version 10.5.1

what did it even do?  in any case, $2d is just a user zero page location it should no longer be associated with the monitor nowadays.

.gitignore

@@ -1,9 +1,12 @@
 .idea/workspace.xml
 .idea/discord.xml
 .idea/developer-tools.xml
+.idea/usage.statistics.xml
+.idea/shelf/
 build/
 dist/
 output/
+out/
 .*cache/
 *.directory
 *.prg
@@ -12,7 +15,6 @@ output/
 *.vm.txt
 *.vice-mon-list
 docs/build
-out/
 parser/**/*.interp
 parser/**/*.tokens
 parser/**/*.java
@@ -23,6 +25,7 @@ compiler/src/prog8/buildversion/*
 .eggs/
 /MANIFEST
 .tox/
+.kotlin/
 __pycache__/
 parser.out
 parsetab.py
@@ -31,7 +34,6 @@ parsetab.py
 compiler/lib/
 
 .gradle
-/prog8compiler.jar
 sd*.img
 *.d64
 

.idea/kotlinc.xml

@@ -4,6 +4,6 @@
     <option name="jvmTarget" value="11" />
   </component>
   <component name="KotlinJpsPluginSettings">
-    <option name="version" value="1.9.24" />
+    <option name="version" value="2.0.21" />
   </component>
 </project>

.idea/libraries/KotlinJavaRuntime.xml

@@ -1,23 +1,23 @@
 <component name="libraryTable">
   <library name="KotlinJavaRuntime" type="repository">
-    <properties maven-id="org.jetbrains.kotlin:kotlin-stdlib-jdk8:2.0.20" />
+    <properties maven-id="org.jetbrains.kotlin:kotlin-stdlib-jdk8:2.0.21" />
     <CLASSES>
-      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk8/2.0.20/kotlin-stdlib-jdk8-2.0.20.jar!/" />
-      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib/2.0.20/kotlin-stdlib-2.0.20.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk8/2.0.21/kotlin-stdlib-jdk8-2.0.21.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib/2.0.21/kotlin-stdlib-2.0.21.jar!/" />
       <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/annotations/13.0/annotations-13.0.jar!/" />
-      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk7/2.0.20/kotlin-stdlib-jdk7-2.0.20.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk7/2.0.21/kotlin-stdlib-jdk7-2.0.21.jar!/" />
     </CLASSES>
     <JAVADOC>
-      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk8/2.0.20/kotlin-stdlib-jdk8-2.0.20-javadoc.jar!/" />
-      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib/2.0.20/kotlin-stdlib-2.0.20-javadoc.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk8/2.0.21/kotlin-stdlib-jdk8-2.0.21-javadoc.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib/2.0.21/kotlin-stdlib-2.0.21-javadoc.jar!/" />
       <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/annotations/13.0/annotations-13.0-javadoc.jar!/" />
-      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk7/2.0.20/kotlin-stdlib-jdk7-2.0.20-javadoc.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk7/2.0.21/kotlin-stdlib-jdk7-2.0.21-javadoc.jar!/" />
     </JAVADOC>
     <SOURCES>
-      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk8/2.0.20/kotlin-stdlib-jdk8-2.0.20-sources.jar!/" />
-      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib/2.0.20/kotlin-stdlib-2.0.20-sources.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk8/2.0.21/kotlin-stdlib-jdk8-2.0.21-sources.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib/2.0.21/kotlin-stdlib-2.0.21-sources.jar!/" />
       <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/annotations/13.0/annotations-13.0-sources.jar!/" />
-      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk7/2.0.20/kotlin-stdlib-jdk7-2.0.20-sources.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/jetbrains/kotlin/kotlin-stdlib-jdk7/2.0.21/kotlin-stdlib-jdk7-2.0.21-sources.jar!/" />
     </SOURCES>
   </library>
 </component>

.idea/libraries/eclipse_lsp4j.xml

@@ -0,0 +1,13 @@
+<component name="libraryTable">
+  <library name="eclipse.lsp4j" type="repository">
+    <properties maven-id="org.eclipse.lsp4j:org.eclipse.lsp4j:0.23.1" />
+    <CLASSES>
+      <root url="jar://$MAVEN_REPOSITORY$/org/eclipse/lsp4j/org.eclipse.lsp4j/0.23.1/org.eclipse.lsp4j-0.23.1.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/org/eclipse/lsp4j/org.eclipse.lsp4j.jsonrpc/0.23.1/org.eclipse.lsp4j.jsonrpc-0.23.1.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/com/google/code/gson/gson/2.11.0/gson-2.11.0.jar!/" />
+      <root url="jar://$MAVEN_REPOSITORY$/com/google/errorprone/error_prone_annotations/2.27.0/error_prone_annotations-2.27.0.jar!/" />
+    </CLASSES>
+    <JAVADOC />
+    <SOURCES />
+  </library>
+</component>

.idea/modules.xml

@@ -2,6 +2,8 @@
 <project version="4">
   <component name="ProjectModuleManager">
     <modules>
+      <module fileurl="file://$PROJECT_DIR$/beanshell/beanshell.iml" filepath="$PROJECT_DIR$/beanshell/beanshell.iml" />
+      <module fileurl="file://$PROJECT_DIR$/benchmark-program/benchmark-program.iml" filepath="$PROJECT_DIR$/benchmark-program/benchmark-program.iml" />
       <module fileurl="file://$PROJECT_DIR$/codeCore/codeCore.iml" filepath="$PROJECT_DIR$/codeCore/codeCore.iml" />
       <module fileurl="file://$PROJECT_DIR$/codeGenCpu6502/codeGenCpu6502.iml" filepath="$PROJECT_DIR$/codeGenCpu6502/codeGenCpu6502.iml" />
       <module fileurl="file://$PROJECT_DIR$/codeGenExperimental/codeGenExperimental.iml" filepath="$PROJECT_DIR$/codeGenExperimental/codeGenExperimental.iml" />
@@ -12,10 +14,10 @@
       <module fileurl="file://$PROJECT_DIR$/compilerAst/compilerAst.iml" filepath="$PROJECT_DIR$/compilerAst/compilerAst.iml" />
       <module fileurl="file://$PROJECT_DIR$/docs/docs.iml" filepath="$PROJECT_DIR$/docs/docs.iml" />
       <module fileurl="file://$PROJECT_DIR$/examples/examples.iml" filepath="$PROJECT_DIR$/examples/examples.iml" />
-      <module fileurl="file://$PROJECT_DIR$/benchmark-program/benchmark-program.iml" filepath="$PROJECT_DIR$/benchmark-program/benchmark-program.iml" />
       <module fileurl="file://$PROJECT_DIR$/intermediate/intermediate.iml" filepath="$PROJECT_DIR$/intermediate/intermediate.iml" />
       <module fileurl="file://$PROJECT_DIR$/parser/parser.iml" filepath="$PROJECT_DIR$/parser/parser.iml" />
+      <module fileurl="file://$PROJECT_DIR$/.idea/modules/prog8.iml" filepath="$PROJECT_DIR$/.idea/modules/prog8.iml" />
       <module fileurl="file://$PROJECT_DIR$/virtualmachine/virtualmachine.iml" filepath="$PROJECT_DIR$/virtualmachine/virtualmachine.iml" />
     </modules>
   </component>
-</project>
+</project>

Makefile

@@ -4,9 +4,9 @@
 
 all:
 	gradle installdist installshadowdist
-	@echo "compiler launch script can be found here: compiler/build/install/compiler-shadow/bin/p8compile"
+	@echo "compiler launch script can be found here: compiler/build/install/compiler-shadow/bin/prog8c"
 
 test:
 	gradle build
-	@echo "compiler launch script can be found here: compiler/build/install/compiler-shadow/bin/p8compile"
+	@echo "compiler launch script can be found here: compiler/build/install/compiler-shadow/bin/prog8c"
 

README.md

@@ -66,11 +66,14 @@ What does Prog8 provide?
 - conditional branches that map 1:1 to cpu status flags
 - ``when`` statement to provide a concise jump table alternative to if/elseif chains
 - ``in`` expression for concise and efficient multi-value/containment check 
+- ``defer`` statement to help write concise and robust subroutine cleanup logic
 - several specialized built-in functions such as ``lsb``, ``msb``, ``min``, ``max``, ``rol``, ``ror``
 - various powerful built-in libraries to do I/O, number conversions, graphics and more  
 - inline assembly allows you to have full control when every cycle or byte matters
 - supports the sixteen 'virtual' 16-bit registers R0 - R15 from the Commander X16 (also available on other targets)
 - encode strings and characters into petscii or screencodes or even other encodings
+- Automatic ROM/RAM bank switching on certain compiler targets when calling routines in other banks
+- 50 Kb of available program RAM size on the C64 by default; because Basic ROM is banked out altogether 
 
 *Rapid edit-compile-run-debug cycle:*
 

beanshell/beanshell.iml

@@ -0,0 +1,24 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<module type="JAVA_MODULE" version="4">
+  <component name="NewModuleRootManager" inherit-compiler-output="true">
+    <exclude-output />
+    <content url="file://$MODULE_DIR$">
+      <sourceFolder url="file://$MODULE_DIR$/src" isTestSource="false" />
+      <excludeFolder url="file://$MODULE_DIR$/build" />
+    </content>
+    <orderEntry type="inheritedJdk" />
+    <orderEntry type="sourceFolder" forTests="false" />
+    <orderEntry type="library" name="KotlinJavaRuntime" level="project" />
+    <orderEntry type="library" name="jetbrains.kotlinx.cli.jvm" level="project" />
+    <orderEntry type="library" name="michael.bull.kotlin.result.jvm" level="project" />
+    <orderEntry type="module-library">
+      <library>
+        <CLASSES>
+          <root url="jar://$MODULE_DIR$/lib/bsh-3.0.0-SNAPSHOT.jar!/" />
+        </CLASSES>
+        <JAVADOC />
+        <SOURCES />
+      </library>
+    </orderEntry>
+  </component>
+</module>

beanshell/build.gradle.kts

@@ -0,0 +1,91 @@
+import org.jetbrains.kotlin.gradle.dsl.JvmTarget
+
+plugins {
+    kotlin("jvm")
+    id("application")
+}
+
+val serverMainClassName = "prog8lsp.MainKt"
+val applicationName = "prog8-beanshell"
+val javaVersion: String by project
+
+application {
+    mainClass.set(serverMainClassName)
+    description = "Code completions, diagnostics and more for Prog8"
+    // applicationDefaultJvmArgs = listOf("-DkotlinLanguageServer.version=$version")
+    applicationDistribution.into("bin") {
+        filePermissions {
+            user {
+                read=true
+                execute=true
+                write=true
+            }
+            other.execute = true
+            group.execute = true
+        }
+    }
+}
+
+repositories {
+    mavenCentral()
+}
+
+dependencies {
+    implementation(files("lib/bsh-3.0.0-SNAPSHOT.jar"))
+}
+
+configurations.forEach { config ->
+    config.resolutionStrategy {
+        preferProjectModules()
+    }
+}
+
+sourceSets.main {
+    java.srcDir("src")
+    resources.srcDir("resources")
+}
+
+sourceSets.test {
+    java.srcDir("src")
+    resources.srcDir("resources")
+}
+
+tasks.startScripts {
+    applicationName = "prog8-beanshell"
+}
+
+tasks.register<Exec>("fixFilePermissions") {
+    // When running on macOS or Linux the start script
+    // needs executable permissions to run.
+
+    onlyIf { !System.getProperty("os.name").lowercase().contains("windows") }
+    commandLine("chmod", "+x", "${tasks.installDist.get().destinationDir}/bin/prog8-beanshell")
+}
+
+tasks.withType<Test>() {
+    testLogging {
+        events("failed")
+        exceptionFormat = org.gradle.api.tasks.testing.logging.TestExceptionFormat.FULL
+    }
+}
+
+tasks.installDist {
+    finalizedBy("fixFilePermissions")
+}
+
+tasks.build {
+    finalizedBy("installDist")
+}
+
+java {
+    targetCompatibility = JavaVersion.VERSION_11
+    sourceCompatibility = JavaVersion.VERSION_11
+}
+
+kotlin {
+    compilerOptions.jvmTarget = JvmTarget.JVM_11
+//    jvmToolchain {
+//        languageVersion = JavaLanguageVersion.of(javaVersion.toInt())
+//    }
+}
+

beanshell/src/prog8beanshell/CommandLineReader.kt

@@ -0,0 +1,48 @@
+package prog8beanshell
+
+import java.io.FilterReader
+import java.io.Reader
+
+
+class CommandLineReader(val input: Reader): FilterReader(input)  {
+    private val normal = 0
+    private val lastCharNL = 1
+    private val sentSemi = 2
+    private var state = lastCharNL
+
+    override fun read(): Int {
+        if (state == sentSemi) {
+            this.state = lastCharNL
+            return 10
+        } else {
+            var b = input.read()
+            while(b==13) b = input.read()
+
+            if (b == 10) {
+                if (this.state == lastCharNL) {
+                    b = 59
+                    this.state = sentSemi
+                } else {
+                    this.state = lastCharNL
+                }
+            } else {
+                this.state = normal
+            }
+
+            return b
+        }
+
+    }
+
+    override fun read(buff: CharArray, off: Int, len: Int): Int {
+        val b = read()
+        if (b == -1) {
+            return -1
+        } else {
+            buff[off] = b.toChar()
+            return 1
+        }
+
+    }
+}
+

beanshell/src/prog8beanshell/Main.kt

@@ -0,0 +1,23 @@
+package prog8beanshell
+
+import bsh.FileReader
+import bsh.Interpreter
+
+
+class BeanshellInterpreter {
+
+    fun run(symbols: Map<String, Any>) {
+        val interpreter = Interpreter(CommandLineReader(FileReader(System.`in`)), System.out, System.err, true)
+        interpreter.setExitOnEOF(false)
+        symbols.forEach { (name, value) -> interpreter.set(name, value) }
+        interpreter.run()
+    }
+}
+
+fun main(args: Array<String>) {
+    val i = BeanshellInterpreter()
+    i.run(mapOf(
+        "env" to System.getenv(),
+        "args" to args
+    ))
+}

benchmark-program/Makefile

@@ -1,12 +1,10 @@
-.PHONY:  all clean emu
+.PHONY:  clean run
 
-all:  benchmark.prg
+run:
+	prog8c -target cx16 benchmark.p8
+	x16emu -run -prg benchmark.prg -warp
 
 clean:
 	rm -f *.prg *.PRG *.asm *.vice-* *.BIN *.PAL *.zip *.7z
 
-emu:  benchmark.prg
-	x16emu -run -prg $< -warp
 
-benchmark.prg: benchmark.p8 b_3d.p8 b_adpcm.p8 b_circles.p8 b_life.p8 b_mandelbrot.p8 b_maze.p8 b_queens.p8 b_textelite.p8
-	p8compile $< -target cx16 

benchmark-program/b_circles.p8

@@ -1,4 +1,4 @@
-%import gfx2
+%import gfx_lores
 %import math
 
 circles {
@@ -14,7 +14,7 @@ circles {
         if use_kernal
             void cx16.set_screen_mode(128)
         else
-            gfx2.screen_mode(1)
+            gfx_lores.graphics_mode()
 
         math.rndseed(12345,6789)
         cbm.SETTIM(0,0,0)
@@ -28,14 +28,14 @@ circles {
                 cx16.GRAPH_clear()
             }
             else
-                gfx2.clear_screen(0)
+                gfx_lores.clear_screen(0)
             total_num_circles += draw_circles(use_kernal, max_time)
         }
 
         if use_kernal
             void cx16.set_screen_mode(3)
         else {
-            gfx2.screen_mode(0)
+            gfx_lores.text_mode()
         }
 
         return total_num_circles
@@ -66,7 +66,7 @@ circles {
                         cx16.GRAPH_draw_oval(x-radius, y-radius, radius*2, radius*2, true)
                     }
                     else
-                        gfx2.disc(x, y as ubyte, radius, color)
+                        gfx_lores.disc(x, y as ubyte, radius, color)
                     circle_x[num_circles] = x
                     circle_y[num_circles] = y
                     circle_radius[num_circles] = radius

benchmark-program/benchmark-program.iml

@@ -1,10 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<module type="WEB_MODULE" version="4">
-  <component name="NewModuleRootManager" inherit-compiler-output="true">
-    <exclude-output />
-    <content url="file://$MODULE_DIR$">
-      <excludeFolder url="file://$MODULE_DIR$/compiled" />
-    </content>
-    <orderEntry type="sourceFolder" forTests="false" />
-  </component>
-</module>

benchmark-program/benchmark.p8

@@ -62,7 +62,7 @@ main {
         benchmark_score[benchmark_number]  = adpcm.decode_benchmark(300)
         benchmark_number++
 
-        announce_benchmark("circles with gfx2")
+        announce_benchmark("circles with gfx_lores")
         benchmark_score[benchmark_number]  = circles.draw(false, 300)
         benchmark_number++
 

codeCore/src/prog8/code/SymbolTable.kt

@@ -1,5 +1,6 @@
 package prog8.code
 
+import prog8.code.ast.PtAsmSub
 import prog8.code.ast.PtNode
 import prog8.code.ast.PtProgram
 import prog8.code.core.*
@@ -97,7 +98,7 @@ enum class StNodeType {
     // MODULE,     // not used with current scoping rules
     BLOCK,
     SUBROUTINE,
-    ROMSUB,
+    EXTSUB,
     LABEL,
     STATICVAR,
     MEMVAR,
@@ -178,31 +179,48 @@ open class StNode(val name: String,
 
 class StStaticVariable(name: String,
                        val dt: DataType,
-                       val onetimeInitializationNumericValue: Double?,      // regular (every-run-time) initialization is done via regular assignments
-                       val onetimeInitializationStringValue: StString?,
-                       val onetimeInitializationArrayValue: StArray?,
+                       val initializationStringValue: StString?,
+                       val initializationArrayValue: StArray?,
                        val length: Int?,            // for arrays: the number of elements, for strings: number of characters *including* the terminating 0-byte
                        val zpwish: ZeropageWish,    // used in the variable allocator
+                       val align: Int,
                        astNode: PtNode) : StNode(name, StNodeType.STATICVAR, astNode) {
 
-    val uninitialized = onetimeInitializationArrayValue==null && onetimeInitializationStringValue==null && onetimeInitializationNumericValue==null
+    var initializationNumericValue: Double? = null
+        private set
+
+    fun setOnetimeInitNumeric(number: Double) {
+        // In certain cases the init value of an existing var should be updated,
+        // so we can't ask this as a constructor parameter.
+        // This has to do with the way Prog8 does the (re)initialization of such variables: via code assignment statements.
+        // Certain codegens might want to put them back into the variable directly.
+        // For strings and arrays this doesn't occur - these are always already specced at creation time.
+        initializationNumericValue = number
+    }
+
+    val uninitialized: Boolean
+        get() = initializationArrayValue==null && initializationStringValue==null && initializationNumericValue==null
 
     init {
         if(length!=null) {
-            require(onetimeInitializationNumericValue == null)
-            if(onetimeInitializationArrayValue!=null)
-                require(onetimeInitializationArrayValue.isEmpty() ||onetimeInitializationArrayValue.size==length)
+            require(initializationNumericValue == null)
+            if(initializationArrayValue!=null)
+                require(initializationArrayValue.isEmpty() ||initializationArrayValue.size==length)
         }
-        if(onetimeInitializationNumericValue!=null) {
+        if(initializationNumericValue!=null) {
             require(dt in NumericDatatypes || dt==DataType.BOOL)
         }
-        if(onetimeInitializationArrayValue!=null) {
+        if(initializationArrayValue!=null) {
             require(dt in ArrayDatatypes)
-            require(length==onetimeInitializationArrayValue.size)
+            require(length==initializationArrayValue.size)
         }
-        if(onetimeInitializationStringValue!=null) {
+        if(initializationStringValue!=null) {
             require(dt == DataType.STR)
-            require(length == onetimeInitializationStringValue.first.length+1)
+            require(length == initializationStringValue.first.length+1)
+        }
+        if(align > 0) {
+            require(dt == DataType.STR || dt in ArrayDatatypes)
+            require(zpwish != ZeropageWish.REQUIRE_ZEROPAGE && zpwish != ZeropageWish.PREFER_ZEROPAGE)
         }
     }
 }
@@ -239,18 +257,24 @@ class StSub(name: String, val parameters: List<StSubroutineParameter>, val retur
         StNode(name, StNodeType.SUBROUTINE, astNode)
 
 
-class StRomSub(name: String,
-               val address: UInt?,      // null in case of asmsub, specified in case of romsub
-               val parameters: List<StRomSubParameter>,
-               val returns: List<StRomSubParameter>,
+class StExtSub(name: String,
+               val address: PtAsmSub.Address?,      // null in case of asmsub, specified in case of extsub.
+               val parameters: List<StExtSubParameter>,
+               val returns: List<StExtSubParameter>,
                astNode: PtNode) :
-    StNode(name, StNodeType.ROMSUB, astNode)
+    StNode(name, StNodeType.EXTSUB, astNode)
 
 
 
 class StSubroutineParameter(val name: String, val type: DataType)
-class StRomSubParameter(val register: RegisterOrStatusflag, val type: DataType)
-class StArrayElement(val number: Double?, val addressOfSymbol: String?, val boolean: Boolean?)
+class StExtSubParameter(val register: RegisterOrStatusflag, val type: DataType)
+class StArrayElement(val number: Double?, val addressOfSymbol: String?, val boolean: Boolean?) {
+    init {
+        if(number!=null) require(addressOfSymbol==null && boolean==null)
+        if(addressOfSymbol!=null) require(number==null && boolean==null)
+        if(boolean!=null) require(addressOfSymbol==null && number==null)
+    }
+}
 
 typealias StString = Pair<String, Encoding>
 typealias StArray = List<StArrayElement>

codeCore/src/prog8/code/SymbolTableMaker.kt

@@ -3,7 +3,6 @@ package prog8.code
 import prog8.code.ast.*
 import prog8.code.core.*
 import prog8.code.target.VMTarget
-import kotlin.collections.ArrayDeque
 
 
 class SymbolTableMaker(private val program: PtProgram, private val options: CompilationOptions) {
@@ -39,9 +38,9 @@ class SymbolTableMaker(private val program: PtProgram, private val options: Comp
     private fun addToSt(node: PtNode, scope: ArrayDeque<StNode>) {
         val stNode = when(node) {
             is PtAsmSub -> {
-                val parameters = node.parameters.map { StRomSubParameter(it.first, it.second.type) }
-                val returns = node.returns.map { StRomSubParameter(it.first, it.second) }
-                StRomSub(node.name, node.address, parameters, returns, node)
+                val parameters = node.parameters.map { StExtSubParameter(it.first, it.second.type) }
+                val returns = node.returns.map { StExtSubParameter(it.first, it.second) }
+                StExtSub(node.name, node.address, parameters, returns, node)
             }
             is PtBlock -> {
                 StNode(node.name, StNodeType.BLOCK, node)
@@ -66,23 +65,26 @@ class SymbolTableMaker(private val program: PtProgram, private val options: Comp
                 val numElements: Int?
                 val value = node.value
                 if(value!=null) {
-                    val number = (value as? PtNumber)?.number
-                    initialNumeric = if(number==0.0) null else number       // 0 as init value -> just uninitialized
                     when (value) {
                         is PtString -> {
                             initialString = StString(value.value, value.encoding)
                             initialArray = null
+                            initialNumeric = null
                             numElements = value.value.length + 1   // include the terminating 0-byte
                         }
                         is PtArray -> {
                             initialArray = makeInitialArray(value)
                             initialString = null
+                            initialNumeric = null
                             numElements = initialArray.size
                             require(node.arraySize?.toInt()==numElements)
                         }
                         else -> {
+                            require(value is PtNumber)
                             initialString = null
                             initialArray = null
+                            val number = value.number
+                            initialNumeric = number
                             numElements = node.arraySize?.toInt()
                         }
                     }
@@ -95,7 +97,10 @@ class SymbolTableMaker(private val program: PtProgram, private val options: Comp
 //                if(node.type in SplitWordArrayTypes) {
 //                    ... split array also add _lsb and _msb to symboltable?
 //                }
-                StStaticVariable(node.name, node.type, initialNumeric, initialString, initialArray, numElements, node.zeropage, node)
+                val stVar = StStaticVariable(node.name, node.type, initialString, initialArray, numElements, node.zeropage, node.align.toInt(), node)
+                if(initialNumeric!=null)
+                    stVar.setOnetimeInitNumeric(initialNumeric)
+                stVar
             }
             is PtBuiltinFunctionCall -> {
                 if(node.name=="memory") {
@@ -131,7 +136,6 @@ class SymbolTableMaker(private val program: PtProgram, private val options: Comp
                         TODO("address-of array element $it in initial array value")
                     StArrayElement(null, it.identifier.name, null)
                 }
-                is PtIdentifier -> StArrayElement(null, it.name, null)
                 is PtNumber -> StArrayElement(it.number, null, null)
                 is PtBool -> StArrayElement(null, null, it.value)
                 else -> throw AssemblyError("invalid array element $it")

codeCore/src/prog8/code/ast/AstBase.kt

@@ -80,18 +80,11 @@ class PtBlock(name: String,
               val options: Options,
               position: Position
 ) : PtNamedNode(name, position), IPtStatementContainer {
-    enum class BlockAlignment {
-        NONE,
-        WORD,
-        PAGE
-    }
-
     class Options(val address: UInt? = null,
                   val forceOutput: Boolean = false,
                   val noSymbolPrefixing: Boolean = false,
                   val veraFxMuls: Boolean = false,
-                  val ignoreUnused: Boolean = false,
-                  val alignment: BlockAlignment = BlockAlignment.NONE)
+                  val ignoreUnused: Boolean = false)
 }
 
 
@@ -103,12 +96,20 @@ class PtInlineAssembly(val assembly: String, val isIR: Boolean, position: Positi
 }
 
 
-class PtLabel(name: String, position: Position) : PtNamedNode(name, position)
+class PtLabel(name: String, position: Position) : PtNamedNode(name, position) {
+    companion object {
+        // all automatically generated labels everywhere need to have the same label name prefix:
+        const val GeneratedLabelPrefix = "p8_label_gen_"
+    }
+}
 
 
 class PtBreakpoint(position: Position): PtNode(position)
 
 
+class PtAlign(val align: UInt, position: Position): PtNode(position)
+
+
 class PtIncludeBinary(val file: Path, val offset: UInt?, val length: UInt?, position: Position) : PtNode(position)
 
 

codeCore/src/prog8/code/ast/AstExpressions.kt

@@ -82,10 +82,11 @@ sealed class PtExpression(val type: DataType, position: Position) : PtNode(posit
     }
 
     fun asConstInteger(): Int? = (this as? PtNumber)?.number?.toInt() ?: (this as? PtBool)?.asInt()
+    fun asConstValue(): Double? = (this as? PtNumber)?.number ?: (this as? PtBool)?.asInt()?.toDouble()
 
     fun isSimple(): Boolean {
         return when(this) {
-            is PtAddressOf -> true
+            is PtAddressOf -> this.arrayIndexExpr==null || this.arrayIndexExpr?.isSimple()==true
             is PtArray -> true
             is PtArrayIndexer -> index is PtNumber || index is PtIdentifier
             is PtBinaryExpression -> false
@@ -106,6 +107,7 @@ sealed class PtExpression(val type: DataType, position: Position) : PtNode(posit
             is PtRange -> true
             is PtString -> true
             is PtTypeCast -> value.isSimple()
+            is PtIfExpression -> condition.isSimple() && truevalue.isSimple() && falsevalue.isSimple()
         }
     }
 
@@ -163,6 +165,7 @@ class PtArrayIndexer(elementType: DataType, position: Position): PtExpression(el
 
 
 class PtArray(type: DataType, position: Position): PtExpression(type, position) {
+    // children are always one of 3 types: PtBool, PtNumber or PtAddressOf.
     override fun hashCode(): Int = Objects.hash(children, type)
     override fun equals(other: Any?): Boolean {
         if(other==null || other !is PtArray)
@@ -205,6 +208,16 @@ class PtBinaryExpression(val operator: String, type: DataType, position: Positio
 }
 
 
+class PtIfExpression(type: DataType, position: Position): PtExpression(type, position) {
+    val condition: PtExpression
+        get() = children[0] as PtExpression
+    val truevalue: PtExpression
+        get() = children[1] as PtExpression
+    val falsevalue: PtExpression
+        get() = children[2] as PtExpression
+}
+
+
 class PtContainmentCheck(position: Position): PtExpression(DataType.BOOL, position) {
     val needle: PtExpression
         get() = children[0] as PtExpression
@@ -226,6 +239,13 @@ class PtFunctionCall(val name: String,
                      position: Position) : PtExpression(type, position) {
     val args: List<PtExpression>
         get() = children.map { it as PtExpression }
+
+    init {
+        if(void) require(type==DataType.UNDEFINED) {
+            "void fcall should have undefined datatype"
+        }
+        // note: non-void calls can have UNDEFINED type: is if they return more than 1 value
+    }
 }
 
 
@@ -330,12 +350,11 @@ class PtRange(type: DataType, position: Position) : PtExpression(type, position)
 
         val fromLv = from as? PtNumber
         val toLv = to as? PtNumber
-        val stepLv = step as? PtNumber
-        if(fromLv==null || toLv==null || stepLv==null)
+        if(fromLv==null || toLv==null)
             return null
         val fromVal = fromLv.number.toInt()
         val toVal = toLv.number.toInt()
-        val stepVal = stepLv.number.toInt()
+        val stepVal = step.number.toInt()
         return makeRange(fromVal, toVal, stepVal)
     }
 }
@@ -359,7 +378,3 @@ class PtTypeCast(type: DataType, position: Position) : PtExpression(type, positi
 
 // special node that isn't created from compiling user code, but used internally in the Intermediate Code
 class PtIrRegister(val register: Int, type: DataType, position: Position) : PtExpression(type, position)
-
-
-fun constValue(expr: PtExpression): Double? = if(expr is PtNumber) expr.number else if(expr is PtBool) expr.asInt().toDouble() else null
-fun constIntValue(expr: PtExpression): Int? = if(expr is PtNumber) expr.number.toInt() else if(expr is PtBool) expr.asInt() else null

codeCore/src/prog8/code/ast/AstPrinter.kt

@@ -10,6 +10,7 @@ fun printAst(root: PtNode, skipLibraries: Boolean, output: (text: String) -> Uni
     fun type(dt: DataType) = "!${dt.name.lowercase()}!"
     fun txt(node: PtNode): String {
         return when(node) {
+            is PtAlign -> "%align ${node.align}"
             is PtAssignTarget -> if(node.void) "<void>" else "<target>"
             is PtAssignment -> "<assign>"
             is PtAugmentedAssign -> "<inplace-assign> ${node.operator}"
@@ -24,6 +25,7 @@ fun printAst(root: PtNode, skipLibraries: Boolean, output: (text: String) -> Uni
             is PtArray -> {
                 val valuelist = node.children.map {
                     when (it) {
+                        is PtBool -> it.toString()
                         is PtNumber -> it.number.toString()
                         is PtIdentifier -> it.name
                         else -> "?"
@@ -87,16 +89,18 @@ fun printAst(root: PtNode, skipLibraries: Boolean, output: (text: String) -> Uni
                     }"
                 }
                 val str = if (node.inline) "inline " else ""
-                if(node.address==null) {
+                if(node.address == null) {
                     str + "asmsub ${node.name}($params) $clobbers $returns"
                 } else {
-                    str + "romsub ${node.address.toHex()} = ${node.name}($params) $clobbers $returns"
+                    val bank = if(node.address.constbank!=null) "@bank ${node.address.constbank}"
+                        else if(node.address.varbank!=null) "@bank ${node.address.varbank?.name}"
+                        else ""
+                    str + "extsub $bank ${node.address.address.toHex()} = ${node.name}($params) $clobbers $returns"
                 }
             }
             is PtBlock -> {
                 val addr = if(node.options.address==null) "" else "@${node.options.address.toHex()}"
-                val align = if(node.options.alignment==PtBlock.BlockAlignment.NONE) "" else "align=${node.options.alignment}"
-                "\nblock '${node.name}' $addr $align"
+                "\nblock '${node.name}' $addr"
             }
             is PtConstant -> {
                 val value = when(node.type) {
@@ -125,13 +129,20 @@ fun printAst(root: PtNode, skipLibraries: Boolean, output: (text: String) -> Uni
             }
             is PtVariable -> {
                 val split = if(node.type in SplitWordArrayTypes) "@split" else ""
+                val align = when(node.align) {
+                    0u -> ""
+                    2u -> "@alignword"
+                    64u -> "@align64"
+                    256u -> "@alignpage"
+                    else -> throw IllegalArgumentException("invalid alignment size")
+                }
                 val str = if(node.arraySize!=null) {
                     val eltType = ArrayToElementTypes.getValue(node.type)
-                    "${eltType.name.lowercase()}[${node.arraySize}] $split ${node.name}"
+                    "${eltType.name.lowercase()}[${node.arraySize}] $split $align ${node.name}"
                 }
                 else if(node.type in ArrayDatatypes) {
                     val eltType = ArrayToElementTypes.getValue(node.type)
-                    "${eltType.name.lowercase()}[] $split ${node.name}"
+                    "${eltType.name.lowercase()}[] $split $align ${node.name}"
                 }
                 else
                     "${node.type.name.lowercase()} ${node.name}"
@@ -153,6 +164,8 @@ fun printAst(root: PtNode, skipLibraries: Boolean, output: (text: String) -> Uni
                 else
                     "->"
             }
+            is PtDefer -> "<defer>"
+            is PtIfExpression -> "<ifexpr>"
         }
     }
 

codeCore/src/prog8/code/ast/AstStatements.kt

@@ -10,13 +10,16 @@ sealed interface IPtSubroutine {
 
 class PtAsmSub(
     name: String,
-    val address: UInt?,
+    val address: Address?,
     val clobbers: Set<CpuRegister>,
     val parameters: List<Pair<RegisterOrStatusflag, PtSubroutineParameter>>,
     val returns: List<Pair<RegisterOrStatusflag, DataType>>,
     val inline: Boolean,
     position: Position
-) : PtNamedNode(name, position), IPtSubroutine
+) : PtNamedNode(name, position), IPtSubroutine {
+
+    class Address(val constbank: UByte?, var varbank: PtIdentifier?, val address: UInt)
+}
 
 
 class PtSub(
@@ -131,7 +134,8 @@ class PtRepeatLoop(position: Position) : PtNode(position) {
 
 
 class PtReturn(position: Position) : PtNode(position) {
-    val hasValue = children.any()
+    val hasValue: Boolean
+        get() = children.any()
     val value: PtExpression?
         get() {
             return if(children.any())
@@ -148,7 +152,15 @@ sealed interface IPtVariable {
 }
 
 
-class PtVariable(name: String, override val type: DataType, val zeropage: ZeropageWish, val value: PtExpression?, val arraySize: UInt?, position: Position) : PtNamedNode(name, position), IPtVariable {
+class PtVariable(
+    name: String,
+    override val type: DataType,
+    val zeropage: ZeropageWish,
+    val align: UInt,
+    val value: PtExpression?,
+    val arraySize: UInt?,
+    position: Position
+) : PtNamedNode(name, position), IPtVariable {
     init {
         value?.let {it.parent=this}
     }
@@ -179,3 +191,6 @@ class PtWhenChoice(val isElse: Boolean, position: Position) : PtNode(position) {
     val statements: PtNodeGroup
         get() = children[1] as PtNodeGroup
 }
+
+
+class PtDefer(position: Position): PtNode(position), IPtStatementContainer

codeCore/src/prog8/code/ast/Verify.kt

@@ -0,0 +1,16 @@
+package prog8.code.ast
+
+import prog8.code.SymbolTable
+import prog8.code.core.*
+
+fun verifyFinalAstBeforeAsmGen(program: PtProgram, options: CompilationOptions, st: SymbolTable, errors: IErrorReporter) {
+/*
+    walkAst(program) { node, _ ->
+        if(node is PtVariable) {
+            if(node.value!=null) {
+                // require(node.type in ArrayDatatypes || node.type==DataType.STR) { "final check before asmgen: only string and array variables can still have an init value ${node.name} ${node.type} ${node.position}"}
+            }
+        }
+    }
+*/
+}

codeCore/src/prog8/code/core/BuiltinFunctions.kt

@@ -49,11 +49,14 @@ class FSignature(val pure: Boolean,      // does it have side effects?
         return when {
             actualParamTypes.isEmpty() -> CallConvention(emptyList(), returns)
             actualParamTypes.size==1 -> {
-                // one parameter goes via register/registerpair
+                // One parameter goes via register/registerpair.
+                // this avoids repeated code for every caller to store the value in the subroutine's argument variable.
+                // (that store is still done, but only coded once at the start at the subroutine itself rather than at every call site).
+                // TODO can we start using the X register as well for a third byte or word+byte combo
                 val paramConv = when(val paramType = actualParamTypes[0]) {
                     DataType.UBYTE, DataType.BYTE -> ParamConvention(paramType, RegisterOrPair.A, false)
                     DataType.UWORD, DataType.WORD -> ParamConvention(paramType, RegisterOrPair.AY, false)
-                    DataType.FLOAT -> ParamConvention(paramType, RegisterOrPair.AY, false)
+                    DataType.FLOAT -> ParamConvention(paramType, RegisterOrPair.AY, false)      // TODO is this correct? shouldn't it be FAC1?
                     in PassByReferenceDatatypes -> ParamConvention(paramType, RegisterOrPair.AY, false)
                     else -> ParamConvention(paramType, null, false)
                 }
@@ -79,7 +82,6 @@ val BuiltinFunctions: Map<String, FSignature> = mapOf(
     // cmp returns a status in the carry flag, but not a proper return value
     "cmp"       to FSignature(false, listOf(FParam("value1", IntegerDatatypes), FParam("value2", NumericDatatypes)), null),
     "prog8_lib_stringcompare"     to FSignature(true, listOf(FParam("str1", arrayOf(DataType.STR)), FParam("str2", arrayOf(DataType.STR))), DataType.BYTE),
-    "prog8_lib_arraycopy"         to FSignature(false, listOf(FParam("source", ArrayDatatypes), FParam("target", ArrayDatatypes)), null),
     "prog8_lib_square_byte"       to FSignature(true, listOf(FParam("value", arrayOf(DataType.BYTE, DataType.UBYTE))), DataType.UBYTE),
     "prog8_lib_square_word"       to FSignature(true, listOf(FParam("value", arrayOf(DataType.WORD, DataType.UWORD))), DataType.UWORD),
     "prog8_ifelse_bittest_set"    to FSignature(true, listOf(FParam("variable", ByteDatatypes), FParam("bitnumber", arrayOf(DataType.UBYTE))), DataType.BOOL),

codeCore/src/prog8/code/core/CompilationOptions.kt

@@ -14,6 +14,7 @@ class CompilationOptions(val output: OutputType,
                          val compTarget: ICompilationTarget,
                          // these are set later, based on command line arguments or options in the source code:
                          var loadAddress: UInt,
+                         var memtopAddress: UInt,
                          var warnSymbolShadowing: Boolean = false,
                          var optimize: Boolean = false,
                          var asmQuiet: Boolean = false,
@@ -27,6 +28,7 @@ class CompilationOptions(val output: OutputType,
                          var slabsHighBank: Int? = null,
                          var slabsGolden: Boolean = false,
                          var splitWordArrays: Boolean = false,
+                         var addMissingRts: Boolean = false,        // deprecated, will likely go way in future version
                          var breakpointCpuInstruction: String? = null,
                          var outputDir: Path = Path(""),
                          var symbolDefs: Map<String, String> = emptyMap()

codeCore/src/prog8/code/core/IMachineDefinition.kt

@@ -13,7 +13,9 @@ interface IMachineDefinition {
     val FLOAT_MAX_NEGATIVE: Double
     val FLOAT_MAX_POSITIVE: Double
     val FLOAT_MEM_SIZE: Int
+    val STARTUP_CODE_RESERVED_SIZE: UInt        // this is here, so that certain compiler targets are able to tune this
     val PROGRAM_LOAD_ADDRESS : UInt
+    val PROGRAM_MEMTOP_ADDRESS: UInt
     val BSSHIGHRAM_START: UInt
     val BSSHIGHRAM_END: UInt
     val BSSGOLDENRAM_START: UInt

codeCore/src/prog8/code/optimize/Optimizer.kt

@@ -1,6 +1,6 @@
 package prog8.code.optimize
 
-import prog8.code.StRomSub
+import prog8.code.StExtSub
 import prog8.code.SymbolTable
 import prog8.code.ast.*
 import prog8.code.core.*
@@ -39,7 +39,7 @@ private fun optimizeAssignTargets(program: PtProgram, st: SymbolTable, errors: I
             }
             if(functionName!=null) {
                 val stNode = st.lookup(functionName)
-                if (stNode is StRomSub) {
+                if (stNode is StExtSub) {
                     require(node.children.size==stNode.returns.size+1) {
                         "number of targets must match return values"
                     }
@@ -60,7 +60,7 @@ private fun optimizeAssignTargets(program: PtProgram, st: SymbolTable, errors: I
                 if(node.children.dropLast(1).all { (it as PtAssignTarget).void }) {
                     // all targets are now void, the whole assignment can be discarded and replaced by just a (void) call to the subroutine
                     val index = node.parent.children.indexOf(node)
-                    val voidCall = PtFunctionCall(functionName, true, value.type, value.position)
+                    val voidCall = PtFunctionCall(functionName, true, DataType.UNDEFINED, value.position)
                     value.children.forEach { voidCall.add(it) }
                     node.parent.children[index] = voidCall
                     voidCall.parent = node.parent
@@ -78,38 +78,73 @@ private fun optimizeBitTest(program: PtProgram, options: CompilationOptions): In
     if(options.compTarget.machine.cpu == CpuType.VIRTUAL)
         return 0        // the special bittest optimization is not yet valid for the IR
 
+    fun makeBittestCall(condition: PtBinaryExpression, and: PtBinaryExpression, variable: PtIdentifier, bitmask: Int): PtBuiltinFunctionCall {
+        require(bitmask==128 || bitmask==64)
+        val setOrNot = if(condition.operator=="!=") "set" else "notset"
+        val bittestCall = PtBuiltinFunctionCall("prog8_ifelse_bittest_$setOrNot", false, true, DataType.BOOL, condition.position)
+        bittestCall.add(variable)
+        if(bitmask==128)
+            bittestCall.add(PtNumber(DataType.UBYTE, 7.0, and.right.position))
+        else
+            bittestCall.add(PtNumber(DataType.UBYTE, 6.0, and.right.position))
+        return bittestCall
+    }
+
+    fun isAndByteCondition(condition: PtBinaryExpression?): Triple<PtBinaryExpression, PtIdentifier, Int>? {
+        if(condition!=null && (condition.operator=="==" || condition.operator=="!=")) {
+            if (condition.right.asConstInteger() == 0) {
+                val and = condition.left as? PtBinaryExpression
+                if (and != null && and.operator == "&" && and.type == DataType.UBYTE) {
+                    val bitmask = and.right.asConstInteger()
+                    if(bitmask==128 || bitmask==64) {
+                        val variable = and.left as? PtIdentifier
+                        if (variable != null && variable.type in ByteDatatypes) {
+                            return Triple(and, variable, bitmask)
+                        }
+                        val typecast = and.left as? PtTypeCast
+                        if (typecast != null && typecast.type == DataType.UBYTE) {
+                            val castedVariable = typecast.value as? PtIdentifier
+                            if(castedVariable!=null && castedVariable.type in ByteDatatypes)
+                                return Triple(and, castedVariable, bitmask)
+                        }
+                    }
+                }
+            }
+        }
+        return null
+    }
+
     var changes = 0
     var recurse = true
     walkAst(program) { node: PtNode, depth: Int ->
         if(node is PtIfElse) {
             val condition = node.condition as? PtBinaryExpression
-            if(condition!=null && (condition.operator=="==" || condition.operator=="!=")) {
-                if(condition.right.asConstInteger()==0) {
-                    val and = condition.left as? PtBinaryExpression
-                    if(and != null && and.operator=="&" && and.type == DataType.UBYTE) {
-                        val variable = and.left as? PtIdentifier
-                        val bitmask = and.right.asConstInteger()
-                        if(variable!=null && variable.type in ByteDatatypes && (bitmask==128 || bitmask==64)) {
-                            val setOrNot = if(condition.operator=="!=") "set" else "notset"
-                            val index = node.parent.children.indexOf(node)
-                            val bittestCall = PtBuiltinFunctionCall("prog8_ifelse_bittest_$setOrNot", false, true, DataType.BOOL, node.condition.position)
-                            bittestCall.add(variable)
-                            if(bitmask==128)
-                                bittestCall.add(PtNumber(DataType.UBYTE, 7.0, and.right.position))
-                            else
-                                bittestCall.add(PtNumber(DataType.UBYTE, 6.0, and.right.position))
-                            val ifElse = PtIfElse(node.position)
-                            ifElse.add(bittestCall)
-                            ifElse.add(node.ifScope)
-                            if(node.hasElse())
-                                ifElse.add(node.elseScope)
-                            node.parent.children[index] = ifElse
-                            ifElse.parent = node.parent
-                            changes++
-                            recurse = false
-                        }
-                    }
-                }
+            val check = isAndByteCondition(condition)
+            if(check!=null) {
+                val (and, variable, bitmask) = check
+                val bittestCall = makeBittestCall(condition!!, and, variable, bitmask)
+                val ifElse = PtIfElse(node.position)
+                ifElse.add(bittestCall)
+                ifElse.add(node.ifScope)
+                if (node.hasElse())
+                    ifElse.add(node.elseScope)
+                val index = node.parent.children.indexOf(node)
+                node.parent.children[index] = ifElse
+                ifElse.parent = node.parent
+                changes++
+                recurse = false
+            }
+        }
+        if (node is PtIfExpression) {
+            val condition = node.condition as? PtBinaryExpression
+            val check = isAndByteCondition(condition)
+            if(check!=null) {
+                val (and, variable, bitmask) = check
+                val bittestCall = makeBittestCall(condition!!, and, variable, bitmask)
+                node.children[0] = bittestCall
+                bittestCall.parent = node
+                changes++
+                recurse = false
             }
         }
         recurse

codeCore/src/prog8/code/target/C64Target.kt

@@ -27,6 +27,7 @@ val CompilationTargets = listOf(
     Cx16Target.NAME,
     PETTarget.NAME,
     AtariTarget.NAME,
+    Neo6502Target.NAME,
     VMTarget.NAME
 )
 
@@ -37,5 +38,6 @@ fun getCompilationTargetByName(name: String) = when(name.lowercase()) {
     PETTarget.NAME -> PETTarget()
     AtariTarget.NAME -> AtariTarget()
     VMTarget.NAME -> VMTarget()
+    Neo6502Target.NAME -> Neo6502Target()
     else -> throw IllegalArgumentException("invalid compilation target")
-}
+}

codeCore/src/prog8/code/target/Neo6502Target.kt

@@ -0,0 +1,30 @@
+package prog8.code.target
+
+import prog8.code.core.*
+import prog8.code.target.neo6502.Neo6502MachineDefinition
+
+
+class Neo6502Target: ICompilationTarget, IStringEncoding by Encoder, IMemSizer {
+    override val name = NAME
+    override val machine = Neo6502MachineDefinition()
+    override val defaultEncoding = Encoding.ISO
+
+    companion object {
+        const val NAME = "neo"
+    }
+
+    override fun memorySize(dt: DataType): Int {
+        return when(dt) {
+            in ByteDatatypesWithBoolean -> 1
+            in WordDatatypes, in PassByReferenceDatatypes -> 2
+            DataType.FLOAT -> machine.FLOAT_MEM_SIZE
+            else -> throw IllegalArgumentException("invalid datatype")
+        }
+    }
+
+    override fun memorySize(arrayDt: DataType, numElements: Int) =
+        if(arrayDt== DataType.UWORD)
+            numElements    // pointer to bytes.
+        else
+            memorySize(ArrayToElementTypes.getValue(arrayDt)) * numElements
+}

codeCore/src/prog8/code/target/atari/AtariMachineDefinition.kt

@@ -11,7 +11,9 @@ class AtariMachineDefinition: IMachineDefinition {
     override val FLOAT_MAX_POSITIVE = 9.999999999e97
     override val FLOAT_MAX_NEGATIVE = -9.999999999e97
     override val FLOAT_MEM_SIZE = 6
+    override val STARTUP_CODE_RESERVED_SIZE = 20u
     override val PROGRAM_LOAD_ADDRESS = 0x2000u
+    override val PROGRAM_MEMTOP_ADDRESS = 0xffffu  // TODO what's memtop?
 
     override val BSSHIGHRAM_START = 0u    // TODO
     override val BSSHIGHRAM_END = 0u      // TODO

codeCore/src/prog8/code/target/c128/C128MachineDefinition.kt

@@ -13,7 +13,9 @@ class C128MachineDefinition: IMachineDefinition {
     override val FLOAT_MAX_POSITIVE = Mflpt5.FLOAT_MAX_POSITIVE
     override val FLOAT_MAX_NEGATIVE = Mflpt5.FLOAT_MAX_NEGATIVE
     override val FLOAT_MEM_SIZE = Mflpt5.FLOAT_MEM_SIZE
+    override val STARTUP_CODE_RESERVED_SIZE = 20u
     override val PROGRAM_LOAD_ADDRESS = 0x1c01u
+    override val PROGRAM_MEMTOP_ADDRESS = 0xff00u
 
     override val BSSHIGHRAM_START = 0u    // TODO
     override val BSSHIGHRAM_END = 0u      // TODO

codeCore/src/prog8/code/target/c64/C64MachineDefinition.kt

@@ -14,7 +14,10 @@ class C64MachineDefinition: IMachineDefinition {
     override val FLOAT_MAX_POSITIVE = Mflpt5.FLOAT_MAX_POSITIVE
     override val FLOAT_MAX_NEGATIVE = Mflpt5.FLOAT_MAX_NEGATIVE
     override val FLOAT_MEM_SIZE = Mflpt5.FLOAT_MEM_SIZE
+    override val STARTUP_CODE_RESERVED_SIZE = 20u
     override val PROGRAM_LOAD_ADDRESS = 0x0801u
+    override val PROGRAM_MEMTOP_ADDRESS = 0xcfe0u      // $a000  if floats are used
+            // note that at $cfe0-$cfff are the 16 'virtual registers' R0-R15
 
     override val BSSHIGHRAM_START = 0xc000u
     override val BSSHIGHRAM_END = 0xcfffu

codeCore/src/prog8/code/target/cx16/CX16MachineDefinition.kt

@@ -13,7 +13,9 @@ class CX16MachineDefinition: IMachineDefinition {
     override val FLOAT_MAX_POSITIVE = Mflpt5.FLOAT_MAX_POSITIVE
     override val FLOAT_MAX_NEGATIVE = Mflpt5.FLOAT_MAX_NEGATIVE
     override val FLOAT_MEM_SIZE = Mflpt5.FLOAT_MEM_SIZE
+    override val STARTUP_CODE_RESERVED_SIZE = 20u
     override val PROGRAM_LOAD_ADDRESS = 0x0801u
+    override val PROGRAM_MEMTOP_ADDRESS = 0x9f00u
 
     override val BSSHIGHRAM_START = 0xa000u     // hiram bank 1, 8Kb, assumed to be active
     override val BSSHIGHRAM_END = 0xbfffu       // Rom starts at $c000

codeCore/src/prog8/code/target/neo6502/Neo6502MachineDefinition.kt

@@ -0,0 +1,50 @@
+package prog8.code.target.neo6502
+
+import prog8.code.core.*
+import java.nio.file.Path
+
+
+class Neo6502MachineDefinition: IMachineDefinition {
+
+    override val cpu = CpuType.CPU65c02
+
+    override val FLOAT_MAX_POSITIVE = 9.999999999e97
+    override val FLOAT_MAX_NEGATIVE = -9.999999999e97
+    override val FLOAT_MEM_SIZE = 6
+    override val STARTUP_CODE_RESERVED_SIZE = 20u
+    override val PROGRAM_LOAD_ADDRESS = 0x0800u
+    override val PROGRAM_MEMTOP_ADDRESS = 0xfc00u       // kernal starts here
+
+    override val BSSHIGHRAM_START = 0u    // TODO
+    override val BSSHIGHRAM_END = 0u      // TODO
+    override val BSSGOLDENRAM_START = 0u  // TODO
+    override val BSSGOLDENRAM_END = 0u    // TODO
+
+    override lateinit var zeropage: Zeropage
+    override lateinit var golden: GoldenRam
+
+    override fun getFloatAsmBytes(num: Number) = TODO("atari float asm bytes from number")
+    override fun convertFloatToBytes(num: Double): List<UByte> = TODO("atari float to bytes")
+    override fun convertBytesToFloat(bytes: List<UByte>): Double = TODO("atari bytes to float")
+
+    override fun launchEmulator(selectedEmulator: Int, programNameWithPath: Path) {
+        if(selectedEmulator!=1) {
+            System.err.println("The neo target only supports the main emulator (neo).")
+            return
+        }
+
+        val cmdline = listOf("neo", "${programNameWithPath}.bin@800", "cold")
+
+        println("\nStarting Neo6502 emulator...")
+        val processb = ProcessBuilder(cmdline).inheritIO()
+        val process: Process = processb.start()
+        process.waitFor()
+    }
+
+    override fun isIOAddress(address: UInt): Boolean = address in 0xff00u..0xff0fu
+
+    override fun initializeMemoryAreas(compilerOptions: CompilationOptions) {
+        zeropage = Neo6502Zeropage(compilerOptions)
+        golden = GoldenRam(compilerOptions, UIntRange.EMPTY)
+    }
+}

codeCore/src/prog8/code/target/neo6502/Neo6502Zeropage.kt

@@ -0,0 +1,48 @@
+package prog8.code.target.neo6502
+
+import prog8.code.core.*
+
+class Neo6502Zeropage(options: CompilationOptions) : Zeropage(options) {
+
+    override val SCRATCH_B1 = 0xfau      // temp storage for a single byte
+    override val SCRATCH_REG = 0xfbu     // temp storage for a register, must be B1+1
+    override val SCRATCH_W1 = 0xfcu      // temp storage 1 for a word  $fc+$fd
+    override val SCRATCH_W2 = 0xfeu      // temp storage 2 for a word  $fe+$ff
+
+    init {
+        if (options.floats) {
+            throw InternalCompilerException("Neo6502 target doesn't support floating point routines")
+        }
+
+        when (options.zeropage) {
+            ZeropageType.DONTUSE -> {
+                free.clear()  // don't use zeropage at all
+            }
+            else -> {
+                free.addAll(0x22u..0xffu)
+            }
+        }
+
+        val distinctFree = free.distinct()
+        free.clear()
+        free.addAll(distinctFree)
+
+        removeReservedFromFreePool()
+        allocateCx16VirtualRegisters()
+        retainAllowed()
+    }
+
+    override fun allocateCx16VirtualRegisters() {
+        // Note: the 16 virtual registers R0-R15 are not regular allocated variables, they're *memory mapped* elsewhere to fixed addresses.
+        // However, to be able for the compiler to "see" them as zero page variables, we have to register them here as well.
+        // This is important because the compiler sometimes treats ZP variables more efficiently (for example if it's a pointer)
+        for(reg in 0..15) {
+            allocatedVariables["cx16.r${reg}"]   = VarAllocation((2+reg*2).toUInt(), DataType.UWORD, 2)       // cx16.r0 .. cx16.r15
+            allocatedVariables["cx16.r${reg}s"]  = VarAllocation((2+reg*2).toUInt(), DataType.WORD, 2)        // cx16.r0s .. cx16.r15s
+            allocatedVariables["cx16.r${reg}L"]  = VarAllocation((2+reg*2).toUInt(), DataType.UBYTE, 1)       // cx16.r0L .. cx16.r15L
+            allocatedVariables["cx16.r${reg}H"]  = VarAllocation((3+reg*2).toUInt(), DataType.UBYTE, 1)       // cx16.r0H .. cx16.r15H
+            allocatedVariables["cx16.r${reg}sL"] = VarAllocation((2+reg*2).toUInt(), DataType.BYTE, 1)        // cx16.r0sL .. cx16.r15sL
+            allocatedVariables["cx16.r${reg}sH"] = VarAllocation((3+reg*2).toUInt(), DataType.BYTE, 1)        // cx16.r0sH .. cx16.r15sH
+        }
+    }
+}

codeCore/src/prog8/code/target/pet/PETMachineDefinition.kt

@@ -13,7 +13,9 @@ class PETMachineDefinition: IMachineDefinition {
     override val FLOAT_MAX_POSITIVE = Mflpt5.FLOAT_MAX_POSITIVE
     override val FLOAT_MAX_NEGATIVE = Mflpt5.FLOAT_MAX_NEGATIVE
     override val FLOAT_MEM_SIZE = Mflpt5.FLOAT_MEM_SIZE
+    override val STARTUP_CODE_RESERVED_SIZE = 20u
     override val PROGRAM_LOAD_ADDRESS = 0x0401u
+    override val PROGRAM_MEMTOP_ADDRESS = 0x8000u
 
     override val BSSHIGHRAM_START = 0u
     override val BSSHIGHRAM_END = 0u

codeCore/src/prog8/code/target/virtual/VirtualMachineDefinition.kt

@@ -13,7 +13,9 @@ class VirtualMachineDefinition: IMachineDefinition {
     override val FLOAT_MAX_POSITIVE = Float.MAX_VALUE.toDouble()
     override val FLOAT_MAX_NEGATIVE = -Float.MAX_VALUE.toDouble()
     override val FLOAT_MEM_SIZE = 8             // 64-bits double
+    override val STARTUP_CODE_RESERVED_SIZE = 0u  // not actually used
     override val PROGRAM_LOAD_ADDRESS = 0u      // not actually used
+    override val PROGRAM_MEMTOP_ADDRESS = 0xffffu  // not actually used
 
     override val BSSHIGHRAM_START = 0u          // not actually used
     override val BSSHIGHRAM_END = 0u            // not actually used

codeGenCpu6502/src/prog8/codegen/cpu6502/AsmGen.kt

@@ -9,7 +9,6 @@ import prog8.code.ast.*
 import prog8.code.core.*
 import prog8.code.target.Cx16Target
 import prog8.codegen.cpu6502.assignment.*
-import kotlin.collections.ArrayDeque
 import kotlin.io.path.Path
 import kotlin.io.path.writeLines
 
@@ -17,15 +16,15 @@ import kotlin.io.path.writeLines
 internal const val subroutineFloatEvalResultVar1 = "prog8_float_eval_result1"
 internal const val subroutineFloatEvalResultVar2 = "prog8_float_eval_result2"
 
-class AsmGen6502(val prefixSymbols: Boolean): ICodeGeneratorBackend {
+class AsmGen6502(val prefixSymbols: Boolean, private val lastGeneratedLabelSequenceNr: Int): ICodeGeneratorBackend {
     override fun generate(
         program: PtProgram,
         symbolTable: SymbolTable,
         options: CompilationOptions,
-        errors: IErrorReporter
+        errors: IErrorReporter,
     ): IAssemblyProgram? {
         val st = if(prefixSymbols) prefixSymbols(program, options, symbolTable) else symbolTable
-        val asmgen = AsmGen6502Internal(program, st, options, errors)
+        val asmgen = AsmGen6502Internal(program, st, options, errors, lastGeneratedLabelSequenceNr)
         return asmgen.compileToAssembly()
     }
 
@@ -37,10 +36,9 @@ class AsmGen6502(val prefixSymbols: Boolean): ICodeGeneratorBackend {
             when(node) {
                 is PtAsmSub, is PtSub -> node.name = "p8s_${node.name}"
                 is PtBlock -> node.name = "p8b_${node.name}"
-                is PtLabel -> node.name = "p8l_${node.name}"
+                is PtLabel -> if(!node.name.startsWith(PtLabel.GeneratedLabelPrefix)) node.name = "p8l_${node.name}"   // don't prefix autogenerated labels
                 is PtConstant -> node.name = "p8c_${node.name}"
                 is PtVariable, is PtMemMapped, is PtSubroutineParameter -> node.name = "p8v_${node.name}"
-                else -> node.name = "p8_${node.name}"
             }
         }
 
@@ -49,6 +47,9 @@ class AsmGen6502(val prefixSymbols: Boolean): ICodeGeneratorBackend {
                 is PtAsmSub -> {
                     prefixNamedNode(node)
                     node.parameters.forEach { (_, param) -> prefixNamedNode(param) }
+                    if(node.address?.varbank!=null) {
+                        node.address!!.varbank = node.address!!.varbank!!.prefix(node, st)
+                    }
                 }
                 is PtSub -> {
                     prefixNamedNode(node)
@@ -124,11 +125,15 @@ class AsmGen6502(val prefixSymbols: Boolean): ICodeGeneratorBackend {
 }
 
 private fun prefixScopedName(name: String, type: Char): String {
-    if('.' !in name)
+    if('.' !in name) {
+        if(name.startsWith(PtLabel.GeneratedLabelPrefix))
+            return name
         return "p8${type}_$name"
+    }
     val parts = name.split('.')
     val firstPrefixed = "p8b_${parts[0]}"
-    val lastPrefixed = "p8${type}_${parts.last()}"
+    val lastPart = parts.last()
+    val lastPrefixed = if(lastPart.startsWith(PtLabel.GeneratedLabelPrefix)) lastPart else  "p8${type}_$lastPart"
     // the parts in between are assumed to be subroutine scopes.
     val inbetweenPrefixed = parts.drop(1).dropLast(1).map{ "p8s_$it" }
     val prefixed = listOf(firstPrefixed) + inbetweenPrefixed + listOf(lastPrefixed)
@@ -145,7 +150,6 @@ private fun PtVariable.prefix(parent: PtNode, st: SymbolTable): PtVariable {
         val newValue = PtArray(arrayValue.type, arrayValue.position)
         arrayValue.children.forEach { elt ->
             when(elt) {
-                is PtIdentifier -> newValue.add(elt.prefix(arrayValue, st))
                 is PtBool -> newValue.add(elt)
                 is PtNumber -> newValue.add(elt)
                 is PtAddressOf -> {
@@ -163,7 +167,7 @@ private fun PtVariable.prefix(parent: PtNode, st: SymbolTable): PtVariable {
                 else -> throw AssemblyError("weird array value element $elt")
             }
         }
-        val result = PtVariable(name, type, zeropage, newValue, arraySize, position)
+        val result = PtVariable(name, type, zeropage, align, newValue, arraySize, position)
         result.parent = parent
         result
     }
@@ -201,7 +205,7 @@ private fun PtIdentifier.prefix(parent: PtNode, st: SymbolTable): PtIdentifier {
 
     val prefixType = when(target!!.type) {
         StNodeType.BLOCK -> 'b'
-        StNodeType.SUBROUTINE, StNodeType.ROMSUB -> 's'
+        StNodeType.SUBROUTINE, StNodeType.EXTSUB -> 's'
         StNodeType.LABEL -> 'l'
         StNodeType.STATICVAR, StNodeType.MEMVAR -> 'v'
         StNodeType.CONSTANT -> 'c'
@@ -220,7 +224,8 @@ class AsmGen6502Internal (
     val program: PtProgram,
     internal val symbolTable: SymbolTable,
     internal val options: CompilationOptions,
-    internal val errors: IErrorReporter
+    internal val errors: IErrorReporter,
+    private var generatedLabelSequenceNumber: Int
 ) {
 
     internal val optimizedByteMultiplications = setOf(3,5,6,7,9,10,11,12,13,14,15,20,25,40,50,80,100)
@@ -582,6 +587,7 @@ class AsmGen6502Internal (
     internal fun translate(stmt: PtNode) {
         outputSourceLine(stmt)
         when(stmt) {
+            is PtAlign -> if(stmt.align > 1u) out("  .align  ${stmt.align.toHex()}")
             is PtReturn -> translate(stmt)
             is PtSub -> programGen.translateSubroutine(stmt)
             is PtAsmSub -> programGen.translateAsmSubroutine(stmt)
@@ -604,9 +610,10 @@ class AsmGen6502Internal (
             is PtRepeatLoop -> translate(stmt)
             is PtWhen -> translate(stmt)
             is PtIncludeBinary -> translate(stmt)
-            is PtBreakpoint -> translate(stmt)
+            is PtBreakpoint -> translateBrk()
             is PtVariable, is PtConstant, is PtMemMapped -> { /* do nothing; variables are handled elsewhere */ }
             is PtBlock -> throw AssemblyError("block should have been handled elsewhere")
+            is PtDefer -> throw AssemblyError("defer should have been transformed")
             is PtNodeGroup -> stmt.children.forEach { translate(it) }
             is PtNop -> {}
             else -> throw AssemblyError("missing asm translation for $stmt")
@@ -782,7 +789,7 @@ class AsmGen6502Internal (
                 when {
                     iterations == 0 -> {}
                     iterations == 1 -> translate(stmt.statements)
-                    iterations<0 || iterations>65535 -> throw AssemblyError("invalid number of iterations")
+                    iterations<0 || iterations>65536 -> throw AssemblyError("invalid number of iterations")
                     iterations <= 256 -> repeatByteCount(iterations, stmt)
                     else -> repeatWordCount(iterations, stmt)
                 }
@@ -822,10 +829,10 @@ class AsmGen6502Internal (
     }
 
     private fun repeatWordCount(iterations: Int, stmt: PtRepeatLoop) {
-        require(iterations in 257..65535) { "invalid repeat count ${stmt.position}" }
+        require(iterations in 257..65536) { "invalid repeat count ${stmt.position}" }
         val repeatLabel = makeLabel("repeat")
         val counterVar = createRepeatCounterVar(DataType.UWORD, isTargetCpu(CpuType.CPU65c02), stmt)
-        val loopcount = if(iterations and 0x00ff == 0) iterations else iterations + 0x0100   // so that the loop can simply use a double-dec
+        val loopcount = if(iterations==65536) 0 else if(iterations and 0x00ff == 0) iterations else iterations + 0x0100   // so that the loop can simply use a double-dec
         out("""
             ldy  #>$loopcount
             lda  #<$loopcount
@@ -1040,7 +1047,7 @@ $repeatLabel""")
         }
     }
 
