBlock*>& lblock);
 
 	bool SingleTailLoopOptimization(const NumberSet& aliasedParams, const GrowingVariableArray& staticVars);
 	bool MergeLoopTails(void);
 
+	bool ChangeTrueJump(InterCodeBasicBlock* block);
+	bool ChangeFalseJump(InterCodeBasicBlock* block);
+
+	InterCodeBasicBlock* CheckIsSimpleIntRangeBranch(const GrowingIntegerValueRangeArray & irange);
 	InterCodeBasicBlock* CheckIsConstBranch(const GrowingInstructionPtrArray& cins);
 	bool ShortcutConstBranches(const GrowingInstructionPtrArray& cins);
+	bool ShortcutDuplicateBranches(void);
 
 	InterCodeBasicBlock* BuildLoopPrefix(void);
 	void BuildLoopSuffix(void);
@@ -656,6 +676,7 @@ public:
 	void FollowJumps(void);
 
 	bool ShortLeaMerge(int& spareTemps);
+	bool ShortLeaCleanup(void);
 
 	bool IsEqual(const InterCodeBasicBlock* block) const;
 
@@ -692,6 +713,7 @@ protected:
 	GrowingIntegerValueRangeArray		mLocalValueRange, mReverseValueRange;
 
 	void ResetVisited(void);
+	void ResetPatched(void);
 	void ResetEntryBlocks(void);
 public:
 	InterCodeBasicBlock				*	mEntryBlock;
@@ -753,7 +775,7 @@ public:
 	void Disassemble(const char* name, bool dumpSets = false);
 protected:
 	void BuildLocalAliasTable(void);
-	void BuildTraces(bool expand, bool dominators = true, bool compact = false);
+	void BuildTraces(int expand, bool dominators = true, bool compact = false);
 	void TrimBlocks(void);
 	void EarlyBranchElimination(void);
 	void BuildDataFlowSets(void);
@@ -774,7 +796,9 @@ protected:
 	void SimplifyIntegerNumeric(FastNumberSet& activeSet);
 	void SingleBlockLoopPointerSplit(FastNumberSet& activeSet);
 	void SingleBlockLoopPointerToByte(FastNumberSet& activeSet);
+	void SingleBlockLoopSinking(FastNumberSet& activeSet);
 	void MergeIndexedLoadStore(void);
+	void EliminateIntegerSumAliasTemps(void);
 	void EliminateAliasValues();
 	void LoadStoreForwarding(InterMemory paramMemory);
 	void ReduceRecursionTempSpilling(InterMemory paramMemory);
@@ -794,9 +818,10 @@ protected:
 	void CheckUsedDefinedTemps(void);
 	void WarnUsedUndefinedVariables(void);
 	void WarnInvalidValueRanges(void);
-	void PropagateMemoryAliasingInfo(void);
+	void PropagateMemoryAliasingInfo(bool loops);
 	void MoveConditionsOutOfLoop(void);
 	void ShortcutConstBranches(void);
+	void ShortcutDuplicateBranches(void);
 	void EliminateDoubleLoopCounter(void);
 	void StructReturnPropagation(void);
 

oscar64/InterCodeGenerator.h

@@ -45,7 +45,7 @@ public:
 	uint64		mCompilerOptions;
 
 	InterCodeProcedure* TranslateProcedure(InterCodeModule* mod, Expression* exp, Declaration * dec);
-	void TranslateAssembler(InterCodeModule* mod, Expression * exp, GrowingArray<Declaration *>	* refvars);
+	void TranslateAssembler(InterCodeModule* mod, Declaration* adec,GrowingArray<Declaration *>	* refvars);
 	void InitGlobalVariable(InterCodeModule* mod, Declaration* dec);
 	void InitLocalVariable(InterCodeProcedure* proc, Declaration* dec, int index);
 	void InitParameter(InterCodeProcedure* proc, Declaration* dec, int index);
@@ -94,12 +94,12 @@ protected:
 	void BuildSwitchTree(InterCodeProcedure* proc, Expression* exp, InterCodeBasicBlock* block, InlineMapper * inlineMapper, ExValue v, const SwitchNodeArray& nodes, int left, int right, int vleft, int vright, InterCodeBasicBlock* dblock);
 
 	ExValue ToValue(InterCodeProcedure* proc, Expression* exp, InterCodeBasicBlock*& block, InlineMapper* inlineMapper, ExValue v);
-	ExValue Dereference(InterCodeProcedure* proc, Expression* exp, InterCodeBasicBlock*& block, InlineMapper* inlineMapper, ExValue v, int level = 0);
+	ExValue Dereference(InterCodeProcedure* proc, Expression* exp, InterCodeBasicBlock*& block, InlineMapper* inlineMapper, ExValue v, int level = 0, int limit = 0);
 	ExValue CoerceType(InterCodeProcedure* proc, Expression* exp, InterCodeBasicBlock*& block, InlineMapper* inlineMapper, ExValue v, Declaration * type, bool checkTrunc = true);
 	ExValue TranslateExpression(Declaration * procType, InterCodeProcedure * proc, InterCodeBasicBlock*& block, Expression* exp, DestructStack*& destack, GotoNode*& gotos, const BranchTarget & breakBlock, const BranchTarget& continueBlock, InlineMapper * inlineMapper, ExValue * lrexp = nullptr);
 	void TranslateLogic(Declaration* procType, InterCodeProcedure* proc, InterCodeBasicBlock* block, InterCodeBasicBlock* tblock, InterCodeBasicBlock* fblock, Expression* exp, DestructStack*& destack, GotoNode*& gotos, InlineMapper* inlineMapper);
 	ExValue TranslateInline(Declaration* procType, InterCodeProcedure* proc, InterCodeBasicBlock*& block, Expression* exp, const BranchTarget& breakBlock, const BranchTarget& continueBlock, InlineMapper* inlineMapper, bool inlineConstexpr, ExValue* lrexp);
-	void CopyStruct(InterCodeProcedure* proc, Expression* exp, InterCodeBasicBlock*& block, ExValue vl, ExValue vr, InlineMapper* inlineMapper, bool moving);
+	void CopyStruct      (InterCodeProcedure* proc, Expression* exp, InterCodeBasicBlock*& block, ExValue vl, ExValue vr, InlineMapper* inlineMapper, bool moving);
 	void CopyStructSimple(InterCodeProcedure* proc, Expression* exp, InterCodeBasicBlock * block, InlineMapper* inlineMapper, ExValue vl, ExValue vr);
 	void StoreValue(InterCodeProcedure* proc, Expression* exp, InterCodeBasicBlock*& block, InlineMapper* inlineMapper, ExValue vl, ExValue vr);
 

oscar64/Linker.cpp

@@ -45,7 +45,7 @@ bool LinkerReference::operator!=(const LinkerReference& ref)
 }
 
 LinkerObject::LinkerObject(void)
-	: mReferences(nullptr), mNumTemporaries(0), mSize(0), mAlignment(1), mStackSection(nullptr), mIdent(nullptr), mFullIdent(nullptr), mStartUsed(0x10000), mEndUsed(0x00000), mMemory(nullptr)
+	: mReferences(nullptr), mNumTemporaries(0), mSize(0), mStripe(0), mAlignment(1), mStackSection(nullptr), mIdent(nullptr), mFullIdent(nullptr), mStartUsed(0x10000), mEndUsed(0x00000), mMemory(nullptr)
 	, mPrefix(nullptr), mSuffix(nullptr), mProc(nullptr), mNativeProc(nullptr)
 {}
 
@@ -54,6 +54,23 @@ LinkerObject::~LinkerObject(void)
 
 }
 
+LinkerObject* LinkerObject::CloneAssembler(Linker* linker) const
+{
+	LinkerObject* lo = linker->AddObject(mLocation, mIdent, mSection, mType, mAlignment);
+	lo->AddData(mData, mSize);
+	lo->mFlags = mFlags;
+
+	for (int i = 0; i < mReferences.Size(); i++)
+	{
+		LinkerReference		ref = *(mReferences[i]);
+		if (ref.mObject == this) ref.mObject = lo;
+		if (ref.mRefObject == this) ref.mRefObject = lo;
+		lo->AddReference(ref);
+	}
+
+	return lo;
+}
+
 void LinkerObject::AddReference(const LinkerReference& ref)
 {
 	LinkerReference* nref = new LinkerReference(ref);
@@ -633,7 +650,7 @@ bool LinkerRegion::Allocate(Linker * linker, LinkerObject* lobj, bool merge, boo
 		int start = (mFreeChunks[i].mStart + lobj->mAlignment - 1) & ~(lobj->mAlignment - 1);
 		int end = start + lobj->mSize;
 
-		if (!(linker->mCompilerOptions & COPT_OPTIMIZE_CODE_SIZE) && (lobj->mFlags & LOBJF_NO_CROSS) && lobj->mSize <= 256 && (start & 0xff00) != ((end - 1) & 0xff00) && !(lobj->mSection->mFlags & LSECF_PACKED))
+		if (((lobj->mFlags & LOBJF_NEVER_CROSS) || !(linker->mCompilerOptions & COPT_OPTIMIZE_CODE_SIZE) && (lobj->mFlags & LOBJF_NO_CROSS) && !(lobj->mSection->mFlags & LSECF_PACKED)) && lobj->mSize <= 256 && (start & 0xff00) != ((end - 1) & 0xff00))
 			;
 		else if (end <= mFreeChunks[i].mEnd)
 		{
@@ -685,7 +702,7 @@ bool LinkerRegion::Allocate(Linker * linker, LinkerObject* lobj, bool merge, boo
 	int start = (mStart + mUsed + lobj->mAlignment - 1) & ~(lobj->mAlignment - 1);
 	int end = start + lobj->mSize;
 
-	if (!(linker->mCompilerOptions & COPT_OPTIMIZE_CODE_SIZE) && !retry && (lobj->mFlags & LOBJF_NO_CROSS) && !(lobj->mFlags & LOBJF_FORCE_ALIGN) && lobj->mSize <= 256 && (start & 0xff00) != ((end - 1) & 0xff00) && !(lobj->mSection->mFlags & LSECF_PACKED))
+	if (((lobj->mFlags & LOBJF_NEVER_CROSS) || !(linker->mCompilerOptions & COPT_OPTIMIZE_CODE_SIZE) && !retry && (lobj->mFlags & LOBJF_NO_CROSS) && !(lobj->mSection->mFlags & LSECF_PACKED)) && !(lobj->mFlags & LOBJF_FORCE_ALIGN) && lobj->mSize <= 256 && (start & 0xff00) != ((end - 1) & 0xff00))
 	{
 		start = (start + 0x00ff) & 0xff00;
 		end = start + lobj->mSize;
@@ -1886,6 +1903,8 @@ bool Linker::WriteDbjFile(FILE* file)
 	return true;
 }
 