-    private fun translate(ret: PtReturn, withRts: Boolean=true) {
+    private fun translate(ret: PtReturn) {
         ret.value?.let { returnvalue ->
             val sub = ret.definingSub()!!
             val returnReg = sub.returnRegister()!!
@@ -1057,9 +1064,7 @@ $repeatLabel""")
                 }
             }
         }
-
-        if(withRts)
-            out("  rts")
+        out("  rts")
     }
 
     private fun translate(asm: PtInlineAssembly) {
@@ -1084,7 +1089,7 @@ $repeatLabel""")
         out("  .binary \"$pathForAssembler\" $offset $length")
     }
 
-    private fun translate(brk: PtBreakpoint) {
+    private fun translateBrk() {
         val label = "_prog8_breakpoint_${breakpointLabels.size+1}"
         breakpointLabels.add(label)
         out(label)
@@ -1363,11 +1368,9 @@ $repeatLabel""")
             extra
     }
 
-    private var generatedLabelSequenceNumber: Int = 0
-
     internal fun makeLabel(postfix: String): String {
         generatedLabelSequenceNumber++
-        return "label_asm_${generatedLabelSequenceNumber}_$postfix"
+        return "${PtLabel.GeneratedLabelPrefix}${generatedLabelSequenceNumber}_$postfix"
     }
 
     fun assignConstFloatToPointerAY(number: PtNumber) {
@@ -1381,6 +1384,10 @@ $repeatLabel""")
             pla
             jsr  floats.copy_float""")
     }
+
+    internal fun assignIfExpression(target: AsmAssignTarget, value: PtIfExpression) {
+        ifElseAsmgen.assignIfExpression(target, value)
+    }
 }
 
 /**

codeGenCpu6502/src/prog8/codegen/cpu6502/AsmOptimizer.kt

@@ -3,6 +3,7 @@ package prog8.codegen.cpu6502
 import prog8.code.StConstant
 import prog8.code.StMemVar
 import prog8.code.SymbolTable
+import prog8.code.ast.PtLabel
 import prog8.code.core.IMachineDefinition
 
 
@@ -361,15 +362,16 @@ or *_afterif labels.
 This gets generated after certain if conditions, and only the branch instruction is needed in these cases.
          */
 
+        val autoLabelPrefix = PtLabel.GeneratedLabelPrefix
         if(first=="beq  +" && second=="lda  #1" && third=="+") {
-            if((fourth.startsWith("beq  label_") || fourth.startsWith("bne  label_")) &&
+            if((fourth.startsWith("beq  $autoLabelPrefix") || fourth.startsWith("bne  $autoLabelPrefix")) &&
                 (fourth.endsWith("_shortcut") || fourth.endsWith("_afterif") || fourth.endsWith("_shortcut:") || fourth.endsWith("_afterif:"))) {
                 mods.add(Modification(lines[0].index, true, null))
                 mods.add(Modification(lines[1].index, true, null))
                 mods.add(Modification(lines[2].index, true, null))
             }
-            else if(fourth.startsWith("label_") && (fourth.endsWith("_shortcut") || fourth.endsWith("_shortcut:"))) {
-                if((fifth.startsWith("beq  label_") || fifth.startsWith("bne  label_")) &&
+            else if(fourth.startsWith(autoLabelPrefix) && (fourth.endsWith("_shortcut") || fourth.endsWith("_shortcut:"))) {
+                if((fifth.startsWith("beq  $autoLabelPrefix") || fifth.startsWith("bne  $autoLabelPrefix")) &&
                     (fifth.endsWith("_shortcut") || fifth.endsWith("_afterif") || fifth.endsWith("_shortcut:") || fifth.endsWith("_afterif:"))) {
                     mods.add(Modification(lines[0].index, true, null))
                     mods.add(Modification(lines[1].index, true, null))
@@ -519,8 +521,10 @@ private fun optimizeJsrRtsAndOtherCombinations(linesByFour: Sequence<List<Indexe
 
         if(!haslabel(second)) {
             if ((" jsr" in first || "\tjsr" in first ) && (" rts" in second || "\trts" in second)) {
-                mods += Modification(lines[0].index, false, lines[0].value.replace("jsr", "jmp"))
-                mods += Modification(lines[1].index, true, null)
+                if("floats.pushFAC" !in first && "floats.popFAC" !in first) {       // these 2 routines depend on being called with JSR!!
+                    mods += Modification(lines[0].index, false, lines[0].value.replace("jsr", "jmp"))
+                    mods += Modification(lines[1].index, true, null)
+                }
             }
             else if (" rts" in first || "\trts" in first) {
                 if (" jmp" in second || "\tjmp" in second)

codeGenCpu6502/src/prog8/codegen/cpu6502/AsmsubHelpers.kt

@@ -11,18 +11,22 @@ fun asmsub6502ArgsEvalOrder(sub: PtAsmSub): List<Int> {
     //  1) cx16 virtual word registers,
     //  2) paired CPU registers,
     //  3) single CPU registers (order Y,X,A),
-    //  4) CPU Carry status flag
+    //  4) floating point registers (FAC1, FAC2),
+    //  5) CPU Carry status flag
     val args = sub.parameters.withIndex()
     val (cx16regs, args2) = args.partition { it.value.first.registerOrPair in Cx16VirtualRegisters }
     val pairedRegisters = arrayOf(RegisterOrPair.AX, RegisterOrPair.AY, RegisterOrPair.XY)
     val (pairedRegs , args3) = args2.partition { it.value.first.registerOrPair in pairedRegisters }
-    val (singleRegs, rest) = args3.partition { it.value.first.registerOrPair != null }
+    val (singleRegsMixed, rest) = args3.partition { it.value.first.registerOrPair != null }
+    val (singleCpuRegs, floatRegs) = singleRegsMixed.partition {it.value.first.registerOrPair != RegisterOrPair.FAC1 && it.value.first.registerOrPair != RegisterOrPair.FAC2  }
 
     cx16regs.forEach { order += it.index }
     pairedRegs.forEach { order += it.index }
-    singleRegs.sortedBy { it.value.first.registerOrPair!!.asCpuRegister() }.asReversed().forEach { order += it.index }
+    singleCpuRegs.sortedBy { it.value.first.registerOrPair!!.asCpuRegister() }.asReversed().forEach { order += it.index }
     require(rest.all { it.value.first.registerOrPair==null && it.value.first.statusflag!=null})
+    floatRegs.forEach { order += it.index }
     rest.forEach { order += it.index }
     require(order.size==sub.parameters.size)
+
     return order
 }

codeGenCpu6502/src/prog8/codegen/cpu6502/AssemblyProgram.kt

@@ -1,13 +1,16 @@
 package prog8.codegen.cpu6502
 
+import prog8.code.ast.PtLabel
 import prog8.code.core.*
+import prog8.code.target.AtariTarget
 import prog8.code.target.C64Target
+import prog8.code.target.Neo6502Target
 import java.nio.file.Path
 
 
 internal class AssemblyProgram(
         override val name: String,
-        outputDir: Path,
+        private val outputDir: Path,
         private val compTarget: ICompilationTarget) : IAssemblyProgram {
 
     private val assemblyFile = outputDir.resolve("$name.asm")
@@ -16,6 +19,7 @@ internal class AssemblyProgram(
     private val binFile = outputDir.resolve("$name.bin")
     private val viceMonListFile = outputDir.resolve(C64Target.viceMonListName(name))
     private val listFile = outputDir.resolve("$name.list")
+    private val targetWithoutBreakpointsForEmulator = setOf(AtariTarget.NAME, Neo6502Target.NAME)
 
     override fun assemble(options: CompilationOptions, errors: IErrorReporter): Boolean {
 
@@ -96,12 +100,47 @@ internal class AssemblyProgram(
                 command.addAll(listOf("--output", outFile.toString(), assemblyFile.toString()))
                 assemblerCommand = command
             }
+            "neo" -> {
+                // Neo6502 raw program generation.
+
+                if(options.output!=OutputType.RAW || options.loadAddress!=0x0800u || options.launcher!=CbmPrgLauncherType.NONE) {
+                    throw AssemblyError("invalid program compilation options. Neo6502 requires %output raw, %launcher none, %address $0800")
+                }
+
+                // TODO are these options okay for neo?
+                val command = mutableListOf("64tass", "--case-sensitive", "--long-branch",
+                    "-Wall", // "-Werror", "-Wno-strict-bool"
+                    "--no-monitor"
+                )
+
+                if(options.warnSymbolShadowing)
+                    command.add("-Wshadow")
+                else
+                    command.add("-Wno-shadow")
+
+                if(options.asmQuiet)
+                    command.add("--quiet")
+
+                if(options.asmListfile)
+                    command.add("--list=$listFile")
+
+                val outFile = when (options.output) {
+                    OutputType.RAW -> {
+                        command.add("--nostart")
+                        println("\nCreating raw binary for target ${compTarget.name}.")
+                        binFile
+                    }
+                    else -> throw AssemblyError("invalid output type, need 'raw'")
+                }
+                command.addAll(listOf("--output", outFile.toString(), assemblyFile.toString()))
+                assemblerCommand = command
+            }
             else -> throw AssemblyError("invalid compilation target")
         }
 
         val proc = ProcessBuilder(assemblerCommand).inheritIO().start()
         val result = proc.waitFor()
-        if (result == 0 && compTarget.name!="atari") {
+        if (result == 0 && compTarget.name !in targetWithoutBreakpointsForEmulator) {
             removeGeneratedLabelsFromMonlist()
             generateBreakpointList()
         }
@@ -109,7 +148,7 @@ internal class AssemblyProgram(
     }
 
     private fun removeGeneratedLabelsFromMonlist() {
-        val pattern = Regex("""al (\w+) \S+prog8_label_.+?""")
+        val pattern = Regex("""al (\w+) \S+${PtLabel.GeneratedLabelPrefix}.+?""")
         val lines = viceMonListFile.toFile().readLines()
         viceMonListFile.toFile().outputStream().bufferedWriter().use {
             for (line in lines) {

codeGenCpu6502/src/prog8/codegen/cpu6502/BuiltinFunctionsAsmGen.kt

@@ -1,7 +1,5 @@
 package prog8.codegen.cpu6502
 
-import prog8.code.StMemVar
-import prog8.code.StStaticVariable
 import prog8.code.ast.*
 import prog8.code.core.*
 import prog8.codegen.cpu6502.assignment.*
@@ -71,114 +69,25 @@ internal class BuiltinFunctionsAsmGen(private val program: PtProgram,
             "prog8_lib_stringcompare" -> funcStringCompare(fcall, resultRegister)
             "prog8_lib_square_byte" -> funcSquare(fcall, DataType.UBYTE, resultRegister)
             "prog8_lib_square_word" -> funcSquare(fcall, DataType.UWORD, resultRegister)
-            "prog8_lib_arraycopy" -> funcArrayCopy(fcall)
             else -> throw AssemblyError("missing asmgen for builtin func ${fcall.name}")
         }
 
         return BuiltinFunctions.getValue(fcall.name).returnType
     }
 
-    private fun funcArrayCopy(fcall: PtBuiltinFunctionCall) {
-        val source = fcall.args[0] as PtIdentifier
-        val target = fcall.args[1] as PtIdentifier
-
-        val numElements = when(val sourceSymbol = asmgen.symbolTable.lookup(source.name)) {
-            is StStaticVariable -> sourceSymbol.length!!
-            is StMemVar -> sourceSymbol.length!!
-            else -> 0
-        }
-        val sourceAsm = asmgen.asmVariableName(source)
-        val targetAsm = asmgen.asmVariableName(target)
-
-        if(source.type in SplitWordArrayTypes && target.type in SplitWordArrayTypes) {
-            // split -> split words (copy lsb and msb arrays separately)
-            asmgen.out("""
-                lda  #<${sourceAsm}_lsb
-                ldy  #>${sourceAsm}_lsb
-                sta  P8ZP_SCRATCH_W1
-                sty  P8ZP_SCRATCH_W1+1
-                lda  #<${targetAsm}_lsb
-                ldy  #>${targetAsm}_lsb
-                sta  P8ZP_SCRATCH_W2
-                sty  P8ZP_SCRATCH_W2+1
-                ldy  #${numElements and 255}
-                jsr  prog8_lib.memcopy_small
-                lda  #<${sourceAsm}_msb
-                ldy  #>${sourceAsm}_msb
-                sta  P8ZP_SCRATCH_W1
-                sty  P8ZP_SCRATCH_W1+1
-                lda  #<${targetAsm}_msb
-                ldy  #>${targetAsm}_msb
-                sta  P8ZP_SCRATCH_W2
-                sty  P8ZP_SCRATCH_W2+1
-                ldy  #${numElements and 255}
-                jsr  prog8_lib.memcopy_small""")
-        }
-        else if(source.type in SplitWordArrayTypes) {
-            // split word array to normal word array (copy lsb and msb arrays separately)
-            require(target.type==DataType.ARRAY_UW || target.type==DataType.ARRAY_W)
-            asmgen.out("""
-                lda  #<${sourceAsm}_lsb
-                ldy  #>${sourceAsm}_lsb
-                sta  P8ZP_SCRATCH_W1
-                sty  P8ZP_SCRATCH_W1+1
-                lda  #<${sourceAsm}_msb
-                ldy  #>${sourceAsm}_msb
-                sta  P8ZP_SCRATCH_W2
-                sty  P8ZP_SCRATCH_W2+1
-                lda  #<${targetAsm}
-                ldy  #>${targetAsm}
-                ldx  #${numElements and 255}
-                jsr  prog8_lib.arraycopy_split_to_normal_words""")
-        }
-        else if(target.type in SplitWordArrayTypes) {
-            // normal word array to split array
-            require(source.type==DataType.ARRAY_UW || source.type==DataType.ARRAY_W)
-            asmgen.out("""
-                lda  #<${targetAsm}_lsb
-                ldy  #>${targetAsm}_lsb
-                sta  P8ZP_SCRATCH_W1
-                sty  P8ZP_SCRATCH_W1+1
-                lda  #<${targetAsm}_msb
-                ldy  #>${targetAsm}_msb
-                sta  P8ZP_SCRATCH_W2
-                sty  P8ZP_SCRATCH_W2+1
-                lda  #<${sourceAsm}
-                ldy  #>${sourceAsm}
-                ldx  #${numElements and 255}
-                jsr  prog8_lib.arraycopy_normal_to_split_words""")
-        }
-        else {
-            // normal array to array copy, various element types
-            val eltsize = asmgen.options.compTarget.memorySize(ArrayToElementTypes.getValue(source.type))
-            val numBytes = numElements * eltsize
-            asmgen.out("""
-                lda  #<${sourceAsm}
-                ldy  #>${sourceAsm}
-                sta  P8ZP_SCRATCH_W1
-                sty  P8ZP_SCRATCH_W1+1
-                lda  #<${targetAsm}
-                ldy  #>${targetAsm}
-                sta  P8ZP_SCRATCH_W2
-                sty  P8ZP_SCRATCH_W2+1
-                ldy  #${numBytes and 255}
-                jsr  prog8_lib.memcopy_small""")
-        }
-    }
-
     private fun funcSquare(fcall: PtBuiltinFunctionCall, resultType: DataType, resultRegister: RegisterOrPair?) {
         // square of word value is faster with dedicated routine, square of byte just use the regular multiplication routine.
         when (resultType) {
             DataType.UBYTE -> {
                 asmgen.assignExpressionToRegister(fcall.args[0], RegisterOrPair.A)
-                asmgen.out("  tay  |  jsr  math.multiply_bytes")
+                asmgen.out("  tay  |  jsr  prog8_math.multiply_bytes")
                 if(resultRegister!=null)  {
                     assignAsmGen.assignRegisterByte(AsmAssignTarget.fromRegisters(resultRegister, false, fcall.position, null, asmgen), CpuRegister.A, false, false)
                 }
             }
             DataType.UWORD -> {
                 asmgen.assignExpressionToRegister(fcall.args[0], RegisterOrPair.AY)
-                asmgen.out("  jsr  math.square")
+                asmgen.out("  jsr  prog8_math.square")
                 if(resultRegister!=null)  {
                     assignAsmGen.assignRegisterpairWord(AsmAssignTarget.fromRegisters(resultRegister, false, fcall.position, null, asmgen), RegisterOrPair.AY)
                 }
@@ -195,7 +104,7 @@ internal class BuiltinFunctionsAsmGen(private val program: PtProgram,
         assignAsmGen.assignExpressionToRegister(fcall.args[1], RegisterOrPair.Y, false)
         asmgen.restoreRegisterStack(CpuRegister.A ,false)
         // math.divmod_ub_asm: -- divide A by Y, result quotient in Y, remainder in A   (unsigned)
-        asmgen.out("  jsr  math.divmod_ub_asm")
+        asmgen.out("  jsr  prog8_math.divmod_ub_asm")
         val var2name = asmgen.asmVariableName(fcall.args[2] as PtIdentifier)
         val var3name = asmgen.asmVariableName(fcall.args[3] as PtIdentifier)
         val divisionTarget = AsmAssignTarget(TargetStorageKind.VARIABLE, asmgen, DataType.UBYTE, fcall.definingISub(), fcall.args[2].position, var2name)
@@ -209,7 +118,7 @@ internal class BuiltinFunctionsAsmGen(private val program: PtProgram,
         // math.divmod_uw_asm: -- divide two unsigned words (16 bit each) into 16 bit results
         //    input:  P8ZP_SCRATCH_W1 in ZP: 16 bit number, A/Y: 16 bit divisor
         //    output: P8ZP_SCRATCH_W2 in ZP: 16 bit remainder, A/Y: 16 bit division result
-        asmgen.out("  jsr  math.divmod_uw_asm")
+        asmgen.out("  jsr  prog8_math.divmod_uw_asm")
         val var2name = asmgen.asmVariableName(fcall.args[2] as PtIdentifier)
         val divisionTarget = AsmAssignTarget(TargetStorageKind.VARIABLE, asmgen, DataType.UBYTE, fcall.definingISub(), fcall.args[2].position, var2name)
         val remainderVar = asmgen.asmVariableName(fcall.args[3] as PtIdentifier)
@@ -299,15 +208,22 @@ internal class BuiltinFunctionsAsmGen(private val program: PtProgram,
     }
 
     private fun funcCallFar(fcall: PtBuiltinFunctionCall, resultRegister: RegisterOrPair?) {
-        if(asmgen.options.compTarget.name != "cx16")
-            throw AssemblyError("callfar only works on cx16 target at this time")
+        val targetName = asmgen.options.compTarget.name
+        if(targetName !in arrayOf("cx16", "c64", "c128"))
+            throw AssemblyError("callfar only works on cx16, c64 and c128 targets at this time")
 
+        val jsrfar = when(targetName) {
+            "cx16" -> "cx16.JSRFAR"
+            "c64" -> "c64.x16jsrfar"
+            "c128" -> "c128.x16jsrfar"
+            else -> TODO("jsrfar routine")
+        }
         val constBank = fcall.args[0].asConstInteger()
         val constAddress = fcall.args[1].asConstInteger()
         if(constBank!=null && constAddress!=null) {
             asmgen.assignExpressionToRegister(fcall.args[2], RegisterOrPair.AY)     // uword argument
             asmgen.out("""
-                jsr  cx16.JSRFAR
+                jsr  $jsrfar
                 .word  ${constAddress.toHex()}
                 .byte  $constBank""")
         } else {
@@ -317,7 +233,7 @@ internal class BuiltinFunctionsAsmGen(private val program: PtProgram,
             asmgen.out("  sta  (+)+0 |  sty  (+)+1")
             asmgen.assignExpressionToRegister(fcall.args[2], RegisterOrPair.AY)     // uword argument
             asmgen.out("""
-                jsr  cx16.JSRFAR
+                jsr  $jsrfar
 +               .word  0
 +               .byte  0""")
         }
@@ -329,8 +245,9 @@ internal class BuiltinFunctionsAsmGen(private val program: PtProgram,
     }
 
     private fun funcCallFar2(fcall: PtBuiltinFunctionCall, resultRegister: RegisterOrPair?) {
-        if(asmgen.options.compTarget.name != "cx16")
-            throw AssemblyError("callfar2 only works on cx16 target at this time")
+        val targetName = asmgen.options.compTarget.name
+        if(targetName !in arrayOf("cx16", "c64", "c128"))
+            throw AssemblyError("callfar2 only works on cx16, c64 and c128 targets at this time")
 
         fun assignArgs() {
             fun assign(value: PtExpression, register: Char) {
@@ -351,12 +268,18 @@ internal class BuiltinFunctionsAsmGen(private val program: PtProgram,
                 TODO("callfar2: support non-const argument values")
         }
 
+        val jsrfar = when(targetName) {
+            "cx16" -> "cx16.JSRFAR"
+            "c64" -> "c64.x16jsrfar"
+            "c128" -> "c128.x16jsrfar"
+            else -> TODO("jsrfar routine")
+        }
         val constBank = fcall.args[0].asConstInteger()
         val constAddress = fcall.args[1].asConstInteger()
         if(constBank!=null && constAddress!=null) {
             assignArgs()
             asmgen.out("""
-                jsr  cx16.JSRFAR
+                jsr  $jsrfar
                 .word  ${constAddress.toHex()}
                 .byte  $constBank""")
         } else {
@@ -366,7 +289,7 @@ internal class BuiltinFunctionsAsmGen(private val program: PtProgram,
             asmgen.out("  sta  (+)+0 |  sty  (+)+1")
             assignArgs()
             asmgen.out("""
-                jsr  cx16.JSRFAR
+                jsr  $jsrfar
 +               .word  0
 +               .byte  0""")
         }
@@ -1357,22 +1280,6 @@ internal class BuiltinFunctionsAsmGen(private val program: PtProgram,
         }
     }
 
-    private fun outputAddressAndLengthOfArray(arg: PtIdentifier) {
-        // address goes in P8ZP_SCRATCH_W1,  number of elements in A
-        val numElements = when(val symbol = asmgen.symbolTable.lookup(arg.name)) {
-            is StStaticVariable -> symbol.length!!
-            is StMemVar -> symbol.length!!
-            else -> 0
-        }
-        val identifierName = asmgen.asmVariableName(arg)
-        asmgen.out("""
-                    lda  #<$identifierName
-                    ldy  #>$identifierName
-                    sta  P8ZP_SCRATCH_W1
-                    sty  P8ZP_SCRATCH_W1+1
-                    lda  #${numElements and 255}""")
-    }
-
     private fun translateArguments(call: PtBuiltinFunctionCall, scope: IPtSubroutine?) {
         val signature = BuiltinFunctions.getValue(call.name)
         val callConv = signature.callConvention(call.args.map { it.type})

codeGenCpu6502/src/prog8/codegen/cpu6502/FunctionCallAsmGen.kt

@@ -15,6 +15,7 @@ internal class FunctionCallAsmGen(private val program: PtProgram, private val as
         // just ignore any result values from the function call.
     }
 
+    // TODO tweak subroutine call convention to also make use of X register to pass a third byte?
     internal fun optimizeIntArgsViaRegisters(sub: PtSub) =
         (sub.parameters.size==1 && sub.parameters[0].type in IntegerDatatypesWithBoolean)
                 || (sub.parameters.size==2 && sub.parameters[0].type in ByteDatatypesWithBoolean && sub.parameters[1].type in ByteDatatypesWithBoolean)
@@ -40,7 +41,84 @@ internal class FunctionCallAsmGen(private val program: PtProgram, private val as
                 sub.children.forEach { asmgen.translate(it as PtInlineAssembly) }
                 asmgen.out("  \t; inlined routine end: ${sub.name}")
             } else {
-                asmgen.out("  jsr  $subAsmName")
+                val bank = sub.address?.constbank?.toString()
+                if(bank==null) {
+                    val varbank = if(sub.address?.varbank==null) null else asmgen.asmVariableName(sub.address!!.varbank!!)
+                    if(varbank!=null) {
+                        when(asmgen.options.compTarget.name) {
+                            "cx16" -> {
+                                // JSRFAR can jump to a banked RAM address as well!
+                                asmgen.out("""
+                                php
+                                pha
+                                lda  $varbank
+                                sta  +
+                                pla
+                                plp
+                                jsr  cx16.JSRFAR
+                                .word  $subAsmName    ; ${sub.address!!.address.toHex()}
++                               .byte  0    ; modified"""
+                                )
+                            }
+                            "c64" -> {
+                                asmgen.out("""
+                                php
+                                pha
+                                lda  $varbank
+                                sta  +
+                                pla
+                                plp
+                                jsr  c64.x16jsrfar
+                                .word  $subAsmName    ; ${sub.address!!.address.toHex()}
++                               .byte  0    ; modified"""
+                                )
+                            }
+                            "c128" -> {
+                                asmgen.out("""
+                                php
+                                pha
+                                lda  $varbank
+                                sta  +
+                                pla
+                                plp
+                                jsr  c128.x16jsrfar
+                                .word  $subAsmName    ; ${sub.address!!.address.toHex()}
++                               .byte  0    ; modified"""
+                                )
+                            }
+                            else -> throw AssemblyError("callfar is not supported on the selected compilation target")
+                        }
+                    } else {
+                        asmgen.out("  jsr  $subAsmName")
+                    }
+                }
+                else {
+                    when(asmgen.options.compTarget.name) {
+                        "cx16" -> {
+                            // JSRFAR can jump to a banked RAM address as well!
+                            asmgen.out("""
+                                jsr cx16.JSRFAR
+                                .word  $subAsmName    ; ${sub.address!!.address.toHex()}
+                                .byte  $bank"""
+                            )
+                        }
+                        "c64" -> {
+                            asmgen.out("""
+                                jsr  c64.x16jsrfar
+                                .word  $subAsmName    ; ${sub.address!!.address.toHex()}
+                                .byte  $bank"""
+                            )
+                        }
+                        "c128" -> {
+                            asmgen.out("""
+                                jsr  c128.x16jsrfar
+                                .word  $subAsmName    ; ${sub.address!!.address.toHex()}
+                                .byte  $bank"""
+                            )
+                        }
+                        else -> throw AssemblyError("callfar is not supported on the selected compilation target")
+                    }
+                }
             }
         }
         else if(sub is PtSub) {

codeGenCpu6502/src/prog8/codegen/cpu6502/IfElseAsmGen.kt

@@ -1,6 +1,6 @@
 package prog8.codegen.cpu6502
 
-import prog8.code.StRomSub
+import prog8.code.StExtSub
 import prog8.code.SymbolTable
 import prog8.code.ast.*
 import prog8.code.core.*
@@ -17,7 +17,7 @@ internal class IfElseAsmGen(private val program: PtProgram,
 
     fun translate(stmt: PtIfElse) {
         require(stmt.condition.type== DataType.BOOL)
-        checkNotRomsubReturnsStatusReg(stmt.condition)
+        checkNotExtsubReturnsStatusReg(stmt.condition)
 
         val jumpAfterIf = stmt.ifScope.children.singleOrNull() as? PtJump
 
@@ -43,23 +43,105 @@ internal class IfElseAsmGen(private val program: PtProgram,
 
         val prefixCond = stmt.condition as? PtPrefix
         if(prefixCond?.operator=="not") {
-            checkNotRomsubReturnsStatusReg(prefixCond.value)
-            assignConditionValueToRegisterAndTest(prefixCond.value)
-            return if(jumpAfterIf!=null)
-                translateJumpElseBodies("beq", "bne", jumpAfterIf, stmt.elseScope)
-            else
-                translateIfElseBodies("bne", stmt)
+            if(stmt.hasElse())
+                throw AssemblyError("not prefix in ifelse should have been replaced by swapped if-else blocks")
+            else {
+                checkNotExtsubReturnsStatusReg(prefixCond.value)
+                assignConditionValueToRegisterAndTest(prefixCond.value)
+                return if (jumpAfterIf != null)
+                    translateJumpElseBodies("beq", "bne", jumpAfterIf, stmt.elseScope)
+                else
+                    translateIfElseBodies("bne", stmt)
+            }
         }
 
         throw AssemblyError("weird non-boolean condition node type ${stmt.condition} at ${stmt.condition.position}")
     }
 
-    private fun checkNotRomsubReturnsStatusReg(condition: PtExpression) {
+    internal fun assignIfExpression(target: AsmAssignTarget, expr: PtIfExpression) {
+        // this is NOT for the if-else STATEMENT, but this is code for the IF-EXPRESSION.
+        require(target.datatype==expr.type)
+        val falseLabel = asmgen.makeLabel("ifexpr_false")
+        val endLabel = asmgen.makeLabel("ifexpr_end")
+        evalConditonAndBranchWhenFalse(expr.condition, falseLabel)
+        when(expr.type) {
+            in ByteDatatypesWithBoolean -> {
+                asmgen.assignExpressionToRegister(expr.truevalue, RegisterOrPair.A, false)
+                asmgen.jmp(endLabel)
+                asmgen.out(falseLabel)
+                asmgen.assignExpressionToRegister(expr.falsevalue, RegisterOrPair.A, false)
+                asmgen.out(endLabel)
+                assignmentAsmGen.assignRegisterByte(target, CpuRegister.A, false, false)
+            }
+            in WordDatatypes -> {
+                asmgen.assignExpressionToRegister(expr.truevalue, RegisterOrPair.AY, false)
+                asmgen.jmp(endLabel)
+                asmgen.out(falseLabel)
+                asmgen.assignExpressionToRegister(expr.falsevalue, RegisterOrPair.AY, false)
+                asmgen.out(endLabel)
+                assignmentAsmGen.assignRegisterpairWord(target, RegisterOrPair.AY)
+            }
+            DataType.FLOAT -> {
+                asmgen.assignExpressionToRegister(expr.truevalue, RegisterOrPair.FAC1, true)
+                asmgen.jmp(endLabel)
+                asmgen.out(falseLabel)
+                asmgen.assignExpressionToRegister(expr.falsevalue, RegisterOrPair.FAC1, true)
+                asmgen.out(endLabel)
+                asmgen.assignRegister(RegisterOrPair.FAC1, target)
+            }
+            else -> throw AssemblyError("weird dt")
+        }
+    }
+
+    private fun evalConditonAndBranchWhenFalse(condition: PtExpression, falseLabel: String) {
+        if (condition is PtBinaryExpression) {
+            return when(condition.right.type) {
+                in ByteDatatypesWithBoolean -> translateIfByteConditionBranch(condition, falseLabel)
+                in WordDatatypes -> translateIfWordConditionBranch(condition, falseLabel)
+                DataType.FLOAT -> translateFloatConditionBranch(condition, falseLabel)
+                else -> throw AssemblyError("weird dt")
+            }
+        }
+        else if(condition is PtPrefix && condition.operator=="not") {
+            throw AssemblyError("not prefix in ifexpression should have been replaced by swapped values")
+        } else {
+            // 'simple' condition, check if it is a byte bittest
+            val bittest = condition as? PtBuiltinFunctionCall
+            if(bittest!=null && bittest.name.startsWith("prog8_ifelse_bittest_")) {
+                val variable = bittest.args[0] as PtIdentifier
+                val bitnumber = (bittest.args[1] as PtNumber).number.toInt()
+                val testForBitSet = bittest.name.endsWith("_set")
+                when (bitnumber) {
+                    7 -> {
+                        // test via bit + N flag
+                        asmgen.out("  bit  ${variable.name}")
+                        if(testForBitSet) asmgen.out("  bpl  $falseLabel")
+                        else asmgen.out("  bmi  $falseLabel")
+                        return
+                    }
+                    6 -> {
+                        // test via bit + V flag
+                        asmgen.out("  bit  ${variable.name}")
+                        if(testForBitSet) asmgen.out("  bvc  $falseLabel")
+                        else asmgen.out("  bvs  $falseLabel")
+                        return
+                    }
+                    else -> throw AssemblyError("prog8_ifelse_bittest can only work on bits 7 and 6")
+                }
+            }
+
+            // the condition is "simple" enough to just assign its 0/1 value to a register and branch on that
+            assignConditionValueToRegisterAndTest(condition)
+            asmgen.out("  beq  $falseLabel")
+        }
+    }
+
+    private fun checkNotExtsubReturnsStatusReg(condition: PtExpression) {
         val fcall = condition as? PtFunctionCall
         if(fcall!=null && fcall.type==DataType.BOOL) {
-            val romsub = st.lookup(fcall.name) as? StRomSub
-            if(romsub!=null && romsub.returns.any { it.register.statusflag!=null }) {
-                throw AssemblyError("if romsub() that returns a status register boolean should have been changed into a Conditional branch such as if_cc")
+            val extsub = st.lookup(fcall.name) as? StExtSub
+            if(extsub!=null && extsub.returns.any { it.register.statusflag!=null }) {
+                throw AssemblyError("if extsub() returning a status register boolean should have been changed into a Conditional branch such as if_cc")
             }
         }
     }
@@ -73,6 +155,7 @@ internal class IfElseAsmGen(private val program: PtProgram,
             is PtIrRegister,
             is PtArrayIndexer,
             is PtPrefix,
+            is PtIfExpression,
             is PtBinaryExpression -> { /* no cmp necessary the lda has been done just prior */ }
             is PtTypeCast -> {
                 if(condition.value.type !in ByteDatatypes && condition.value.type !in WordDatatypes)
@@ -276,6 +359,230 @@ internal class IfElseAsmGen(private val program: PtProgram,
         }
     }
 
+    private fun translateIfByteConditionBranch(condition: PtBinaryExpression, falseLabel: String) {
+        val signed = condition.left.type in SignedDatatypes
+        val constValue = condition.right.asConstInteger()
+        if(constValue==0) {
+            return translateIfCompareWithZeroByteBranch(condition, signed, falseLabel)
+        }
+
+        when(condition.operator) {
+            "==" -> {
+                // if X==value
+                asmgen.assignExpressionToRegister(condition.left, RegisterOrPair.A, signed)
+                cmpAwithByteValue(condition.right, false)
+                asmgen.out("  bne  $falseLabel")
+            }
+            "!=" -> {
+                // if X!=value
+                asmgen.assignExpressionToRegister(condition.left, RegisterOrPair.A, signed)
+                cmpAwithByteValue(condition.right, false)
+                asmgen.out("  beq  $falseLabel")
+            }
+            in LogicalOperators -> {
+                val regAtarget = AsmAssignTarget(TargetStorageKind.REGISTER, asmgen, DataType.BOOL, condition.definingISub(), condition.position, register=RegisterOrPair.A)
+                if (assignmentAsmGen.optimizedLogicalExpr(condition, regAtarget)) {
+                    asmgen.out("  beq  $falseLabel")
+                } else {
+                    errors.warn("SLOW FALLBACK FOR 'IFEXPR' CODEGEN - ask for support", condition.position)      //  should not occur ;-)
+                    assignConditionValueToRegisterAndTest(condition)
+                    asmgen.out("  beq  $falseLabel")
+                }
+            }
+            else -> {
+                // TODO don't store condition as expression result but just use the flags, like a normal PtIfElse translation does
+                // TODO: special cases for <, <=, >, >= above.
+                assignConditionValueToRegisterAndTest(condition)
+                asmgen.out("  beq  $falseLabel")
+            }
+        }
+    }
+
+    private fun translateIfWordConditionBranch(condition: PtBinaryExpression, falseLabel: String) {
+        // TODO can we reuse this whole thing from IfElse ?
+        val constValue = condition.right.asConstInteger()
+        if(constValue!=null) {
+            if (constValue == 0) {
+                when (condition.operator) {
+                    "==" -> return translateWordExprIsZero(condition.left, falseLabel)
+                    "!=" -> return translateWordExprIsNotZero(condition.left, falseLabel)
+                }
+            }
+            if (constValue != 0) {
+                when (condition.operator) {
+                    "==" -> return translateWordExprEqualsNumber(condition.left, constValue, falseLabel)
+                    "!=" -> return translateWordExprNotEqualsNumber(condition.left, constValue, falseLabel)
+                }
+            }
+        }
+        val variable = condition.right as? PtIdentifier
+        if(variable!=null) {
+            when (condition.operator) {
+                "==" -> return translateWordExprEqualsVariable(condition.left, variable, falseLabel)
+                "!=" -> return translateWordExprNotEqualsVariable(condition.left, variable, falseLabel)
+            }
+        }
+
+        // TODO don't store condition as expression result but just use the flags, like a normal PtIfElse translation does
+        assignConditionValueToRegisterAndTest(condition)
+        asmgen.out("  beq  $falseLabel")
+    }
+
+    private fun translateWordExprNotEqualsVariable(expr: PtExpression, variable: PtIdentifier, falseLabel: String) {
+        // if w!=variable
+        // TODO reuse code from ifElse?
+        val varRight = asmgen.asmVariableName(variable)
+        if(expr is PtIdentifier) {
+            val varLeft = asmgen.asmVariableName(expr)
+            asmgen.out("""
+                lda  $varLeft
+                cmp  $varRight
+                bne  +
+                lda  $varLeft+1
+                cmp  $varRight+1
+                beq  $falseLabel
++""")
+        } else {
+            asmgen.assignExpressionToRegister(expr, RegisterOrPair.AY, false)
+            asmgen.out("""
+                cmp  $varRight
+                bne  +
+                cpy  $varRight+1
+                beq  $falseLabel
++""")
+        }
+    }
+
+    private fun translateWordExprEqualsVariable(expr: PtExpression, variable: PtIdentifier, falseLabel: String) {
+        // if w==variable
+        // TODO reuse code from ifElse?
+        val varRight = asmgen.asmVariableName(variable)
+        if(expr is PtIdentifier) {
+            val varLeft = asmgen.asmVariableName(expr)
+            asmgen.out("""
+                lda  $varLeft
+                cmp  $varRight
+                bne  $falseLabel
+                lda  $varLeft+1
+                cmp  $varRight+1
+                bne  $falseLabel""")
+        } else {
+            asmgen.assignExpressionToRegister(expr, RegisterOrPair.AY, false)
+            asmgen.out("""
+                cmp  $varRight
+                bne  $falseLabel
+                cpy  $varRight+1
+                bne  $falseLabel""")
+        }
+    }
+
+    private fun translateWordExprNotEqualsNumber(expr: PtExpression, number: Int, falseLabel: String) {
+        // if w!=number
+        // TODO reuse code from ifElse?
+        if(expr is PtIdentifier) {
+            val varname = asmgen.asmVariableName(expr)
+            asmgen.out("""
+                lda  $varname
+                cmp  #<$number
+                bne  +
+                lda  $varname+1
+                cmp  #>$number
+                beq  $falseLabel
++""")
+        } else {
+            asmgen.assignExpressionToRegister(expr, RegisterOrPair.AY, false)
+            asmgen.out("""
+                cmp  #<$number
+                bne  +
+                cpy  #>$number
+                beq  $falseLabel
++""")
+        }
+    }
+
+    private fun translateWordExprEqualsNumber(expr: PtExpression, number: Int, falseLabel: String) {
+        // if w==number
+        // TODO reuse code from ifElse?
+        if(expr is PtIdentifier) {
+            val varname = asmgen.asmVariableName(expr)
+            asmgen.out("""
+                lda  $varname
+                cmp  #<$number
+                bne  $falseLabel
+                lda  $varname+1
+                cmp  #>$number
+                bne  $falseLabel""")
+        } else {
+            asmgen.assignExpressionToRegister(expr, RegisterOrPair.AY, false)
+            asmgen.out(                """
+                cmp  #<$number
+                bne  $falseLabel
+                cpy  #>$number
+                bne  $falseLabel""")
+        }
+    }
+
+    private fun translateWordExprIsNotZero(expr: PtExpression, falseLabel: String) {
+        // if w!=0
+        // TODO reuse code from ifElse?
+        if(expr is PtIdentifier) {
+            val varname = asmgen.asmVariableName(expr)
+            asmgen.out("""
+                lda  $varname
+                ora  $varname+1
+                beq  $falseLabel""")
+        } else {
+            asmgen.assignExpressionToRegister(expr, RegisterOrPair.AY, false)
+            asmgen.out("  sty  P8ZP_SCRATCH_REG |  ora  P8ZP_SCRATCH_REG |  beq  $falseLabel")
+        }
+    }
+
+    private fun translateWordExprIsZero(expr: PtExpression, falseLabel: String) {
+        // if w==0
+        // TODO reuse code from ifElse?
+        if(expr is PtIdentifier) {
+            val varname = asmgen.asmVariableName(expr)
+            asmgen.out("""
+                lda  $varname
+                ora  $varname+1
+                bne  $falseLabel""")
+        } else {
+            asmgen.assignExpressionToRegister(expr, RegisterOrPair.AY, false)
+            asmgen.out("  sty  P8ZP_SCRATCH_REG |  ora  P8ZP_SCRATCH_REG |  bne  $falseLabel")
+        }
+    }
+
+    private fun translateIfCompareWithZeroByteBranch(condition: PtBinaryExpression, signed: Boolean, falseLabel: String) {
+        // optimized code for byte comparisons with 0
+        assignConditionValueToRegisterAndTest(condition.left)
+        when (condition.operator) {
+            "==" -> asmgen.out("  bne  $falseLabel")
+            "!=" -> asmgen.out("  beq  $falseLabel")
+            ">" -> {
+                if(signed) asmgen.out("  bmi  $falseLabel |  beq  $falseLabel")
+                else asmgen.out("  beq  $falseLabel")
+            }
+            ">=" -> {
+                if(signed) asmgen.out("  bmi  $falseLabel")
+                else { /* always true for unsigned */ }
+            }
+            "<" -> {
+                if(signed) asmgen.out("  bpl  $falseLabel")
+                else asmgen.jmp(falseLabel)
+            }
+            "<=" -> {
+                if(signed) {
+                    // inverted '>'
+                    asmgen.out("""
+                        beq  +
+                        bpl  $falseLabel
++""")
+                } else asmgen.out("  bne  $falseLabel")
+            }
+            else -> throw AssemblyError("expected comparison operator")
+        }
+    }
+
     private fun translateIfCompareWithZeroByte(stmt: PtIfElse, signed: Boolean, jumpAfterIf: PtJump?) {
         // optimized code for byte comparisons with 0
         val condition = stmt.condition as PtBinaryExpression
@@ -1741,6 +2048,51 @@ _jump                       jmp  ($asmLabel)
         }
     }
 
+    private fun translateFloatConditionBranch(condition: PtBinaryExpression, elseLabel: String) {
+        val constValue = (condition.right as? PtNumber)?.number
+        if(constValue==0.0) {
+            if (condition.operator == "==") {
+                // if FL==0.0
+                asmgen.assignExpressionToRegister(condition.left, RegisterOrPair.FAC1, true)
+                asmgen.out("  jsr  floats.SIGN |  cmp  #0 |  bne  $elseLabel")
+                return
+            } else if(condition.operator=="!=") {
+                // if FL!=0.0
+                asmgen.assignExpressionToRegister(condition.left, RegisterOrPair.FAC1, true)
+                asmgen.out("  jsr  floats.SIGN |  cmp  #0 |  beq  $elseLabel")
+                return
+            }
+        }
+
+        when(condition.operator) {
+            "==" -> {
+                translateFloatsEqualsConditionIntoA(condition.left, condition.right)
+                asmgen.out("  beq  $elseLabel")
+            }
+            "!=" -> {
+                translateFloatsEqualsConditionIntoA(condition.left, condition.right)
+                asmgen.out("  bne  $elseLabel")
+            }
+            "<" -> {
+                translateFloatsLessConditionIntoA(condition.left, condition.right, false)
+                asmgen.out("  beq  $elseLabel")
+            }
+            "<=" -> {
+                translateFloatsLessConditionIntoA(condition.left, condition.right, true)
+                asmgen.out("  beq  $elseLabel")
+            }
+            ">" -> {
+                translateFloatsLessConditionIntoA(condition.left, condition.right, true)
+                asmgen.out("  bne  $elseLabel")
+            }
+            ">=" -> {
+                translateFloatsLessConditionIntoA(condition.left, condition.right, false)
+                asmgen.out("  bne  $elseLabel")
+            }
+            else -> throw AssemblyError("expected comparison operator")
+        }
+    }
+
     private fun translateFloatsEqualsConditionIntoA(left: PtExpression, right: PtExpression) {
         fun equalf(leftName: String, rightName: String) {
             asmgen.out("""

codeGenCpu6502/src/prog8/codegen/cpu6502/ProgramAndVarsGen.kt

@@ -32,10 +32,6 @@ internal class ProgramAndVarsGen(
 
     internal fun generate() {
         header()
-        val allBlocks = program.allBlocks()
-
-        if(allBlocks.first().name != "p8b_main" && allBlocks.first().name != "main")
-            throw AssemblyError("first block should be 'main' or 'p8b_main'")
 
         if(errors.noErrors())  {
             program.allBlocks().forEach { block2asm(it) }
@@ -75,6 +71,12 @@ internal class ProgramAndVarsGen(
         asmgen.out("P8ZP_SCRATCH_REG = ${zp.SCRATCH_REG}")
         asmgen.out("P8ZP_SCRATCH_W1 = ${zp.SCRATCH_W1}    ; word")
         asmgen.out("P8ZP_SCRATCH_W2 = ${zp.SCRATCH_W2}    ; word")
+        if(compTarget.name=="c64") {
+            if(options.floats)
+                asmgen.out("PROG8_C64_BANK_CONFIG=31  ; basic+IO+kernal")
+            else
+                asmgen.out("PROG8_C64_BANK_CONFIG=30  ; IO+kernal, no basic")
+        }
         asmgen.out(".weak")   // hack to allow user to override the following two with command line redefinition (however, just use '-esa' command line option instead!)
         asmgen.out(".endweak")
 
@@ -89,12 +91,13 @@ internal class ProgramAndVarsGen(
             OutputType.RAW -> {
                 asmgen.out("; ---- raw assembler program ----")
                 asmgen.out("* = ${options.loadAddress.toHex()}")
+                asmgen.out("prog8_program_start\t; start of program label")
                 asmgen.out("  cld")
                 asmgen.out("  tsx  ; save stackpointer for sys.exit()")
                 asmgen.out("  stx  prog8_lib.orig_stackpointer")
                 if(!options.noSysInit)
-                    asmgen.out("  jsr  sys.init_system")
-                asmgen.out("  jsr  sys.init_system_phase2")
+                    asmgen.out("  jsr  p8_sys_startup.init_system")
+                asmgen.out("  jsr  p8_sys_startup.init_system_phase2")
             }
             OutputType.PRG -> {
                 when(options.launcher) {
@@ -104,40 +107,43 @@ internal class ProgramAndVarsGen(
                         }
                         asmgen.out("; ---- basic program with sys call ----")
                         asmgen.out("* = ${options.loadAddress.toHex()}")
+                        asmgen.out("prog8_program_start\t; start of program label")
                         val year = LocalDate.now().year
                         asmgen.out("  .word  (+), $year")
                         asmgen.out("  .null  $9e, format(' %d ', prog8_entrypoint), $3a, $8f, ' prog8'")
                         asmgen.out("+\t.word  0")
-                        asmgen.out("prog8_entrypoint\t; assembly code starts here")
+                        asmgen.out("prog8_entrypoint")
                         asmgen.out("  cld")
                         asmgen.out("  tsx  ; save stackpointer for sys.exit()")
                         asmgen.out("  stx  prog8_lib.orig_stackpointer")
                         if(!options.noSysInit)
-                            asmgen.out("  jsr  sys.init_system")
-                        asmgen.out("  jsr  sys.init_system_phase2")
+                            asmgen.out("  jsr  p8_sys_startup.init_system")
+                        asmgen.out("  jsr  p8_sys_startup.init_system_phase2")
                     }
                     CbmPrgLauncherType.NONE -> {
                         // this is the same as RAW
                         asmgen.out("; ---- program without basic sys call ----")
                         asmgen.out("* = ${options.loadAddress.toHex()}")
+                        asmgen.out("prog8_program_start\t; start of program label")
                         asmgen.out("  cld")
                         asmgen.out("  tsx  ; save stackpointer for sys.exit()")
                         asmgen.out("  stx  prog8_lib.orig_stackpointer")
                         if(!options.noSysInit)
-                            asmgen.out("  jsr  sys.init_system")
-                        asmgen.out("  jsr  sys.init_system_phase2")
+                            asmgen.out("  jsr  p8_sys_startup.init_system")
+                        asmgen.out("  jsr  p8_sys_startup.init_system_phase2")
                     }
                 }
             }
             OutputType.XEX -> {
                 asmgen.out("; ---- atari xex program ----")
                 asmgen.out("* = ${options.loadAddress.toHex()}")
+                asmgen.out("prog8_program_start\t; start of program label")
                 asmgen.out("  cld")
                 asmgen.out("  tsx  ; save stackpointer for sys.exit()")
                 asmgen.out("  stx  prog8_lib.orig_stackpointer")
                 if(!options.noSysInit)
-                    asmgen.out("  jsr  sys.init_system")
-                asmgen.out("  jsr  sys.init_system_phase2")
+                    asmgen.out("  jsr  p8_sys_startup.init_system")
+                asmgen.out("  jsr  p8_sys_startup.init_system_phase2")
             }
         }
 
@@ -155,19 +161,19 @@ internal class ProgramAndVarsGen(
                 if(options.floats)
                     asmgen.out("  lda  #4 |  sta  $01")    // to use floats, make sure Basic rom is banked in
                 asmgen.out("  jsr  p8b_main.p8s_start")
-                asmgen.out("  jmp  sys.cleanup_at_exit")
+                asmgen.out("  jmp  p8_sys_startup.cleanup_at_exit")
             }
             "c64" -> {
-                asmgen.out("  jsr  p8b_main.p8s_start |  lda  #31 |  sta  $01")
-                asmgen.out("  jmp  sys.cleanup_at_exit")
+                asmgen.out("  jsr  p8b_main.p8s_start")
+                asmgen.out("  jmp  p8_sys_startup.cleanup_at_exit")
             }
             "c128" -> {
-                asmgen.out("  jsr  p8b_main.p8s_start |  lda  #0 |  sta ${"$"}ff00")
-                asmgen.out("  jmp  sys.cleanup_at_exit")
+                asmgen.out("  jsr  p8b_main.p8s_start")
+                asmgen.out("  jmp  p8_sys_startup.cleanup_at_exit")
             }
             else -> {
                 asmgen.out("  jsr  p8b_main.p8s_start")
-                asmgen.out("  jmp  sys.cleanup_at_exit")
+                asmgen.out("  jmp  p8_sys_startup.cleanup_at_exit")
             }
         }
     }
@@ -284,6 +290,8 @@ internal class ProgramAndVarsGen(
                 asmgen.out("  .cerror * > ${relocatedBssEnd.toHex()}, \"too many data for slabs_BSS section\"")
             }
         }
+        asmgen.out("  ; memtop check")
+        asmgen.out("  .cerror * >= ${options.memtopAddress.toHex()}, \"Program too long by \", * - ${(options.memtopAddress-1u).toHex()}, \" bytes, memtop=${options.memtopAddress.toHex()}\"")
     }
 
     private fun block2asm(block: PtBlock) {
@@ -291,12 +299,6 @@ internal class ProgramAndVarsGen(
         asmgen.out("; ---- block: '${block.name}' ----")
         if(block.options.address!=null)
             asmgen.out("* = ${block.options.address!!.toHex()}")
-        else {
-            if(block.options.alignment==PtBlock.BlockAlignment.WORD)
-                asmgen.out("\t.align 2")
-            else if(block.options.alignment==PtBlock.BlockAlignment.PAGE)
-                asmgen.out("\t.align $100")
-        }
 
         asmgen.out("${block.name}\t" + (if(block.options.forceOutput) ".block" else ".proc"))
         asmgen.outputSourceLine(block)
@@ -471,6 +473,7 @@ internal class ProgramAndVarsGen(
 
     private fun entrypointInitialization() {
         // zero out the BSS area first, before setting the variable init values
+        // this is mainly to make sure the arrays are all zero'd out at program startup
         asmgen.out("  jsr  prog8_lib.program_startup_clear_bss")
 
         // initialize block-level (global) variables at program start
@@ -516,12 +519,12 @@ internal class ProgramAndVarsGen(
 
         stringVarsWithInitInZp.forEach {
             val varname = asmgen.asmVariableName(it.name)+"_init_value"
-            outputStringvar(varname, it.value.second, it.value.first)
+            outputStringvar(varname, 0, it.value.second, it.value.first)
         }
 
         arrayVarsWithInitInZp.forEach {
             val varname = asmgen.asmVariableName(it.name)+"_init_value"
-            arrayVariable2asm(varname, it.alloc.dt, it.value, null)
+            arrayVariable2asm(varname, it.alloc.dt, 0, it.value, null)
         }
 
         asmgen.out("+")
@@ -545,8 +548,8 @@ internal class ProgramAndVarsGen(
         for (variable in vars) {
             val scopedName = variable.key
             val svar = symboltable.lookup(scopedName) as? StStaticVariable
-            if(svar?.onetimeInitializationStringValue!=null)
-                result.add(ZpStringWithInitial(scopedName, variable.value, svar.onetimeInitializationStringValue!!))
+            if(svar?.initializationStringValue!=null)
+                result.add(ZpStringWithInitial(scopedName, variable.value, svar.initializationStringValue!!))
         }
         return result
     }
@@ -557,8 +560,8 @@ internal class ProgramAndVarsGen(
         for (variable in vars) {
             val scopedName = variable.key
             val svar = symboltable.lookup(scopedName) as? StStaticVariable
-            if(svar?.onetimeInitializationArrayValue!=null)
-                result.add(ZpArrayWithInitial(scopedName, variable.value, svar.onetimeInitializationArrayValue!!))
+            if(svar?.initializationArrayValue!=null)
+                result.add(ZpArrayWithInitial(scopedName, variable.value, svar.initializationArrayValue!!))
         }
         return result
     }
@@ -586,7 +589,11 @@ internal class ProgramAndVarsGen(
         if(varsNoInit.isNotEmpty()) {
             asmgen.out("; non-zeropage variables")
             asmgen.out("  .section BSS")
-            varsNoInit.sortedWith(compareBy<StStaticVariable> { it.name }.thenBy { it.dt }).forEach {
+            val (notAligned, aligned) = varsNoInit.partition { it.align==0 }
+            notAligned.sortedWith(compareBy<StStaticVariable> { it.name }.thenBy { it.dt }).forEach {
+                uninitializedVariable2asm(it)
+            }
+            aligned.sortedWith(compareBy<StStaticVariable> { it.align }.thenBy { it.name }.thenBy { it.dt }).forEach {
                 uninitializedVariable2asm(it)
             }
             asmgen.out("  .send BSS")
@@ -595,14 +602,29 @@ internal class ProgramAndVarsGen(
         if(varsWithInit.isNotEmpty()) {
             asmgen.out("; non-zeropage variables with init value")
             val (stringvars, othervars) = varsWithInit.sortedBy { it.name }.partition { it.dt == DataType.STR }
-            stringvars.forEach {
+            val (notAlignedStrings, alignedStrings) = stringvars.partition { it.align==0 }
+            val (notAlignedOther, alignedOther) = othervars.partition { it.align==0 }
+            notAlignedStrings.forEach {
                 outputStringvar(
                     it.name,
-                    it.onetimeInitializationStringValue!!.second,
-                    it.onetimeInitializationStringValue!!.first
+                    it.align,
+                    it.initializationStringValue!!.second,
+                    it.initializationStringValue!!.first
                 )
             }
-            othervars.sortedBy { it.type }.forEach {
+            alignedStrings.sortedBy { it.align }.forEach {
+                outputStringvar(
+                    it.name,
+                    it.align,
+                    it.initializationStringValue!!.second,
+                    it.initializationStringValue!!.first
+                )
+            }
+
+            notAlignedOther.sortedBy { it.type }.forEach {
+                staticVariable2asm(it)
+            }
+            alignedOther.sortedBy { it.align }.sortedBy { it.type }.forEach {
                 staticVariable2asm(it)
             }
         }
@@ -616,11 +638,13 @@ internal class ProgramAndVarsGen(
             DataType.WORD -> asmgen.out("${variable.name}\t.sint  ?")
             DataType.FLOAT -> asmgen.out("${variable.name}\t.fill  ${compTarget.machine.FLOAT_MEM_SIZE}")
             in SplitWordArrayTypes -> {
+                alignVar(variable.align)
                 val numbytesPerHalf = compTarget.memorySize(variable.dt, variable.length!!) / 2
                 asmgen.out("${variable.name}_lsb\t.fill  $numbytesPerHalf")
                 asmgen.out("${variable.name}_msb\t.fill  $numbytesPerHalf")
             }
             in ArrayDatatypes -> {
+                alignVar(variable.align)
                 val numbytes = compTarget.memorySize(variable.dt, variable.length!!)
                 asmgen.out("${variable.name}\t.fill  $numbytes")
             }
@@ -630,13 +654,18 @@ internal class ProgramAndVarsGen(
         }
     }
 
+    private fun alignVar(align: Int) {
+        if(align > 1)
+            asmgen.out("  .align  ${align.toHex()}")
+    }
+
     private fun staticVariable2asm(variable: StStaticVariable) {
         val initialValue: Number =
-            if(variable.onetimeInitializationNumericValue!=null) {
+            if(variable.initializationNumericValue!=null) {
                 if(variable.dt== DataType.FLOAT)
-                    variable.onetimeInitializationNumericValue!!
+                    variable.initializationNumericValue!!
                 else
-                    variable.onetimeInitializationNumericValue!!.toInt()
+                    variable.initializationNumericValue!!.toInt()
             } else 0
 
         when (variable.dt) {
@@ -655,14 +684,17 @@ internal class ProgramAndVarsGen(
             DataType.STR -> {
                 throw AssemblyError("all string vars should have been interned into prog")
             }
-            in ArrayDatatypes -> arrayVariable2asm(variable.name, variable.dt, variable.onetimeInitializationArrayValue, variable.length)
+            in ArrayDatatypes -> {
+                arrayVariable2asm(variable.name, variable.dt, variable.align, variable.initializationArrayValue, variable.length)
+            }
             else -> {
                 throw AssemblyError("weird dt")
             }
         }
     }
 
-    private fun arrayVariable2asm(varname: String, dt: DataType, value: StArray?, orNumberOfZeros: Int?) {
+    private fun arrayVariable2asm(varname: String, dt: DataType, align: Int, value: StArray?, orNumberOfZeros: Int?) {
+        alignVar(align)
         when(dt) {
             DataType.ARRAY_UB, DataType.ARRAY_BOOL -> {
                 val data = makeArrayFillDataUnsigned(dt, value, orNumberOfZeros)
@@ -754,11 +786,16 @@ internal class ProgramAndVarsGen(
             .filter { it is PtAsmSub && it.address!=null }
             .forEach { asmsub ->
                 asmsub as PtAsmSub
-                asmgen.out("  ${asmsub.name} = ${asmsub.address!!.toHex()}")
+                val address = asmsub.address!!
+                val bank = if(address.constbank!=null) "; @bank ${address.constbank}"
+                    else if(address.varbank!=null) "; @bank ${address.varbank?.name}"
+                    else ""
+                asmgen.out("  ${asmsub.name} = ${address.address.toHex()} $bank")
             }
     }
 
-    private fun outputStringvar(varname: String, encoding: Encoding, value: String) {
+    private fun outputStringvar(varname: String, align: Int, encoding: Encoding, value: String) {
+        alignVar(align)
         asmgen.out("$varname\t; $encoding:\"${value.escape().replace("\u0000", "<NULL>")}\"", false)
         val bytes = compTarget.encodeString(value, encoding).plus(0.toUByte())
         val outputBytes = bytes.map { "$" + it.toString(16).padStart(2, '0') }

codeGenCpu6502/src/prog8/codegen/cpu6502/VariableAllocator.kt

@@ -49,8 +49,8 @@ internal class VariableAllocator(private val symboltable: SymbolTable,
         val varsRequiringZp = allVariables.filter { it.zpwish == ZeropageWish.REQUIRE_ZEROPAGE }
         val varsPreferringZp = allVariables.filter { it.zpwish == ZeropageWish.PREFER_ZEROPAGE }
         val varsNotZp = allVariables.filter { it.zpwish == ZeropageWish.NOT_IN_ZEROPAGE }
-        val varsDontCare = allVariables.filter { it.zpwish == ZeropageWish.DONTCARE }
-        require(varsDontCare.size + varsRequiringZp.size + varsPreferringZp.size + varsNotZp.size == numberOfAllocatableVariables)
+        val (varsDontCareWithoutAlignment, varsDontCareWithAlignment) = allVariables.filter { it.zpwish == ZeropageWish.DONTCARE }.partition { it.align == 0 }
+        require(varsDontCareWithAlignment.size + varsDontCareWithoutAlignment.size + varsRequiringZp.size + varsPreferringZp.size + varsNotZp.size == numberOfAllocatableVariables)
 
         var numVariablesAllocatedInZP = 0
         var numberOfNonIntegerVariables = 0
@@ -89,7 +89,7 @@ internal class VariableAllocator(private val symboltable: SymbolTable,
             // try to allocate the "don't care" interger variables into the zeropage until it is full.
             // TODO some form of intelligent priorization? most often used variables first? loopcounter vars first? ...?
             if(errors.noErrors()) {
-                val sortedList = varsDontCare.sortedByDescending { it.scopedName }
+                val sortedList = varsDontCareWithoutAlignment.sortedByDescending { it.scopedName }
                 for (variable in sortedList) {
                     if(variable.dt in IntegerDatatypesWithBoolean) {
                         if(zeropage.free.isEmpty()) {

codeGenCpu6502/src/prog8/codegen/cpu6502/assignment/AssignmentAsmGen.kt

@@ -1,8 +1,8 @@
 package prog8.codegen.cpu6502.assignment
 
 import prog8.code.StMemVar
-import prog8.code.StRomSub
-import prog8.code.StRomSubParameter
+import prog8.code.StExtSub
+import prog8.code.StExtSubParameter
 import prog8.code.StStaticVariable
 import prog8.code.ast.*
 import prog8.code.core.*
@@ -39,7 +39,7 @@ internal class AssignmentAsmGen(
         val values = assignment.value as? PtFunctionCall
             ?: throw AssemblyError("only function calls can return multiple values in a multi-assign")
 
-        val sub = asmgen.symbolTable.lookup(values.name) as? StRomSub
+        val sub = asmgen.symbolTable.lookup(values.name) as? StExtSub
             ?: throw AssemblyError("only asmsubs can return multiple values")
 
         require(sub.returns.size>=2)
@@ -64,11 +64,11 @@ internal class AssignmentAsmGen(
     }
 
     private fun assignStatusFlagsAndRegistersResults(
-        statusFlagResults: List<Pair<StRomSubParameter, PtNode>>,
-        registersResults: List<Pair<StRomSubParameter, PtNode>>
+        statusFlagResults: List<Pair<StExtSubParameter, PtNode>>,
+        registersResults: List<Pair<StExtSubParameter, PtNode>>
     ) {
 
-        fun needsToSaveA(registersResults: List<Pair<StRomSubParameter, PtNode>>): Boolean =
+        fun needsToSaveA(registersResults: List<Pair<StExtSubParameter, PtNode>>): Boolean =
             if(registersResults.isEmpty())
                 false
             else if(registersResults.all { (it.second as PtAssignTarget).identifier!=null})
@@ -91,12 +91,12 @@ internal class AssignmentAsmGen(
         }
     }
 
-    private fun assignOnlyTheStatusFlagsResults(saveA: Boolean, statusFlagResults: List<Pair<StRomSubParameter, PtNode>>) {
+    private fun assignOnlyTheStatusFlagsResults(saveA: Boolean, statusFlagResults: List<Pair<StExtSubParameter, PtNode>>) {
         // assigning flags to their variables targets requires load-instructions that destroy flags
         // so if there's more than 1, we need to save and restore the flags
         val saveFlags = statusFlagResults.size>1
 
-        fun hasFlag(statusFlagResults: List<Pair<StRomSubParameter, PtNode>>, flag: Statusflag): PtAssignTarget? {
+        fun hasFlag(statusFlagResults: List<Pair<StExtSubParameter, PtNode>>, flag: Statusflag): PtAssignTarget? {
             for ((returns, target) in statusFlagResults) {
                 if(returns.register.statusflag!! == flag)
                     return target as PtAssignTarget
@@ -121,7 +121,7 @@ internal class AssignmentAsmGen(
         if(saveA) asmgen.out("  pla")
     }
 
-    private fun assignRegisterResults(registersResults: List<Pair<StRomSubParameter, PtNode>>) {
+    private fun assignRegisterResults(registersResults: List<Pair<StExtSubParameter, PtNode>>) {
         registersResults.forEach { (returns, target) ->
             target as PtAssignTarget
             if(!target.void) {
@@ -639,6 +639,7 @@ internal class AssignmentAsmGen(
                     throw AssemblyError("Expression is too complex to translate into assembly. Split it up into several separate statements, introduce a temporary variable, or otherwise rewrite it. Location: $pos")
                 }
             }
+            is PtIfExpression -> asmgen.assignIfExpression(assign.target, value)
             else -> throw AssemblyError("weird assignment value type $value")
         }
     }
@@ -935,7 +936,7 @@ internal class AssignmentAsmGen(
                 if(!directIntoY(expr.right)) asmgen.out("  pha")
                 assignExpressionToRegister(expr.right, RegisterOrPair.Y, false)
                 if(!directIntoY(expr.right)) asmgen.out("  pla")
-                asmgen.out("  jsr  math.divmod_ub_asm")
+                asmgen.out("  jsr  prog8_math.divmod_ub_asm")
                 if(target.register==RegisterOrPair.A)
                     asmgen.out("  cmp  #0")     // fix the status register
                 else
@@ -944,7 +945,7 @@ internal class AssignmentAsmGen(
             }
             DataType.UWORD -> {
                 asmgen.assignWordOperandsToAYAndVar(expr.right, expr.left, "P8ZP_SCRATCH_W1")
-                asmgen.out("  jsr  math.divmod_uw_asm")
+                asmgen.out("  jsr  prog8_math.divmod_uw_asm")
                 assignVariableWord(target, "P8ZP_SCRATCH_W2", DataType.UWORD)
                 return true
             }
@@ -960,7 +961,7 @@ internal class AssignmentAsmGen(
                 if(!directIntoY(expr.right)) asmgen.out("  pha")
                 assignExpressionToRegister(expr.right, RegisterOrPair.Y, false)
                 if(!directIntoY(expr.right)) asmgen.out("  pla")
-                asmgen.out("  jsr  math.divmod_ub_asm")
+                asmgen.out("  jsr  prog8_math.divmod_ub_asm")
                 assignRegisterByte(target, CpuRegister.Y, false, true)
                 return true
             }
@@ -969,19 +970,19 @@ internal class AssignmentAsmGen(
                 if(!directIntoY(expr.right)) asmgen.out("  pha")
                 assignExpressionToRegister(expr.right, RegisterOrPair.Y, true)
                 if(!directIntoY(expr.right)) asmgen.out("  pla")
-                asmgen.out("  jsr  math.divmod_b_asm")
+                asmgen.out("  jsr  prog8_math.divmod_b_asm")
                 assignRegisterByte(target, CpuRegister.Y, true, true)
                 return true
             }
             DataType.UWORD -> {
                 asmgen.assignWordOperandsToAYAndVar(expr.right, expr.left, "P8ZP_SCRATCH_W1")
-                asmgen.out("  jsr  math.divmod_uw_asm")
+                asmgen.out("  jsr  prog8_math.divmod_uw_asm")
                 assignRegisterpairWord(target, RegisterOrPair.AY)
                 return true
             }
             DataType.WORD -> {
                 asmgen.assignWordOperandsToAYAndVar(expr.right, expr.left, "P8ZP_SCRATCH_W1")
-                asmgen.out("  jsr  math.divmod_w_asm")
+                asmgen.out("  jsr  prog8_math.divmod_w_asm")
                 assignRegisterpairWord(target, RegisterOrPair.AY)
                 return true
             }
@@ -998,7 +999,7 @@ internal class AssignmentAsmGen(
                     if(!directIntoY(expr.right)) asmgen.out("  pha")
                     assignExpressionToRegister(expr.right, RegisterOrPair.Y, expr.type in SignedDatatypes)
                     if(!directIntoY(expr.right)) asmgen.out("  pla")
-                    asmgen.out("  jsr  math.multiply_bytes")
+                    asmgen.out("  jsr  prog8_math.multiply_bytes")
                     assignRegisterByte(target, CpuRegister.A, false, true)
                     return true
                 }
@@ -1021,8 +1022,8 @@ internal class AssignmentAsmGen(
                         assignRegisterpairWord(target, RegisterOrPair.AY)
                         return true
                     } else {
-                        asmgen.assignWordOperandsToAYAndVar(expr.right, expr.left, "math.multiply_words.multiplier")
-                        asmgen.out("  jsr  math.multiply_words")
+                        asmgen.assignWordOperandsToAYAndVar(expr.right, expr.left, "prog8_math.multiply_words.multiplier")
+                        asmgen.out("  jsr  prog8_math.multiply_words")
                         assignRegisterpairWord(target, RegisterOrPair.AY)
                     }
                     return true
@@ -1034,16 +1035,16 @@ internal class AssignmentAsmGen(
                 in ByteDatatypes -> {
                     assignExpressionToRegister(expr.left, RegisterOrPair.A, expr.type in SignedDatatypes)
                     if (value in asmgen.optimizedByteMultiplications)
-                        asmgen.out("  jsr  math.mul_byte_${value}")
+                        asmgen.out("  jsr  prog8_math.mul_byte_${value}")
                     else
-                        asmgen.out("  ldy  #$value |  jsr  math.multiply_bytes")
+                        asmgen.out("  ldy  #$value |  jsr  prog8_math.multiply_bytes")
                     assignRegisterByte(target, CpuRegister.A, false, true)
                     return true
                 }
                 in WordDatatypes -> {
                     if (value in asmgen.optimizedWordMultiplications) {
                         assignExpressionToRegister(expr.left, RegisterOrPair.AY, expr.type in SignedDatatypes)
-                        asmgen.out("  jsr  math.mul_word_${value}")
+                        asmgen.out("  jsr  prog8_math.mul_word_${value}")
                     }
                     else {
                         if(expr.definingBlock()!!.options.veraFxMuls){
@@ -1054,8 +1055,8 @@ internal class AssignmentAsmGen(
                                 sty  cx16.r0+1
                                 jsr  verafx.muls""")
                         } else {
-                            asmgen.assignWordOperandsToAYAndVar(expr.right, expr.left, "math.multiply_words.multiplier")
-                            asmgen.out("  jsr  math.multiply_words")
+                            asmgen.assignWordOperandsToAYAndVar(expr.right, expr.left, "prog8_math.multiply_words.multiplier")
+                            asmgen.out("  jsr  prog8_math.multiply_words")
                         }
                     }
                     assignRegisterpairWord(target, RegisterOrPair.AY)
@@ -1092,11 +1093,11 @@ internal class AssignmentAsmGen(
                 asmgen.restoreRegisterStack(CpuRegister.Y, true)
                 if(expr.operator==">>")
                     if(signed)
-                        asmgen.out("  jsr  math.lsr_byte_A")
+                        asmgen.out("  jsr  prog8_math.lsr_byte_A")
                     else
-                        asmgen.out("  jsr  math.lsr_ubyte_A")
+                        asmgen.out("  jsr  prog8_math.lsr_ubyte_A")
                 else
-                    asmgen.out("  jsr  math.asl_byte_A")
+                    asmgen.out("  jsr  prog8_math.asl_byte_A")
                 assignRegisterByte(target, CpuRegister.A, signed, true)
                 return true
             } else {
@@ -1104,11 +1105,11 @@ internal class AssignmentAsmGen(
                 asmgen.restoreRegisterStack(CpuRegister.X, true)
                 if(expr.operator==">>")
                     if(signed)
-                        asmgen.out("  jsr  math.lsr_word_AY")
+                        asmgen.out("  jsr  prog8_math.lsr_word_AY")
                     else
-                        asmgen.out("  jsr  math.lsr_uword_AY")
+                        asmgen.out("  jsr  prog8_math.lsr_uword_AY")
                 else
-                    asmgen.out("  jsr  math.asl_word_AY")
+                    asmgen.out("  jsr  prog8_math.asl_word_AY")
                 assignRegisterpairWord(target, RegisterOrPair.AY)
                 return true
             }
@@ -1128,7 +1129,7 @@ internal class AssignmentAsmGen(
                                 if(shifts==1)
                                     asmgen.out("  cmp  #$80 |  ror  a")
                                 else
-                                    asmgen.out("  ldy  #$shifts |  jsr  math.lsr_byte_A")
+                                    asmgen.out("  ldy  #$shifts |  jsr  prog8_math.lsr_byte_A")
                             } else {
                                 repeat(shifts) {
                                     asmgen.out("  lsr  a")
@@ -1140,7 +1141,7 @@ internal class AssignmentAsmGen(
                     }
                     else -> {
                         if(signed && expr.operator==">>") {
-                            asmgen.out("  ldy  #$shifts |  jsr  math.lsr_byte_A")
+                            asmgen.out("  ldy  #$shifts |  jsr  prog8_math.lsr_byte_A")
                         } else {
                             asmgen.out("  lda  #0")
                         }
@@ -1173,7 +1174,7 @@ internal class AssignmentAsmGen(
                                         ror  a""")
                                 }
                                 else
-                                    asmgen.out("  ldx  #$shifts |  jsr  math.lsr_word_AY")
+                                    asmgen.out("  ldx  #$shifts |  jsr  prog8_math.lsr_word_AY")
                             } else {
                                 if(shifts>0) {
                                     asmgen.out("  sty  P8ZP_SCRATCH_B1")
@@ -1197,16 +1198,16 @@ internal class AssignmentAsmGen(
                         } else {
                             asmgen.out("  ldx  #$shifts")
                             if(signed)
-                                asmgen.out("  jsr  math.lsr_word_AY")
+                                asmgen.out("  jsr  prog8_math.lsr_word_AY")
                             else
-                                asmgen.out("  jsr  math.lsr_uword_AY")
+                                asmgen.out("  jsr  prog8_math.lsr_uword_AY")
                         }
                         assignRegisterpairWord(target, RegisterOrPair.AY)
                         return true
                     }
                     else -> {
                         if(signed && expr.operator==">>") {
-                            asmgen.out("  ldx  #$shifts |  jsr  math.lsr_word_AY")
+                            asmgen.out("  ldx  #$shifts |  jsr  prog8_math.lsr_word_AY")
                         } else {
                             asmgen.out("  lda  #0 |  ldy  #0")
                         }

codeGenCpu6502/src/prog8/codegen/cpu6502/assignment/AugmentableAssignmentAsmGen.kt

@@ -904,9 +904,9 @@ internal class AugmentableAssignmentAsmGen(private val program: PtProgram,
         when (operator) {
             "+" -> asmgen.out("  clc |  adc  $otherName")
             "-" -> asmgen.out("  sec |  sbc  $otherName")
-            "*" -> asmgen.out("  ldy  $otherName |  jsr  math.multiply_bytes")
-            "/" -> asmgen.out("  ldy  $otherName |  jsr  math.divmod_ub_asm |  tya")
-            "%" -> asmgen.out("  ldy  $otherName |  jsr  math.divmod_ub_asm")
+            "*" -> asmgen.out("  ldy  $otherName |  jsr  prog8_math.multiply_bytes")
+            "/" -> asmgen.out("  ldy  $otherName |  jsr  prog8_math.divmod_ub_asm |  tya")
+            "%" -> asmgen.out("  ldy  $otherName |  jsr  prog8_math.divmod_ub_asm")
             "<<" -> {
                 asmgen.out("""
                         ldy  $otherName
@@ -981,23 +981,23 @@ internal class AugmentableAssignmentAsmGen(private val program: PtProgram,
             "*" -> {
                 val sourceName = asmgen.loadByteFromPointerIntoA(pointervar)
                 if(value in asmgen.optimizedByteMultiplications)
-                    asmgen.out("  jsr  math.mul_byte_${value}")
+                    asmgen.out("  jsr  prog8_math.mul_byte_${value}")
                 else
-                    asmgen.out("  ldy  #$value |  jsr  math.multiply_bytes")
+                    asmgen.out("  ldy  #$value |  jsr  prog8_math.multiply_bytes")
                 asmgen.storeAIntoZpPointerVar(sourceName, false)
             }
             "/" -> {
                 val sourceName = asmgen.loadByteFromPointerIntoA(pointervar)
                 if(value==0)
                     throw AssemblyError("division by zero")
-                asmgen.out("  ldy  #$value |  jsr  math.divmod_ub_asm |  tya")
+                asmgen.out("  ldy  #$value |  jsr  prog8_math.divmod_ub_asm |  tya")
                 asmgen.storeAIntoZpPointerVar(sourceName, false)
             }
             "%" -> {
                 val sourceName = asmgen.loadByteFromPointerIntoA(pointervar)
                 if(value==0)
                     throw AssemblyError("division by zero")
-                asmgen.out("  ldy  #$value |  jsr  math.divmod_ub_asm")
+                asmgen.out("  ldy  #$value |  jsr  prog8_math.divmod_ub_asm")
                 asmgen.storeAIntoZpPointerVar(sourceName, false)
             }
             "<<" -> {
@@ -1143,17 +1143,17 @@ $shortcutLabel:""")
         when (operator) {
             "+" -> asmgen.out("  clc |  adc  $variable")
             "-" -> asmgen.out("  sec |  sbc  $variable")
-            "*" -> asmgen.out("  ldy  $variable  |  jsr  math.multiply_bytes")
+            "*" -> asmgen.out("  ldy  $variable  |  jsr  prog8_math.multiply_bytes")
             "/" -> {
                 if(signed)
-                    asmgen.out("  ldy  $variable  |  jsr  math.divmod_b_asm |  tya")
+                    asmgen.out("  ldy  $variable  |  jsr  prog8_math.divmod_b_asm |  tya")
                 else
-                    asmgen.out("  ldy  $variable  |  jsr  math.divmod_ub_asm |  tya")
+                    asmgen.out("  ldy  $variable  |  jsr  prog8_math.divmod_ub_asm |  tya")
             }
             "%" -> {
                 if(signed)
                     throw AssemblyError("remainder of signed integers is not properly defined/implemented, use unsigned instead")
-                asmgen.out("  ldy  $variable  |  jsr  math.divmod_ub_asm")
+                asmgen.out("  ldy  $variable  |  jsr  prog8_math.divmod_ub_asm")
             }
             "<<" -> {
                 asmgen.out("""
@@ -1318,14 +1318,14 @@ $shortcutLabel:""")
             }
             "/" -> {
                 if(signed)
-                    asmgen.out("  tay |  lda  $variable  |  jsr  math.divmod_b_asm |  tya")
+                    asmgen.out("  tay |  lda  $variable  |  jsr  prog8_math.divmod_b_asm |  tya")
                 else
-                    asmgen.out("  tay |  lda  $variable  |  jsr  math.divmod_ub_asm |  tya")
+                    asmgen.out("  tay |  lda  $variable  |  jsr  prog8_math.divmod_ub_asm |  tya")
             }
             "%" -> {
                 if(signed)
                     throw AssemblyError("remainder of signed integers is not properly defined/implemented, use unsigned instead")
-                asmgen.out("  tay |  lda  $variable  |  jsr  math.divmod_ub_asm")
+                asmgen.out("  tay |  lda  $variable  |  jsr  prog8_math.divmod_ub_asm")
             }
             "<<" -> {
                 asmgen.out("""
@@ -1478,16 +1478,16 @@ $shortcutLabel:""")
             }
             "*" -> {
                 if(value in asmgen.optimizedByteMultiplications)
-                    asmgen.out("  lda  $name |  jsr  math.mul_byte_$value |  sta  $name")
+                    asmgen.out("  lda  $name |  jsr  prog8_math.mul_byte_$value |  sta  $name")
                 else
-                    asmgen.out("  lda  $name |  ldy  #$value |  jsr  math.multiply_bytes |  sta  $name")
+                    asmgen.out("  lda  $name |  ldy  #$value |  jsr  prog8_math.multiply_bytes |  sta  $name")
             }
             "/" -> {
                 // replacing division by shifting is done in an optimizer step.
                 if (dt == DataType.UBYTE)
-                    asmgen.out("  lda  $name |  ldy  #$value |  jsr  math.divmod_ub_asm |  sty  $name")
+                    asmgen.out("  lda  $name |  ldy  #$value |  jsr  prog8_math.divmod_ub_asm |  sty  $name")
                 else
-                    asmgen.out("  lda  $name |  ldy  #$value |  jsr  math.divmod_b_asm |  sty  $name")
+                    asmgen.out("  lda  $name |  ldy  #$value |  jsr  prog8_math.divmod_b_asm |  sty  $name")
             }
             "%" -> {
                 if(dt==DataType.BYTE)
@@ -1495,7 +1495,7 @@ $shortcutLabel:""")
                 asmgen.out("""
                     lda  $name
                     ldy  #$value
-                    jsr  math.divmod_ub_asm
+                    jsr  prog8_math.divmod_ub_asm
                     sta  $name""")
             }
             "<<" -> {
@@ -1529,7 +1529,7 @@ $shortcutLabel:""")
                             value>3 -> asmgen.out("""
                                 lda  $name
                                 ldy  #$value
-                                jsr  math.lsr_byte_A
+                                jsr  prog8_math.lsr_byte_A
                                 sta  $name""")
                             else -> repeat(value) { asmgen.out("  lda  $name | asl  a |  ror  $name") }
                         }
@@ -1802,7 +1802,7 @@ $shortcutLabel:""")
             "*" -> {
                 // the mul code works for both signed and unsigned
                 if(value in asmgen.optimizedWordMultiplications) {
-                    asmgen.out("  lda  $lsb |  ldy  $msb |  jsr  math.mul_word_$value |  sta  $lsb |  sty  $msb")
+                    asmgen.out("  lda  $lsb |  ldy  $msb |  jsr  prog8_math.mul_word_$value |  sta  $lsb |  sty  $msb")
                 } else {
                     if(block?.options?.veraFxMuls==true)
                         // cx16 verafx hardware mul
@@ -1821,12 +1821,12 @@ $shortcutLabel:""")
                     else
                         asmgen.out("""
                             lda  $lsb
-                            sta  math.multiply_words.multiplier
+                            sta  prog8_math.multiply_words.multiplier
                             lda  $msb
-                            sta  math.multiply_words.multiplier+1
+                            sta  prog8_math.multiply_words.multiplier+1
                             lda  #<$value
                             ldy  #>$value
-                            jsr  math.multiply_words
+                            jsr  prog8_math.multiply_words
                             sta  $lsb
                             sty  $msb""")
                 }
@@ -1844,7 +1844,7 @@ $shortcutLabel:""")
                             sty  P8ZP_SCRATCH_W1+1
                             lda  #<$value
                             ldy  #>$value
-                            jsr  math.divmod_w_asm
+                            jsr  prog8_math.divmod_w_asm
                             sta  $lsb
                             sty  $msb
                         """)
@@ -1857,7 +1857,7 @@ $shortcutLabel:""")
                             sty  P8ZP_SCRATCH_W1+1
                             lda  #<$value
                             ldy  #>$value
-                            jsr  math.divmod_uw_asm
+                            jsr  prog8_math.divmod_uw_asm
                             sta  $lsb
                             sty  $msb
                         """)
@@ -1876,7 +1876,7 @@ $shortcutLabel:""")
                     sty  P8ZP_SCRATCH_W1+1
                     lda  #<$value
                     ldy  #>$value
-                    jsr  math.divmod_uw_asm
+                    jsr  prog8_math.divmod_uw_asm
                     lda  P8ZP_SCRATCH_W2
                     ldy  P8ZP_SCRATCH_W2+1
                     sta  $lsb
@@ -2297,20 +2297,20 @@ $shortcutLabel:""")
                                 sty  $name+1""")
                         } else {
                             if(valueDt==DataType.UBYTE) {
-                                asmgen.out("  lda  $otherName |  sta  math.multiply_words.multiplier")
+                                asmgen.out("  lda  $otherName |  sta  prog8_math.multiply_words.multiplier")
                                 if(asmgen.isTargetCpu(CpuType.CPU65c02))
-                                    asmgen.out("  stz  math.multiply_words.multiplier+1")
+                                    asmgen.out("  stz  prog8_math.multiply_words.multiplier+1")
                                 else
-                                    asmgen.out("  lda  #0 |  sta  math.multiply_words.multiplier+1")
+                                    asmgen.out("  lda  #0 |  sta  prog8_math.multiply_words.multiplier+1")
                             } else {
                                 asmgen.out("  lda  $otherName")
                                 asmgen.signExtendAYlsb(valueDt)
-                                asmgen.out("  sta  math.multiply_words.multiplier |  sty  math.multiply_words.multiplier+1")
+                                asmgen.out("  sta  prog8_math.multiply_words.multiplier |  sty  prog8_math.multiply_words.multiplier+1")
                             }
                             asmgen.out("""
                                     lda  $name
                                     ldy  $name+1
-                                    jsr  math.multiply_words
+                                    jsr  prog8_math.multiply_words
                                     sta  $name
                                     sty  $name+1""")
                         }
@@ -2324,7 +2324,7 @@ $shortcutLabel:""")
                                 sty  P8ZP_SCRATCH_W1+1
                                 lda  $otherName
                                 ldy  #0
-                                jsr  math.divmod_uw_asm
+                                jsr  prog8_math.divmod_uw_asm
                                 sta  $name
                                 sty  $name+1
                             """)
@@ -2336,7 +2336,7 @@ $shortcutLabel:""")
                                 sty  P8ZP_SCRATCH_W1+1
                                 lda  $otherName
                                 ldy  #0
-                                jsr  math.divmod_w_asm
+                                jsr  prog8_math.divmod_w_asm
                                 sta  $name
                                 sty  $name+1
                             """)
@@ -2352,7 +2352,7 @@ $shortcutLabel:""")
                             sty  P8ZP_SCRATCH_W1+1
                             lda  $otherName
                             ldy  #0
-                            jsr  math.divmod_uw_asm
+                            jsr  prog8_math.divmod_uw_asm
                             lda  P8ZP_SCRATCH_W2
                             sta  $name
                             lda  P8ZP_SCRATCH_W2+1
@@ -2460,11 +2460,11 @@ $shortcutLabel:""")
                             asmgen.out("""
                                 lda  $otherName
                                 ldy  $otherName+1
-                                sta  math.multiply_words.multiplier
-                                sty  math.multiply_words.multiplier+1
+                                sta  prog8_math.multiply_words.multiplier
+                                sty  prog8_math.multiply_words.multiplier+1
                                 lda  $name
                                 ldy  $name+1
-                                jsr  math.multiply_words
+                                jsr  prog8_math.multiply_words
                                 sta  $name
                                 sty  $name+1""")
                     }
@@ -2477,7 +2477,7 @@ $shortcutLabel:""")
                                 sty  P8ZP_SCRATCH_W1+1
                                 lda  $otherName
                                 ldy  $otherName+1
-                                jsr  math.divmod_w_asm
+                                jsr  prog8_math.divmod_w_asm
                                 sta  $name
                                 sty  $name+1""")
                         }
@@ -2489,7 +2489,7 @@ $shortcutLabel:""")
                                 sty  P8ZP_SCRATCH_W1+1
                                 lda  $otherName
                                 ldy  $otherName+1
-                                jsr  math.divmod_uw_asm
+                                jsr  prog8_math.divmod_uw_asm
                                 sta  $name
                                 sty  $name+1""")
                         }
@@ -2504,7 +2504,7 @@ $shortcutLabel:""")
                             sty  P8ZP_SCRATCH_W1+1
                             lda  $otherName
                             ldy  $otherName+1
-                            jsr  math.divmod_uw_asm
+                            jsr  prog8_math.divmod_uw_asm
                             lda  P8ZP_SCRATCH_W2
                             sta  $name
                             lda  P8ZP_SCRATCH_W2+1
@@ -2661,11 +2661,11 @@ $shortcutLabel:""")
                 """)
             else
                 asmgen.out("""
-                    sta  math.multiply_words.multiplier
-                    sty  math.multiply_words.multiplier+1
+                    sta  prog8_math.multiply_words.multiplier
+                    sty  prog8_math.multiply_words.multiplier+1
                     lda  $name
                     ldy  $name+1
-                    jsr  math.multiply_words
+                    jsr  prog8_math.multiply_words
                     sta  $name
                     sty  $name+1
                 """)
@@ -2680,9 +2680,9 @@ $shortcutLabel:""")
                     sta  P8ZP_SCRATCH_W1+1
                     txa""")
             if (dt == DataType.WORD)
-                asmgen.out("  jsr  math.divmod_w_asm")
+                asmgen.out("  jsr  prog8_math.divmod_w_asm")
             else
-                asmgen.out("  jsr  math.divmod_uw_asm")
+                asmgen.out("  jsr  prog8_math.divmod_uw_asm")
             asmgen.out("  sta  $name |  sty  $name+1")
         }
 
@@ -2696,7 +2696,7 @@ $shortcutLabel:""")
                 lda  $name+1
                 sta  P8ZP_SCRATCH_W1+1
                 txa
-                jsr  math.divmod_uw_asm
+                jsr  prog8_math.divmod_uw_asm
                 lda  P8ZP_SCRATCH_W2
                 ldy  P8ZP_SCRATCH_W2+1
                 sta  $name

codeGenCpu6502/test/TestCodegen.kt

@@ -26,7 +26,8 @@ class TestCodegen: FunSpec({
             floats = true,
             noSysInit = false,
             compTarget = target,
-            loadAddress = target.machine.PROGRAM_LOAD_ADDRESS
+            loadAddress = target.machine.PROGRAM_LOAD_ADDRESS,
+            memtopAddress = 0xffffu
         )
     }
 
@@ -43,14 +44,46 @@ class TestCodegen: FunSpec({
 //        xx += cx16.r0
 //    }
 //}
-        val codegen = AsmGen6502(prefixSymbols = false)
+        val codegen = AsmGen6502(prefixSymbols = false, 0)
         val program = PtProgram("test", DummyMemsizer, DummyStringEncoder)
         val block = PtBlock("main",false, SourceCode.Generated("test"), PtBlock.Options(), Position.DUMMY)
         val sub = PtSub("start", emptyList(), null, Position.DUMMY)
-        sub.add(PtVariable("pi", DataType.UBYTE, ZeropageWish.DONTCARE, PtNumber(DataType.UBYTE, 0.0, Position.DUMMY), null, Position.DUMMY))
-        sub.add(PtVariable("particleX", DataType.ARRAY_UB, ZeropageWish.DONTCARE, null, 3u, Position.DUMMY))
-        sub.add(PtVariable("particleDX", DataType.ARRAY_UB, ZeropageWish.DONTCARE, null, 3u, Position.DUMMY))
-        sub.add(PtVariable("xx", DataType.WORD, ZeropageWish.DONTCARE, PtNumber(DataType.WORD, 1.0, Position.DUMMY), null, Position.DUMMY))
+        sub.add(PtVariable(
+            "pi",
+            DataType.UBYTE,
+            ZeropageWish.DONTCARE,
+            0u,
+            PtNumber(DataType.UBYTE, 0.0, Position.DUMMY),
+            null,
+            Position.DUMMY
+        ))
+        sub.add(PtVariable(
+            "particleX",
+            DataType.ARRAY_UB,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            3u,
+            Position.DUMMY
+        ))
+        sub.add(PtVariable(
+            "particleDX",
+            DataType.ARRAY_UB,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            3u,
+            Position.DUMMY
+        ))
+        sub.add(PtVariable(
+            "xx",
+            DataType.WORD,
+            ZeropageWish.DONTCARE,
+            0u,
+            PtNumber(DataType.WORD, 1.0, Position.DUMMY),
+            null,
+            Position.DUMMY
+        ))
 
         val assign = PtAugmentedAssign("+=", Position.DUMMY)
         val target = PtAssignTarget(false, Position.DUMMY).also {

codeGenIntermediate/src/prog8/codegen/intermediate/AssignmentGen.kt

@@ -1,7 +1,7 @@
 package prog8.codegen.intermediate
 
-import prog8.code.StRomSub
-import prog8.code.StRomSubParameter
+import prog8.code.StExtSub
+import prog8.code.StExtSubParameter
 import prog8.code.ast.*
 import prog8.code.core.*
 import prog8.intermediate.*
@@ -14,7 +14,7 @@ internal class AssignmentGen(private val codeGen: IRCodeGen, private val express
             val values = assignment.value as? PtFunctionCall
                 ?: throw AssemblyError("only function calls can return multiple values in a multi-assign")
 
-            val sub = codeGen.symbolTable.lookup(values.name) as? StRomSub
+            val sub = codeGen.symbolTable.lookup(values.name) as? StExtSub
                 ?: throw AssemblyError("only asmsubs can return multiple values")
 
             val result = mutableListOf<IRCodeChunkBase>()
@@ -46,7 +46,7 @@ internal class AssignmentGen(private val codeGen: IRCodeGen, private val express
         }
     }
 
-    private fun assignCpuRegister(returns: StRomSubParameter, regNum: Int, target: PtAssignTarget): IRCodeChunks {
+    private fun assignCpuRegister(returns: StExtSubParameter, regNum: Int, target: PtAssignTarget): IRCodeChunks {
         val result = mutableListOf<IRCodeChunkBase>()
         val loadCpuRegInstr = when(returns.register.registerOrPair) {
             RegisterOrPair.A -> IRInstruction(Opcode.LOADHA, IRDataType.BYTE, reg1=regNum)
@@ -278,7 +278,6 @@ internal class AssignmentGen(private val codeGen: IRCodeGen, private val express
                 }
             }
             "not" -> {
-                // TODO: in boolean branch, is 'not' handled ok like this?
                 val register = codeGen.registers.nextFree()
                 if(constIndex!=null) {
                     result += IRCodeChunk(null, null).also {
@@ -360,7 +359,7 @@ internal class AssignmentGen(private val codeGen: IRCodeGen, private val express
             val variable = targetArray.variable.name
             val itemsize = codeGen.program.memsizer.memorySize(targetArray.type)
 
-            val fixedIndex = constIntValue(targetArray.index)
+            val fixedIndex = targetArray.index.asConstInteger()
             val arrayLength = codeGen.symbolTable.getLength(targetArray.variable.name)
             if(zero) {
                 if(fixedIndex!=null) {
@@ -884,7 +883,7 @@ internal class AssignmentGen(private val codeGen: IRCodeGen, private val express
     }
 
     private fun operatorMultiplyInplaceSplitArray(array: PtArrayIndexer, operand: PtExpression): IRCodeChunks? {
-        return null //    TODO("inplace split word array *")
+        return null  // TODO("inplace split word array *")
     }
 
     private fun operatorMinusInplaceSplitArray(array: PtArrayIndexer, operand: PtExpression): IRCodeChunks? {
@@ -1371,7 +1370,7 @@ internal class AssignmentGen(private val codeGen: IRCodeGen, private val express
 
     private fun createInplaceArrayComparison(array: PtArrayIndexer, value: PtExpression, comparisonOperator: String): IRCodeChunks? {
         if(array.type==DataType.FLOAT)
-            return null  // TODO("optimized in-place compare on float arrays"))   // TODO?
+            return null  // TODO("optimized in-place compare on float arrays"))
 
         val eltSize = codeGen.program.memsizer.memorySize(array.type)
         val result = mutableListOf<IRCodeChunkBase>()
@@ -1390,10 +1389,10 @@ internal class AssignmentGen(private val codeGen: IRCodeGen, private val express
                     it += when(comparisonOperator) {
                         "==" -> IRInstruction(Opcode.SZ, vmDt, reg1=cmpResultReg, reg2=valueReg)
                         "!=" -> IRInstruction(Opcode.SNZ, vmDt, reg1=cmpResultReg, reg2=valueReg)
-                        "<" -> return null  // TODO("array <0 inplace")) // TODO?
-                        "<=" -> return null  // TODO("array <=0 inplace")) // TODO?
-                        ">" -> return null  // TODO("array >0 inplace")) // TODO?
-                        ">=" -> return null  // TODO("array >=0 inplace")) // TODO?
+                        "<" -> return null  // TODO("array <0 inplace"))
+                        "<=" -> return null  // TODO("array <=0 inplace"))
+                        ">" -> return null  // TODO("array >0 inplace"))
+                        ">=" -> return null  // TODO("array >=0 inplace"))
                         else -> throw AssemblyError("invalid operator")
                     }
                     it += IRInstruction(Opcode.STOREM, vmDt, reg1 = cmpResultReg, labelSymbol = array.variable.name, symbolOffset = constIndex*eltSize)
@@ -1413,10 +1412,10 @@ internal class AssignmentGen(private val codeGen: IRCodeGen, private val express
                     it += when(comparisonOperator) {
                         "==" -> IRInstruction(Opcode.SZ, vmDt, reg1=cmpResultReg, reg2=valueReg)
                         "!=" -> IRInstruction(Opcode.SNZ, vmDt, reg1=cmpResultReg, reg2=valueReg)
-                        "<" -> return null  // TODO("array <0 inplace")) // TODO?
-                        "<=" -> return null  // TODO("array <=0 inplace")) // TODO?
-                        ">" -> return null  // TODO("array >0 inplace")) // TODO?
-                        ">=" -> return null  // TODO("array >=0 inplace")) // TODO?
+                        "<" -> return null  // TODO("array <0 inplace"))
+                        "<=" -> return null  // TODO("array <=0 inplace"))
+                        ">" -> return null  // TODO("array >0 inplace"))
+                        ">=" -> return null  // TODO("array >=0 inplace"))
                         else -> throw AssemblyError("invalid operator")
                     }
                     it += IRInstruction(Opcode.STOREX, vmDt, reg1=valueReg, reg2=indexTr.resultReg, labelSymbol = array.variable.name)

codeGenIntermediate/src/prog8/codegen/intermediate/BuiltinFuncGen.kt

@@ -39,77 +39,15 @@ internal class BuiltinFuncGen(private val codeGen: IRCodeGen, private val exprGe
             "ror" -> funcRolRor(call)
             "rol2" -> funcRolRor(call)
             "ror2" -> funcRolRor(call)
-            "prog8_ifelse_bittest_set" -> throw AssemblyError("prog8_ifelse_bittest_set() should have been translated as part of an ifElse statement")
-            "prog8_ifelse_bittest_notset" -> throw AssemblyError("prog8_ifelse_bittest_notset() should have been translated as part of an ifElse statement")
+            "prog8_ifelse_bittest_set" -> throw AssemblyError("prog8_ifelse_bittest_set() should have been translated as part of an ifElse/ifExpression statement")
+            "prog8_ifelse_bittest_notset" -> throw AssemblyError("prog8_ifelse_bittest_notset() should have been translated as part of an ifElse/ifExpression statement")
             "prog8_lib_stringcompare" -> funcStringCompare(call)
             "prog8_lib_square_byte" -> funcSquare(call, IRDataType.BYTE)
             "prog8_lib_square_word" -> funcSquare(call, IRDataType.WORD)
-            "prog8_lib_arraycopy" -> funcArrayCopy(call)
             else -> throw AssemblyError("missing builtinfunc for ${call.name}")
         }
     }
 
-    private fun funcArrayCopy(call: PtBuiltinFunctionCall): ExpressionCodeResult {
-        val source = call.args[0] as PtIdentifier
-        val target = call.args[1] as PtIdentifier
-        val sourceLength = codeGen.symbolTable.getLength(source.name)!!
-        val targetLength = codeGen.symbolTable.getLength(target.name)!!
-        require(sourceLength==targetLength)
-        val result = mutableListOf<IRCodeChunkBase>()
-        val fromReg = codeGen.registers.nextFree()
-        val toReg = codeGen.registers.nextFree()
-        val countReg = codeGen.registers.nextFree()
-        if(source.type in SplitWordArrayTypes && target.type in SplitWordArrayTypes) {
-            // split words -> split words, copy lsb and msb arrays separately
-            result += IRCodeChunk(null, null).also {
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=fromReg, labelSymbol = source.name+"_lsb")
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=toReg, labelSymbol = target.name+"_lsb")
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=countReg, immediate = sourceLength)
-                it += codeGen.makeSyscall(IMSyscall.MEMCOPY_SMALL, listOf(IRDataType.WORD to fromReg, IRDataType.WORD to toReg, IRDataType.BYTE to (countReg and 255)), returns = null)
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=fromReg, labelSymbol = source.name+"_msb")
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=toReg, labelSymbol = target.name+"_msb")
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=countReg, immediate = sourceLength)
-                it += codeGen.makeSyscall(IMSyscall.MEMCOPY_SMALL, listOf(IRDataType.WORD to fromReg, IRDataType.WORD to toReg, IRDataType.BYTE to (countReg and 255)), returns = null)
-            }
-        }
-        else if(source.type in SplitWordArrayTypes) {
-            // split -> normal words
-            require(target.type==DataType.ARRAY_UW || target.type==DataType.ARRAY_W)
-            val fromRegMsb = codeGen.registers.nextFree()
-            result += IRCodeChunk(null, null).also {
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=fromReg, labelSymbol = source.name+"_lsb")
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=fromRegMsb, labelSymbol = source.name+"_msb")
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=toReg, labelSymbol = target.name)
-                it += IRInstruction(Opcode.LOAD, IRDataType.BYTE, reg1=countReg, immediate = sourceLength)
-            }
-            result += codeGen.makeSyscall(IMSyscall.ARRAYCOPY_SPLITW_TO_NORMAL, listOf(IRDataType.WORD to fromReg, IRDataType.WORD to fromRegMsb, IRDataType.WORD to toReg, IRDataType.BYTE to countReg), returns = null)
-        }
-        else if(target.type in SplitWordArrayTypes) {
-            // normal -> split words
-            require(source.type==DataType.ARRAY_UW || source.type==DataType.ARRAY_W)
-            val toRegMsb = codeGen.registers.nextFree()
-            result += IRCodeChunk(null, null).also {
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=fromReg, labelSymbol = source.name)
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=toReg, labelSymbol = target.name+"_lsb")
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=toRegMsb, labelSymbol = target.name+"_msb")
-                it += IRInstruction(Opcode.LOAD, IRDataType.BYTE, reg1=countReg, immediate = sourceLength)
-            }
-            result += codeGen.makeSyscall(IMSyscall.ARRAYCOPY_NORMAL_TO_SPLITW, listOf(IRDataType.WORD to fromReg, IRDataType.WORD to toReg, IRDataType.WORD to toRegMsb, IRDataType.BYTE to countReg), returns = null)
-        }
-        else {
-            // normal array to array copy (various element types)
-            val eltsize = codeGen.options.compTarget.memorySize(ArrayToElementTypes.getValue(source.type))
-            result += IRCodeChunk(null, null).also {
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=fromReg, labelSymbol = source.name)
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=toReg, labelSymbol = target.name)
-                it += IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=countReg, immediate = sourceLength * eltsize)
-            }
-            result += codeGen.makeSyscall(IMSyscall.MEMCOPY_SMALL, listOf(IRDataType.WORD to fromReg, IRDataType.WORD to toReg, IRDataType.BYTE to (countReg and 255)), returns = null)
-        }
-
-        return ExpressionCodeResult(result, IRDataType.BYTE, -1, -1)
-    }
-
     private fun funcSquare(call: PtBuiltinFunctionCall, resultType: IRDataType): ExpressionCodeResult {
         val result = mutableListOf<IRCodeChunkBase>()
         val valueTr = exprGen.translateExpression(call.args[0])

codeGenIntermediate/src/prog8/codegen/intermediate/ExpressionGen.kt

@@ -1,7 +1,7 @@
 package prog8.codegen.intermediate
 
 import prog8.code.StNode
-import prog8.code.StRomSub
+import prog8.code.StExtSub
 import prog8.code.StSub
 import prog8.code.ast.*
 import prog8.code.core.*
@@ -79,6 +79,7 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
             is PtPrefix -> translate(expr)
             is PtArrayIndexer -> translate(expr)
             is PtBinaryExpression -> translate(expr)
+            is PtIfExpression -> translate(expr)
             is PtBuiltinFunctionCall -> codeGen.translateBuiltinFunc(expr)
             is PtFunctionCall -> translate(expr)
             is PtContainmentCheck -> translate(expr)
@@ -89,6 +90,41 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
         }
     }
 
+    private fun translate(ifExpr: PtIfExpression): ExpressionCodeResult {
+
+        if((ifExpr.condition as? PtPrefix)?.operator=="not")
+            throw AssemblyError("not prefix in ifexpression should have been replaced by swapped values")
+
+        // TODO don't store condition as expression result but just use the flags, like a normal PtIfElse translation does
+        val condTr = translateExpression(ifExpr.condition)
+        val trueTr = translateExpression(ifExpr.truevalue)
+        val falseTr = translateExpression(ifExpr.falsevalue)
+        val irDt = irType(ifExpr.type)
+        val result = mutableListOf<IRCodeChunkBase>()
+        val falseLabel = codeGen.createLabelName()
+        val endLabel = codeGen.createLabelName()
+
+        addToResult(result, condTr, condTr.resultReg, -1)
+        addInstr(result, IRInstruction(Opcode.CMPI, IRDataType.BYTE, reg1=condTr.resultReg, immediate = 0), null)
+        addInstr(result, IRInstruction(Opcode.BSTEQ, labelSymbol = falseLabel), null)
+
+        if (irDt != IRDataType.FLOAT) {
+            addToResult(result, trueTr, trueTr.resultReg, -1)
+            addInstr(result, IRInstruction(Opcode.JUMP, labelSymbol = endLabel), null)
+            result += IRCodeChunk(falseLabel, null)
+            addToResult(result, falseTr, trueTr.resultReg, -1)
+            result += IRCodeChunk(endLabel, null)
+            return ExpressionCodeResult(result, irDt, trueTr.resultReg, -1)
+        } else {
+            addToResult(result, trueTr, -1, trueTr.resultFpReg)
+            addInstr(result, IRInstruction(Opcode.JUMP, labelSymbol = endLabel), null)
+            result += IRCodeChunk(falseLabel, null)
+            addToResult(result, falseTr, -1, trueTr.resultFpReg)
+            result += IRCodeChunk(endLabel, null)
+            return ExpressionCodeResult(result, irDt, -1, trueTr.resultFpReg)
+        }
+    }
+
     private fun translate(expr: PtAddressOf): ExpressionCodeResult {
         val vmDt = irType(expr.type)
         val symbol = expr.identifier.name
@@ -492,7 +528,7 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
     fun translate(fcall: PtFunctionCall): ExpressionCodeResult {
         val callTarget = codeGen.symbolTable.flat.getValue(fcall.name)
 
-        if(callTarget.scopedName in listOf("sys.push", "sys.pushw", "sys.pop", "sys.popw")) {
+        if(callTarget.scopedName in listOf("sys.push", "sys.pushw", "sys.pop", "sys.popw", "floats.push", "floats.pop")) {
             // special case, these should be inlined, or even use specialized instructions. Instead of doing a normal subroutine call.
             return translateStackFunctions(fcall, callTarget)
         }
@@ -552,7 +588,7 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
                 else
                     ExpressionCodeResult(result, returnRegSpec!!.dt, returnRegSpec.registerNum, -1)
             }
-            is StRomSub -> {
+            is StExtSub -> {
                 val result = mutableListOf<IRCodeChunkBase>()
                 addInstr(result, IRInstruction(Opcode.PREPARECALL, immediate = callTarget.parameters.size), null)
                 // assign the arguments
@@ -572,6 +608,7 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
                         RegisterOrPair.AX -> addInstr(result, IRInstruction(Opcode.STOREHAX, IRDataType.WORD, reg1=tr.resultReg), null)
                         RegisterOrPair.AY -> addInstr(result, IRInstruction(Opcode.STOREHAY, IRDataType.WORD, reg1=tr.resultReg), null)
                         RegisterOrPair.XY -> addInstr(result, IRInstruction(Opcode.STOREHXY, IRDataType.WORD, reg1=tr.resultReg), null)
+                        RegisterOrPair.FAC1, RegisterOrPair.FAC2 -> TODO("floating point register parameters not supported")
                         in Cx16VirtualRegisters -> {
                             addInstr(result, IRInstruction(Opcode.STOREM, paramDt, reg1=tr.resultReg, labelSymbol = "cx16.${parameter.register.registerOrPair.toString().lowercase()}"), null)
                         }
@@ -585,7 +622,7 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
                 }
 
                 if(callTarget.returns.size>1)
-                    return callRomSubWithMultipleReturnValues(callTarget, fcall, argRegisters, result)
+                    return callExtSubWithMultipleReturnValues(callTarget, fcall, argRegisters, result)
 
                 // return a single value (or nothing)
                 val returnRegSpec = if(fcall.void) null else {
@@ -607,8 +644,18 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
                 val call =
                     if(callTarget.address==null)
                         IRInstruction(Opcode.CALL, labelSymbol = fcall.name, fcallArgs = FunctionCallArgs(argRegisters, returnRegs))
-                    else
-                        IRInstruction(Opcode.CALL, address = callTarget.address!!.toInt(), fcallArgs = FunctionCallArgs(argRegisters, returnRegs))
+                    else {
+                        val address = callTarget.address!!
+                        if(address.constbank==null && address.varbank==null) {
+                            IRInstruction(
+                                Opcode.CALL,
+                                address = address.address.toInt(),
+                                fcallArgs = FunctionCallArgs(argRegisters, returnRegs))
+                        }
+                        else {
+                            TODO("callfar into another bank is not implemented for the selected compilation target")
+                        }
+                    }
                 addInstr(result, call, null)
                 var finalReturnRegister = returnRegSpec?.registerNum ?: -1
 
@@ -687,12 +734,25 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
                 addInstr(chunk, IRInstruction(Opcode.POP, IRDataType.WORD, reg1=popReg), null)
                 return ExpressionCodeResult(chunk, IRDataType.WORD, popReg, -1)
             }
+            "floats.push" -> {
+                // push float
+                val tr = translateExpression(fcall.args.single())
+                chunk += tr.chunks
+                addInstr(chunk, IRInstruction(Opcode.PUSH, IRDataType.FLOAT, fpReg1 = tr.resultFpReg), null)
+                return ExpressionCodeResult(chunk, IRDataType.FLOAT, -1, -1)
+            }
+            "floats.pop" -> {
+                // pop float
+                val popReg = codeGen.registers.nextFreeFloat()
+                addInstr(chunk, IRInstruction(Opcode.POP, IRDataType.FLOAT, fpReg1 = popReg), null)
+                return ExpressionCodeResult(chunk, IRDataType.FLOAT, -1, resultFpReg = popReg)
+            }
             else -> throw AssemblyError("unknown stack subroutine called")
         }
     }
 
-    private fun callRomSubWithMultipleReturnValues(
-        callTarget: StRomSub,
+    private fun callExtSubWithMultipleReturnValues(
+        callTarget: StExtSub,
         fcall: PtFunctionCall,
         argRegisters: MutableList<FunctionCallArgs.ArgumentSpec>,
         result: MutableList<IRCodeChunkBase>
@@ -706,8 +766,17 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
         val call =
             if(callTarget.address==null)
                 IRInstruction(Opcode.CALL, labelSymbol = fcall.name, fcallArgs = FunctionCallArgs(argRegisters, returnRegisters))
-            else
-                IRInstruction(Opcode.CALL, address = callTarget.address!!.toInt(), fcallArgs = FunctionCallArgs(argRegisters, returnRegisters))
+            else {
+                val address = callTarget.address!!
+                if(address.constbank==null && address.varbank==null) {
+                    IRInstruction(
+                        Opcode.CALL,
+                        address = address.address.toInt(),
+                        fcallArgs = FunctionCallArgs(argRegisters, returnRegisters)
+                    )
+                }
+                else TODO("extsub with banked address got called ${callTarget.name}")
+            }
         addInstr(result, call, null)
         val resultRegs = returnRegisters.filter{it.dt!=IRDataType.FLOAT}.map{it.registerNum}
         val resultFpRegs = returnRegisters.filter{it.dt==IRDataType.FLOAT}.map{it.registerNum}
@@ -826,7 +895,7 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
             if(binExpr.left.type==DataType.STR || binExpr.right.type==DataType.STR) {
                 throw AssemblyError("str compares should have been replaced with builtin function call to do the compare")
             } else {
-                return if(constValue(binExpr.right)==0.0) {
+                return if(binExpr.right.asConstValue()==0.0) {
                     val tr = translateExpression(binExpr.left)
                     addToResult(result, tr, tr.resultReg, -1)
                     val opcode = if (notEquals) Opcode.SNZ else Opcode.SZ

codeGenIntermediate/src/prog8/codegen/intermediate/IRCodeGen.kt

@@ -23,6 +23,7 @@ class IRCodeGen(
         makeAllNodenamesScoped(program)
         moveAllNestedSubroutinesToBlockScope(program)
         verifyNameScoping(program, symbolTable)
+        changeGlobalVarInits(symbolTable)
 
         val irSymbolTable = IRSymbolTable.fromStDuringCodegen(symbolTable)
         val irProg = IRProgram(program.name, irSymbolTable, options, program.encoding)
@@ -54,12 +55,43 @@ class IRCodeGen(
         return irProg
     }
 
+    private fun changeGlobalVarInits(symbolTable: SymbolTable) {
+        // Normally, block level (global) variables that have a numeric initialization value
+        // are initialized via an assignment statement.
+        val initsToRemove = mutableListOf<Pair<PtBlock, PtAssignment>>()
+
+        symbolTable.allVariables.forEach { variable ->
+            if(variable.uninitialized && variable.parent.type==StNodeType.BLOCK) {
+                val block = variable.parent.astNode as PtBlock
+                val initialization = (block.children.firstOrNull {
+                    it is PtAssignment && it.target.identifier?.name==variable.scopedName
+                } as PtAssignment?)
+                val initValue = initialization?.value
+                when(initValue){
+                    is PtBool -> {
+                        variable.setOnetimeInitNumeric(initValue.asInt().toDouble())
+                        initsToRemove += block to initialization
+                    }
+                    is PtNumber -> {
+                        variable.setOnetimeInitNumeric(initValue.number)
+                        initsToRemove += block to initialization
+                    }
+                    is PtArray, is PtString -> throw AssemblyError("array or string initialization values should already be part of the vardecl, not a separate assignment")
+                    else -> {}
+                }
+            }
+        }
+
+        for((block, assign) in initsToRemove) {
+            block.children.remove(assign)
+        }
+    }
+
     private fun verifyNameScoping(program: PtProgram, symbolTable: SymbolTable) {
         fun verifyPtNode(node: PtNode) {
             when (node) {
                 is PtBuiltinFunctionCall -> require('.' !in node.name) { "builtin function call name should not be scoped: ${node.name}" }
                 is PtFunctionCall -> require('.' in node.name) { "node $node name is not scoped: ${node.name}" }
-                is PtIdentifier -> require('.' in node.name) { "node $node name is not scoped: ${node.name}" }
                 is PtAsmSub -> require('.' in node.name) { "node $node name is not scoped: ${node.name}" }
                 is PtBlock -> require('.' !in node.name) { "block name should not be scoped: ${node.name}" }
                 is PtConstant -> require('.' in node.name) { "node $node name is not scoped: ${node.name}" }
@@ -69,8 +101,7 @@ class IRCodeGen(
                 is PtVariable -> require('.' in node.name) { "node $node name is not scoped: ${node.name}" }
                 is PtProgram -> require('.' !in node.name) { "program name should not be scoped: ${node.name}" }
                 is PtSubroutineParameter -> require('.' in node.name) { "node $node name is not scoped: ${node.name}" }
-                else -> { /* node has no name */
-                }
+                else -> { /* node has no name or is ok to have no dots in the name */ }
             }
             node.children.forEach { verifyPtNode(it) }
         }
@@ -201,6 +232,7 @@ class IRCodeGen(
                 chunk += IRInstruction(Opcode.BREAKPOINT)
                 listOf(chunk)
             }
+            is PtAlign -> TODO("ir support for inline %align")
             is PtConditionalBranch -> translate(node)
             is PtInlineAssembly -> listOf(IRInlineAsmChunk(null, node.assembly, node.isIR, null))
             is PtIncludeBinary -> listOf(IRInlineBinaryChunk(null, readBinaryData(node), null))
@@ -221,6 +253,7 @@ class IRCodeGen(
             is PtBool,
             is PtArray,
             is PtBlock,
+            is PtDefer -> throw AssemblyError("should have been transformed")
             is PtString -> throw AssemblyError("should not occur as separate statement node ${node.position}")
             is PtSub -> throw AssemblyError("nested subroutines should have been flattened ${node.position}")
             else -> TODO("missing codegen for $node")
@@ -1295,6 +1328,9 @@ class IRCodeGen(
     }
 
     private fun translateIfElse(ifElse: PtIfElse): IRCodeChunks {
+        if((ifElse.condition as? PtPrefix)?.operator=="not" && ifElse.hasElse())
+            throw AssemblyError("not prefix in ifelse should have been replaced by swapped if-else blocks")
+
         val condition = ifElse.condition as? PtBinaryExpression
         if(condition==null || condition.left.type != DataType.FLOAT) {
             return ifWithElse_IntegerCond(ifElse)
@@ -1360,12 +1396,14 @@ class IRCodeGen(
     private fun ifWithElse_IntegerCond(ifElse: PtIfElse): List<IRCodeChunkBase> {
         val result = mutableListOf<IRCodeChunkBase>()
 
-        fun translateSimple(condition: PtExpression, jumpFalseOpcode: Opcode) {
+        fun translateSimple(condition: PtExpression, jumpFalseOpcode: Opcode, addCmpiZero: Boolean) {
 
             if(condition is PtBuiltinFunctionCall && condition.name.startsWith("prog8_ifelse_bittest_"))
                 throw AssemblyError("IR codegen doesn't have special instructions for dedicated BIT tests and should just still use normal AND")
 
             val tr = expressionEval.translateExpression(condition)
+            if(addCmpiZero)
+                tr.chunks.last().instructions.add(IRInstruction(Opcode.CMPI, tr.dt, reg1 = tr.resultReg, immediate = 0))
             result += tr.chunks
             if(ifElse.hasElse()) {
                 val elseLabel = createLabelName()
@@ -1385,7 +1423,7 @@ class IRCodeGen(
 
         fun translateBinExpr(condition: PtBinaryExpression) {
             if(condition.operator in LogicalOperators)
-                return translateSimple(condition, Opcode.BSTEQ)
+                return translateSimple(condition, Opcode.BSTEQ, false)
 
             val signed = condition.left.type in SignedDatatypes
             val elseBranchFirstReg: Int
@@ -1524,18 +1562,21 @@ class IRCodeGen(
         when(val cond=ifElse.condition) {
             is PtBool -> {
                 // normally this will be optimized away, but not with -noopt
-                translateSimple(cond, Opcode.BSTEQ)
+                translateSimple(cond, Opcode.BSTEQ, false)
             }
             is PtTypeCast -> {
                 require(cond.type==DataType.BOOL && cond.value.type in NumericDatatypes)
-                translateSimple(cond, Opcode.BSTEQ)
+                translateSimple(cond, Opcode.BSTEQ, false)
             }
-            is PtIdentifier, is PtArrayIndexer, is PtBuiltinFunctionCall, is PtFunctionCall, is PtContainmentCheck -> {
-                translateSimple(cond, Opcode.BSTEQ)
+            is PtIdentifier, is PtArrayIndexer, is PtContainmentCheck -> {
+                translateSimple(cond, Opcode.BSTEQ, false)
+            }
+            is PtBuiltinFunctionCall, is PtFunctionCall -> {
+                translateSimple(cond, Opcode.BSTEQ, true)
             }
             is PtPrefix -> {
                 require(cond.operator=="not")
-                translateSimple(cond.value, Opcode.BSTNE)
+                translateSimple(cond.value, Opcode.BSTNE, false)
             }
             is PtBinaryExpression -> {
                 translateBinExpr(cond)
@@ -1546,7 +1587,7 @@ class IRCodeGen(
     }
 
     private fun translate(repeat: PtRepeatLoop): IRCodeChunks {
-        when (constIntValue(repeat.count)) {
+        when (repeat.count.asConstInteger()) {
             0 -> return emptyList()
             1 -> return translateGroup(repeat.children)
             256 -> {
@@ -1557,18 +1598,30 @@ class IRCodeGen(
 
         val repeatLabel = createLabelName()
         val skipRepeatLabel = createLabelName()
-        val irDt = irType(repeat.count.type)
+        val constRepeats = repeat.count.asConstInteger()
         val result = mutableListOf<IRCodeChunkBase>()
-        val countTr = expressionEval.translateExpression(repeat.count)
-        addToResult(result, countTr, countTr.resultReg, -1)
-        if(constIntValue(repeat.count)==null) {
-            // check if the counter is already zero
-            addInstr(result, IRInstruction(Opcode.BSTEQ, labelSymbol = skipRepeatLabel), null)
-        }
-        result += labelFirstChunk(translateNode(repeat.statements), repeatLabel)
-        result += IRCodeChunk(null, null).also {
-            it += IRInstruction(Opcode.DEC, irDt, reg1 = countTr.resultReg)  // sets status bits
-            it += IRInstruction(Opcode.BSTNE, labelSymbol = repeatLabel)
+        if(constRepeats==65536) {
+            // make use of the word wrap around to count to 65536
+            val resultRegister = registers.nextFree()
+            addInstr(result, IRInstruction(Opcode.LOAD, IRDataType.WORD, reg1=resultRegister, immediate = 0), null)
+            result += labelFirstChunk(translateNode(repeat.statements), repeatLabel)
+            result += IRCodeChunk(null, null).also {
+                it += IRInstruction(Opcode.DEC, IRDataType.WORD, reg1 = resultRegister)  // sets status bits
+                it += IRInstruction(Opcode.BSTNE, labelSymbol = repeatLabel)
+            }
+        } else {
+            val irDt = irType(repeat.count.type)
+            val countTr = expressionEval.translateExpression(repeat.count)
+            addToResult(result, countTr, countTr.resultReg, -1)
+            if (repeat.count.asConstValue() == null) {
+                // check if the counter is already zero
+                addInstr(result, IRInstruction(Opcode.BSTEQ, labelSymbol = skipRepeatLabel), null)
+            }
+            result += labelFirstChunk(translateNode(repeat.statements), repeatLabel)
+            result += IRCodeChunk(null, null).also {
+                it += IRInstruction(Opcode.DEC, irDt, reg1 = countTr.resultReg)  // sets status bits
+                it += IRInstruction(Opcode.BSTNE, labelSymbol = repeatLabel)
+            }
         }
         result += IRCodeChunk(skipRepeatLabel, null)
         return result
@@ -1613,14 +1666,22 @@ class IRCodeGen(
             addInstr(result, IRInstruction(Opcode.RETURN), null)
         } else {
             if(value.type==DataType.FLOAT) {
-                val tr = expressionEval.translateExpression(value)
-                addToResult(result, tr, -1, tr.resultFpReg)
-                addInstr(result, IRInstruction(Opcode.RETURNR, IRDataType.FLOAT, fpReg1 = tr.resultFpReg), null)
+                if(value is PtNumber) {
+                    addInstr(result, IRInstruction(Opcode.RETURNI, IRDataType.FLOAT, immediateFp = value.number), null)
+                } else {
+                    val tr = expressionEval.translateExpression(value)
+                    addToResult(result, tr, -1, tr.resultFpReg)
+                    addInstr(result, IRInstruction(Opcode.RETURNR, IRDataType.FLOAT, fpReg1 = tr.resultFpReg), null)
+                }
             }
             else {
-                val tr = expressionEval.translateExpression(value)
-                addToResult(result, tr, tr.resultReg, -1)
-                addInstr(result, IRInstruction(Opcode.RETURNR, irType(value.type) , reg1=tr.resultReg), null)
+                if(value.asConstInteger()!=null) {
+                    addInstr(result, IRInstruction(Opcode.RETURNI, irType(value.type), immediate = value.asConstInteger()), null)
+                } else {
+                    val tr = expressionEval.translateExpression(value)
+                    addToResult(result, tr, tr.resultReg, -1)
+                    addInstr(result, IRInstruction(Opcode.RETURNR, irType(value.type), reg1 = tr.resultReg), null)
+                }
             }
         }
         return result
@@ -1633,14 +1694,14 @@ class IRCodeGen(
                 block.options.forceOutput,
                 block.options.noSymbolPrefixing,
                 block.options.veraFxMuls,
-                block.options.ignoreUnused,
-                translate(block.options.alignment)
+                block.options.ignoreUnused
             ), block.position)
         for (child in block.children) {
             when(child) {
                 is PtNop -> { /* nothing */ }
                 is PtAssignment, is PtAugmentedAssign -> { /* global variable initialization is done elsewhere */ }
                 is PtVariable, is PtConstant, is PtMemMapped -> { /* vars should be looked up via symbol table */ }
+                is PtAlign -> TODO("ir support for inline %align")
                 is PtSub -> {
                     val sub = IRSubroutine(child.name, translate(child.parameters), child.returntype, child.position)
                     for (subchild in child.children) {
@@ -1650,9 +1711,9 @@ class IRCodeGen(
                 }
                 is PtAsmSub -> {
                     if(child.address!=null) {
-                        // romsub. No codegen needed: calls to this are jumping straight to the address.
+                        // extmsub. No codegen needed: calls to this are jumping straight to the address.
                         require(child.children.isEmpty()) {
-                            "romsub should be empty at ${child.position}"
+                            "extsub should be empty at ${child.position}"
                         }
                     } else {
                         // regular asmsub
@@ -1665,7 +1726,7 @@ class IRCodeGen(
                         )
                         irBlock += IRAsmSubroutine(
                             child.name,
-                            child.address,
+                            null,
                             child.clobbers,
                             child.parameters.map { IRAsmSubroutine.IRAsmParam(it.first, it.second.type) },        // note: the name of the asmsub param is not used here anymore
                             child.returns.map { IRAsmSubroutine.IRAsmParam(it.first, it.second)},
@@ -1698,18 +1759,10 @@ class IRCodeGen(
             IRSubroutine.IRParam(flattenedName, orig.dt)
         }
 
-    private fun translate(alignment: PtBlock.BlockAlignment): IRBlock.BlockAlignment {
-        return when(alignment) {
-            PtBlock.BlockAlignment.NONE -> IRBlock.BlockAlignment.NONE
-            PtBlock.BlockAlignment.WORD -> IRBlock.BlockAlignment.WORD
-            PtBlock.BlockAlignment.PAGE -> IRBlock.BlockAlignment.PAGE
-        }
-    }
-
     private var labelSequenceNumber = 0
     internal fun createLabelName(): String {
         labelSequenceNumber++
-        return "label_gen_$labelSequenceNumber"
+        return "${PtLabel.GeneratedLabelPrefix}$labelSequenceNumber"
     }
 
     internal fun translateBuiltinFunc(call: PtBuiltinFunctionCall): ExpressionCodeResult

codeGenIntermediate/src/prog8/codegen/intermediate/IRPeepholeOptimizer.kt

@@ -64,7 +64,7 @@ class IRPeepholeOptimizer(private val irprog: IRProgram) {
     private fun replaceConcatZeroMsbWithExt(chunk: IRCodeChunk, indexedInstructions: List<IndexedValue<IRInstruction>>): Boolean {
         var changed = false
         indexedInstructions.reversed().forEach { (idx, ins) ->
-            if (ins.opcode == Opcode.CONCAT) {
+            if (ins.opcode == Opcode.CONCAT && idx>0) {
                 // if the previous instruction loads a zero in the msb, this can be turned into EXT.B instead
                 val prev = indexedInstructions[idx-1].value
                 if(prev.opcode==Opcode.LOAD && prev.immediate==0 && prev.reg1==ins.reg2) {
@@ -299,7 +299,7 @@ class IRPeepholeOptimizer(private val irprog: IRProgram) {
             }
 
             // remove useless RETURN
-            if(idx>0 && (ins.opcode == Opcode.RETURN || ins.opcode==Opcode.RETURNR)) {
+            if(idx>0 && (ins.opcode == Opcode.RETURN || ins.opcode==Opcode.RETURNR || ins.opcode==Opcode.RETURNI)) {
                 val previous = chunk.instructions[idx-1]
                 if(previous.opcode in OpcodesThatJump) {
                     chunk.instructions.removeAt(idx)
@@ -449,7 +449,7 @@ class IRPeepholeOptimizer(private val irprog: IRProgram) {
     private fun removeDoubleLoadsAndStores(chunk: IRCodeChunk, indexedInstructions: List<IndexedValue<IRInstruction>>): Boolean {
         var changed = false
         indexedInstructions.forEach { (idx, ins) ->
-            if(ins.opcode==Opcode.STOREM) {
+            if(ins.opcode==Opcode.STOREM && idx>0) {
                 val prev = indexedInstructions[idx-1].value
                 if(prev.opcode==Opcode.LOADM) {
                     // loadm.X rX,something | storem.X rX,something ?? -> get rid of the store.