+static const char hexchars[] = "0123456789abcdef";
+
 bool Linker::WriteLblFile(const char* filename)
 {
 	FILE* file;
@@ -1899,7 +1918,28 @@ bool Linker::WriteLblFile(const char* filename)
 			if (obj->mFlags & LOBJF_REFERENCED)
 			{
 				if (obj->mIdent)
-					fprintf(file, "al %04x .%s\n", obj->mAddress, obj->mIdent->mString);
+				{
+					char buffer[400];
+					char nbuffer[500];
+
+					strcpy_s(buffer, obj->mIdent->mString);
+					int i = 0, j = 0;
+					while (buffer[i])
+					{
+						if (buffer[i] >= '0' && buffer[i] <= '9' || buffer[i] >= 'a' && buffer[i] <= 'z' || buffer[i] >= 'A' && buffer[i] <= 'Z' || buffer[i] == '_' || buffer[i] == ':')
+							nbuffer[j++] = buffer[i];
+						else
+						{
+							nbuffer[j++] = '?';
+							nbuffer[j++] = hexchars[(buffer[i] >> 4) & 0x0f];
+							nbuffer[j++] = hexchars[buffer[i] & 0x0f];
+
+						}
+						i++;
+					}
+					nbuffer[j] = 0;
+					fprintf(file, "al %04x .%s\n", obj->mAddress, nbuffer);
+				}
 			}
 		}
 

oscar64/Linker.h

@@ -156,6 +156,7 @@ static const uint32 LOBJF_NO_CROSS		= 0x00000080;
 static const uint32 LOBJF_ZEROPAGE		= 0x00000100;
 static const uint32 LOBJF_FORCE_ALIGN	= 0x00000200;
 static const uint32 LOBJF_ZEROPAGESET = 0x00000400;
+static const uint32 LOBJF_NEVER_CROSS = 0x00000800;
 
 static const uint32 LOBJF_ARG_REG_A = 0x00001000;
 static const uint32 LOBJF_ARG_REG_X = 0x00002000;
@@ -163,6 +164,7 @@ static const uint32 LOBJF_ARG_REG_Y = 0x00004000;
 
 static const uint32 LOBJF_RET_REG_A = 0x00010000;
 static const uint32 LOBJF_RET_REG_X = 0x00020000;
+static const uint32 LOBJF_RET_REG_Y = 0x00020000;
 
 static const uint32 LOBJF_LOCAL_VAR = 0x00100000;
 static const uint32 LOBJF_LOCAL_USED = 0x00200000;
@@ -190,7 +192,7 @@ public:
 	LinkerObjectType					mType;
 	int									mID, mMapID;
 	int									mAddress, mRefAddress;
-	int									mSize, mAlignment, mStartUsed, mEndUsed;
+	int									mSize, mAlignment, mStripe, mStartUsed, mEndUsed;
 	LinkerSection					*	mSection;
 	LinkerRegion					*	mRegion;
 	uint8							*	mData, * mMemory;
@@ -212,6 +214,8 @@ public:
 	LinkerObject(void);
 	~LinkerObject(void);
 
+	LinkerObject* CloneAssembler(Linker * linker) const;
+
 	void AddData(const uint8* data, int size);
 	void AddLocations(const ExpandingArray<CodeLocation>& locations);
 	uint8* AddSpace(int size);

oscar64/NativeCodeGenerator.h

@@ -9,6 +9,9 @@ class NativeCodeBasicBlock;
 class NativeCodeGenerator;
 class NativeCodeInstruction;
 
+class NativeCodeMapper;
+class SuffixTree;
+
 enum NativeRegisterDataMode
 {
 	NRDM_UNKNOWN,
@@ -130,19 +133,25 @@ static const uint32 NCIF_ALIASING = 0x00000800;
 static const uint32 NCIF_USE_CPU_REG_A = 0x00001000;
 static const uint32 NCIF_USE_CPU_REG_X = 0x00002000;
 static const uint32 NCIF_USE_CPU_REG_Y = 0x00004000;
+static const uint32 NCIF_USE_CPU_REG_C = 0x00008000;
+
+static const uint32 NCIF_PROVIDE_CPU_REG_A = 0x00010000;
+static const uint32 NCIF_PROVIDE_CPU_REG_X = 0x00020000;
+static const uint32 NCIF_PROVIDE_CPU_REG_Y = 0x00040000;
+static const uint32 NCIF_PROVIDE_CPU_REG_C = 0x00080000;
 
 // use a 32bit zero page register indexed by X for JSR
-static const uint32 NCIF_USE_ZP_32_X = 0x00008000;
+static const uint32 NCIF_USE_ZP_32_X = 0x00100000;
-static const uint32 NICF_USE_ZP_ADDR = 0x00010000;
+static const uint32 NICF_USE_ZP_ADDR = 0x00200000;
-static const uint32 NICF_USE_WORKREGS = 0x00020000;
+static const uint32 NICF_USE_WORKREGS = 0x00400000;
 
-static const uint32 NCIF_BREAKPOINT = 0x00040000;
+static const uint32 NCIF_BREAKPOINT = 0x00800000;
 
 class NativeCodeInstruction
 {
 public:
 	NativeCodeInstruction(void);
-	NativeCodeInstruction(const InterInstruction * ins, AsmInsType type, AsmInsMode mode = ASMIM_IMPLIED, int64 address = 0, LinkerObject * linkerObject = nullptr, uint32 flags = NCIF_LOWER | NCIF_UPPER, int param = 0);
+	NativeCodeInstruction(const InterInstruction * ins, AsmInsType type, AsmInsMode mode = ASMIM_IMPLIED, int64 address = 0, LinkerObject * linkerObject = nullptr, uint32 flags = NCIF_LOWER | NCIF_UPPER, int param = 0, int minv = 0, int maxv = 255);
 	NativeCodeInstruction(const InterInstruction* ins, AsmInsType type, const NativeCodeInstruction & addr);
 
 	AsmInsType					mType;
@@ -153,8 +162,13 @@ public:
 	uint32						mLive;
 	LinkerObject			*	mLinkerObject;
 	const InterInstruction	*	mIns;
+	uint8						mMinVal, mMaxVal;
+
+	void Disassemble(FILE* file) const;
+	void DisassembleAddress(FILE* file) const;
 
 	void CopyMode(const NativeCodeInstruction& ins);
+	void CopyModeAndRange(const NativeCodeInstruction& ins);
 
 	void Assemble(NativeCodeBasicBlock* block);
 	void FilterRegUsage(NumberSet& requiredTemps, NumberSet& providedTemps);
@@ -191,6 +205,7 @@ public:
 	bool UsesZeroPage(int address) const;
 	bool ReferencesZeroPage(int address) const;
 
+	bool SameLinkerObjectVariableRange(const NativeCodeInstruction& ins, bool sameXY = false) const;
 
 	bool ChangesGlobalMemory(void) const;
 	bool UsesMemoryOf(const NativeCodeInstruction& ins) const;
@@ -205,7 +220,7 @@ public:
 	bool IsShift(void) const;
 	bool IsShiftOrInc(void) const;
 	bool IsSimpleJSR(void) const;
-	bool MayBeMovedBefore(const NativeCodeInstruction& ins);
+	bool MayBeMovedBefore(const NativeCodeInstruction& ins) const;
 
 	bool ReplaceYRegWithXReg(void);
 	bool ReplaceXRegWithYReg(void);
@@ -216,6 +231,12 @@ public:
 	void BuildCollisionTable(NumberSet& liveTemps, NumberSet* collisionSets);
 
 	uint32 NeedsLive(void) const;
+
+	uint32 CodeHash(void) const;
+	bool CodeSame(const NativeCodeInstruction& ins);
+
+protected:
+	const char* AddrName(char* buffer) const;
 };
 
 struct NativeCodeLoadStorePair
@@ -246,7 +267,7 @@ public:
 	ExpandingArray<NativeCodeBasicBlock*>	mEntryBlocks;
 
 	int							mOffset, mSize, mPlace, mNumEntries, mNumEntered, mFrameOffset, mTemp;
-	bool						mPlaced, mCopied, mKnownShortBranch, mBypassed, mAssembled, mNoFrame, mVisited, mLoopHead, mVisiting, mLocked, mPatched, mPatchFail, mPatchChecked, mPatchStart, mPatchLoop, mPatchLoopChanged, mPatchExit;
+	bool						mPlaced, mCopied, mKnownShortBranch, mBypassed, mAssembled, mNoFrame, mVisited, mLoopHead, mVisiting, mLocked, mPatched, mPatchFail, mPatchChecked, mPatchUsed, mPatchStart, mPatchLoop, mPatchLoopChanged, mPatchExit;
 	bool						mEntryRegA, mEntryRegX, mEntryRegY, mExitRegA, mExitRegX, mChecked;
 	NativeCodeBasicBlock	*	mDominator, * mSameBlock;
 
@@ -264,12 +285,16 @@ public:
 
 	NativeCodeInstruction		mALSIns, mXLSIns, mYLSIns;
 
+	void Disassemble(FILE* file);
+	void DisassembleBody(FILE* file);
+
 	NativeCodeInstruction DecodeNative(const InterInstruction* ins, LinkerObject * lobj, int& offset) const;
 
 	int PutBranch(NativeCodeProcedure* proc, NativeCodeBasicBlock* target, AsmInsType code, int from, int to);
 	int PutJump(NativeCodeProcedure* proc, NativeCodeBasicBlock* target, int from, int to, AsmInsType code = ASMIT_INV);
 	int JumpByteSize(NativeCodeBasicBlock * target, int from, int to, bool second, bool final);
-	int BranchByteSize(NativeCodeBasicBlock* target, int from, int to, bool final);
+	int CheckFinalBranchByteSize(NativeCodeBasicBlock* target, int from, int to) const;
+	int BranchByteSize(NativeCodeBasicBlock* target, int from, int to, bool final) const;
 
 	NativeCodeBasicBlock* SplitAt(int at);
 
@@ -278,6 +303,7 @@ public:
 	int LeadsInto(NativeCodeBasicBlock* block, int dist);
 	NativeCodeBasicBlock* PlaceSequence(ExpandingArray<NativeCodeBasicBlock*>& placement, NativeCodeBasicBlock* block);
 	void BuildPlacement(ExpandingArray<NativeCodeBasicBlock*>& placement);
+	void OptimizePlacement(void);
 	void InitialOffset(int& total);
 	bool CalculateOffset(int& total, bool final);
 
@@ -302,6 +328,8 @@ public:
 	bool UsesZeroPage(int address, int from = 0, int to = 65536) const;
 	bool ReferencesZeroPage(int address, int from = 0, int to = 65536) const;
 
+	bool ChangesMemory(const NativeCodeInstruction& ins, int from = 0, int to = 65536) const;
+	bool ReferencesMemory(const NativeCodeInstruction& ins, int from = 0, int to = 65536) const;
 
 	bool RemoveNops(void);
 	bool PeepHoleOptimizer(int pass);
@@ -320,18 +348,20 @@ public:
 	bool PeepHoleOptimizerIterate(int pass);
 	bool PeepHoleOptimizerExits(int pass);
 
-	void BlockSizeReduction(NativeCodeProcedure* proc, int xenter, int yenter);
+	void BlockSizeReduction(NativeCodeProcedure* proc, int xenter, int yenter, int center);
 	bool BlockSizeCopyReduction(NativeCodeProcedure* proc, int & si, int & di);
 
 	bool OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc, bool full);
 	bool OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc, NativeCodeBasicBlock * prevBlock, NativeCodeBasicBlock* exitBlock, bool full);
 	bool RemoveSimpleLoopUnusedIndex(void);
 	bool OptimizeLoopCarryOver(void);
+	bool OptimizeLoopRegisterWrapAround(void);
 
 	bool OptimizeSingleEntryLoopInvariant(NativeCodeProcedure* proc, NativeCodeBasicBlock* prev, NativeCodeBasicBlock* tail, ExpandingArray<NativeCodeBasicBlock*>& blocks);
 	bool OptimizeSingleEntryLoop(NativeCodeProcedure* proc);
 
 	bool OptimizeSimpleLoop(NativeCodeProcedure* proc, bool full);
+	bool OptimizeSimpleForLoop(void);
 	bool SimpleLoopReversal(NativeCodeProcedure* proc);
 	bool OptimizeInnerLoop(NativeCodeProcedure* proc, NativeCodeBasicBlock* head, NativeCodeBasicBlock* tail, ExpandingArray<NativeCodeBasicBlock*>& blocks);
 	bool OptimizeXYSimpleLoop(void);
@@ -343,9 +373,11 @@ public:
 	bool OptimizeInnerLoops(NativeCodeProcedure* proc);
 	NativeCodeBasicBlock* CollectInnerLoop(NativeCodeBasicBlock* head, ExpandingArray<NativeCodeBasicBlock*>& lblocks);
 
-	bool OptimizeGenericLoop(NativeCodeProcedure* proc);
+	int CorrectXOffset(const InterInstruction * ins, int yoffset, int at);
-	bool CollectGenericLoop(NativeCodeProcedure* proc, ExpandingArray<NativeCodeBasicBlock*>& lblocks);
+	int CorrectYOffset(const InterInstruction * ins, int yoffset, int at);
-	bool CollectSingleEntryGenericLoop(NativeCodeProcedure* proc, ExpandingArray<NativeCodeBasicBlock*>& lblocks);
+	bool OptimizeGenericLoop(void);
+	bool CollectGenericLoop(ExpandingArray<NativeCodeBasicBlock*>& lblocks);
+	bool CollectSingleEntryGenericLoop(ExpandingArray<NativeCodeBasicBlock*>& lblocks);
 	void CollectReachable(ExpandingArray<NativeCodeBasicBlock*>& lblock);
 
 	bool OptimizeFindLoop(NativeCodeProcedure* proc);
@@ -379,7 +411,7 @@ public:
 	NativeCodeBasicBlock* BinaryOperator(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction * ins, const InterInstruction* sins1, const InterInstruction* sins0);
 	void UnaryOperator(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction * ins);
 	void RelationalOperator(InterCodeProcedure* proc, const InterInstruction * ins, NativeCodeProcedure * nproc, NativeCodeBasicBlock* trueJump, NativeCodeBasicBlock * falseJump);
-	void LoadEffectiveAddress(InterCodeProcedure* proc, const InterInstruction * ins, const InterInstruction* sins1, const InterInstruction* sins0, bool addrvalid);
+	void LoadEffectiveAddress(InterCodeProcedure* proc, const InterInstruction * ins, const InterInstruction* sins1, const InterInstruction* sins0, bool addrvalid, bool addrused = false);
 	void LoadStoreOpAbsolute2D(InterCodeProcedure* proc, const InterInstruction* lins1, const InterInstruction* lins2, const InterInstruction* mins);
 	void SignExtendAddImmediate(InterCodeProcedure* proc, const InterInstruction* xins, const InterInstruction* ains);
 	void BinaryDivModPair(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction* ins1, const InterInstruction* ins2);
@@ -396,6 +428,7 @@ public:
 
 	void LoadByteIndexedValue(InterCodeProcedure* proc, const InterInstruction* iins, const InterInstruction* rins);
 	void StoreByteIndexedValue(InterCodeProcedure* proc, const InterInstruction* iins, const InterInstruction* rins);
+	void CopyByteIndexedValue(InterCodeProcedure* proc, const InterInstruction* riins, const InterInstruction* wiins, const InterInstruction* rins, const InterInstruction* wins);
 
 	void CallAssembler(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction * ins);
 	void CallFunction(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction * ins);
@@ -408,6 +441,7 @@ public:
 	int ShortSignedDivide(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction* ins, const InterInstruction* sins, int mul);
 
 	bool CheckPredAccuStore(int reg);
+	bool CheckIsInAccu(int reg);
 
 	NumberSet		mLocalRequiredRegs, mLocalProvidedRegs;
 	NumberSet		mEntryRequiredRegs, mEntryProvidedRegs;
@@ -429,9 +463,7 @@ public:
 	void CountEntries(NativeCodeBasicBlock* fromJump);
 	NativeCodeBasicBlock * ForwardAccuBranch(bool eq, bool ne, bool pl, bool mi, int limit);
 	bool MergeBasicBlocks(void);
-	void RemoveJumpToBranch(void);
+	bool RemoveJumpToBranch(void);
-
-	void MarkLoopHead(void);
 
 	struct DominatorStacks
 	{
@@ -505,7 +537,9 @@ public:
 	bool CombineImmediateADCUp(int at);
 	bool CombineImmediateADCUpX(int at);
 	bool MoveTXADCDown(int at);
+	bool MoveTXALogicTAXDown(int at);
 	bool FoldShiftORAIntoLoadImmUp(int at);
+	bool ReverseShiftByteOrder(int at);
 
 	bool FindAccuExitValue(int& at);
 	bool MoveLoadXAbsUpCrossBlock(int at);
@@ -539,6 +573,7 @@ public:
 	bool CombineSameYtoX(int xpos, int ypos, int end);
 
 	bool FindImmediateStore(int at, int reg, const NativeCodeInstruction*& ains);
+	int FindImmediateGlobalStore(int at, const NativeCodeInstruction& ins);
 
 	bool JoinXYCrossBlock(void);
 	bool CanCombineSameXtoYCrossBlock(int from);
@@ -559,6 +594,13 @@ public:
 
 	bool ReverseReplaceTAX(int at);
 	
+	bool CanReplaceExitAccuWithX(NativeCodeBasicBlock * target);
+	bool CanReplaceExitAccuWithY(NativeCodeBasicBlock* target);
+	void ReplaceExitAccuWithX(NativeCodeBasicBlock* target);
+	void ReplaceExitAccuWithY(NativeCodeBasicBlock* target);
+
+	bool ReverseLoadAccuToRegXY(void);
+
 	void ResetModifiedDataSet(NativeRegisterDataSet& data);
 
 	bool ValueForwarding(NativeCodeProcedure* proc, const NativeRegisterDataSet& data, bool global, bool final);
@@ -568,6 +610,9 @@ public:
 	bool OffsetValueForwarding(const ValueNumberingDataSet & data);
 	bool AbsoluteValueForwarding(const ExpandingArray<NativeCodeLoadStorePair>& npairs);
 	bool IndexXYValueForwarding(int xreg, int xoffset, int xvalue, int yreg, int yoffset, int yvalue);
+	bool ReduceIndexXYZeroShuffle(NativeCodeBasicBlock* from, int xreg, int yreg);
+	bool CheckLoopIndexXRegisters(NativeCodeBasicBlock* head, int xreg);
+	bool CheckLoopIndexYRegisters(NativeCodeBasicBlock* head, int yreg);
 
 	void MarkLocalUsedLinkerObjects(void);
 	bool RemoveLocalUnusedLinkerObjects(void);
@@ -587,7 +632,14 @@ public:
 	bool HasTailSTY(int& addr, int& index) const;
 	bool HasTailSTAX16(int& addr, int& index0) const;
 