codeGenIntermediate/test/TestIRPeepholeOpt.kt

@@ -22,7 +22,8 @@ class TestIRPeepholeOpt: FunSpec({
             floats = false,
             noSysInit = true,
             compTarget = target,
-            loadAddress = target.machine.PROGRAM_LOAD_ADDRESS
+            loadAddress = target.machine.PROGRAM_LOAD_ADDRESS,
+            memtopAddress = 0xffffu
         )
         val prog = IRProgram("test", IRSymbolTable(), options, target)
         prog.addBlock(block)

codeGenIntermediate/test/TestVmCodeGen.kt

@@ -23,7 +23,8 @@ class TestVmCodeGen: FunSpec({
             floats = true,
             noSysInit = false,
             compTarget = target,
-            loadAddress = target.machine.PROGRAM_LOAD_ADDRESS
+            loadAddress = target.machine.PROGRAM_LOAD_ADDRESS,
+            memtopAddress = 0xffffu
         )
     }
 
@@ -44,10 +45,42 @@ class TestVmCodeGen: FunSpec({
         val program = PtProgram("test", DummyMemsizer, DummyStringEncoder)
         val block = PtBlock("main", false, SourceCode.Generated("test"), PtBlock.Options(), Position.DUMMY)
         val sub = PtSub("start", emptyList(), null, Position.DUMMY)
-        sub.add(PtVariable("pi", DataType.UBYTE, ZeropageWish.DONTCARE, PtNumber(DataType.UBYTE, 0.0, Position.DUMMY), null, Position.DUMMY))
-        sub.add(PtVariable("particleX", DataType.ARRAY_UB, ZeropageWish.DONTCARE, null, 3u, Position.DUMMY))
-        sub.add(PtVariable("particleDX", DataType.ARRAY_UB, ZeropageWish.DONTCARE, null, 3u, Position.DUMMY))
-        sub.add(PtVariable("xx", DataType.WORD, ZeropageWish.DONTCARE, PtNumber(DataType.WORD, 1.0, Position.DUMMY), null, Position.DUMMY))
+        sub.add(PtVariable(
+            "pi",
+            DataType.UBYTE,
+            ZeropageWish.DONTCARE,
+            0u,
+            PtNumber(DataType.UBYTE, 0.0, Position.DUMMY),
+            null,
+            Position.DUMMY
+        ))
+        sub.add(PtVariable(
+            "particleX",
+            DataType.ARRAY_UB,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            3u,
+            Position.DUMMY
+        ))
+        sub.add(PtVariable(
+            "particleDX",
+            DataType.ARRAY_UB,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            3u,
+            Position.DUMMY
+        ))
+        sub.add(PtVariable(
+            "xx",
+            DataType.WORD,
+            ZeropageWish.DONTCARE,
+            0u,
+            PtNumber(DataType.WORD, 1.0, Position.DUMMY),
+            null,
+            Position.DUMMY
+        ))
 
         val assign = PtAugmentedAssign("+=", Position.DUMMY)
         val target = PtAssignTarget(false, Position.DUMMY).also {
@@ -123,7 +156,15 @@ class TestVmCodeGen: FunSpec({
         val program = PtProgram("test", DummyMemsizer, DummyStringEncoder)
         val block = PtBlock("main", false, SourceCode.Generated("test"), PtBlock.Options(), Position.DUMMY)
         val sub = PtSub("start", emptyList(), null, Position.DUMMY)
-        sub.add(PtVariable("f1", DataType.FLOAT, ZeropageWish.DONTCARE, null, null, Position.DUMMY))
+        sub.add(PtVariable(
+            "f1",
+            DataType.FLOAT,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            null,
+            Position.DUMMY
+        ))
         val if1 = PtIfElse(Position.DUMMY)
         val cmp1 = PtBinaryExpression("==", DataType.BOOL, Position.DUMMY)
         cmp1.add(PtIdentifier("main.start.f1", DataType.FLOAT, Position.DUMMY))
@@ -186,7 +227,15 @@ class TestVmCodeGen: FunSpec({
         val program = PtProgram("test", DummyMemsizer, DummyStringEncoder)
         val block = PtBlock("main", false, SourceCode.Generated("test"), PtBlock.Options(), Position.DUMMY)
         val sub = PtSub("start", emptyList(), null, Position.DUMMY)
-        sub.add(PtVariable("f1", DataType.FLOAT, ZeropageWish.DONTCARE, null, null, Position.DUMMY))
+        sub.add(PtVariable(
+            "f1",
+            DataType.FLOAT,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            null,
+            Position.DUMMY
+        ))
         val if1 = PtIfElse(Position.DUMMY)
         val cmp1 = PtBinaryExpression("==", DataType.BOOL, Position.DUMMY)
         cmp1.add(PtIdentifier("main.start.f1", DataType.FLOAT, Position.DUMMY))
@@ -245,7 +294,15 @@ class TestVmCodeGen: FunSpec({
         val program = PtProgram("test", DummyMemsizer, DummyStringEncoder)
         val block = PtBlock("main", false, SourceCode.Generated("test"), PtBlock.Options(), Position.DUMMY)
         val sub = PtSub("start", emptyList(), null, Position.DUMMY)
-        sub.add(PtVariable("f1", DataType.FLOAT, ZeropageWish.DONTCARE, null, null, Position.DUMMY))
+        sub.add(PtVariable(
+            "f1",
+            DataType.FLOAT,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            null,
+            Position.DUMMY
+        ))
         val if1 = PtIfElse(Position.DUMMY)
         val cmp1 = PtBinaryExpression("==", DataType.BOOL, Position.DUMMY)
         cmp1.add(PtIdentifier("main.start.f1", DataType.FLOAT, Position.DUMMY))
@@ -292,7 +349,15 @@ class TestVmCodeGen: FunSpec({
         val program = PtProgram("test", DummyMemsizer, DummyStringEncoder)
         val block = PtBlock("main", false, SourceCode.Generated("test"), PtBlock.Options(), Position.DUMMY)
         val sub = PtSub("start", emptyList(), null, Position.DUMMY)
-        sub.add(PtVariable("sb1", DataType.BYTE, ZeropageWish.DONTCARE, null, null, Position.DUMMY))
+        sub.add(PtVariable(
+            "sb1",
+            DataType.BYTE,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            null,
+            Position.DUMMY
+        ))
         val if1 = PtIfElse(Position.DUMMY)
         val cmp1 = PtBinaryExpression("==", DataType.BOOL, Position.DUMMY)
         cmp1.add(PtIdentifier("main.start.sb1", DataType.BYTE, Position.DUMMY))
@@ -355,7 +420,15 @@ class TestVmCodeGen: FunSpec({
         val program = PtProgram("test", DummyMemsizer, DummyStringEncoder)
         val block = PtBlock("main", false, SourceCode.Generated("test"), PtBlock.Options(), Position.DUMMY)
         val sub = PtSub("start", emptyList(), null, Position.DUMMY)
-        sub.add(PtVariable("sb1", DataType.BYTE, ZeropageWish.DONTCARE, null, null, Position.DUMMY))
+        sub.add(PtVariable(
+            "sb1",
+            DataType.BYTE,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            null,
+            Position.DUMMY
+        ))
         val if1 = PtIfElse(Position.DUMMY)
         val cmp1 = PtBinaryExpression("==", DataType.BOOL, Position.DUMMY)
         cmp1.add(PtIdentifier("main.start.sb1", DataType.BYTE, Position.DUMMY))
@@ -414,7 +487,15 @@ class TestVmCodeGen: FunSpec({
         val program = PtProgram("test", DummyMemsizer, DummyStringEncoder)
         val block = PtBlock("main", false, SourceCode.Generated("test"), PtBlock.Options(), Position.DUMMY)
         val sub = PtSub("start", emptyList(), null, Position.DUMMY)
-        sub.add(PtVariable("ub1", DataType.UBYTE, ZeropageWish.DONTCARE, null, null, Position.DUMMY))
+        sub.add(PtVariable(
+            "ub1",
+            DataType.UBYTE,
+            ZeropageWish.DONTCARE,
+            0u,
+            null,
+            null,
+            Position.DUMMY
+        ))
         val if1 = PtIfElse(Position.DUMMY)
         val cmp1 = PtBinaryExpression("==", DataType.BOOL, Position.DUMMY)
         cmp1.add(PtIdentifier("main.start.ub1", DataType.UBYTE, Position.DUMMY))
@@ -442,9 +523,9 @@ class TestVmCodeGen: FunSpec({
         irChunks.size shouldBe 1
     }
 
-    test("romsub allowed in ir-codegen") {
+    test("extsub allowed in ir-codegen") {
 //main {
-//    romsub $5000 = routine()
+//    extsub $5000 = routine()
 //
 //    sub start()  {
 //        routine()
@@ -453,8 +534,8 @@ class TestVmCodeGen: FunSpec({
         val codegen = VmCodeGen()
         val program = PtProgram("test", DummyMemsizer, DummyStringEncoder)
         val block = PtBlock("main", false, SourceCode.Generated("test"), PtBlock.Options(), Position.DUMMY)
-        val romsub = PtAsmSub("routine", 0x5000u, setOf(CpuRegister.Y), emptyList(), emptyList(), false, Position.DUMMY)
-        block.add(romsub)
+        val extsub = PtAsmSub("routine", PtAsmSub.Address(null, null, 0x5000u), setOf(CpuRegister.Y), emptyList(), emptyList(), false, Position.DUMMY)
+        block.add(extsub)
         val sub = PtSub("start", emptyList(), null, Position.DUMMY)
         val call = PtFunctionCall("main.routine", true, DataType.UNDEFINED, Position.DUMMY)
         sub.add(call)

codeOptimizers/src/prog8/optimizer/ConstantFoldingOptimizer.kt

@@ -2,6 +2,7 @@ package prog8.optimizer
 
 import prog8.ast.Node
 import prog8.ast.Program
+import prog8.ast.base.FatalAstException
 import prog8.ast.expressions.*
 import prog8.ast.maySwapOperandOrder
 import prog8.ast.statements.*
@@ -112,6 +113,28 @@ class ConstantFoldingOptimizer(private val program: Program, private val errors:
             }
         }
 
+        if(expr.left.inferType(program).isArray) {
+            if (expr.operator=="*" && rightconst!=null) {
+                if (expr.left is ArrayLiteral) {
+                    // concatenate array literal.
+                    val part = expr.left as ArrayLiteral
+                    if(part.value.isEmpty())
+                        errors.warn("resulting array has length zero", part.position)
+                    val tmp = mutableListOf<Expression>()
+                    repeat(rightconst.number.toInt()) {
+                        part.value.forEach { tmp += it.copy() }
+                    }
+                    val newArray = ArrayLiteral(part.type, tmp.toTypedArray(), part.position)
+                    return listOf(IAstModification.ReplaceNode(expr, newArray, parent))
+                }
+                else {
+                    val leftTarget = (expr.left as? IdentifierReference)?.targetVarDecl(program)
+                    if(leftTarget!=null && leftTarget.origin==VarDeclOrigin.ARRAYLITERAL)
+                        throw FatalAstException("shouldn't see an array literal converted to an autovar here")
+                }
+            }
+        }
+
         if(expr.operator=="==" && rightconst!=null) {
             val leftExpr = expr.left as? BinaryExpression
             // only do this shuffling when the LHS is not a constant itself (otherwise problematic nested replacements)
@@ -298,6 +321,31 @@ class ConstantFoldingOptimizer(private val program: Program, private val errors:
         return noModifications
     }
 
+    override fun after(arrayIndexedExpression: ArrayIndexedExpression, parent: Node): Iterable<IAstModification> {
+        if(parent is VarDecl && parent.parent is Block) {
+            // only block level (global) initializers are considered here, because they're run just once at program startup
+
+            val const = arrayIndexedExpression.constValue(program)
+            if (const != null)
+                return listOf(IAstModification.ReplaceNode(arrayIndexedExpression, const, parent))
+
+            val constIndex = arrayIndexedExpression.indexer.constIndex()
+            if (constIndex != null) {
+                val arrayVar = arrayIndexedExpression.arrayvar.targetVarDecl(program)
+                if(arrayVar!=null) {
+                    val array =arrayVar.value as? ArrayLiteral
+                    if(array!=null) {
+                        val value = array.value[constIndex].constValue(program)
+                        if(value!=null) {
+                            return listOf(IAstModification.ReplaceNode(arrayIndexedExpression, value, parent))
+                        }
+                    }
+                }
+            }
+        }
+        return noModifications
+    }
+
     override fun after(functionCallExpr: FunctionCallExpression, parent: Node): Iterable<IAstModification> {
         val constvalue = functionCallExpr.constValue(program)
         return if(constvalue!=null)

codeOptimizers/src/prog8/optimizer/ConstantIdentifierReplacer.kt

@@ -76,7 +76,7 @@ class VarConstantValueTypeAdjuster(
                     if (declValue != null) {
                         // variable is never written to, so it can be replaced with a constant, IF the value is a constant
                         errors.info("variable '${decl.name}' is never written to and was replaced by a constant", decl.position)
-                        val const = VarDecl(VarDeclType.CONST, decl.origin, decl.datatype, decl.zeropage, decl.arraysize, decl.name, decl.names, declValue, decl.sharedWithAsm, decl.splitArray, decl.position)
+                        val const = VarDecl(VarDeclType.CONST, decl.origin, decl.datatype, decl.zeropage, decl.arraysize, decl.name, decl.names, declValue, decl.sharedWithAsm, decl.splitArray, decl.alignment, decl.dirty, decl.position)
                         decl.value = null
                         return listOf(
                             IAstModification.ReplaceNode(decl, const, parent)
@@ -96,7 +96,7 @@ class VarConstantValueTypeAdjuster(
                     }
                     // variable only has a single write and it is the initialization value, so it can be replaced with a constant, IF the value is a constant
                     errors.info("variable '${decl.name}' is never written to and was replaced by a constant", decl.position)
-                    val const = VarDecl(VarDeclType.CONST, decl.origin, decl.datatype, decl.zeropage, decl.arraysize, decl.name, decl.names, singleAssignment.value, decl.sharedWithAsm, decl.splitArray, decl.position)
+                    val const = VarDecl(VarDeclType.CONST, decl.origin, decl.datatype, decl.zeropage, decl.arraysize, decl.name, decl.names, singleAssignment.value, decl.sharedWithAsm, decl.splitArray, decl.alignment, decl.dirty, decl.position)
                     return listOf(
                         IAstModification.ReplaceNode(decl, const, parent),
                         IAstModification.Remove(singleAssignment, singleAssignment.parent as IStatementContainer)
@@ -268,11 +268,9 @@ class VarConstantValueTypeAdjuster(
 
 
 // Replace all constant identifiers with their actual value,
-// and the array var initializer values and sizes.
 // This is needed because further constant optimizations depend on those.
 internal class ConstantIdentifierReplacer(
     private val program: Program,
-    private val options: CompilationOptions,
     private val errors: IErrorReporter
 ) : AstWalker() {
 
@@ -287,7 +285,7 @@ internal class ConstantIdentifierReplacer(
     override fun after(identifier: IdentifierReference, parent: Node): Iterable<IAstModification> {
         // replace identifiers that refer to const value, with the value itself
         // if it's a simple type and if it's not a left hand side variable
-        if(identifier.parent is AssignTarget)
+        if(identifier.parent is AssignTarget || identifier.parent is Alias)
             return noModifications
         var forloop = identifier.parent as? ForLoop
         if(forloop==null)
@@ -400,7 +398,7 @@ internal class ConstantIdentifierReplacer(
             if(targetDatatype.isArray) {
                 val decl = VarDecl(VarDeclType.VAR, VarDeclOrigin.ARRAYLITERAL, targetDatatype.getOr(DataType.UNDEFINED),
                     ZeropageWish.DONTCARE, null, "dummy", emptyList(),
-                    assignment.value, false, false, Position.DUMMY)
+                    assignment.value, false, false, 0u, false, Position.DUMMY)
                 val replaceValue = createConstArrayInitializerValue(decl)
                 if(replaceValue!=null) {
                     return listOf(IAstModification.ReplaceNode(assignment.value, replaceValue, assignment))
@@ -421,104 +419,41 @@ internal class ConstantIdentifierReplacer(
             return null
         }
 
-        // convert the initializer range expression from a range or int, to an actual array.
+        val rangeExpr = decl.value as? RangeExpression ?: return null
+
+        // convert the initializer range expression from a range, to an actual array literal.
+        val declArraySize = decl.arraysize?.constIndex()
+        val constRange = rangeExpr.toConstantIntegerRange()
+        if(constRange?.isEmpty()==true) {
+            if(constRange.first>constRange.last && constRange.step>=0)
+                errors.err("descending range with positive step", decl.value?.position!!)
+            else if(constRange.first<constRange.last && constRange.step<=0)
+                errors.err("ascending range with negative step", decl.value?.position!!)
+        }
         when(decl.datatype) {
             DataType.ARRAY_UB, DataType.ARRAY_B, DataType.ARRAY_UW, DataType.ARRAY_W, DataType.ARRAY_W_SPLIT, DataType.ARRAY_UW_SPLIT -> {
-                val rangeExpr = decl.value as? RangeExpression
-                if(rangeExpr!=null) {
-                    val constRange = rangeExpr.toConstantIntegerRange()
-                    if(constRange?.isEmpty()==true) {
-                        if(constRange.first>constRange.last && constRange.step>=0)
-                            errors.err("descending range with positive step", decl.value?.position!!)
-                        else if(constRange.first<constRange.last && constRange.step<=0)
-                            errors.err("ascending range with negative step", decl.value?.position!!)
-                    }
-                    val declArraySize = decl.arraysize?.constIndex()
-                    if(declArraySize!=null && declArraySize!=rangeExpr.size())
-                        errors.err("range expression size (${rangeExpr.size()}) doesn't match declared array size ($declArraySize)", decl.value?.position!!)
-                    if(constRange!=null) {
-                        val eltType = rangeExpr.inferType(program).getOr(DataType.UBYTE)
-                        return if(eltType in ByteDatatypes) {
-                            ArrayLiteral(InferredTypes.InferredType.known(decl.datatype),
-                                constRange.map { NumericLiteral(eltType, it.toDouble(), decl.value!!.position) }.toTypedArray(),
-                                position = decl.value!!.position)
-                        } else {
-                            ArrayLiteral(InferredTypes.InferredType.known(decl.datatype),
-                                constRange.map { NumericLiteral(eltType, it.toDouble(), decl.value!!.position) }.toTypedArray(),
-                                position = decl.value!!.position)
-                        }
-                    }
-                }
-                val numericLv = decl.value as? NumericLiteral
-                if(numericLv!=null && numericLv.type== DataType.FLOAT)
-                    errors.err("arraysize requires only integers here", numericLv.position)
-                val size = decl.arraysize?.constIndex() ?: return null
-                if (rangeExpr==null && numericLv!=null) {
-                    // arraysize initializer is empty or a single int, and we know the size; create the arraysize.
-                    val fillvalue = numericLv.number.toInt()
-                    when(decl.datatype){
-                        DataType.ARRAY_UB -> {
-                            if(fillvalue !in 0..255)
-                                errors.err("ubyte value overflow", numericLv.position)
-                        }
-                        DataType.ARRAY_B -> {
-                            if(fillvalue !in -128..127)
-                                errors.err("byte value overflow", numericLv.position)
-                        }
-                        DataType.ARRAY_UW -> {
-                            if(fillvalue !in 0..65535)
-                                errors.err("uword value overflow", numericLv.position)
-                        }
-                        DataType.ARRAY_W -> {
-                            if(fillvalue !in -32768..32767)
-                                errors.err("word value overflow", numericLv.position)
-                        }
-                        else -> {}
-                    }
-                    // create the array itself, filled with the fillvalue.
-                    val array = Array(size) {fillvalue}.map { NumericLiteral(ArrayToElementTypes.getValue(decl.datatype), it.toDouble(), numericLv.position) }.toTypedArray<Expression>()
-                    return ArrayLiteral(InferredTypes.InferredType.known(decl.datatype), array, position = numericLv.position)
-                }
-            }
-            DataType.ARRAY_F -> {
-                val rangeExpr = decl.value as? RangeExpression
-                if(rangeExpr!=null) {
-                    // convert the initializer range expression to an actual array of floats
-                    val declArraySize = decl.arraysize?.constIndex()
-                    if(declArraySize!=null && declArraySize!=rangeExpr.size())
-                        errors.err("range expression size (${rangeExpr.size()}) doesn't match declared array size ($declArraySize)", decl.value?.position!!)
-                    val constRange = rangeExpr.toConstantIntegerRange()
-                    if(constRange!=null) {
-                        return ArrayLiteral(InferredTypes.InferredType.known(DataType.ARRAY_F),
-                            constRange.map { NumericLiteral(DataType.FLOAT, it.toDouble(), decl.value!!.position) }.toTypedArray(),
+                if(declArraySize!=null && declArraySize!=rangeExpr.size())
+                    errors.err("range expression size (${rangeExpr.size()}) doesn't match declared array size ($declArraySize)", decl.value?.position!!)
+                if(constRange!=null) {
+                    val eltType = rangeExpr.inferType(program).getOr(DataType.UBYTE)
+                    return if(eltType in ByteDatatypes) {
+                        ArrayLiteral(InferredTypes.InferredType.known(decl.datatype),
+                            constRange.map { NumericLiteral(eltType, it.toDouble(), decl.value!!.position) }.toTypedArray(),
+                            position = decl.value!!.position)
+                    } else {
+                        ArrayLiteral(InferredTypes.InferredType.known(decl.datatype),
+                            constRange.map { NumericLiteral(eltType, it.toDouble(), decl.value!!.position) }.toTypedArray(),
                             position = decl.value!!.position)
                     }
                 }
-
-                val numericLv = decl.value as? NumericLiteral
-                val size = decl.arraysize?.constIndex() ?: return null
-                if(rangeExpr==null && numericLv!=null) {
-                    // arraysize initializer is a single int, and we know the array size.
-                    val fillvalue = numericLv.number
-                    if (fillvalue < options.compTarget.machine.FLOAT_MAX_NEGATIVE || fillvalue > options.compTarget.machine.FLOAT_MAX_POSITIVE)
-                        errors.err("float value overflow", numericLv.position)
-                    else {
-                        val array = Array(size) {fillvalue}.map { NumericLiteral(DataType.FLOAT, it, numericLv.position) }.toTypedArray<Expression>()
-                        return ArrayLiteral(InferredTypes.InferredType.known(DataType.ARRAY_F), array, position = numericLv.position)
-                    }
-                }
             }
-            DataType.ARRAY_BOOL -> {
-                val numericLv = decl.value as? NumericLiteral
-                val size = decl.arraysize?.constIndex() ?: return null
-                if(numericLv!=null) {
-                    // arraysize initializer is a single value, and we know the array size.
-                    if(numericLv.type!=DataType.BOOL) {
-                        errors.err("initializer value is not a boolean", numericLv.position)
-                        return null
-                    }
-                    val array = Array(size) {numericLv.number}.map { NumericLiteral(DataType.BOOL, it, numericLv.position) }.toTypedArray<Expression>()
-                    return ArrayLiteral(InferredTypes.InferredType.known(DataType.ARRAY_BOOL), array, position = numericLv.position)
+            DataType.ARRAY_F -> {
+                if(declArraySize!=null && declArraySize!=rangeExpr.size())
+                    errors.err("range expression size (${rangeExpr.size()}) doesn't match declared array size ($declArraySize)", decl.value?.position!!)
+                if(constRange!=null) {
+                    return ArrayLiteral(InferredTypes.InferredType.known(DataType.ARRAY_F),
+                        constRange.map { NumericLiteral(DataType.FLOAT, it.toDouble(), decl.value!!.position) }.toTypedArray(),
+                        position = decl.value!!.position)
                 }
             }
             else -> return null

codeOptimizers/src/prog8/optimizer/ExpressionSimplifier.kt

@@ -238,11 +238,8 @@ class ExpressionSimplifier(private val program: Program, private val errors: IEr
         if(expr.operator=="==") {
             if(rightDt==DataType.BOOL && leftDt==DataType.BOOL) {
                 val rightConstBool = rightVal?.asBooleanValue
-                if(rightConstBool!=null) {
-                    return if(rightConstBool)
-                        listOf(IAstModification.ReplaceNode(expr, expr.left, parent))
-                    else
-                        listOf(IAstModification.ReplaceNode(expr, PrefixExpression("not", expr.left, expr.position), parent))
+                if(rightConstBool==true) {
+                    return listOf(IAstModification.ReplaceNode(expr, expr.left, parent))
                 }
             }
             if (rightVal?.number == 1.0) {
@@ -261,11 +258,8 @@ class ExpressionSimplifier(private val program: Program, private val errors: IEr
         if (expr.operator=="!=") {
             if(rightDt==DataType.BOOL && leftDt==DataType.BOOL) {
                 val rightConstBool = rightVal?.asBooleanValue
-                if(rightConstBool!=null) {
-                    return if (rightConstBool)
-                        listOf(IAstModification.ReplaceNode(expr, PrefixExpression("not", expr.left, expr.position), parent))
-                    else
-                        listOf(IAstModification.ReplaceNode(expr, expr.left, parent))
+                if(rightConstBool==false) {
+                    listOf(IAstModification.ReplaceNode(expr, expr.left, parent))
                 }
             }
             if (rightVal?.number == 1.0) {
@@ -416,21 +410,6 @@ class ExpressionSimplifier(private val program: Program, private val errors: IEr
         return noModifications
     }
 
-    override fun after(expr: PrefixExpression, parent: Node): Iterable<IAstModification> {
-        if(expr.operator=="not") {
-            // not  X <compare> Y  ->   X <invertedcompare> Y
-            val binExpr = expr.expression as? BinaryExpression
-            if(binExpr!=null) {
-                val invertedOperator = invertedComparisonOperator(binExpr.operator)
-                if(invertedOperator!=null) {
-                    val inverted = BinaryExpression(binExpr.left, invertedOperator, binExpr.right, binExpr.position)
-                    return listOf(IAstModification.ReplaceNode(expr, inverted, parent))
-                }
-            }
-        }
-        return noModifications
-    }
-
     private fun applyAbsorptionLaws(expr: BinaryExpression): Expression? {
         // NOTE: only when the terms are not function calls!!!
         if(expr.left is IFunctionCall || expr.right is IFunctionCall)

codeOptimizers/src/prog8/optimizer/Extensions.kt

@@ -12,7 +12,7 @@ fun Program.constantFold(errors: IErrorReporter, options: CompilationOptions) {
     if(errors.noErrors()) {
         valuetypefixer.applyModifications()
 
-        val replacer = ConstantIdentifierReplacer(this, options, errors)
+        val replacer = ConstantIdentifierReplacer(this, errors)
         replacer.visit(this)
         if (errors.noErrors()) {
             replacer.applyModifications()

codeOptimizers/src/prog8/optimizer/StatementOptimizer.kt

@@ -127,19 +127,6 @@ class StatementOptimizer(private val program: Program,
                     )
                 }
             }
-
-            // switch if/else around if the condition is a not
-            val prefixCond = ifElse.condition as? PrefixExpression
-            if(prefixCond?.operator=="not") {
-                errors.info("invert conditon and swap if/else blocks", ifElse.condition.position)
-                val newTruePart = AnonymousScope(ifElse.elsepart.statements, ifElse.elsepart.position)
-                val newElsePart = AnonymousScope(ifElse.truepart.statements, ifElse.truepart.position)
-                return listOf(
-                    IAstModification.ReplaceNode(ifElse.elsepart, newElsePart, ifElse),
-                    IAstModification.ReplaceNode(ifElse.truepart, newTruePart, ifElse),
-                    IAstModification.ReplaceNode(ifElse.condition, prefixCond.expression, ifElse)
-                )
-            }
         }
 
         return noModifications

codeOptimizers/src/prog8/optimizer/UnusedCodeRemover.kt

@@ -21,6 +21,31 @@ class UnusedCodeRemover(private val program: Program,
 ): AstWalker() {
 
     private lateinit var callgraph: CallGraph
+    private val neverRemoveSubroutines = mutableListOf<Subroutine>()
+
+    init {
+        neverRemoveSubroutines.add(program.entrypoint)
+
+        program.allBlocks.singleOrNull { it.name=="sys" } ?.let {
+            val subroutines = it.statements.filterIsInstance<Subroutine>()
+            val push = subroutines.single { it.name == "push" }
+            val pushw = subroutines.single { it.name == "pushw" }
+            val pop = subroutines.single { it.name == "pop" }
+            val popw = subroutines.single { it.name == "popw" }
+            neverRemoveSubroutines.add(push)
+            neverRemoveSubroutines.add(pushw)
+            neverRemoveSubroutines.add(pop)
+            neverRemoveSubroutines.add(popw)
+        }
+
+        program.allBlocks.singleOrNull { it.name=="floats" } ?.let {
+            val subroutines = it.statements.filterIsInstance<Subroutine>()
+            val push = subroutines.single { it.name == "push" }
+            val pop = subroutines.single { it.name == "pop" }
+            neverRemoveSubroutines.add(push)
+            neverRemoveSubroutines.add(pop)
+        }
+    }
 
     override fun before(program: Program): Iterable<IAstModification> {
         callgraph = CallGraph(program)
@@ -92,7 +117,7 @@ class UnusedCodeRemover(private val program: Program,
 
     override fun after(subroutine: Subroutine, parent: Node): Iterable<IAstModification> {
         val forceOutput = "force_output" in subroutine.definingBlock.options()
-        if (subroutine !== program.entrypoint && !forceOutput && !subroutine.isAsmSubroutine) {
+        if (subroutine !in neverRemoveSubroutines && !forceOutput && !subroutine.isAsmSubroutine) {
             if(callgraph.unused(subroutine)) {
                 if(subroutine.containsNoCodeNorVars) {
                     if("ignore_unused" !in subroutine.definingBlock.options())

compiler/build.gradle

@@ -32,6 +32,7 @@ dependencies {
     implementation project(':codeGenIntermediate')
     implementation project(':codeGenExperimental')
     implementation project(':virtualmachine')
+    // implementation project(':beanshell')
     implementation "org.antlr:antlr4-runtime:4.13.2"
     implementation "org.jetbrains.kotlin:kotlin-stdlib-jdk8"
     // implementation "org.jetbrains.kotlin:kotlin-reflect"
@@ -80,11 +81,11 @@ startScripts.enabled = true
 
 application {
     mainClass = 'prog8.CompilerMainKt'
-    applicationName = 'p8compile'
+    applicationName = 'prog8c'
 }
 
 shadowJar {
-    archiveBaseName = 'prog8compiler'
+    archiveBaseName = 'prog8c'
     archiveVersion = version
     // minimize()
 }

compiler/res/prog8lib/atari/syslib.p8

@@ -12,16 +12,16 @@ atari {
         &uword COLCRS = 85
         &ubyte ROWCRS = 84
 
-    romsub $F24A = getchar() -> ubyte @A
-    romsub $F2B0 = outchar(ubyte character @ A)
-    romsub $F2FD = waitkey() -> ubyte @A
+    extsub $F24A = getchar() -> ubyte @A
+    extsub $F2B0 = outchar(ubyte character @ A)
+    extsub $F2FD = waitkey() -> ubyte @A
 
 }
 
 sys {
     ; ------- lowlevel system routines --------
 
-    const ubyte target = 8         ;  compilation target specifier.  64 = C64, 128 = C128,  16 = CommanderX16, 8 = atari800XL
+    const ubyte target = 8         ;  compilation target specifier.  255=virtual, 128=C128, 64=C64, 32=PET, 16=CommanderX16, 8=atari800XL, 7=Neo6502
 
     const ubyte sizeof_bool = 1
     const ubyte sizeof_byte = 1
@@ -30,43 +30,6 @@ sys {
     const ubyte sizeof_uword = 2
     const ubyte sizeof_float = 0    ; undefined, no float support
 
-    asmsub  init_system()  {
-        ; Initializes the machine to a sane starting state.
-        ; Called automatically by the loader program logic.
-        ; TODO
-        %asm {{
-            sei
-            ; TODO reset screen mode etc etc
-            cli
-            rts
-        }}
-    }
-
-    asmsub  init_system_phase2()  {
-        %asm {{
-            cld
-            clc
-            clv
-            rts
-        }}
-    }
-
-    asmsub  cleanup_at_exit() {
-        ; executed when the main subroutine does rts
-        %asm {{
-_exitcodeCarry = *+1
-            lda  #0
-            lsr  a
-_exitcode = *+1
-            lda  #0        ; exit code possibly modified in exit()
-_exitcodeX = *+1
-            ldx  #0
-_exitcodeY = *+1
-            ldy  #0
-            rts
-        }}
-    }
-
     asmsub  reset_system()  {
         ; Soft-reset the system back to initial power-on Basic prompt.
         ; TODO
@@ -177,6 +140,65 @@ _longcopy
         }}
     }
 
+    asmsub memcmp(uword address1 @R0, uword address2 @R1, uword size @AY) -> byte @A {
+        ; Compares two blocks of memory
+        ; Returns -1 (255), 0 or 1, meaning: block 1 sorts before, equal or after block 2.
+        %asm {{
+            sta  P8ZP_SCRATCH_REG   ; lsb(size)
+            sty  P8ZP_SCRATCH_B1    ; msb(size)
+            lda  cx16.r0
+            ldy  cx16.r0+1
+            sta  P8ZP_SCRATCH_W1
+            sty  P8ZP_SCRATCH_W1+1
+            lda  cx16.r1
+            ldy  cx16.r1+1
+            sta  P8ZP_SCRATCH_W2
+            sty  P8ZP_SCRATCH_W2+1
+
+            ldx  P8ZP_SCRATCH_B1
+            beq  _no_msb_size
+
+_loop_msb_size
+            ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            cmp  (P8ZP_SCRATCH_W2),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            bne  -
+            inc  P8ZP_SCRATCH_W1+1
+            inc  P8ZP_SCRATCH_W2+1
+            dec  P8ZP_SCRATCH_B1        ; msb(size) -= 1
+            dex
+            bne  _loop_msb_size
+
+_no_msb_size
+            lda  P8ZP_SCRATCH_REG       ; lsb(size)
+            bne  +
+            rts
+
++           ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            cmp  (P8ZP_SCRATCH_W2),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            cpy  P8ZP_SCRATCH_REG       ; lsb(size)
+            bne  -
+
+            lda #0
+            rts
+        }}
+    }
+
     inline asmsub read_flags() -> ubyte @A {
         %asm {{
         php
@@ -248,6 +270,7 @@ save_SCRATCH_ZPB1	.byte  0
 save_SCRATCH_ZPREG	.byte  0
 save_SCRATCH_ZPWORD1	.word  0
 save_SCRATCH_ZPWORD2	.word  0
+            ; !notreached!
         }}
     }
 
@@ -272,37 +295,37 @@ save_SCRATCH_ZPWORD2	.word  0
     asmsub exit(ubyte returnvalue @A) {
         ; -- immediately exit the program with a return code in the A register
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit2(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y) {
         ; -- immediately exit the program with result values in the A, X and Y registers.
         %asm {{
-            sta  cleanup_at_exit._exitcode
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit3(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y, bool carry @Pc) {
         ; -- immediately exit the program with result values in the A, X and Y registers, and the Carry flag in the status register.
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             lda  #0
             rol  a
-            sta  cleanup_at_exit._exitcodeCarry
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcodeCarry
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
@@ -313,6 +336,13 @@ save_SCRATCH_ZPWORD2	.word  0
         }}
     }
 
+    inline asmsub progstart() -> uword @AY {
+        %asm {{
+            lda  #<prog8_program_start
+            ldy  #>prog8_program_start
+        }}
+    }
+
     inline asmsub push(ubyte value @A) {
         %asm {{
             pha
@@ -461,6 +491,7 @@ cx16 {
     _cx16_vreg_storage
             .word 0,0,0,0,0,0,0,0
             .word 0,0,0,0,0,0,0,0
+            ; !notreached!
         }}
     }
 
@@ -481,3 +512,44 @@ cx16 {
     }
 
 }
+
+p8_sys_startup {
+    ; program startup and shutdown machinery. Needs to reside in normal system ram.
+
+    asmsub  init_system()  {
+        ; Initializes the machine to a sane starting state.
+        ; Called automatically by the loader program logic.
+        ; TODO
+        %asm {{
+            sei
+            ; TODO reset screen mode etc etc
+            cli
+            rts
+        }}
+    }
+
+    asmsub  init_system_phase2()  {
+        %asm {{
+            cld
+            clc
+            clv
+            rts
+        }}
+    }
+
+    asmsub  cleanup_at_exit() {
+        ; executed when the main subroutine does rts
+        %asm {{
+_exitcodeCarry = *+1
+            lda  #0
+            lsr  a
+_exitcode = *+1
+            lda  #0        ; exit code possibly modified in exit()
+_exitcodeX = *+1
+            ldx  #0
+_exitcodeY = *+1
+            ldy  #0
+            rts
+        }}
+    }
+}

compiler/res/prog8lib/c128/diskio.p8

@@ -0,0 +1,4 @@
+%import shared_cbm_diskio
+
+; No alterations here; everything is taken from the shared module.
+

compiler/res/prog8lib/c128/syslib.p8

@@ -56,51 +56,49 @@ cbm {
 ; CLEARSCR -> use txt.clear_screen
 ; HOMECRSR -> use txt.home or txt.plot
 
-romsub $FA65 = IRQDFRT() clobbers(A,X,Y)                        ; default IRQ routine
-romsub $FF33 = IRQDFEND() clobbers(A,X,Y)                       ; default IRQ end/cleanup
+extsub $FA65 = IRQDFRT() clobbers(A,X,Y)                        ; default IRQ routine
+extsub $FF33 = IRQDFEND() clobbers(A,X,Y)                       ; default IRQ end/cleanup
 
-; TODO c128 a bunch of kernal routines are missing here that are specific to the c128
-
-romsub $FF81 = CINT() clobbers(A,X,Y)                           ; (alias: SCINIT) initialize screen editor and video chip
-romsub $FF84 = IOINIT() clobbers(A, X)                          ; initialize I/O devices (CIA, SID, IRQ)
-romsub $FF87 = RAMTAS() clobbers(A,X,Y)                         ; initialize RAM, tape buffer, screen
-romsub $FF8A = RESTOR() clobbers(A,X,Y)                         ; restore default I/O vectors
-romsub $FF8D = VECTOR(uword userptr @ XY, bool dir @ Pc) clobbers(A,Y)     ; read/set I/O vector table
-romsub $FF90 = SETMSG(ubyte value @ A)                          ; set Kernal message control flag
-romsub $FF93 = SECOND(ubyte address @ A) clobbers(A)            ; (alias: LSTNSA) send secondary address after LISTEN
-romsub $FF96 = TKSA(ubyte address @ A) clobbers(A)              ; (alias: TALKSA) send secondary address after TALK
-romsub $FF99 = MEMTOP(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set top of memory  pointer
-romsub $FF9C = MEMBOT(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set bottom of memory  pointer
-romsub $FF9F = SCNKEY() clobbers(A,X,Y)                         ; scan the keyboard
-romsub $FFA2 = SETTMO(ubyte timeout @ A)                        ; set time-out flag for IEEE bus
-romsub $FFA5 = ACPTR() -> ubyte @ A                             ; (alias: IECIN) input byte from serial bus
-romsub $FFA8 = CIOUT(ubyte databyte @ A)                        ; (alias: IECOUT) output byte to serial bus
-romsub $FFAB = UNTLK() clobbers(A)                              ; command serial bus device to UNTALK
-romsub $FFAE = UNLSN() clobbers(A)                              ; command serial bus device to UNLISTEN
-romsub $FFB1 = LISTEN(ubyte device @ A) clobbers(A)             ; command serial bus device to LISTEN
-romsub $FFB4 = TALK(ubyte device @ A) clobbers(A)               ; command serial bus device to TALK
-romsub $FFB7 = READST() -> ubyte @ A                            ; read I/O status word  (use CLEARST to reset it to 0)
-romsub $FFBA = SETLFS(ubyte logical @ A, ubyte device @ X, ubyte secondary @ Y)   ; set logical file parameters
-romsub $FFBD = SETNAM(ubyte namelen @ A, str filename @ XY)     ; set filename parameters
-romsub $FFC0 = OPEN() clobbers(X,Y) -> bool @Pc, ubyte @A      ; (via 794 ($31A)) open a logical file
-romsub $FFC3 = CLOSE(ubyte logical @ A) clobbers(A,X,Y)         ; (via 796 ($31C)) close a logical file
-romsub $FFC6 = CHKIN(ubyte logical @ X) clobbers(A,X) -> bool @Pc    ; (via 798 ($31E)) define an input channel
-romsub $FFC9 = CHKOUT(ubyte logical @ X) clobbers(A,X)          ; (via 800 ($320)) define an output channel
-romsub $FFCC = CLRCHN() clobbers(A,X)                           ; (via 802 ($322)) restore default devices
-romsub $FFCF = CHRIN() clobbers(X, Y) -> ubyte @ A   ; (via 804 ($324)) input a character (for keyboard, read a whole line from the screen) A=byte read.
-romsub $FFD2 = CHROUT(ubyte character @ A)                           ; (via 806 ($326)) output a character
-romsub $FFD5 = LOAD(ubyte verify @ A, uword address @ XY) -> bool @Pc, ubyte @ A, uword @ XY     ; (via 816 ($330)) load from device
-romsub $FFD8 = SAVE(ubyte zp_startaddr @ A, uword endaddr @ XY) -> bool @ Pc, ubyte @ A          ; (via 818 ($332)) save to a device
-romsub $FFDB = SETTIM(ubyte low @ A, ubyte middle @ X, ubyte high @ Y)      ; set the software clock
-romsub $FFDE = RDTIM() -> ubyte @ A, ubyte @ X, ubyte @ Y       ; read the software clock (A=lo,X=mid,Y=high)
-romsub $FFE1 = STOP() clobbers(X) -> bool @ Pz, ubyte @ A       ; (via 808 ($328)) check the STOP key (and some others in A)     also see STOP2
-romsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A     ; (via 810 ($32A)) get a character       also see GETIN2
-romsub $FFE7 = CLALL() clobbers(A,X)                            ; (via 812 ($32C)) close all files
-romsub $FFEA = UDTIM() clobbers(A,X)                            ; update the software clock
-romsub $FFED = SCREEN() -> ubyte @ X, ubyte @ Y                 ; get current window dimensions into X (columns) and Y (rows)  NOTE: changed behavior compared to VIC/C64/PET SCREEN() routine!
-romsub $FFED = SCRORG() -> ubyte @ X, ubyte @ Y                 ; get current window dimensions into X (columns) and Y (rows)  NOTE: changed behavior compared to VIC/C64/PET SCREEN() routine!
-romsub $FFF0 = PLOT(ubyte col @ Y, ubyte row @ X, bool dir @ Pc) clobbers(A) -> ubyte @ X, ubyte @ Y       ; read/set position of cursor on screen.  Use txt.plot for a 'safe' wrapper that preserves X.
-romsub $FFF3 = IOBASE() -> uword @ XY                           ; read base address of I/O devices
+extsub $FF81 = CINT() clobbers(A,X,Y)                           ; (alias: SCINIT) initialize screen editor and video chip
+extsub $FF84 = IOINIT() clobbers(A, X)                          ; initialize I/O devices (CIA, SID, IRQ)
+extsub $FF87 = RAMTAS() clobbers(A,X,Y)                         ; initialize RAM, tape buffer, screen
+extsub $FF8A = RESTOR() clobbers(A,X,Y)                         ; restore default I/O vectors
+extsub $FF8D = VECTOR(uword userptr @ XY, bool dir @ Pc) clobbers(A,Y)     ; read/set I/O vector table
+extsub $FF90 = SETMSG(ubyte value @ A)                          ; set Kernal message control flag
+extsub $FF93 = SECOND(ubyte address @ A) clobbers(A)            ; (alias: LSTNSA) send secondary address after LISTEN
+extsub $FF96 = TKSA(ubyte address @ A) clobbers(A)              ; (alias: TALKSA) send secondary address after TALK
+extsub $FF99 = MEMTOP(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set top of memory  pointer
+extsub $FF9C = MEMBOT(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set bottom of memory  pointer
+extsub $FF9F = SCNKEY() clobbers(A,X,Y)                         ; scan the keyboard
+extsub $FFA2 = SETTMO(ubyte timeout @ A)                        ; set time-out flag for IEEE bus
+extsub $FFA5 = ACPTR() -> ubyte @ A                             ; (alias: IECIN) input byte from serial bus
+extsub $FFA8 = CIOUT(ubyte databyte @ A)                        ; (alias: IECOUT) output byte to serial bus
+extsub $FFAB = UNTLK() clobbers(A)                              ; command serial bus device to UNTALK
+extsub $FFAE = UNLSN() clobbers(A)                              ; command serial bus device to UNLISTEN
+extsub $FFB1 = LISTEN(ubyte device @ A) clobbers(A)             ; command serial bus device to LISTEN
+extsub $FFB4 = TALK(ubyte device @ A) clobbers(A)               ; command serial bus device to TALK
+extsub $FFB7 = READST() -> ubyte @ A                            ; read I/O status word  (use CLEARST to reset it to 0)
+extsub $FFBA = SETLFS(ubyte logical @ A, ubyte device @ X, ubyte secondary @ Y)   ; set logical file parameters
+extsub $FFBD = SETNAM(ubyte namelen @ A, str filename @ XY)     ; set filename parameters
+extsub $FFC0 = OPEN() clobbers(X,Y) -> bool @Pc, ubyte @A      ; (via 794 ($31A)) open a logical file
+extsub $FFC3 = CLOSE(ubyte logical @ A) clobbers(A,X,Y)         ; (via 796 ($31C)) close a logical file
+extsub $FFC6 = CHKIN(ubyte logical @ X) clobbers(A,X) -> bool @Pc    ; (via 798 ($31E)) define an input channel
+extsub $FFC9 = CHKOUT(ubyte logical @ X) clobbers(A,X)          ; (via 800 ($320)) define an output channel
+extsub $FFCC = CLRCHN() clobbers(A,X)                           ; (via 802 ($322)) restore default devices
+extsub $FFCF = CHRIN() clobbers(X, Y) -> ubyte @ A   ; (via 804 ($324)) input a character (for keyboard, read a whole line from the screen) A=byte read.
+extsub $FFD2 = CHROUT(ubyte character @ A)                           ; (via 806 ($326)) output a character
+extsub $FFD5 = LOAD(ubyte verify @ A, uword address @ XY) -> bool @Pc, ubyte @ A, uword @ XY     ; (via 816 ($330)) load from device
+extsub $FFD8 = SAVE(ubyte zp_startaddr @ A, uword endaddr @ XY) -> bool @ Pc, ubyte @ A          ; (via 818 ($332)) save to a device
+extsub $FFDB = SETTIM(ubyte low @ A, ubyte middle @ X, ubyte high @ Y)      ; set the software clock
+extsub $FFDE = RDTIM() -> ubyte @ A, ubyte @ X, ubyte @ Y       ; read the software clock (A=lo,X=mid,Y=high)
+extsub $FFE1 = STOP() clobbers(X) -> bool @ Pz, ubyte @ A       ; (via 808 ($328)) check the STOP key (and some others in A)     also see STOP2
+extsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A     ; (via 810 ($32A)) get a character       also see GETIN2
+extsub $FFE7 = CLALL() clobbers(A,X)                            ; (via 812 ($32C)) close all files
+extsub $FFEA = UDTIM() clobbers(A,X)                            ; update the software clock
+extsub $FFED = SCREEN() -> ubyte @ X, ubyte @ Y                 ; get current window dimensions into X (columns) and Y (rows)  NOTE: changed behavior compared to VIC/C64/PET SCREEN() routine!
+extsub $FFED = SCRORG() -> ubyte @ X, ubyte @ Y                 ; get current window dimensions into X (columns) and Y (rows)  NOTE: changed behavior compared to VIC/C64/PET SCREEN() routine!
+extsub $FFF0 = PLOT(ubyte col @ Y, ubyte row @ X, bool dir @ Pc) clobbers(A) -> ubyte @ X, ubyte @ Y       ; read/set position of cursor on screen.  Use txt.plot for a 'safe' wrapper that preserves X.
+extsub $FFF3 = IOBASE() -> uword @ XY                           ; read base address of I/O devices
 
 ; ---- end of C64 compatible ROM kernal routines ----
 
@@ -319,8 +317,27 @@ c128 {
     &ubyte  VM4     = $0A2F         ; starting page for VDC attribute mem
 
 
+; TODO c128 a bunch of kernal routines are missing here that are specific to the c128
+
+extsub $FF6E = JSRFAR()
+
+
 ; ---- C128 specific system utility routines: ----
 
+inline asmsub banks(ubyte banks @A) {
+    ; -- set the memory bank configuration MMU register
+    %asm {{
+        sta  $FF00
+    }}
+}
+
+inline asmsub getbanks() -> ubyte @A {
+    ; -- get the current memory bank configuration from the MMU register
+    %asm {{
+        lda  $FF00
+    }}
+}
+
 asmsub  disable_basic() clobbers(A) {
     %asm {{
         lda $0a04   ; disable BASIC shadow registers
@@ -342,6 +359,56 @@ asmsub  disable_basic() clobbers(A) {
     }}
 }
 
+asmsub  x16jsrfar() {
+    %asm {{
+        ; setup a JSRFAR call (using X16 call convention)
+        ; see https://cx16.dk/c128-kernal-routines/jsrfar.html
+        sty  $08                ; save registers
+        stx  $07
+        sta  $06
+        php                    ; including PSR
+        pla
+        sta  $05
+
+        pla                    ; get original return address
+        sta  $fa                ; and store it in a temp ZP pointer
+        pla
+        sta  $fb
+
+        ldy  #$01
+        lda  ($fa),y            ; grab low byte of target address
+        sta  $04
+        iny
+        lda  ($fa),y            ; now the high byte
+        sta  $03
+        iny
+        lda  ($fa),y            ; then the target bank
+        sta  $02
+
+        ; replace the original return address by it + 3 to skip the data bytes
+        clc
+        lda  $fa
+        adc  #3
+        sta  $fa
+        lda  $fb
+        adc  #0
+        pha
+        lda  $fa
+        pha
+
+        jsr c128.JSRFAR        ; call kernal's jsrfar routine
+
+        lda  $05                ; populate registers
+        pha
+        lda  $06
+        ldx  $07
+        ldy  $08
+        plp
+
+        rts                    ; and return
+    }}
+}
+
 ; ---- end of C128 specific system utility routines ----
 
 }
@@ -349,7 +416,7 @@ asmsub  disable_basic() clobbers(A) {
 sys {
     ; ------- lowlevel system routines --------
 
-    const ubyte target = 128         ;  compilation target specifier.  64 = C64, 128 = C128,  16 = CommanderX16.
+    const ubyte target = 128         ;  compilation target specifier.  255=virtual, 128=C128, 64=C64, 32=PET, 16=CommanderX16, 8=atari800XL, 7=Neo6502
 
     const ubyte sizeof_bool = 1
     const ubyte sizeof_byte = 1
@@ -359,80 +426,13 @@ sys {
     const ubyte sizeof_float = 5
 
 
-asmsub  init_system()  {
-    ; Initializes the machine to a sane starting state.
-    ; Called automatically by the loader program logic.
-    ; This means that the BASIC, KERNAL and CHARGEN ROMs are banked in,
-    ; the VIC, SID and CIA chips are reset, screen is cleared, and the default IRQ is set.
-    ; Also a different color scheme is chosen to identify ourselves a little.
-    ; Uppercase charset is activated.
-    %asm {{
-        sei
-        lda  #0
-        sta  $ff00      ; select default bank 15
-        jsr  cbm.IOINIT
-        jsr  cbm.RESTOR
-        jsr  cbm.CINT
-        lda  #6
-        sta  c64.EXTCOL
-        lda  #7
-        sta  cbm.COLOR
-        lda  #0
-        sta  c64.BGCOL0
-        jsr  disable_runstop_and_charsetswitch
-        cli
-        rts
-    }}
-}
+    sub  disable_runstop_and_charsetswitch() {
+        p8_sys_startup.disable_runstop_and_charsetswitch()
+    }
 
-asmsub  init_system_phase2()  {
-    %asm {{
-        cld
-        clc
-        clv
-        rts
-    }}
-}
-
-asmsub  cleanup_at_exit() {
-    ; executed when the main subroutine does rts
-    %asm {{
-        lda  #0
-        sta  $ff00          ; default bank 15
-        jsr  cbm.CLRCHN		; reset i/o channels
-        jsr  enable_runstop_and_charsetswitch
-_exitcodeCarry = *+1
-        lda  #0
-        lsr  a
-_exitcode = *+1
-        lda  #0        ; exit code possibly modified in exit()
-_exitcodeX = *+1
-        ldx  #0
-_exitcodeY = *+1
-        ldy  #0
-        rts
-    }}
-}
-
-asmsub  disable_runstop_and_charsetswitch() clobbers(A) {
-    %asm {{
-        lda  #$80
-        sta  247    ; disable charset switching
-        lda  #112
-        sta  808    ; disable run/stop key
-        rts
-    }}
-}
-
-asmsub  enable_runstop_and_charsetswitch() clobbers(A) {
-    %asm {{
-        lda  #0
-        sta  247    ; enable charset switching
-        lda  #110
-        sta  808    ; enable run/stop key
-        rts
-    }}
-}
+    sub  enable_runstop_and_charsetswitch() {
+        p8_sys_startup.enable_runstop_and_charsetswitch()
+    }
 
     asmsub save_prog8_internals() {
         %asm {{
@@ -453,6 +453,7 @@ save_SCRATCH_ZPB1	.byte  0
 save_SCRATCH_ZPREG	.byte  0
 save_SCRATCH_ZPWORD1	.word  0
 save_SCRATCH_ZPWORD2	.word  0
+            ; !notreached!
         }}
     }
 
@@ -504,8 +505,6 @@ _modified
 		tax
 		pla
 		rti
-
-_use_kernal     .byte  0
 		}}
 }
 
@@ -727,6 +726,65 @@ _longcopy
         }}
     }
 
+    asmsub memcmp(uword address1 @R0, uword address2 @R1, uword size @AY) -> byte @A {
+        ; Compares two blocks of memory
+        ; Returns -1 (255), 0 or 1, meaning: block 1 sorts before, equal or after block 2.
+        %asm {{
+            sta  P8ZP_SCRATCH_REG   ; lsb(size)
+            sty  P8ZP_SCRATCH_B1    ; msb(size)
+            lda  cx16.r0
+            ldy  cx16.r0+1
+            sta  P8ZP_SCRATCH_W1
+            sty  P8ZP_SCRATCH_W1+1
+            lda  cx16.r1
+            ldy  cx16.r1+1
+            sta  P8ZP_SCRATCH_W2
+            sty  P8ZP_SCRATCH_W2+1
+
+            ldx  P8ZP_SCRATCH_B1
+            beq  _no_msb_size
+
+_loop_msb_size
+            ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            cmp  (P8ZP_SCRATCH_W2),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            bne  -
+            inc  P8ZP_SCRATCH_W1+1
+            inc  P8ZP_SCRATCH_W2+1
+            dec  P8ZP_SCRATCH_B1        ; msb(size) -= 1
+            dex
+            bne  _loop_msb_size
+
+_no_msb_size
+            lda  P8ZP_SCRATCH_REG       ; lsb(size)
+            bne  +
+            rts
+
++           ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            cmp  (P8ZP_SCRATCH_W2),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            cpy  P8ZP_SCRATCH_REG       ; lsb(size)
+            bne  -
+
+            lda #0
+            rts
+        }}
+    }
+
     inline asmsub read_flags() -> ubyte @A {
         %asm {{
         php
@@ -788,37 +846,37 @@ _longcopy
     asmsub exit(ubyte returnvalue @A) {
         ; -- immediately exit the program with a return code in the A register
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit2(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y) {
         ; -- immediately exit the program with result values in the A, X and Y registers.
         %asm {{
-            sta  cleanup_at_exit._exitcode
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit3(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y, bool carry @Pc) {
         ; -- immediately exit the program with result values in the A, X and Y registers, and the Carry flag in the status register.
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             lda  #0
             rol  a
-            sta  cleanup_at_exit._exitcodeCarry
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcodeCarry
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
@@ -829,6 +887,13 @@ _longcopy
         }}
     }
 
+    inline asmsub progstart() -> uword @AY {
+        %asm {{
+            lda  #<prog8_program_start
+            ldy  #>prog8_program_start
+        }}
+    }
+
     inline asmsub push(ubyte value @A) {
         %asm {{
             pha
@@ -978,6 +1043,7 @@ cx16 {
     _cx16_vreg_storage
             .word 0,0,0,0,0,0,0,0
             .word 0,0,0,0,0,0,0,0
+            ; !notreached!
         }}
     }
 
@@ -998,3 +1064,85 @@ cx16 {
     }
 
 }
+
+p8_sys_startup {
+    ; program startup and shutdown machinery. Needs to reside in normal system ram.
+
+asmsub  init_system()  {
+    ; Initializes the machine to a sane starting state.
+    ; Called automatically by the loader program logic.
+    ; This means that the BASIC, KERNAL and CHARGEN ROMs are banked in,
+    ; the VIC, SID and CIA chips are reset, screen is cleared, and the default IRQ is set.
+    ; Also a different color scheme is chosen to identify ourselves a little.
+    ; Uppercase charset is activated.
+    %asm {{
+        sei
+        lda  #0
+        sta  $ff00      ; select default bank 15
+        jsr  cbm.IOINIT
+        jsr  cbm.RESTOR
+        jsr  cbm.CINT
+        lda  #%00001110
+        sta  $ff00      ; bank out basic rom so we have ram from $1c00-$bfff
+        lda  #6
+        sta  c64.EXTCOL
+        lda  #7
+        sta  cbm.COLOR
+        lda  #0
+        sta  c64.BGCOL0
+        jsr  disable_runstop_and_charsetswitch
+        cli
+        rts
+    }}
+}
+
+asmsub  init_system_phase2()  {
+    %asm {{
+        cld
+        clc
+        clv
+        rts
+    }}
+}
+
+asmsub  cleanup_at_exit() {
+    ; executed when the main subroutine does rts
+    %asm {{
+        lda  #0
+        sta  $ff00          ; default bank 15
+        jsr  cbm.CLRCHN		; reset i/o channels
+        jsr  enable_runstop_and_charsetswitch
+_exitcodeCarry = *+1
+        lda  #0
+        lsr  a
+_exitcode = *+1
+        lda  #0        ; exit code possibly modified in exit()
+_exitcodeX = *+1
+        ldx  #0
+_exitcodeY = *+1
+        ldy  #0
+        rts
+    }}
+}
+
+asmsub  disable_runstop_and_charsetswitch() clobbers(A) {
+    %asm {{
+        lda  #$80
+        sta  247    ; disable charset switching
+        lda  #112
+        sta  808    ; disable run/stop key
+        rts
+    }}
+}
+
+asmsub  enable_runstop_and_charsetswitch() clobbers(A) {
+    %asm {{
+        lda  #0
+        sta  247    ; enable charset switching
+        lda  #110
+        sta  808    ; enable run/stop key
+        rts
+    }}
+}
+
+}

compiler/res/prog8lib/c128/textio.p8

@@ -13,7 +13,7 @@ txt {
 const ubyte DEFAULT_WIDTH = 40
 const ubyte DEFAULT_HEIGHT = 25
 
-romsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
+extsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
 
 
 sub  clear_screen() {
@@ -314,6 +314,7 @@ _mod		sta  $ffff		; modified
 		rts
 
 _screenrows	.word  cbm.Screen + range(0, 1000, 40)
+        ; !notreached!
 	}}
 }
 
@@ -357,6 +358,7 @@ _mod		sta  $ffff		; modified
 		rts
 
 _colorrows	.word  $d800 + range(0, 1000, 40)
+        ; !notreached!
 	}}
 }
 

compiler/res/prog8lib/c64/diskio.p8

@@ -0,0 +1,4 @@
+%import shared_cbm_diskio
+
+; No alterations here; everything is taken from the shared module.
+

compiler/res/prog8lib/c64/floats.asm

@@ -120,12 +120,8 @@ cast_FAC1_as_uw_into_ya	.proc               ; also used for float 2 ub
 		.pend
 
 cast_FAC1_as_w_into_ay	.proc               ; also used for float 2 b
-		; -- cast fac1 to word into A/Y
-		;    clobbers X
-		jsr  AYINT
-		ldy  floats.AYINT_facmo
-		lda  floats.AYINT_facmo+1
-		rts
+		; -- cast fac1 to word into A/Y.  clobbers X
+		jmp  AYINT2
 		.pend
 
 

compiler/res/prog8lib/c64/floats.p8

@@ -15,21 +15,21 @@ floats {
 ; note: fac1/2 might get clobbered even if not mentioned in the function's name.
 ; note: for subtraction and division, the left operand is in fac2, the right operand in fac1.
 
-romsub $bba2 = MOVFM(uword mflpt @ AY) clobbers(A,Y)        ; load mflpt value from memory  in A/Y into fac1
-romsub $bba6 = FREADMEM() clobbers(A,Y)                     ; load mflpt value from memory  in $22/$23 into fac1
-romsub $ba8c = CONUPK(uword mflpt @ AY) clobbers(A,Y)       ; load mflpt value from memory  in A/Y into fac2
-romsub $ba90 = FAREADMEM() clobbers(A,Y)                    ; load mflpt value from memory  in $22/$23 into fac2
-romsub $bbfc = MOVFA() clobbers(A,X)                        ; copy fac2 to fac1
-romsub $bc0c = MOVAF() clobbers(A,X)                        ; copy fac1 to fac2  (rounded the least significant bit)
-romsub $bc0f = MOVEF() clobbers(A,X)                        ; copy fac1 to fac2
-romsub $bbd4 = MOVMF(uword mflpt @ XY) clobbers(A,Y)        ; store fac1 to memory  X/Y as 5-byte mflpt
+extsub $bba2 = MOVFM(uword mflpt @ AY) clobbers(A,Y)        ; load mflpt value from memory  in A/Y into fac1
+extsub $bba6 = FREADMEM() clobbers(A,Y)                     ; load mflpt value from memory  in $22/$23 into fac1
+extsub $ba8c = CONUPK(uword mflpt @ AY) clobbers(A,Y)       ; load mflpt value from memory  in A/Y into fac2
+extsub $ba90 = FAREADMEM() clobbers(A,Y)                    ; load mflpt value from memory  in $22/$23 into fac2
+extsub $bbfc = MOVFA() clobbers(A,X)                        ; copy fac2 to fac1
+extsub $bc0c = MOVAF() clobbers(A,X)                        ; copy fac1 to fac2  (rounded the least significant bit)
+extsub $bc0f = MOVEF() clobbers(A,X)                        ; copy fac1 to fac2
+extsub $bbd4 = MOVMF(uword mflpt @ XY) clobbers(A,Y)        ; store fac1 to memory  X/Y as 5-byte mflpt
 
 ; fac1 -> unsigned word in Y/A (might throw ILLEGAL QUANTITY) (result also in $14/15)
 ; (tip: use floats.GETADRAY to get A/Y output; lo/hi switched to normal little endian order)
-romsub $b7f7 = GETADR() clobbers(X) -> ubyte @ Y, ubyte @ A
+extsub $b7f7 = GETADR() clobbers(X) -> ubyte @ Y, ubyte @ A
 
-romsub $bc9b = QINT() clobbers(A,X,Y)           ; fac1 -> 4-byte signed integer in 98-101 ($62-$65), with the MSB FIRST.
-romsub $b1bf = AYINT() clobbers(A,X,Y)          ; fac1-> signed word in 100-101 ($64-$65) MSB FIRST. (might throw ILLEGAL QUANTITY)
+extsub $bc9b = QINT() clobbers(A,X,Y)           ; fac1 -> 4-byte signed integer in 98-101 ($62-$65), with the MSB FIRST.
+extsub $b1bf = AYINT() clobbers(A,X,Y)          ; fac1-> signed word in 100-101 ($64-$65) MSB FIRST. (might throw ILLEGAL QUANTITY) DON'T USE THIS, USE WRAPPER 'AYINT2' INSTEAD.
 