+	bool HasTailSTAInto(int& addr, int& index, NativeCodeBasicBlock* tblock) const;
+	bool HasTailSTXInto(int& addr, int& index, NativeCodeBasicBlock* tblock) const;
+	bool HasTailSTYInto(int& addr, int& index, NativeCodeBasicBlock* tblock) const;
+
+	bool HasTailSTAGlobal(NativeCodeInstruction & ins, int& index) const;
+
 	bool MayBeMovedBeforeBlock(int at);
+	bool MayBeMovedBeforeBlock(int at, const NativeCodeInstruction & ins);
 	bool MayBeMovedBeforeBlock(int start, int end);
 	bool SafeInjectSequenceFromBack(NativeCodeBasicBlock* block, int start, int end);
 	bool JoinCommonBranchCodeSequences(void);
@@ -691,6 +743,8 @@ public:
 	bool EliminateDeadLoops(void);
 	bool LoopRegisterWrapAround(void);
 	bool CrossBlockFlagsForwarding(void);
+	bool MoveStoresBeforeDiamond(void);
+	bool MoveStoresBehindCondition(void);
 
 	bool SinglePathRegisterForwardY(NativeCodeBasicBlock* path, int yreg);
 	bool SinglePathRegisterForward(void);
@@ -723,8 +777,9 @@ public:
 	bool CanGlobalSwapXY(void);
 	bool GlobalSwapXY(void);
 	bool LocalSwapXY(void);
-	bool UntangleXYUsage(void);
+	bool UntangleXYUsage(bool final);
 	bool AlternateXXUsage(void);
+	bool ShortSwapXY(void);
 
 	bool IsSimpleSubExpression(int at, NativeSimpleSubExpression & ex);
 	bool PropagateCommonSubExpression(void);
@@ -748,10 +803,20 @@ public:
 	bool CheckPatchFailUse(void);
 
 	bool CheckPatchFailLoop(const NativeCodeBasicBlock* block, const NativeCodeBasicBlock* head, int reg, bool changed);
+	bool CheckPatchFailLoopPair(const NativeCodeBasicBlock* block, const NativeCodeBasicBlock* head, int reg, bool changed);
 
 	bool JoinSameBranch(NativeCodeBasicBlock* block);
 	bool MergeSameBranch(void);
 
+	bool CheckBoolBitPropagation(const NativeCodeBasicBlock* block, int at, int reg);
+	bool PatchBoolBitPropagation(const NativeCodeBasicBlock* block, int at, int reg);
+
+	bool CollectRegBoolInstructionsForward(int reg, ExpandingArray<NativeCodeBasicBlock*>& cblocks, ExpandingArray<NativeCodeInstruction*>& lins);
+	bool CollectRegBoolInstructionsBackward(int reg, ExpandingArray<NativeCodeBasicBlock*>& cblocks, ExpandingArray<NativeCodeInstruction*>& lins);
+
+	bool CollectCheckRegOriginBlocks(int at, int reg, ExpandingArray<NativeCodeBasicBlock*>& lblocks, ExpandingArray<NativeCodeInstruction*>& lins);
+	bool PatchBitBoolConstOrigin(void);
+
 	// reg : base register pair to replace
 	// index: index register
 	// at : start position in block
@@ -761,9 +826,20 @@ public:
 	bool CheckGlobalAddressSumYPointer(const NativeCodeBasicBlock * block, int reg, int index, int at, int yval);
 	bool PatchGlobalAddressSumYPointer(const NativeCodeBasicBlock* block, int reg, int index, int at, int yval, LinkerObject * lobj, int address, uint32 flags = NCIF_LOWER | NCIF_UPPER);
 
+	// reg : register to replace
+	// at : start position in block
+	// ains : instruction loading original data
+	// cycles : max number of cycles saving
 	bool CheckSingleUseGlobalLoad(const NativeCodeBasicBlock* block, int reg, int at, const NativeCodeInstruction& ains, int cycles);
 	bool PatchSingleUseGlobalLoad(const NativeCodeBasicBlock* block, int reg, int at, const NativeCodeInstruction& ains);
 
+	// rins : instruction storing the data
+	// at : start position in block
+	// ains : instruction loading original data
+	// cycles : max number of cycles saving
+	bool CheckSingleUseGlobalLoadStruct(const NativeCodeBasicBlock* block, const NativeCodeInstruction& rins, int at, const NativeCodeInstruction& ains, bool cleared, bool poisoned);
+	bool PatchSingleUseGlobalLoadStruct(const NativeCodeBasicBlock* block, const NativeCodeInstruction& rins, int at, const NativeCodeInstruction& ains);
+
 	// reg : base register pair to replace
 	// base: new base register
 	// iins : indexing instruction
@@ -807,6 +883,10 @@ public:
 
 	void PropagateAddGlobalCarry(void);
 
+	bool EliminateNonAliasedLocalStores(void);
+	bool CheckNonAliasedLocalStore(int at, const NativeCodeInstruction& sins);
+
+
 	void RegisterFunctionCalls(void);
 	bool MergeFunctionCalls(void);
 
@@ -816,6 +896,9 @@ public:
 	void CheckBlocks(bool sequence = false);
 	void CheckAsmCode(void);
 	void CheckVisited(void);
+
+	int* mSuffixString;
+	void AddToSuffixTree(NativeCodeMapper& mapper, SuffixTree * tree);
 };
 
 class NativeCodeProcedure
@@ -830,6 +913,10 @@ class NativeCodeProcedure
 		NativeCodeGenerator* mGenerator;
 
 		InterCodeProcedure* mInterProc;
+		LinkerObject* mLinkerObject;
+		const Ident* mIdent;
+		Location	mLocation;
+		uint64				mCompilerOptions;
 
 		int		mProgStart, mProgSize, mIndex, mFrameOffset, mStackExpand;
 		int		mFastCallBase;
@@ -841,10 +928,15 @@ class NativeCodeProcedure
 		ExpandingArray<CodeLocation>		mCodeLocations;
 
 
+		void DisassembleDebug(const char* name);
+		void Disassemble(FILE* file);
+
 		void Compile(InterCodeProcedure* proc);
 		void Optimize(void);
 		void Assemble(void);
 
+		void AddToSuffixTree(NativeCodeMapper& mapper, SuffixTree* tree);
+
 		NativeCodeBasicBlock* CompileBlock(InterCodeProcedure* iproc, InterCodeBasicBlock* block);
 		NativeCodeBasicBlock* AllocateBlock(void);
 
@@ -910,6 +1002,8 @@ public:
 	Linker* mLinker;
 	LinkerSection* mRuntimeSection;
 
+	ExpandingArray<NativeCodeProcedure*>	mProcedures;
+
 	ExpandingArray<Runtime>		mRuntime;
 	ExpandingArray<MulTable>	mMulTables;
 	ExpandingArray<FloatTable>	mFloatTables;
@@ -928,6 +1022,8 @@ public:
 
 	FunctionCall* mFunctionCalls;
 
+	void OutlineFunctions(void);
+
 	void RegisterFunctionCall(NativeCodeBasicBlock* block, int at);
 	void BuildFunctionProxies(void);
 	bool MergeFunctionCall(NativeCodeBasicBlock* block, int at);

oscar64/NativeCodeOutliner.cpp

@@ -0,0 +1,379 @@
+#include "NativeCodeOutliner.h"
+
+
+NativeCodeMapper::NativeCodeMapper(void)
+{
+	for (int i = 0; i < HashSize; i++)
+		mHash[i] = nullptr;
+}
+
+NativeCodeMapper::~NativeCodeMapper(void)
+{
+	for (int i = 0; i < HashSize; i++)
+	{
+		InsNode* n = mHash[i];
+		while (n)
+		{
+			InsNode* m = n;
+			n = n->mNext;
+			delete m;
+		}
+		mHash[i] = nullptr;
+	}
+
+}
+
+void NativeCodeMapper::Reset(void)
+{
+	mBlocks.SetSize(0);
+}
+
+int NativeCodeMapper::MapBasicBlock(NativeCodeBasicBlock* block)
+{
+	mBlocks.Push(block);
+	return -mBlocks.Size();
+}
+
+int NativeCodeMapper::MapInstruction(const NativeCodeInstruction& ins, LinkerSection* ls)
+{
+	int hash = ins.CodeHash() % HashSize;
+	InsNode* n = mHash[hash];
+	while (n)
+	{
+		if (mIns[n->mIndex].CodeSame(ins) && n->mSection == ls)
+			return n->mIndex;
+		n = n->mNext;
+	}
+	n = new InsNode();
+	n->mIndex = mIns.Size();
+	n->mSection = ls;
+	mIns.Push(ins);
+	n->mNext = mHash[hash];
+	mHash[hash] = n;
+	return n->mIndex;
+}
+
+SuffixTree::SuffixTree(const int* str, int s, SuffixTree* n)
+{
+	mSeg = str;
+	mSize = s;
+	mNext = n;
+	mParent = nullptr;
+	mFirst = nullptr;
+}
+
+SuffixTree::~SuffixTree(void)
+{
+	delete[] mFirst;
+}
+
+void SuffixTree::AddParents(SuffixTree* parent)
+{
+	mParent = parent;
+	if (mFirst)
+	{
+		for (int i = 0; i < HashSize; i++)
+		{
+			SuffixTree* n = mFirst[i];
+			while (n)
+			{
+				n->AddParents(this);
+				n = n->mNext;
+			}
+		}
+	}
+}
+
+void SuffixTree::AddSuffix(const int* str, int s)
+{
+	int hi = str[0] & (HashSize - 1);
+	SuffixTree* c = mFirst ? mFirst[hi] : nullptr;
+	while (c && c->mSeg[0] != str[0])
+		c = c->mNext;
+
+	if (c)
+	{
+		int k = 1;
+		while (k < c->mSize && str[k] == c->mSeg[k])
+			k++;
+		if (k == c->mSize)
+			c->AddSuffix(str + k, s - k);
+		else
+		{
+			SuffixTree * n = new SuffixTree(c->mSeg + k, c->mSize - k, nullptr);
+			if (c->mFirst)
+			{
+				n->mFirst = new SuffixTree * [HashSize];
+				for (int i = 0; i < HashSize; i++)
+				{
+					n->mFirst[i] = c->mFirst[i];
+					c->mFirst[i] = nullptr;
+				}
+			}
+			else
+			{
+				c->mFirst = new SuffixTree * [HashSize];
+				for (int i = 0; i < HashSize; i++)
+					c->mFirst[i] = nullptr;
+			}
+			c->mFirst[c->mSeg[k] & (HashSize - 1)] = n;
+			int hk = str[k] & (HashSize - 1);
+			c->mFirst[hk] = new SuffixTree(str + k, s - k, c->mFirst[hk]);
+			c->mSize = k;
+		}
+	}
+	else
+	{
+		if (!mFirst)
+		{
+			mFirst = new SuffixTree * [HashSize];
+			for (int i = 0; i < HashSize; i++)
+				mFirst[i] = nullptr;
+		}
+
+		mFirst[hi] = new SuffixTree(str, s, mFirst[hi]);
+	}
+}
+
+void SuffixTree::AddString(const int* str)
+{
+	int s = 0;
+	while(str[s] >= 0)
+		s++;
+	s++;
+
+	int i = 0;
+	while (str[i] >= 0)
+	{
+		AddSuffix(str + i, s - i);
+		i++;
+	}
+	AddSuffix(str + i, 1);
+}
+
+void SuffixTree::CollectSuffix(NativeCodeMapper& map, int offset, ExpandingArray<SuffixSegment>& segs)
+{
+	offset += mSize;
+	if (mFirst)
+	{
+		for (int i = 0; i < HashSize; i++)
+		{
+			SuffixTree* t = mFirst[i];
+
+			while (t)
+			{
+				t->CollectSuffix(map, offset, segs);
+				t = t->mNext;
+			}
+		}
+	}
+	else
+	{
+		NativeCodeBasicBlock* block = map.mBlocks[-(mSeg[mSize - 1] + 1)];
+
+		SuffixSegment	seg;
+		seg.mBlock = block;
+		seg.mStart = offset;
+		seg.mEnd = block->mIns.Size() + 1;
+		segs.Push(seg);
+	}
+}
+
+int SuffixTree::LongestMatch(NativeCodeMapper& map, int size, int isize, int& msize, SuffixTree*& mtree)
+{
+	if (mFirst)
+	{
+		isize += mSize;
+
+		for (int i = 0; i < mSize; i++)
+		{
+			if (mSeg[i] >= 0)
+				size += AsmInsModeSize[map.mIns[mSeg[i]].mMode];
+		}
+
+		assert(size < 10000);
+
+		int cnt = 0;
+		for (int i = 0; i < HashSize; i++)
+		{
+			SuffixTree* t = mFirst[i];
+			while (t)
+			{
+				cnt += t->LongestMatch(map, size, isize, msize, mtree);
+				t = t->mNext;
+			}
+		}
+
+		if (size >= 6 && (size - 3) * (cnt - 1) > msize)
+		{
+			// Second run to cross check for overlaps
+			ExpandingArray<SuffixSegment>	segs;
+			for (int i = 0; i < HashSize; i++)
+			{
+				SuffixTree* t = mFirst[i];
+				while (t)
+				{
+					t->CollectSuffix(map, 0, segs);
+					t = t->mNext;
+				}
+			}
+			segs.Sort([](const SuffixSegment& l, const SuffixSegment& r)->bool {
+				return l.mBlock == r.mBlock ? l.mStart < r.mStart : ptrdiff_t(l.mBlock) < ptrdiff_t(r.mBlock);
+			});
+
+			for (int i = 0; i + 1 < segs.Size(); i++)
+			{
+				if (segs[i].mBlock == segs[i + 1].mBlock && segs[i].mStart + isize > segs[i + 1].mStart)
+					cnt--;
+			}
+
+			if (cnt > 1 && (size - 3) * (cnt - 1) > msize)
+			{
+				msize = (size - 3) * (cnt - 1);
+				mtree = this;
+			}
+		}
+
+		return cnt;
+	}
+	else
+		return 1;
+}
+
+void SuffixTree::Print(FILE * file, NativeCodeMapper& map, int depth)
+{
+	for (int i = 0; i < depth; i++)
+		fprintf(file, ".");
+
+	for (int i = 0; i < mSize; i++)
+	{
+		fprintf(file, "[");
+
+		if (mSeg[i] >= 0)
+			map.mIns[mSeg[i]].Disassemble(file);
+		else
+		{
+			NativeCodeBasicBlock* block = map.mBlocks[- (mSeg[i] + 1)];
+			fprintf(file, "%s,%d", block->mProc->mInterProc->mIdent->mString, block->mIndex);
+		}
+		fprintf(file, "]");
+	}
+	fprintf(file, "\n");
+
+	if (mFirst)
+	{
+		for (int i = 0; i < HashSize; i++)
+		{
+			SuffixTree* n = mFirst[i];
+			while (n)
+			{
+				n->Print(file, map, depth + 1);
+				n = n->mNext;
+			}
+		}
+	}
+}
+
+void SuffixTree::ParentPrint(FILE* file, NativeCodeMapper& map)
+{
+	if (mParent)
+		mParent->ParentPrint(file, map);
+	for (int i = 0; i < mSize; i++)
+	{
+		if (mSeg[i] >= 0)
+			map.mIns[mSeg[i]].Disassemble(file);
+		else
+		{
+			NativeCodeBasicBlock* block = map.mBlocks[-(mSeg[i] + 1)];
+			fprintf(file, "%s,%d", block->mProc->mInterProc->mIdent->mString, block->mIndex);
+		}
+		fprintf(file, "\n");
+	}
+}
+
+int SuffixTree::ParentCodeSize(NativeCodeMapper& map) const
+{
+	int size = 0;
+	for (int i = 0; i < mSize; i++)
+	{
+		if (mSeg[i] >= 0)
+			size += AsmInsModeSize[map.mIns[mSeg[i]].mMode];
+	}
+
+	if (mParent)
+		size += mParent->ParentCodeSize(map);
+
+	return size;
+}
+
+void SuffixTree::ParentCollect(NativeCodeMapper& map, NativeCodeBasicBlock* block)
+{
+	if (mParent)
+		mParent->ParentCollect(map, block);
+	for (int i = 0; i < mSize; i++)
+		block->mIns.Push(map.mIns[mSeg[i]]);
+}
+
+void SuffixTree::ReplaceCalls(NativeCodeMapper& map, ExpandingArray<SuffixSegment>& segs)
+{
+	if (mFirst)
+	{
+		for (int i = 0; i < HashSize; i++)
+		{
+			SuffixTree* n = mFirst[i];
+			while (n)
+			{
+				n->ChildReplaceCalls(map, this, 0, segs);
+				n = n->mNext;
+			}
+		}
+	}
+}
+
+void SuffixTree::ChildReplaceCalls(NativeCodeMapper& map, SuffixTree* tree, int offset, ExpandingArray<SuffixSegment>& segs)
+{
+	for (int i = 0; i < mSize; i++)
+	{
+		if (mSeg[i] >= 0)
+			offset ++;
+	}
+
+	if (mFirst)
+	{
+		for (int i = 0; i < HashSize; i++)
+		{
+			SuffixTree* n = mFirst[i];
+			while (n)
+			{
+				n->ChildReplaceCalls(map, tree, offset, segs);
+				n = n->mNext;
+			}
+		}
+	}
+	else
+	{
+		NativeCodeBasicBlock* block = map.mBlocks[-(mSeg[mSize - 1] + 1)];
+		tree->ParentReplaceCalls(map, block, offset, 0, segs);
+	}
+}
+
+void SuffixTree::ParentReplaceCalls(NativeCodeMapper& map, NativeCodeBasicBlock* block, int offset, int size, ExpandingArray<SuffixSegment>& segs)
+{
+	for (int i = 0; i < mSize; i++)
+	{
+		if (mSeg[i] >= 0)
+			size ++;
+	}
+	if (mParent)
+		mParent->ParentReplaceCalls(map, block, offset, size, segs);
+	else
+	{
+		int at = block->mIns.Size() - offset - size;
+
+		SuffixSegment	seg;
+		seg.mBlock = block;
+		seg.mStart = at;
+		seg.mEnd = at + size;
+		segs.Push(seg);
+	}
+}

oscar64/NativeCodeOutliner.h

@@ -0,0 +1,66 @@
+#pragma once
+
+#include "NativeCodeGenerator.h"
+#include "Array.h"
+
+class NativeCodeMapper
+{
+public:
+	NativeCodeMapper(void);
+	~NativeCodeMapper(void);
+
+	void Reset(void);
+
+	int MapInstruction(const NativeCodeInstruction& ins, LinkerSection * ls);
+	int MapBasicBlock(NativeCodeBasicBlock* block);
+
+	ExpandingArray<NativeCodeInstruction>	mIns;
+	ExpandingArray<NativeCodeBasicBlock*>	mBlocks;
+protected:
+	static const int HashSize = 256;
+
+	struct InsNode
+	{
+		int						mIndex;
+		LinkerSection		*	mSection;
+		InsNode* mNext;
+	};
+
+	InsNode							*		mHash[HashSize];
+};
+
+struct SuffixSegment
+{
+	NativeCodeBasicBlock	*	mBlock;
+	int							mStart, mEnd;
+};
+
+class SuffixTree
+{
+public:
+	const int	*	mSeg;
+	int				mSize;
+
+	static const int HashSize = 32;
+
+	SuffixTree* mNext, * mParent, ** mFirst;
+
+	SuffixTree(const int* str, int s, SuffixTree* n);
+	~SuffixTree(void);
+
+	void AddParents(SuffixTree* parent);
+
+	void AddSuffix(const int* str, int s);
+	void AddString(const int* str);
+
+	void Print(FILE* file, NativeCodeMapper & map, int depth);
+	void ParentPrint(FILE* file, NativeCodeMapper& map);
+	int LongestMatch(NativeCodeMapper& map, int size, int isize, int & msize, SuffixTree *& mtree);
+	void CollectSuffix(NativeCodeMapper& map, int offset, ExpandingArray<SuffixSegment>& segs);
+	int ParentCodeSize(NativeCodeMapper& map) const;
+	void ParentCollect(NativeCodeMapper& map, NativeCodeBasicBlock * block);
+	void ReplaceCalls(NativeCodeMapper& map, ExpandingArray<SuffixSegment> & segs);
+	void ChildReplaceCalls(NativeCodeMapper& map, SuffixTree * tree, int offset, ExpandingArray<SuffixSegment>& segs);
+	void ParentReplaceCalls(NativeCodeMapper& map, NativeCodeBasicBlock* block, int offset, int size, ExpandingArray<SuffixSegment>& segs);
+
+};

oscar64/Parser.h

@@ -11,6 +11,7 @@ public:
 	~Parser(void);
 
 	Parser* Clone(void);
+	Parser				*	mParent;
 
 	DeclarationScope	*	mGlobals, * mScope, * mTemplateScope, * mCaptureScope;
 	int						mLocalIndex;
@@ -24,6 +25,8 @@ public:
 	uint64			mCompilerOptionStack[32];
 	int				mCompilerOptionSP;
 
+	Location FullLocation(const Location& loc);
+
 	void Parse(void);
 protected:
 	bool ExpectToken(Token token);
@@ -77,12 +80,13 @@ protected:
 	Expression* DefaultInitExpression(Expression* vexp);
 	Expression* ParseVarInitExpression(Expression* lexp, bool inner = false);
 	Expression* CloneVarInitExpression(Expression* vexp, Expression* iexp, Expression * qexp);
+	Expression* CopyElision(Expression* vexp, Expression* rexp);
 