 ; GIVAYF: signed word in Y/A (note different lsb/msb order) -> float in fac1
 ; (tip: use floats.GIVAYFAY to use A/Y input; lo/hi switched to normal order)
@@ -37,49 +37,63 @@ romsub $b1bf = AYINT() clobbers(A,X,Y)          ; fac1-> signed word in 100-101
 ; there is also floats.FREADS32  that reads from 98-101 ($62-$65) MSB FIRST
 ; there is also floats.FREADUS32  that reads from 98-101 ($62-$65) MSB FIRST
 ; there is also floats.FREAD{S,U}24AXY  that read (un)signed int24 into fac1 from A/X/Y (lo/mid/hi bytes)
-romsub $b391 = GIVAYF(ubyte lo @ Y, ubyte hi @ A) clobbers(A,X,Y)
+extsub $b391 = GIVAYF(ubyte lo @ Y, ubyte hi @ A) clobbers(A,X,Y)
 
-romsub $b3a2 = FREADUY(ubyte value @ Y) clobbers(A,X,Y)     ; 8 bit unsigned Y -> float in fac1
-romsub $bc3c = FREADSA(byte value @ A) clobbers(A,X,Y)      ; 8 bit signed A -> float in fac1
-romsub $b7b5 = FREADSTR(ubyte length @ A) clobbers(A,X,Y)   ; str -> fac1, $22/23 must point to string, A=string length.  Also see parse()
-romsub $aabc = FPRINTLN() clobbers(A,X,Y)                   ; print string of fac1, on one line (= with newline) destroys fac1.  (consider FOUT + STROUT as well)
-romsub $bddd = FOUT() clobbers(X) -> uword @ AY             ; fac1 -> string, address returned in AY ($0100)
+extsub $b3a2 = FREADUY(ubyte value @ Y) clobbers(A,X,Y)     ; 8 bit unsigned Y -> float in fac1
+extsub $bc3c = FREADSA(byte value @ A) clobbers(A,X,Y)      ; 8 bit signed A -> float in fac1
+extsub $b7b5 = FREADSTR(ubyte length @ A) clobbers(A,X,Y)   ; str -> fac1, $22/23 must point to string, A=string length.  Also see parse()
+extsub $aabc = FPRINTLN() clobbers(A,X,Y)                   ; print string of fac1, on one line (= with newline) destroys fac1.  (consider FOUT + STROUT as well)
+extsub $bddd = FOUT() clobbers(X) -> uword @ AY             ; fac1 -> string, address returned in AY ($0100)
 
-romsub $b849 = FADDH() clobbers(A,X,Y)                      ; fac1 += 0.5, for integer rounding- call this before INT
-romsub $bae2 = MUL10() clobbers(A,X,Y)                      ; fac1 *= 10
-romsub $bafe = DIV10() clobbers(A,X,Y)                      ; fac1 /= 10 , CAUTION: result is always positive! You have to fix sign manually!
-romsub $bc5b = FCOMP(uword mflpt @ AY) clobbers(X,Y) -> ubyte @ A   ; A = compare fac1 to mflpt in A/Y, 0=equal 1=fac1 is greater, 255=fac1 is less than
+extsub $b849 = FADDH() clobbers(A,X,Y)                      ; fac1 += 0.5, for integer rounding- call this before INT
+extsub $bae2 = MUL10() clobbers(A,X,Y)                      ; fac1 *= 10
+extsub $bafe = DIV10() clobbers(A,X,Y)                      ; fac1 /= 10 , CAUTION: result is always positive! You have to fix sign manually!
+extsub $bc5b = FCOMP(uword mflpt @ AY) clobbers(X,Y) -> ubyte @ A   ; A = compare fac1 to mflpt in A/Y, 0=equal 1=fac1 is greater, 255=fac1 is less than
 
-romsub $b86a = FADDT() clobbers(A,X,Y)                      ; fac1 += fac2
-romsub $b867 = FADD(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 += mflpt value from A/Y
-romsub $b853 = FSUBT() clobbers(A,X,Y)                      ; fac1 = fac2-fac1   mind the order of the operands
-romsub $b850 = FSUB(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = mflpt from A/Y - fac1
-romsub $ba2b = FMULTT() clobbers(A,X,Y)                     ; fac1 *= fac2
-romsub $ba28 = FMULT(uword mflpt @ AY) clobbers(A,X,Y)      ; fac1 *= mflpt value from A/Y
-romsub $bb12 = FDIVT() clobbers(A,X,Y)                      ; fac1 = fac2/fac1  mind the order of the operands
-romsub $bb0f = FDIV(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = mflpt in A/Y / fac1
-romsub $bf7b = FPWRT() clobbers(A,X,Y)                      ; fac1 = fac2 ** fac1
-romsub $bf78 = FPWR(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = fac2 ** mflpt from A/Y
-romsub $bd7e = FINLOG(byte value @A) clobbers (A, X, Y)     ; fac1 += signed byte in A
+extsub $b86a = FADDT() clobbers(A,X,Y)                      ; fac1 += fac2
+extsub $b867 = FADD(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 += mflpt value from A/Y
+extsub $b853 = FSUBT() clobbers(A,X,Y)                      ; fac1 = fac2-fac1   mind the order of the operands
+extsub $b850 = FSUB(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = mflpt from A/Y - fac1
+extsub $ba2b = FMULTT() clobbers(A,X,Y)                     ; fac1 *= fac2
+extsub $ba28 = FMULT(uword mflpt @ AY) clobbers(A,X,Y)      ; fac1 *= mflpt value from A/Y
+extsub $bb12 = FDIVT() clobbers(A,X,Y)                      ; fac1 = fac2/fac1  mind the order of the operands
+extsub $bb0f = FDIV(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = mflpt in A/Y / fac1
+extsub $bf7b = FPWRT() clobbers(A,X,Y)                      ; fac1 = fac2 ** fac1
+extsub $bf78 = FPWR(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = fac2 ** mflpt from A/Y
+extsub $bd7e = FINLOG(byte value @A) clobbers (A, X, Y)     ; fac1 += signed byte in A
 
-romsub $aed4 = NOTOP() clobbers(A,X,Y)                      ; fac1 = NOT(fac1)
-romsub $bccc = INT() clobbers(A,X,Y)                        ; INT() truncates, use FADDH first to integer round instead of trunc
-romsub $b9ea = LOG() clobbers(A,X,Y)                        ; fac1 = LN(fac1)  (natural log)
-romsub $bc39 = SGN() clobbers(A,X,Y)                        ; fac1 = SGN(fac1), result of SIGN (-1, 0 or 1)
-romsub $bc2b = SIGN() -> ubyte @ A                          ; SIGN(fac1) to A, $ff, $0, $1 for negative, zero, positive
-romsub $bc58 = ABS()                                        ; fac1 = ABS(fac1)
-romsub $bf71 = SQR() clobbers(A,X,Y)                        ; fac1 = SQRT(fac1)
-romsub $bf74 = SQRA() clobbers(A,X,Y)                       ; fac1 = SQRT(fac2)
-romsub $bfed = EXP() clobbers(A,X,Y)                        ; fac1 = EXP(fac1)  (e ** fac1)
-romsub $bfb4 = NEGOP() clobbers(A)                          ; switch the sign of fac1 (fac1 = -fac1)
-romsub $b8d7 = NORMAL() clobbers(A)                         ; normalize FAC1
-romsub $e097 = RND() clobbers(A,X,Y)                        ; fac1 = RND(fac1) float random number generator
-romsub $e264 = COS() clobbers(A,X,Y)                        ; fac1 = COS(fac1)
-romsub $e26b = SIN() clobbers(A,X,Y)                        ; fac1 = SIN(fac1)
-romsub $e2b4 = TAN() clobbers(A,X,Y)                        ; fac1 = TAN(fac1)
-romsub $e30e = ATN() clobbers(A,X,Y)                        ; fac1 = ATN(fac1)
+extsub $aed4 = NOTOP() clobbers(A,X,Y)                      ; fac1 = NOT(fac1)
+extsub $bccc = INT() clobbers(A,X,Y)                        ; INT() truncates, use FADDH first to integer round instead of trunc
+extsub $b9ea = LOG() clobbers(A,X,Y)                        ; fac1 = LN(fac1)  (natural log)
+extsub $bc39 = SGN() clobbers(A,X,Y)                        ; fac1 = SGN(fac1), result of SIGN (-1, 0 or 1)
+extsub $bc2b = SIGN() -> ubyte @ A                          ; SIGN(fac1) to A, $ff, $0, $1 for negative, zero, positive
+extsub $bc58 = ABS()                                        ; fac1 = ABS(fac1)
+extsub $bf71 = SQR() clobbers(A,X,Y)                        ; fac1 = SQRT(fac1)
+extsub $bf74 = SQRA() clobbers(A,X,Y)                       ; fac1 = SQRT(fac2)
+extsub $bfed = EXP() clobbers(A,X,Y)                        ; fac1 = EXP(fac1)  (e ** fac1)
+extsub $bfb4 = NEGOP() clobbers(A)                          ; switch the sign of fac1 (fac1 = -fac1)
+extsub $b8d7 = NORMAL() clobbers(A)                         ; normalize FAC1
+extsub $e097 = RND() clobbers(A,X,Y)                        ; fac1 = RND(fac1) float random number generator
+extsub $e264 = COS() clobbers(A,X,Y)                        ; fac1 = COS(fac1)
+extsub $e26b = SIN() clobbers(A,X,Y)                        ; fac1 = SIN(fac1)
+extsub $e2b4 = TAN() clobbers(A,X,Y)                        ; fac1 = TAN(fac1)
+extsub $e30e = ATN() clobbers(A,X,Y)                        ; fac1 = ATN(fac1)
 
 
+asmsub  AYINT2() clobbers(X) -> word @AY {
+    ; fac1-> signed word in AY. Safe wrapper around the AYINT kernal routine (not reading internal memory locations)
+    ; (might throw ILLEGAL QUANTITY)
+    %asm {{
+		jsr  AYINT
+		ldx  #<floats_temp_var
+		ldy  #>floats_temp_var
+		jsr  MOVMF
+		lda  floats_temp_var+4
+		ldy  floats_temp_var+3
+        rts
+    }}
+}
+
 asmsub  FREADS32() clobbers(A,X,Y)  {
 	; ---- fac1 = signed int32 from $62-$65 big endian (MSB FIRST)
 	%asm {{
@@ -154,7 +168,7 @@ asmsub FREADU24AXY(ubyte lo @ A, ubyte mid @ X, ubyte hi @ Y) clobbers(A, X, Y)
 asmsub  GIVUAYFAY  (uword value @ AY) clobbers(A,X,Y)  {
 	; ---- unsigned 16 bit word in A/Y (lo/hi) to fac1
 	%asm {{
-		sty  $62
+		sty  $62    ; facmo
 		sta  $63
 		ldx  #$90
 		sec
@@ -183,7 +197,6 @@ asmsub  GETADRAY  () clobbers(X) -> uword @ AY  {
 	}}
 }
 
-&uword AYINT_facmo = $64      ; $64/$65 contain result of AYINT
 
 sub rnd() -> float {
     %asm {{

compiler/res/prog8lib/c64/graphics.p8

@@ -347,12 +347,11 @@ hline_zero2
 
     uword @zp internal_plotx     ; 0..319        ; separate 'parameter' for internal_plot()
 
-    asmsub  internal_plot(ubyte ploty @A) clobbers (A, X, Y) {      ; internal_plotx is 16 bits 0 to 319... doesn't fit in a register
+    asmsub  internal_plot(ubyte ploty @Y) clobbers (A, X, Y) {      ; internal_plotx is 16 bits 0 to 319... doesn't fit in a register
         %asm {{
-        tay
         lda  p8v_internal_plotx+1
         sta  P8ZP_SCRATCH_W2+1
-        lsr  a            ; 0
+        lsr  a      ; make a=0
         sta  P8ZP_SCRATCH_W2
         lda  p8v_internal_plotx
         pha
@@ -377,7 +376,7 @@ hline_zero2
 
 _ormask     .byte 128, 64, 32, 16, 8, 4, 2, 1
 
-; note: this can be even faster if we also have a 256 byte x-lookup table, but hey.
+; note: this can be even faster if we also have a 320 word x-lookup table, but hey, that's a lot of memory.
 ; see http://codebase64.org/doku.php?id=base:various_techniques_to_calculate_adresses_fast_common_screen_formats_for_pixel_graphics
 ; the y lookup tables encodes this formula:  BITMAP_ADDRESS + 320*(py>>3) + (py & 7)    (y from 0..199)
 ; We use the 64tass syntax for range expressions to calculate this table on assembly time.
@@ -386,6 +385,7 @@ _plot_y_values := p8c_BITMAP_ADDRESS + 320*(range(200)>>3) + (range(200) & 7)
 
 _y_lookup_lo    .byte  <_plot_y_values
 _y_lookup_hi    .byte  >_plot_y_values
+            ; !notreached!
 
         }}
     }

compiler/res/prog8lib/c64/syslib.p8

@@ -57,50 +57,50 @@ cbm {
 
 ; ---- CBM ROM kernal routines (C64 addresses) ----
 
-romsub $AB1E = STROUT(uword strptr @ AY) clobbers(A, X, Y)      ; print null-terminated string (use txt.print instead)
-romsub $E544 = CLEARSCR() clobbers(A,X,Y)                       ; clear the screen
-romsub $E566 = HOMECRSR() clobbers(A,X,Y)                       ; cursor to top left of screen
-romsub $EA31 = IRQDFRT() clobbers(A,X,Y)                        ; default IRQ routine
-romsub $EA81 = IRQDFEND() clobbers(A,X,Y)                       ; default IRQ end/cleanup
-romsub $FF81 = CINT() clobbers(A,X,Y)                           ; (alias: SCINIT) initialize screen editor and video chip
-romsub $FF84 = IOINIT() clobbers(A, X)                          ; initialize I/O devices (CIA, SID, IRQ)
-romsub $FF87 = RAMTAS() clobbers(A,X,Y)                         ; initialize RAM, tape buffer, screen
-romsub $FF8A = RESTOR() clobbers(A,X,Y)                         ; restore default I/O vectors
-romsub $FF8D = VECTOR(uword userptr @ XY, bool dir @ Pc) clobbers(A,Y)     ; read/set I/O vector table
-romsub $FF90 = SETMSG(ubyte value @ A)                          ; set Kernal message control flag
-romsub $FF93 = SECOND(ubyte address @ A) clobbers(A)            ; (alias: LSTNSA) send secondary address after LISTEN
-romsub $FF96 = TKSA(ubyte address @ A) clobbers(A)              ; (alias: TALKSA) send secondary address after TALK
-romsub $FF99 = MEMTOP(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set top of memory  pointer
-romsub $FF9C = MEMBOT(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set bottom of memory  pointer
-romsub $FF9F = SCNKEY() clobbers(A,X,Y)                         ; scan the keyboard
-romsub $FFA2 = SETTMO(ubyte timeout @ A)                        ; set time-out flag for IEEE bus
-romsub $FFA5 = ACPTR() -> ubyte @ A                             ; (alias: IECIN) input byte from serial bus
-romsub $FFA8 = CIOUT(ubyte databyte @ A)                        ; (alias: IECOUT) output byte to serial bus
-romsub $FFAB = UNTLK() clobbers(A)                              ; command serial bus device to UNTALK
-romsub $FFAE = UNLSN() clobbers(A)                              ; command serial bus device to UNLISTEN
-romsub $FFB1 = LISTEN(ubyte device @ A) clobbers(A)             ; command serial bus device to LISTEN
-romsub $FFB4 = TALK(ubyte device @ A) clobbers(A)               ; command serial bus device to TALK
-romsub $FFB7 = READST() -> ubyte @ A                            ; read I/O status word  (use CLEARST to reset it to 0)
-romsub $FFBA = SETLFS(ubyte logical @ A, ubyte device @ X, ubyte secondary @ Y)   ; set logical file parameters
-romsub $FFBD = SETNAM(ubyte namelen @ A, str filename @ XY)     ; set filename parameters
-romsub $FFC0 = OPEN() clobbers(X,Y) -> bool @Pc, ubyte @A      ; (via 794 ($31A)) open a logical file
-romsub $FFC3 = CLOSE(ubyte logical @ A) clobbers(A,X,Y)         ; (via 796 ($31C)) close a logical file
-romsub $FFC6 = CHKIN(ubyte logical @ X) clobbers(A,X) -> bool @Pc    ; (via 798 ($31E)) define an input channel
-romsub $FFC9 = CHKOUT(ubyte logical @ X) clobbers(A,X)          ; (via 800 ($320)) define an output channel
-romsub $FFCC = CLRCHN() clobbers(A,X)                           ; (via 802 ($322)) restore default devices
-romsub $FFCF = CHRIN() clobbers(X, Y) -> ubyte @ A   ; (via 804 ($324)) input a character (for keyboard, read a whole line from the screen) A=byte read.
-romsub $FFD2 = CHROUT(ubyte character @ A)                      ; (via 806 ($326)) output a character
-romsub $FFD5 = LOAD(ubyte verify @ A, uword address @ XY) -> bool @Pc, ubyte @ A, uword @ XY     ; (via 816 ($330)) load from device
-romsub $FFD8 = SAVE(ubyte zp_startaddr @ A, uword endaddr @ XY) -> bool @ Pc, ubyte @ A          ; (via 818 ($332)) save to a device
-romsub $FFDB = SETTIM(ubyte low @ A, ubyte middle @ X, ubyte high @ Y)      ; set the software clock
-romsub $FFDE = RDTIM() -> ubyte @ A, ubyte @ X, ubyte @ Y       ; read the software clock (A=lo,X=mid,Y=high)
-romsub $FFE1 = STOP() clobbers(X) -> bool @ Pz, ubyte @ A       ; (via 808 ($328)) check the STOP key (and some others in A)     also see STOP2
-romsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A     ; (via 810 ($32A)) get a character       also see GETIN2
-romsub $FFE7 = CLALL() clobbers(A,X)                            ; (via 812 ($32C)) close all files
-romsub $FFEA = UDTIM() clobbers(A,X)                            ; update the software clock
-romsub $FFED = SCREEN() -> ubyte @ X, ubyte @ Y                 ; get size of text screen into X (columns) and Y (rows)
-romsub $FFF0 = PLOT(ubyte col @ Y, ubyte row @ X, bool dir @ Pc) clobbers(A) -> ubyte @ X, ubyte @ Y       ; read/set position of cursor on screen.  Use txt.plot for a 'safe' wrapper that preserves X.
-romsub $FFF3 = IOBASE() -> uword @ XY                           ; read base address of I/O devices
+extsub $AB1E = STROUT(uword strptr @ AY) clobbers(A, X, Y)      ; print null-terminated string (use txt.print instead)
+extsub $E544 = CLEARSCR() clobbers(A,X,Y)                       ; clear the screen
+extsub $E566 = HOMECRSR() clobbers(A,X,Y)                       ; cursor to top left of screen
+extsub $EA31 = IRQDFRT() clobbers(A,X,Y)                        ; default IRQ routine
+extsub $EA81 = IRQDFEND() clobbers(A,X,Y)                       ; default IRQ end/cleanup
+extsub $FF81 = CINT() clobbers(A,X,Y)                           ; (alias: SCINIT) initialize screen editor and video chip
+extsub $FF84 = IOINIT() clobbers(A, X)                          ; initialize I/O devices (CIA, SID, IRQ)
+extsub $FF87 = RAMTAS() clobbers(A,X,Y)                         ; initialize RAM, tape buffer, screen
+extsub $FF8A = RESTOR() clobbers(A,X,Y)                         ; restore default I/O vectors
+extsub $FF8D = VECTOR(uword userptr @ XY, bool dir @ Pc) clobbers(A,Y)     ; read/set I/O vector table
+extsub $FF90 = SETMSG(ubyte value @ A)                          ; set Kernal message control flag
+extsub $FF93 = SECOND(ubyte address @ A) clobbers(A)            ; (alias: LSTNSA) send secondary address after LISTEN
+extsub $FF96 = TKSA(ubyte address @ A) clobbers(A)              ; (alias: TALKSA) send secondary address after TALK
+extsub $FF99 = MEMTOP(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set top of memory  pointer
+extsub $FF9C = MEMBOT(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set bottom of memory  pointer
+extsub $FF9F = SCNKEY() clobbers(A,X,Y)                         ; scan the keyboard
+extsub $FFA2 = SETTMO(ubyte timeout @ A)                        ; set time-out flag for IEEE bus
+extsub $FFA5 = ACPTR() -> ubyte @ A                             ; (alias: IECIN) input byte from serial bus
+extsub $FFA8 = CIOUT(ubyte databyte @ A)                        ; (alias: IECOUT) output byte to serial bus
+extsub $FFAB = UNTLK() clobbers(A)                              ; command serial bus device to UNTALK
+extsub $FFAE = UNLSN() clobbers(A)                              ; command serial bus device to UNLISTEN
+extsub $FFB1 = LISTEN(ubyte device @ A) clobbers(A)             ; command serial bus device to LISTEN
+extsub $FFB4 = TALK(ubyte device @ A) clobbers(A)               ; command serial bus device to TALK
+extsub $FFB7 = READST() -> ubyte @ A                            ; read I/O status word  (use CLEARST to reset it to 0)
+extsub $FFBA = SETLFS(ubyte logical @ A, ubyte device @ X, ubyte secondary @ Y)   ; set logical file parameters
+extsub $FFBD = SETNAM(ubyte namelen @ A, str filename @ XY)     ; set filename parameters
+extsub $FFC0 = OPEN() clobbers(X,Y) -> bool @Pc, ubyte @A      ; (via 794 ($31A)) open a logical file
+extsub $FFC3 = CLOSE(ubyte logical @ A) clobbers(A,X,Y)         ; (via 796 ($31C)) close a logical file
+extsub $FFC6 = CHKIN(ubyte logical @ X) clobbers(A,X) -> bool @Pc    ; (via 798 ($31E)) define an input channel
+extsub $FFC9 = CHKOUT(ubyte logical @ X) clobbers(A,X)          ; (via 800 ($320)) define an output channel
+extsub $FFCC = CLRCHN() clobbers(A,X)                           ; (via 802 ($322)) restore default devices
+extsub $FFCF = CHRIN() clobbers(X, Y) -> ubyte @ A   ; (via 804 ($324)) input a character (for keyboard, read a whole line from the screen) A=byte read.
+extsub $FFD2 = CHROUT(ubyte character @ A)                      ; (via 806 ($326)) output a character
+extsub $FFD5 = LOAD(ubyte verify @ A, uword address @ XY) -> bool @Pc, ubyte @ A, uword @ XY     ; (via 816 ($330)) load from device
+extsub $FFD8 = SAVE(ubyte zp_startaddr @ A, uword endaddr @ XY) -> bool @ Pc, ubyte @ A          ; (via 818 ($332)) save to a device
+extsub $FFDB = SETTIM(ubyte low @ A, ubyte middle @ X, ubyte high @ Y)      ; set the software clock
+extsub $FFDE = RDTIM() -> ubyte @ A, ubyte @ X, ubyte @ Y       ; read the software clock (A=lo,X=mid,Y=high)
+extsub $FFE1 = STOP() clobbers(X) -> bool @ Pz, ubyte @ A       ; (via 808 ($328)) check the STOP key (and some others in A)     also see STOP2
+extsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A     ; (via 810 ($32A)) get a character       also see GETIN2
+extsub $FFE7 = CLALL() clobbers(A,X)                            ; (via 812 ($32C)) close all files
+extsub $FFEA = UDTIM() clobbers(A,X)                            ; update the software clock
+extsub $FFED = SCREEN() -> ubyte @ X, ubyte @ Y                 ; get size of text screen into X (columns) and Y (rows)
+extsub $FFF0 = PLOT(ubyte col @ Y, ubyte row @ X, bool dir @ Pc) clobbers(A) -> ubyte @ X, ubyte @ Y       ; read/set position of cursor on screen.  Use txt.plot for a 'safe' wrapper that preserves X.
+extsub $FFF3 = IOBASE() -> uword @ XY                           ; read base address of I/O devices
 
 
 inline asmsub STOP2() clobbers(X,A) -> bool @Pz  {
@@ -307,6 +307,93 @@ c64 {
 
 ; ---- end of SID registers ----
 
+asmsub banks(ubyte banks @A) {
+    ; -- set the memory bank configuration
+    ;    see https://www.c64-wiki.com/wiki/Bank_Switching
+    %asm {{
+        and  #%00000111
+        sta  P8ZP_SCRATCH_REG
+        sei
+        lda  $01
+        and  #%11111000
+        ora  P8ZP_SCRATCH_REG
+        sta  $01
+        cli
+        rts
+    }}
+}
+
+inline asmsub getbanks() -> ubyte @A {
+    ; -- get the current memory bank configuration
+    ;    see https://www.c64-wiki.com/wiki/Bank_Switching
+    %asm {{
+        lda  $01
+        and  #%00000111
+    }}
+}
+
+    asmsub x16jsrfar() {
+        %asm {{
+            ; setup a JSRFAR call (using X16 call convention)
+            sta  P8ZP_SCRATCH_W2        ; save A
+            sty  P8ZP_SCRATCH_W2+1      ; save Y
+            php
+            pla
+            sta  P8ZP_SCRATCH_REG       ; save Status
+
+            pla
+            sta  P8ZP_SCRATCH_W1
+            pla
+            sta  P8ZP_SCRATCH_W1+1
+
+            ; retrieve arguments
+            ldy  #$01
+            lda  (P8ZP_SCRATCH_W1),y            ; grab low byte of target address
+            sta  _jmpfar+1
+            iny
+            lda  (P8ZP_SCRATCH_W1),y            ; now the high byte
+            sta  _jmpfar+2
+            iny
+            lda  (P8ZP_SCRATCH_W1),y            ; then the target bank
+            sta  P8ZP_SCRATCH_B1
+
+            ; adjust return address to skip over the arguments
+            clc
+            lda  P8ZP_SCRATCH_W1
+            adc  #3
+            sta  P8ZP_SCRATCH_W1
+            lda  P8ZP_SCRATCH_W1+1
+            adc  #0
+            pha
+            lda  P8ZP_SCRATCH_W1
+            pha
+            lda  $01        ; save old ram banks
+            pha
+            ; set target bank, restore A, Y and flags
+            lda  P8ZP_SCRATCH_REG
+            pha
+            lda  P8ZP_SCRATCH_B1
+            jsr  banks
+            lda  P8ZP_SCRATCH_W2
+            ldy  P8ZP_SCRATCH_W2+1
+            plp
+            jsr  _jmpfar        ; do the actual call
+            ; restore bank without clobbering status flags and A register
+            sta  P8ZP_SCRATCH_W1
+            php
+            pla
+            sta  P8ZP_SCRATCH_B1
+            pla
+            jsr  banks
+            lda  P8ZP_SCRATCH_B1
+            pha
+            lda  P8ZP_SCRATCH_W1
+            plp
+            rts
+
+_jmpfar     jmp  $0000          ; modified
+        }}
+    }
 
     sub get_vic_memory_base() -> uword {
         ; one of the 4 possible banks. $0000/$4000/$8000/$c000.
@@ -345,7 +432,7 @@ c64 {
 sys {
     ; ------- lowlevel system routines --------
 
-    const ubyte target = 64         ;  compilation target specifier.  64 = C64, 128 = C128, 16 = CommanderX16.
+    const ubyte target = 64         ;  compilation target specifier.  255=virtual, 128=C128, 64=C64, 32=PET, 16=CommanderX16, 8=atari800XL, 7=Neo6502
 
     const ubyte sizeof_bool = 1
     const ubyte sizeof_byte = 1
@@ -354,83 +441,13 @@ sys {
     const ubyte sizeof_uword = 2
     const ubyte sizeof_float = 5
 
+    sub  disable_runstop_and_charsetswitch() {
+        p8_sys_startup.disable_runstop_and_charsetswitch()
+    }
 
-asmsub  init_system()  {
-    ; Initializes the machine to a sane starting state.
-    ; Called automatically by the loader program logic.
-    ; This means that the BASIC, KERNAL and CHARGEN ROMs are banked in,
-    ; the VIC, SID and CIA chips are reset, screen is cleared, and the default IRQ is set.
-    ; Also a different color scheme is chosen to identify ourselves a little.
-    ; Uppercase charset is activated.
-    %asm {{
-        sei
-        lda  #%00101111
-        sta  $00
-        lda  #%00100111
-        sta  $01
-        jsr  cbm.IOINIT
-        jsr  cbm.RESTOR
-        jsr  cbm.CINT
-        lda  #6
-        sta  c64.EXTCOL
-        lda  #7
-        sta  cbm.COLOR
-        lda  #0
-        sta  c64.BGCOL0
-        jsr  disable_runstop_and_charsetswitch
-        cli
-        rts
-    }}
-}
-
-asmsub  init_system_phase2()  {
-    %asm {{
-        cld
-        clc
-        clv
-        rts
-    }}
-}
-
-asmsub  cleanup_at_exit() {
-    ; executed when the main subroutine does rts
-    %asm {{
-        lda  #31
-        sta  $01        ; bank the kernal in
-        jsr  cbm.CLRCHN		; reset i/o channels
-        jsr  enable_runstop_and_charsetswitch
-_exitcodeCarry = *+1
-        lda  #0
-        lsr  a
-_exitcode = *+1
-        lda  #0        ; exit code possibly modified in exit()
-_exitcodeX = *+1
-        ldx  #0
-_exitcodeY = *+1
-        ldy  #0
-        rts
-    }}
-}
-
-asmsub  disable_runstop_and_charsetswitch() clobbers(A) {
-    %asm {{
-        lda  #$80
-        sta  657    ; disable charset switching
-        lda  #239
-        sta  808    ; disable run/stop key
-        rts
-    }}
-}
-
-asmsub  enable_runstop_and_charsetswitch() clobbers(A) {
-    %asm {{
-        lda  #0
-        sta  657    ; enable charset switching
-        lda  #237
-        sta  808    ; enable run/stop key
-        rts
-    }}
-}
+    sub  enable_runstop_and_charsetswitch() {
+        p8_sys_startup.enable_runstop_and_charsetswitch()
+    }
 
     asmsub save_prog8_internals() {
         %asm {{
@@ -451,6 +468,7 @@ save_SCRATCH_ZPB1	.byte  0
 save_SCRATCH_ZPREG	.byte  0
 save_SCRATCH_ZPWORD1	.word  0
 save_SCRATCH_ZPWORD2	.word  0
+            ; !notreached!
         }}
     }
 
@@ -724,6 +742,65 @@ _longcopy
         }}
     }
 
+    asmsub memcmp(uword address1 @R0, uword address2 @R1, uword size @AY) -> byte @A {
+        ; Compares two blocks of memory
+        ; Returns -1 (255), 0 or 1, meaning: block 1 sorts before, equal or after block 2.
+        %asm {{
+            sta  P8ZP_SCRATCH_REG   ; lsb(size)
+            sty  P8ZP_SCRATCH_B1    ; msb(size)
+            lda  cx16.r0
+            ldy  cx16.r0+1
+            sta  P8ZP_SCRATCH_W1
+            sty  P8ZP_SCRATCH_W1+1
+            lda  cx16.r1
+            ldy  cx16.r1+1
+            sta  P8ZP_SCRATCH_W2
+            sty  P8ZP_SCRATCH_W2+1
+
+            ldx  P8ZP_SCRATCH_B1
+            beq  _no_msb_size
+
+_loop_msb_size
+            ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            cmp  (P8ZP_SCRATCH_W2),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            bne  -
+            inc  P8ZP_SCRATCH_W1+1
+            inc  P8ZP_SCRATCH_W2+1
+            dec  P8ZP_SCRATCH_B1        ; msb(size) -= 1
+            dex
+            bne  _loop_msb_size
+
+_no_msb_size
+            lda  P8ZP_SCRATCH_REG       ; lsb(size)
+            bne  +
+            rts
+
++           ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            cmp  (P8ZP_SCRATCH_W2),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            cpy  P8ZP_SCRATCH_REG       ; lsb(size)
+            bne  -
+
+            lda #0
+            rts
+        }}
+    }
+
     inline asmsub read_flags() -> ubyte @A {
         %asm {{
         php
@@ -785,37 +862,37 @@ _longcopy
     asmsub exit(ubyte returnvalue @A) {
         ; -- immediately exit the program with a return code in the A register
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit2(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y) {
         ; -- immediately exit the program with result values in the A, X and Y registers.
         %asm {{
-            sta  cleanup_at_exit._exitcode
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit3(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y, bool carry @Pc) {
         ; -- immediately exit the program with result values in the A, X and Y registers, and the Carry flag in the status register.
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             lda  #0
             rol  a
-            sta  cleanup_at_exit._exitcodeCarry
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcodeCarry
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
@@ -826,6 +903,13 @@ _longcopy
         }}
     }
 
+    inline asmsub progstart() -> uword @AY {
+        %asm {{
+            lda  #<prog8_program_start
+            ldy  #>prog8_program_start
+        }}
+    }
+
     inline asmsub push(ubyte value @A) {
         %asm {{
             pha
@@ -975,6 +1059,7 @@ cx16 {
     _cx16_vreg_storage
             .word 0,0,0,0,0,0,0,0
             .word 0,0,0,0,0,0,0,0
+            ; !notreached!
         }}
     }
 
@@ -995,3 +1080,101 @@ cx16 {
     }
 
 }
+
+p8_sys_startup {
+    ; program startup and shutdown machinery. Needs to reside in normal system ram.
+
+asmsub  init_system()  {
+    ; Initializes the machine to a sane starting state.
+    ; Called automatically by the loader program logic.
+    ; This means that the KERNAL and CHARGEN ROMs are banked in,
+    ; BASIC ROM is NOT banked in (so we have another 8Kb of RAM at our disposal),
+    ; the VIC, SID and CIA chips are reset, screen is cleared, and the default IRQ is set.
+    ; Also a different color scheme is chosen to identify ourselves a little.
+    ; Uppercase charset is activated.
+    %asm {{
+        sei
+        lda  #%00101111
+        sta  $00
+        lda  #%00100110   ; kernal and i/o banked in, basic off
+        sta  $01
+        jsr  cbm.IOINIT
+        jsr  cbm.RESTOR
+        jsr  cbm.CINT
+        lda  #6
+        sta  c64.EXTCOL
+        lda  #7
+        sta  cbm.COLOR
+        lda  #0
+        sta  c64.BGCOL0
+        jsr  disable_runstop_and_charsetswitch
+        lda  #PROG8_C64_BANK_CONFIG     ; apply bank config
+        sta  $01
+        and  #1
+        bne  +
+        ; basic is not banked in, adjust MEMTOP
+        ldx  #<$d000
+        ldy  #>$d000
+        clc
+        jsr  cbm.MEMTOP
++       cli
+        rts
+    }}
+}
+
+asmsub  init_system_phase2()  {
+    %asm {{
+        cld
+        clc
+        clv
+        rts
+    }}
+}
+
+asmsub  cleanup_at_exit() {
+    ; executed when the main subroutine does rts
+    %asm {{
+        lda  #%00101111
+        sta  $00
+        lda  #31
+        sta  $01            ; bank the kernal and basic in
+        ldx  #<$a000
+        ldy  #>$a000
+        clc
+        jsr  cbm.MEMTOP     ; adjust MEMTOP down again
+        jsr  cbm.CLRCHN		; reset i/o channels
+        jsr  enable_runstop_and_charsetswitch
+_exitcodeCarry = *+1
+        lda  #0
+        lsr  a
+_exitcode = *+1
+        lda  #0        ; exit code possibly modified in exit()
+_exitcodeX = *+1
+        ldx  #0
+_exitcodeY = *+1
+        ldy  #0
+        rts
+    }}
+}
+
+asmsub  disable_runstop_and_charsetswitch() clobbers(A) {
+    %asm {{
+        lda  #$80
+        sta  657    ; disable charset switching
+        lda  #239
+        sta  808    ; disable run/stop key
+        rts
+    }}
+}
+
+asmsub  enable_runstop_and_charsetswitch() clobbers(A) {
+    %asm {{
+        lda  #0
+        sta  657    ; enable charset switching
+        lda  #237
+        sta  808    ; enable run/stop key
+        rts
+    }}
+}
+
+}

compiler/res/prog8lib/c64/textio.p8

@@ -13,7 +13,7 @@ txt {
 const ubyte DEFAULT_WIDTH = 40
 const ubyte DEFAULT_HEIGHT = 25
 
-romsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
+extsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
 
 sub  clear_screen() {
     chrout(147)
@@ -317,6 +317,7 @@ _mod		sta  $ffff		; modified
 		rts
 
 _screenrows	.word  cbm.Screen + range(0, 1000, 40)
+        ; !notreached!
 	}}
 }
 
@@ -360,6 +361,7 @@ _mod		sta  $ffff		; modified
 		rts
 
 _colorrows	.word  $d800 + range(0, 1000, 40)
+        ; !notreached!
 	}}
 }
 

compiler/res/prog8lib/conv.p8

@@ -200,6 +200,7 @@ _allzero    lda  #'0'
         sta  string_out,x
         inx
         bne  _end
+        ; !notreached!
 	}}
 }
 
@@ -402,6 +403,7 @@ _digit
 		bne  _parse
 		; never reached
 _negative	.byte  0
+        ; !notreached!
 	}}
 }
 
@@ -472,6 +474,7 @@ _try_iso
         bcs  _stop
         and  #63
         bne  _add_letter
+		; !notreached!
 	}}
 }
 
@@ -702,7 +705,7 @@ decHundreds		.byte  0
 decTens			.byte  0
 decOnes   		.byte  0
 			.byte  0		; zero-terminate the decimal output string
-
+		; !notreached!
     }}
 }
 
@@ -736,6 +739,7 @@ asmsub  internal_ubyte2hex  (ubyte value @A) clobbers(X) -> ubyte @A, ubyte @Y
 		rts
 
 _hex_digits	.text "0123456789abcdef"	; can probably be reused for other stuff as well
+		; !notreached!
 	}}
 }
 
@@ -753,6 +757,7 @@ asmsub  internal_uword2hex  (uword value @AY) clobbers(A,Y)  {
 		sty  output+3
 		rts
 output		.text  "0000", $00      ; 0-terminated output buffer (to make printing easier)
+		; !notreached!
 	}}
 }
 

compiler/res/prog8lib/cx16/audio.p8

@@ -1,459 +0,0 @@
-%import syslib
-
-; Stubs for the audio routines in rom bank 10
-; They use JSRFAR to select the correct bank, so user code doesn't have to bother.
-;
-; The stubs here still use the romsub definitions of the routines from syslib, but those are
-; just the jump addresses and are unaware of the kernal bank. Wrapping in JSRFAR fixes this.
-
-audio {
-    const ubyte rom_bank = $0a
-    asmsub audio_init() clobbers(A,X,Y) -> bool @Pc { ; (re)initialize both vera PSG and YM audio chips
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.audio_init
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_fmfreq(ubyte channel @A, uword freq @XY, bool noretrigger @Pc) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_fmfreq
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_fmnote(ubyte channel @A, ubyte note @X, ubyte fracsemitone @Y, bool noretrigger @Pc) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_fmnote
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_fmplaystring(ubyte length @A, str string @XY) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_fmplaystring
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_fmvib(ubyte speed @A, ubyte depth @X) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_fmvib
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_playstringvoice(ubyte channel @A) clobbers(Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_playstringvoice
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_psgfreq(ubyte voice @A, uword freq @XY) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_psgfreq
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_psgnote(ubyte voice @A, ubyte note @X, ubyte fracsemitone @Y) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_psgnote
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_psgwav(ubyte voice @A, ubyte waveform @X) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_psgwav
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_psgplaystring(ubyte length @A, str string @XY) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_psgplaystring
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_fmchordstring(ubyte length @A, str string @XY) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_fmchordstring
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub bas_psgchordstring(ubyte length @A, str string @XY) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.bas_psgchordstring
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_bas2fm(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_bas2fm
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_bas2midi(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_bas2midi
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_bas2psg(ubyte note @X, ubyte fracsemitone @Y) clobbers(A) -> uword @XY, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_bas2psg
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_fm2bas(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_fm2bas
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_fm2midi(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_fm2midi
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_fm2psg(ubyte note @X, ubyte fracsemitone @Y) clobbers(A) -> uword @XY, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_fm2psg
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_freq2bas(uword freqHz @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_freq2bas
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_freq2fm(uword freqHz @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_freq2fm
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_freq2midi(uword freqHz @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_freq2midi
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_freq2psg(uword freqHz @XY) clobbers(A) -> uword @XY, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_freq2psg
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_midi2bas(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_midi2bas
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_midi2fm(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_midi2fm
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_midi2psg(ubyte note @X, ubyte fracsemitone @Y) clobbers(A) -> uword @XY, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_midi2psg
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_psg2bas(uword freq @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_psg2bas
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_psg2fm(uword freq @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_psg2fm
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub notecon_psg2midi(uword freq @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.notecon_psg2midi
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_init() clobbers(A,X,Y) {              ; (re)init Vera PSG
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_init
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_playfreq(ubyte voice @A, uword freq @XY) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_playfreq
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_read(ubyte offset @X, bool cookedVol @Pc) clobbers(Y) -> ubyte @A{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_read
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_setatten(ubyte voice @A, ubyte attenuation @X) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_setatten
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_setfreq(ubyte voice @A, uword freq @XY) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_setfreq
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_setpan(ubyte voice @A, ubyte panning @X) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_setpan
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_setvol(ubyte voice @A, ubyte volume @X) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_setvol
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_write(ubyte value @A, ubyte offset @X) clobbers(Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_write
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_write_fast(ubyte value @A, ubyte offset @X) clobbers(Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_write_fast
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_getatten(ubyte voice @A) clobbers(Y) -> ubyte @X{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_getatten
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub psg_getpan(ubyte voice @A) clobbers(Y) -> ubyte @X{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.psg_getpan
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_init() clobbers(A,X,Y) -> bool @Pc {            ; (re)init YM chip
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_init
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_loaddefpatches() clobbers(A,X,Y) -> bool @Pc {   ; load default YM patches
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_loaddefpatches
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_loadpatch(ubyte channel @A, uword patchOrAddress @XY, bool what @Pc) clobbers(A,X,Y){
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_loadpatch
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_loadpatchlfn(ubyte channel @A, ubyte lfn @X) clobbers(X,Y) -> ubyte @A, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_loadpatchlfn
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_playdrum(ubyte channel @A, ubyte note @X) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_playdrum
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_playnote(ubyte channel @A, ubyte kc @X, ubyte kf @Y, bool notrigger @Pc) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_playnote
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_setatten(ubyte channel @A, ubyte attenuation @X) clobbers(Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_setatten
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_setdrum(ubyte channel @A, ubyte note @X) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_setdrum
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_setnote(ubyte channel @A, ubyte kc @X, ubyte kf @Y) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_setnote
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_setpan(ubyte channel @A, ubyte panning @X) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_setpan
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_read(ubyte register @X, bool cooked @Pc) clobbers(Y) -> ubyte @A, bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_read
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_release(ubyte channel @A) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_release
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_trigger(ubyte channel @A, bool noRelease @Pc) clobbers(A,X,Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_trigger
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_write(ubyte value @A, ubyte register @X) clobbers(Y) -> bool @Pc{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_write
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_getatten(ubyte channel @A) clobbers(Y) -> ubyte @X{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_getatten
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_getpan(ubyte channel @A) clobbers(Y) -> ubyte @X{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_getpan
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-    asmsub ym_get_chip_type() clobbers(X) -> ubyte @A{
-        %asm{{
-            jsr cx16.JSRFAR
-            .word cx16.ym_get_chip_type
-            .byte p8c_rom_bank
-            rts
-        }}
-    }
-}

compiler/res/prog8lib/cx16/bmx.p8

@@ -1,4 +1,5 @@
 ; Routines to load and save "BMX" files (commander X16 bitmap format) Version 1.
+; Bitmap data is loaded directly into VRAM without intermediary buffering.
 ; Only uncompressed images are supported for now.
 ; BMX Specification: https://cx16forum.com/forum/viewtopic.php?t=6945
 

compiler/res/prog8lib/cx16/diskio.p8

@@ -145,6 +145,7 @@ io_error:
     uword list_pattern
     uword list_blocks
     bool iteration_in_progress = false
+    bool write_iteration_in_progress = false
     str list_filetype = "???"       ; prg, seq, dir
     str list_filename = "?" * 50
 
@@ -423,6 +424,7 @@ _end        jsr  cbm.READST
 
     sub f_close() {
         ; -- end an iterative file loading session (close channels).
+        ;    it is safe to call this multiple times, or when no file is open for reading.
         if iteration_in_progress {
             cbm.CLRCHN()
             cbm.CLOSE(READ_IO_CHANNEL)
@@ -460,7 +462,10 @@ _end        jsr  cbm.READST
         cbm.SETLFS(WRITE_IO_CHANNEL, drivenumber, WRITE_IO_CHANNEL)
         void cbm.OPEN()             ; open 13,8,13,"filename"
         if_cc {
-            return cbm.READST()==0
+            if cbm.READST()==0 {
+                write_iteration_in_progress = true
+                return true
+            }
         }
         cbm.CLOSE(WRITE_IO_CHANNEL)
         f_close_w()
@@ -498,8 +503,12 @@ no_mciout:
 
     sub f_close_w() {
         ; -- end an iterative file writing session (close channels).
-        cbm.CLRCHN()
-        cbm.CLOSE(WRITE_IO_CHANNEL)
+        ;    it is safe to call this multiple times, or when no file is open for reading.
+        if write_iteration_in_progress {
+            cbm.CLRCHN()
+            cbm.CLOSE(WRITE_IO_CHANNEL)
+            write_iteration_in_progress = false
+        }
     }
 
 
@@ -703,7 +712,7 @@ io_error:
     ; CommanderX16 extensions over the basic C64/C128 diskio routines:
 
     ; For use directly after a load or load_raw call (don't mess with the ram bank yet):
-    ; Calculates the number of bytes loaded (files > 64Kb ar truncated to 16 bits)
+    ; Calculates the number of bytes loaded (files > 64Kb are truncated to 16 bits)
     sub load_size(ubyte startbank, uword startaddress, uword endaddress) -> uword {
         return $2000 * (cx16.getrambank() - startbank) + endaddress - startaddress
     }
@@ -869,6 +878,15 @@ io_error:
         send_command(list_filename)
     }
 
+    sub exists(str filename) -> bool {
+        ; -- returns true if the given file exists on the disk, otherwise false
+        if f_open(filename) {
+            f_close()
+            return true
+        }
+        return false
+    }
+
     sub f_seek(uword pos_hiword, uword pos_loword) {
         ; -- seek in the reading file opened with f_open, to the given 32-bits position
         ;    Note: this will not work if you have already read the last byte of the file! Then you must close and reopen the file first.

compiler/res/prog8lib/cx16/floats.p8

@@ -14,69 +14,82 @@ floats {
 ; note: fac1/2 might get clobbered even if not mentioned in the function's name.
 ; note: for subtraction and division, the left operand is in fac2, the right operand in fac1.
 
-romsub $fe00 = AYINT() clobbers(A,X,Y)          ; fac1-> signed word in 'facmo' and 'faclo', MSB FIRST. (might throw ILLEGAL QUANTITY)  See "basic.sym" kernal symbol file for their memory locations.
+extsub $fe00 = AYINT() clobbers(A,X,Y)          ; fac1-> signed word in 'facmo' and 'faclo', MSB FIRST. DON'T USE THIS, USE WRAPPER 'AYINT2' INSTEAD.  (might throw ILLEGAL QUANTITY)
 
 ; GIVAYF: signed word in Y/A (note different lsb/msb order) -> float in fac1
 ; there is also floats.GIVUAYFAY - unsigned word in A/Y (lo/hi) to fac1
 ; (tip: use GIVAYFAY to use A/Y input; lo/hi switched to normal order)
-romsub $fe03 = GIVAYF(ubyte lo @ Y, ubyte hi @ A) clobbers(A,X,Y)
+extsub $fe03 = GIVAYF(ubyte lo @ Y, ubyte hi @ A) clobbers(A,X,Y)
 
-romsub $fe06 = FOUT() clobbers(X) -> uword @ AY             ; fac1 -> string, address returned in AY
-romsub $fe09 = VAL_1(uword string @XY, ubyte length @A) clobbers(A,X,Y) -> float @FAC1      ; convert ASCII string in XY and length in A, to floating point in FAC1. WARNING: only implemented in ROM 47+. Safer to use floats.parse() instead.
+extsub $fe06 = FOUT() clobbers(X) -> uword @ AY             ; fac1 -> string, address returned in AY
+extsub $fe09 = VAL_1(uword string @XY, ubyte length @A) clobbers(A,X,Y) -> float @FAC1      ; convert ASCII string in XY and length in A, to floating point in FAC1. WARNING: only implemented in ROM 47+. Safer to use floats.parse() instead.
 
 ; GETADR: fac1 -> unsigned word in Y/A (might throw ILLEGAL QUANTITY) (result also in $14/15)
 ; (tip: use GETADRAY to get A/Y output; lo/hi switched to normal little endian order)
-romsub $fe0c = GETADR() clobbers(X) -> ubyte @ Y, ubyte @ A
-romsub $fe0f = FLOATC() clobbers(A,X,Y)                     ; convert address to floating point
+extsub $fe0c = GETADR() clobbers(X) -> ubyte @ Y, ubyte @ A
+extsub $fe0f = FLOATC() clobbers(A,X,Y)                     ; convert address to floating point
 
-romsub $fe12 = FSUB(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = mflpt from A/Y - fac1
-romsub $fe15 = FSUBT() clobbers(A,X,Y)                      ; fac1 = fac2-fac1   mind the order of the operands
-romsub $fe18 = FADD(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 += mflpt value from A/Y
-romsub $fe1b = FADDT() clobbers(A,X,Y)                      ; fac1 += fac2
-romsub $fe1e = FMULT(uword mflpt @ AY) clobbers(A,X,Y)      ; fac1 *= mflpt value from A/Y
-romsub $fe21 = FMULTT() clobbers(A,X,Y)                     ; fac1 *= fac2
-romsub $fe24 = FDIV(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = mflpt in A/Y / fac1
-romsub $fe27 = FDIVT() clobbers(A,X,Y)                      ; fac1 = fac2/fac1  mind the order of the operands
-romsub $fe2a = LOG() clobbers(A,X,Y)                        ; fac1 = LN(fac1)  (natural log)
-romsub $fe2d = INT() clobbers(A,X,Y)                        ; INT() truncates, use FADDH first to integer round instead of trunc
-romsub $fe30 = SQR() clobbers(A,X,Y)                        ; fac1 = SQRT(fac1)
-romsub $fe33 = NEGOP() clobbers(A)                          ; switch the sign of fac1 (fac1 = -fac1)
-romsub $fe36 = FPWR(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = fac2 ** float in A/Y
-romsub $fe39 = FPWRT() clobbers(A,X,Y)                      ; fac1 = fac2 ** fac1
-romsub $fe3c = EXP() clobbers(A,X,Y)                        ; fac1 = EXP(fac1)  (e ** fac1)
-romsub $fe3f = COS() clobbers(A,X,Y)                        ; fac1 = COS(fac1)
-romsub $fe42 = SIN() clobbers(A,X,Y)                        ; fac1 = SIN(fac1)
-romsub $fe45 = TAN() clobbers(A,X,Y)                        ; fac1 = TAN(fac1)
-romsub $fe48 = ATN() clobbers(A,X,Y)                        ; fac1 = ATN(fac1)
-romsub $fe4b = ROUND() clobbers(A,X,Y)                      ; round least significant bit of fac1
-romsub $fe4e = ABS() clobbers(A,X,Y)                        ; fac1 = ABS(fac1)
-romsub $fe51 = SIGN() clobbers(X,Y) -> ubyte @ A            ; SIGN(fac1) to A, $ff, $0, $1 for negative, zero, positive
-romsub $fe54 = FCOMP(uword mflpt @ AY) clobbers(X,Y) -> ubyte @ A   ; A = compare fac1 to mflpt in A/Y, 0=equal 1=fac1 is greater, 255=fac1 is less than
-romsub $fe57 = RND_0() clobbers(A,X,Y)                      ; fac1 = RND(fac1) float random number generator  NOTE: incompatible with C64's RND routine
-romsub $fe57 = RND() clobbers(A,X,Y)                        ; alias for RND_0
-romsub $fe5a = CONUPK(uword mflpt @ AY) clobbers(A,X,Y)     ; load mflpt value from memory in A/Y into fac2
-romsub $fe5d = ROMUPK(uword mflpt @ AY) clobbers(A,X,Y)     ; load mflpt value from memory in current bank in A/Y into fac2
-romsub $fe60 = MOVFRM(uword mflpt @ AY) clobbers(A,X,Y)     ; load mflpt value from memory in A/Y into fac1  (use MOVFM instead)
-romsub $fe63 = MOVFM(uword mflpt @ AY) clobbers(A,X,Y)      ; load mflpt value from memory in A/Y into fac1
-romsub $fe66 = MOVMF(uword mflpt @ XY) clobbers(A,X,Y)      ; store fac1 to memory  X/Y as 5-byte mflpt
-romsub $fe69 = MOVFA() clobbers(A,X)                        ; copy fac2 to fac1
-romsub $fe6c = MOVAF() clobbers(A,X)                        ; copy fac1 to fac2  (rounded the least significant bit)
+extsub $fe12 = FSUB(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = mflpt from A/Y - fac1
+extsub $fe15 = FSUBT() clobbers(A,X,Y)                      ; fac1 = fac2-fac1   mind the order of the operands
+extsub $fe18 = FADD(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 += mflpt value from A/Y
+extsub $fe1b = FADDT() clobbers(A,X,Y)                      ; fac1 += fac2
+extsub $fe1e = FMULT(uword mflpt @ AY) clobbers(A,X,Y)      ; fac1 *= mflpt value from A/Y
+extsub $fe21 = FMULTT() clobbers(A,X,Y)                     ; fac1 *= fac2
+extsub $fe24 = FDIV(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = mflpt in A/Y / fac1
+extsub $fe27 = FDIVT() clobbers(A,X,Y)                      ; fac1 = fac2/fac1  mind the order of the operands
+extsub $fe2a = LOG() clobbers(A,X,Y)                        ; fac1 = LN(fac1)  (natural log)
+extsub $fe2d = INT() clobbers(A,X,Y)                        ; INT() truncates, use FADDH first to integer round instead of trunc
+extsub $fe30 = SQR() clobbers(A,X,Y)                        ; fac1 = SQRT(fac1)
+extsub $fe33 = NEGOP() clobbers(A)                          ; switch the sign of fac1 (fac1 = -fac1)
+extsub $fe36 = FPWR(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = fac2 ** float in A/Y
+extsub $fe39 = FPWRT() clobbers(A,X,Y)                      ; fac1 = fac2 ** fac1
+extsub $fe3c = EXP() clobbers(A,X,Y)                        ; fac1 = EXP(fac1)  (e ** fac1)
+extsub $fe3f = COS() clobbers(A,X,Y)                        ; fac1 = COS(fac1)
+extsub $fe42 = SIN() clobbers(A,X,Y)                        ; fac1 = SIN(fac1)
+extsub $fe45 = TAN() clobbers(A,X,Y)                        ; fac1 = TAN(fac1)
+extsub $fe48 = ATN() clobbers(A,X,Y)                        ; fac1 = ATN(fac1)
+extsub $fe4b = ROUND() clobbers(A,X,Y)                      ; round least significant bit of fac1
+extsub $fe4e = ABS() clobbers(A,X,Y)                        ; fac1 = ABS(fac1)
+extsub $fe51 = SIGN() clobbers(X,Y) -> ubyte @ A            ; SIGN(fac1) to A, $ff, $0, $1 for negative, zero, positive
+extsub $fe54 = FCOMP(uword mflpt @ AY) clobbers(X,Y) -> ubyte @ A   ; A = compare fac1 to mflpt in A/Y, 0=equal 1=fac1 is greater, 255=fac1 is less than
+extsub $fe57 = RND_0() clobbers(A,X,Y)                      ; fac1 = RND(fac1) float random number generator  NOTE: incompatible with C64's RND routine
+extsub $fe57 = RND() clobbers(A,X,Y)                        ; alias for RND_0
+extsub $fe5a = CONUPK(uword mflpt @ AY) clobbers(A,X,Y)     ; load mflpt value from memory in A/Y into fac2
+extsub $fe5d = ROMUPK(uword mflpt @ AY) clobbers(A,X,Y)     ; load mflpt value from memory in current bank in A/Y into fac2
+extsub $fe60 = MOVFRM(uword mflpt @ AY) clobbers(A,X,Y)     ; load mflpt value from memory in A/Y into fac1  (use MOVFM instead)
+extsub $fe63 = MOVFM(uword mflpt @ AY) clobbers(A,X,Y)      ; load mflpt value from memory in A/Y into fac1
+extsub $fe66 = MOVMF(uword mflpt @ XY) clobbers(A,X,Y)      ; store fac1 to memory  X/Y as 5-byte mflpt
+extsub $fe69 = MOVFA() clobbers(A,X)                        ; copy fac2 to fac1
+extsub $fe6c = MOVAF() clobbers(A,X)                        ; copy fac1 to fac2  (rounded the least significant bit)
 
 ; X16 additions
-romsub $fe6f = FADDH() clobbers(A,X,Y)                      ; fac1 += 0.5, for integer rounding- call this before INT
-romsub $fe72 = ZEROFC() clobbers(A,X,Y)                     ; fac1 = 0
-romsub $fe75 = NORMAL() clobbers(A,X,Y)                     ; normalize fac1
-romsub $fe78 = NEGFAC() clobbers(A)                         ; switch the sign of fac1 (fac1 = -fac1) (doesn't work, juse use NEGOP() instead!)
-romsub $fe7b = MUL10() clobbers(A,X,Y)                      ; fac1 *= 10
-romsub $fe7e = DIV10() clobbers(A,X,Y)                      ; fac1 /= 10 , CAUTION: result is always positive! Have to restore sign manually!
-romsub $fe81 = MOVEF() clobbers(A,X)                        ; copy fac1 to fac2
-romsub $fe84 = SGN() clobbers(A,X,Y)                        ; fac1 = SGN(fac1), result of SIGN (-1, 0 or 1)
-romsub $fe87 = FLOAT() clobbers(A,X,Y)                      ; FAC = (s8).A
-romsub $fe8a = FLOATS() clobbers(A,X,Y)                     ; FAC = (s16)facho+1:facho
-romsub $fe8d = QINT() clobbers(A,X,Y)                       ; facho:facho+1:facho+2:facho+3 = u32(FAC)
-romsub $fe90 = FINLOG(byte value @A) clobbers (A, X, Y)     ; fac1 += signed byte in A
+extsub $fe6f = FADDH() clobbers(A,X,Y)                      ; fac1 += 0.5, for integer rounding- call this before INT
+extsub $fe72 = ZEROFC() clobbers(A,X,Y)                     ; fac1 = 0
+extsub $fe75 = NORMAL() clobbers(A,X,Y)                     ; normalize fac1
+extsub $fe78 = NEGFAC() clobbers(A)                         ; switch the sign of fac1 (fac1 = -fac1) (doesn't work, juse use NEGOP() instead!)
+extsub $fe7b = MUL10() clobbers(A,X,Y)                      ; fac1 *= 10
+extsub $fe7e = DIV10() clobbers(A,X,Y)                      ; fac1 /= 10 , CAUTION: result is always positive! Have to restore sign manually!
+extsub $fe81 = MOVEF() clobbers(A,X)                        ; copy fac1 to fac2
+extsub $fe84 = SGN() clobbers(A,X,Y)                        ; fac1 = SGN(fac1), result of SIGN (-1, 0 or 1)
+extsub $fe87 = FLOAT() clobbers(A,X,Y)                      ; FAC = (s8).A
+extsub $fe8a = FLOATS() clobbers(A,X,Y)                     ; FAC = (s16)facho+1:facho
+extsub $fe8d = QINT() clobbers(A,X,Y)                       ; facho:facho+1:facho+2:facho+3 = u32(FAC)
+extsub $fe90 = FINLOG(byte value @A) clobbers (A, X, Y)     ; fac1 += signed byte in A
 
 
+asmsub  AYINT2() clobbers(X) -> word @AY {
+    ; fac1-> signed word in AY. Safe wrapper around the AYINT kernal routine (not reading internal memory locations)
+    ; (might throw ILLEGAL QUANTITY)
+    %asm {{
+		jsr  AYINT
+		ldx  #<floats_temp_var
+		ldy  #>floats_temp_var
+		jsr  MOVMF
+		lda  floats_temp_var+4
+		ldy  floats_temp_var+3
+        rts
+    }}
+}
 
 asmsub  FREADSA  (byte value @A) clobbers(A,X,Y) {
     ; ---- 8 bit signed A -> float in fac1
@@ -124,6 +137,8 @@ asmsub FREADU24AXY(ubyte lo @ A, ubyte mid @ X, ubyte hi @ Y) clobbers(A, X, Y)
 
 asmsub  GIVUAYFAY  (uword value @ AY) clobbers(A,X,Y)  {
 	; ---- unsigned 16 bit word in A/Y (lo/hi) to fac1
+	; See "basic.sym" kernal symbol file for the facmo memory location.
+	; TODO find a way to not depend on that internal fac memory location. MOVFM doesn't work.
 	%asm {{
 	sty  $c4        ; facmo
 	sta  $c5        ; facmo+1
@@ -186,10 +201,10 @@ _borked
 +       jmp  sys.exit
 
 _msg    .text 13,"?rom 47+ required for val1",13,0
+        ; !notreached!
     }}
 }
 
-&uword AYINT_facmo = $c6      ; $c6/$c7 contain result of AYINT   (See "basic.sym" kernal symbol file)
 
 sub rnd() -> float {
     %asm {{

compiler/res/prog8lib/cx16/gfx_hires.p8

@@ -1,117 +1,49 @@
 ; Bitmap pixel graphics routines for the CommanderX16
-; Custom routines to use the full-screen 640x480 and 320x240 screen modes.
+; Custom routines to use the full-screen 640x480 in 4 color mode.
+; (monochrome mode can be used via the monogfx module)
 ; (These modes are not supported by the documented GRAPH_xxxx kernal routines)
 ;
 ; No text layer is currently shown, text can be drawn as part of the bitmap itself.
 ; Note: for similar graphics routines that also work on the C-64, use the "graphics" module instead.
 ; Note: for identical routines for a monochrome 1 bpp screen, use the "monogfx" module instead.
 ; Note: for color palette manipulation, use the "palette" module or write Vera registers yourself.
-; Note: this library implements code for various resolutions and color depths. This takes up memory.
-;       If you're memory constrained you should probably not use this built-in library,
-;       but make a copy in your project only containing the code for the required resolution.
 ;
-; NOTE: For sake of speed, NO BOUNDS CHECKING is performed in most routines!
-;       You'll have to make sure yourself that you're not writing outside of bitmap boundaries!
+; NOTE: the bitmap screen data is positioned in vram at $00000
+; This is compatible with the CX16's screen mode 128.  (void cx16.set_screen_mode(128))
 ;
-; NOTE: the bitmap screen data is positioned in vram at $0:0000
-;
-; NOTE: In the future, these routines might be split out to separate modules, 1 for each screen mode,
-;       so they can be optimized a lot better.  There's already a "gfx_lores" module with a few routines for lores 256C mode.
-;
-; SCREEN MODE LIST:
-;   mode 0 = reset back to default text mode
-;   mode 1 = bitmap 320 x 240 x 256c (8 bpp)
-;   mode 2 = bitmap 640 x 480 x 4c (2 bpp)
-;   mode 3 = bitmap 320 x 240 x 4c (not yet implemented: just use 256c, there's enough vram for that)
-;   mode 4 = bitmap 320 x 240 x 16c (not yet implemented: just use 256c, there's enough vram for that)
-;   mode 5 = bitmap 640 x 400 x 16c (not yet implemented)
 
-gfx2 {
+gfx_hires {
 
     %option ignore_unused
 
-    ; read-only control variables:
-    ubyte active_mode = 0
-    uword width = 0
-    uword height = 0
-    ubyte bpp = 0
+    const uword WIDTH = 640
+    const uword HEIGHT = 480
 