 	Declaration* ParsePostfixDeclaration(void);
 	Declaration* ReverseDeclaration(Declaration* odec, Declaration* bdec);
 
 	Expression* ParseFunction(Declaration* dec);
-	Expression* ParseAssembler(void);
+	Expression* ParseAssembler(Declaration * vdassm = nullptr);
 
 	Expression* ParseAssemblerBaseOperand(Declaration* pcasm, int pcoffset);
 	Expression* ParseAssemblerMulOperand(Declaration* pcasm, int pcoffset);
@@ -133,7 +137,10 @@ protected:
 
 	Declaration* ParseTypeID(bool tid, Declaration * bdec = nullptr);
 
+	Expression* ParseConstruction(Declaration* type);
+
 	Expression* ParseCastExpression(Expression* exp);
+	Expression* ParseIdentExpression(const Location & eloc, Declaration* dec, bool lhs, bool tid = false);
 	Expression* ParseSimpleExpression(bool lhs, bool tid = false);
 	Expression* ParsePrefixExpression(bool lhs);
 	Expression* ParsePostfixExpression(bool lhs);
@@ -152,9 +159,10 @@ protected:
 	Expression* ParseRExpression(void);
 	
 	Expression* ExpandArgumentPack(Expression * exp, Declaration* dec);
-	Expression* ParseListExpression(bool lhs);
+	Expression* ParseListExpression(bool lhs, Declaration * params = nullptr);
 
 	Expression* ParseParenthesisExpression(void);
+	void ParseStaticAssert(void);
 
 	Errors* mErrors;
 	Scanner* mScanner;

oscar64/Preprocessor.cpp

@@ -44,7 +44,7 @@ bool SourceFile::ReadLineLZO(char* line, ptrdiff_t limit)
 		while (pi < 127 && mPos < mFill)
 		{
 			int	bi = pi, bj = 0;
-			for (int i = 1; i < mPos; i++)
+			for (int i = 1; i <= (mPos < 255 ? mPos : 255); i++)
 			{
 				int j = 0;
 				while (j < 127 && mPos + j < mFill && mBuffer[mPos - i + j] == mBuffer[mPos + j])
@@ -327,6 +327,16 @@ struct CTMHeader9
 	uint8	mColors[7];
 };
 
+struct CTTHeader9
+{
+	uint8	mDispMode;
+	uint8	mColorMethod;
+	uint8	mFlags;
+	uint8	mFgridWidth[2], mFGridHeight[2];
+	char	mFGridConfig;
+	uint8	mColors[6];
+};
+
 #if 0
 #pragma pack(push, 1)
 struct SPDHeader5
@@ -486,6 +496,7 @@ void SourceFile::ReadCharPad(Errors* errors, const Location& location, SourceFil
 	CTMHeader	ctmHeader;
 	CTMHeader8	ctmHeader8;
 	CTMHeader9	ctmHeader9;
+	CTTHeader9	cttHeader9;
 	uint16		ctmMarker, numChars, numTiles;
 	char		tileWidth, tileHeight;
 
@@ -496,10 +507,20 @@ void SourceFile::ReadCharPad(Errors* errors, const Location& location, SourceFil
 		fread(&ctmHeader8, sizeof(CTMHeader8), 1, mFile);
 		break;
 	case 9:
+		if (ctmHeader.mID[2] == 'T')
+		{
+			fread(&cttHeader9, sizeof(CTTHeader9), 1, mFile);
+			ctmHeader8.mDispMode = cttHeader9.mDispMode;
+			ctmHeader8.mColorMethod = cttHeader9.mColorMethod;
+			ctmHeader8.mFlags = cttHeader9.mFlags;
+		}
+		else
+		{
 			fread(&ctmHeader9, sizeof(CTMHeader9), 1, mFile);
 			ctmHeader8.mDispMode = ctmHeader9.mDispMode;
 			ctmHeader8.mColorMethod = ctmHeader9.mColorMethod;
 			ctmHeader8.mFlags = ctmHeader9.mFlags;
+		}
 
 		break;
 	}

oscar64/Scanner.cpp

@@ -51,6 +51,7 @@ const char* TokenNames[] =
 	"'extern'",
 	"'inline'",
 	"'__assume'",
+	"'static_assert'",
 
 	"__asm",
 	"__interrupt",
@@ -173,6 +174,7 @@ const char* TokenNames[] =
 	"'friend'",
 	"'constexpr'",
 	"'typename'",
+	"'decltype'",
 };
 
 
@@ -355,7 +357,7 @@ Scanner::Scanner(Errors* errors, Preprocessor* preprocessor)
 	mToken = TK_NONE;
 	mUngetToken = TK_NONE;
 	mReplay = nullptr;
-	mRecord = mRecordLast = nullptr;
+	mRecord = mRecordLast = mRecordPrev = nullptr;
 
 	mOnceDict = new MacroDict();
 
@@ -372,13 +374,13 @@ Scanner::~Scanner(void)
 
 void Scanner::BeginRecord(void)
 {
-	mRecord = mRecordLast = new TokenSequence(this);
+	mRecord = mRecordLast = mRecordPrev = new TokenSequence(this);
 }
 
 TokenSequence* Scanner::CompleteRecord(void)
 {
 	TokenSequence* seq = mRecord;
-	mRecord = mRecordLast = nullptr;
+	mRecord = mRecordLast = mRecordPrev = nullptr;
 	return seq;
 }
 
@@ -483,6 +485,8 @@ void Scanner::UngetToken(Token token)
 {
 	mUngetToken = mToken;
 	mToken = token;
+	if (mRecord)
+		mRecordLast = mRecordPrev;
 }
 
 void Scanner::NextToken(void)
@@ -491,10 +495,8 @@ void Scanner::NextToken(void)
 	{
 		mToken = mUngetToken;
 		mUngetToken = TK_NONE;
-		return;
 	}
-
+	else if (mReplay)
-	if (mReplay)
 	{
 		mLocation = mReplay->mLocation;
 		mToken = mReplay->mToken;
@@ -515,6 +517,7 @@ void Scanner::NextToken(void)
 
 	if (mRecord)
 	{
+		mRecordPrev = mRecordLast;
 		mRecordLast->mNext = new TokenSequence(this);
 		mRecordLast = mRecordLast->mNext;			
 	}
@@ -819,6 +822,8 @@ void Scanner::NextPreToken(void)
 			}
 		}
 		else if (mToken == TK_PREP_IDENT)
+		{
+			if (mTokenIdent->mString[0])
 			{
 				Macro* def = nullptr;
 				if (mDefineArguments)
@@ -843,6 +848,12 @@ void Scanner::NextPreToken(void)
 				else
 					mErrors->Error(mLocation, EERR_INVALID_PREPROCESSOR, "Invalid preprocessor command", mTokenIdent);
 			}
+			else
+			{
+				mToken = TK_HASH;
+				return;
+			}
+		}
 		else if (mToken == TK_PREP_UNDEF)
 		{
 			NextRawToken();
@@ -1114,14 +1125,15 @@ void Scanner::NextPreToken(void)
 				while (mTokenChar == ' ')
 					NextChar();
 
-				while (mTokenChar == '#' && mLine[mOffset] == '#')
+				if (mTokenChar == '#' && mLine[mOffset] == '#')
 				{
-					mOffset++;
-					NextChar();
-
 					char	tkbase[256];
 					strcpy_s(tkbase, mTokenIdent->mString);
 
+					do {
+						mOffset++;
+						NextChar();
+
 						ptrdiff_t	n = 0;
 						char		tkident[256];
 						while (IsIdentChar(mTokenChar))
@@ -1150,9 +1162,32 @@ void Scanner::NextPreToken(void)
 							n--;
 						tkbase[n] = 0;
 
-					mTokenIdent = Ident::Unique(tkbase);
+						while (mTokenChar == ' ')
-				}
+							NextChar();
 
+					} while (mTokenChar == '#' && mLine[mOffset] == '#');
+
+					ptrdiff_t n = strlen(tkbase);
+					char* str = new char[n + 1];
+					strcpy_s(str, n + 1, tkbase);
+
+					MacroExpansion* ex = new MacroExpansion();
+					ex->mDefinedArguments = mDefineArguments;
+
+					ex->mLine = mLine;
+					ex->mOffset = mOffset;
+					ex->mLink = mMacroExpansion;
+					ex->mChar = mTokenChar;
+
+					mMacroExpansion = ex;
+					mMacroExpansionDepth++;
+					if (mMacroExpansionDepth > 1024)
+						mErrors->Error(mLocation, EFATAL_MACRO_EXPANSION_DEPTH, "Maximum macro expansion depth exceeded", mTokenIdent);
+					mLine = str;
+					mOffset = 0;
+					NextChar();
+				}
+				else
 					return;
 			}
 		}
@@ -1798,6 +1833,8 @@ void Scanner::NextRawToken(void)
 					mToken = TK_ASM;
 				else if (!strcmp(tkident, "__assume"))
 					mToken = TK_ASSUME;
+				else if (!strcmp(tkident, "static_assert"))
+					mToken = TK_STATIC_ASSERT;
 				else if (!strcmp(tkident, "__interrupt"))
 					mToken = TK_INTERRUPT;
 				else if (!strcmp(tkident, "__hwinterrupt"))
@@ -1846,6 +1883,8 @@ void Scanner::NextRawToken(void)
 					mToken = TK_CONSTEXPR;
 				else if ((mCompilerOptions & COPT_CPLUSPLUS) && !strcmp(tkident, "typename"))
 					mToken = TK_TYPENAME;
+				else if ((mCompilerOptions & COPT_CPLUSPLUS) && !strcmp(tkident, "decltype"))
+					mToken = TK_DECLTYPE;
 				else if ((mCompilerOptions & COPT_CPLUSPLUS) && !strcmp(tkident, "operator"))
 				{
 					NextRawToken();
@@ -1975,7 +2014,7 @@ void Scanner::NextRawToken(void)
 
 					default:
 						// dirty little hack to implement token preview, got to fix
-						// this with an infinit preview sequence at one point
+						// this with an infinite preview sequence at one point
 						mUngetToken = mToken;
 						mToken = TK_OPERATOR;
 						return;

oscar64/Scanner.h

@@ -49,6 +49,7 @@ enum Token
 	TK_EXTERN,
 	TK_INLINE,
 	TK_ASSUME,
+	TK_STATIC_ASSERT,
 
 	TK_ASM,
 	TK_INTERRUPT,
@@ -172,6 +173,7 @@ enum Token
 	TK_FRIEND,
 	TK_CONSTEXPR,
 	TK_TYPENAME,
+	TK_DECLTYPE,
 
 	NUM_TOKENS
 };
@@ -315,7 +317,7 @@ protected:
 	Token		mUngetToken;
 
 	const TokenSequence* mReplay;
-	TokenSequence* mRecord, * mRecordLast;
+	TokenSequence* mRecord, * mRecordLast, * mRecordPrev;
 
 	void StringToken(char terminator, char mode);
 	void CharToken(char mode);

oscar64/oscar64.cpp

@@ -76,7 +76,7 @@ int main2(int argc, const char** argv)
 