-    sub screen_mode(ubyte mode) {
+    sub graphics_mode() {
+        ; set hires 4c mode
         cx16.VERA_CTRL=0
-        when mode {
-            1 -> {
-                ; lores 256c
-                cx16.VERA_DC_VIDEO = (cx16.VERA_DC_VIDEO & %11001111) | %00100000      ; enable only layer 1
-                cx16.VERA_DC_HSCALE = 64
-                cx16.VERA_DC_VSCALE = 64
-                cx16.VERA_L1_CONFIG = %00000111
-                cx16.VERA_L1_MAPBASE = 0
-                cx16.VERA_L1_TILEBASE = 0
-            }
-            2 -> {
-                ; highres 4c
-                cx16.VERA_DC_VIDEO = (cx16.VERA_DC_VIDEO & %11001111) | %00100000      ; enable only layer 1
-                cx16.VERA_DC_HSCALE = 128
-                cx16.VERA_DC_VSCALE = 128
-                cx16.VERA_L1_CONFIG = %00000101
-                cx16.VERA_L1_MAPBASE = 0
-                cx16.VERA_L1_TILEBASE = %00000001
-            }
-            else -> {
-                ; back to default text mode
-                if active_mode!=0 {
-                    cx16.r15L = cx16.VERA_DC_VIDEO & %00000111 ; retain chroma + output mode
-                    cbm.CINT()
-                    cx16.VERA_DC_VIDEO = (cx16.VERA_DC_VIDEO & %11111000) | cx16.r15L
-                }
-            }
-        }
-
-        init_mode(mode)
-        if active_mode!=0
-            clear_screen(0)
+        cx16.VERA_DC_VIDEO = (cx16.VERA_DC_VIDEO & %11001111) | %00100000      ; enable only layer 1
+        cx16.VERA_DC_HSCALE = 128
+        cx16.VERA_DC_VSCALE = 128
+        cx16.VERA_L1_CONFIG = %00000101
+        cx16.VERA_L1_MAPBASE = 0
+        cx16.VERA_L1_TILEBASE = %00000001
+        clear_screen(0)
     }
 
-    sub init_mode(ubyte mode) {
-        ; set the internal configuration variables corresponding to the given screenmode
-        ; doesn't manipulate Vera / the actual display mode
-        active_mode = mode
-        when mode {
-            1 -> {
-                width = 320
-                height = 240
-                bpp = 8
-            }
-            2 -> {
-                width = 640
-                height = 480
-                bpp = 2
-            }
-            else -> {
-                width = 0
-                height = 0
-                bpp = 0
-                active_mode = 0
-            }
-        }
+    sub text_mode() {
+        ; back to normal text mode
+        cx16.r15L = cx16.VERA_DC_VIDEO & %00000111 ; retain chroma + output mode
+        cbm.CINT()
+        cx16.VERA_DC_VIDEO = (cx16.VERA_DC_VIDEO & %11111000) | cx16.r15L
     }
 
     sub clear_screen(ubyte color) {
         position(0, 0)
-        when active_mode {
-            1 -> {
-                ; lores 256c
-                repeat 240/2
-                    cs_innerloop640(color)
-            }
-            2 -> {
-                ; highres 4c
-                ubyte[] colors = [%00000000, %01010101, %10101010, %11111111]
-                color = colors[color&3]
-                repeat 480/4
-                    cs_innerloop640(color)
-            }
-        }
+        ubyte[] colors = [%00000000, %01010101, %10101010, %11111111]
+        color = colors[color&3]
+        repeat 480/4
+            cs_innerloop640(color)
         position(0, 0)
     }
 
@@ -128,6 +60,18 @@ gfx2 {
         vertical_line(xx+rwidth-1, yy+1, rheight-2, color)
     }
 
+    sub safe_rect(uword xx, uword yy, uword rwidth, uword rheight, ubyte color) {
+        ; does bounds checking and clipping
+        safe_horizontal_line(xx, yy, rwidth, color)
+        if rheight==1
+            return
+        safe_horizontal_line(xx, yy+rheight-1, rwidth, color)
+        safe_vertical_line(xx, yy+1, rheight-2, color)
+        if rwidth==1
+            return
+        safe_vertical_line(xx+rwidth-1, yy+1, rheight-2, color)
+    }
+
     sub fillrect(uword xx, uword yy, uword rwidth, uword rheight, ubyte color) {
         ; Draw a filled rectangle of the given size and color.
         ; To fill the whole screen, use clear_screen(color) instead - it is much faster.
@@ -139,140 +83,124 @@ gfx2 {
         }
     }
 
+    sub safe_fillrect(uword xx, uword yy, uword rwidth, uword rheight, ubyte color) {
+        ; Draw a filled rectangle of the given size and color.
+        ; To fill the whole screen, use clear_screen(color) instead - it is much faster.
+        ; This safe version does bounds checking and clipping.
+        if xx>=WIDTH or yy>=WIDTH
+            return
+        if msb(xx)&$80!=0 {
+            rwidth += xx
+            xx = 0
+        }
+        if xx>=WIDTH
+            return
+        if xx+rwidth>WIDTH
+            rwidth = WIDTH-xx
+        if rwidth>WIDTH
+            return
+
+        if yy + rheight > HEIGHT
+            rheight = HEIGHT-yy
+        if rheight>HEIGHT
+            return
+
+        repeat rheight {
+            horizontal_line(xx, yy, rwidth, color)
+            yy++
+        }
+    }
+
     sub horizontal_line(uword xx, uword yy, uword length, ubyte color) {
         if length==0
             return
-        when active_mode {
-            1 -> {
-                ; lores 256c
-                position(xx, yy)
-                %asm {{
-                    lda  p8v_color
-                    ldx  p8v_length+1
-                    beq  +
-                    ldy  #0
--                   sta  cx16.VERA_DATA0
-                    iny
-                    bne  -
-                    dex
-                    bne  -
-+                   ldy  p8v_length     ; remaining
-                    beq  +
--                   sta  cx16.VERA_DATA0
-                    dey
-                    bne  -
-+
-                }}
-            }
-            2 -> {
-                ; highres 4c ....also mostly usable for lores 4c?
-                color &= 3
-                ubyte[4] colorbits
-                ubyte ii
-                for ii in 3 downto 0 {
-                    colorbits[ii] = color
-                    color <<= 2
-                }
-                void addr_mul_24_for_highres_4c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
-                %asm {{
-                    lda  cx16.VERA_ADDR_H
-                    and  #%00000111         ; no auto advance
-                    sta  cx16.VERA_ADDR_H
-                    stz  cx16.VERA_CTRL     ; setup vera addr 0
-                    lda  cx16.r1
-                    and  #1
-                    sta  cx16.VERA_ADDR_H
-                    lda  cx16.r0
-                    sta  cx16.VERA_ADDR_L
-                    lda  cx16.r0+1
-                    sta  cx16.VERA_ADDR_M
-                    ldx  p8v_xx
-                }}
+        ; highres 4c ....also mostly usable for lores 4c?
+        color &= 3
+        ubyte[4] colorbits
+        ubyte ii
+        for ii in 3 downto 0 {
+            colorbits[ii] = color
+            color <<= 2
+        }
+        void addr_mul_24_for_highres_4c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
+        %asm {{
+            lda  cx16.VERA_ADDR_H
+            and  #%00000111         ; no auto advance
+            sta  cx16.VERA_ADDR_H
+            stz  cx16.VERA_CTRL     ; setup vera addr 0
+            lda  cx16.r1
+            and  #1
+            sta  cx16.VERA_ADDR_H
+            lda  cx16.r0
+            sta  cx16.VERA_ADDR_L
+            lda  cx16.r0+1
+            sta  cx16.VERA_ADDR_M
+            ldx  p8v_xx
+        }}
 
-                repeat length {
-                    %asm {{
-                        txa
-                        and  #3
-                        tay
-                        lda  cx16.VERA_DATA0
-                        and  p8b_gfx2.p8s_plot.p8v_mask4c,y
-                        ora  p8v_colorbits,y
-                        sta  cx16.VERA_DATA0
-                        cpy  #%00000011         ; next vera byte?
-                        bne  ++
-                        inc  cx16.VERA_ADDR_L
-                        bne  ++
-                        inc  cx16.VERA_ADDR_M
+        repeat length {
+            %asm {{
+                txa
+                and  #3
+                tay
+                lda  cx16.VERA_DATA0
+                and  p8b_gfx_hires.p8s_plot.p8v_mask4c,y
+                ora  p8v_colorbits,y
+                sta  cx16.VERA_DATA0
+                cpy  #%00000011         ; next vera byte?
+                bne  ++
+                inc  cx16.VERA_ADDR_L
+                bne  ++
+                inc  cx16.VERA_ADDR_M
 +                       bne  +
-                        inc  cx16.VERA_ADDR_H
+                inc  cx16.VERA_ADDR_H
 +                       inx                     ; next pixel
-                    }}
-                }
-            }
+            }}
         }
     }
 
     sub safe_horizontal_line(uword xx, uword yy, uword length, ubyte color) {
         ; does bounds checking and clipping
-        if msb(yy)&$80!=0 or yy>=height
+        if msb(yy)&$80!=0 or yy>=HEIGHT
             return
         if msb(xx)&$80!=0 {
             length += xx
             xx = 0
         }
-        if xx>=width
+        if xx>=WIDTH
             return
-        if xx+length>width
-            length = width-xx
-        if length>width
+        if xx+length>WIDTH
+            length = WIDTH-xx
+        if length>WIDTH
             return
 
         horizontal_line(xx, yy, length, color)
     }
 
     sub vertical_line(uword xx, uword yy, uword lheight, ubyte color) {
-        when active_mode {
-            1 -> {
-                ; lores 256c
-                ; set vera auto-increment to 320 pixel increment (=next line)
-                position(xx,yy)
-                cx16.VERA_ADDR_H = cx16.VERA_ADDR_H & %00000111 | (14<<4)
-                %asm {{
-                    ldy  p8v_lheight
-                    beq  +
-                    lda  p8v_color
--                   sta  cx16.VERA_DATA0
-                    dey
-                    bne  -
-+
-                }}
-            }
-            2 -> {
-                ; highres 4c
-                ; use TWO vera adress pointers simultaneously one for reading, one for writing, so auto-increment is possible
-                if lheight==0
-                    return
-                position2(xx,yy,true)
-                set_both_strides(13)    ; 160 increment = 1 line in 640 px 4c mode
-                ;; color &= 3
-                ;; color <<= gfx2.plot.shift4c[lsb(xx) & 3]
-                cx16.r2L = lsb(xx) & 3
-                when color & 3 {
-                    0 -> color = 0
-                    1 -> color = gfx2.plot.shiftedleft_4c_1[cx16.r2L]
-                    2 -> color = gfx2.plot.shiftedleft_4c_2[cx16.r2L]
-                    3 -> color = gfx2.plot.shiftedleft_4c_3[cx16.r2L]
-                }
-                ubyte @shared mask = gfx2.plot.mask4c[lsb(xx) & 3]
-                repeat lheight {
-                    %asm {{
-                        lda  cx16.VERA_DATA0
-                        and  p8v_mask
-                        ora  p8v_color
-                        sta  cx16.VERA_DATA1
-                    }}
-                }
-            }
+        ; highres 4c
+        ; use TWO vera adress pointers simultaneously one for reading, one for writing, so auto-increment is possible
+        if lheight==0
+            return
+        position2(xx,yy,true)
+        set_both_strides(13)    ; 160 increment = 1 line in 640 px 4c mode
+        ;; color &= 3
+        ;; color <<= plot.shift4c[lsb(xx) & 3]
+        cx16.r2L = lsb(xx) & 3
+        when color & 3 {
+            0 -> color = 0
+            1 -> color = gfx_hires.plot.shiftedleft_4c_1[cx16.r2L]
+            2 -> color = gfx_hires.plot.shiftedleft_4c_2[cx16.r2L]
+            3 -> color = gfx_hires.plot.shiftedleft_4c_3[cx16.r2L]
+        }
+        ubyte @shared mask = gfx_hires.plot.mask4c[lsb(xx) & 3]
+        repeat lheight {
+            %asm {{
+                lda  cx16.VERA_DATA0
+                and  p8v_mask
+                ora  p8v_color
+                sta  cx16.VERA_DATA1
+            }}
         }
 
         sub set_both_strides(ubyte stride) {
@@ -285,6 +213,24 @@ gfx2 {
 
     }
 
+    sub safe_vertical_line(uword xx, uword yy, uword lheight, ubyte color) {
+        ; does bounds checking and clipping
+        if msb(yy)&$80!=0 or yy>=HEIGHT
+            return
+        if msb(xx)&$80!=0 or xx>=WIDTH
+            return
+        if msb(yy)&$80!=0 {
+            lheight += yy
+            yy = 0
+        }
+        if yy+lheight>HEIGHT
+            lheight = HEIGHT-yy
+        if lheight>HEIGHT
+            return
+
+        vertical_line(xx, yy, lheight, color)
+    }
+
     sub line(uword @zp x1, uword @zp y1, uword @zp x2, uword @zp y2, ubyte color) {
         ; Bresenham algorithm.
         ; This code special-cases various quadrant loops to allow simple ++ and -- operations.
@@ -564,107 +510,68 @@ gfx2 {
         ubyte[4] shiftedleft_4c_2 = [2<<6, 2<<4, 2<<2, 2<<0]
         ubyte[4] shiftedleft_4c_3 = [3<<6, 3<<4, 3<<2, 3<<0]
 
-        when active_mode {
-            1 -> {
-                ; lores 256c
-                void addr_mul_24_for_lores_256c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
-                %asm {{
-                    stz  cx16.VERA_CTRL
-                    lda  cx16.r1
-                    ora  #%00010000     ; enable auto-increment so next_pixel() can be used after this
-                    sta  cx16.VERA_ADDR_H
-                    lda  cx16.r0+1
-                    sta  cx16.VERA_ADDR_M
-                    lda  cx16.r0
-                    sta  cx16.VERA_ADDR_L
-                    lda  p8v_color
-                    sta  cx16.VERA_DATA0
-                }}
-            }
-            2 -> {
-                ; highres 4c   ....also mostly usable for lores 4c?
-                void addr_mul_24_for_highres_4c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
-                cx16.r2L = lsb(xx) & 3       ; xbits
-                ; color &= 3
-                ; color <<= shift4c[cx16.r2L]
-                when color & 3 {
-                    0 -> color = 0
-                    1 -> color = shiftedleft_4c_1[cx16.r2L]
-                    2 -> color = shiftedleft_4c_2[cx16.r2L]
-                    3 -> color = shiftedleft_4c_3[cx16.r2L]
-                }
-                %asm {{
-                    stz  cx16.VERA_CTRL
-                    lda  cx16.r1L
-                    sta  cx16.VERA_ADDR_H
-                    lda  cx16.r0H
-                    sta  cx16.VERA_ADDR_M
-                    lda  cx16.r0L
-                    sta  cx16.VERA_ADDR_L
-                    ldy  cx16.r2L           ; xbits
-                    lda  p8v_mask4c,y
-                    and  cx16.VERA_DATA0
-                    ora  p8v_color
-                    sta  cx16.VERA_DATA0
-                }}
-            }
+        ; highres 4c   ....also mostly usable for lores 4c?
+        void addr_mul_24_for_highres_4c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
+        cx16.r2L = lsb(xx) & 3       ; xbits
+        ; color &= 3
+        ; color <<= shift4c[cx16.r2L]
+        when color & 3 {
+            0 -> color = 0
+            1 -> color = shiftedleft_4c_1[cx16.r2L]
+            2 -> color = shiftedleft_4c_2[cx16.r2L]
+            3 -> color = shiftedleft_4c_3[cx16.r2L]
         }
+        %asm {{
+            stz  cx16.VERA_CTRL
+            lda  cx16.r1L
+            sta  cx16.VERA_ADDR_H
+            lda  cx16.r0H
+            sta  cx16.VERA_ADDR_M
+            lda  cx16.r0L
+            sta  cx16.VERA_ADDR_L
+            ldy  cx16.r2L           ; xbits
+            lda  p8v_mask4c,y
+            and  cx16.VERA_DATA0
+            ora  p8v_color
+            sta  cx16.VERA_DATA0
+        }}
     }
 
     sub safe_plot(uword xx, uword yy, ubyte color) {
         ; A plot that does bounds checks to see if the pixel is inside the screen.
         if msb(xx)&$80!=0 or msb(yy)&$80!=0
             return
-        if xx >= width or yy >= height
+        if xx >= WIDTH or yy >= HEIGHT
             return
         plot(xx, yy, color)
     }
 
     sub pget(uword @zp xx, uword yy) -> ubyte {
-        when active_mode {
-            1 -> {
-                ; lores 256c
-                void addr_mul_24_for_lores_256c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
-                %asm {{
-                    stz  cx16.VERA_CTRL
-                    lda  cx16.r1
-                    sta  cx16.VERA_ADDR_H
-                    lda  cx16.r0+1
-                    sta  cx16.VERA_ADDR_M
-                    lda  cx16.r0
-                    sta  cx16.VERA_ADDR_L
-                    lda  cx16.VERA_DATA0
-                }}
-            }
-            2 -> {
-                ; hires 4c
-                void addr_mul_24_for_highres_4c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
-                %asm {{
-                    stz  cx16.VERA_CTRL
-                    lda  cx16.r1L
-                    sta  cx16.VERA_ADDR_H
-                    lda  cx16.r0H
-                    sta  cx16.VERA_ADDR_M
-                    lda  cx16.r0L
-                    sta  cx16.VERA_ADDR_L
-                    lda  cx16.VERA_DATA0
-                    pha
-                    lda  p8v_xx
-                    and  #3
-                    tay
-                    lda  p8b_gfx2.p8s_plot.p8v_shift4c,y
-                    tay
-                    pla
-                    cpy  #0
-                    beq  +
--                   lsr  a
-                    dey
-                    bne  -
-+                   and #3
-                }}
-            }
-            else -> return 0
-        }
+        void addr_mul_24_for_highres_4c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
+        %asm {{
+            stz  cx16.VERA_CTRL
+            lda  cx16.r1L
+            sta  cx16.VERA_ADDR_H
+            lda  cx16.r0H
+            sta  cx16.VERA_ADDR_M
+            lda  cx16.r0L
+            sta  cx16.VERA_ADDR_L
+            lda  cx16.VERA_DATA0
+            pha
+            lda  p8v_xx
+            and  #3
+            tay
+            lda  p8b_gfx_hires.p8s_plot.p8v_shift4c,y
+            tay
+            pla
+            cpy  #0
+            beq  +
+-           lsr  a
+            dey
+            bne  -
++           and #3
+            rts
+        }}
     }
 
     sub fill(uword x, uword y, ubyte new_color) {
@@ -696,7 +603,7 @@ gfx2 {
             if cx16.r12L==MAXDEPTH
                 return
             cx16.r0s = sy+sdy
-            if cx16.r0s>=0 and cx16.r0s<=height-1 {
+            if cx16.r0s>=0 and cx16.r0s<=HEIGHT-1 {
 ;;                stack_xl[cx16.r12L] = sxl
 ;;                stack_xr[cx16.r12L] = sxr
 ;;                stack_y[cx16.r12L] = sy
@@ -753,29 +660,23 @@ gfx2 {
         cx16.r11L = pget(xx as uword, yy as uword)        ; old_color
         if cx16.r11L == cx16.r10L
             return
-        if xx<0 or xx>width-1 or yy<0 or yy>height-1
+        if xx<0 or xx>WIDTH-1 or yy<0 or yy>HEIGHT-1
             return
-        if gfx2.active_mode == 2 set_color_masks()
+        set_color_masks()
         push_stack(xx, xx, yy, 1)
         push_stack(xx, xx, yy + 1, -1)
         word left = 0
         while cx16.r12L!=0 {
             pop_stack()
             xx = x1
-            when active_mode {
-                1 -> if fill_scanline_left_8bpp() goto skip
-                2 -> if fill_scanline_left_2bpp() goto skip
-            }
+            if fill_scanline_left_2bpp() goto skip
             left = xx + 1
             if left < x1
                 push_stack(left, x1 - 1, yy, -dy)
             xx = x1 + 1
 
             do {
-                when active_mode {
-                    1 -> fill_scanline_right_8bpp()
-                    2 -> fill_scanline_right_2bpp()
-                }
+                fill_scanline_right_2bpp()
                 push_stack(left, xx - 1, yy, dy)
                 if xx > x2 + 1
                     push_stack(x2 + 1, xx - 1, yy, -dy)
@@ -801,35 +702,9 @@ skip:
             cx16.VERA_CTRL = 0
         }
 
-        sub fill_scanline_left_8bpp() -> bool {
-            void addr_mul_24_for_lores_256c(yy as uword, xx as uword)      ; 24 bits result is in r0 and r1L (highest byte)
-            cx16.r1L |= %00011000     ; auto decrement enabled
-            set_vera_address()
-            cx16.r9s = xx
-            while xx >= 0 {
-                if cx16.VERA_DATA0 != cx16.r11L
-                    break
-                cx16.VERA_DATA1 = cx16.r10L
-                xx--
-            }
-            return xx==cx16.r9s
-        }
-
-        sub fill_scanline_right_8bpp() {
-            void addr_mul_24_for_lores_256c(yy as uword, xx as uword)      ; 24 bits result is in r0 and r1L (highest byte)
-            cx16.r1L |= %00010000    ; auto increment enabled
-            set_vera_address()
-            while xx <= width-1 {
-                if cx16.VERA_DATA0 != cx16.r11L
-                    break
-                cx16.VERA_DATA1 = cx16.r10L
-                xx++
-            }
-        }
-
         sub fill_scanline_left_2bpp() -> bool {
             uword vx = xx as uword
-            void gfx2.addr_mul_24_for_highres_4c(yy as uword,vx)
+            void gfx_hires.addr_mul_24_for_highres_4c(yy as uword,vx)
             cx16.r1L |= %0001_1000  ; auto decrement
             set_vera_address()
             cmask = amask[lsb(vx) & 3]  ; set the color mask for the first color pel
@@ -873,7 +748,7 @@ set_byte:
         }
 
         sub fill_scanline_right_2bpp() {
-            void gfx2.addr_mul_24_for_highres_4c(yy as uword,xx as uword)
+            void gfx_hires.addr_mul_24_for_highres_4c(yy as uword,xx as uword)
             cx16.r1L |= %00010000    ; auto increment
             set_vera_address()
             cmask = amask[lsb(xx) & 3]  ; set the color mask for the first color pel
@@ -907,7 +782,7 @@ set_byte:
                 }
 set_byte:
                 cx16.VERA_DATA1 = nvub
-                if xx >= gfx2.width-1 break
+                if xx >= WIDTH-1 break
                 cmask = $C0
             }
         }
@@ -919,16 +794,7 @@ set_byte:
     }
 
     sub position(uword @zp xx, uword yy) {
-        when active_mode {
-            1 -> {
-                ; lores 256c
-                void addr_mul_24_for_lores_256c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
-            }
-            2 -> {
-                ; highres 4c
-                void addr_mul_24_for_highres_4c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
-            }
-        }
+        void addr_mul_24_for_highres_4c(yy, xx)      ; 24 bits result is in r0 and r1L (highest byte)
         cx16.r2L = cx16.r1L
         cx16.vaddr(cx16.r2L, cx16.r0, 0, 1)
     }
@@ -1010,83 +876,36 @@ set_byte:
         ubyte[8] @shared char_bitmap_bytes_left
         ubyte[8] @shared char_bitmap_bytes_right
 
-        when active_mode {
-            1 -> {
-                ; lores 256c
-                while @(sctextptr)!=0 {
-                    chardataptr = charset_addr + (@(sctextptr) as uword)*8
-                    cx16.vaddr(charset_bank, chardataptr, 1, 1)
-                    repeat 8 {
-                        position(xx,yy)
-                        yy++
-                        %asm {{
-                            ldx  p8v_color
-                            lda  cx16.VERA_DATA1
-                            sta  P8ZP_SCRATCH_B1
-                            ldy  #8
--                           asl  P8ZP_SCRATCH_B1
-                            bcc  +
-                            stx  cx16.VERA_DATA0    ; write a pixel
-                            bra  ++
-+                           lda  cx16.VERA_DATA0    ; don't write a pixel, but do advance to the next address
-+                           dey
-                            bne  -
-                        }}
-                    }
-                    xx+=8
-                    yy-=8
-                    sctextptr++
-                }
-            }
-            2 -> {
-                ; hires 4c
-                ; we're going to use a few cx16 registers to make sure every variable is in zeropage in the inner loop.
-                cx16.r11L = color
-                while @(sctextptr)!=0 {
-                    chardataptr = charset_addr + (@(sctextptr) as uword)*8
-                    cx16.vaddr(charset_bank, chardataptr, 1, 1)  ; for reading the chardata from Vera data channel 1
-                    position(xx, yy)              ; only calculated once, we update vera address in the loop instead
-                    cx16.VERA_ADDR_H &= $0f     ; no auto increment
-                    repeat 8 {
-                        cx16.r10L = cx16.VERA_DATA1  ; get the next 8 horizontal character bits
-                        cx16.r7 = xx
-                        repeat 8 {
-                            cx16.r10L <<= 1
-                            if_cs {
-                                cx16.r2L = cx16.r7L & 3       ; xbits
-                                when cx16.r11L & 3 {
-                                    1 -> cx16.r12L = gfx2.plot.shiftedleft_4c_1[cx16.r2L]
-                                    2 -> cx16.r12L = gfx2.plot.shiftedleft_4c_2[cx16.r2L]
-                                    3 -> cx16.r12L = gfx2.plot.shiftedleft_4c_3[cx16.r2L]
-                                    else -> cx16.r12L = 0
-                                }
-                                cx16.VERA_DATA0 = cx16.VERA_DATA0 & gfx2.plot.mask4c[cx16.r2L] | cx16.r12L
-                            }
-                            cx16.r7++
-                            if (cx16.r7 & 3) == 0 {
-                                ; increment the pixel address by one
-                                %asm {{
-                                    stz  cx16.VERA_CTRL
-                                    clc
-                                    lda  cx16.VERA_ADDR_L
-                                    adc  #1
-                                    sta  cx16.VERA_ADDR_L
-                                    lda  cx16.VERA_ADDR_M
-                                    adc  #0
-                                    sta  cx16.VERA_ADDR_M
-                                    lda  cx16.VERA_ADDR_H
-                                    adc  #0
-                                    sta  cx16.VERA_ADDR_H
-                                }}
-                            }
+        ; we're going to use a few cx16 registers to make sure every variable is in zeropage in the inner loop.
+        cx16.r11L = color
+        while @(sctextptr)!=0 {
+            chardataptr = charset_addr + (@(sctextptr) as uword)*8
+            cx16.vaddr(charset_bank, chardataptr, 1, 1)  ; for reading the chardata from Vera data channel 1
+            position(xx, yy)              ; only calculated once, we update vera address in the loop instead
+            cx16.VERA_ADDR_H &= $0f     ; no auto increment
+            repeat 8 {
+                cx16.r10L = cx16.VERA_DATA1  ; get the next 8 horizontal character bits
+                cx16.r7 = xx
+                repeat 8 {
+                    cx16.r10L <<= 1
+                    if_cs {
+                        cx16.r2L = cx16.r7L & 3       ; xbits
+                        when cx16.r11L & 3 {
+                            1 -> cx16.r12L = gfx_hires.plot.shiftedleft_4c_1[cx16.r2L]
+                            2 -> cx16.r12L = gfx_hires.plot.shiftedleft_4c_2[cx16.r2L]
+                            3 -> cx16.r12L = gfx_hires.plot.shiftedleft_4c_3[cx16.r2L]
+                            else -> cx16.r12L = 0
                         }
-
-                        ; increment pixel address to the next line
+                        cx16.VERA_DATA0 = cx16.VERA_DATA0 & gfx_hires.plot.mask4c[cx16.r2L] | cx16.r12L
+                    }
+                    cx16.r7++
+                    if (cx16.r7 & 3) == 0 {
+                        ; increment the pixel address by one
                         %asm {{
                             stz  cx16.VERA_CTRL
                             clc
                             lda  cx16.VERA_ADDR_L
-                            adc  #(640-8)/4
+                            adc  #1
                             sta  cx16.VERA_ADDR_L
                             lda  cx16.VERA_ADDR_M
                             adc  #0
@@ -1096,10 +915,25 @@ set_byte:
                             sta  cx16.VERA_ADDR_H
                         }}
                     }
-                    xx+=8
-                    sctextptr++
                 }
+
+                ; increment pixel address to the next line
+                %asm {{
+                    stz  cx16.VERA_CTRL
+                    clc
+                    lda  cx16.VERA_ADDR_L
+                    adc  #(640-8)/4
+                    sta  cx16.VERA_ADDR_L
+                    lda  cx16.VERA_ADDR_M
+                    adc  #0
+                    sta  cx16.VERA_ADDR_M
+                    lda  cx16.VERA_ADDR_H
+                    adc  #0
+                    sta  cx16.VERA_ADDR_H
+                }}
             }
+            xx+=8
+            sctextptr++
         }
     }
 
@@ -1170,50 +1004,4 @@ set_byte:
             rts
         }}
     }
-
-    asmsub addr_mul_24_for_lores_256c(uword yy @R0, uword xx @AY) clobbers(A) -> uword @R0, ubyte @R1  {
-        ; yy * 320 + xx (24 bits calculation)
-        %asm {{
-            sta  P8ZP_SCRATCH_W1
-            sty  P8ZP_SCRATCH_W1+1
-            lda  cx16.r0
-            sta  P8ZP_SCRATCH_B1
-            lda  cx16.r0+1
-            sta  cx16.r1
-            sta  P8ZP_SCRATCH_REG
-            lda  cx16.r0
-            asl  a
-            rol  P8ZP_SCRATCH_REG
-            asl  a
-            rol  P8ZP_SCRATCH_REG
-            asl  a
-            rol  P8ZP_SCRATCH_REG
-            asl  a
-            rol  P8ZP_SCRATCH_REG
-            asl  a
-            rol  P8ZP_SCRATCH_REG
-            asl  a
-            rol  P8ZP_SCRATCH_REG
-            sta  cx16.r0
-            lda  P8ZP_SCRATCH_B1
-            clc
-            adc  P8ZP_SCRATCH_REG
-            sta  cx16.r0+1
-            bcc  +
-            inc  cx16.r1
-+           ; now add the value to this 24-bits number
-            lda  cx16.r0
-            clc
-            adc  P8ZP_SCRATCH_W1
-            sta  cx16.r0
-            lda  cx16.r0+1
-            adc  P8ZP_SCRATCH_W1+1
-            sta  cx16.r0+1
-            bcc  +
-            inc  cx16.r1
-+           lda  cx16.r1
-            rts
-        }}
-    }
-
 }

compiler/res/prog8lib/cx16/gfx_lores.p8

@@ -2,14 +2,18 @@
 ; bitmap image needs to start at VRAM addres $00000.
 ; This is compatible with the CX16's screen mode 128.  (void cx16.set_screen_mode(128))
 
+
+%import syslib
 %import verafx
 
-
 gfx_lores {
-
     %option ignore_unused
 
-    sub set_screen_mode() {
+    const uword WIDTH = 320
+    const ubyte HEIGHT = 240
+
+    sub graphics_mode() {
+        ; enable 320x240 256c bitmap graphics mode
         cx16.VERA_CTRL=0
         cx16.VERA_DC_VIDEO = (cx16.VERA_DC_VIDEO & %11001111) | %00100000      ; enable only layer 1
         cx16.VERA_DC_HSCALE = 64
@@ -18,8 +22,23 @@ gfx_lores {
         cx16.VERA_L1_MAPBASE = 0
         cx16.VERA_L1_TILEBASE = 0
         clear_screen(0)
+        drawmode_eor(false)
     }
 
+    sub text_mode() {
+        ; back to normal text mode
+        cx16.r15L = cx16.VERA_DC_VIDEO & %00000111 ; retain chroma + output mode
+        cbm.CINT()
+        cx16.VERA_DC_VIDEO = (cx16.VERA_DC_VIDEO & %11111000) | cx16.r15L
+    }
+
+    sub drawmode_eor(bool enabled) {
+        ; with EOR drawing mode you can have non destructive drawing (2*EOR=restore original)
+        eor_mode = enabled
+    }
+
+    bool eor_mode
+
     sub clear_screen(ubyte color) {
         if verafx.available() {
             ; use verafx cache writes to quicly clear the screen
@@ -56,10 +75,10 @@ gfx_lores {
         cx16.VERA_CTRL=0
         cx16.VERA_ADDR=0
         cx16.VERA_ADDR_H = 1<<4    ; 1 pixel auto increment
-        repeat 240 {
+        repeat HEIGHT {
             %asm {{
                 lda  p8v_color
-                ldy  #320/8
+                ldy  #p8c_WIDTH/8
 -               .rept 8
                 sta  cx16.VERA_DATA0
                 .endrept
@@ -71,28 +90,186 @@ gfx_lores {
         cx16.VERA_ADDR_H = 0
     }
 
-    asmsub plot(uword x @AX, ubyte y @Y, ubyte color @R0) {
-        ; x in r0,  y in r1,   color.
-        %asm {{
-            clc
-            adc  times320_lo,y
-            sta  cx16.VERA_ADDR_L
-            txa
-            adc  times320_mid,y
-            sta  cx16.VERA_ADDR_M
-            lda  #0
-            adc  times320_hi,y
-            sta  cx16.VERA_ADDR_H
-            lda  cx16.r0L
-            sta  cx16.VERA_DATA0
-            rts
-        }}
+    sub rect(uword xx, ubyte yy, uword rwidth, ubyte rheight, ubyte color) {
+        if rwidth==0 or rheight==0
+            return
+        horizontal_line(xx, yy, rwidth, color)
+        if rheight==1
+            return
+        horizontal_line(xx, yy+rheight-1, rwidth, color)
+        vertical_line(xx, yy+1, rheight-2, color)
+        if rwidth==1
+            return
+        vertical_line(xx+rwidth-1, yy+1, rheight-2, color)
+    }
+
+    sub safe_rect(uword xx, ubyte yy, uword rwidth, ubyte rheight, ubyte color) {
+        ; does bounds checking and clipping
+        safe_horizontal_line(xx, yy, rwidth, color)
+        if rheight==1
+            return
+        uword bottomyy = yy as uword + rheight -1
+        if bottomyy<HEIGHT
+            safe_horizontal_line(xx, lsb(bottomyy), rwidth, color)
+        safe_vertical_line(xx, yy+1, rheight-2, color)
+        if rwidth==1
+            return
+        safe_vertical_line(xx+rwidth-1, yy+1, rheight-2, color)
+    }
+
+    sub fillrect(uword xx, ubyte yy, uword rwidth, ubyte rheight, ubyte color) {
+        ; Draw a filled rectangle of the given size and color.
+        ; To fill the whole screen, use clear_screen(color) instead - it is much faster.
+        if rwidth==0
+            return
+        repeat rheight {
+            horizontal_line(xx, yy, rwidth, color)
+            yy++
+        }
+    }
+
+    sub safe_fillrect(uword xx, ubyte yy, uword rwidth, ubyte rheight, ubyte color) {
+        ; Draw a filled rectangle of the given size and color.
+        ; To fill the whole screen, use clear_screen(color) instead - it is much faster.
+        ; This safe version does bounds checking and clipping.
+        if xx>=WIDTH or yy>=HEIGHT
+            return
+        if msb(xx)&$80!=0 {
+            rwidth += xx
+            xx = 0
+        }
+        if xx>=WIDTH
+            return
+        if xx+rwidth>WIDTH
+            rwidth = WIDTH-xx
+        if rwidth>WIDTH
+            return
+
+        if yy as uword + rheight > HEIGHT
+            rheight = HEIGHT-yy
+        if rheight>HEIGHT
+            return
+
+        repeat rheight {
+            horizontal_line(xx, yy, rwidth, color)
+            yy++
+        }
+    }
+
+    sub horizontal_line(uword xx, ubyte yy, uword length, ubyte color) {
+        if length==0
+            return
+        position(xx, yy)
+        ; set vera auto-increment to 1 pixel
+        cx16.VERA_ADDR_H = cx16.VERA_ADDR_H & %00000111 | (1<<4)
+        if eor_mode {
+            cx16.vaddr_clone(0)      ; also setup port 1, for reading
+            %asm {{
+                ldx  p8v_length+1
+                beq  +
+                ldy  #0
+-               lda  p8v_color
+                eor  cx16.VERA_DATA1
+                sta  cx16.VERA_DATA0
+                iny
+                bne  -
+                dex
+                bne  -
++               ldy  p8v_length     ; remaining
+                beq  +
+-               lda  p8v_color
+                eor  cx16.VERA_DATA1
+                sta  cx16.VERA_DATA0
+                dey
+                bne  -
++
+            }}
+        } else {
+            %asm {{
+                lda  p8v_color
+                ldx  p8v_length+1
+                beq  +
+                ldy  #0
+-               sta  cx16.VERA_DATA0
+                iny
+                bne  -
+                dex
+                bne  -
++               ldy  p8v_length     ; remaining
+                beq  +
+-               sta  cx16.VERA_DATA0
+                dey
+                bne  -
++
+            }}
+        }
+    }
+
+    sub safe_horizontal_line(uword xx, ubyte yy, uword length, ubyte color) {
+        ; does bounds checking and clipping
+        if yy>=HEIGHT
+            return
+        if msb(xx)&$80!=0 {
+            length += xx
+            xx = 0
+        }
+        if xx>=WIDTH
+            return
+        if xx+length>WIDTH
+            length = WIDTH-xx
+        if length>WIDTH
+            return
+
+        horizontal_line(xx, yy, length, color)
+    }
+
+    sub vertical_line(uword xx, ubyte yy, ubyte lheight, ubyte color) {
+        if lheight==0
+            return
+        position(xx, yy)
+        ; set vera auto-increment to 320 pixel increment (=next line)
+        cx16.VERA_ADDR_H = cx16.VERA_ADDR_H & %00000111 | (14<<4)
+        if eor_mode {
+            cx16.vaddr_clone(0)      ; also setup port 1, for reading
+            %asm {{
+                ldy  p8v_lheight
+                beq  +
+-               lda  p8v_color
+                eor  cx16.VERA_DATA1
+                sta  cx16.VERA_DATA0
+                dey
+                bne  -
++
+            }}
+        } else {
+            %asm {{
+                ldy  p8v_lheight
+                lda  p8v_color
+-               sta  cx16.VERA_DATA0
+                dey
+                bne  -
+            }}
+        }
+    }
+
+    sub safe_vertical_line(uword xx, ubyte yy, ubyte lheight, ubyte color) {
+        ; does bounds checking and clipping
+        if yy>=HEIGHT
+            return
+        if msb(xx)&$80!=0 or xx>=WIDTH
+            return
+        if yy as uword + lheight > HEIGHT
+            lheight = HEIGHT-yy
+        if lheight>HEIGHT
+            return
+
+        vertical_line(xx, yy, lheight, color)
     }
 
     sub line(uword x1, ubyte y1, uword x2, ubyte y2, ubyte color) {
         ; Bresenham algorithm.
         ; This code special-cases various quadrant loops to allow simple ++ and -- operations.
-        ; NOTE:  this is about twice as fast as the kernal routine GRAPH_draw_line, and ~3-4 times as fast as gfx2.line()
+        ; NOTE:  this is about twice as fast as the kernal routine GRAPH_draw_line
         ;        it trades memory for speed (uses inline plot routine and multiplication lookup tables)
         ;
         ; NOTE:  is currently still a regular 6502 routine, could likely be made much faster with the VeraFX line helper.
@@ -207,55 +384,512 @@ gfx_lores {
                 lda  #0
                 adc  times320_hi,y
                 sta  cx16.VERA_ADDR_H
+
+                lda  p8v_eor_mode
+                bne  +
                 lda  p8v_color
                 sta  cx16.VERA_DATA0
                 rts
++               lda  p8v_color
+                eor  cx16.VERA_DATA0
+                sta  cx16.VERA_DATA0
+                rts
             }}
         }
     }
 
-    sub horizontal_line(uword xx, ubyte yy, uword length, ubyte color) {
-        if length==0
+    sub circle(uword @zp xcenter, ubyte @zp ycenter, ubyte radius, ubyte color) {
+        ; Warning: NO BOUNDS CHECKS. Make sure circle fits in the screen.
+        ; Midpoint algorithm.
+        if radius==0
             return
-        plot(xx, yy, color)  ; set starting position by reusing plot routine
-        ; set vera auto-increment to 1 pixel
-        cx16.VERA_ADDR_H = cx16.VERA_ADDR_H & %00000111 | (1<<4)
 
+        ubyte @zp xx = radius
+        ubyte @zp yy = 0
+        word @zp decisionOver2 = (1 as word)-xx
+        ; R14 = internal plot X
+        ; R15 = internal plot Y
+
+        while xx>=yy {
+            cx16.r14 = xcenter + xx
+            cx16.r15 = ycenter + yy
+            plotq()
+            cx16.r14 = xcenter - xx
+            plotq()
+            cx16.r14 = xcenter + xx
+            cx16.r15 = ycenter - yy
+            plotq()
+            cx16.r14 = xcenter - xx
+            plotq()
+            cx16.r14 = xcenter + yy
+            cx16.r15 = ycenter + xx
+            plotq()
+            cx16.r14 = xcenter - yy
+            plotq()
+            cx16.r14 = xcenter + yy
+            cx16.r15 = ycenter - xx
+            plotq()
+            cx16.r14 = xcenter - yy
+            plotq()
+
+            yy++
+            if decisionOver2>=0 {
+                xx--
+                decisionOver2 -= xx*$0002
+            }
+            decisionOver2 += yy*$0002
+            decisionOver2++
+        }
+
+        sub plotq() {
+            ; cx16.r14 = x, cx16.r15 = y, color=color.
+            plot(cx16.r14, cx16.r15L, color)
+        }
+    }
+
+    sub safe_circle(uword @zp xcenter, uword @zp ycenter, ubyte radius, ubyte color) {
+        ; This version does bounds checks and clipping, but is a lot slower.
+        ; Midpoint algorithm.
+        if radius==0
+            return
+
+        ubyte @zp xx = radius
+        ubyte @zp yy = 0
+        word @zp decisionOver2 = (1 as word)-xx
+        ; R14 = internal plot X
+        ; R15 = internal plot Y
+
+        while xx>=yy {
+            cx16.r14 = xcenter + xx
+            cx16.r15 = ycenter + yy
+            plotq()
+            cx16.r14 = xcenter - xx
+            plotq()
+            cx16.r14 = xcenter + xx
+            cx16.r15 = ycenter - yy
+            plotq()
+            cx16.r14 = xcenter - xx
+            plotq()
+            cx16.r14 = xcenter + yy
+            cx16.r15 = ycenter + xx
+            plotq()
+            cx16.r14 = xcenter - yy
+            plotq()
+            cx16.r14 = xcenter + yy
+            cx16.r15 = ycenter - xx
+            plotq()
+            cx16.r14 = xcenter - yy
+            plotq()
+
+            yy++
+            if decisionOver2>=0 {
+                xx--
+                decisionOver2 -= xx*$0002
+            }
+            decisionOver2 += yy*$0002
+            decisionOver2++
+        }
+
+        sub plotq() {
+            ; cx16.r14 = x, cx16.r15 = y, color=color.
+            if cx16.r15 < HEIGHT
+                safe_plot(cx16.r14, cx16.r15L, color)
+        }
+    }
+
+    sub disc(uword @zp xcenter, ubyte @zp ycenter, ubyte @zp radius, ubyte color) {
+        ; Warning: NO BOUNDS CHECKS. Make sure circle fits in the screen.
+        ; Midpoint algorithm, filled
+        if radius==0
+            return
+        ubyte @zp yy = 0
+        word @zp decisionOver2 = (1 as word)-radius
+        ubyte last_y3 = ycenter+radius
+        ubyte last_y4 = ycenter-radius
+        ubyte new_y3, new_y4
+
+        while radius>=yy {
+            horizontal_line(xcenter-radius, ycenter+yy, radius*$0002+1, color)
+            horizontal_line(xcenter-radius, ycenter-yy, radius*$0002+1, color)
+
+            new_y3 = ycenter+radius
+            if new_y3 != last_y3 {
+                horizontal_line(xcenter-yy, last_y3, yy*$0002+1, color)
+                last_y3 = new_y3
+            }
+            new_y4 = ycenter-radius
+            if new_y4 != last_y4 {
+                horizontal_line(xcenter-yy, last_y4, yy*$0002+1, color)
+                last_y4 = new_y4
+            }
+
+            yy++
+            if decisionOver2>=0 {
+                radius--
+                decisionOver2 -= radius*$0002
+            }
+            decisionOver2 += yy*$0002
+            decisionOver2++
+        }
+
+        ; draw the final two spans
+        yy--
+        horizontal_line(xcenter-yy, last_y3, yy*$0002+1, color)
+        horizontal_line(xcenter-yy, last_y4, yy*$0002+1, color)
+    }
+
+    sub safe_disc(uword @zp xcenter, uword @zp ycenter, ubyte @zp radius, ubyte color) {
+        ; This version does bounds checks and clipping, but is a lot slower.
+        ; Midpoint algorithm, filled
+        if radius==0
+            return
+        ubyte @zp yy = 0
+        word @zp decisionOver2 = (1 as word)-radius
+        uword last_y3 = ycenter+radius
+        uword last_y4 = ycenter-radius
+        uword new_y3, new_y4
+
+        while radius>=yy {
+            uword liney = ycenter+yy
+            if msb(liney)==0
+                safe_horizontal_line(xcenter-radius, lsb(ycenter+yy), radius*$0002+1, color)
+            liney = ycenter-yy
+            if msb(liney)==0
+                safe_horizontal_line(xcenter-radius, lsb(ycenter-yy), radius*$0002+1, color)
+            new_y3 = ycenter+radius
+            if new_y3 != last_y3 {
+                if msb(last_y3)==0
+                    safe_horizontal_line(xcenter-yy, lsb(last_y3), yy*$0002+1, color)
+                last_y3 = new_y3
+            }
+            new_y4 = ycenter-radius
+            if new_y4 != last_y4 {
+                if msb(last_y4)==0
+                    safe_horizontal_line(xcenter-yy, lsb(last_y4), yy*$0002+1, color)
+                last_y4 = new_y4
+            }
+            yy++
+            if decisionOver2>=0 {
+                radius--
+                decisionOver2 -= radius*$0002
+            }
+            decisionOver2 += yy*$0002
+            decisionOver2++
+        }
+        ; draw the final two spans
+        yy--
+        if msb(last_y3)==0
+            safe_horizontal_line(xcenter-yy, lsb(last_y3), yy*$0002+1, color)
+        if msb(last_y4)==0
+            safe_horizontal_line(xcenter-yy, lsb(last_y4), yy*$0002+1, color)
+    }
+
+    asmsub plot(uword x @AX, ubyte y @Y, ubyte color @R0) {
+        ; x in r0,  y in r1,   color.
         %asm {{
-            lda  p8v_color
-            ldx  p8v_length+1
+            clc
+            adc  times320_lo,y
+            sta  cx16.VERA_ADDR_L
+            txa
+            adc  times320_mid,y
+            sta  cx16.VERA_ADDR_M
+            lda  #0
+            adc  times320_hi,y
+            sta  cx16.VERA_ADDR_H
+
+            lda  p8v_eor_mode
+            bne  +
+            lda  cx16.r0L
+            sta  cx16.VERA_DATA0
+            rts
++           lda  cx16.r0L
+            eor  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            rts
+        }}
+    }
+
+    sub safe_plot(uword xx, ubyte yy, ubyte color) {
+        ; A plot that does bounds checks to see if the pixel is inside the screen.
+        if msb(xx)&$80!=0
+            return
+        if xx >= WIDTH or yy >= HEIGHT
+            return
+        plot(xx, yy, color)
+    }
+
+    asmsub pget(uword x @AX, ubyte y @Y) -> ubyte @A {
+        ; returns the color of the pixel
+        %asm {{
+            jsr  p8s_position
+            lda  cx16.VERA_DATA0
+            rts
+        }}
+    }
+
+    sub fill(uword x, ubyte y, ubyte new_color) {
+        ; reuse a few virtual registers in ZP for variables
+        &ubyte fillm = &cx16.r7L
+        &ubyte seedm = &cx16.r8L
+        &ubyte cmask = &cx16.r8H
+        &ubyte vub   = &cx16.r13L
+        &ubyte nvub  = &cx16.r13H
+        ubyte[4] amask = [$c0,$30,$0c,$03] ; array of cmask bytes
+
+        ; Non-recursive scanline flood fill.
+        ; based loosely on code found here https://www.codeproject.com/Articles/6017/QuickFill-An-efficient-flood-fill-algorithm
+        ; with the fixes applied to the seedfill_4 routine as mentioned in the comments.
+        const ubyte MAXDEPTH = 100
+        word @zp xx = x as word
+        word @zp yy = y as word
+        word[MAXDEPTH] @split @shared stack_xl
+        word[MAXDEPTH] @split @shared stack_xr
+        word[MAXDEPTH] @split @shared stack_y
+        byte[MAXDEPTH] @shared stack_dy
+        cx16.r12L = 0       ; stack pointer
+        word x1
+        word x2
+        byte dy
+        cx16.r10L = new_color
+
+        sub push_stack(word sxl, word sxr, word sy, byte sdy) {
+            if cx16.r12L==MAXDEPTH
+                return
+            cx16.r0s = sy+sdy
+            if cx16.r0s>=0 and cx16.r0s<=HEIGHT-1 {
+;;                stack_xl[cx16.r12L] = sxl
+;;                stack_xr[cx16.r12L] = sxr
+;;                stack_y[cx16.r12L] = sy
+;;                stack_dy[cx16.r12L] = sdy
+;;                cx16.r12L++
+                %asm {{
+                    ldy  cx16.r12L
+                    lda  p8v_sxl
+                    sta  p8v_stack_xl_lsb,y
+                    lda  p8v_sxl+1
+                    sta  p8v_stack_xl_msb,y
+                    lda  p8v_sxr
+                    sta  p8v_stack_xr_lsb,y
+                    lda  p8v_sxr+1
+                    sta  p8v_stack_xr_msb,y
+                    lda  p8v_sy
+                    sta  p8v_stack_y_lsb,y
+                    lda  p8v_sy+1
+                    sta  p8v_stack_y_msb,y
+                    ldy  cx16.r12L
+                    lda  p8v_sdy
+                    sta  p8v_stack_dy,y
+                    inc  cx16.r12L
+                }}
+            }
+        }
+        sub pop_stack() {
+;;            cx16.r12L--
+;;            x1 = stack_xl[cx16.r12L]
+;;            x2 = stack_xr[cx16.r12L]
+;;            y = stack_y[cx16.r12L]
+;;            dy = stack_dy[cx16.r12L]
+            %asm {{
+                dec  cx16.r12L
+                ldy  cx16.r12L
+                lda  p8v_stack_xl_lsb,y
+                sta  p8v_x1
+                lda  p8v_stack_xl_msb,y
+                sta  p8v_x1+1
+                lda  p8v_stack_xr_lsb,y
+                sta  p8v_x2
+                lda  p8v_stack_xr_msb,y
+                sta  p8v_x2+1
+                lda  p8v_stack_y_lsb,y
+                sta  p8v_yy
+                lda  p8v_stack_y_msb,y
+                sta  p8v_yy+1
+                ldy  cx16.r12L
+                lda  p8v_stack_dy,y
+                sta  p8v_dy
+            }}
+            yy+=dy
+        }
+        cx16.r11L = pget(xx as uword, lsb(yy))        ; old_color
+        if cx16.r11L == cx16.r10L
+            return
+        if xx<0 or xx>WIDTH-1 or yy<0 or yy>HEIGHT-1
+            return
+        push_stack(xx, xx, yy, 1)
+        push_stack(xx, xx, yy + 1, -1)
+        word left = 0
+        while cx16.r12L!=0 {
+            pop_stack()
+            xx = x1
+            if fill_scanline_left_8bpp() goto skip
+            left = xx + 1
+            if left < x1
+                push_stack(left, x1 - 1, yy, -dy)
+            xx = x1 + 1
+
+            do {
+                fill_scanline_right_8bpp()
+                push_stack(left, xx - 1, yy, dy)
+                if xx > x2 + 1
+                    push_stack(x2 + 1, xx - 1, yy, -dy)
+skip:
+                xx++
+                while xx <= x2 {
+                    if pget(xx as uword, lsb(yy)) == cx16.r11L
+                        break
+                    xx++
+                }
+                left = xx
+            } until xx>x2
+        }
+
+        sub set_vera_address(bool decr) {
+            ; set both data0 and data1 addresses
+            position(xx as uword, lsb(yy))
+            cx16.r0 = cx16.VERA_ADDR
+            cx16.r1L = cx16.VERA_ADDR_H & 1 | if decr %00011000 else %00010000
+            cx16.VERA_ADDR_H = cx16.r1L
+            cx16.VERA_CTRL = 1
+            cx16.VERA_ADDR = cx16.r0
+            cx16.VERA_ADDR_H = cx16.r1L
+            cx16.VERA_CTRL = 0
+        }
+
+        sub fill_scanline_left_8bpp() -> bool {
+            set_vera_address(true)
+            cx16.r9s = xx
+            while xx >= 0 {
+                if cx16.VERA_DATA0 != cx16.r11L
+                    break
+                cx16.VERA_DATA1 = cx16.r10L
+                xx--
+            }
+            return xx==cx16.r9s
+        }
+
+        sub fill_scanline_right_8bpp() {
+            set_vera_address(false)
+            while xx <= WIDTH-1 {
+                if cx16.VERA_DATA0 != cx16.r11L
+                    break
+                cx16.VERA_DATA1 = cx16.r10L
+                xx++
+            }
+        }
+    }
+
+    sub text_charset(ubyte charset) {
+        ; -- select the text charset to use with the text() routine
+        ;    the charset number is the same as for the cx16.screen_set_charset() ROM function.
+        ;    1 = ISO charset, 2 = PETSCII uppercase+graphs, 3= PETSCII uppercase+lowercase etc. etc.
+        cx16.screen_set_charset(charset, 0)
+    }
+
+    const ubyte charset_bank = $1
+    const uword charset_addr = $f000       ; in bank 1, so $1f000
+
+    sub text(uword @zp xx, uword yy, ubyte color, uword textptr) {
+        ; -- Write some text at the given pixel position. The text string must be in an encoding approprite for the charset.
+        ;    You must also have called text_charset() first to select and prepare the character set to use.
+        uword chardataptr
+        ubyte[8] @shared char_bitmap_bytes_left
+        ubyte[8] @shared char_bitmap_bytes_right
+
+        while @(textptr)!=0 {
+            chardataptr = charset_addr + (@(textptr) as uword)*8
+            cx16.vaddr(charset_bank, chardataptr, 1, 1)
+            repeat 8 {
+                position(xx,lsb(yy))
+                yy++
+                %asm {{
+                    ldx  p8v_color
+                    lda  cx16.VERA_DATA1
+                    sta  P8ZP_SCRATCH_B1
+                    ldy  #8
+-                   asl  P8ZP_SCRATCH_B1
+                    bcc  +
+                    stx  cx16.VERA_DATA0    ; write a pixel
+                    bra  ++
++                   lda  cx16.VERA_DATA0    ; don't write a pixel, but do advance to the next address
++                   dey
+                    bne  -
+                }}
+            }
+            xx+=8
+            yy-=8
+            textptr++
+        }
+    }
+
+    asmsub position(uword x @AX, ubyte y @Y) {
+        %asm {{
+            clc
+            adc  times320_lo,y
+            sta  cx16.VERA_ADDR_L
+            txa
+            adc  times320_mid,y
+            sta  cx16.VERA_ADDR_M
+            lda  #%00010000         ; auto increment on
+            adc  times320_hi,y
+            sta  cx16.VERA_ADDR_H
+            rts
+        }}
+    }
+
+    inline asmsub next_pixel(ubyte color @A) {
+        ; -- sets the next pixel byte to the graphics chip.
+        ;    for 8 bpp screens this will plot 1 pixel.
+        ;    for 2 bpp screens it will plot 4 pixels at once (color = bit pattern).
+        %asm {{
+            sta  cx16.VERA_DATA0
+        }}
+    }
+
+    asmsub next_pixels(uword pixels @AY, uword amount @R0) clobbers(A, X, Y)  {
+        ; -- sets the next bunch of pixels from a prepared array of bytes.
+        ;    for 8 bpp screens this will plot 1 pixel per byte.
+        ;    for 2 bpp screens it will plot 4 pixels at once (colors are the bit patterns per byte).
+        %asm {{
+            sta  P8ZP_SCRATCH_W1
+            sty  P8ZP_SCRATCH_W1+1
+            ldx  cx16.r0+1
             beq  +
             ldy  #0
--           sta  cx16.VERA_DATA0
+-           lda  (P8ZP_SCRATCH_W1),y
+            sta  cx16.VERA_DATA0
             iny
             bne  -
+            inc  P8ZP_SCRATCH_W1+1       ; next page of 256 pixels
             dex
             bne  -
-+           ldy  p8v_length     ; remaining
+
++           ldx  cx16.r0           ; remaining pixels
             beq  +
--           sta  cx16.VERA_DATA0
-            dey
+            ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            sta  cx16.VERA_DATA0
+            iny
+            dex
             bne  -
-+
++           rts
         }}
     }
 
-    sub vertical_line(uword xx, ubyte yy, ubyte lheight, ubyte color) {
-        if lheight==0
-            return
-        plot(xx, yy, color)  ; set starting position by reusing plot routine
-        ; set vera auto-increment to 320 pixel increment (=next line)
-        cx16.VERA_ADDR_H = cx16.VERA_ADDR_H & %00000111 | (14<<4)
+    asmsub set_8_pixels_from_bits(ubyte bits @R0, ubyte oncolor @A, ubyte offcolor @Y) clobbers(X) {
+        ; this is only useful in 256 color mode where one pixel equals one byte value.
         %asm {{
-            ldy  p8v_lheight
-            lda  p8v_color
--           sta  cx16.VERA_DATA0
-            dey
+            ldx  #8
+-           asl  cx16.r0
+            bcc  +
+            sta  cx16.VERA_DATA0
+            bra  ++
++           sty  cx16.VERA_DATA0
++           dex
             bne  -
+            rts
         }}
     }
 
-
     %asm {{
 ; multiplication by 320 lookup table
 times320 := 320*range(240)
@@ -265,4 +899,3 @@ times320_mid    .byte >times320
 times320_hi     .byte `times320
     }}
 }
-

compiler/res/prog8lib/cx16/graphics.p8

@@ -4,7 +4,7 @@
 ; Wraps the graphics functions that are in ROM.
 ; Only lo-res 320x240 256 color mode for now.
 ; Unlike graphics module on the C64, you can use colors() to set new drawing colors for every draw operation.
-; For other resolutions or other color modes, use the "gfx2" or "monogfx" module instead. (which is Cx16-specific)
+; For other resolutions or other color modes, use the "gfx_lores", "gfx_hires", or "monogfx" module instead.
 ; Note: there is no color palette manipulation here, you have to do that yourself or use the "palette" module.
 ;
 ; NOTE: For sake of speed, NO BOUNDS CHECKING is performed in most routines!
@@ -15,7 +15,7 @@
 graphics {
     %option ignore_unused
 
-    romsub $feff = FB_cursor_position2()  clobbers(A,X,Y)     ; alias for the normal FB_cursor_position() call but reuses existing r0 and r1
+    extsub $feff = FB_cursor_position2()  clobbers(A,X,Y)     ; alias for the normal FB_cursor_position() call but reuses existing r0 and r1
 
     const uword WIDTH = 320
     const ubyte HEIGHT = 240

compiler/res/prog8lib/cx16/monogfx.p8

@@ -2,9 +2,8 @@
 ; Using the full-screen 640x480 and 320x240 screen modes, in 1 bpp mode (black/white).
 ;
 ; No text layer is currently shown, but text can be drawn as part of the bitmap itself.
-; For color bitmap graphics, see the gfx2 library.
+; For color bitmap graphics, see the gfx_lores or gfx_hires libraries.
 ;
-; NOTE: a lot of the code here is similar or the same to that in gfx2
 ; NOTE: For sake of speed, NO BOUNDS CHECKING is performed in most routines!
 ;       You'll have to make sure yourself that you're not writing outside of bitmap boundaries!
 

compiler/res/prog8lib/cx16/sprites.p8

@@ -34,6 +34,22 @@ sprites {
         cx16.vpoke(1, sprite_reg+7, height_flag<<6 | width_flag<<4 | palette_offset&15) ; 64x64 pixels, palette offset
     }
 
+    sub reset(ubyte spritenum_start, ubyte count)  {
+        ; resets all sprite attributes for the given sprite range
+        ; this removes these sprites from the screen completely
+        ; (without disabling sprites globally so the mouse cursor remains visible)
+        if spritenum_start > 127
+            return
+        if count + spritenum_start > 128
+            return
+        cx16.VERA_CTRL   = $00
+        cx16.VERA_ADDR_H = $11
+        cx16.VERA_ADDR   = VERA_SPRITEREGS + spritenum_start * $0008
+        repeat count {
+            unroll 8 cx16.VERA_DATA0 = $00
+        }
+    }
+
     sub data(ubyte spritenum, ubyte bank, uword addr) {
         addr >>= 5
         addr |= (bank as uword)<<11

compiler/res/prog8lib/cx16/syslib.p8

@@ -44,45 +44,45 @@ cbm {
 ; CLEARSCR -> use txt.clear_screen
 ; HOMECRSR -> use txt.home or txt.plot
 