 #else
 		strcpy(strProductName, "oscar64");
-		strcpy(strProductVersion, "1.31.257");
+		strcpy(strProductVersion, "1.31.261");
 
 #ifdef __APPLE__
 		uint32_t length = sizeof(basePath);
@@ -111,38 +111,12 @@ int main2(int argc, const char** argv)
 		GrowingArray<const char*>	dataFiles(nullptr);
 		GrowingArray<bool>			dataFileCompressed(false);
 
+		bool		emulate = false, profile = false, customCRT = false;
+		int			trace = 0;
+
 		compiler->mPreprocessor->AddPath(basePath);
 		strcpy_s(includePath, basePath);
-		{
+		strcat_s(includePath, "include");
-			FILE* crtFile;
-			char crtFileNamePath[FILENAME_MAX];
-			crtFileNamePath[FILENAME_MAX - 1] = '\0';
-			strcpy_s(crtFileNamePath, basePath);
-			strcat_s(crtFileNamePath, "include/crt.h");
-
-			if (!fopen_s(&crtFile, crtFileNamePath, "r"))
-				strcat_s(includePath, "include/");
-			else
-			{
-				strcpy_s(crtFileNamePath, basePath);
-				strcat_s(crtFileNamePath, "include/oscar64/crt.h");
-
-				if (!fopen_s(&crtFile, crtFileNamePath, "r"))
-					strcat_s(includePath, "include/oscar64/");
-				else
-				{
-					printf("Could not locate Oscar64 includes under %s\n", basePath);
-					return 20;
-				}
-			}
-			fclose(crtFile);
-		}
-		compiler->mPreprocessor->AddPath(includePath);
-		strcpy_s(crtPath, includePath);
-		strcat_s(crtPath, "crt.c");
-
-		bool		emulate = false, profile = false;
-		int			trace = 0;
 
 		targetPath[0] = 0;
 		diskPath[0] = 0;
@@ -185,6 +159,10 @@ int main2(int argc, const char** argv)
 				{
 					compiler->mPreprocessor->AddPath(arg + 3);
 				}
+				else if (arg[1] == 'i' && arg[2] == 'i' && arg[3] == '=')
+				{
+					strcpy_s(includePath, arg + 4);
+				}
 				else if (arg[1] == 'f' && arg[2] == '=')
 				{
 					dataFiles.Push(arg + 3);
@@ -206,6 +184,7 @@ int main2(int argc, const char** argv)
 				else if (arg[1] == 'r' && arg[2] == 't' && arg[3] == '=')
 				{
 					strcpy_s(crtPath, arg + 4);
+					customCRT = true;
 				}
 				else if (arg[1] == 'd' && arg[2] == '6' && arg[3] == '4' && arg[4] == '=')
 				{
@@ -253,6 +232,8 @@ int main2(int argc, const char** argv)
 						compiler->mCompilerOptions |= COPT_OPTIMIZE_ASSEMBLER;
 					else if (arg[2] == 'i' && !arg[3])
 						compiler->mCompilerOptions |= COPT_OPTIMIZE_AUTO_INLINE;
+					else if (arg[2] == 'i' && arg[3] == 'i' && !arg[4])
+						compiler->mCompilerOptions |= COPT_OPTIMIZE_AUTO_INLINE | COPT_OPTIMIZE_AUTO_INLINE_ALL;
 					else if (arg[2] == 'z' && !arg[3])
 						compiler->mCompilerOptions |= COPT_OPTIMIZE_AUTO_ZEROPAGE;
 					else if (arg[2] == 'p' && !arg[3])
@@ -261,6 +242,10 @@ int main2(int argc, const char** argv)
 						compiler->mCompilerOptions |= COPT_OPTIMIZE_GLOBAL;
 					else if (arg[2] == 'm' && !arg[3])
 						compiler->mCompilerOptions |= COPT_OPTIMIZE_MERGE_CALLS;
+					else if (arg[2] == 'o' && !arg[3])
+						compiler->mCompilerOptions |= COPT_OPTIMIZE_OUTLINE;
+					else if (arg[2] == 'x' && !arg[3])
+						compiler->mCompilerOptions |= COPT_OPTIMIZE_PAGE_CROSSING;
 					else
 						compiler->mErrors->Error(loc, EERR_COMMAND_LINE, "Invalid command line argument", arg);
 				}
@@ -321,6 +306,12 @@ int main2(int argc, const char** argv)
 				{
 					compiler->mCompilerOptions |= COPT_CPLUSPLUS;
 					compiler->AddDefine(Ident::Unique("__cplusplus"), "1");
+					compiler->AddDefine(Ident::Unique("__cplusplus__"), "1");
+				}
+				else if (arg[1] == 's' && arg[2] == 't' && arg[3] == 'r' && arg[4] == 'i' && arg[5] == 'c' && arg[6] == 't' && !arg[7])
+				{
+					compiler->mCompilerOptions |= COPT_STRICT;
+					compiler->AddDefine(Ident::Unique("__strict__"), "1");
 					}
 				else if (arg[1] == 'r' && arg[2] == 'm' && arg[3] == 'p' && !arg[4])
 				{
@@ -339,6 +330,7 @@ int main2(int argc, const char** argv)
 				{
 					compiler->mCompilerOptions |= COPT_CPLUSPLUS;
 					compiler->AddDefine(Ident::Unique("__cplusplus"), "1");
+					compiler->AddDefine(Ident::Unique("__cplusplus__"), "1");
 				}
 
 				compiler->mCompilationUnits->AddUnit(loc, argv[i], nullptr);
@@ -435,6 +427,12 @@ int main2(int argc, const char** argv)
 			compiler->mTargetMachine = TMACH_PLUS4;
 			compiler->AddDefine(Ident::Unique("__PLUS4__"), "1");
 		}
+		else if (!strcmp(targetMachine, "mega65"))
+		{
+			strcpy_s(basicStart, "0x2001");		
+			compiler->mTargetMachine = TMACH_MEGA65;
+			compiler->AddDefine(Ident::Unique("__MEGA65__"), "1");
+		}
 		else if (!strcmp(targetMachine, "x16"))
 		{
 			strcpy_s(basicStart, "0x0801");
@@ -559,6 +557,35 @@ int main2(int argc, const char** argv)
 
 			// Add runtime module
 
+			compiler->mPreprocessor->AddPath(includePath);
+
+			if (!customCRT)
+			{
+				FILE* crtFile;
+				char crtFileNamePath[FILENAME_MAX];
+				crtFileNamePath[FILENAME_MAX - 1] = '\0';
+				strcpy_s(crtFileNamePath, includePath);
+				strcat_s(crtFileNamePath, "/crt.h");
+
+				if (fopen_s(&crtFile, crtFileNamePath, "r"))
+				{
+					strcpy_s(crtFileNamePath, basePath);
+					strcat_s(crtFileNamePath, "include/oscar64/crt.h");
+
+					if (!fopen_s(&crtFile, crtFileNamePath, "r"))
+						strcat_s(includePath, "include/oscar64/");
+					else
+					{
+						printf("Could not locate Oscar64 includes under %s\n", basePath);
+						return 20;
+					}
+				}
+				fclose(crtFile);
+
+				strcpy_s(crtPath, includePath);
+				strcat_s(crtPath, "/crt.c");
+			}
+
 			if (crtPath[0])
 				compiler->mCompilationUnits->AddUnit(loc, crtPath, nullptr);
 

oscar64/oscar64.rc

@@ -25,8 +25,8 @@ LANGUAGE LANG_ENGLISH, SUBLANG_NEUTRAL
 //
 
 VS_VERSION_INFO VERSIONINFO
- FILEVERSION 1,31,257,0
+ FILEVERSION 1,31,261,0
- PRODUCTVERSION 1,31,257,0
+ PRODUCTVERSION 1,31,261,0
  FILEFLAGSMASK 0x3fL
 #ifdef _DEBUG
  FILEFLAGS 0x1L
@@ -43,12 +43,12 @@ BEGIN
         BEGIN
             VALUE "CompanyName", "oscar64"
             VALUE "FileDescription", "oscar64 compiler"
-            VALUE "FileVersion", "1.31.257.0"
+            VALUE "FileVersion", "1.31.261.0"
             VALUE "InternalName", "oscar64.exe"
             VALUE "LegalCopyright", "Copyright (C) 2021"
             VALUE "OriginalFilename", "oscar64.exe"
             VALUE "ProductName", "oscar64"
-            VALUE "ProductVersion", "1.31.257.0"
+            VALUE "ProductVersion", "1.31.261.0"
         END
     END
     BLOCK "VarFileInfo"

oscar64/oscar64.vcxproj

@@ -189,6 +189,7 @@
     <ClCompile Include="Linker.cpp" />
     <ClCompile Include="MachineTypes.cpp" />
     <ClCompile Include="NativeCodeGenerator.cpp" />
+    <ClCompile Include="NativeCodeOutliner.cpp" />
     <ClCompile Include="NumberSet.cpp" />
     <ClCompile Include="oscar64.cpp" />
     <ClCompile Include="Parser.cpp" />
@@ -218,6 +219,7 @@
     <ClInclude Include="Linker.h" />
     <ClInclude Include="MachineTypes.h" />
     <ClInclude Include="NativeCodeGenerator.h" />
+    <ClInclude Include="NativeCodeOutliner.h" />
     <ClInclude Include="NumberSet.h" />
     <ClInclude Include="Parser.h" />
     <ClInclude Include="Preprocessor.h" />