-romsub $FF81 = CINT() clobbers(A,X,Y)                           ; (alias: SCINIT) initialize screen editor and video chip, including resetting to the default color palette. Note: also sets the video mode back to VGA
-romsub $FF84 = IOINIT() clobbers(A, X)                          ; initialize I/O devices (CIA, IRQ, ...)
-romsub $FF87 = RAMTAS() clobbers(A,X,Y)                         ; initialize RAM, screen
-romsub $FF8A = RESTOR() clobbers(A,X,Y)                         ; restore default I/O vectors
-romsub $FF8D = VECTOR(uword userptr @ XY, bool dir @ Pc) clobbers(A,Y)     ; read/set I/O vector table
-romsub $FF90 = SETMSG(ubyte value @ A)                          ; set Kernal message control flag
-romsub $FF93 = SECOND(ubyte address @ A) clobbers(A)            ; (alias: LSTNSA) send secondary address after LISTEN
-romsub $FF96 = TKSA(ubyte address @ A) clobbers(A)              ; (alias: TALKSA) send secondary address after TALK
-romsub $FF99 = MEMTOP(uword address @ XY, bool dir @ Pc) -> uword @ XY, ubyte @A     ; read/set top of memory  pointer.   NOTE: on the Cx16 also returns the number of RAM memory banks in A!  Also see cx16.numbanks()
-romsub $FF9C = MEMBOT(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set bottom of memory  pointer
-romsub $FF9F = SCNKEY() clobbers(A,X,Y)                         ; scan the keyboard, also called  kbd_scan
-romsub $FFA2 = SETTMO(ubyte timeout @ A)                        ; set time-out flag for IEEE bus
-romsub $FFA5 = ACPTR() -> ubyte @ A                             ; (alias: IECIN) input byte from serial bus
-romsub $FFA8 = CIOUT(ubyte databyte @ A)                        ; (alias: IECOUT) output byte to serial bus
-romsub $FFAB = UNTLK() clobbers(A)                              ; command serial bus device to UNTALK
-romsub $FFAE = UNLSN() clobbers(A)                              ; command serial bus device to UNLISTEN
-romsub $FFB1 = LISTEN(ubyte device @ A) clobbers(A)             ; command serial bus device to LISTEN
-romsub $FFB4 = TALK(ubyte device @ A) clobbers(A)               ; command serial bus device to TALK
-romsub $FFB7 = READST() -> ubyte @ A                            ; read I/O status word  (use CLEARST to reset it to 0)
-romsub $FFBA = SETLFS(ubyte logical @ A, ubyte device @ X, ubyte secondary @ Y)   ; set logical file parameters
-romsub $FFBD = SETNAM(ubyte namelen @ A, str filename @ XY)     ; set filename parameters
-romsub $FFC0 = OPEN() clobbers(X,Y) -> bool @Pc, ubyte @A       ; (via 794 ($31A)) open a logical file
-romsub $FFC3 = CLOSE(ubyte logical @ A) clobbers(A,X,Y)         ; (via 796 ($31C)) close a logical file
-romsub $FFC6 = CHKIN(ubyte logical @ X) clobbers(A,X) -> bool @Pc    ; (via 798 ($31E)) define an input channel
-romsub $FFC9 = CHKOUT(ubyte logical @ X) clobbers(A,X)          ; (via 800 ($320)) define an output channel
-romsub $FFCC = CLRCHN() clobbers(A,X)                           ; (via 802 ($322)) restore default devices
-romsub $FFCF = CHRIN() clobbers(X, Y) -> ubyte @ A   ; (via 804 ($324)) input a character (for keyboard, read a whole line from the screen) A=byte read.
-romsub $FFD2 = CHROUT(ubyte character @ A)                           ; (via 806 ($326)) output a character
-romsub $FFD5 = LOAD(ubyte verify @ A, uword address @ XY) -> bool @Pc, ubyte @ A, uword @ XY     ; (via 816 ($330)) load from device
-romsub $FFD8 = SAVE(ubyte zp_startaddr @ A, uword endaddr @ XY) clobbers (X, Y) -> bool @ Pc, ubyte @ A       ; (via 818 ($332)) save to a device.  See also BSAVE
-romsub $FFDB = SETTIM(ubyte low @ A, ubyte middle @ X, ubyte high @ Y)      ; set the software clock
-romsub $FFDE = RDTIM() -> ubyte @ A, ubyte @ X, ubyte @ Y       ; read the software clock (in little endian order: A=lo,X=mid,Y=high) , however use RDTIM_safe() instead
-romsub $FFE1 = STOP() clobbers(X) -> bool @ Pz, ubyte @ A       ; (via 808 ($328)) check the STOP key (and some others in A)        also see STOP2
-romsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A     ; (via 810 ($32A)) get a character      also see GETIN2
-romsub $FFE7 = CLALL() clobbers(A,X)                            ; (via 812 ($32C)) close all files
-romsub $FFEA = UDTIM() clobbers(A,X)                            ; update the software clock
-romsub $FFED = SCREEN() -> ubyte @ X, ubyte @ Y                 ; get size of text screen into X (columns) and Y (rows)
-romsub $FFF0 = PLOT(ubyte col @ Y, ubyte row @ X, bool dir @ Pc) clobbers(A) -> ubyte @ X, ubyte @ Y       ; read/set position of cursor on screen.  Also see txt.plot
-romsub $FFF3 = IOBASE() -> uword @ XY                           ; read base address of I/O devices
+extsub $FF81 = CINT() clobbers(A,X,Y)                           ; (alias: SCINIT) initialize screen editor and video chip, including resetting to the default color palette. Note: also sets the video mode back to VGA
+extsub $FF84 = IOINIT() clobbers(A, X)                          ; initialize I/O devices (CIA, IRQ, ...)
+extsub $FF87 = RAMTAS() clobbers(A,X,Y)                         ; initialize RAM, screen
+extsub $FF8A = RESTOR() clobbers(A,X,Y)                         ; restore default I/O vectors
+extsub $FF8D = VECTOR(uword userptr @ XY, bool dir @ Pc) clobbers(A,Y)     ; read/set I/O vector table
+extsub $FF90 = SETMSG(ubyte value @ A)                          ; set Kernal message control flag
+extsub $FF93 = SECOND(ubyte address @ A) clobbers(A)            ; (alias: LSTNSA) send secondary address after LISTEN
+extsub $FF96 = TKSA(ubyte address @ A) clobbers(A)              ; (alias: TALKSA) send secondary address after TALK
+extsub $FF99 = MEMTOP(uword address @ XY, bool dir @ Pc) -> uword @ XY, ubyte @A     ; read/set top of memory  pointer.   NOTE: on the Cx16 also returns the number of RAM memory banks in A!  Also see cx16.numbanks()
+extsub $FF9C = MEMBOT(uword address @ XY, bool dir @ Pc) -> uword @ XY     ; read/set bottom of memory  pointer
+extsub $FF9F = SCNKEY() clobbers(A,X,Y)                         ; scan the keyboard, also called  kbd_scan
+extsub $FFA2 = SETTMO(ubyte timeout @ A)                        ; set time-out flag for IEEE bus
+extsub $FFA5 = ACPTR() -> ubyte @ A                             ; (alias: IECIN) input byte from serial bus
+extsub $FFA8 = CIOUT(ubyte databyte @ A)                        ; (alias: IECOUT) output byte to serial bus
+extsub $FFAB = UNTLK() clobbers(A)                              ; command serial bus device to UNTALK
+extsub $FFAE = UNLSN() clobbers(A)                              ; command serial bus device to UNLISTEN
+extsub $FFB1 = LISTEN(ubyte device @ A) clobbers(A)             ; command serial bus device to LISTEN
+extsub $FFB4 = TALK(ubyte device @ A) clobbers(A)               ; command serial bus device to TALK
+extsub $FFB7 = READST() -> ubyte @ A                            ; read I/O status word  (use CLEARST to reset it to 0)
+extsub $FFBA = SETLFS(ubyte logical @ A, ubyte device @ X, ubyte secondary @ Y)   ; set logical file parameters
+extsub $FFBD = SETNAM(ubyte namelen @ A, str filename @ XY)     ; set filename parameters
+extsub $FFC0 = OPEN() clobbers(X,Y) -> bool @Pc, ubyte @A       ; (via 794 ($31A)) open a logical file
+extsub $FFC3 = CLOSE(ubyte logical @ A) clobbers(A,X,Y)         ; (via 796 ($31C)) close a logical file
+extsub $FFC6 = CHKIN(ubyte logical @ X) clobbers(A,X) -> bool @Pc    ; (via 798 ($31E)) define an input channel
+extsub $FFC9 = CHKOUT(ubyte logical @ X) clobbers(A,X)          ; (via 800 ($320)) define an output channel
+extsub $FFCC = CLRCHN() clobbers(A,X)                           ; (via 802 ($322)) restore default devices
+extsub $FFCF = CHRIN() clobbers(X, Y) -> ubyte @ A   ; (via 804 ($324)) input a character (for keyboard, read a whole line from the screen) A=byte read.
+extsub $FFD2 = CHROUT(ubyte character @ A)                           ; (via 806 ($326)) output a character
+extsub $FFD5 = LOAD(ubyte verify @ A, uword address @ XY) -> bool @Pc, ubyte @ A, uword @ XY     ; (via 816 ($330)) load from device
+extsub $FFD8 = SAVE(ubyte zp_startaddr @ A, uword endaddr @ XY) clobbers (X, Y) -> bool @ Pc, ubyte @ A       ; (via 818 ($332)) save to a device.  See also BSAVE
+extsub $FFDB = SETTIM(ubyte low @ A, ubyte middle @ X, ubyte high @ Y)      ; set the software clock
+extsub $FFDE = RDTIM() -> ubyte @ A, ubyte @ X, ubyte @ Y       ; read the software clock (in little endian order: A=lo,X=mid,Y=high) , however use RDTIM_safe() instead
+extsub $FFE1 = STOP() clobbers(X) -> bool @ Pz, ubyte @ A       ; (via 808 ($328)) check the STOP key (and some others in A)        also see STOP2
+extsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A     ; (via 810 ($32A)) get a character      also see GETIN2
+extsub $FFE7 = CLALL() clobbers(A,X)                            ; (via 812 ($32C)) close all files
+extsub $FFEA = UDTIM() clobbers(A,X)                            ; update the software clock
+extsub $FFED = SCREEN() -> ubyte @ X, ubyte @ Y                 ; get size of text screen into X (columns) and Y (rows)
+extsub $FFF0 = PLOT(ubyte col @ Y, ubyte row @ X, bool dir @ Pc) clobbers(A) -> ubyte @ X, ubyte @ Y       ; read/set position of cursor on screen.  Also see txt.plot
+extsub $FFF3 = IOBASE() -> uword @ XY                           ; read base address of I/O devices
 
 ; ---- utility
 
@@ -332,7 +332,7 @@ cx16 {
     &ubyte  VERA_FX_X_POS_S     = VERA_BASE + $0009
     &ubyte  VERA_FX_Y_POS_S     = VERA_BASE + $000a
     &ubyte  VERA_FX_POLY_FILL_L = VERA_BASE + $000b
-    &ubyte  VERA_FX_POLY_FILL_H = VERA_BASE + $000c
+    &ubyte  VERA_FX_POLY_FILL_H = VERA_BASE + $000cF
     &uword  VERA_FX_POLY_FILL   = VERA_BASE + $000b
     &ubyte  VERA_FX_CACHE_L     = VERA_BASE + $0009
     &ubyte  VERA_FX_CACHE_M     = VERA_BASE + $000a
@@ -394,152 +394,152 @@ cx16 {
 ; ---- Commander X-16 additions on top of C64 kernal routines ----
 ; spelling of the names is taken from the Commander X-16 rom sources
 
-romsub $ff4a = CLOSE_ALL(ubyte device @A)  clobbers(A,X,Y)
-romsub $ff59 = LKUPLA(ubyte la @A)  clobbers(A,X,Y)
-romsub $ff5c = LKUPSA(ubyte sa @Y)  clobbers(A,X,Y)
-romsub $ff5f = screen_mode(ubyte mode @A, bool getCurrent @Pc) -> ubyte @A, ubyte @X, ubyte @Y, bool @Pc        ; also see SCREEN or get_screen_mode()
-romsub $ff62 = screen_set_charset(ubyte charset @A, uword charsetptr @XY)  clobbers(A,X,Y)
-romsub $ff6e = JSRFAR()  ; following word = address to call, byte after that=rom/ram bank it is in
-romsub $ff74 = fetch(ubyte zp_startaddr @A, ubyte bank @X, ubyte index @Y)  clobbers(X) -> ubyte @A
-romsub $ff77 = stash(ubyte data @A, ubyte bank @X, ubyte index @Y)  clobbers(X)     ;  note: The the zero page address containing the base address is passed in stavec ($03B2)
-romsub $ff7d = PRIMM()
+extsub $ff4a = CLOSE_ALL(ubyte device @A)  clobbers(A,X,Y)
+extsub $ff59 = LKUPLA(ubyte la @A)  clobbers(A,X,Y)
+extsub $ff5c = LKUPSA(ubyte sa @Y)  clobbers(A,X,Y)
+extsub $ff5f = screen_mode(ubyte mode @A, bool getCurrent @Pc) -> ubyte @A, ubyte @X, ubyte @Y, bool @Pc        ; also see SCREEN or get_screen_mode()
+extsub $ff62 = screen_set_charset(ubyte charset @A, uword charsetptr @XY)  clobbers(A,X,Y)
+extsub $ff6e = JSRFAR()  ; following word = address to call, byte after that=rom/ram bank it is in
+extsub $ff74 = fetch(ubyte zp_startaddr @A, ubyte bank @X, ubyte index @Y)  clobbers(X) -> ubyte @A
+extsub $ff77 = stash(ubyte data @A, ubyte bank @X, ubyte index @Y)  clobbers(X)     ;  note: The the zero page address containing the base address is passed in stavec ($03B2)
+extsub $ff7d = PRIMM()
 
 ; high level graphics & fonts
-romsub $ff20 = GRAPH_init(uword vectors @R0)  clobbers(A,X,Y)
-romsub $ff23 = GRAPH_clear()  clobbers(A,X,Y)
-romsub $ff26 = GRAPH_set_window(uword x @R0, uword y @R1, uword width @R2, uword height @R3)  clobbers(A,X,Y)
-romsub $ff29 = GRAPH_set_colors(ubyte stroke @A, ubyte fill @X, ubyte background @Y)  clobbers (A,X,Y)
-romsub $ff2c = GRAPH_draw_line(uword x1 @R0, uword y1 @R1, uword x2 @R2, uword y2 @R3)  clobbers(A,X,Y)
-romsub $ff2f = GRAPH_draw_rect(uword x @R0, uword y @R1, uword width @R2, uword height @R3, uword cornerradius @R4, bool fill @Pc)  clobbers(A,X,Y)
-romsub $ff32 = GRAPH_move_rect(uword sx @R0, uword sy @R1, uword tx @R2, uword ty @R3, uword width @R4, uword height @R5)  clobbers(A,X,Y)
-romsub $ff35 = GRAPH_draw_oval(uword x @R0, uword y @R1, uword width @R2, uword height @R3, bool fill @Pc)  clobbers(A,X,Y)
-romsub $ff38 = GRAPH_draw_image(uword x @R0, uword y @R1, uword ptr @R2, uword width @R3, uword height @R4)  clobbers(A,X,Y)
-romsub $ff3b = GRAPH_set_font(uword fontptr @R0)  clobbers(A,X,Y)
-romsub $ff3e = GRAPH_get_char_size(ubyte baseline @A, ubyte width @X, ubyte height_or_style @Y, bool is_control @Pc)  clobbers(A,X,Y)
-romsub $ff41 = GRAPH_put_char(uword x @R0, uword y @R1, ubyte character @A)  clobbers(A,X,Y)
-romsub $ff41 = GRAPH_put_next_char(ubyte character @A)  clobbers(A,X,Y)     ; alias for the routine above that doesn't reset the position of the initial character
+extsub $ff20 = GRAPH_init(uword vectors @R0)  clobbers(A,X,Y)
+extsub $ff23 = GRAPH_clear()  clobbers(A,X,Y)
+extsub $ff26 = GRAPH_set_window(uword x @R0, uword y @R1, uword width @R2, uword height @R3)  clobbers(A,X,Y)
+extsub $ff29 = GRAPH_set_colors(ubyte stroke @A, ubyte fill @X, ubyte background @Y)  clobbers (A,X,Y)
+extsub $ff2c = GRAPH_draw_line(uword x1 @R0, uword y1 @R1, uword x2 @R2, uword y2 @R3)  clobbers(A,X,Y)
+extsub $ff2f = GRAPH_draw_rect(uword x @R0, uword y @R1, uword width @R2, uword height @R3, uword cornerradius @R4, bool fill @Pc)  clobbers(A,X,Y)
+extsub $ff32 = GRAPH_move_rect(uword sx @R0, uword sy @R1, uword tx @R2, uword ty @R3, uword width @R4, uword height @R5)  clobbers(A,X,Y)
+extsub $ff35 = GRAPH_draw_oval(uword x @R0, uword y @R1, uword width @R2, uword height @R3, bool fill @Pc)  clobbers(A,X,Y)
+extsub $ff38 = GRAPH_draw_image(uword x @R0, uword y @R1, uword ptr @R2, uword width @R3, uword height @R4)  clobbers(A,X,Y)
+extsub $ff3b = GRAPH_set_font(uword fontptr @R0)  clobbers(A,X,Y)
+extsub $ff3e = GRAPH_get_char_size(ubyte baseline @A, ubyte width @X, ubyte height_or_style @Y, bool is_control @Pc)  clobbers(A,X,Y)
+extsub $ff41 = GRAPH_put_char(uword x @R0, uword y @R1, ubyte character @A)  clobbers(A,X,Y)
+extsub $ff41 = GRAPH_put_next_char(ubyte character @A)  clobbers(A,X,Y)     ; alias for the routine above that doesn't reset the position of the initial character
 
 ; framebuffer
-romsub $fef6 = FB_init()  clobbers(A,X,Y)
-romsub $fef9 = FB_get_info()  clobbers(X,Y) -> byte @A, uword @R0, uword @R1    ; width=r0, height=r1
-romsub $fefc = FB_set_palette(uword pointer @R0, ubyte index @A, ubyte colorcount @X)  clobbers(A,X,Y)
-romsub $feff = FB_cursor_position(uword x @R0, uword y @R1)  clobbers(A,X,Y)
-romsub $ff02 = FB_cursor_next_line(uword x @R0)  clobbers(A,X,Y)
-romsub $ff05 = FB_get_pixel()  clobbers(X,Y) -> ubyte @A
-romsub $ff08 = FB_get_pixels(uword pointer @R0, uword count @R1)  clobbers(A,X,Y)
-romsub $ff0b = FB_set_pixel(ubyte color @A)  clobbers(A,X,Y)
-romsub $ff0e = FB_set_pixels(uword pointer @R0, uword count @R1)  clobbers(A,X,Y)
-romsub $ff11 = FB_set_8_pixels(ubyte pattern @A, ubyte color @X)  clobbers(A,X,Y)
-romsub $ff14 = FB_set_8_pixels_opaque(ubyte pattern @R0, ubyte mask @A, ubyte color1 @X, ubyte color2 @Y)  clobbers(A,X,Y)
-romsub $ff17 = FB_fill_pixels(uword count @R0, uword pstep @R1, ubyte color @A)  clobbers(A,X,Y)
-romsub $ff1a = FB_filter_pixels(uword pointer @ R0, uword count @R1)  clobbers(A,X,Y)
-romsub $ff1d = FB_move_pixels(uword sx @R0, uword sy @R1, uword tx @R2, uword ty @R3, uword count @R4)  clobbers(A,X,Y)
+extsub $fef6 = FB_init()  clobbers(A,X,Y)
+extsub $fef9 = FB_get_info()  clobbers(X,Y) -> byte @A, uword @R0, uword @R1    ; width=r0, height=r1
+extsub $fefc = FB_set_palette(uword pointer @R0, ubyte index @A, ubyte colorcount @X)  clobbers(A,X,Y)
+extsub $feff = FB_cursor_position(uword x @R0, uword y @R1)  clobbers(A,X,Y)
+extsub $ff02 = FB_cursor_next_line(uword x @R0)  clobbers(A,X,Y)
+extsub $ff05 = FB_get_pixel()  clobbers(X,Y) -> ubyte @A
+extsub $ff08 = FB_get_pixels(uword pointer @R0, uword count @R1)  clobbers(A,X,Y)
+extsub $ff0b = FB_set_pixel(ubyte color @A)  clobbers(A,X,Y)
+extsub $ff0e = FB_set_pixels(uword pointer @R0, uword count @R1)  clobbers(A,X,Y)
+extsub $ff11 = FB_set_8_pixels(ubyte pattern @A, ubyte color @X)  clobbers(A,X,Y)
+extsub $ff14 = FB_set_8_pixels_opaque(ubyte pattern @R0, ubyte mask @A, ubyte color1 @X, ubyte color2 @Y)  clobbers(A,X,Y)
+extsub $ff17 = FB_fill_pixels(uword count @R0, uword pstep @R1, ubyte color @A)  clobbers(A,X,Y)
+extsub $ff1a = FB_filter_pixels(uword pointer @ R0, uword count @R1)  clobbers(A,X,Y)
+extsub $ff1d = FB_move_pixels(uword sx @R0, uword sy @R1, uword tx @R2, uword ty @R3, uword count @R4)  clobbers(A,X,Y)
 
 ; misc
-romsub $fec6 = i2c_read_byte(ubyte device @X, ubyte offset @Y) clobbers (X,Y) -> ubyte @A, bool @Pc
-romsub $fec9 = i2c_write_byte(ubyte device @X, ubyte offset @Y, ubyte data @A) clobbers (A,X,Y) -> bool @Pc
-romsub $feb4 = i2c_batch_read(ubyte device @X, uword buffer @R0, uword length @R1, bool advance @Pc) clobbers(A,Y) -> bool @Pc
-romsub $feb7 = i2c_batch_write(ubyte device @X, uword buffer @R0, uword length @R1, bool advance @Pc) clobbers(A,Y) -> bool @Pc
+extsub $fec6 = i2c_read_byte(ubyte device @X, ubyte offset @Y) clobbers (X,Y) -> ubyte @A, bool @Pc
+extsub $fec9 = i2c_write_byte(ubyte device @X, ubyte offset @Y, ubyte data @A) clobbers (A,X,Y) -> bool @Pc
+extsub $feb4 = i2c_batch_read(ubyte device @X, uword buffer @R0, uword length @R1, bool advance @Pc) clobbers(A,Y) -> bool @Pc
+extsub $feb7 = i2c_batch_write(ubyte device @X, uword buffer @R0, uword length @R1, bool advance @Pc) clobbers(A,Y) -> bool @Pc
 
-romsub $fef0 = sprite_set_image(uword pixels @R0, uword mask @R1, ubyte bpp @R2, ubyte number @A, ubyte width @X, ubyte height @Y, bool apply_mask @Pc)  clobbers(A,X,Y) -> bool @Pc
-romsub $fef3 = sprite_set_position(uword x @R0, uword y @R1, ubyte number @A)  clobbers(A,X,Y)
-romsub $fee4 = memory_fill(uword address @R0, uword num_bytes @R1, ubyte value @A)  clobbers(A,X,Y)
-romsub $fee7 = memory_copy(uword source @R0, uword target @R1, uword num_bytes @R2)  clobbers(A,X,Y)
-romsub $feea = memory_crc(uword address @R0, uword num_bytes @R1)  clobbers(A,X,Y) -> uword @R2
-romsub $feed = memory_decompress(uword input @R0, uword output @R1)  clobbers(A,X,Y) -> uword @R1       ; last address +1 is result in R1
-romsub $fedb = console_init(uword x @R0, uword y @R1, uword width @R2, uword height @R3)  clobbers(A,X,Y)
-romsub $fede = console_put_char(ubyte character @A, bool wrapping @Pc)  clobbers(A,X,Y)
-romsub $fee1 = console_get_char()  clobbers(X,Y) -> ubyte @A
-romsub $fed8 = console_put_image(uword pointer @R0, uword width @R1, uword height @R2)  clobbers(A,X,Y)
-romsub $fed5 = console_set_paging_message(uword msgptr @R0)  clobbers(A,X,Y)
-romsub $fecf = entropy_get() -> ubyte @A, ubyte @X, ubyte @Y
-;; romsub $fea8 = extapi16(ubyte callnumber @A) clobbers (A,X,Y)    ; not useful yet because is for 65816 cpu
-romsub $feab = extapi(ubyte callnumber @A) clobbers (A,X,Y)
-romsub $fecc = monitor()  clobbers(A,X,Y)
+extsub $fef0 = sprite_set_image(uword pixels @R0, uword mask @R1, ubyte bpp @R2, ubyte number @A, ubyte width @X, ubyte height @Y, bool apply_mask @Pc)  clobbers(A,X,Y) -> bool @Pc
+extsub $fef3 = sprite_set_position(uword x @R0, uword y @R1, ubyte number @A)  clobbers(A,X,Y)
+extsub $fee4 = memory_fill(uword address @R0, uword num_bytes @R1, ubyte value @A)  clobbers(A,X,Y)
+extsub $fee7 = memory_copy(uword source @R0, uword target @R1, uword num_bytes @R2)  clobbers(A,X,Y)
+extsub $feea = memory_crc(uword address @R0, uword num_bytes @R1)  clobbers(A,X,Y) -> uword @R2
+extsub $feed = memory_decompress(uword input @R0, uword output @R1)  clobbers(A,X,Y) -> uword @R1       ; last address +1 is result in R1
+extsub $fedb = console_init(uword x @R0, uword y @R1, uword width @R2, uword height @R3)  clobbers(A,X,Y)
+extsub $fede = console_put_char(ubyte character @A, bool wrapping @Pc)  clobbers(A,X,Y)
+extsub $fee1 = console_get_char()  clobbers(X,Y) -> ubyte @A
+extsub $fed8 = console_put_image(uword pointer @R0, uword width @R1, uword height @R2)  clobbers(A,X,Y)
+extsub $fed5 = console_set_paging_message(uword msgptr @R0)  clobbers(A,X,Y)
+extsub $fecf = entropy_get() -> ubyte @A, ubyte @X, ubyte @Y
+;; extsub $fea8 = extapi16(ubyte callnumber @A) clobbers (A,X,Y)    ; not useful yet because is for 65816 cpu
+extsub $feab = extapi(ubyte callnumber @A) clobbers (A,X,Y)
+extsub $fecc = monitor()  clobbers(A,X,Y)
 
-romsub $ff44 = MACPTR(ubyte length @A, uword buffer @XY, bool dontAdvance @Pc)  clobbers(A) -> bool @Pc, uword @XY
-romsub $feb1 = MCIOUT(ubyte length @A, uword buffer @XY, bool dontAdvance @Pc)  clobbers(A) -> bool @Pc, uword @XY
-romsub $FEBA = BSAVE(ubyte zp_startaddr @ A, uword endaddr @ XY) clobbers (X, Y) -> bool @ Pc, ubyte @ A      ; like cbm.SAVE, but omits the 2-byte prg header
-romsub $ff47 = enter_basic(bool cold_or_warm @Pc)  clobbers(A,X,Y)
-romsub $ff4d = clock_set_date_time(uword yearmonth @R0, uword dayhours @R1, uword minsecs @R2, uword jiffiesweekday @R3)  clobbers(A, X, Y)
-romsub $ff50 = clock_get_date_time()  clobbers(A, X, Y)  -> uword @R0, uword @R1, uword @R2, uword @R3   ; result registers see clock_set_date_time()
+extsub $ff44 = MACPTR(ubyte length @A, uword buffer @XY, bool dontAdvance @Pc)  clobbers(A) -> bool @Pc, uword @XY
+extsub $feb1 = MCIOUT(ubyte length @A, uword buffer @XY, bool dontAdvance @Pc)  clobbers(A) -> bool @Pc, uword @XY
+extsub $FEBA = BSAVE(ubyte zp_startaddr @ A, uword endaddr @ XY) clobbers (X, Y) -> bool @ Pc, ubyte @ A      ; like cbm.SAVE, but omits the 2-byte prg header
+extsub $ff47 = enter_basic(bool cold_or_warm @Pc)  clobbers(A,X,Y)
+extsub $ff4d = clock_set_date_time(uword yearmonth @R0, uword dayhours @R1, uword minsecs @R2, uword jiffiesweekday @R3)  clobbers(A, X, Y)
+extsub $ff50 = clock_get_date_time()  clobbers(A, X, Y)  -> uword @R0, uword @R1, uword @R2, uword @R3   ; result registers see clock_set_date_time()
 
 ; keyboard, mouse, joystick
 ; note: also see the cbm.kbdbuf_clear() helper routine
-romsub $febd = kbdbuf_peek() -> ubyte @A, ubyte @X     ; key in A, queue length in X
-romsub $fec0 = kbdbuf_get_modifiers() -> ubyte @A
-romsub $fec3 = kbdbuf_put(ubyte key @A) clobbers(X)
-romsub $fed2 = keymap(uword identifier @XY, bool read @Pc) -> bool @Pc
-romsub $ff68 = mouse_config(byte shape @A, ubyte resX @X, ubyte resY @Y)  clobbers (A, X, Y)
-romsub $ff6b = mouse_get(ubyte zdataptr @X) -> ubyte @A, byte @X    ;  use mouse_pos() instead
-romsub $ff71 = mouse_scan()  clobbers(A, X, Y)
-romsub $ff53 = joystick_scan()  clobbers(A, X, Y)
-romsub $ff56 = joystick_get(ubyte joynr @A) -> uword @AX, bool @Y   ; note: everything is inverted
+extsub $febd = kbdbuf_peek() -> ubyte @A, ubyte @X     ; key in A, queue length in X
+extsub $fec0 = kbdbuf_get_modifiers() -> ubyte @A
+extsub $fec3 = kbdbuf_put(ubyte key @A) clobbers(X)
+extsub $fed2 = keymap(uword identifier @XY, bool read @Pc) -> bool @Pc
+extsub $ff68 = mouse_config(byte shape @A, ubyte resX @X, ubyte resY @Y)  clobbers (A, X, Y)
+extsub $ff6b = mouse_get(ubyte zdataptr @X) -> ubyte @A, byte @X    ;  use mouse_pos() instead
+extsub $ff71 = mouse_scan()  clobbers(A, X, Y)
+extsub $ff53 = joystick_scan()  clobbers(A, X, Y)
+extsub $ff56 = joystick_get(ubyte joynr @A) -> uword @AX, bool @Y   ; note: everything is inverted
 
 ; X16Edit (rom bank 13/14 but you ideally should use the routine search_x16edit() to search for the correct bank)
-romsub $C000 = x16edit_default() clobbers(A,X,Y)
-romsub $C003 = x16edit_loadfile(ubyte firstbank @X, ubyte lastbank @Y, str filename @R0, ubyte filenameLength @R1) clobbers(A,X,Y)
-romsub $C006 = x16edit_loadfile_options(ubyte firstbank @X, ubyte lastbank @Y, str filename @R0,
+extsub $C000 = x16edit_default() clobbers(A,X,Y)
+extsub $C003 = x16edit_loadfile(ubyte firstbank @X, ubyte lastbank @Y, str filename @R0, ubyte filenameLength @R1) clobbers(A,X,Y)
+extsub $C006 = x16edit_loadfile_options(ubyte firstbank @X, ubyte lastbank @Y, str filename @R0,
                 uword filenameLengthAndOptions @R1, uword tabstopAndWordwrap @R2,
                 uword disknumberAndColors @R3, uword headerAndStatusColors @R4) clobbers(A,X,Y)
 
-; Audio (rom bank 10 - you have to activate this bank manually first! Or use the stubs in the audio module. Or use callfar/JSRFAR manually.)
-romsub $C09F = audio_init() clobbers(A,X,Y) -> bool @Pc     ; (re)initialize both vera PSG and YM audio chips
-romsub $C000 = bas_fmfreq(ubyte channel @A, uword freq @XY, bool noretrigger @Pc) clobbers(A,X,Y) -> bool @Pc
-romsub $C003 = bas_fmnote(ubyte channel @A, ubyte note @X, ubyte fracsemitone @Y, bool noretrigger @Pc) clobbers(A,X,Y) -> bool @Pc
-romsub $C006 = bas_fmplaystring(ubyte length @A, str string @XY) clobbers(A,X,Y)
-romsub $C009 = bas_fmvib(ubyte speed @A, ubyte depth @X) clobbers(A,X,Y) -> bool @Pc
-romsub $C00C = bas_playstringvoice(ubyte channel @A) clobbers(Y)
-romsub $C00F = bas_psgfreq(ubyte voice @A, uword freq @XY) clobbers(A,X,Y) -> bool @Pc
-romsub $C012 = bas_psgnote(ubyte voice @A, ubyte note @X, ubyte fracsemitone @Y) clobbers(A,X,Y) -> bool @Pc
-romsub $C015 = bas_psgwav(ubyte voice @A, ubyte waveform @X) clobbers(A,X,Y) -> bool @Pc
-romsub $C018 = bas_psgplaystring(ubyte length @A, str string @XY) clobbers(A,X,Y)
-romsub $C08D = bas_fmchordstring(ubyte length @A, str string @XY) clobbers(A,X,Y)
-romsub $C090 = bas_psgchordstring(ubyte length @A, str string @XY) clobbers(A,X,Y)
-romsub $C01B = notecon_bas2fm(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
-romsub $C01E = notecon_bas2midi(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
-romsub $C021 = notecon_bas2psg(ubyte note @X, ubyte fracsemitone @Y) clobbers(A) -> uword @XY, bool @Pc
-romsub $C024 = notecon_fm2bas(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
-romsub $C027 = notecon_fm2midi(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
-romsub $C02A = notecon_fm2psg(ubyte note @X, ubyte fracsemitone @Y) clobbers(A) -> uword @XY, bool @Pc
-romsub $C02D = notecon_freq2bas(uword freqHz @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
-romsub $C030 = notecon_freq2fm(uword freqHz @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
-romsub $C033 = notecon_freq2midi(uword freqHz @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
-romsub $C036 = notecon_freq2psg(uword freqHz @XY) clobbers(A) -> uword @XY, bool @Pc
-romsub $C039 = notecon_midi2bas(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
-romsub $C03C = notecon_midi2fm(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
-romsub $C03F = notecon_midi2psg(ubyte note @X, ubyte fracsemitone @Y) clobbers(A) -> uword @XY, bool @Pc
-romsub $C042 = notecon_psg2bas(uword freq @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
-romsub $C045 = notecon_psg2fm(uword freq @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
-romsub $C048 = notecon_psg2midi(uword freq @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
-romsub $C04B = psg_init() clobbers(A,X,Y)               ; (re)init Vera PSG
-romsub $C04E = psg_playfreq(ubyte voice @A, uword freq @XY) clobbers(A,X,Y)
-romsub $C051 = psg_read(ubyte offset @X, bool cookedVol @Pc) clobbers(Y) -> ubyte @A
-romsub $C054 = psg_setatten(ubyte voice @A, ubyte attenuation @X) clobbers(A,X,Y)
-romsub $C057 = psg_setfreq(ubyte voice @A, uword freq @XY) clobbers(A,X,Y)
-romsub $C05A = psg_setpan(ubyte voice @A, ubyte panning @X) clobbers(A,X,Y)
-romsub $C05D = psg_setvol(ubyte voice @A, ubyte volume @X) clobbers(A,X,Y)
-romsub $C060 = psg_write(ubyte value @A, ubyte offset @X) clobbers(Y)
-romsub $C0A2 = psg_write_fast(ubyte value @A, ubyte offset @X) clobbers(Y)
-romsub $C093 = psg_getatten(ubyte voice @A) clobbers(Y) -> ubyte @X
-romsub $C096 = psg_getpan(ubyte voice @A) clobbers(Y) -> ubyte @X
-romsub $C063 = ym_init() clobbers(A,X,Y) -> bool @Pc              ; (re)init YM chip
-romsub $C066 = ym_loaddefpatches() clobbers(A,X,Y) -> bool @Pc    ; load default YM patches
-romsub $C069 = ym_loadpatch(ubyte channel @A, uword patchOrAddress @XY, bool what @Pc) clobbers(A,X,Y)
-romsub $C06C = ym_loadpatchlfn(ubyte channel @A, ubyte lfn @X) clobbers(X,Y) -> ubyte @A, bool @Pc
-romsub $C06F = ym_playdrum(ubyte channel @A, ubyte note @X) clobbers(A,X,Y) -> bool @Pc
-romsub $C072 = ym_playnote(ubyte channel @A, ubyte kc @X, ubyte kf @Y, bool notrigger @Pc) clobbers(A,X,Y) -> bool @Pc
-romsub $C075 = ym_setatten(ubyte channel @A, ubyte attenuation @X) clobbers(Y) -> bool @Pc
-romsub $C078 = ym_setdrum(ubyte channel @A, ubyte note @X) clobbers(A,X,Y) -> bool @Pc
-romsub $C07B = ym_setnote(ubyte channel @A, ubyte kc @X, ubyte kf @Y) clobbers(A,X,Y) -> bool @Pc
-romsub $C07E = ym_setpan(ubyte channel @A, ubyte panning @X) clobbers(A,X,Y) -> bool @Pc
-romsub $C081 = ym_read(ubyte register @X, bool cooked @Pc) clobbers(Y) -> ubyte @A, bool @Pc
-romsub $C084 = ym_release(ubyte channel @A) clobbers(A,X,Y) -> bool @Pc
-romsub $C087 = ym_trigger(ubyte channel @A, bool noRelease @Pc) clobbers(A,X,Y) -> bool @Pc
-romsub $C08A = ym_write(ubyte value @A, ubyte register @X) clobbers(Y) -> bool @Pc
-romsub $C099 = ym_getatten(ubyte channel @A) clobbers(Y) -> ubyte @X
-romsub $C09C = ym_getpan(ubyte channel @A) clobbers(Y) -> ubyte @X
-romsub $C0A5 = ym_get_chip_type() clobbers(X) -> ubyte @A
+; Audio (rom bank 10)
+extsub @bank 10  $C09F = audio_init() clobbers(A,X,Y) -> bool @Pc     ; (re)initialize both vera PSG and YM audio chips
+extsub @bank 10  $C000 = bas_fmfreq(ubyte channel @A, uword freq @XY, bool noretrigger @Pc) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C003 = bas_fmnote(ubyte channel @A, ubyte note @X, ubyte fracsemitone @Y, bool noretrigger @Pc) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C006 = bas_fmplaystring(ubyte length @A, str string @XY) clobbers(A,X,Y)
+extsub @bank 10  $C009 = bas_fmvib(ubyte speed @A, ubyte depth @X) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C00C = bas_playstringvoice(ubyte channel @A) clobbers(Y)
+extsub @bank 10  $C00F = bas_psgfreq(ubyte voice @A, uword freq @XY) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C012 = bas_psgnote(ubyte voice @A, ubyte note @X, ubyte fracsemitone @Y) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C015 = bas_psgwav(ubyte voice @A, ubyte waveform @X) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C018 = bas_psgplaystring(ubyte length @A, str string @XY) clobbers(A,X,Y)
+extsub @bank 10  $C08D = bas_fmchordstring(ubyte length @A, str string @XY) clobbers(A,X,Y)
+extsub @bank 10  $C090 = bas_psgchordstring(ubyte length @A, str string @XY) clobbers(A,X,Y)
+extsub @bank 10  $C01B = notecon_bas2fm(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
+extsub @bank 10  $C01E = notecon_bas2midi(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
+extsub @bank 10  $C021 = notecon_bas2psg(ubyte note @X, ubyte fracsemitone @Y) clobbers(A) -> uword @XY, bool @Pc
+extsub @bank 10  $C024 = notecon_fm2bas(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
+extsub @bank 10  $C027 = notecon_fm2midi(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
+extsub @bank 10  $C02A = notecon_fm2psg(ubyte note @X, ubyte fracsemitone @Y) clobbers(A) -> uword @XY, bool @Pc
+extsub @bank 10  $C02D = notecon_freq2bas(uword freqHz @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
+extsub @bank 10  $C030 = notecon_freq2fm(uword freqHz @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
+extsub @bank 10  $C033 = notecon_freq2midi(uword freqHz @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
+extsub @bank 10  $C036 = notecon_freq2psg(uword freqHz @XY) clobbers(A) -> uword @XY, bool @Pc
+extsub @bank 10  $C039 = notecon_midi2bas(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
+extsub @bank 10  $C03C = notecon_midi2fm(ubyte note @X) clobbers(A) -> ubyte @X, bool @Pc
+extsub @bank 10  $C03F = notecon_midi2psg(ubyte note @X, ubyte fracsemitone @Y) clobbers(A) -> uword @XY, bool @Pc
+extsub @bank 10  $C042 = notecon_psg2bas(uword freq @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
+extsub @bank 10  $C045 = notecon_psg2fm(uword freq @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
+extsub @bank 10  $C048 = notecon_psg2midi(uword freq @XY) clobbers(A) -> ubyte @X, ubyte @Y, bool @Pc
+extsub @bank 10  $C04B = psg_init() clobbers(A,X,Y)               ; (re)init Vera PSG
+extsub @bank 10  $C04E = psg_playfreq(ubyte voice @A, uword freq @XY) clobbers(A,X,Y)
+extsub @bank 10  $C051 = psg_read(ubyte offset @X, bool cookedVol @Pc) clobbers(Y) -> ubyte @A
+extsub @bank 10  $C054 = psg_setatten(ubyte voice @A, ubyte attenuation @X) clobbers(A,X,Y)
+extsub @bank 10  $C057 = psg_setfreq(ubyte voice @A, uword freq @XY) clobbers(A,X,Y)
+extsub @bank 10  $C05A = psg_setpan(ubyte voice @A, ubyte panning @X) clobbers(A,X,Y)
+extsub @bank 10  $C05D = psg_setvol(ubyte voice @A, ubyte volume @X) clobbers(A,X,Y)
+extsub @bank 10  $C060 = psg_write(ubyte value @A, ubyte offset @X) clobbers(Y)
+extsub @bank 10  $C0A2 = psg_write_fast(ubyte value @A, ubyte offset @X) clobbers(Y)
+extsub @bank 10  $C093 = psg_getatten(ubyte voice @A) clobbers(Y) -> ubyte @X
+extsub @bank 10  $C096 = psg_getpan(ubyte voice @A) clobbers(Y) -> ubyte @X
+extsub @bank 10  $C063 = ym_init() clobbers(A,X,Y) -> bool @Pc              ; (re)init YM chip
+extsub @bank 10  $C066 = ym_loaddefpatches() clobbers(A,X,Y) -> bool @Pc    ; load default YM patches
+extsub @bank 10  $C069 = ym_loadpatch(ubyte channel @A, uword patchOrAddress @XY, bool what @Pc) clobbers(A,X,Y)
+extsub @bank 10  $C06C = ym_loadpatchlfn(ubyte channel @A, ubyte lfn @X) clobbers(X,Y) -> ubyte @A, bool @Pc
+extsub @bank 10  $C06F = ym_playdrum(ubyte channel @A, ubyte note @X) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C072 = ym_playnote(ubyte channel @A, ubyte kc @X, ubyte kf @Y, bool notrigger @Pc) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C075 = ym_setatten(ubyte channel @A, ubyte attenuation @X) clobbers(Y) -> bool @Pc
+extsub @bank 10  $C078 = ym_setdrum(ubyte channel @A, ubyte note @X) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C07B = ym_setnote(ubyte channel @A, ubyte kc @X, ubyte kf @Y) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C07E = ym_setpan(ubyte channel @A, ubyte panning @X) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C081 = ym_read(ubyte register @X, bool cooked @Pc) clobbers(Y) -> ubyte @A, bool @Pc
+extsub @bank 10  $C084 = ym_release(ubyte channel @A) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C087 = ym_trigger(ubyte channel @A, bool noRelease @Pc) clobbers(A,X,Y) -> bool @Pc
+extsub @bank 10  $C08A = ym_write(ubyte value @A, ubyte register @X) clobbers(Y) -> bool @Pc
+extsub @bank 10  $C099 = ym_getatten(ubyte channel @A) clobbers(Y) -> ubyte @X
+extsub @bank 10  $C09C = ym_getpan(ubyte channel @A) clobbers(Y) -> ubyte @X
+extsub @bank 10  $C0A5 = ym_get_chip_type() clobbers(X) -> ubyte @A
 
 ; extapi call numbers
 const ubyte  EXTAPI_clear_status = $01
@@ -734,9 +734,11 @@ asmsub vpeek(ubyte bank @A, uword address @XY) -> ubyte @A {
 }
 
 asmsub vaddr(ubyte bank @A, uword address @R0, ubyte addrsel @R1, byte autoIncrOrDecrByOne @Y) clobbers(A) {
-        ; -- setup the VERA's data address register 0 or 1
-        ;    with optional auto increment or decrement of 1.
+        ; -- setup the VERA's data address register 0 or 1 with optional auto increment or decrement of 1.
+        ;    This is a convenience routine, and not very efficient if you call it often;
+        ;    it's usually better to write a tailor made version of it that accounts for the repeated values.
         ;    Note that the vaddr_autoincr() and vaddr_autodecr() routines allow to set all possible strides, not just 1.
+        ;    Note also that Vera's addrset is reset to 0 on exit, even if you set port #1's address.
         %asm {{
             pha
             lda  cx16.r1
@@ -752,15 +754,19 @@ asmsub vaddr(ubyte bank @A, uword address @R0, ubyte addrsel @R1, byte autoIncrO
             beq  +
             ora  #%00010000
 +           sta  cx16.VERA_ADDR_H
+            stz  cx16.VERA_CTRL
             rts
 +           ora  #%00011000
             sta  cx16.VERA_ADDR_H
+            stz  cx16.VERA_CTRL
             rts
         }}
 }
 
 asmsub vaddr_clone(ubyte port @A) clobbers (A,X,Y) {
     ; -- clones Vera addresses from the given source port to the other one.
+    ;    This is a convenience routine, and not very efficient if you call it often;
+    ;    it's usually better to write a tailor made version of it that accounts for the repeated values.
     %asm {{
         sta  VERA_CTRL
         ldx  VERA_ADDR_L
@@ -774,15 +780,16 @@ asmsub vaddr_clone(ubyte port @A) clobbers (A,X,Y) {
         ply
         sty  VERA_ADDR_H
         eor  #1
-        sta  VERA_CTRL
+        stz  VERA_CTRL
         rts
     }}
 }
 
 asmsub vaddr_autoincr(ubyte bank @A, uword address @R0, ubyte addrsel @R1, uword autoIncrAmount @R2) clobbers(A,Y) {
-        ; -- setup the VERA's data address register 0 or 1
-        ;    including setting up optional auto increment amount.
+        ; -- setup the VERA's data address register 0 or 1, including setting up optional auto increment amount.
         ;    Specifiying an unsupported amount results in amount of zero. See the Vera docs about what amounts are possible.
+        ;    This is a convenience routine, and not very efficient if you call it often;
+        ;    it's usually better to write a tailor made version of it that accounts for the repeated values.
         %asm {{
             jsr  _setup
             lda  cx16.r2H
@@ -791,6 +798,7 @@ asmsub vaddr_autoincr(ubyte bank @A, uword address @R0, ubyte addrsel @R1, uword
             jsr  _determine_incr_bits
 +           ora  P8ZP_SCRATCH_REG
             sta  cx16.VERA_ADDR_H
+            stz  cx16.VERA_CTRL
             rts
 
 _setup      sta  P8ZP_SCRATCH_REG
@@ -838,13 +846,15 @@ _large      ora  cx16.r2L
 +           lda  #0
             rts
 _strides_lsb    .byte   0,1,2,4,8,16,32,64,128,255,255,40,80,160,255,255
+            ; !notreached!
         }}
 }
 
 asmsub vaddr_autodecr(ubyte bank @A, uword address @R0, ubyte addrsel @R1, uword autoDecrAmount @R2) clobbers(A,Y) {
-        ; -- setup the VERA's data address register 0 or 1
-        ;    including setting up optional auto decrement amount.
+        ; -- setup the VERA's data address register 0 or 1 including setting up optional auto decrement amount.
         ;    Specifiying an unsupported amount results in amount of zero. See the Vera docs about what amounts are possible.
+        ;    This is a convenience routine, and not very efficient if you call it often;
+        ;    it's usually better to write a tailor made version of it that accounts for the repeated values.
         %asm {{
             jsr  vaddr_autoincr._setup
             lda  cx16.r2H
@@ -854,6 +864,7 @@ asmsub vaddr_autodecr(ubyte bank @A, uword address @R0, ubyte addrsel @R1, uword
             ora  #%00001000         ; autodecrement
 +           ora  P8ZP_SCRATCH_REG
             sta  cx16.VERA_ADDR_H
+            stz  cx16.VERA_CTRL
             rts
         }}
 }
@@ -955,6 +966,7 @@ asmsub save_virtual_registers() clobbers(A,Y) {
 _cx16_vreg_storage
         .word 0,0,0,0,0,0,0,0
         .word 0,0,0,0,0,0,0,0
+        ; !notreached!
     }}
 }
 
@@ -993,6 +1005,7 @@ asmsub save_vera_context() clobbers(A) {
         sta  _vera_storage+6
         rts
 _vera_storage:  .byte 0,0,0,0,0,0,0,0
+        ; !notreached!
     }}
 }
 
@@ -1315,6 +1328,7 @@ _continue   iny
             bne  _continue
 +           pla
             sta  $00
+            rts
         }}
     }
 
@@ -1358,7 +1372,7 @@ _continue   iny
 sys {
     ; ------- lowlevel system routines --------
 
-    const ubyte target = 16         ;  compilation target specifier.  64 = C64,  128 = C128,  16 = CommanderX16.
+    const ubyte target = 16         ;  compilation target specifier.  255=virtual, 128=C128, 64=C64, 32=PET, 16=CommanderX16, 8=atari800XL, 7=Neo6502
 
     const ubyte sizeof_bool = 1
     const ubyte sizeof_byte = 1
@@ -1367,90 +1381,6 @@ sys {
     const ubyte sizeof_uword = 2
     const ubyte sizeof_float = 5
 
-asmsub  init_system()  {
-    ; Initializes the machine to a sane starting state.
-    ; Called automatically by the loader program logic.
-    %asm {{
-        sei
-        lda  #0
-        tax
-        tay
-        jsr  cx16.mouse_config  ; disable mouse
-        lda  cx16.VERA_DC_VIDEO
-        and  #%00000111 ; retain chroma + output mode
-        sta  P8ZP_SCRATCH_REG
-        lda  #$0a
-        sta  $01        ; rom bank 10 (audio)
-        jsr  cx16.audio_init ; silence
-        stz  $01        ; rom bank 0 (kernal)
-        jsr  cbm.IOINIT
-        jsr  cbm.RESTOR
-        jsr  cbm.CINT
-        lda  cx16.VERA_DC_VIDEO
-        and  #%11111000
-        ora  P8ZP_SCRATCH_REG
-        sta  cx16.VERA_DC_VIDEO  ; restore old output mode
-        lda  #$90       ; black
-        jsr  cbm.CHROUT
-        lda  #1
-        jsr  cbm.CHROUT ; swap fg/bg
-        lda  #$9e       ; yellow
-        jsr  cbm.CHROUT
-        lda  #147       ; clear screen
-        jsr  cbm.CHROUT
-        lda  #8         ; disable charset case switch
-        jsr  cbm.CHROUT
-        lda  #PROG8_VARSHIGH_RAMBANK
-        sta  $00    ; select ram bank
-        lda  #0
-        sta  $01    ; set ROM bank to kernal bank to speed up kernal calls
-        tax
-        tay
-        cli
-        rts
-    }}
-}
-
-asmsub  init_system_phase2()  {
-    %asm {{
-        sei
-        lda  cbm.CINV
-        sta  restore_irq._orig_irqvec
-        lda  cbm.CINV+1
-        sta  restore_irq._orig_irqvec+1
-        lda  #PROG8_VARSHIGH_RAMBANK
-        sta  $00    ; select ram bank
-        stz  $01    ; set ROM bank to kernal bank to speed up kernal calls
-        cli
-        cld
-        clc
-        clv
-        rts
-    }}
-}
-
-asmsub  cleanup_at_exit() {
-    ; executed when the main subroutine does rts
-    %asm {{
-        lda  #1
-        sta  $00        ; ram bank 1
-        lda  #4
-        sta  $01        ; rom bank 4 (basic)
-        stz  $2d        ; hack to reset machine code monitor bank to 0
-        jsr  cbm.CLRCHN		; reset i/o channels
-_exitcodeCarry = *+1
-        lda  #0
-        lsr  a
-_exitcode = *+1
-        lda  #0        ; exit code possibly modified in exit()
-_exitcodeX = *+1
-        ldx  #0
-_exitcodeY = *+1
-        ldy  #0
-        rts
-    }}
-}
-
 asmsub  set_irq(uword handler @AY) clobbers(A)  {
     ; Sets the handler for the VSYNC interrupt, and enable that interrupt.
 	%asm {{
@@ -1499,7 +1429,8 @@ asmsub  restore_irq() clobbers(A) {
 	    cli
 	    rts
 _orig_irqvec    .word  0
-        }}
+        ; !notreached!
+    }}
 }
 
 asmsub  set_rasterirq(uword handler @AY, uword rasterpos @R0) clobbers(A) {
@@ -1698,6 +1629,56 @@ _longcopy
         }}
     }
 
+    asmsub memcmp(uword address1 @R0, uword address2 @R1, uword size @AY) -> byte @A {
+        ; Compares two blocks of memory
+        ; Returns -1 (255), 0 or 1, meaning: block 1 sorts before, equal or after block 2.
+        %asm {{
+            sta  P8ZP_SCRATCH_W1
+            sty  P8ZP_SCRATCH_W1+1
+            ldx  P8ZP_SCRATCH_W1+1
+            beq  _no_msb_size
+
+_loop_msb_size
+            ldy  #0
+-           lda  (cx16.r0),y
+            cmp  (cx16.r1),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            bne  -
+            inc  cx16.r0+1
+            inc  cx16.r1+1
+            dec  P8ZP_SCRATCH_W1+1
+            dex
+            bne  _loop_msb_size
+
+_no_msb_size
+            lda  P8ZP_SCRATCH_W1
+            bne  +
+            rts
+
++           ldy  #0
+-           lda  (cx16.r0),y
+            cmp  (cx16.r1),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            cpy  P8ZP_SCRATCH_W1
+            bne  -
+
+            lda #0
+            rts
+        }}
+    }
+
     inline asmsub read_flags() -> ubyte @A {
         %asm {{
             php
@@ -1775,6 +1756,7 @@ save_SCRATCH_ZPB1	.byte  0
 save_SCRATCH_ZPREG	.byte  0
 save_SCRATCH_ZPWORD1	.word  0
 save_SCRATCH_ZPWORD2	.word  0
+            ; !notreached!
         }}
     }
 
@@ -1799,37 +1781,37 @@ save_SCRATCH_ZPWORD2	.word  0
     asmsub exit(ubyte returnvalue @A) {
         ; -- immediately exit the program with a return code in the A register
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit2(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y) {
         ; -- immediately exit the program with result values in the A, X and Y registers.
         %asm {{
-            sta  cleanup_at_exit._exitcode
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit3(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y, bool carry @Pc) {
         ; -- immediately exit the program with result values in the A, X and Y registers, and the Carry flag in the status register.
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             lda  #0
             rol  a
-            sta  cleanup_at_exit._exitcodeCarry
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcodeCarry
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
@@ -1840,6 +1822,13 @@ save_SCRATCH_ZPWORD2	.word  0
         }}
     }
 
+    inline asmsub progstart() -> uword @AY {
+        %asm {{
+            lda  #<prog8_program_start
+            ldy  #>prog8_program_start
+        }}
+    }
+
     inline asmsub push(ubyte value @A) {
         %asm {{
             pha
@@ -1867,3 +1856,91 @@ save_SCRATCH_ZPWORD2	.word  0
     }
 
 }
+
+p8_sys_startup {
+    ; program startup and shutdown machinery. Needs to reside in normal system ram.
+
+asmsub  init_system()  {
+    ; Initializes the machine to a sane starting state.
+    ; Called automatically by the loader program logic.
+    %asm {{
+        sei
+        lda  #0
+        tax
+        tay
+        jsr  cx16.mouse_config  ; disable mouse
+        lda  cx16.VERA_DC_VIDEO
+        and  #%00000111 ; retain chroma + output mode
+        sta  P8ZP_SCRATCH_REG
+        lda  #$0a
+        sta  $01        ; rom bank 10 (audio)
+        jsr  cx16.audio_init ; silence
+        stz  $01        ; rom bank 0 (kernal)
+        jsr  cbm.IOINIT
+        jsr  cbm.RESTOR
+        jsr  cbm.CINT
+        lda  cx16.VERA_DC_VIDEO
+        and  #%11111000
+        ora  P8ZP_SCRATCH_REG
+        sta  cx16.VERA_DC_VIDEO  ; restore old output mode
+        lda  #$90       ; black
+        jsr  cbm.CHROUT
+        lda  #1
+        jsr  cbm.CHROUT ; swap fg/bg
+        lda  #$9e       ; yellow
+        jsr  cbm.CHROUT
+        lda  #147       ; clear screen
+        jsr  cbm.CHROUT
+        lda  #8         ; disable charset case switch
+        jsr  cbm.CHROUT
+        lda  #PROG8_VARSHIGH_RAMBANK
+        sta  $00    ; select ram bank
+        lda  #0
+        sta  $01    ; set ROM bank to kernal bank to speed up kernal calls
+        tax
+        tay
+        cli
+        rts
+    }}
+}
+
+asmsub  init_system_phase2()  {
+    %asm {{
+        sei
+        lda  cbm.CINV
+        sta  sys.restore_irq._orig_irqvec
+        lda  cbm.CINV+1
+        sta  sys.restore_irq._orig_irqvec+1
+        lda  #PROG8_VARSHIGH_RAMBANK
+        sta  $00    ; select ram bank
+        stz  $01    ; set ROM bank to kernal bank to speed up kernal calls
+        cli
+        cld
+        clc
+        clv
+        rts
+    }}
+}
+
+asmsub  cleanup_at_exit() {
+    ; executed when the main subroutine does rts
+    %asm {{
+        lda  #1
+        sta  $00        ; ram bank 1
+        lda  #4
+        sta  $01        ; rom bank 4 (basic)
+        jsr  cbm.CLRCHN		; reset i/o channels
+_exitcodeCarry = *+1
+        lda  #0
+        lsr  a
+_exitcode = *+1
+        lda  #0        ; exit code possibly modified in exit()
+_exitcodeX = *+1
+        ldx  #0
+_exitcodeY = *+1
+        ldy  #0
+        rts
+    }}
+}
+
+}

compiler/res/prog8lib/cx16/textio.p8

@@ -16,7 +16,7 @@ const ubyte DEFAULT_HEIGHT = 60
 const ubyte VERA_TEXTMATRIX_BANK = 1
 const uword VERA_TEXTMATRIX_ADDR = $b000
 
-romsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
+extsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
 
 
 sub clear_screen() {

compiler/res/prog8lib/cx16/verafx.p8

@@ -28,7 +28,7 @@ verafx {
 
     sub clear(ubyte vbank, uword vaddr, ubyte data, uword num_longwords) {
         ; use cached 4-byte write to quickly clear a portion of the video memory to a given byte value
-        ; this routine is around 3 times faster as gfx2.clear_screen()
+        ; this routine is around 3 times faster as gfx_hires/gfx_lores.clear_screen()
         cx16.VERA_CTRL = 0
         cx16.VERA_ADDR_H = vbank | %00110000       ; 4-byte increment
         cx16.VERA_ADDR_M = msb(vaddr)

compiler/res/prog8lib/math.p8

@@ -5,14 +5,13 @@
 math {
     %option no_symbol_prefixing, ignore_unused
 
-	%asminclude "library:math.asm"
-
     asmsub sin8u(ubyte angle @A) clobbers(Y) -> ubyte @A {
         %asm {{
 		tay
 		lda  _sinecos8u,y
 		rts
 _sinecos8u	.byte  trunc(128.0 + 127.5 * sin(range(256+64) * rad(360.0/256.0)))
+        ; !notreached!
         }}
     }
 
@@ -30,6 +29,7 @@ _sinecos8u	.byte  trunc(128.0 + 127.5 * sin(range(256+64) * rad(360.0/256.0)))
 		lda  _sinecos8,y
 		rts
 _sinecos8	.char  trunc(127.0 * sin(range(256+64) * rad(360.0/256.0)))
+        ; !notreached!
         }}
     }
 
@@ -47,6 +47,7 @@ _sinecos8	.char  trunc(127.0 * sin(range(256+64) * rad(360.0/256.0)))
 		lda  _sinecosR8u,y
 		rts
 _sinecosR8u	.byte  trunc(128.0 + 127.5 * sin(range(180+45) * rad(360.0/180.0)))
+        ; !notreached!
         }}
     }
 
@@ -64,6 +65,7 @@ _sinecosR8u	.byte  trunc(128.0 + 127.5 * sin(range(180+45) * rad(360.0/180.0)))
 		lda  _sinecosR8,y
 		rts
 _sinecosR8	.char  trunc(127.0 * sin(range(180+45) * rad(360.0/180.0)))
+        ; !notreached!
         }}
     }
 
@@ -77,13 +79,13 @@ _sinecosR8	.char  trunc(127.0 * sin(range(180+45) * rad(360.0/180.0)))
 
     asmsub rnd() clobbers(Y) -> ubyte @A {
         %asm {{
-            jmp  math.randbyte
+            jmp  prog8_math.randbyte
         }}
     }
 
     asmsub rndw() -> uword @AY {
         %asm {{
-            jmp  math.randword
+            jmp  prog8_math.randword
         }}
     }
 
@@ -108,12 +110,12 @@ _sinecosR8	.char  trunc(127.0 * sin(range(180+45) * rad(360.0/180.0)))
     asmsub rndseed(uword seed1 @AY, uword seed2 @R0) clobbers(A,Y) {
         ; -- set new pseudo RNG's seed values. Defaults are: $00c2, $1137
         %asm {{
-            sta  math.randword.x1
-            sty  math.randword.c1
+            sta  prog8_math.randword.x1
+            sty  prog8_math.randword.c1
             lda  cx16.r0L
-            sta  math.randword.a1
+            sta  prog8_math.randword.a1
             lda  cx16.r0H
-            sta  math.randword.b1
+            sta  prog8_math.randword.b1
             rts
         }}
     }
@@ -131,6 +133,7 @@ _sinecosR8	.char  trunc(127.0 * sin(range(180+45) * rad(360.0/180.0)))
             rts
 +           lsr  a
             bne  -
+            ; !notreached!
         }}
     }
 
@@ -172,8 +175,8 @@ _sinecosR8	.char  trunc(127.0 * sin(range(180+45) * rad(360.0/180.0)))
         ;     as soon as a negative word value (or 2) was used in the multiplication, these upper 16 bits are not valid!!
         ;     Suggestion (if you are on the Commander X16): use verafx.muls() to get a hardware accelerated 32 bit signed multplication.
         %asm {{
-            lda  multiply_words.result+2
-            ldy  multiply_words.result+3
+            lda  prog8_math.multiply_words.result+2
+            ldy  prog8_math.multiply_words.result+3
             rts
         }}
     }
@@ -331,7 +334,7 @@ _quadrant_region_to_direction:
   .byte  8, 4,16,20
   .byte  7, 5,17,19
   .byte  6, 6,18,18
-
+        ; !notreached!
     }}
 }
 
@@ -464,7 +467,7 @@ log2_tab
 		.byte $fb,$fb,$fb,$fc,$fc,$fc,$fc,$fc
 		.byte $fd,$fd,$fd,$fd,$fd,$fd,$fe,$fe
 		.byte $fe,$fe,$fe,$ff,$ff,$ff,$ff,$ff
-
+        ; !notreached!
     }}
 }
 
@@ -640,4 +643,23 @@ log2_tab
         cx16.r15 ^= $ffff
         cx16.r14 ^= $ffff
     }
+
+
+    sub lerp(ubyte v0, ubyte v1, ubyte t) -> ubyte {
+        ; Linear interpolation (LERP)
+        ; returns an interpolation between two inputs (v0, v1) for a parameter t in the interval [0, 255]
+        ; guarantees v = v1 when t = 255
+        return v0 + msb(t as uword * (v1 - v0) + 255)
+    }
+
+    sub lerpw(uword v0, uword v1, uword t) -> uword {
+        ; Linear interpolation (LERP) on word values
+        ; returns an interpolation between two inputs (v0, v1) for a parameter t in the interval [0, 65535]
+        ; guarantees v = v1 when t = 65535
+        t *= v1-v0
+        cx16.r0 = math.mul16_last_upper()
+        if t!=0
+            cx16.r0++
+        return v0 + cx16.r0
+    }
 }

compiler/res/prog8lib/neo/neo6502.asm

@@ -0,0 +1,500 @@
+; Convenience macros for Neo6502 applications programming
+; SPDX-License-Identifier: CC0-1.0
+; obtained from https://github.com/paulscottrobson/neo6502-firmware
+
+;-----------------------------------------------------;
+; Neo6502 Kernel jump vectors (see kernel/kernel.asm) ;
+;-----------------------------------------------------;
+
+ChainBasicOnlyProgram = $ffdf
+ChainBasicProgram = $ffe2
+WriteCharacterInline = $ffe5
+LoadExtended = $ffe8
+ReadLine       = $FFEB
+ReadCharacter  = $FFEE
+WriteCharacter = $FFF1
+WaitMessage    = $FFF4
+SendMessage    = $FFF7
+
+
+;--------------------------------------;
+; Neo6502 Kernel API control addresses ;
+;--------------------------------------;
+
+ControlPort = $FF00
+API_COMMAND    = ControlPort + 0 ; function group address
+API_FUNCTION   = ControlPort + 1 ; function address
+API_ERROR      = ControlPort + 2 ; function error codes
+API_STATUS     = ControlPort + 3 ; misc hardware status codes (bit-field)
+API_PARAMETERS = ControlPort + 4 ; function parameters base address (+0-7)
+
+
+;------------------------------------------------;
+; Neo6502 Kernel API control codes (see api.pdf) ;
+;------------------------------------------------;
+
+; Status Information
+API_ERROR_NONE = #$00 ; error code
+API_STATUS_ESC = #$07 ; flag
+
+; System functions (Group 1)
+API_GROUP_SYSTEM     = #$01 ; API function group
+API_FN_TIMER         = #$01 ; API function
+API_FN_KEY_STATUS    = #$02 ; API function
+API_FN_BASIC         = #$03 ; API function
+API_FN_CREDITS       = #$04 ; API function
+API_FN_SERIAL_STATUS = #$05 ; API function
+API_FN_LOCALE        = #$06 ; API function
+API_FN_RESET         = #$07 ; API function
+API_FN_MOS           = #$08 ; API function
+API_FN_SWEET16       = #$09 ; API function
+
+; Console functions (Group 2)
+API_GROUP_CONSOLE     = #$02 ; API function group
+API_FN_READ_CHAR      = #$01 ; API function
+API_FN_CONSOLE_STATUS = #$02 ; API function
+API_FN_READ_LINE      = #$03 ; API function
+API_FN_DEFINE_HOTKEY  = #$04 ; API function
+API_FN_DEFINE_CHAR    = #$05 ; API function
+API_FN_WRITE_CHAR     = #$06 ; API function
+API_FN_SET_CURSOR_POS = #$07 ; API function
+API_FN_LIST_HOTKEYS   = #$08 ; API function
+API_FN_SCREEN_SIZE    = #$09 ; API function
+API_FN_INSERT_LINE    = #$0A ; API function
+API_FN_DELETE_LINE    = #$0B ; API function
+API_FN_CLEAR_SCREEN   = #$0C ; API function
+API_FN_CURSOR_POS     = #$0D ; API function
+API_FN_CLEAR_REGIION  = #$0E ; API function
+API_FN_SET_TEXT_COLOR = #$0F ; API function
+API_FN_CURSOR_INVERSE = #$10 ; API function
+
+; Console results (Group 2 Function 2)
+API_QUEUE_EMPTY = #$FF ; API result (status code)
+
+; File I/O functions (Group 3)
+API_GROUP_FILEIO      = #$03 ; API function group
+API_FN_LIST_DIRECTORY = #$01 ; API function
+API_FN_LOAD_FILENAME  = #$02 ; API function
+API_FN_STORE_FILENAME = #$03 ; API function
+API_FN_FILE_OPEN      = #$04 ; API function
+API_FN_FILE_CLOSE     = #$05 ; API function
+API_FN_FILE_SEEK      = #$06 ; API function
+API_FN_FILE_TELL      = #$07 ; API function
+API_FN_FILE_READ      = #$08 ; API function
+API_FN_FILE_WRITE     = #$09 ; API function
+API_FN_FILE_SIZE      = #$0A ; API function
+API_FN_FILE_SET_SIZE  = #$0B ; API function
+API_FN_FILE_RENAME    = #$0C ; API function
+API_FN_FILE_DELETE    = #$0D ; API function
+API_FN_DIR_CHDIR      = #$0E ; API function
+API_FN_DIR_MKDIR      = #$0F ; API function
+API_FN_FILE_STAT      = #$10 ; API function
+API_FN_DIR_OPEN       = #$11 ; API function
+API_FN_DIR_READ       = #$12 ; API function
+API_FN_DIR_CLOSE      = #$13 ; API function
+API_FN_FILE_COPY      = #$14 ; API function
+API_FN_SET_ATTR       = #$15 ; API function
+API_FN_LIST_FILTERED  = #$20 ; API function
+
+; File I/O parameters (Group 3 Function 2)
+API_FILE_TO_SCREEN = #$FFFF ; API parameter
+
+; Mathematics functions (Group 4)
+API_GROUP_MATH    = #$04 ; API function group
+API_FN_ADD        = #$00 ; API function
+API_FN_SUB        = #$01 ; API function
+API_FN_MUL        = #$02 ; API function
+API_FN_DIV_DEC    = #$03 ; API function
+API_FN_DIV_INT    = #$04 ; API function
+API_FN_MOD        = #$05 ; API function
+API_FN_COMP       = #$06 ; API function
+API_FN_NEG        = #$10 ; API function
+API_FN_FLOOR      = #$11 ; API function
+API_FN_SQRT       = #$12 ; API function
+API_FN_SINE       = #$13 ; API function
+API_FN_COS        = #$14 ; API function
+API_FN_TAN        = #$15 ; API function
+API_FN_ATAN       = #$16 ; API function
+API_FN_EXP        = #$17 ; API function
+API_FN_LOG        = #$18 ; API function
+API_FN_ABS        = #$19 ; API function
+API_FN_SIGN       = #$1A ; API function
+API_FN_RND_DEC    = #$1B ; API function
+API_FN_RND_INT    = #$1C ; API function
+API_FN_INT_TO_DEC = #$20 ; API function
+API_FN_STR_TO_NUM = #$21 ; API function
+API_FN_NUM_TO_STR = #$22 ; API function
+API_FN_SET_DEG_RAD_MODE = #$23 ; API function
+
+; Graphics functions (Group 5)
+API_GROUP_GRAPHICS     = #$05 ; API function group
+API_FN_SET_GFX         = #$01 ; API function
+API_FN_DRAW_LINE       = #$02 ; API function
+API_FN_DRAW_RECT       = #$03 ; API function
+API_FN_DRAW_ELLIPSE    = #$04 ; API function
+API_FN_DRAW_PIXEL      = #$05 ; API function
+API_FN_DRAW_TEXT       = #$06 ; API function
+API_FN_DRAW_IMG        = #$07 ; API function
+API_FN_DRAW_TILEMAP    = #$08 ; API function
+API_FN_SET_PALETTE     = #$20 ; API function
+API_FN_READ_PIXEL      = #$21 ; API function
+API_FN_RESET_PALETTE   = #$22 ; API function
+API_FN_SET_TILEMAP     = #$23 ; API function
+API_FN_READ_SPRITE_PXL = #$24 ; API function
+API_FN_FRAME_COUNT     = #$25 ; API function
+API_FN_GET_PALETTE     = #$26 ; API function
+API_FN_WRITE_PIXEL     = #$27 ; API function
+API_FN_SET_COLOR       = #$40 ; API function
+API_FN_SET_SOLID       = #$41 ; API function
+API_FN_SET_DRAW_SIZE   = #$42 ; API function
+API_FN_SET_FLIP        = #$43 ; API function
+
+; Graphics parameters (Group 5, Function 1 - Group 6, Function 2)
+API_FLIP_HORZ = #$00 ; API parameter (flag)
+API_FLIP_VERT = #$01 ; API parameter (flag)
+
+; Graphics results (Group 5, Functions 33,36)
+API_PIXEL_TRANSPARENT = #$00 ; API result (flag)
+
+; Sprites functions (Group 6)
+API_GROUP_SPRITES       = #$06 ; API function group
+API_FN_SPRITE_RESET     = #$01 ; API function
+API_FN_SPRITE_SET       = #$02 ; API function
+API_FN_SPRITE_HIDE      = #$03 ; API function
+API_FN_SPRITE_COLLISION = #$04 ; API function
+API_FN_SPRITE_POS       = #$05 ; API function
+
+; Sprites parameters (Group 6, Function 2)
+API_SPRITE_TURTLE = #$00 ; API parameter (sprite index)
+API_SPRITE_32BIT  = #$40 ; API parameter (bit-mask)
+API_SPRITE_CLEAR  = #$80 ; API parameter (bit-mask)
+API_ANCHOR_BL     = #$01 ; API parameter (anchor position)
+API_ANCHOR_B      = #$02 ; API parameter (anchor position)
+API_ANCHOR_BR     = #$03 ; API parameter (anchor position)
+API_ANCHOR_L      = #$04 ; API parameter (anchor position)
+API_ANCHOR_C      = #$05 ; API parameter (anchor position)
+API_ANCHOR_R      = #$06 ; API parameter (anchor position)
+API_ANCHOR_TL     = #$07 ; API parameter (anchor position)
+API_ANCHOR_T      = #$08 ; API parameter (anchor position)
+API_ANCHOR_TR     = #$09 ; API parameter (anchor position)
+
+; Sprites results (Group 6, Function 4)
+API_COLLISION_NONE = #$00 ; API result (flag)
+
+; Controller functions (Group 7)
+API_GROUP_CONTROLLER   = #$07 ; API function group
+API_FN_READ_DEFAULT_CONTROLLER = #$01 ; API function
+API_FN_READ_CONTROLLER_COUNT = #$02 ; API function
+API_FN_READ_CONTROLLER = #$03 ; API function
+
+; Controller results (Group 7, Function 1)
+API_CONTROLLER_LEFT  = #$01 ; API result (status bit-mask)
+API_CONTROLLER_RIGHT = #$02 ; API result (status bit-mask)
+API_CONTROLLER_UP    = #$04 ; API result (status bit-mask)
+API_CONTROLLER_DOWN  = #$08 ; API result (status bit-mask)
+API_CONTROLLER_BTNA  = #$10 ; API result (status bit-mask)
+API_CONTROLLER_BTNB  = #$20 ; API result (status bit-mask)
+
+; Sound functions (Group 8)
+API_GROUP_SOUND      = #$08 ; API function group
+API_FN_RESET_SOUND   = #$01 ; API function
+API_FN_RESET_CHANNEL = #$02 ; API function
+API_FN_BEEP          = #$03 ; API function
+API_FN_QUEUE_SOUND   = #$04 ; API function
+API_FN_PLAY_SOUND    = #$05 ; API function
+API_FN_SOUND_STATUS  = #$06 ; API function
+
+; Sound parameters (Group 8, Functions 2,4,5)
+API_SOUND_CH_00 = #$00 ; API parameter (channel index)
+
+; Sound parameters (Group 8, Function 4)
+API_NOTE_REST      = #$0000 ; API parameter (musical rest)
+API_NOTE_C0        = #$0010 ; API parameter (musical note)
+API_NOTE_Cs0       = #$0011 ; API parameter (musical note)
+API_NOTE_Df0       = #$0011 ; API parameter (musical note)
+API_NOTE_D0        = #$0012 ; API parameter (musical note)
+API_NOTE_Ds0       = #$0013 ; API parameter (musical note)
+API_NOTE_Ef0       = #$0013 ; API parameter (musical note)
+API_NOTE_E0        = #$0015 ; API parameter (musical note)
+API_NOTE_F0        = #$0016 ; API parameter (musical note)
+API_NOTE_Fs0       = #$0017 ; API parameter (musical note)
+API_NOTE_Gf0       = #$0017 ; API parameter (musical note)
+API_NOTE_G0        = #$0018 ; API parameter (musical note)
+API_NOTE_Af0       = #$001A ; API parameter (musical note)
+API_NOTE_Gs0       = #$001A ; API parameter (musical note)
+API_NOTE_A0        = #$001C ; API parameter (musical note)
+API_NOTE_As0       = #$001D ; API parameter (musical note)
+API_NOTE_Bf0       = #$001D ; API parameter (musical note)
+API_NOTE_B0        = #$001F ; API parameter (musical note)
+API_NOTE_C1        = #$0021 ; API parameter (musical note)
+API_NOTE_Cs1       = #$0023 ; API parameter (musical note)
+API_NOTE_Df1       = #$0023 ; API parameter (musical note)
+API_NOTE_D1        = #$0025 ; API parameter (musical note)
+API_NOTE_Ds1       = #$0027 ; API parameter (musical note)
+API_NOTE_Ef1       = #$0027 ; API parameter (musical note)
+API_NOTE_E1        = #$0029 ; API parameter (musical note)
+API_NOTE_F1        = #$002C ; API parameter (musical note)
+API_NOTE_Fs1       = #$002E ; API parameter (musical note)
+API_NOTE_Gf1       = #$002E ; API parameter (musical note)
+API_NOTE_G1        = #$0031 ; API parameter (musical note)
+API_NOTE_Af1       = #$0034 ; API parameter (musical note)
+API_NOTE_Gs1       = #$0034 ; API parameter (musical note)
+API_NOTE_A1        = #$0037 ; API parameter (musical note)
+API_NOTE_As1       = #$003A ; API parameter (musical note)
+API_NOTE_Bf1       = #$003A ; API parameter (musical note)
+API_NOTE_B1        = #$003E ; API parameter (musical note)
+API_NOTE_C2        = #$0041 ; API parameter (musical note)
+API_NOTE_Cs2       = #$0045 ; API parameter (musical note)
+API_NOTE_Df2       = #$0045 ; API parameter (musical note)
+API_NOTE_D2        = #$0049 ; API parameter (musical note)
+API_NOTE_Ds2       = #$004E ; API parameter (musical note)
+API_NOTE_Ef2       = #$004E ; API parameter (musical note)
+API_NOTE_E2        = #$0052 ; API parameter (musical note)
+API_NOTE_F2        = #$0057 ; API parameter (musical note)
+API_NOTE_Fs2       = #$005C ; API parameter (musical note)
+API_NOTE_Gf2       = #$005C ; API parameter (musical note)
+API_NOTE_G2        = #$0062 ; API parameter (musical note)
+API_NOTE_Af2       = #$0068 ; API parameter (musical note)
+API_NOTE_Gs2       = #$0068 ; API parameter (musical note)
+API_NOTE_A2        = #$006E ; API parameter (musical note)
+API_NOTE_As2       = #$0075 ; API parameter (musical note)
+API_NOTE_Bf2       = #$0075 ; API parameter (musical note)
+API_NOTE_B2        = #$007B ; API parameter (musical note)
+API_NOTE_C3        = #$0083 ; API parameter (musical note)
+API_NOTE_Cs3       = #$008B ; API parameter (musical note)
+API_NOTE_Df3       = #$008B ; API parameter (musical note)
+API_NOTE_D3        = #$0093 ; API parameter (musical note)
+API_NOTE_Ds3       = #$009C ; API parameter (musical note)
+API_NOTE_Ef3       = #$009C ; API parameter (musical note)
+API_NOTE_E3        = #$00A5 ; API parameter (musical note)
+API_NOTE_F3        = #$00AF ; API parameter (musical note)
+API_NOTE_Fs3       = #$00B9 ; API parameter (musical note)
+API_NOTE_Gf3       = #$00B9 ; API parameter (musical note)
+API_NOTE_G3        = #$00C4 ; API parameter (musical note)
+API_NOTE_Af3       = #$00D0 ; API parameter (musical note)
+API_NOTE_Gs3       = #$00D0 ; API parameter (musical note)
+API_NOTE_A3        = #$00DC ; API parameter (musical note)
+API_NOTE_As3       = #$00E9 ; API parameter (musical note)
+API_NOTE_Bf3       = #$00E9 ; API parameter (musical note)
+API_NOTE_B3        = #$00F7 ; API parameter (musical note)
+API_NOTE_C4        = #$0106 ; API parameter (musical note)
+API_NOTE_Cs4       = #$0115 ; API parameter (musical note)
+API_NOTE_Df4       = #$0115 ; API parameter (musical note)
+API_NOTE_D4        = #$0126 ; API parameter (musical note)
+API_NOTE_Ds4       = #$0137 ; API parameter (musical note)
+API_NOTE_Ef4       = #$0137 ; API parameter (musical note)
+API_NOTE_E4        = #$014A ; API parameter (musical note)
+API_NOTE_F4        = #$015D ; API parameter (musical note)
+API_NOTE_Fs4       = #$0172 ; API parameter (musical note)
+API_NOTE_Gf4       = #$0172 ; API parameter (musical note)
+API_NOTE_G4        = #$0188 ; API parameter (musical note)
+API_NOTE_Af4       = #$019F ; API parameter (musical note)
+API_NOTE_Gs4       = #$019F ; API parameter (musical note)
+API_NOTE_A4        = #$01B8 ; API parameter (musical note)
+API_NOTE_As4       = #$01D2 ; API parameter (musical note)
+API_NOTE_Bf4       = #$01D2 ; API parameter (musical note)
+API_NOTE_B4        = #$01EE ; API parameter (musical note)
+API_NOTE_C5        = #$020B ; API parameter (musical note)
+API_NOTE_Cs5       = #$022A ; API parameter (musical note)
+API_NOTE_Df5       = #$022A ; API parameter (musical note)
+API_NOTE_D5        = #$024B ; API parameter (musical note)
+API_NOTE_Ds5       = #$026E ; API parameter (musical note)
+API_NOTE_Ef5       = #$026E ; API parameter (musical note)
+API_NOTE_E5        = #$0293 ; API parameter (musical note)
+API_NOTE_F5        = #$02BA ; API parameter (musical note)
+API_NOTE_Fs5       = #$02E4 ; API parameter (musical note)
+API_NOTE_Gf5       = #$02E4 ; API parameter (musical note)
+API_NOTE_G5        = #$0310 ; API parameter (musical note)
+API_NOTE_Af5       = #$033F ; API parameter (musical note)
+API_NOTE_Gs5       = #$033F ; API parameter (musical note)
+API_NOTE_A5        = #$0370 ; API parameter (musical note)
+API_NOTE_As5       = #$03A4 ; API parameter (musical note)
+API_NOTE_Bf5       = #$03A4 ; API parameter (musical note)
+API_NOTE_B5        = #$03DC ; API parameter (musical note)
+API_NOTE_C6        = #$0417 ; API parameter (musical note)
+API_NOTE_Cs6       = #$0455 ; API parameter (musical note)
+API_NOTE_Df6       = #$0455 ; API parameter (musical note)
+API_NOTE_D6        = #$0497 ; API parameter (musical note)
+API_NOTE_Ds6       = #$04DD ; API parameter (musical note)
+API_NOTE_Ef6       = #$04DD ; API parameter (musical note)
+API_NOTE_E6        = #$0527 ; API parameter (musical note)
+API_NOTE_F6        = #$0575 ; API parameter (musical note)
+API_NOTE_Fs6       = #$05C8 ; API parameter (musical note)
+API_NOTE_Gf6       = #$05C8 ; API parameter (musical note)
+API_NOTE_G6        = #$0620 ; API parameter (musical note)
+API_NOTE_Af6       = #$067D ; API parameter (musical note)
+API_NOTE_Gs6       = #$067D ; API parameter (musical note)
+API_NOTE_A6        = #$06E0 ; API parameter (musical note)
+API_NOTE_As6       = #$0749 ; API parameter (musical note)
+API_NOTE_Bf6       = #$0749 ; API parameter (musical note)
+API_NOTE_B6        = #$07B8 ; API parameter (musical note)
+API_NOTE_C7        = #$082D ; API parameter (musical note)
+API_NOTE_Cs7       = #$08A9 ; API parameter (musical note)
+API_NOTE_Df7       = #$08A9 ; API parameter (musical note)
+API_NOTE_D7        = #$092D ; API parameter (musical note)
+API_NOTE_Ds7       = #$09B9 ; API parameter (musical note)
+API_NOTE_Ef7       = #$09B9 ; API parameter (musical note)
+API_NOTE_E7        = #$0A4D ; API parameter (musical note)
+API_NOTE_F7        = #$0AEA ; API parameter (musical note)
+API_NOTE_Fs7       = #$0B90 ; API parameter (musical note)
+API_NOTE_Gf7       = #$0B90 ; API parameter (musical note)
+API_NOTE_G7        = #$0C40 ; API parameter (musical note)
+API_NOTE_Af7       = #$0CFA ; API parameter (musical note)
+API_NOTE_Gs7       = #$0CFA ; API parameter (musical note)
+API_NOTE_A7        = #$0DC0 ; API parameter (musical note)
+API_NOTE_As7       = #$0E91 ; API parameter (musical note)
+API_NOTE_Bf7       = #$0E91 ; API parameter (musical note)
+API_NOTE_B7        = #$0F6F ; API parameter (musical note)
+API_NOTE_C8        = #$105A ; API parameter (musical note)
+API_NOTE_Cs8       = #$1153 ; API parameter (musical note)
+API_NOTE_Df8       = #$1153 ; API parameter (musical note)
+API_NOTE_D8        = #$125B ; API parameter (musical note)
+API_NOTE_Ds8       = #$1372 ; API parameter (musical note)
+API_NOTE_Ef8       = #$1372 ; API parameter (musical note)
+API_NOTE_E8        = #$149A ; API parameter (musical note)
+API_NOTE_F8        = #$15D4 ; API parameter (musical note)
+API_NOTE_Fs8       = #$1720 ; API parameter (musical note)
+API_NOTE_Gf8       = #$1720 ; API parameter (musical note)
+API_NOTE_G8        = #$1880 ; API parameter (musical note)
+API_NOTE_Af8       = #$19F5 ; API parameter (musical note)
+API_NOTE_Gs8       = #$19F5 ; API parameter (musical note)
+API_NOTE_A8        = #$1B80 ; API parameter (musical note)
+API_NOTE_As8       = #$1D23 ; API parameter (musical note)
+API_NOTE_Bf8       = #$1D23 ; API parameter (musical note)
+API_NOTE_B8        = #$1EDE ; API parameter (musical note)
+API_NOTE_C9        = #$20B4 ; API parameter (musical note)
+API_NOTE_Cs9       = #$22A6 ; API parameter (musical note)
+API_NOTE_Df9       = #$22A6 ; API parameter (musical note)
+API_NOTE_D9        = #$24B5 ; API parameter (musical note)
+API_NOTE_Ds9       = #$26E4 ; API parameter (musical note)
+API_NOTE_Ef9       = #$26E4 ; API parameter (musical note)
+API_NOTE_E9        = #$2934 ; API parameter (musical note)
+API_NOTE_F9        = #$2BA7 ; API parameter (musical note)
+API_NOTE_Fs9       = #$2E40 ; API parameter (musical note)
+API_NOTE_Gf9       = #$2E40 ; API parameter (musical note)
+API_NOTE_G9        = #$3100 ; API parameter (musical note)
+API_NOTE_Af9       = #$33EA ; API parameter (musical note)
+API_NOTE_Gs9       = #$33EA ; API parameter (musical note)
+API_NOTE_A9        = #$3700 ; API parameter (musical note)
+API_NOTE_As9       = #$3A45 ; API parameter (musical note)
+API_NOTE_Bf9       = #$3A45 ; API parameter (musical note)
+API_NOTE_B9        = #$3DBC ; API parameter (musical note)
+API_NOTE_C10       = #$4168 ; API parameter (musical note)
+API_NOTE_Cs10      = #$454C ; API parameter (musical note)
+API_NOTE_Df10      = #$454C ; API parameter (musical note)
+API_NOTE_D10       = #$496B ; API parameter (musical note)
+API_NOTE_Ds10      = #$4DC8 ; API parameter (musical note)
+API_NOTE_Ef10      = #$4DC8 ; API parameter (musical note)
+API_TEMPO_60       = #$0064 ; API parameter (musical note duration, 60BPM)
+API_TEMPO_80       = #$004B ; API parameter (musical note duration, 80BPM)
+API_TEMPO_90       = #$0042 ; API parameter (musical note duration, 90BPM)
+API_TEMPO_120      = #$0032 ; API parameter (musical note duration, 120BPM)
+API_SLIDE_NONE     = #$0000 ; API parameter (slide value)
+API_SLIDE_SLOW     = #$0004 ; API parameter (slide range)
+API_SLIDE_MED      = #$0008 ; API parameter (slide range)
+API_SLIDE_FAST     = #$0010 ; API parameter (slide range)
+API_SOUND_SRC_BEEP = #$00   ; API parameter (sound generator)
+
+; Sound parameters (Group 8, Function 5)
+API_SFX_POSITIVE     = #$00 ; API parameter (sound effect)
+API_SFX_NEGATIVE     = #$01 ; API parameter (sound effect)
+API_SFX_ERROR        = #$02 ; API parameter (sound effect)
+API_SFX_CONFIRM      = #$03 ; API parameter (sound effect)
+API_SFX_REJECT       = #$04 ; API parameter (sound effect)
+API_SFX_SWEEP        = #$05 ; API parameter (sound effect)
+API_SFX_COIN         = #$06 ; API parameter (sound effect)
+API_SFX_LASER_LONG   = #$07 ; API parameter (sound effect)
+API_SFX_POWERUP      = #$08 ; API parameter (sound effect)
+API_SFX_VICTORY      = #$09 ; API parameter (sound effect)
+API_SFX_DEFEAT       = #$0A ; API parameter (sound effect)
+API_SFX_FANFARE      = #$0B ; API parameter (sound effect)
+API_SFX_ALARM1       = #$0C ; API parameter (sound effect)
+API_SFX_ALARM2       = #$0D ; API parameter (sound effect)
+API_SFX_ALARM3       = #$0E ; API parameter (sound effect)
+API_SFX_RING1        = #$0F ; API parameter (sound effect)
+API_SFX_RING2        = #$10 ; API parameter (sound effect)
+API_SFX_RING3        = #$11 ; API parameter (sound effect)
+API_SFX_DANGER       = #$12 ; API parameter (sound effect)
+API_SFX_EXPL_LONG    = #$13 ; API parameter (sound effect)
+API_SFX_EXPL_MEDIUM  = #$14 ; API parameter (sound effect)
+API_SFX_EXPL_SHORT   = #$15 ; API parameter (sound effect)
+API_SFX_LASER_MEDIUM = #$16 ; API parameter (sound effect)
+API_SFX_LASER_SHORT  = #$17 ; API parameter (sound effect)
+
+; Turtle Graphics functions (Group 9)
+API_GROUP_TURTLE   = #$09 ; API function group
+API_FN_TURTLE_INIT = #$01 ; API function
+API_FN_TURTLE_TURN = #$02 ; API function
+API_FN_TURTLE_MOVE = #$03 ; API function
+API_FN_TURTLE_HIDE = #$04 ; API function
+API_FN_TURTLE_HOME = #$05 ; API function
+
+; Turtle Graphics parameters (Group 9, Function 2)
+API_TURTLE_LEFT  = #$010E ; API parameter (turn -90 degrees)
+API_TURTLE_RIGHT = #$005A ; API parameter (turn +90 degrees)
+API_TURTLE_FLIP  = #$00B4 ; API parameter (turn 180 degrees)
+
+; Turtle Graphics parameters (Group 9, Function 3)
+API_PEN_UP   = #$00 ; API parameter (turtle tracks on)
+API_PEN_DOWN = #$01 ; API parameter (turtle tracks off)
+
+; UExt functions (Group 10)
+API_GROUP_UEXT      = #$09 ; API function group
+API_FN_UEXT_INIT    = #$01 ; API function
+API_FN_GPIO_WRITE   = #$02 ; API function
+API_FN_GPIO_READ    = #$03 ; API function
+API_FN_SET_PORT_DIR = #$04 ; API function
+API_FN_I2C_WRITE    = #$05 ; API function
+API_FN_I2C_READ     = #$06 ; API function
+API_FN_ANALOG_READ  = #$07 ; API function
+API_FN_I2C_STATUS   = #$08 ; API function
+API_FN_I2C_READ_BLOCK   = #$09 ; API function
+API_FN_I2C_WRITE_BLCOK  = #$0a ; API function
+API_FN_SPI_READ_BLOCK   = #$0b ; API function
+API_FN_SPI_WRITE_BLOCK  = #$0c ; API function
+API_FN_UART_READ_BLOCK  = #$0d ; API function
+API_FN_UART_WRITE_BLOPCK= #$0e ; API function
+API_FN_UART_SET_SPEED   = #$0f ; API function
+API_FN_UART_WRITE   = #$10 ; API function
+API_FN_UART_READ    = #$11 ; API function
+API_FN_UART_CHECK_BYTE_AVAILABLE = #$12 ; API function
+
+; TODO group 11 MOUSE
+; TODO group 12 BLITTER
+; TODO group 13 EDITOR
+
+
+;--------;
+; colors ;
+;--------;
+
+COLOR_BLACK       = #$80
+COLOR_RED         = #$81
+COLOR_GREEN       = #$82
+COLOR_YELLOW      = #$83
+COLOR_BLUE        = #$84
+COLOR_MAGENTA     = #$85
+COLOR_CYAN        = #$86
+COLOR_WHITE       = #$87
+COLOR_ALT_BLACK   = #$88
+COLOR_DARK_GREY   = #$89
+COLOR_DARK_GREEN  = #$8A
+COLOR_ORANGE      = #$8B
+COLOR_DARK_ORANGE = #$8C
+COLOR_BROWN       = #$8D
+COLOR_PINK        = #$8E
+COLOR_LIGHT_GREY  = #$8F
+
+
+;---------------------------------;
+; limited-case convenience macros ;
+;---------------------------------;
+
+DoSendMessage .macro
+  jsr  neo.SendMessage
+  .endm
+
+DoWaitMessage .macro
+  jsr  neo.WaitMessage
+  .endm
+

compiler/res/prog8lib/neo/syslib.p8

@@ -0,0 +1,488 @@
+; Prog8 definitions for the Neo6502
+
+%option no_symbol_prefixing, ignore_unused
+
+neo {
+    &uword  NMI_VEC         = $FFFA     ; 6502 nmi vector, determined by the kernal if banked in
+    &uword  RESET_VEC       = $FFFC     ; 6502 reset vector, determined by the kernal if banked in
+    &uword  IRQ_VEC         = $FFFE     ; 6502 interrupt vector, determined by the kernal if banked in
+
+    %asminclude "library:neo/neo6502.asm"
+}
+
+sys {
+    ; ------- lowlevel system routines --------
+
+    const ubyte target = 7         ;  compilation target specifier.  255=virtual, 128=C128, 64=C64, 32=PET, 16=CommanderX16, 8=atari800XL, 7=Neo6502
+
+    const ubyte sizeof_bool = 1
+    const ubyte sizeof_byte = 1
+    const ubyte sizeof_ubyte = 1
+    const ubyte sizeof_word = 2
+    const ubyte sizeof_uword = 2
+    const ubyte sizeof_float = 0    ; undefined, no floats supported
+
+
+    asmsub  reset_system()  {
+        ; Soft-reset the system back to initial power-on status
+        ; TODO
+        %asm {{
+            sei
+            jmp  (neo.RESET_VEC)
+        }}
+    }
+
+    sub wait(uword jiffies) {
+        ; --- wait approximately the given number of jiffies (1/60th seconds)
+        ;     TODO
+    }
+
+    asmsub waitvsync() clobbers(A) {
+        ; --- busy wait till the next vsync has occurred (approximately), without depending on custom irq handling.
+        ;     TODO
+        %asm {{
+            nop
+            rts
+        }}
+    }
+
+    asmsub internal_stringcopy(uword source @R0, uword target @AY) clobbers (A,Y) {
+        ; Called when the compiler wants to assign a string value to another string.
+        %asm {{
+		sta  P8ZP_SCRATCH_W1
+		sty  P8ZP_SCRATCH_W1+1
+		lda  cx16.r0
+		ldy  cx16.r0+1
+		jmp  prog8_lib.strcpy
+        }}
+    }
+
+    asmsub memcopy(uword source @R0, uword target @R1, uword count @AY) clobbers(A,X,Y) {
+        ; note: only works for NON-OVERLAPPING memory regions!
+        ; note: can't be inlined because is called from asm as well
+        %asm {{
+            ldx  cx16.r0
+            stx  P8ZP_SCRATCH_W1        ; source in ZP
+            ldx  cx16.r0+1
+            stx  P8ZP_SCRATCH_W1+1
+            ldx  cx16.r1
+            stx  P8ZP_SCRATCH_W2        ; target in ZP
+            ldx  cx16.r1+1
+            stx  P8ZP_SCRATCH_W2+1
+            cpy  #0
+            bne  _longcopy
+
+            ; copy <= 255 bytes
+            tay
+            bne  _copyshort
+            rts     ; nothing to copy
+
+_copyshort
+            dey
+            beq  +
+-           lda  (P8ZP_SCRATCH_W1),y
+            sta  (P8ZP_SCRATCH_W2),y
+            dey
+            bne  -
++           lda  (P8ZP_SCRATCH_W1),y
+            sta  (P8ZP_SCRATCH_W2),y
+            rts
+
+_longcopy
+            sta  P8ZP_SCRATCH_B1        ; lsb(count) = remainder in last page
+            tya
+            tax                         ; x = num pages (1+)
+            ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            sta  (P8ZP_SCRATCH_W2),y
+            iny
+            bne  -
+            inc  P8ZP_SCRATCH_W1+1
+            inc  P8ZP_SCRATCH_W2+1
+            dex
+            bne  -
+            ldy  P8ZP_SCRATCH_B1
+            bne  _copyshort
+            rts
+        }}
+    }
+
+    asmsub memset(uword mem @R0, uword numbytes @R1, ubyte value @A) clobbers(A,X,Y) {
+        %asm {{
+            ldy  cx16.r0
+            sty  P8ZP_SCRATCH_W1
+            ldy  cx16.r0+1
+            sty  P8ZP_SCRATCH_W1+1
+            ldx  cx16.r1
+            ldy  cx16.r1+1
+            jmp  prog8_lib.memset
+        }}
+    }
+
+    asmsub memsetw(uword mem @R0, uword numwords @R1, uword value @AY) clobbers(A,X,Y) {
+        %asm {{
+            ldx  cx16.r0
+            stx  P8ZP_SCRATCH_W1
+            ldx  cx16.r0+1
+            stx  P8ZP_SCRATCH_W1+1
+            ldx  cx16.r1
+            stx  P8ZP_SCRATCH_W2
+            ldx  cx16.r1+1
+            stx  P8ZP_SCRATCH_W2+1
+            jmp  prog8_lib.memsetw
+        }}
+    }
+
+    inline asmsub read_flags() -> ubyte @A {
+        %asm {{
+        php
+        pla
+        }}
+    }
+
+    inline asmsub clear_carry() {
+        %asm {{
+        clc
+        }}
+    }
+
+    inline asmsub set_carry() {
+        %asm {{
+        sec
+        }}
+    }
+
+    inline asmsub clear_irqd() {
+        %asm {{
+        cli
+        }}
+    }
+
+    inline asmsub set_irqd() {
+        %asm {{
+        sei
+        }}
+    }
+
+    inline asmsub irqsafe_set_irqd() {
+        %asm {{
+        php
+        sei
+        }}
+    }
+
+    inline asmsub irqsafe_clear_irqd() {
+        %asm {{
+        plp
+        }}
+    }
+
+    sub disable_caseswitch() {
+        ; no-op
+    }
+
+    sub enable_caseswitch() {
+        ; no-op
+    }
+
+    asmsub save_prog8_internals() {
+        %asm {{
+            lda  P8ZP_SCRATCH_B1
+            sta  save_SCRATCH_ZPB1
+            lda  P8ZP_SCRATCH_REG
+            sta  save_SCRATCH_ZPREG
+            lda  P8ZP_SCRATCH_W1
+            sta  save_SCRATCH_ZPWORD1
+            lda  P8ZP_SCRATCH_W1+1
+            sta  save_SCRATCH_ZPWORD1+1
+            lda  P8ZP_SCRATCH_W2
+            sta  save_SCRATCH_ZPWORD2
+            lda  P8ZP_SCRATCH_W2+1
+            sta  save_SCRATCH_ZPWORD2+1
+            rts
+save_SCRATCH_ZPB1	.byte  0
+save_SCRATCH_ZPREG	.byte  0
+save_SCRATCH_ZPWORD1	.word  0
+save_SCRATCH_ZPWORD2	.word  0
+            ; !notreached!
+        }}
+    }
+
+    asmsub restore_prog8_internals() {
+        %asm {{
+            lda  save_prog8_internals.save_SCRATCH_ZPB1
+            sta  P8ZP_SCRATCH_B1
+            lda  save_prog8_internals.save_SCRATCH_ZPREG
+            sta  P8ZP_SCRATCH_REG
+            lda  save_prog8_internals.save_SCRATCH_ZPWORD1
+            sta  P8ZP_SCRATCH_W1
+            lda  save_prog8_internals.save_SCRATCH_ZPWORD1+1
+            sta  P8ZP_SCRATCH_W1+1
+            lda  save_prog8_internals.save_SCRATCH_ZPWORD2
+            sta  P8ZP_SCRATCH_W2
+            lda  save_prog8_internals.save_SCRATCH_ZPWORD2+1
+            sta  P8ZP_SCRATCH_W2+1
+            rts
+        }}
+    }
+
+    asmsub exit(ubyte returnvalue @A) {
+        ; -- immediately exit the program with a return code in the A register
+        %asm {{
+            sta  cleanup_at_exit._exitcode
+            ldx  prog8_lib.orig_stackpointer
+            txs
+            jmp  cleanup_at_exit
+        }}
+    }
+
+    asmsub exit2(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y) {
+        ; -- immediately exit the program with result values in the A, X and Y registers.
+        %asm {{
+            sta  cleanup_at_exit._exitcode
+            stx  cleanup_at_exit._exitcodeX
+            sty  cleanup_at_exit._exitcodeY
+            ldx  prog8_lib.orig_stackpointer
+            txs
+            jmp  cleanup_at_exit
+        }}
+    }
+
+    asmsub exit3(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y, bool carry @Pc) {
+        ; -- immediately exit the program with result values in the A, X and Y registers, and the Carry flag in the status register.
+        %asm {{
+            sta  cleanup_at_exit._exitcode
+            lda  #0
+            rol  a
+            sta  cleanup_at_exit._exitcodeCarry
+            stx  cleanup_at_exit._exitcodeX
+            sty  cleanup_at_exit._exitcodeY
+            ldx  prog8_lib.orig_stackpointer
+            txs
+            jmp  cleanup_at_exit
+        }}
+    }
+
+    inline asmsub progend() -> uword @AY {
+        %asm {{
+            lda  #<prog8_program_end
+            ldy  #>prog8_program_end
+        }}
+    }
+
+    inline asmsub progstart() -> uword @AY {
+        %asm {{
+            lda  #<prog8_program_start
+            ldy  #>prog8_program_start
+        }}
+    }
+
+    inline asmsub push(ubyte value @A) {
+        %asm {{
+            pha
+        }}
+    }
+
+    inline asmsub pushw(uword value @AY) {
+        %asm {{
+            pha
+            tya
+            pha
+        }}
+    }
+
+    inline asmsub pop() -> ubyte @A {
+        %asm {{
+            pla
+        }}
+    }
+
+    inline asmsub popw() -> uword @AY {
+        %asm {{
+            pla
+            tay
+            pla
+        }}
+    }
+
+}
+
+cx16 {
+    ; the sixteen virtual 16-bit registers that the CX16 has defined in the zeropage
+; the sixteen virtual 16-bit registers in both normal unsigned mode and signed mode (s)
+    &uword r0  = $0002
+    &uword r1  = $0004
+    &uword r2  = $0006
+    &uword r3  = $0008
+    &uword r4  = $000a
+    &uword r5  = $000c
+    &uword r6  = $000e
+    &uword r7  = $0010
+    &uword r8  = $0012
+    &uword r9  = $0014
+    &uword r10 = $0016
+    &uword r11 = $0018
+    &uword r12 = $001a
+    &uword r13 = $001c
+    &uword r14 = $001e
+    &uword r15 = $0020
+
+    &word r0s  = $0002
+    &word r1s  = $0004
+    &word r2s  = $0006
+    &word r3s  = $0008
+    &word r4s  = $000a
+    &word r5s  = $000c
+    &word r6s  = $000e
+    &word r7s  = $0010
+    &word r8s  = $0012
+    &word r9s  = $0014
+    &word r10s = $0016
+    &word r11s = $0018
+    &word r12s = $001a
+    &word r13s = $001c
+    &word r14s = $001e
+    &word r15s = $0020
+
+    &ubyte r0L  = $0002
+    &ubyte r1L  = $0004
+    &ubyte r2L  = $0006
+    &ubyte r3L  = $0008
+    &ubyte r4L  = $000a
+    &ubyte r5L  = $000c
+    &ubyte r6L  = $000e
+    &ubyte r7L  = $0010
+    &ubyte r8L  = $0012
+    &ubyte r9L  = $0014
+    &ubyte r10L = $0016
+    &ubyte r11L = $0018
+    &ubyte r12L = $001a
+    &ubyte r13L = $001c
+    &ubyte r14L = $001e
+    &ubyte r15L = $0020
+
+    &ubyte r0H  = $0003
+    &ubyte r1H  = $0005
+    &ubyte r2H  = $0007
+    &ubyte r3H  = $0009
+    &ubyte r4H  = $000b
+    &ubyte r5H  = $000d
+    &ubyte r6H  = $000f
+    &ubyte r7H  = $0011
+    &ubyte r8H  = $0013
+    &ubyte r9H  = $0015
+    &ubyte r10H = $0017
+    &ubyte r11H = $0019
+    &ubyte r12H = $001b
+    &ubyte r13H = $001d
+    &ubyte r14H = $001f
+    &ubyte r15H = $0021
+
+    &byte r0sL  = $0002
+    &byte r1sL  = $0004
+    &byte r2sL  = $0006
+    &byte r3sL  = $0008
+    &byte r4sL  = $000a
+    &byte r5sL  = $000c
+    &byte r6sL  = $000e
+    &byte r7sL  = $0010
+    &byte r8sL  = $0012
+    &byte r9sL  = $0014
+    &byte r10sL = $0016
+    &byte r11sL = $0018
+    &byte r12sL = $001a
+    &byte r13sL = $001c
+    &byte r14sL = $001e
+    &byte r15sL = $0020
+
+    &byte r0sH  = $0003
+    &byte r1sH  = $0005
+    &byte r2sH  = $0007
+    &byte r3sH  = $0009
+    &byte r4sH  = $000b
+    &byte r5sH  = $000d
+    &byte r6sH  = $000f
+    &byte r7sH  = $0011
+    &byte r8sH  = $0013
+    &byte r9sH  = $0015
+    &byte r10sH = $0017
+    &byte r11sH = $0019
+    &byte r12sH = $001b
+    &byte r13sH = $001d
+    &byte r14sH = $001f
+    &byte r15sH = $0021
+
+    asmsub save_virtual_registers() clobbers(A,Y) {
+        %asm {{
+            ldy  #31
+    -       lda  cx16.r0,y
+            sta  _cx16_vreg_storage,y
+            dey
+            bpl  -
+            rts
+
+    _cx16_vreg_storage
+            .word 0,0,0,0,0,0,0,0
+            .word 0,0,0,0,0,0,0,0
+            ; !notreached!
+        }}
+    }
+
+    asmsub restore_virtual_registers() clobbers(A,Y) {
+        %asm {{
+            ldy  #31
+    -       lda  save_virtual_registers._cx16_vreg_storage,y
+            sta  cx16.r0,y
+            dey
+            bpl  -
+            rts
+        }}
+    }
+
+    sub cpu_is_65816() -> bool {
+        ; Returns true when you have a 65816 cpu, false when it's a 6502.
+        return false
+    }
+
+}
+
+p8_sys_startup {
+    ; program startup and shutdown machinery. Needs to reside in normal system ram.
+
+    asmsub  init_system()  {
+        ; Initializes the machine to a sane starting state.
+        ; Called automatically by the loader program logic.
+        %asm {{
+            sei
+            cld
+            clc
+            ; TODO reset screen mode etc etc?
+            clv
+            ; TODO what about IRQ handler?  cli
+            rts
+        }}
+    }
+
+    asmsub  init_system_phase2()  {
+        %asm {{
+            rts     ; no phase 2 steps on the Neo6502
+        }}
+    }
+
+    asmsub  cleanup_at_exit() {
+        ; executed when the main subroutine does rts
+        %asm {{
+_exitcodeCarry = *+1
+            lda  #0
+            lsr  a
+_exitcode = *+1
+            lda  #0        ; exit code possibly modified in exit()
+_exitcodeX = *+1
+            ldx  #0
+_exitcodeY = *+1
+            ldy  #0
+            rts
+        }}
+    }
+
+}

compiler/res/prog8lib/pet32/syslib.p8

@@ -24,15 +24,15 @@ cbm {
         const  uword  Screen    = $8000     ; to have this as an array[40*25] the compiler would have to support array size > 255
 
 
-romsub $FFC6 = CHKIN(ubyte logical @ X) clobbers(A,X) -> bool @Pc    ; define an input channel
-romsub $FFC9 = CHKOUT(ubyte logical @ X) clobbers(A,X)          ; define an output channel
-romsub $FFCC = CLRCHN() clobbers(A,X)                           ; restore default devices
-romsub $FFCF = CHRIN() clobbers(X, Y) -> ubyte @ A              ; input a character (for keyboard, read a whole line from the screen) A=byte read.
-romsub $FFD2 = CHROUT(ubyte character @ A)                           ; output a character
-romsub $FFE1 = STOP() clobbers(X) -> bool @ Pz, ubyte @ A       ; check the STOP key (and some others in A)     also see STOP2
-romsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A     ; get a character       also see GETIN2
-romsub $FFE7 = CLALL() clobbers(A,X)                            ; close all files
-romsub $FFEA = UDTIM() clobbers(A,X)                            ; update the software clock
+extsub $FFC6 = CHKIN(ubyte logical @ X) clobbers(A,X) -> bool @Pc    ; define an input channel
+extsub $FFC9 = CHKOUT(ubyte logical @ X) clobbers(A,X)          ; define an output channel
+extsub $FFCC = CLRCHN() clobbers(A,X)                           ; restore default devices
+extsub $FFCF = CHRIN() clobbers(X, Y) -> ubyte @ A              ; input a character (for keyboard, read a whole line from the screen) A=byte read.
+extsub $FFD2 = CHROUT(ubyte character @ A)                           ; output a character
+extsub $FFE1 = STOP() clobbers(X) -> bool @ Pz, ubyte @ A       ; check the STOP key (and some others in A)     also see STOP2
+extsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A     ; get a character       also see GETIN2
+extsub $FFE7 = CLALL() clobbers(A,X)                            ; close all files
+extsub $FFEA = UDTIM() clobbers(A,X)                            ; update the software clock
 
 
 inline asmsub STOP2() clobbers(X,A) -> bool @Pz  {
@@ -96,7 +96,7 @@ asmsub kbdbuf_clear() {
 sys {
     ; ------- lowlevel system routines --------
 
-    const ubyte target = 32         ;  compilation target specifier.  64 = C64, 128 = C128,  16 = CommanderX16, 32=PET
+    const ubyte target = 32         ;  compilation target specifier.  255=virtual, 128=C128, 64=C64, 32=PET, 16=CommanderX16, 8=atari800XL, 7=Neo6502
 
     const ubyte sizeof_bool = 1
     const ubyte sizeof_byte = 1
@@ -106,46 +106,6 @@ sys {
     const ubyte sizeof_float = 0    ; undefined, no floats supported
 
 
-asmsub  init_system()  {
-    ; Initializes the machine to a sane starting state.
-    ; Called automatically by the loader program logic.
-    ; Uppercase charset is activated.
-    %asm {{
-        sei
-        lda  #142
-        jsr  cbm.CHROUT     ; uppercase
-        lda  #147
-        jsr  cbm.CHROUT     ; clear screen
-        cli
-        rts
-    }}
-}
-
-asmsub  init_system_phase2()  {
-    %asm {{
-        cld
-        clc
-        clv
-        rts
-    }}
-}
-
-asmsub  cleanup_at_exit() {
-    ; executed when the main subroutine does rts
-    %asm {{
-_exitcodeCarry = *+1
-        lda  #0
-        lsr  a
-_exitcode = *+1
-        lda  #0        ; exit code possibly modified in exit()
-_exitcodeX = *+1
-        ldx  #0
-_exitcodeY = *+1
-        ldy  #0
-        rts
-    }}
-}
-
     asmsub  reset_system()  {
         ; Soft-reset the system back to initial power-on Basic prompt.
         %asm {{
@@ -279,6 +239,65 @@ _longcopy
         }}
     }
 
+    asmsub memcmp(uword address1 @R0, uword address2 @R1, uword size @AY) -> byte @A {
+        ; Compares two blocks of memory
+        ; Returns -1 (255), 0 or 1, meaning: block 1 sorts before, equal or after block 2.
+        %asm {{
+            sta  P8ZP_SCRATCH_REG   ; lsb(size)
+            sty  P8ZP_SCRATCH_B1    ; msb(size)
+            lda  cx16.r0
+            ldy  cx16.r0+1
+            sta  P8ZP_SCRATCH_W1
+            sty  P8ZP_SCRATCH_W1+1
+            lda  cx16.r1
+            ldy  cx16.r1+1
+            sta  P8ZP_SCRATCH_W2
+            sty  P8ZP_SCRATCH_W2+1
+
+            ldx  P8ZP_SCRATCH_B1
+            beq  _no_msb_size
+
+_loop_msb_size
+            ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            cmp  (P8ZP_SCRATCH_W2),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            bne  -
+            inc  P8ZP_SCRATCH_W1+1
+            inc  P8ZP_SCRATCH_W2+1
+            dec  P8ZP_SCRATCH_B1        ; msb(size) -= 1
+            dex
+            bne  _loop_msb_size
+
+_no_msb_size
+            lda  P8ZP_SCRATCH_REG       ; lsb(size)
+            bne  +
+            rts
+
++           ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            cmp  (P8ZP_SCRATCH_W2),y
+            bcs  +
+            lda  #-1
+            rts
++           beq  +
+            lda  #1
+            rts
++           iny
+            cpy  P8ZP_SCRATCH_REG       ; lsb(size)
+            bne  -
+
+            lda #0
+            rts
+        }}
+    }
+
     inline asmsub read_flags() -> ubyte @A {
         %asm {{
         php
@@ -350,6 +369,7 @@ save_SCRATCH_ZPB1	.byte  0
 save_SCRATCH_ZPREG	.byte  0
 save_SCRATCH_ZPWORD1	.word  0
 save_SCRATCH_ZPWORD2	.word  0
+            ; !notreached!
         }}
     }
 
@@ -374,37 +394,37 @@ save_SCRATCH_ZPWORD2	.word  0
     asmsub exit(ubyte returnvalue @A) {
         ; -- immediately exit the program with a return code in the A register
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit2(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y) {
         ; -- immediately exit the program with result values in the A, X and Y registers.
         %asm {{
-            sta  cleanup_at_exit._exitcode
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
     asmsub exit3(ubyte resulta @A, ubyte resultx @X, ubyte resulty @Y, bool carry @Pc) {
         ; -- immediately exit the program with result values in the A, X and Y registers, and the Carry flag in the status register.
         %asm {{
-            sta  cleanup_at_exit._exitcode
+            sta  p8_sys_startup.cleanup_at_exit._exitcode
             lda  #0
             rol  a
-            sta  cleanup_at_exit._exitcodeCarry
-            stx  cleanup_at_exit._exitcodeX
-            sty  cleanup_at_exit._exitcodeY
+            sta  p8_sys_startup.cleanup_at_exit._exitcodeCarry
+            stx  p8_sys_startup.cleanup_at_exit._exitcodeX
+            sty  p8_sys_startup.cleanup_at_exit._exitcodeY
             ldx  prog8_lib.orig_stackpointer
             txs
-            jmp  cleanup_at_exit
+            jmp  p8_sys_startup.cleanup_at_exit
         }}
     }
 
@@ -415,6 +435,13 @@ save_SCRATCH_ZPWORD2	.word  0
         }}
     }
 
+    inline asmsub progstart() -> uword @AY {
+        %asm {{
+            lda  #<prog8_program_start
+            ldy  #>prog8_program_start
+        }}
+    }
+
     inline asmsub push(ubyte value @A) {
         %asm {{
             pha
@@ -564,6 +591,7 @@ cx16 {
     _cx16_vreg_storage
             .word 0,0,0,0,0,0,0,0
             .word 0,0,0,0,0,0,0,0
+            ; !notreached!
         }}
     }
 
@@ -584,3 +612,48 @@ cx16 {
     }
 
 }
+
+p8_sys_startup {
+    ; program startup and shutdown machinery. Needs to reside in normal system ram.
+
+asmsub  init_system()  {
+    ; Initializes the machine to a sane starting state.
+    ; Called automatically by the loader program logic.
+    ; Uppercase charset is activated.
+    %asm {{
+        sei
+        lda  #142
+        jsr  cbm.CHROUT     ; uppercase
+        lda  #147
+        jsr  cbm.CHROUT     ; clear screen
+        cli
+        rts
+    }}
+}
+
+asmsub  init_system_phase2()  {
+    %asm {{
+        cld
+        clc
+        clv
+        rts
+    }}
+}
+
+asmsub  cleanup_at_exit() {
+    ; executed when the main subroutine does rts
+    %asm {{
+_exitcodeCarry = *+1
+        lda  #0
+        lsr  a
+_exitcode = *+1
+        lda  #0        ; exit code possibly modified in exit()
+_exitcodeX = *+1
+        ldx  #0
+_exitcodeY = *+1
+        ldy  #0
+        rts
+    }}
+}
+
+}

compiler/res/prog8lib/pet32/textio.p8

@@ -12,7 +12,7 @@ txt {
 const ubyte DEFAULT_WIDTH = 40
 const ubyte DEFAULT_HEIGHT = 25
 
-romsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
+extsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
 
 
 sub  clear_screen() {
@@ -171,6 +171,7 @@ _mod		sta  $ffff		; modified
 		rts
 
 _screenrows	.word  cbm.Screen + range(0, 1000, 40)
+        ; !notreached!
 	}}
 }
 

compiler/res/prog8lib/prog8_math.p8

@@ -0,0 +1,5 @@
+prog8_math {
+    %option no_symbol_prefixing
+
+	%asminclude "library:math.asm"
+}

compiler/res/prog8lib/shared_cbm_diskio.p8

@@ -1,4 +1,4 @@
-; C64/C128 disk drive I/O routines.
+; shared CBM (C64/C128) disk drive I/O routines.
 
 %import textio
 %import conv
@@ -6,7 +6,7 @@
 %import syslib
 
 diskio {
-    %option no_symbol_prefixing, ignore_unused
+    %option merge, no_symbol_prefixing, ignore_unused
 
     const ubyte READ_IO_CHANNEL=12
     const ubyte WRITE_IO_CHANNEL=13
@@ -124,6 +124,7 @@ io_error:
     uword list_pattern
     uword list_blocks
     bool iteration_in_progress = false
+    bool write_iteration_in_progress = false
     str list_filetype = "???"       ; prg, seq, dir
     str list_filename = "?" * 50
 
@@ -377,6 +378,7 @@ _end        jsr  cbm.READST
 
     sub f_close() {
         ; -- end an iterative file loading session (close channels).
+        ;    it is safe to call this multiple times, or when no file is open for reading.
         if iteration_in_progress {
             cbm.CLRCHN()
             cbm.CLOSE(READ_IO_CHANNEL)
@@ -399,8 +401,12 @@ _end        jsr  cbm.READST
         cbm.SETNAM(string.length(filenameptr), filenameptr)
         cbm.SETLFS(WRITE_IO_CHANNEL, drivenumber, 1)
         void cbm.OPEN()             ; open 13,8,1,"filename"
-        if_cc
-            return cbm.READST()==0
+        if_cc {
+            if cbm.READST()==0 {
+                write_iteration_in_progress = true
+                return true
+            }
+        }
         cbm.CLOSE(WRITE_IO_CHANNEL)
         f_close_w()
         return false
@@ -422,8 +428,12 @@ _end        jsr  cbm.READST
 
     sub f_close_w() {
         ; -- end an iterative file writing session (close channels).
-        cbm.CLRCHN()
-        cbm.CLOSE(WRITE_IO_CHANNEL)
+        ;    it is safe to call this multiple times, or when no file is open for reading.
+        if write_iteration_in_progress {
+            cbm.CLRCHN()
+            cbm.CLOSE(WRITE_IO_CHANNEL)
+            write_iteration_in_progress = false
+        }
     }
 
 
@@ -590,6 +600,15 @@ io_error:
         cbm.CLOSE(1)
     }
 
+    sub exists(str filename) -> bool {
+        ; -- returns true if the given file exists on the disk, otherwise false
+        if f_open(filename) {
+            f_close()
+            return true
+        }
+        return false
+    }
+
     sub send_command(uword commandptr) {
         ; -- send a dos command to the drive
         cbm.SETNAM(string.length(commandptr), commandptr)

compiler/res/prog8lib/shared_cbm_textio_functions.p8

@@ -171,6 +171,7 @@ _allzero    lda  #'0'
             eor  P8ZP_SCRATCH_REG
             rts
 _offsets    .byte  128, 0, 64, 32, 64, 192, 128, 128
+            ; !notreached!
         }}
     }
 

compiler/res/prog8lib/shared_floats_functions.p8

@@ -256,4 +256,18 @@ inline asmsub pop() -> float @FAC1 {
     }}
 }
 
+sub lerp(float v0, float v1, float t) -> float {
+    ; Linear interpolation (LERP)
+    ; Precise method, which guarantees v = v1 when t = 1.
+    ; returns an interpolation between two inputs (v0, v1) for a parameter t in the closed unit interval [0, 1]
+    return (1 - t) * v0 + t * v1
+}
+
+sub lerp_fast(float v0, float v1, float t) -> float {
+    ; Linear interpolation (LERP)
+    ; Imprecise (but slightly faster) method, which does not guarantee v = v1 when t = 1
+    ; returns an interpolation between two inputs (v0, v1) for a parameter t in the closed unit interval [0, 1]
+    return v0 + t * (v1 - v0)
+}
+
 }

compiler/res/prog8lib/string.p8

@@ -178,7 +178,7 @@ _found      tya
             rts
 
 _str    .word 0
-
+            ; !notreached!
         }}
     }
 

compiler/res/prog8lib/virtual/diskio.p8

@@ -1,4 +1,8 @@
-; file I/O routines. (partially implemented)
+; File I/O routines for the VM target
+;
+; NOTE: not all is implemented.
+; NOTE: some calls are slightly different from the "official" diskio library because for example,
+;       here we cannot deal with multiple return values.
 
 %import textio
 %import syslib
@@ -11,7 +15,7 @@ diskio {
         %ir {{
             loadm.w r65534,diskio.load.filenameptr
             loadm.w r65535,diskio.load.address_override
-            syscall 48 (): r0.b
+            syscall 45 (): r0.b
             returnr.b r0
         }}
     }
@@ -59,37 +63,87 @@ diskio {
         ;    NOTE: the default input isn't yet set to this logical file, you must use reset_read_channel() to do this,
         ;          if you're going to read from it yourself instead of using f_read()!
 
-        txt.print("@TODO: f_open\n")
-        return false
+        %ir {{
+            loadm.w r65535,diskio.f_open.filenameptr
+            syscall 52 (r65535.w): r0.b
+            returnr.b r0
+        }}
     }
 
     sub f_read(uword bufferpointer, uword num_bytes) -> uword {
         ; -- read from the currently open file, up to the given number of bytes.
         ;    returns the actual number of bytes read.  (checks for End-of-file and error conditions)
-        txt.print("@TODO: f_read\n")
-        return 0
+        uword actual
+        repeat num_bytes {
+            %ir {{
+                syscall 54 (): r0.w
+                storem.w r0,$ff02
+            }}
+            if cx16.r0H==0
+                return actual
+            @(bufferpointer) = cx16.r0L
+            bufferpointer++
+            actual++
+        }
+        return actual
     }
 
     sub f_read_all(uword bufferpointer) -> uword {
         ; -- read the full contents of the file, returns number of bytes read.
         ;    It is assumed the file size is less than 64 K.
-        txt.print("@TODO: f_read_all\n")
-        return 0
+        uword actual
+        repeat {
+            %ir {{
+                syscall 54 (): r0.w
+                storem.w r0,$ff02
+            }}
+            if cx16.r0H==0
+                return actual
+            @(bufferpointer) = cx16.r0L
+            bufferpointer++
+            actual++
+        }
     }
 
     sub f_readline(uword bufptr) -> ubyte {
         ; Routine to read text lines from a text file. Lines must be less than 255 characters.
         ; Reads characters from the input file UNTIL a newline or return character (or EOF).
         ; The line read will be 0-terminated in the buffer (and not contain the end of line character).
-        ; The length of the line is returned in Y. Note that an empty line is okay and is length 0!
-        ; This routine is not able here to return the status as well in a secondary return value, so you have to do that yourself.
-        txt.print("@TODO: f_readline\n")
-        return 0
+        ; The length of the line is returned. Note that an empty line is okay and is length 0!
+        ; The success status is returned in the Carry flag instead: C set = success, C clear = failure/endoffile
+        ubyte size
+        repeat {
+            %ir {{
+                syscall 54 (): r0.w
+                storem.w r0,$ff02
+            }}
+
+            if cx16.r0H==0 {
+                sys.clear_carry()
+                return size
+            } else {
+                if cx16.r0L == '\n' or cx16.r0L=='\r' {
+                    @(bufptr) = 0
+                    sys.set_carry()
+                    return size
+                }
+                @(bufptr) = cx16.r0L
+                bufptr++
+                size++
+                if_z {
+                    @(bufptr) = 0
+                    return 255
+                }
+            }
+        }
     }
 
     sub f_close() {
         ; -- end an iterative file loading session (close channels).
-        txt.print("@TODO: f_close\n")
+        ;    it is safe to call this multiple times, or when no file is open for reading.
+        %ir {{
+            syscall 56 ()
+        }}
     }
 
 
@@ -102,20 +156,36 @@ diskio {
         ;    (for example, if it already exists). This is different than f_open()!
         ;    To be 100% sure if this call was successful, you have to use status()
         ;    and check the drive's status message!
-        txt.print("@TODO: f_open_w\n")
-        return false
+        %ir {{
+            loadm.w r65535,diskio.f_open_w.filenameptr
+            syscall 53 (r65535.w): r0.b
+            returnr.b r0
+        }}
     }
 
     sub f_write(uword bufferpointer, uword num_bytes) -> bool {
         ; -- write the given number of bytes to the currently open file
         ;    you can call this multiple times to append more data
-        txt.print("@TODO: f_write\n")
-        return false
+        repeat num_bytes {
+            %ir {{
+                loadm.w r0,diskio.f_write.bufferpointer
+                loadi.b r1,r0
+                syscall 55 (r1.b): r0.b
+                storem.b r0,$ff02
+            }}
+            if cx16.r0L==0
+                return false
+            bufferpointer++
+        }
+        return true
     }
 
     sub f_close_w() {
         ; -- end an iterative file writing session (close channels).
-        txt.print("@TODO: f_close_w\n")
+        ;    it is safe to call this multiple times, or when no file is open for reading.
+        %ir {{
+            syscall 57 ()
+        }}
     }
 
 
@@ -123,23 +193,34 @@ diskio {
 
     sub chdir(str path) {
         ; -- change current directory.
-        txt.print("@TODO: chdir\n")
+        %ir {{
+            loadm.w r65535,diskio.chdir.path
+            syscall 50 (r65535.w)
+        }}
     }
 
     sub mkdir(str name) {
         ; -- make a new subdirectory.
-        txt.print("@TODO: mkdir\n")
+        %ir {{
+            loadm.w r65535,diskio.mkdir.name
+            syscall 49 (r65535.w)
+        }}
     }
 
     sub rmdir(str name) {
         ; -- remove a subdirectory.
-        txt.print("@TODO: rmdir\n")
+        %ir {{
+            loadm.w r65535,diskio.rmdir.name
+            syscall 51 (r65535.w)
+        }}
     }
 
     sub curdir() -> uword {
         ; return current directory name or 0 if error
-        txt.print("@TODO: curdir\n")
-        return 0
+        %ir {{
+            syscall 48 (): r0.w
+            returnr.w r0
+        }}
     }
 
     sub status() -> str {
@@ -159,7 +240,7 @@ diskio {
             loadm.w r65533,diskio.save.filenameptr
             loadm.w r65534,diskio.save.start_address
             loadm.w r65535,diskio.save.savesize
-            syscall 45 (r65532.b, r65533.w, r65534.w, r65535.w): r0.b
+            syscall 42 (r65532.b, r65533.w, r65534.w, r65535.w): r0.b
             returnr.b r0
         }}
     }
@@ -171,7 +252,7 @@ diskio {
             loadm.w r65533,diskio.save.filenameptr
             loadm.w r65534,diskio.save.start_address
             loadm.w r65535,diskio.save.savesize
-            syscall 45 (r65532.b, r65533.w, r65534.w, r65535.w): r0.b
+            syscall 42 (r65532.b, r65533.w, r65534.w, r65535.w): r0.b
             returnr.b r0
         }}
     }
@@ -186,7 +267,7 @@ diskio {
         %ir {{
             loadm.w r65534,diskio.load.filenameptr
             loadm.w r65535,diskio.load.address_override
-            syscall 43 (r65534.w, r65535.w): r0.w
+            syscall 40 (r65534.w, r65535.w): r0.w
             returnr.w r0
         }}
     }
@@ -198,7 +279,7 @@ diskio {
         %ir {{
             loadm.w r65534,diskio.load_raw.filenameptr
             loadm.w r65535,diskio.load_raw.start_address
-            syscall 44 (r65534.w, r65535.w): r0.w
+            syscall 41 (r65534.w, r65535.w): r0.w
             returnr.w r0
         }}
     }
@@ -207,7 +288,7 @@ diskio {
         ; -- delete a file on the drive
         %ir {{
             loadm.w r65535,diskio.delete.filenameptr
-            syscall 46 (r65535.w)
+            syscall 43 (r65535.w)
         }}
     }
 
@@ -216,7 +297,17 @@ diskio {
         %ir {{
             loadm.w r65534,diskio.rename.oldfileptr
             loadm.w r65535,diskio.rename.newfileptr
-            syscall 47 (r65534.w, r65535.w)
+            syscall 44 (r65534.w, r65535.w)
         }}
     }
+
+    sub exists(str filename) -> bool {
+        ; -- returns true if the given file exists on the disk, otherwise false
+        if f_open(filename) {
+            f_close()
+            return true
+        }
+        return false
+    }
+
 }

compiler/res/prog8lib/virtual/floats.p8

@@ -192,7 +192,6 @@ sub push(float value) {
     %ir {{
         loadm.f  fr65535,floats.push.value
         push.f  fr65535
-        return
     }}
 }
 
@@ -204,4 +203,18 @@ sub pop() -> float {
     }}
 }
 
+sub lerp(float v0, float v1, float t) -> float {
+    ; Linear interpolation (LERP)
+    ; Precise method, which guarantees v = v1 when t = 1.
+    ; returns an interpolation between two inputs (v0, v1) for a parameter t in the closed unit interval [0, 1]
+    return (1 - t) * v0 + t * v1
+}
+
+sub lerp_fast(float v0, float v1, float t) -> float {
+    ; Linear interpolation (LERP)
+    ; Imprecise (but slightly faster) method, which does not guarantee v = v1 when t = 1
+    ; returns an interpolation between two inputs (v0, v1) for a parameter t in the closed unit interval [0, 1]
+    return v0 + t * (v1 - v0)
+}
+
 }

compiler/res/prog8lib/virtual/math.p8

@@ -398,4 +398,23 @@ math {
         cx16.r15 ^= $ffff
         cx16.r14 ^= $ffff
     }
+
+
+    sub lerp(ubyte v0, ubyte v1, ubyte t) -> ubyte {
+        ; Linear interpolation (LERP)
+        ; returns an interpolation between two inputs (v0, v1) for a parameter t in the interval [0, 255]
+        ; guarantees v = v1 when t = 255
+        return v0 + msb(t as uword * (v1 - v0) + 255)
+    }
+
+    sub lerpw(uword v0, uword v1, uword t) -> uword {
+        ; Linear interpolation (LERP) on word values
+        ; returns an interpolation between two inputs (v0, v1) for a parameter t in the interval [0, 65535]
+        ; guarantees v = v1 when t = 65535
+        t *= v1-v0
+        cx16.r0 = math.mul16_last_upper()
+        if t!=0
+            cx16.r0++
+        return v0 + cx16.r0
+    }
 }

compiler/res/prog8lib/virtual/prog8_math.p8

@@ -0,0 +1 @@
+; for the VM this is just empty, everything math is in math.p8