oscar64/oscar64.vcxproj.filters

@@ -87,6 +87,9 @@
     <ClCompile Include="Compression.cpp">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="NativeCodeOutliner.cpp">
+      <Filter>Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="Array.h">
@@ -170,6 +173,9 @@
     <ClInclude Include="Compression.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="NativeCodeOutliner.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <ResourceCompile Include="oscar64.rc">

oscar64setup/oscar64setup.vdproj

@@ -6310,15 +6310,15 @@
         {
         "Name" = "8:Microsoft Visual Studio"
         "ProductName" = "8:oscar64"
-        "ProductCode" = "8:{6792319A-5F61-4488-AF65-0F772832E2AC}"
+        "ProductCode" = "8:{1B9A0E10-DEA6-4276-AC91-17782FE44D94}"
-        "PackageCode" = "8:{2E154137-8761-471C-A56E-D8F5BED781E5}"
+        "PackageCode" = "8:{51BFE576-50D1-4F8F-AE7A-9FF07396F53D}"
         "UpgradeCode" = "8:{9AB61EFF-ACAC-4079-9950-8D96615CD4EF}"
         "AspNetVersion" = "8:2.0.50727.0"
         "RestartWWWService" = "11:FALSE"
         "RemovePreviousVersions" = "11:TRUE"
         "DetectNewerInstalledVersion" = "11:TRUE"
         "InstallAllUsers" = "11:FALSE"
-        "ProductVersion" = "8:1.31.257"
+        "ProductVersion" = "8:1.31.261"
         "Manufacturer" = "8:oscar64"
         "ARPHELPTELEPHONE" = "8:"
         "ARPHELPLINK" = "8:"

samples/games/missile.c

@@ -29,7 +29,7 @@ const char MissileChars[] = {
 #define Charset ((char *)0xd800)
 
 // Joystick and crosshair control
-int		CrossX = 160, CrossY = 100;
+volatile int		CrossX = 160, CrossY = 100;
 bool	CrossP = false;
 char	CrossDelay = 0;
 
@@ -553,20 +553,22 @@ __interrupt void joy_interrupt()
 	joy_poll(0);
 
 	// Move crosshair coordinates
-	CrossX += 2 * joyx[0]; CrossY += 2 * joyy[0];
+	int cx = CrossX + 2 * joyx[0], cy = CrossY + 2 * joyy[0];
 
 	// Stop at edges of screen
-	if (CrossX < 8)
+	if (cx < 8)
-		CrossX = 8;
+		cx = 8;
-	else if (CrossX > 312)
+	else if (cx > 312)
-		CrossX = 312;
+		cx = 312;
-	if (CrossY < 20)
+	if (cy < 20)
-		CrossY = 20;
+		cy = 20;
-	else if (CrossY > 172)
+	else if (cy > 172)
-		CrossY = 172;
+		cy = 172;
 
 	// Move crosshair sprite
-	spr_move(0, CrossX + 14, CrossY + 40);	
+	spr_move(0, cx + 14, cy + 40);	
+	CrossX = cx;
+	CrossY = cy;
 
 	// Check button
 	if (joyb[0])
@@ -697,15 +699,17 @@ void game_play(void)
 	// Check if fire request
 	if (CrossP)
 	{
+		int cx = CrossX, cy = CrossY;
+
 		// Find launch site
 		int	sx = 160;
-		if (CrossX < 120)
+		if (cx < 120)
 			sx = 24;
-		else if (CrossX > 200)
+		else if (cx > 200)
 			sx = 296;
 
 		// Fire missile
-		missile_start(sx, 184, CrossX, CrossY);
+		missile_start(sx, 184, cx, cy);
 
 		// Reset request
 		CrossP = false;

samples/hires/cube3d.c

@@ -48,7 +48,7 @@ struct Point
 };
 
 
-Point	tcorners[8], pcorners[8];
+__striped	Point	tcorners[8], pcorners[8];
 
 void drawCube(void)
 {
@@ -77,6 +77,45 @@ void hideCube(void)
 	}
 }
 
+void xorCube(void)
+{
+	for(char i=0; i<8; i++)
+	{
+		if (!(i & 1))
+			bm_line(&Screen, &cr, tcorners[i].x, tcorners[i].y, tcorners[i | 1].x, tcorners[i | 1].y, 0xff, LINOP_XOR);
+		if (!(i & 2))
+			bm_line(&Screen, &cr, tcorners[i].x, tcorners[i].y, tcorners[i | 2].x, tcorners[i | 2].y, 0xff, LINOP_XOR);
+		if (!(i & 4))
+			bm_line(&Screen, &cr, tcorners[i].x, tcorners[i].y, tcorners[i | 4].x, tcorners[i | 4].y, 0xff, LINOP_XOR);
+		pcorners[i] = tcorners[i];
+	}
+}
+
+void xor2Cube(void)
+{
+	for(char i=0; i<8; i++)
+	{
+		if (!(i & 1))
+		{
+			bm_line(&Screen, &cr, pcorners[i].x, pcorners[i].y, pcorners[i | 1].x, pcorners[i | 1].y, 0xff, LINOP_XOR);
+			bm_line(&Screen, &cr, tcorners[i].x, tcorners[i].y, tcorners[i | 1].x, tcorners[i | 1].y, 0xff, LINOP_XOR);
+		}
+		if (!(i & 2))
+		{
+			bm_line(&Screen, &cr, pcorners[i].x, pcorners[i].y, pcorners[i | 2].x, pcorners[i | 2].y, 0xff, LINOP_XOR);
+			bm_line(&Screen, &cr, tcorners[i].x, tcorners[i].y, tcorners[i | 2].x, tcorners[i | 2].y, 0xff, LINOP_XOR);
+		}
+		if (!(i & 4))
+		{
+			bm_line(&Screen, &cr, pcorners[i].x, pcorners[i].y, pcorners[i | 4].x, pcorners[i | 4].y, 0xff, LINOP_XOR);
+			bm_line(&Screen, &cr, tcorners[i].x, tcorners[i].y, tcorners[i | 4].x, tcorners[i | 4].y, 0xff, LINOP_XOR);
+		}
+	}
+
+	for(char i=0; i<8; i++)
+		pcorners[i] = tcorners[i];
+}
+
 #if 1
 
 F12Vector3	corners[8];
@@ -112,8 +151,15 @@ int main(void)
 			tcorners[i].y = lmuldiv16s(vd.v[1], 140, vd.v[2] + 4 * FIX12_ONE) + 100;
 		}
 
+#if 1
+		if (k)
+			xor2Cube();
+		else
+			xorCube();
+#else
 		hideCube();
 		drawCube();
+#endif
 
 	}
 

samples/rasterirq/autocrawler.c

@@ -22,7 +22,7 @@ __interrupt void doscroll(void)
 	vic.color_border++;
 
 	// Update raster IRQ for scroll line with new horizontal scroll offset		
-	rirq_data(&scroll, 0, 7 - (x & 7));
+	rirq_data(&scroll, 1, 7 - (x & 7));
 	// Copy scrolled version of text when switching over char border
 	if ((x & 7) == 0)
 		memcpy((char *)0x0400 + 40 * 24, Text + ((x >> 3) & 255), 40);
@@ -37,11 +37,13 @@ int main(void)
 	rirq_init(true);
 
 	// Build switch to scroll line IRQ
-	rirq_build(&scroll, 1);
+	rirq_build(&scroll, 2);
+	// Delay for one line to get to right border
+	rirq_delay(&scroll, 11);
 	// Change control register two with this IRQ
-	rirq_write(&scroll, 0, &vic.ctrl2, 0);
+	rirq_write(&scroll, 1, &vic.ctrl2, 0);
 	// Put it onto the scroll line
-	rirq_set(0, 50 + 24 * 8, &scroll);
+	rirq_set(0, 49 + 24 * 8, &scroll);
 
 	// Build the switch to normal IRQ
 	rirq_build(&restore, 2);

samples/rasterirq/textcrawler.c

@@ -20,11 +20,13 @@ int main(void)
 	rirq_init(true);
 
 	// Build switch to scroll line IRQ
-	rirq_build(&scroll, 1);
+	rirq_build(&scroll, 2);
+	// Delay for one line to get to right border
+	rirq_delay(&scroll, 11);
 	// Change control register two with this IRQ
-	rirq_write(&scroll, 0, &vic.ctrl2, 0);
+	rirq_write(&scroll, 1, &vic.ctrl2, 0);
 	// Put it onto the scroll line
-	rirq_set(0, 50 + 24 * 8, &scroll);
+	rirq_set(0, 49 + 24 * 8, &scroll);
 
 	// Build the switch to normal IRQ
 	rirq_build(&bottom, 1);
@@ -46,7 +48,7 @@ int main(void)
 		// wait for raster reaching bottom of screen
 		rirq_wait();
 		// Update raster IRQ for scroll line with new horizontal scroll offset		
-		rirq_data(&scroll, 0, 7 - (x & 7));
+		rirq_data(&scroll, 1, 7 - (x & 7));
 		// Copy scrolled version of text when switching over char border
 		if ((x & 7) == 0)
 			memcpy((char *)0x0400 + 40 * 24, Text + ((x >> 3) & 255), 40);

samples/scrolling/bigfont.c

@@ -27,7 +27,7 @@ ScreenRow * const color = (ScreenRow *)0xd800;
 // Move the screen one character to the left
 void scrollLeft(void)
 {
-	// Loop horizontaly
+	// Loop horizontally
 	for(char x=0; x<39; x++)
 	{
 		// Unroll vertical loop 16 times

samples/sprites/creditroll.c

@@ -64,6 +64,7 @@ const char * scrolltext[] = {
 "",
 "",
 "",
+""
 };
 
 
@@ -198,7 +199,7 @@ int main(void)
 		// Update interrupt position
 		for(char i=0; i<5; i++)
 		{
-			int ty = 48 * i + 46 + oy;
+			int ty = 48 * i + 45 + oy;
 
 			// No interrupts below screen bottom
 			if (ty < 250)
@@ -246,7 +247,7 @@ int main(void)
 		case 42:
 			readline(line, lpos);
 			lpos++;
-			if (lpos == 28)
+			if (lpos == 27)
 				lpos = 0;
 			break;
 		case 45: