fix sieve example

* feat: additional trig functions

* fix: 64tass won't assemble a proc named 'sec'

* fix: indentation

.github/workflows/all-ci.yml

@@ -18,30 +18,24 @@ jobs:
           git clone --depth=1 https://github.com/irmen/64tass
           cd 64tass
           make -j4
-          sudo make install
+          sudo make install          
 
       - name: Set up JDK 11
         uses: actions/setup-java@v4
         with:
           java-version: 11
-          distribution: adopt
+          distribution: temurin
 
       - name: Build and test with Gradle
-        run: ./gradlew build shadowJar --no-daemon
+        run: |
+          ./gradlew build shadowJar --no-daemon
+          sha256sum -b compiler/build/libs/*-all.jar > compiler/build/libs/hash.txt
         
       - name: Create compiler shadowJar artifact
         uses: actions/upload-artifact@v4
         with:
           name: prog8-compiler-jar-zipped
-          path: compiler/build/libs/*-all.jar
+          path: |
+            compiler/build/libs/*-all.jar
+            compiler/build/libs/hash.txt
 
-      - uses: MCJack123/ghaction-generate-release-hashes@v4
-        if: "github.event_name == 'release'"
-        with:
-            hash-type: sha256
-            file-name: hashes.txt
-      - uses: actions/upload-artifact@v4
-        if: "github.event_name == 'release'"
-        with:
-            name: Artifact Hashes
-            path: hashes.txt        

README.md

@@ -52,7 +52,7 @@ What does Prog8 provide?
 ------------------------
 
 - all advantages of a higher level language over having to write assembly code manually
-- programs run very fast because compilation to native machine code. It's possible to write games purely in Prog8, and even certain raster interrupt 'demoscene' effects.
+- programs run very fast because compilation to native machine code
 - modularity, symbol scoping, subroutines
 - various data types other than just bytes (16-bit words, floats, strings)
 - floating point math is supported if the target system provides floating point library routines (C64 and Cx16 both do)
@@ -71,7 +71,7 @@ What does Prog8 provide?
 - convenience abstractions for low level aspects such as ZeroPage handling, program startup, explicit memory addresses
 - 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, and provides them also on the C64.
-- encode strings and characters into petscii or screencodes as desired (C64/Cx16)
+- encode strings and characters into petscii or screencodes or even other encodings, as desired (C64/Cx16)
 
 *Rapid edit-compile-run-debug cycle:*
 
@@ -119,14 +119,13 @@ This code calculates prime numbers using the Sieve of Eratosthenes algorithm::
 
     %import textio
     %zeropage basicsafe
-    
+
     main {
-    
-        ubyte[256] sieve
+        bool[256] sieve
         ubyte candidate_prime = 2       ; is increased in the loop
     
         sub start() {
-            sys.memset(sieve, 256, false)   ; clear the sieve
+            sys.memset(sieve, 256, 0)   ; clear the sieve
             txt.print("prime numbers up to 255:\n\n")
             ubyte amount=0
             repeat {
@@ -162,9 +161,6 @@ This code calculates prime numbers using the Sieve of Eratosthenes algorithm::
         }
     }
 
-
-
-
 when compiled an ran on a C-64 you'll get:
 
 ![c64 screen](docs/source/_static/primes_example.png)

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

@@ -44,7 +44,7 @@ sealed class PtExpression(val type: DataType, position: Position) : PtNode(posit
             }
             is PtContainmentCheck -> other is PtContainmentCheck && other.type==type && other.element isSameAs element && other.iterable isSameAs iterable
             is PtIdentifier -> other is PtIdentifier && other.type==type && other.name==name
-            is PtMachineRegister -> other is PtMachineRegister && other.type==type && other.register==register
+            is PtIrRegister -> other is PtIrRegister && other.type==type && other.register==register
             is PtMemoryByte -> other is PtMemoryByte && other.address isSameAs address
             is PtNumber -> other is PtNumber && other.type==type && other.number==number
             is PtBool -> other is PtBool && other.value==value
@@ -87,7 +87,7 @@ sealed class PtExpression(val type: DataType, position: Position) : PtNode(posit
             is PtContainmentCheck -> false
             is PtFunctionCall -> false
             is PtIdentifier -> true
-            is PtMachineRegister -> true
+            is PtIrRegister -> true
             is PtMemoryByte -> address is PtNumber || address is PtIdentifier
             is PtBool -> true
             is PtNumber -> true
@@ -139,13 +139,9 @@ class PtAddressOf(position: Position) : PtExpression(DataType.UWORD, position) {
 
 class PtArrayIndexer(elementType: DataType, position: Position): PtExpression(elementType, position) {
     val variable: PtIdentifier
-        get() {
-            require((children[0] as? PtIdentifier)?.type in ArrayDatatypes+DataType.STR)        // TODO remove
-            return children[0] as PtIdentifier
-        }
+        get() = children[0] as PtIdentifier
     val index: PtExpression
         get() = children[1] as PtExpression
-
     val splitWords: Boolean
         get() = variable.type in SplitWordArrayTypes
 
@@ -210,11 +206,6 @@ class PtFunctionCall(val name: String,
                      val void: Boolean,
                      type: DataType,
                      position: Position) : PtExpression(type, position) {
-    init {
-        if(!void)
-            require(type!=DataType.UNDEFINED)
-    }
-
     val args: List<PtExpression>
         get() = children.map { it as PtExpression }
 }
@@ -354,7 +345,7 @@ 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 PtMachineRegister(val register: Int, type: DataType, position: Position) : PtExpression(type, position)
+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

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

@@ -10,7 +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 PtAssignTarget -> "<target>"
+            is PtAssignTarget -> if(node.void) "<void>" else "<target>"
             is PtAssignment -> "<assign>"
             is PtAugmentedAssign -> "<inplace-assign> ${node.operator}"
             is PtBreakpoint -> "%breakpoint"
@@ -34,7 +34,7 @@ fun printAst(root: PtNode, skipLibraries: Boolean, output: (text: String) -> Uni
                 str + node.name + "()"
             }
             is PtIdentifier -> "${node.name} ${type(node.type)}"
-            is PtMachineRegister -> "VMREG#${node.register} ${type(node.type)}"
+            is PtIrRegister -> "IRREG#${node.register} ${type(node.type)}"
             is PtMemoryByte -> "@()"
             is PtNumber -> {
                 val numstr = if(node.type == DataType.FLOAT) node.number.toString() else node.number.toHex()

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

@@ -41,9 +41,18 @@ class PtSubroutineParameter(name: String, val type: DataType, position: Position
 sealed interface IPtAssignment {
     val children: MutableList<PtNode>
     val target: PtAssignTarget
-        get() = children[0] as PtAssignTarget
+        get() {
+            if(children.size==2)
+                return children[0] as PtAssignTarget
+            else if(children.size<2)
+                throw AssemblyError("incomplete node")
+            else
+                throw AssemblyError("no singular target")
+        }
     val value: PtExpression
-        get() = children[1] as PtExpression
+        get() = children.last() as PtExpression
+    val multiTarget: Boolean
+        get() = children.size>2
 }
 
 class PtAssignment(position: Position) : PtNode(position), IPtAssignment
@@ -51,7 +60,7 @@ class PtAssignment(position: Position) : PtNode(position), IPtAssignment
 class PtAugmentedAssign(val operator: String, position: Position) : PtNode(position), IPtAssignment
 
 
-class PtAssignTarget(position: Position) : PtNode(position) {
+class PtAssignTarget(val void: Boolean, position: Position) : PtNode(position) {
     val identifier: PtIdentifier?
         get() = children.single() as? PtIdentifier
     val array: PtArrayIndexer?
@@ -69,7 +78,7 @@ class PtAssignTarget(position: Position) : PtNode(position) {
             }
         }
 
-    infix fun isSameAs(expression: PtExpression): Boolean = expression.isSameAs(this)
+    infix fun isSameAs(expression: PtExpression): Boolean = !void && expression.isSameAs(this)
 }
 
 

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

@@ -96,9 +96,9 @@ val BuiltinFunctions: Map<String, FSignature> = mapOf(
     "sqrt__ubyte" to FSignature(true, listOf(FParam("value", arrayOf(DataType.UBYTE))), DataType.UBYTE),
     "sqrt__uword" to FSignature(true, listOf(FParam("value", arrayOf(DataType.UWORD))), DataType.UBYTE),
     "sqrt__float" to FSignature(true, listOf(FParam("value", arrayOf(DataType.FLOAT))), DataType.FLOAT),
-    "divmod"           to FSignature(false, listOf(FParam("number", arrayOf(DataType.UBYTE, DataType.UWORD)), FParam("divident", arrayOf(DataType.UBYTE, DataType.UWORD)), FParam("division", arrayOf(DataType.UBYTE, DataType.UWORD)), FParam("remainder", arrayOf(DataType.UBYTE, DataType.UWORD))), null),
-    "divmod__ubyte"    to FSignature(false, listOf(FParam("number", arrayOf(DataType.UBYTE)), FParam("divident", arrayOf(DataType.UBYTE)), FParam("division", arrayOf(DataType.UBYTE)), FParam("remainder", arrayOf(DataType.UBYTE))), null),
-    "divmod__uword"    to FSignature(false, listOf(FParam("number", arrayOf(DataType.UWORD)), FParam("divident", arrayOf(DataType.UWORD)), FParam("division", arrayOf(DataType.UWORD)), FParam("remainder", arrayOf(DataType.UWORD))), null),
+    "divmod"           to FSignature(false, listOf(FParam("dividend", arrayOf(DataType.UBYTE, DataType.UWORD)), FParam("divisor", arrayOf(DataType.UBYTE, DataType.UWORD)), FParam("quotient", arrayOf(DataType.UBYTE, DataType.UWORD)), FParam("remainder", arrayOf(DataType.UBYTE, DataType.UWORD))), null),
+    "divmod__ubyte"    to FSignature(false, listOf(FParam("dividend", arrayOf(DataType.UBYTE)), FParam("divisor", arrayOf(DataType.UBYTE)), FParam("quotient", arrayOf(DataType.UBYTE)), FParam("remainder", arrayOf(DataType.UBYTE))), null),
+    "divmod__uword"    to FSignature(false, listOf(FParam("dividend", arrayOf(DataType.UWORD)), FParam("divisor", arrayOf(DataType.UWORD)), FParam("quotient", arrayOf(DataType.UWORD)), FParam("remainder", arrayOf(DataType.UWORD))), null),
     "any"       to FSignature(true, listOf(FParam("values", ArrayDatatypes)), DataType.BOOL),
     "all"       to FSignature(true, listOf(FParam("values", ArrayDatatypes)), DataType.BOOL),
     "lsb"       to FSignature(true, listOf(FParam("value", arrayOf(DataType.UWORD, DataType.WORD))), DataType.UBYTE),
@@ -133,4 +133,9 @@ val BuiltinFunctions: Map<String, FSignature> = mapOf(
     "call"      to FSignature(false, listOf(FParam("address", arrayOf(DataType.UWORD))), DataType.UWORD),
 )
 
-val InplaceModifyingBuiltinFunctions = setOf("setlsb", "setmsb", "rol", "ror", "rol2", "ror2", "sort", "reverse")
+val InplaceModifyingBuiltinFunctions = setOf(
+    "setlsb", "setmsb",
+    "rol", "ror", "rol2", "ror2",
+    "sort", "reverse",
+    "divmod", "divmod__ubyte", "divmod__uword"
+)

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

@@ -20,6 +20,7 @@ class CompilationOptions(val output: OutputType,
                          var asmListfile: Boolean = false,
                          var includeSourcelines: Boolean = false,
                          var dumpVariables: Boolean = false,
+                         var dumpSymbols: Boolean = false,
                          var experimentalCodegen: Boolean = false,
                          var varsHighBank: Int? = null,
                          var varsGolden: Boolean = false,

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

@@ -26,6 +26,9 @@ interface IMachineDefinition {
     fun initializeMemoryAreas(compilerOptions: CompilationOptions)
     fun getFloatAsmBytes(num: Number): String
 
+    fun convertFloatToBytes(num: Double): List<UByte>
+    fun convertBytesToFloat(bytes: List<UByte>): Double
+
     fun importLibs(compilerOptions: CompilationOptions, compilationTargetName: String): List<String>
     fun launchEmulator(selectedEmulator: Int, programNameWithPath: Path)
     fun isIOAddress(address: UInt): Boolean

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

@@ -5,7 +5,10 @@ enum class Encoding(val prefix: String) {
     PETSCII("petscii"),         // c64/c128/cx16
     SCREENCODES("sc"),          // c64/c128/cx16
     ATASCII("atascii"),         // atari
-    ISO("iso")                  // cx16
+    ISO("iso"),                 // cx16  (iso-8859-15)
+    ISO5("iso5"),               // cx16  (iso-8859-5, cyrillic)
+    ISO16("iso16"),             // cx16  (iso-8859-16, eastern european)
+    CP437("cp437")              // cx16  (ibm pc, codepage 437)
 }
 
 interface IStringEncoding {

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

@@ -1,6 +1,6 @@
 package prog8.code.core
 
-val AssociativeOperators = setOf("+", "*", "&", "|", "^", "==", "!=", "and", "or", "xor")
+val AssociativeOperators = setOf("+", "*", "&", "|", "^", "==", "!=", "xor")       // note: and,or are no longer associative because of Shortcircuit/McCarthy evaluation
 val ComparisonOperators = setOf("==", "!=", "<", ">", "<=", ">=")
 val LogicalOperators = setOf("and", "or", "xor", "not", "in")
 val BitwiseOperators = setOf("&", "|", "^", "~")

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

@@ -1,13 +1,18 @@
 package prog8.code.optimize
 
+import prog8.code.StRomSub
+import prog8.code.SymbolTable
 import prog8.code.ast.*
 import prog8.code.core.*
 
 
-fun optimizeIntermediateAst(program: PtProgram, options: CompilationOptions, errors: IErrorReporter) {
+fun optimizeIntermediateAst(program: PtProgram, options: CompilationOptions, st: SymbolTable, errors: IErrorReporter) {
     if (!options.optimize)
         return
-    while(errors.noErrors() && optimizeCommonSubExpressions(program, errors)>0) {
+    while (errors.noErrors() &&
+        (optimizeCommonSubExpressions(program, errors)
+                + optimizeAssignTargets(program, st, errors)) > 0
+    ) {
         // keep rolling
     }
 }
@@ -27,11 +32,15 @@ private var tempVarCounter = 0
 private fun optimizeCommonSubExpressions(program: PtProgram, errors: IErrorReporter): Int {
 
     fun extractableSubExpr(expr: PtExpression): Boolean {
+        if(expr is PtArrayIndexer && expr.index.isSimple())
+            return false
+        if (expr is PtMemoryByte && expr.address.isSimple())
+            return false
+
         val result = if(expr is PtBinaryExpression)
             expr.type !in ByteDatatypes ||
-                    !expr.left.isSimple() ||
-                    !expr.right.isSimple() ||
-                    (expr.operator !in LogicalOperators && expr.operator !in BitwiseOperators)
+            !(expr.left.isSimple() && expr.right.isSimple()) ||
+            (expr.operator !in LogicalOperators && expr.operator !in BitwiseOperators)
         else if (expr is PtArrayIndexer && expr.type !in ByteDatatypes)
             true
         else
@@ -89,7 +98,7 @@ private fun optimizeCommonSubExpressions(program: PtProgram, errors: IErrorRepor
         singleReplacement2.parent = occurrence2.parent
 
         val tempassign = PtAssignment(binexpr.position).also { assign ->
-            assign.add(PtAssignTarget(binexpr.position).also { tgt->
+            assign.add(PtAssignTarget(false, binexpr.position).also { tgt->
                 tgt.add(PtIdentifier("$containerScopedName.$tempvarName", datatype, binexpr.position))
             })
             assign.add(occurrence1)
@@ -109,6 +118,79 @@ private fun optimizeCommonSubExpressions(program: PtProgram, errors: IErrorRepor
 }
 
 
+private fun optimizeAssignTargets(program: PtProgram, st: SymbolTable, errors: IErrorReporter): Int {
+    var changes = 0
+    walkAst(program) { node: PtNode, depth: Int ->
+        if(node is PtAssignment) {
+            val value = node.value
+            val functionName = when(value) {
+                is PtBuiltinFunctionCall -> value.name
+                is PtFunctionCall -> value.name
+                else -> null
+            }
+            if(functionName!=null) {
+                val stNode = st.lookup(functionName)
+                if (stNode is StRomSub) {
+                    require(node.children.size==stNode.returns.size+1) {
+                        "number of targets must match return values"
+                    }
+                    node.children.zip(stNode.returns).withIndex().forEach { (index, xx) ->
+                        val target = xx.first as PtAssignTarget
+                        val returnedRegister = xx.second.register.registerOrPair
+                        if(returnedRegister!=null && !target.void && target.identifier!=null) {
+                            if(isSame(target.identifier!!, xx.second.type, returnedRegister)) {
+                                // output register is already identical to target register, so it can become void
+                                val voidTarget = PtAssignTarget(true, target.position)
+                                node.children[index] = voidTarget
+                                voidTarget.parent = node
+                                changes++
+                            }
+                        }
+                    }
+                }
+                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)
+                    value.children.forEach { voidCall.add(it) }
+                    node.parent.children[index] = voidCall
+                    voidCall.parent = node.parent
+                    changes++
+                }
+            }
+        }
+        true
+    }
+    return changes
+}
+
+internal fun isSame(identifier: PtIdentifier, type: DataType, returnedRegister: RegisterOrPair): Boolean {
+    if(returnedRegister in Cx16VirtualRegisters) {
+        val regname = returnedRegister.name.lowercase()
+        val identifierRegName = identifier.name.substringAfterLast('.')
+        /*
+            cx16.r?    UWORD
+            cx16.r?s   WORD
+            cx16.r?L   UBYTE
+            cx16.r?H   UBYTE
+            cx16.r?sL  BYTE
+            cx16.r?sH  BYTE
+         */
+        if(identifier.type in ByteDatatypes && type in ByteDatatypes) {
+            if(identifier.name.startsWith("cx16.$regname") && identifierRegName.startsWith(regname)) {
+                return identifierRegName.substring(2) in arrayOf("", "L", "sL")     // note: not the -H (msb) variants!
+            }
+        }
+        else if(identifier.type in WordDatatypes && type in WordDatatypes) {
+            if(identifier.name.startsWith("cx16.$regname") && identifierRegName.startsWith(regname)) {
+                return identifierRegName.substring(2) in arrayOf("", "s")
+            }
+        }
+    }
+    return false   // there are no identifiers directly corresponding to cpu registers
+}
+
+
 internal fun findScopeName(node: PtNode): String {
     var parent=node
     while(parent !is PtNamedNode)

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

@@ -19,3 +19,15 @@ class C64Target: ICompilationTarget, IStringEncoding by Encoder, IMemSizer by Cb
         fun viceMonListName(baseFilename: String) = "$baseFilename.vice-mon-list"
     }
 }
+
+
+
+fun getCompilationTargetByName(name: String) = when(name.lowercase()) {
+        C64Target.NAME -> C64Target()
+        C128Target.NAME -> C128Target()
+        Cx16Target.NAME -> Cx16Target()
+        PETTarget.NAME -> PETTarget()
+        AtariTarget.NAME -> AtariTarget()
+        VMTarget.NAME -> VMTarget()
+        else -> throw IllegalArgumentException("invalid compilation target")
+    }

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

@@ -4,9 +4,7 @@ import com.github.michaelbull.result.fold
 import prog8.code.core.Encoding
 import prog8.code.core.IStringEncoding
 import prog8.code.core.InternalCompilerException
-import prog8.code.target.cbm.AtasciiEncoding
-import prog8.code.target.cbm.IsoEncoding
-import prog8.code.target.cbm.PetsciiEncoding
+import prog8.code.target.encodings.*
 
 
 object Encoder: IStringEncoding {
@@ -18,6 +16,9 @@ object Encoder: IStringEncoding {
             Encoding.SCREENCODES -> PetsciiEncoding.encodeScreencode(str, true)
             Encoding.ISO -> IsoEncoding.encode(str)
             Encoding.ATASCII -> AtasciiEncoding.encode(str)
+            Encoding.ISO5 -> IsoCyrillicEncoding.encode(str)
+            Encoding.ISO16 -> IsoEasternEncoding.encode(str)
+            Encoding.CP437 -> Cp437Encoding.encode(str)
             else -> throw InternalCompilerException("unsupported encoding $encoding")
         }
         return coded.fold(
@@ -31,6 +32,9 @@ object Encoder: IStringEncoding {
             Encoding.SCREENCODES -> PetsciiEncoding.decodeScreencode(bytes, true)
             Encoding.ISO -> IsoEncoding.decode(bytes)
             Encoding.ATASCII -> AtasciiEncoding.decode(bytes)
+            Encoding.ISO5 -> IsoCyrillicEncoding.decode(bytes)
+            Encoding.ISO16 -> IsoEasternEncoding.decode(bytes)
+            Encoding.CP437 -> Cp437Encoding.decode(bytes)
             else -> throw InternalCompilerException("unsupported encoding $encoding")
         }
         return decoded.fold(

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

@@ -22,6 +22,8 @@ class AtariMachineDefinition: IMachineDefinition {
     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 importLibs(compilerOptions: CompilationOptions, compilationTargetName: String): List<String> {
         return if (compilerOptions.output == OutputType.XEX)

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

@@ -25,6 +25,17 @@ class C128MachineDefinition: IMachineDefinition {
 
     override fun getFloatAsmBytes(num: Number) = Mflpt5.fromNumber(num).makeFloatFillAsm()
 
+    override fun convertFloatToBytes(num: Double): List<UByte> {
+        val m5 = Mflpt5.fromNumber(num)
+        return listOf(m5.b0, m5.b1, m5.b2, m5.b3, m5.b4)
+    }
+
+    override fun convertBytesToFloat(bytes: List<UByte>): Double {
+        require(bytes.size==5) { "need 5 bytes" }
+        val m5 = Mflpt5(bytes[0], bytes[1], bytes[2], bytes[3], bytes[4])
+        return m5.toDouble()
+    }
+
     override fun importLibs(compilerOptions: CompilationOptions, compilationTargetName: String): List<String> {
         return if (compilerOptions.launcher == CbmPrgLauncherType.BASIC || compilerOptions.output == OutputType.PRG)
             listOf("syslib")

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

@@ -26,6 +26,17 @@ class C64MachineDefinition: IMachineDefinition {
 
     override fun getFloatAsmBytes(num: Number) = Mflpt5.fromNumber(num).makeFloatFillAsm()
 
+    override fun convertFloatToBytes(num: Double): List<UByte> {
+        val m5 = Mflpt5.fromNumber(num)
+        return listOf(m5.b0, m5.b1, m5.b2, m5.b3, m5.b4)
+    }
+
+    override fun convertBytesToFloat(bytes: List<UByte>): Double {
+        require(bytes.size==5) { "need 5 bytes" }
+        val m5 = Mflpt5(bytes[0], bytes[1], bytes[2], bytes[3], bytes[4])
+        return m5.toDouble()
+    }
+
     override fun importLibs(compilerOptions: CompilationOptions, compilationTargetName: String): List<String> {
         return if (compilerOptions.launcher == CbmPrgLauncherType.BASIC || compilerOptions.output == OutputType.PRG)
             listOf("syslib")

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

@@ -24,6 +24,18 @@ class CX16MachineDefinition: IMachineDefinition {
     override lateinit var golden: GoldenRam
 
     override fun getFloatAsmBytes(num: Number) = Mflpt5.fromNumber(num).makeFloatFillAsm()
+
+    override fun convertFloatToBytes(num: Double): List<UByte> {
+        val m5 = Mflpt5.fromNumber(num)
+        return listOf(m5.b0, m5.b1, m5.b2, m5.b3, m5.b4)
+    }
+
+    override fun convertBytesToFloat(bytes: List<UByte>): Double {
+        require(bytes.size==5) { "need 5 bytes" }
+        val m5 = Mflpt5(bytes[0], bytes[1], bytes[2], bytes[3], bytes[4])
+        return m5.toDouble()
+    }
+
     override fun importLibs(compilerOptions: CompilationOptions, compilationTargetName: String): List<String> {
         return if (compilerOptions.launcher == CbmPrgLauncherType.BASIC || compilerOptions.output == OutputType.PRG)
             listOf("syslib")

codeCore/src/prog8/code/target/encodings/AtasciiEncoding.kt

@@ -1,4 +1,4 @@
-package prog8.code.target.cbm
+package prog8.code.target.encodings
 
 import com.github.michaelbull.result.Ok
 import com.github.michaelbull.result.Result

codeCore/src/prog8/code/target/encodings/Cp437Encoding.kt

@@ -0,0 +1,69 @@
+package prog8.code.target.encodings
+
+import com.github.michaelbull.result.Err
+import com.github.michaelbull.result.Ok
+import com.github.michaelbull.result.Result
+import java.io.CharConversionException
+import java.nio.charset.Charset
+
+object Cp437Encoding {
+    val charset: Charset = Charset.forName("IBM437")
+
+    fun encode(str: String): Result<List<UByte>, CharConversionException> {
+        return try {
+            val mapped = str.map { chr ->
+                when (chr) {
+                    '\u0000' -> 0u
+                    '\u00a0' -> 255u
+                    '☺' -> 1u
+                    '☻' -> 2u
+                    '♥' -> 3u
+                    '♦' -> 4u
+                    '♣' -> 5u
+                    '♠' -> 6u
+                    '•' -> 7u
+                    '◘' -> 8u
+                    '○' -> 9u
+                    '◙' -> 10u
+                    '♂' -> 11u
+                    '♀' -> 12u
+                    '♪' -> 13u
+                    '♫' -> 14u
+                    '☼' -> 15u
+                    '►' -> 16u
+                    '◄' -> 17u
+                    '↕' -> 18u
+                    '‼' -> 19u
+                    '¶' -> 20u
+                    '§' -> 21u
+                    '▬' -> 22u
+                    '↨' -> 23u
+                    '↑' -> 24u
+                    '↓' -> 25u
+                    '→' -> 26u
+                    '←' -> 27u
+                    '∟' -> 28u
+                    '↔' -> 29u
+                    '▲' -> 30u
+                    '▼' -> 31u
+                    in '\u8000'..'\u80ff' -> {
+                        // special case: take the lower 8 bit hex value directly
+                        (chr.code - 0x8000).toUByte()
+                    }
+                    else -> charset.encode(chr.toString())[0].toUByte()
+                }
+            }
+            Ok(mapped)
+        } catch (ce: CharConversionException) {
+            Err(ce)
+        }
+    }
+
+    fun decode(bytes: Iterable<UByte>): Result<String, CharConversionException> {
+        return try {
+            Ok(String(bytes.map { it.toByte() }.toByteArray(), charset))
+        } catch (ce: CharConversionException) {
+            Err(ce)
+        }
+    }
+}

codeCore/src/prog8/code/target/encodings/IsoEncoding.kt

@@ -1,4 +1,4 @@
-package prog8.code.target.cbm
+package prog8.code.target.encodings
 
 import com.github.michaelbull.result.Err
 import com.github.michaelbull.result.Ok
@@ -6,8 +6,8 @@ import com.github.michaelbull.result.Result
 import java.io.CharConversionException
 import java.nio.charset.Charset
 
-object IsoEncoding {
-    val charset: Charset = Charset.forName("ISO-8859-15")
+open class IsoEncodingBase(charsetName: String) {
+    val charset: Charset = Charset.forName(charsetName)
 
     fun encode(str: String): Result<List<UByte>, CharConversionException> {
         return try {
@@ -35,3 +35,8 @@ object IsoEncoding {
         }
     }
 }
+
+
+object IsoEncoding: IsoEncodingBase("ISO-8859-15")
+object IsoCyrillicEncoding: IsoEncodingBase("ISO-8859-5")
+object IsoEasternEncoding: IsoEncodingBase("ISO-8859-16")

codeCore/src/prog8/code/target/encodings/PetsciiEncoding.kt

@@ -1,4 +1,4 @@
-package prog8.code.target.cbm
+package prog8.code.target.encodings
 
 import com.github.michaelbull.result.Err
 import com.github.michaelbull.result.Ok

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

@@ -25,6 +25,17 @@ class PETMachineDefinition: IMachineDefinition {
 
     override fun getFloatAsmBytes(num: Number) = Mflpt5.fromNumber(num).makeFloatFillAsm()
 
+    override fun convertFloatToBytes(num: Double): List<UByte> {
+        val m5 = Mflpt5.fromNumber(num)
+        return listOf(m5.b0, m5.b1, m5.b2, m5.b3, m5.b4)
+    }
+
+    override fun convertBytesToFloat(bytes: List<UByte>): Double {
+        require(bytes.size==5) { "need 5 bytes" }
+        val m5 = Mflpt5(bytes[0], bytes[1], bytes[2], bytes[3], bytes[4])
+        return m5.toDouble()
+    }
+
     override fun importLibs(compilerOptions: CompilationOptions, compilationTargetName: String): List<String> {
         return if (compilerOptions.launcher == CbmPrgLauncherType.BASIC || compilerOptions.output == OutputType.PRG)
             listOf("syslib")

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

@@ -30,6 +30,26 @@ class VirtualMachineDefinition: IMachineDefinition {
         return parts.joinToString(", ")
     }
 
+    override fun convertFloatToBytes(num: Double): List<UByte> {
+        val bits = num.toBits().toULong()
+        val hexStr = bits.toString(16).padStart(16, '0')
+        val parts = hexStr.chunked(2).map { it.toInt(16).toUByte() }
+        return parts
+    }
+
+    override fun convertBytesToFloat(bytes: List<UByte>): Double {
+        require(bytes.size==8) { "need 8 bytes" }
+        val b0 = bytes[0].toLong() shl (8*7)
+        val b1 = bytes[1].toLong() shl (8*6)
+        val b2 = bytes[2].toLong() shl (8*5)
+        val b3 = bytes[3].toLong() shl (8*4)
+        val b4 = bytes[4].toLong() shl (8*3)
+        val b5 = bytes[5].toLong() shl (8*2)
+        val b6 = bytes[6].toLong() shl (8*1)
+        val b7 = bytes[7].toLong() shl (8*0)
+        return Double.fromBits(b0 or b1 or b2 or b3 or b4 or b5 or b6 or b7)
+    }
+
     override fun importLibs(compilerOptions: CompilationOptions, compilationTargetName: String): List<String> {
         return listOf("syslib")
     }

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

@@ -66,7 +66,7 @@ class AsmGen6502(val prefixSymbols: Boolean): ICodeGeneratorBackend {
                     if(node.type in SplitWordArrayTypes && (lookupName.endsWith("_lsb") || lookupName.endsWith("_msb"))) {
                         lookupName = lookupName.dropLast(4)
                     }
-                    val stNode = st.lookup(lookupName)!!
+                    val stNode = st.lookup(lookupName) ?: throw AssemblyError("unknown identifier $node")
                     if(stNode.astNode.definingBlock()?.options?.noSymbolPrefixing!=true) {
                         val index = node.parent.children.indexOf(node)
                         nodesToPrefix += node.parent to index
@@ -587,7 +587,10 @@ class AsmGen6502Internal (
             is PtInlineAssembly -> translate(stmt)
             is PtBuiltinFunctionCall -> builtinFunctionsAsmGen.translateFunctioncallStatement(stmt)
             is PtFunctionCall -> functioncallAsmGen.translateFunctionCallStatement(stmt)
-            is PtAssignment -> assignmentAsmGen.translate(stmt)
+            is PtAssignment -> {
+                if(stmt.multiTarget) assignmentAsmGen.translateMultiAssign(stmt)
+                else assignmentAsmGen.translate(stmt)
+            }
             is PtAugmentedAssign -> assignmentAsmGen.translate(stmt)
             is PtJump -> {
                 val (asmLabel, indirect) = getJumpTarget(stmt)

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

@@ -65,8 +65,6 @@ internal fun optimizeAssembly(lines: MutableList<String>, machine: IMachineDefin
         numberOfOptimizations++
     }
 
-    // TODO more assembly peephole optimizations
-
     return numberOfOptimizations
 }
 

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

@@ -81,7 +81,7 @@ internal class FunctionCallAsmGen(private val program: PtProgram, private val as
             }
             is PtAddressOf -> false
             is PtIdentifier -> false
-            is PtMachineRegister -> false
+            is PtIrRegister -> false
             is PtMemoryByte -> return usesOtherRegistersWhileEvaluating(arg.address)
             is PtNumber -> false
             is PtBool -> false

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

@@ -70,7 +70,7 @@ internal class IfElseAsmGen(private val program: PtProgram,
             is PtNumber,
             is PtBool,
             is PtIdentifier,
-            is PtMachineRegister,
+            is PtIrRegister,
             is PtArrayIndexer,
             is PtPrefix,
             is PtBinaryExpression -> { /* no cmp necessary the lda has been done just prior */ }
@@ -193,7 +193,7 @@ internal class IfElseAsmGen(private val program: PtProgram,
                     else
                         translateIfElseBodies("beq", stmt)
                 } else {
-                    errors.warn("SLOW FALLBACK FOR 'IF' CODEGEN - ask for support", stmt.position)      // TODO should have no more of these at all
+                    errors.warn("SLOW FALLBACK FOR 'IF' CODEGEN - ask for support", stmt.position)      //  should not occur ;-)
                     assignConditionValueToRegisterAndTest(stmt.condition)
                     if(jumpAfterIf!=null)
                         translateJumpElseBodies("bne", "beq", jumpAfterIf, stmt.elseScope)

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

@@ -635,8 +635,7 @@ internal class ProgramAndVarsGen(
             } else 0
 
         when (variable.dt) {
-            DataType.BOOL -> TODO("bool var to asm")
-            DataType.UBYTE -> asmgen.out("${variable.name}\t.byte  ${initialValue.toHex()}")
+            DataType.BOOL, DataType.UBYTE -> asmgen.out("${variable.name}\t.byte  ${initialValue.toHex()}")
             DataType.BYTE -> asmgen.out("${variable.name}\t.char  $initialValue")
             DataType.UWORD -> asmgen.out("${variable.name}\t.word  ${initialValue.toHex()}")
             DataType.WORD -> asmgen.out("${variable.name}\t.sint  $initialValue")

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

@@ -35,7 +35,7 @@ internal class AnyExprAsmGen(
                 require(expr.left.type in WordDatatypes && expr.right.type in WordDatatypes) {
                     "both operands must be words"
                 }
-                return assignWordBinExpr(expr)
+                throw AssemblyError("expression should have been handled otherwise: word ${expr.operator} at ${expr.position}")
             }
             DataType.FLOAT -> {
                 require(expr.left.type==DataType.FLOAT && expr.right.type==DataType.FLOAT) {
@@ -47,30 +47,6 @@ internal class AnyExprAsmGen(
         }
     }
 
-    private fun assignWordBinExpr(expr: PtBinaryExpression): Boolean {
-        when(expr.operator) {
-            "+" -> TODO("word + at ${expr.position}")
-            "-" -> TODO("word - at ${expr.position}")
-            "*" -> TODO("word * at ${expr.position}")
-            "/" -> TODO("word / at ${expr.position}")
-            "<<" -> TODO("word << at ${expr.position}")
-            ">>" -> TODO("word >> at ${expr.position}")
-            "%" -> TODO("word % at ${expr.position}")
-            "and" -> TODO("word logical and (with optional shortcircuit) ${expr.position}")
-            "or" -> TODO("word logical or (with optional shortcircuit) ${expr.position}")
-            "&" -> TODO("word and at ${expr.position}")
-            "|" -> TODO("word or at ${expr.position}")
-            "^", "xor" -> TODO("word xor at ${expr.position}")
-            "==" -> TODO("word == at ${expr.position}")
-            "!=" -> TODO("word != at ${expr.position}")
-            "<" -> TODO("word < at ${expr.position}")
-            "<=" -> TODO("word <= at ${expr.position}")
-            ">" -> TODO("word > at ${expr.position}")
-            ">=" -> TODO("word >= at ${expr.position}")
-            else -> return false
-        }
-    }
-
     private fun assignByteBinExpr(expr: PtBinaryExpression, assign: AsmAssignment): Boolean {
         when(expr.operator) {
             "+" -> {
@@ -89,21 +65,11 @@ internal class AnyExprAsmGen(
                 asmgen.assignRegister(RegisterOrPair.A, assign.target)
                 return true
             }
-            "*" -> {
-                TODO("byte * at ${expr.position}")
-            }
-            "/" -> {
-                TODO("byte / at ${expr.position}")
-            }
-            "<<" -> {
-                TODO("byte << at ${expr.position}")
-            }
-            ">>" -> {
-                TODO("byte >> at ${expr.position}")
-            }
-            "%" -> {
-                TODO("byte % at ${expr.position}")
-            }
+            "*" -> TODO("byte * at ${expr.position}")
+            "/" -> TODO("byte / at ${expr.position}")
+            "<<" -> TODO("byte << at ${expr.position}")
+            ">>" -> TODO("byte >> at ${expr.position}")
+            "%" -> TODO("byte % at ${expr.position}")
             "and" -> TODO("logical and (with optional shortcircuit) ${expr.position}")
             "or" -> TODO("logical or (with optional shortcircuit) ${expr.position}")
             "&" -> {
@@ -130,24 +96,12 @@ internal class AnyExprAsmGen(
                 asmgen.assignRegister(RegisterOrPair.A, assign.target)
                 return true
             }
-            "==" -> {
-                TODO("byte == at ${expr.position}")
-            }
-            "!=" -> {
-                TODO("byte != at ${expr.position}")
-            }
-            "<" -> {
-                TODO("byte < at ${expr.position}")
-            }
-            "<=" -> {
-                TODO("byte <= at ${expr.position}")
-            }
-            ">" -> {
-                TODO("byte > at ${expr.position}")
-            }
-            ">=" -> {
-                TODO("byte >= at ${expr.position}")
-            }
+            "==" -> TODO("byte == at ${expr.position}")
+            "!=" -> TODO("byte != at ${expr.position}")
+            "<" -> TODO("byte < at ${expr.position}")
+            "<=" -> TODO("byte <= at ${expr.position}")
+            ">" -> TODO("byte > at ${expr.position}")
+            ">=" -> TODO("byte >= at ${expr.position}")
             else -> return false
         }
     }

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

@@ -1,6 +1,8 @@
 package prog8.codegen.cpu6502.assignment
 
 import prog8.code.StMemVar
+import prog8.code.StRomSub
+import prog8.code.StRomSubParameter
 import prog8.code.StStaticVariable
 import prog8.code.ast.*
 import prog8.code.core.*
@@ -32,6 +34,145 @@ internal class AssignmentAsmGen(private val program: PtProgram,
         augmentableAsmGen.translate(assign, augmentedAssign.definingISub())
     }
 
+    fun translateMultiAssign(assignment: PtAssignment) {
+        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
+            ?: throw AssemblyError("only asmsubs can return multiple values")
+
+        require(sub.returns.size>=2)
+        if(sub.returns.any { it.type==DataType.FLOAT })
+            TODO("deal with (multiple?) FP return registers")
+
+        asmgen.translate(values)
+
+        fun needsToSaveA(registersResults: List<Pair<StRomSubParameter, PtNode>>): Boolean =
+            if(registersResults.isEmpty())
+                false
+            else if(registersResults.all { (it.second as PtAssignTarget).identifier!=null})
+                false
+            else
+                true
+
+        fun assignCarryFlagResult(target: PtAssignTarget, saveA: Boolean) {
+            if(saveA) asmgen.out("  pha")
+            asmgen.out("  lda  #0  |  rol  a")
+            val tgt = AsmAssignTarget.fromAstAssignment(target, target.definingISub(), asmgen)
+            assignRegisterByte(tgt, CpuRegister.A, false, false)
+            if(saveA) asmgen.out("  pla")
+        }
+
+        fun assignZeroFlagResult(target: PtAssignTarget, saveA: Boolean) {
+            if(saveA) asmgen.out("  pha")
+            asmgen.out("""
+                beq  +
+                lda  #0
+                beq  ++
++               lda  #1
++""")
+            val tgt = AsmAssignTarget.fromAstAssignment(target, target.definingISub(), asmgen)
+            assignRegisterByte(tgt, CpuRegister.A, false, false)
+            if(saveA) asmgen.out("  pla")
+        }
+
+        fun assignNegativeFlagResult(target: PtAssignTarget, saveA: Boolean) {
+            if(saveA) asmgen.out("  pha")
+            asmgen.out("""
+                bmi  +
+                lda  #0
+                beq  ++
++               lda  #1
++""")
+            val tgt = AsmAssignTarget.fromAstAssignment(target, target.definingISub(), asmgen)
+            assignRegisterByte(tgt, CpuRegister.A, false, false)
+            if(saveA) asmgen.out("  pla")
+        }
+
+        fun assignOverflowFlagResult(target: PtAssignTarget, saveA: Boolean) {
+            if(saveA) asmgen.out("  pha")
+            asmgen.out("""
+                bvs  +
+                lda  #0
+                beq  ++
++               lda  #1
++""")
+            val tgt = AsmAssignTarget.fromAstAssignment(target, target.definingISub(), asmgen)
+            assignRegisterByte(tgt, CpuRegister.A, false, false)
+            if(saveA) asmgen.out("  pla")
+        }
+
+        fun assignRegisterResults(registersResults: List<Pair<StRomSubParameter, PtNode>>) {
+            registersResults.forEach { (returns, target) ->
+                target as PtAssignTarget
+                if(!target.void) {
+                    val targetIdent = target.identifier
+                    val targetMem = target.memory
+                    if(targetIdent!=null || targetMem!=null) {
+                        val tgt = AsmAssignTarget.fromAstAssignment(target, target.definingISub(), asmgen)
+                        when(returns.type) {
+                            in ByteDatatypesWithBoolean -> {
+                                if(returns.register.registerOrPair in Cx16VirtualRegisters) {
+                                    assignVirtualRegister(tgt, returns.register.registerOrPair!!)
+                                } else {
+                                    assignRegisterByte(tgt, returns.register.registerOrPair!!.asCpuRegister(), false, false)
+                                }
+                            }
+                            in WordDatatypes -> {
+                                assignRegisterpairWord(tgt, returns.register.registerOrPair!!)
+                            }
+                            else -> throw AssemblyError("weird dt")
+                        }
+                    }
+                    else TODO("array target for multi-value assignment")        // Not done yet due to result register clobbering complexity
+                }
+            }
+        }
+
+        val assignmentTargets = assignment.children.dropLast(1)
+        if(sub.returns.size==assignmentTargets.size) {
+            // because we can only handle integer results right now we can just zip() it all up
+            val (statusFlagResults, registersResults) = sub.returns.zip(assignmentTargets).partition { it.first.register.statusflag!=null }
+            if(statusFlagResults.isNotEmpty()) {
+                if(statusFlagResults.size>1)
+                    TODO("handle multiple status flag results")
+                val (returns, target) = statusFlagResults.single()
+                target as PtAssignTarget
+                if(target.void) {
+                    // forget about the Carry status flag, only assign the normal return values
+                    assignRegisterResults(registersResults)
+                    return
+                }
+                if(registersResults.all {
+                    val tgt = it.second as PtAssignTarget
+                    tgt.void || tgt.identifier!=null})
+                {
+                    // all other results are just stored into identifiers directly so first handle those
+                    // (simple store instructions that don't modify the carry flag)
+                    assignRegisterResults(registersResults)
+                    return when(returns.register.statusflag!!) {
+                        Statusflag.Pc -> assignCarryFlagResult(target, false)
+                        Statusflag.Pz -> assignZeroFlagResult(target, false)
+                        Statusflag.Pv -> assignOverflowFlagResult(target, false)
+                        Statusflag.Pn -> assignNegativeFlagResult(target, false)
+                    }
+                }
+
+                val saveA = needsToSaveA(registersResults)
+                when(returns.register.statusflag!!) {
+                    Statusflag.Pc -> assignCarryFlagResult(target, saveA)
+                    Statusflag.Pz -> assignZeroFlagResult(target, saveA)
+                    Statusflag.Pv -> assignOverflowFlagResult(target, saveA)
+                    Statusflag.Pn -> assignNegativeFlagResult(target, saveA)
+                }
+            }
+            assignRegisterResults(registersResults)
+        } else {
+            throw AssemblyError("number of values and targets don't match")
+        }
+    }
+
+
     fun translateNormalAssignment(assign: AsmAssignment, scope: IPtSubroutine?) {
         when(assign.source.kind) {
             SourceStorageKind.LITERALBOOLEAN -> {
@@ -597,7 +738,6 @@ internal class AssignmentAsmGen(private val program: PtProgram,
             assignTrue.add(target)
             assignTrue.add(PtNumber.fromBoolean(true, assign.position))
         } else {
-            require(assign.target.datatype in ByteDatatypes)
             when(assign.target.kind) {
                 TargetStorageKind.VARIABLE -> {
                     if(assign.target.datatype in WordDatatypes) {
@@ -618,7 +758,7 @@ internal class AssignmentAsmGen(private val program: PtProgram,
                     }
                 }
                 TargetStorageKind.MEMORY -> {
-                    val tgt = PtAssignTarget(assign.target.position)
+                    val tgt = PtAssignTarget(false, assign.target.position)
                     val targetmem = assign.target.memory!!
                     val mem = PtMemoryByte(targetmem.position)
                     mem.add(targetmem.address)
@@ -628,7 +768,7 @@ internal class AssignmentAsmGen(private val program: PtProgram,
                     assignTrue.add(PtNumber.fromBoolean(true, assign.position))
                 }
                 TargetStorageKind.ARRAY -> {
-                    val tgt = PtAssignTarget(assign.target.position)
+                    val tgt = PtAssignTarget(false, assign.target.position)
                     val targetarray = assign.target.array!!
                     val array = PtArrayIndexer(assign.target.datatype, targetarray.position)
                     array.add(targetarray.variable)
@@ -667,7 +807,7 @@ internal class AssignmentAsmGen(private val program: PtProgram,
     internal fun directIntoY(expr: PtExpression): Boolean {
         return when(expr) {
             is PtIdentifier -> true
-            is PtMachineRegister -> true
+            is PtIrRegister -> true
             is PtNumber -> true
             is PtBuiltinFunctionCall -> expr.name in arrayOf("lsb", "msb")
             else -> false
@@ -1454,6 +1594,7 @@ internal class AssignmentAsmGen(private val program: PtProgram,
                     in ByteDatatypes -> {
                         assignExpressionToRegister(expr.left, RegisterOrPair.A, dt==DataType.BYTE)
                         asmgen.out("""
+                            cmp  #0
                             beq  +
                             lda  #1
 +                           eor  #1""")
@@ -1662,8 +1803,8 @@ internal class AssignmentAsmGen(private val program: PtProgram,
 
             when (valueDt) {
                 DataType.BOOL -> {
-                    if(targetDt in ByteDatatypes) {
-                        // optimization to assign boolean expression to byte target (just assign the 0 or 1 directly, no cast needed)
+                    if(targetDt in IntegerDatatypes) {
+                        // optimization to assign boolean expression to integer target (just assign the 0 or 1 directly)
                         val assignDirect = AsmAssignment(
                             AsmAssignSource.fromAstSource(value, program, asmgen),
                             target,
@@ -1672,7 +1813,7 @@ internal class AssignmentAsmGen(private val program: PtProgram,
                         )
                         assignExpression(assignDirect, target.scope)
                     } else {
-                        throw AssemblyError("expected bool or byte target type")
+                        TODO("assign bool to non-integer type $targetDt")
                     }
                 }
                 in ByteDatatypes -> {

codeGenCpu6502/test/TestCodegen.kt

@@ -53,7 +53,7 @@ class TestCodegen: FunSpec({
         sub.add(PtVariable("xx", DataType.WORD, ZeropageWish.DONTCARE, PtNumber(DataType.WORD, 1.0, Position.DUMMY), null, Position.DUMMY))
 
         val assign = PtAugmentedAssign("+=", Position.DUMMY)
-        val target = PtAssignTarget(Position.DUMMY).also {
+        val target = PtAssignTarget(false, Position.DUMMY).also {
             val targetIdx = PtArrayIndexer(DataType.UBYTE, Position.DUMMY).also { idx ->
                 idx.add(PtIdentifier("main.start.particleX", DataType.ARRAY_UB, Position.DUMMY))
                 idx.add(PtNumber(DataType.UBYTE, 2.0, Position.DUMMY))
@@ -68,7 +68,7 @@ class TestCodegen: FunSpec({
         sub.add(assign)
 
         val prefixAssign = PtAugmentedAssign("-", Position.DUMMY)
-        val prefixTarget = PtAssignTarget(Position.DUMMY).also {
+        val prefixTarget = PtAssignTarget(false, Position.DUMMY).also {
             it.add(PtIdentifier("main.start.xx", DataType.WORD, Position.DUMMY))
         }
         prefixAssign.add(prefixTarget)
@@ -76,7 +76,7 @@ class TestCodegen: FunSpec({
         sub.add(prefixAssign)
 
         val numberAssign = PtAugmentedAssign("-=", Position.DUMMY)
-        val numberAssignTarget = PtAssignTarget(Position.DUMMY).also {
+        val numberAssignTarget = PtAssignTarget(false, Position.DUMMY).also {
             it.add(PtIdentifier("main.start.xx", DataType.WORD, Position.DUMMY))
         }
         numberAssign.add(numberAssignTarget)
@@ -84,7 +84,7 @@ class TestCodegen: FunSpec({
         sub.add(numberAssign)
 
         val cxregAssign = PtAugmentedAssign("+=", Position.DUMMY)
-        val cxregAssignTarget = PtAssignTarget(Position.DUMMY).also {
+        val cxregAssignTarget = PtAssignTarget(false, Position.DUMMY).also {
             it.add(PtIdentifier("main.start.xx", DataType.WORD, Position.DUMMY))
         }
         cxregAssign.add(cxregAssignTarget)

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

@@ -1,5 +1,7 @@
 package prog8.codegen.intermediate
 
+import prog8.code.StRomSub
+import prog8.code.StRomSubParameter
 import prog8.code.ast.*
 import prog8.code.core.*
 import prog8.intermediate.*
@@ -8,16 +10,73 @@ import prog8.intermediate.*
 internal class AssignmentGen(private val codeGen: IRCodeGen, private val expressionEval: ExpressionGen) {
 
     internal fun translate(assignment: PtAssignment): IRCodeChunks {
-        if(assignment.target.children.single() is PtMachineRegister)
-            throw AssemblyError("assigning to a register should be done by just evaluating the expression into resultregister")
+        if(assignment.multiTarget) {
+            val values = assignment.value as? PtFunctionCall
+                ?: throw AssemblyError("only function calls can return multiple values in a multi-assign")
 
-        val chunks = translateRegularAssign(assignment)
-        chunks.filterIsInstance<IRCodeChunk>().firstOrNull()?.appendSrcPosition(assignment.position)
-        return chunks
+            val sub = codeGen.symbolTable.lookup(values.name) as? StRomSub
+                ?: throw AssemblyError("only asmsubs can return multiple values")
+
+            val result = mutableListOf<IRCodeChunkBase>()
+            val funcCall = this.expressionEval.translate(values)
+            require(funcCall.multipleResultRegs.size + funcCall.multipleResultFpRegs.size >= 2)
+            if(funcCall.multipleResultFpRegs.isNotEmpty())
+                TODO("deal with (multiple?) FP return registers")
+            val assignmentTargets = assignment.children.dropLast(1)
+            addToResult(result, funcCall, funcCall.resultReg, funcCall.resultFpReg)
+            if(sub.returns.size==assignmentTargets.size) {
+                // Targets and values match. Assign all the things. Skip 'void' targets.
+                sub.returns.zip(assignmentTargets).zip(funcCall.multipleResultRegs).forEach {
+                    val target = it.first.second as PtAssignTarget
+                    if(!target.void) {
+                        val regNumber = it.second
+                        val returns = it.first.first
+                        result += assignCpuRegister(returns, regNumber, target)
+                    }
+                }
+            } else {
+                throw AssemblyError("number of values and targets don't match")
+            }
+            return result
+        } else {
+            if (assignment.target.children.single() is PtIrRegister)
+                throw AssemblyError("assigning to a register should be done by just evaluating the expression into resultregister")
+
+            return translateRegularAssign(assignment)
+        }
+    }
+
+    private fun assignCpuRegister(returns: StRomSubParameter, regNum: Int, target: PtAssignTarget): IRCodeChunks {
+        val result = mutableListOf<IRCodeChunkBase>()
+        val loadCpuRegInstr = when(returns.register.registerOrPair) {
+            RegisterOrPair.A -> IRInstruction(Opcode.LOADHA, IRDataType.BYTE, reg1=regNum)
+            RegisterOrPair.X -> IRInstruction(Opcode.LOADHX, IRDataType.BYTE, reg1=regNum)
+            RegisterOrPair.Y -> IRInstruction(Opcode.LOADHY, IRDataType.BYTE, reg1=regNum)
+            RegisterOrPair.AX -> IRInstruction(Opcode.LOADHAX, IRDataType.WORD, reg1=regNum)
+            RegisterOrPair.AY -> IRInstruction(Opcode.LOADHAY, IRDataType.WORD, reg1=regNum)
+            RegisterOrPair.XY -> IRInstruction(Opcode.LOADHXY, IRDataType.WORD, reg1=regNum)
+            in Cx16VirtualRegisters -> IRInstruction(Opcode.LOADM, IRDataType.WORD, reg1=regNum, labelSymbol = "cx16.${returns.register.registerOrPair.toString().lowercase()}")
+            null -> {
+                when(returns.register.statusflag) {
+                    Statusflag.Pc -> IRInstruction(Opcode.LOADHA, IRDataType.BYTE, reg1=regNum)
+                    else -> throw AssemblyError("weird statusflag as returnvalue")
+                }
+            }
+            else -> throw AssemblyError("cannot load register")
+        }
+        addInstr(result, loadCpuRegInstr, null)
+
+        // build an assignment to store the value in the actual target.
+        val assign = PtAssignment(target.position)
+        assign.add(target)
+        assign.add(PtIrRegister(regNum, target.type, target.position))
+        result += translate(assign)
+        return result
     }
 
     internal fun translate(augAssign: PtAugmentedAssign): IRCodeChunks {
-        if(augAssign.target.children.single() is PtMachineRegister)
+        // augmented assignment always has just a single target
+        if(augAssign.target.children.single() is PtIrRegister)
             throw AssemblyError("assigning to a register should be done by just evaluating the expression into resultregister")
 
         val target = augAssign.target
@@ -270,11 +329,11 @@ internal class AssignmentGen(private val codeGen: IRCodeGen, private val express
                     else
                         throw AssemblyError("assignment value and target dt mismatch")
                 } else false
-                if (assignment.value is PtMachineRegister) {
-                    valueRegister = (assignment.value as PtMachineRegister).register
+                if (assignment.value is PtIrRegister) {
+                    valueRegister = (assignment.value as PtIrRegister).register
                     if(extendByteToWord) {
                         valueRegister = codeGen.registers.nextFree()
-                        addInstr(result, IRInstruction(Opcode.EXT, IRDataType.BYTE, reg1=valueRegister, reg2=(assignment.value as PtMachineRegister).register), null)
+                        addInstr(result, IRInstruction(Opcode.EXT, IRDataType.BYTE, reg1=valueRegister, reg2=(assignment.value as PtIrRegister).register), null)
                     }
                 } else {
                     val tr = expressionEval.translateExpression(assignment.value)

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

@@ -854,10 +854,10 @@ internal class BuiltinFuncGen(private val codeGen: IRCodeGen, private val exprGe
 
     private fun assignRegisterTo(target: PtExpression, register: Int): IRCodeChunks {
         val assignment = PtAssignment(target.position)
-        val assignTarget = PtAssignTarget(target.position)
+        val assignTarget = PtAssignTarget(false, target.position)
         assignTarget.children.add(target)
         assignment.children.add(assignTarget)
-        assignment.children.add(PtMachineRegister(register, target.type, target.position))
+        assignment.children.add(PtIrRegister(register, target.type, target.position))
         val result = mutableListOf<IRCodeChunkBase>()
         result += codeGen.translateNode(assignment)
         return result

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

@@ -7,8 +7,11 @@ import prog8.code.core.*
 import prog8.intermediate.*
 
 
-internal class ExpressionCodeResult(val chunks: IRCodeChunks, val dt: IRDataType, val resultReg: Int, val resultFpReg: Int) {
-    constructor(chunks: IRCodeChunk, dt: IRDataType, resultReg: Int, resultFpReg: Int) : this(listOf(chunks), dt, resultReg, resultFpReg)
+internal class ExpressionCodeResult(val chunks: IRCodeChunks, val dt: IRDataType, val resultReg: Int, val resultFpReg: Int,
+                                    val multipleResultRegs: List<Int> = emptyList(), val multipleResultFpRegs: List<Int> = emptyList()
+) {
+    constructor(chunk: IRCodeChunk, dt: IRDataType, resultReg: Int, resultFpReg: Int, multipleResultRegs: List<Int> = emptyList(), multipleResultFpRegs: List<Int> = emptyList())
+            : this(listOf(chunk), dt, resultReg, resultFpReg, multipleResultRegs, multipleResultFpRegs)
 
     companion object {
         val EMPTY: ExpressionCodeResult = ExpressionCodeResult(emptyList(), IRDataType.BYTE, -1, -1)
@@ -24,7 +27,7 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
 
     fun translateExpression(expr: PtExpression): ExpressionCodeResult {
         return when (expr) {
-            is PtMachineRegister -> {
+            is PtIrRegister -> {
                 ExpressionCodeResult(emptyList(), irType(expr.type), expr.register, -1)
             }
             is PtBool -> {
@@ -463,7 +466,7 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
                         argRegisters.add(FunctionCallArgs.ArgumentSpec(parameter.name, null, FunctionCallArgs.RegSpec(paramDt, tr.resultReg, null)))
                     result += tr.chunks
                 }
-                // return value
+                // return value (always singular for normal Subs)
                 val returnRegSpec = if(fcall.void) null else {
                     val returnIrType = irType(callTarget.returnType!!)
                     if(returnIrType==IRDataType.FLOAT)
@@ -472,7 +475,8 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
                         FunctionCallArgs.RegSpec(returnIrType, codeGen.registers.nextFree(), null)
                 }
                 // create the call
-                addInstr(result, IRInstruction(Opcode.CALL, labelSymbol = fcall.name, fcallArgs = FunctionCallArgs(argRegisters, returnRegSpec)), null)
+                addInstr(result, IRInstruction(Opcode.CALL, labelSymbol = fcall.name,
+                    fcallArgs = FunctionCallArgs(argRegisters, if(returnRegSpec==null) emptyList() else listOf(returnRegSpec))), null)
                 return if(fcall.void)
                     ExpressionCodeResult(result, IRDataType.BYTE, -1, -1)
                 else if(fcall.type==DataType.FLOAT)
@@ -493,44 +497,57 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
                     else
                         argRegisters.add(FunctionCallArgs.ArgumentSpec("", null, FunctionCallArgs.RegSpec(paramDt, tr.resultReg, parameter.register)))
                     result += tr.chunks
+                    when(parameter.register.registerOrPair) {
+                        RegisterOrPair.A -> addInstr(result, IRInstruction(Opcode.STOREHA, IRDataType.BYTE, reg1=tr.resultReg), null)
+                        RegisterOrPair.X -> addInstr(result, IRInstruction(Opcode.STOREHX, IRDataType.BYTE, reg1=tr.resultReg), null)
+                        RegisterOrPair.Y -> addInstr(result, IRInstruction(Opcode.STOREHY, IRDataType.BYTE, reg1=tr.resultReg), null)
+                        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)
+                        in Cx16VirtualRegisters -> {
+                            addInstr(result, IRInstruction(Opcode.STOREM, paramDt, reg1=tr.resultReg, labelSymbol = "cx16.${parameter.register.registerOrPair.toString().lowercase()}"), null)
+                        }
+                        null -> when(parameter.register.statusflag) {
+                            // TODO: do the statusflag argument as last
+                            Statusflag.Pc -> addInstr(result, IRInstruction(Opcode.LSR, paramDt, reg1=tr.resultReg), null)
+                            else -> throw AssemblyError("weird statusflag as param")
+                        }
+                        else -> throw AssemblyError("unsupported register arg")
+                    }
                 }
-                // return value
-                var statusFlagResult: Statusflag? = null
+
+                if(callTarget.returns.size>1)
+                    return callRomSubWithMultipleReturnValues(callTarget, fcall, argRegisters, result)
+
+                // return a single value (or nothing)
                 val returnRegSpec = if(fcall.void) null else {
                     if(callTarget.returns.isEmpty())
                         null
-                    else if(callTarget.returns.size==1) {
+                    else {
                         val returns = callTarget.returns[0]
                         val returnIrType = irType(returns.type)
-                        if(returnIrType==IRDataType.FLOAT)
+                        if (returnIrType == IRDataType.FLOAT)
                             FunctionCallArgs.RegSpec(returnIrType, codeGen.registers.nextFreeFloat(), returns.register)
                         else {
-                            statusFlagResult = returns.register.statusflag
-                            val returnRegister = if(statusFlagResult==null) codeGen.registers.nextFree() else -1
+                            val returnRegister = codeGen.registers.nextFree()
                             FunctionCallArgs.RegSpec(returnIrType, returnRegister, returns.register)
                         }
-                    } else {
-                        // multiple return values: take the first *register* (not status flag) return value and ignore the rest.
-                        val returns = callTarget.returns.first { it.register.registerOrPair!=null }
-                        val returnIrType = irType(returns.type)
-                        if(returnIrType==IRDataType.FLOAT)
-                            FunctionCallArgs.RegSpec(returnIrType, codeGen.registers.nextFreeFloat(), returns.register)
-                        else
-                            FunctionCallArgs.RegSpec(returnIrType, codeGen.registers.nextFree(), returns.register)
                     }
                 }
                 // create the call
+                val returnRegs = if(returnRegSpec==null) emptyList() else listOf(returnRegSpec)
                 val call =
                     if(callTarget.address==null)
-                        IRInstruction(Opcode.CALL, labelSymbol = fcall.name, fcallArgs = FunctionCallArgs(argRegisters, returnRegSpec))
+                        IRInstruction(Opcode.CALL, labelSymbol = fcall.name, fcallArgs = FunctionCallArgs(argRegisters, returnRegs))
                     else
-                        IRInstruction(Opcode.CALL, address = callTarget.address!!.toInt(), fcallArgs = FunctionCallArgs(argRegisters, returnRegSpec))
+                        IRInstruction(Opcode.CALL, address = callTarget.address!!.toInt(), fcallArgs = FunctionCallArgs(argRegisters, returnRegs))
                 addInstr(result, call, null)
                 var finalReturnRegister = returnRegSpec?.registerNum ?: -1
 
                 if(fcall.parent is PtAssignment || fcall.parent is PtTypeCast) {
                     // look if the status flag bit should actually be returned as a 0/1 byte value in a result register (so it can be assigned)
-                    if(statusFlagResult!=null && returnRegSpec!=null) {
+                    val statusFlagResult = returnRegSpec?.cpuRegister?.statusflag
+                    if(statusFlagResult!=null) {
                         // assign status flag bit to the return value register
                         finalReturnRegister = returnRegSpec.registerNum
                         if(finalReturnRegister<0)
@@ -573,6 +590,29 @@ internal class ExpressionGen(private val codeGen: IRCodeGen) {
         }
     }
 
+    private fun callRomSubWithMultipleReturnValues(
+        callTarget: StRomSub,
+        fcall: PtFunctionCall,
+        argRegisters: MutableList<FunctionCallArgs.ArgumentSpec>,
+        result: MutableList<IRCodeChunkBase>
+    ): ExpressionCodeResult {
+        // return multiple values
+        val returnRegisters = callTarget.returns.map {
+            val regnum = if(it.type==DataType.FLOAT) codeGen.registers.nextFreeFloat() else codeGen.registers.nextFree()
+            FunctionCallArgs.RegSpec(irType(it.type), regnum, it.register)
+        }
+        // create the call
+        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))
+        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}
+        return ExpressionCodeResult(result, IRDataType.BYTE, -1, -1, resultRegs, resultFpRegs)
+    }
+
     private fun operatorGreaterThan(
         binExpr: PtBinaryExpression,
         vmDt: IRDataType,

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

@@ -1451,6 +1451,10 @@ class IRCodeGen(
         }
 
         when(val cond=ifElse.condition) {
+            is PtBool -> {
+                // normally this will be optimized away, but not with -noopt
+                translateSimple(cond, Opcode.BSTEQ)
+            }
             is PtTypeCast -> {
                 require(cond.type==DataType.BOOL && cond.value.type in NumericDatatypes)
                 translateSimple(cond, Opcode.BSTEQ)
@@ -1650,7 +1654,8 @@ class IRCodeGen(
             val args = params.map { (dt, reg)->
                 FunctionCallArgs.ArgumentSpec("", null, FunctionCallArgs.RegSpec(dt, reg, null))
             }
-            val returnSpec = if(returns==null) null else FunctionCallArgs.RegSpec(returns.first, returns.second, null)
+            // for now, syscalls have 0 or 1 return value
+            val returnSpec = if(returns==null) emptyList() else listOf(FunctionCallArgs.RegSpec(returns.first, returns.second, null))
             it += IRInstruction(Opcode.SYSCALL, immediate = syscall.number, fcallArgs = FunctionCallArgs(args, returnSpec))
         }
     }

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

@@ -52,15 +52,12 @@ class IRPeepholeOptimizer(private val irprog: IRProgram) {
                                 || removeWeirdBranches(chunk1, chunk2, indexedInstructions)
                                 || removeDoubleSecClc(chunk1, indexedInstructions)
                                 || cleanupPushPop(chunk1, indexedInstructions)
-                        // TODO other optimizations
                     } while (changed)
                 }
             }
             removeEmptyChunks(sub)
         }
 
-        // TODO also do register optimization step here at the end?
-
         irprog.linkChunks()  // re-link
     }
 
@@ -433,21 +430,34 @@ class IRPeepholeOptimizer(private val irprog: IRProgram) {
     }
 
     private fun removeDoubleLoadsAndStores(chunk: IRCodeChunk, indexedInstructions: List<IndexedValue<IRInstruction>>): Boolean {
-        return false
-
-/*
         var changed = false
         indexedInstructions.forEach { (idx, ins) ->
+            if(ins.opcode==Opcode.STOREM) {
+                val prev = indexedInstructions[idx-1].value
+                if(prev.opcode==Opcode.LOADM) {
+                    // loadm.X rX,something | storem.X rX,something ?? -> get rid of the store.
+                    if(ins.labelSymbol!=null && ins.labelSymbol==prev.labelSymbol && ins.labelSymbolOffset==prev.labelSymbolOffset) {
+                        changed=true
+                        chunk.instructions.removeAt(idx)
+                    }
+                    else if(ins.address!=null && ins.address==prev.address) {
+                        changed=true
+                        chunk.instructions.removeAt(idx)
+                    }
+                }
+            }
 
-            // TODO: detect multiple loads to the same target registers, only keep first (if source is not I/O memory)
-            // TODO: detect multiple stores to the same target, only keep first (if target is not I/O memory)
-            // TODO: detect multiple float ffrom/fto to the same target, only keep first
-            // TODO: detect subsequent same xors/nots/negs, remove the pairs completely as they cancel out
-            // TODO: detect multiple same ands, ors; only keep first
-            // TODO: (hard) detect multiple registers being assigned the same value (and not changed) - use only 1 of them
+/*
+    Possible other optimizations:
+            // detect multiple loads to the same target registers, only keep first (if source is not I/O memory)
+            // detect multiple stores to the same target, only keep first (if target is not I/O memory)
+            // detect multiple float ffrom/fto to the same target, only keep first
+            // detect subsequent same xors/nots/negs, remove the pairs completely as they cancel out
+            // detect multiple same ands, ors; only keep first
+            // detect multiple registers being assigned the same value (and not changed) - use only 1 of them  (hard!)
             // ...
+*/
         }
         return changed
-*/
     }
 }

codeGenIntermediate/test/TestVmCodeGen.kt

@@ -50,7 +50,7 @@ class TestVmCodeGen: FunSpec({
         sub.add(PtVariable("xx", DataType.WORD, ZeropageWish.DONTCARE, PtNumber(DataType.WORD, 1.0, Position.DUMMY), null, Position.DUMMY))
 
         val assign = PtAugmentedAssign("+=", Position.DUMMY)
-        val target = PtAssignTarget(Position.DUMMY).also {
+        val target = PtAssignTarget(false, Position.DUMMY).also {
             val targetIdx = PtArrayIndexer(DataType.UBYTE, Position.DUMMY).also { idx ->
                 idx.add(PtIdentifier("main.start.particleX", DataType.ARRAY_UB, Position.DUMMY))
                 idx.add(PtNumber(DataType.UBYTE, 2.0, Position.DUMMY))
@@ -65,7 +65,7 @@ class TestVmCodeGen: FunSpec({
         sub.add(assign)
 
         val prefixAssign = PtAugmentedAssign("-", Position.DUMMY)
-        val prefixTarget = PtAssignTarget(Position.DUMMY).also {
+        val prefixTarget = PtAssignTarget(false, Position.DUMMY).also {
             it.add(PtIdentifier("main.start.xx", DataType.WORD, Position.DUMMY))
         }
         prefixAssign.add(prefixTarget)
@@ -73,7 +73,7 @@ class TestVmCodeGen: FunSpec({
         sub.add(prefixAssign)
 
         val numberAssign = PtAugmentedAssign("+=", Position.DUMMY)
-        val numberAssignTarget = PtAssignTarget(Position.DUMMY).also {
+        val numberAssignTarget = PtAssignTarget(false, Position.DUMMY).also {
             it.add(PtIdentifier("main.start.xx", DataType.WORD, Position.DUMMY))
         }
         numberAssign.add(numberAssignTarget)
@@ -81,7 +81,7 @@ class TestVmCodeGen: FunSpec({
         sub.add(numberAssign)
 
         val cxregAssign = PtAugmentedAssign("+=", Position.DUMMY)
-        val cxregAssignTarget = PtAssignTarget(Position.DUMMY).also {
+        val cxregAssignTarget = PtAssignTarget(false, Position.DUMMY).also {
             it.add(PtIdentifier("main.start.xx", DataType.WORD, Position.DUMMY))
         }
         cxregAssign.add(cxregAssignTarget)

codeOptimizers/src/prog8/optimizer/ConstantFoldingOptimizer.kt

@@ -435,7 +435,6 @@ class ConstantFoldingOptimizer(private val program: Program, private val errors:
        subrightIsConst: Boolean): IAstModification?
     {
         // NOTE: THESE REORDERINGS ARE ONLY VALID FOR FLOATING POINT CONSTANTS
-        // TODO: this implements only a small set of possible reorderings at this time, we could perhaps add more
 
         if(expr.operator==subExpr.operator) {
             // both operators are the same.

codeOptimizers/src/prog8/optimizer/ConstantIdentifierReplacer.kt

@@ -75,10 +75,11 @@ class VarConstantValueTypeAdjuster(
                     val declValue = decl.value?.constValue(program)
                     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 is never written to and was replaced by a constant", decl.position)
+                        errors.info("variable '${decl.name}' is never written to and was replaced by a constant", decl.position)
                         val const = VarDecl(VarDeclType.CONST, decl.origin, decl.datatype, decl.zeropage, decl.arraysize, decl.name, decl.names, declValue, decl.sharedWithAsm, decl.splitArray, decl.position)
+                        decl.value = null
                         return listOf(
-                            IAstModification.ReplaceNode(decl, const, parent),
+                            IAstModification.ReplaceNode(decl, const, parent)
                         )
                     }
                 }
@@ -94,7 +95,7 @@ class VarConstantValueTypeAdjuster(
                         )
                     }
                     // variable only has a single write and it is the initialization value, so it can be replaced with a constant, IF the value is a constant
-                    errors.info("variable is never written to and was replaced by a constant", decl.position)
+                    errors.info("variable '${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)
                     return listOf(
                         IAstModification.ReplaceNode(decl, const, parent),
@@ -103,15 +104,15 @@ class VarConstantValueTypeAdjuster(
                 }
             }
             /*
-        TODO: need to check if there are no variable usages between the declaration and the assignment (because these rely on the original initialization value)
-        if(writes.size==2) {
-            val firstAssignment = writes[0].parent as? Assignment
-            val secondAssignment = writes[1].parent as? Assignment
-            if(firstAssignment?.origin==AssignmentOrigin.VARINIT && secondAssignment?.value?.constValue(program)!=null) {
-                errors.warn("variable is only assigned once here, consider using this as the initialization value in the declaration instead", secondAssignment.position)
+            TODO: need to check if there are no variable usages between the declaration and the assignment (because these rely on the original initialization value)
+            if(writes.size==2) {
+                val firstAssignment = writes[0].parent as? Assignment
+                val secondAssignment = writes[1].parent as? Assignment
+                if(firstAssignment?.origin==AssignmentOrigin.VARINIT && secondAssignment?.value?.constValue(program)!=null) {
+                    errors.warn("variable is only assigned once here, consider using this as the initialization value in the declaration instead", secondAssignment.position)
+                }
             }
-        }
-        */
+            */
         }
 
         return noModifications
@@ -309,7 +310,7 @@ internal class ConstantIdentifierReplacer(
                 val add = BinaryExpression(NumericLiteral(cval.type, cval.number, identifier.position), "+", arrayIdx.indexer.indexExpr, identifier.position)
                 return if(arrayIdx.parent is AssignTarget) {
                     val memwrite = DirectMemoryWrite(add, identifier.position)
-                    val assignTarget = AssignTarget(null, null, memwrite, identifier.position)
+                    val assignTarget = AssignTarget(null, null, memwrite, null, false, identifier.position)
                     listOf(IAstModification.ReplaceNode(arrayIdx.parent, assignTarget, arrayIdx.parent.parent))
                 } else {
                     val memread = DirectMemoryRead(add, identifier.position)

codeOptimizers/src/prog8/optimizer/ExpressionSimplifier.kt

@@ -4,7 +4,6 @@ import prog8.ast.*
 import prog8.ast.base.FatalAstException
 import prog8.ast.expressions.*
 import prog8.ast.statements.AnonymousScope
-import prog8.ast.statements.Assignment
 import prog8.ast.statements.IfElse
 import prog8.ast.statements.Jump
 import prog8.ast.walk.AstWalker
@@ -14,8 +13,6 @@ import kotlin.math.abs
 import kotlin.math.log2
 import kotlin.math.pow
 
-// TODO add more peephole expression optimizations? Investigate what optimizations binaryen has?
-
 class ExpressionSimplifier(private val program: Program, private val options: CompilationOptions, private val errors: IErrorReporter) : AstWalker() {
     private val powersOfTwo = (1..16).map { (2.0).pow(it) }.toSet()
     private val negativePowersOfTwo = powersOfTwo.map { -it }.toSet()
@@ -109,12 +106,6 @@ class ExpressionSimplifier(private val program: Program, private val options: Co
         if (!leftIDt.isKnown || !rightIDt.isKnown)
             throw FatalAstException("can't determine datatype of both expression operands $expr")
 
-        // NonBinaryExpression  <associativeoperator>  BinaryExpression  -->  BinaryExpression  <associativeoperator>  NonBinaryExpression
-        if (expr.operator in AssociativeOperators && expr.left !is BinaryExpression && expr.right is BinaryExpression) {
-            if(parent !is Assignment || !(expr.left isSameAs parent.target) && maySwapOperandOrder(expr))
-                return listOf(IAstModification.SwapOperands(expr))
-        }
-
         // X + (-A)  -->  X - A
         if (expr.operator == "+" && (expr.right as? PrefixExpression)?.operator == "-") {
             return listOf(IAstModification.ReplaceNode(
@@ -249,10 +240,13 @@ class ExpressionSimplifier(private val program: Program, private val options: Co
         // optimize boolean constant comparisons
         if(expr.operator=="==") {
             if(rightDt==DataType.BOOL && leftDt==DataType.BOOL) {
-                if(rightVal?.asBooleanValue==true)
-                    return listOf(IAstModification.ReplaceNode(expr, expr.left, parent))
-                else
-                    return listOf(IAstModification.ReplaceNode(expr, PrefixExpression("not", expr.left, expr.position), parent))
+                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 (rightVal?.number == 1.0) {
                 if (options.strictBool) {
@@ -747,7 +741,7 @@ class ExpressionSimplifier(private val program: Program, private val options: Co
                 in powersOfTwo -> {
                     if (leftDt==DataType.UBYTE || leftDt==DataType.UWORD) {
                         // Unsigned number divided by a power of two => shift right
-                        // Signed number can't simply be bitshifted in this case (due to rounding issues for negative values),  TODO is this correct???
+                        // Signed number can't simply be bitshifted in this case (due to rounding issues for negative values),
                         // so we leave that as is and let the code generator deal with it.
                         val numshifts = log2(cv).toInt()
                         return BinaryExpression(expr.left, ">>", NumericLiteral.optimalInteger(numshifts, expr.position), expr.position)

codeOptimizers/src/prog8/optimizer/StatementOptimizer.kt

@@ -162,7 +162,7 @@ class StatementOptimizer(private val program: Program,
                 // for loop over a (constant) range of just a single value-- optimize the loop away
                 // loopvar/reg = range value , follow by block
                 val scope = AnonymousScope(mutableListOf(), forLoop.position)
-                scope.statements.add(Assignment(AssignTarget(forLoop.loopVar, null, null, forLoop.position), range.from, AssignmentOrigin.OPTIMIZER, forLoop.position))
+                scope.statements.add(Assignment(AssignTarget(forLoop.loopVar, null, null, null, false, forLoop.position), range.from, AssignmentOrigin.OPTIMIZER, forLoop.position))
                 scope.statements.addAll(forLoop.body.statements)
                 return listOf(IAstModification.ReplaceNode(forLoop, scope, parent))
             }
@@ -177,7 +177,7 @@ class StatementOptimizer(private val program: Program,
                     val character = options.compTarget.encodeString(sv.value, sv.encoding)[0]
                     val byte = NumericLiteral(DataType.UBYTE, character.toDouble(), iterable.position)
                     val scope = AnonymousScope(mutableListOf(), forLoop.position)
-                    scope.statements.add(Assignment(AssignTarget(forLoop.loopVar, null, null, forLoop.position), byte, AssignmentOrigin.OPTIMIZER, forLoop.position))
+                    scope.statements.add(Assignment(AssignTarget(forLoop.loopVar, null, null, null, false, forLoop.position), byte, AssignmentOrigin.OPTIMIZER, forLoop.position))
                     scope.statements.addAll(forLoop.body.statements)
                     return listOf(IAstModification.ReplaceNode(forLoop, scope, parent))
                 }
@@ -190,7 +190,7 @@ class StatementOptimizer(private val program: Program,
                     if(av!=null) {
                         val scope = AnonymousScope(mutableListOf(), forLoop.position)
                         scope.statements.add(Assignment(
-                                AssignTarget(forLoop.loopVar, null, null, forLoop.position), NumericLiteral.optimalInteger(av.toInt(), iterable.position),
+                                AssignTarget(forLoop.loopVar, null, null, null, false, forLoop.position), NumericLiteral.optimalInteger(av.toInt(), iterable.position),
                                 AssignmentOrigin.OPTIMIZER, forLoop.position))
                         scope.statements.addAll(forLoop.body.statements)
                         return listOf(IAstModification.ReplaceNode(forLoop, scope, parent))
@@ -206,7 +206,7 @@ class StatementOptimizer(private val program: Program,
                 val pos = forLoop.position
                 val loopVar = forLoop.loopVar
                 val addSubOne = BinaryExpression(loopVar.copy(), if(inc) "+" else "-", NumericLiteral.optimalInteger(1, pos), pos, false)
-                return Assignment(AssignTarget(loopVar.copy(), null, null, pos), addSubOne, AssignmentOrigin.USERCODE, pos)
+                return Assignment(AssignTarget(loopVar.copy(), null, null, null, false, pos), addSubOne, AssignmentOrigin.USERCODE, pos)
             }
 
             if (range != null && range.to.constValue(program)?.number == 0.0 && range.step.constValue(program)?.number==-1.0) {
@@ -219,7 +219,7 @@ class StatementOptimizer(private val program: Program,
                     val decOne = incOrDec(false)
                     forLoop.body.statements.add(decOne)
                     val replacement = AnonymousScope(mutableListOf(
-                        Assignment(AssignTarget(forLoop.loopVar.copy(), null, null, pos),
+                        Assignment(AssignTarget(forLoop.loopVar.copy(), null, null, null, false, pos),
                             fromExpr, AssignmentOrigin.OPTIMIZER, pos),
                         UntilLoop(forLoop.body, condition, pos)
                     ), pos)
@@ -454,12 +454,6 @@ class StatementOptimizer(private val program: Program,
 
     override fun after(whenStmt: When, parent: Node): Iterable<IAstModification> {
 
-        fun replaceWithIf(condition: Expression, trueBlock: AnonymousScope, elseBlock: AnonymousScope?): List<IAstModification> {
-            val ifStmt = IfElse(condition, trueBlock, elseBlock ?: AnonymousScope(mutableListOf(), whenStmt.position), whenStmt.position)
-            errors.info("for boolean condition a normal if statement is preferred", whenStmt.position)
-            return listOf(IAstModification.ReplaceNode(whenStmt, ifStmt, parent))
-        }
-
         val constantValue = whenStmt.condition.constValue(program)?.number
         if(constantValue!=null) {
             // when condition is a constant

codeOptimizers/src/prog8/optimizer/UnusedCodeRemover.kt

@@ -69,14 +69,14 @@ class UnusedCodeRemover(private val program: Program,
     override fun after(block: Block, parent: Node): Iterable<IAstModification> {
         if("force_output" !in block.options()) {
             if (block.containsNoCodeNorVars) {
-                if(block.name != internedStringsModuleName) {
+                if(block.name != internedStringsModuleName && "ignore_unused" !in block.options()) {
                     if(!block.statements.any { it is Subroutine && it.hasBeenInlined })
                         errors.info("removing unused block '${block.name}'", block.position)
                 }
                 return listOf(IAstModification.Remove(block, parent as IStatementContainer))
             }
             if(callgraph.unused(block)) {
-                if(block.statements.any{ it !is VarDecl || it.type== VarDeclType.VAR}) {
+                if(block.statements.any{ it !is VarDecl || it.type== VarDeclType.VAR} && "ignore_unused" !in block.options()) {
                     if(!block.statements.any { it is Subroutine && it.hasBeenInlined })
                         errors.info("removing unused block '${block.name}'", block.position)
                 }

compiler/res/prog8lib/atari/textio.p8

@@ -3,7 +3,6 @@
 
 %import syslib
 %import conv
-%import shared_textio_functions
 
 txt {
 
@@ -17,6 +16,10 @@ sub  clear_screen() {
     chrout(125)
 }
 
+sub  cls() {
+    chrout(125)
+}
+
 sub nl() {
     chrout('\n')
 }
@@ -148,15 +151,15 @@ _tmp_outchar
 }
 
 asmsub  print (str text @ AY) clobbers(A,Y)  {
-	; ---- print null terminated string from A/Y
+	; ---- print zero terminated string from A/Y
 	; note: the compiler contains an optimization that will replace
 	;       a call to this subroutine with a string argument of just one char,
 	;       by just one call to CHROUT of that single char.
 	%asm {{
-		sta  P8ZP_SCRATCH_B1
-		sty  P8ZP_SCRATCH_REG
+		sta  P8ZP_SCRATCH_W2
+		sty  P8ZP_SCRATCH_W2+1
 		ldy  #0
--		lda  (P8ZP_SCRATCH_B1),y
+-		lda  (P8ZP_SCRATCH_W2),y
 		beq  +
 		jsr  chrout
 		iny
@@ -168,13 +171,13 @@ asmsub  print (str text @ AY) clobbers(A,Y)  {
 asmsub  print_ub0  (ubyte value @ A) clobbers(A,X,Y)  {
 	; ---- print the ubyte in A in decimal form, with left padding 0s (3 positions total)
 	%asm {{
-		jsr  conv.ubyte2decimal
+		jsr  conv.internal_ubyte2decimal
 		pha
 		tya
 		jsr  chrout
-		pla
-		jsr  chrout
 		txa
+		jsr  chrout
+		pla
 		jmp  chrout
 	}}
 }
@@ -182,21 +185,21 @@ asmsub  print_ub0  (ubyte value @ A) clobbers(A,X,Y)  {
 asmsub  print_ub  (ubyte value @ A) clobbers(A,X,Y)  {
 	; ---- print the ubyte in A in decimal form, without left padding 0s
 	%asm {{
-		jsr  conv.ubyte2decimal
+		jsr  conv.internal_ubyte2decimal
 _print_byte_digits
 		pha
 		cpy  #'0'
 		beq  +
 		tya
 		jsr  chrout
-		pla
+		txa
 		jsr  chrout
 		jmp  _ones
-+       pla
-        cmp  #'0'
++       cpx  #'0'
         beq  _ones
+        txa
         jsr  chrout
-_ones   txa
+_ones   pla
 		jmp  chrout
 	}}
 }
@@ -210,7 +213,7 @@ asmsub  print_b  (byte value @ A) clobbers(A,X,Y)  {
 		lda  #'-'
 		jsr  chrout
 +		pla
-		jsr  conv.byte2decimal
+		jsr  conv.internal_byte2decimal
 		jmp  print_ub._print_byte_digits
 	}}
 }
@@ -223,7 +226,7 @@ asmsub  print_ubhex  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
 		lda  #'$'
 		jsr  chrout
 		pla
-+		jsr  conv.ubyte2hex
++		jsr  conv.internal_ubyte2hex
 		jsr  chrout
 		tya
 		jmp  chrout
@@ -277,9 +280,9 @@ asmsub  print_uwhex  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
 asmsub  print_uw0  (uword value @ AY) clobbers(A,X,Y)  {
 	; ---- print the uword in A/Y in decimal form, with left padding 0s (5 positions total)
 	%asm {{
-		jsr  conv.uword2decimal
+		jsr  conv.internal_uword2decimal
 		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
+-		lda  conv.internal_uword2decimal.decTenThousands,y
         beq  +
 		jsr  chrout
 		iny
@@ -291,9 +294,9 @@ asmsub  print_uw0  (uword value @ AY) clobbers(A,X,Y)  {
 asmsub  print_uw  (uword value @ AY) clobbers(A,X,Y)  {
 	; ---- print the uword in A/Y in decimal form, without left padding 0s
 	%asm {{
-		jsr  conv.uword2decimal
+		jsr  conv.internal_uword2decimal
 		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
+-		lda  conv.internal_uword2decimal.decTenThousands,y
 		beq  _allzero
 		cmp  #'0'
 		bne  _gotdigit
@@ -303,7 +306,7 @@ asmsub  print_uw  (uword value @ AY) clobbers(A,X,Y)  {
 _gotdigit
 		jsr  chrout
 		iny
-		lda  conv.uword2decimal.decTenThousands,y
+		lda  conv.internal_uword2decimal.decTenThousands,y
 		bne  _gotdigit
 		rts
 _allzero

compiler/res/prog8lib/c128/syslib.p8

@@ -93,11 +93,12 @@ romsub $FFD5 = LOAD(ubyte verify @ A, uword address @ XY) -> bool @Pc, ubyte @ A
 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)
-romsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A    ; (via 810 ($32A)) get a character
+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                 ; read number of screen rows and columns
+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
 
@@ -105,15 +106,20 @@ romsub $FFF3 = IOBASE() -> uword @ XY                           ; read base addr
 
 ; ---- utilities -----
 
-asmsub STOP2() clobbers(X) -> bool @A  {
-    ; -- check if STOP key was pressed, returns true if so.  More convenient to use than STOP() because that only sets the carry status flag.
+inline asmsub STOP2() clobbers(X,A) -> bool @Pz  {
+    ; -- just like STOP, but omits the special keys result value in A.
+    ;    just for convenience because most of the times you're only interested in the stop pressed or not status.
     %asm {{
         jsr  cbm.STOP
-        beq  +
-        lda  #0
-        rts
-+       lda  #1
-        rts
+    }}
+}
+
+inline asmsub GETIN2() clobbers(X,Y) -> ubyte @A {
+    ; -- just like GETIN, but omits the carry flag result value.
+    ;    just for convenience because GETIN is so often used to just read keyboard input,
+    ;    where you don't have to deal with a potential error status
+    %asm {{
+        jsr  cbm.GETIN
     }}
 }
 

compiler/res/prog8lib/c128/textio.p8

@@ -4,7 +4,7 @@
 
 %import syslib
 %import conv
-%import shared_textio_functions
+%import shared_cbm_textio_functions
 
 
 txt {
@@ -13,11 +13,17 @@ 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.
+
 
 sub  clear_screen() {
     chrout(147)
 }
 
+sub  cls() {
+    chrout(147)
+}
+
 sub home() {
     chrout(19)
 }
@@ -288,216 +294,6 @@ _scroll_screen  ; scroll only the screen memory
 	}}
 }
 
-romsub $FFD2 = chrout(ubyte char @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
-
-asmsub  print (str text @ AY) clobbers(A,Y)  {
-	; ---- print null terminated string, in PETSCII encoding, from A/Y
-	; note: the compiler contains an optimization that will replace
-	;       a call to this subroutine with a string argument of just one char,
-	;       by just one call to cbm.CHROUT of that single char.
-	%asm {{
-		sta  P8ZP_SCRATCH_B1
-		sty  P8ZP_SCRATCH_REG
-		ldy  #0
--		lda  (P8ZP_SCRATCH_B1),y
-		beq  +
-		jsr  cbm.CHROUT
-		iny
-		bne  -
-+		rts
-	}}
-}
-
-asmsub  print_ub0  (ubyte value @ A) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in decimal form, with left padding 0s (3 positions total)
-	%asm {{
-		jsr  conv.ubyte2decimal
-		pha
-		tya
-		jsr  cbm.CHROUT
-		pla
-		jsr  cbm.CHROUT
-		txa
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_ub  (ubyte value @ A) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in decimal form, without left padding 0s
-	%asm {{
-		jsr  conv.ubyte2decimal
-_print_byte_digits
-		pha
-		cpy  #'0'
-		beq  +
-		tya
-		jsr  cbm.CHROUT
-		pla
-		jsr  cbm.CHROUT
-		jmp  _ones
-+       pla
-        cmp  #'0'
-        beq  _ones
-        jsr  cbm.CHROUT
-_ones   txa
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_b  (byte value @ A) clobbers(A,X,Y)  {
-	; ---- print the byte in A in decimal form, without left padding 0s
-	%asm {{
-		pha
-		cmp  #0
-		bpl  +
-		lda  #'-'
-		jsr  cbm.CHROUT
-+		pla
-		jsr  conv.byte2decimal
-		jmp  print_ub._print_byte_digits
-	}}
-}
-
-asmsub  print_ubhex  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in hex form (if Carry is set, a radix prefix '$' is printed as well)
-	%asm {{
-		bcc  +
-		pha
-		lda  #'$'
-		jsr  cbm.CHROUT
-		pla
-+		jsr  conv.ubyte2hex
-		jsr  cbm.CHROUT
-		tya
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_ubbin  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in binary form (if Carry is set, a radix prefix '%' is printed as well)
-	%asm {{
-		sta  P8ZP_SCRATCH_B1
-		bcc  +
-		lda  #'%'
-		jsr  cbm.CHROUT
-+		ldy  #8
--		lda  #'0'
-		asl  P8ZP_SCRATCH_B1
-		bcc  +
-		lda  #'1'
-+		jsr  cbm.CHROUT
-		dey
-		bne  -
-		rts
-	}}
-}
-
-asmsub  print_uwbin  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in binary form (if Carry is set, a radix prefix '%' is printed as well)
-	%asm {{
-		pha
-		tya
-		jsr  print_ubbin
-		pla
-		clc
-		jmp  print_ubbin
-	}}
-}
-
-asmsub  print_uwhex  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in hexadecimal form (4 digits)
-	;      (if Carry is set, a radix prefix '$' is printed as well)
-	%asm {{
-		pha
-		tya
-		jsr  print_ubhex
-		pla
-		clc
-		jmp  print_ubhex
-	}}
-}
-
-asmsub  print_uw0  (uword value @ AY) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in decimal form, with left padding 0s (5 positions total)
-	%asm {{
-		jsr  conv.uword2decimal
-		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
-        beq  +
-		jsr  cbm.CHROUT
-		iny
-		bne  -
-+		rts
-	}}
-}
-
-asmsub  print_uw  (uword value @ AY) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in decimal form, without left padding 0s
-	%asm {{
-		jsr  conv.uword2decimal
-		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
-		beq  _allzero
-		cmp  #'0'
-		bne  _gotdigit
-		iny
-		bne  -
-
-_gotdigit
-		jsr  cbm.CHROUT
-		iny
-		lda  conv.uword2decimal.decTenThousands,y
-		bne  _gotdigit
-		rts
-_allzero
-        lda  #'0'
-        jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_w  (word value @ AY) clobbers(A,X,Y)  {
-	; ---- print the (signed) word in A/Y in decimal form, without left padding 0's
-	%asm {{
-		cpy  #0
-		bpl  +
-		pha
-		lda  #'-'
-		jsr  cbm.CHROUT
-		tya
-		eor  #255
-		tay
-		pla
-		eor  #255
-		clc
-		adc  #1
-		bcc  +
-		iny
-+		jmp  print_uw
-	}}
-}
-
-asmsub  input_chars  (uword buffer @ AY) clobbers(A) -> ubyte @ Y  {
-	; ---- Input a string (max. 80 chars) from the keyboard, in PETSCII encoding.
-	;      Returns length in Y. (string is terminated with a 0 byte as well)
-	;      It assumes the keyboard is selected as I/O channel!
-
-	%asm {{
-		sta  P8ZP_SCRATCH_W1
-		sty  P8ZP_SCRATCH_W1+1
-		ldy  #0				; char counter = 0
--		jsr  cbm.CHRIN
-		cmp  #$0d			; return (ascii 13) pressed?
-		beq  +				; yes, end.
-		sta  (P8ZP_SCRATCH_W1),y	; else store char in buffer
-		iny
-		bne  -
-+		lda  #0
-		sta  (P8ZP_SCRATCH_W1),y	; finish string with 0 byte
-		rts
-
-	}}
-}
-
 asmsub  setchr  (ubyte col @X, ubyte row @Y, ubyte character @A) clobbers(A, Y)  {
 	; ---- sets the character in the screen matrix at the given position
 	%asm {{

compiler/res/prog8lib/c64/floats.p8

@@ -4,11 +4,7 @@
 %import shared_floats_functions
 
 floats {
-	; ---- this block contains C-64 floating point related functions ----
-        const float  π      = 3.141592653589793
-        const float  PI     = π
-        const float  TWOPI  = 2*π
-
+; ---- this block contains C-64 floating point related functions ----
 
 ; ---- C64 basic and kernal ROM float constants and functions ----
 
@@ -24,7 +20,7 @@ romsub $bba6 = FREADMEM() clobbers(A,Y)                     ; load mflpt value f
 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)
+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
 
@@ -49,7 +45,7 @@ romsub $b7b5 = FREADSTR(ubyte length @ A) clobbers(A,X,Y)   ; str -> fac1, $22/2
 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)
 
-romsub $b849 = FADDH() clobbers(A,X,Y)                      ; fac1 += 0.5, for rounding- call this before INT
+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
@@ -67,7 +63,7 @@ romsub $bf78 = FPWR(uword mflpt @ AY) clobbers(A,X,Y)       ; fac1 = fac2 ** mfl
 romsub $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 round instead of trunc
+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

compiler/res/prog8lib/c64/syslib.p8

@@ -94,23 +94,29 @@ romsub $FFD5 = LOAD(ubyte verify @ A, uword address @ XY) -> bool @Pc, ubyte @ A
 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)
-romsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A    ; (via 810 ($32A)) get a character
+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                 ; read number of screen rows and columns
+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
 
-asmsub STOP2() clobbers(X) -> bool @A  {
-    ; -- check if STOP key was pressed, returns true if so.  More convenient to use than STOP() because that only sets the carry status flag.
+
+inline asmsub STOP2() clobbers(X,A) -> bool @Pz  {
+    ; -- just like STOP, but omits the special keys result value in A.
+    ;    just for convenience because most of the times you're only interested in the stop pressed or not status.
     %asm {{
         jsr  cbm.STOP
-        beq  +
-        lda  #0
-        rts
-+       lda  #1
-        rts
+    }}
+}
+
+inline asmsub GETIN2() clobbers(X,Y) -> ubyte @A {
+    ; -- just like GETIN, but omits the carry flag result value.
+    ;    just for convenience because GETIN is so often used to just read keyboard input,
+    ;    where you don't have to deal with a potential error status
+    %asm {{
+        jsr  cbm.GETIN
     }}
 }
 

compiler/res/prog8lib/c64/textio.p8

@@ -4,7 +4,7 @@
 
 %import syslib
 %import conv
-%import shared_textio_functions
+%import shared_cbm_textio_functions
 
 txt {
 
@@ -13,11 +13,16 @@ 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.
 
 sub  clear_screen() {
     chrout(147)
 }
 
+sub  cls() {
+    chrout(147)
+}
+
 sub home() {
     chrout(19)
 }
@@ -292,216 +297,6 @@ _scroll_screen  ; scroll only the screen memory
 	}}
 }
 
-romsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
-
-asmsub  print (str text @ AY) clobbers(A,Y)  {
-	; ---- print null terminated string, in PESCII encoding, from A/Y
-	; note: the compiler contains an optimization that will replace
-	;       a call to this subroutine with a string argument of just one char,
-	;       by just one call to cbm.CHROUT of that single char.
-	%asm {{
-		sta  P8ZP_SCRATCH_B1
-		sty  P8ZP_SCRATCH_REG
-		ldy  #0
--		lda  (P8ZP_SCRATCH_B1),y
-		beq  +
-		jsr  cbm.CHROUT
-		iny
-		bne  -
-+		rts
-	}}
-}
-
-asmsub  print_ub0  (ubyte value @ A) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in decimal form, with left padding 0s (3 positions total)
-	%asm {{
-		jsr  conv.ubyte2decimal
-		pha
-		tya
-		jsr  cbm.CHROUT
-		pla
-		jsr  cbm.CHROUT
-		txa
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_ub  (ubyte value @ A) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in decimal form, without left padding 0s
-	%asm {{
-		jsr  conv.ubyte2decimal
-_print_byte_digits
-		pha
-		cpy  #'0'
-		beq  +
-		tya
-		jsr  cbm.CHROUT
-		pla
-		jsr  cbm.CHROUT
-		jmp  _ones
-+       pla
-        cmp  #'0'
-        beq  _ones
-        jsr  cbm.CHROUT
-_ones   txa
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_b  (byte value @ A) clobbers(A,X,Y)  {
-	; ---- print the byte in A in decimal form, without left padding 0s
-	%asm {{
-		pha
-		cmp  #0
-		bpl  +
-		lda  #'-'
-		jsr  cbm.CHROUT
-+		pla
-		jsr  conv.byte2decimal
-		jmp  print_ub._print_byte_digits
-	}}
-}
-
-asmsub  print_ubhex  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in hex form (if Carry is set, a radix prefix '$' is printed as well)
-	%asm {{
-		bcc  +
-		pha
-		lda  #'$'
-		jsr  cbm.CHROUT
-		pla
-+		jsr  conv.ubyte2hex
-		jsr  cbm.CHROUT
-		tya
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_ubbin  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in binary form (if Carry is set, a radix prefix '%' is printed as well)
-	%asm {{
-		sta  P8ZP_SCRATCH_B1
-		bcc  +
-		lda  #'%'
-		jsr  cbm.CHROUT
-+		ldy  #8
--		lda  #'0'
-		asl  P8ZP_SCRATCH_B1
-		bcc  +
-		lda  #'1'
-+		jsr  cbm.CHROUT
-		dey
-		bne  -
-		rts
-	}}
-}
-
-asmsub  print_uwbin  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in binary form (if Carry is set, a radix prefix '%' is printed as well)
-	%asm {{
-		pha
-		tya
-		jsr  print_ubbin
-		pla
-		clc
-		jmp  print_ubbin
-	}}
-}
-
-asmsub  print_uwhex  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in hexadecimal form (4 digits)
-	;      (if Carry is set, a radix prefix '$' is printed as well)
-	%asm {{
-		pha
-		tya
-		jsr  print_ubhex
-		pla
-		clc
-		jmp  print_ubhex
-	}}
-}
-
-asmsub  print_uw0  (uword value @ AY) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in decimal form, with left padding 0s (5 positions total)
-	%asm {{
-		jsr  conv.uword2decimal
-		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
-        beq  +
-		jsr  cbm.CHROUT
-		iny
-		bne  -
-+		rts
-	}}
-}
-
-asmsub  print_uw  (uword value @ AY) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in decimal form, without left padding 0s
-	%asm {{
-		jsr  conv.uword2decimal
-		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
-		beq  _allzero
-		cmp  #'0'
-		bne  _gotdigit
-		iny
-		bne  -
-
-_gotdigit
-		jsr  cbm.CHROUT
-		iny
-		lda  conv.uword2decimal.decTenThousands,y
-		bne  _gotdigit
-		rts
-_allzero
-        lda  #'0'
-        jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_w  (word value @ AY) clobbers(A,X,Y)  {
-	; ---- print the (signed) word in A/Y in decimal form, without left padding 0's
-	%asm {{
-		cpy  #0
-		bpl  +
-		pha
-		lda  #'-'
-		jsr  cbm.CHROUT
-		tya
-		eor  #255
-		tay
-		pla
-		eor  #255
-		clc
-		adc  #1
-		bcc  +
-		iny
-+		jmp  print_uw
-	}}
-}
-
-asmsub  input_chars  (uword buffer @ AY) clobbers(A) -> ubyte @ Y  {
-	; ---- Input a string (max. 80 chars) from the keyboard, in PETSCII encoding.
-	;      Returns length in Y. (string is terminated with a 0 byte as well)
-	;      It assumes the keyboard is selected as I/O channel!
-
-	%asm {{
-		sta  P8ZP_SCRATCH_W1
-		sty  P8ZP_SCRATCH_W1+1
-		ldy  #0				; char counter = 0
--		jsr  cbm.CHRIN
-		cmp  #$0d			; return (ascii 13) pressed?
-		beq  +				; yes, end.
-		sta  (P8ZP_SCRATCH_W1),y	; else store char in buffer
-		iny
-		bne  -
-+		lda  #0
-		sta  (P8ZP_SCRATCH_W1),y	; finish string with 0 byte
-		rts
-
-	}}
-}
-
 asmsub  setchr  (ubyte col @X, ubyte row @Y, ubyte character @A) clobbers(A, Y)  {
 	; ---- sets the character in the screen matrix at the given position
 	%asm {{

compiler/res/prog8lib/conv.p8

@@ -8,97 +8,79 @@ conv {
 
     str  @shared string_out = "????????????????"       ; result buffer for the string conversion routines
 
-asmsub  str_ub0  (ubyte value @ A) clobbers(X) -> str @AY  {
-	; ---- convert the ubyte in A in decimal string form, with left padding 0s (3 positions total)
-	%asm {{
-        jsr  conv.ubyte2decimal
-        sty  string_out
-        sta  string_out+1
-        stx  string_out+2
-        lda  #0
-        sta  string_out+3
-        lda  #<string_out
-        ldy  #>string_out
-        rts
-	}}
-}
+    asmsub str_ub0(ubyte value @A) clobbers(X) -> str @AY {
+        ; ---- convert the ubyte in A in decimal string form, with left padding 0s (3 positions total)
+        %asm {{
+            jsr  internal_ubyte2decimal
+            sty  conv.string_out
+            stx  conv.string_out+1
+            sta  conv.string_out+2
+            lda  #0
+            sta  conv.string_out+3
+            lda  #<conv.string_out
+            ldy  #>conv.string_out
+            rts
+        }}
+    }
 
-asmsub  str_ub  (ubyte value @ A) clobbers(X) -> str @AY  {
-	; ---- convert the ubyte in A in decimal string form, without left padding 0s
-	%asm {{
-		ldy  #0
-		sty  P8ZP_SCRATCH_B1
-		jsr  ubyte2decimal     ; result in Y/A/X (100s, 10s, 1s).
-        ; hundreds?
-		cpy  #'0'
-		beq  +
-		sty  string_out
-		sta  string_out+1
-		stx  string_out+2
-		lda  #0
-		sta  string_out+3
-		jmp  _done
-		; tens?
-+		cmp  #'0'
-		beq  +
-		sta  string_out
-		stx  string_out+1
-		lda  #0
-		sta  string_out+2
-		jmp  _done
-+       ; ones.
-        stx  string_out
-		lda  #0
-		sta  string_out+1
-_done   lda  #<string_out
-        ldy  #>string_out
-        rts
-	}}
-}
+    asmsub str_ub(ubyte value @A) clobbers(X) -> str @AY {
+        ; ---- convert the ubyte in A in decimal string form, without left padding 0s
+        %asm {{
+            jsr  internal_ubyte2decimal
+            cpy  #'0'
+            beq  +
+            sty  conv.string_out
+            stx  conv.string_out+1
+            sta  conv.string_out+2
+            lda  #0
+            sta  conv.string_out+3
+            jmp  _done
++           cpx  #'0'
+            beq  +
+            stx  conv.string_out
+            sta  conv.string_out+1
+            lda  #0
+            sta  conv.string_out+2
+            jmp  _done
++           sta  conv.string_out
+            lda  #0
+            sta  conv.string_out+1
+_done       lda  #<conv.string_out
+            ldy  #>conv.string_out
+            rts
+        }}
+    }
 
-asmsub  str_b  (byte value @ A) clobbers(X) -> str @AY  {
-	; ---- convert the byte in A in decimal string form, without left padding 0s
-	%asm {{
-	    ldy  #0
-        cmp  #0
-        bpl  +
-        ldy  #'-'
-        sty  string_out
-        ldy  #1
-+       sty  P8ZP_SCRATCH_REG
-   	    jsr  conv.byte2decimal      ; result in Y/A/X (100s, 10s, 1s).  and in uword2decimal.decHundreds, decTens, decOnes.
-        ; hundreds?
-		cpy  #'0'
-		bne  _out_hundreds
-		ldy  P8ZP_SCRATCH_REG
-        cmp  #'0'
-        bne  _out_tens
-        beq  _out_ones
-_out_hundreds
-		ldy  P8ZP_SCRATCH_REG
-		lda  uword2decimal.decHundreds
-		sta  string_out,y
-		iny
-_out_tens
-		lda  uword2decimal.decTens
-		sta  string_out,y
-		iny
-_out_ones
-		lda  uword2decimal.decOnes
-        sta  string_out,y
-        iny
-        lda  #0
-        sta  string_out,y
-        lda  #<string_out
-        ldy  #>string_out
-   	    rts
-	}}
-}
+    asmsub str_b(byte value @A) clobbers(X) -> str @AY {
+        ; ---- convert the byte in A in decimal string form, without left padding 0s
+        %asm {{
+            cmp  #0
+            bpl  str_ub
+            eor  #255
+            clc
+            adc  #1
+            jsr  str_ub
+            ; insert a minus sign at the start
+            lda  #0
+            sta  conv.string_out+4
+            lda  conv.string_out+2
+            sta  conv.string_out+3
+            lda  conv.string_out+1
+            sta  conv.string_out+2
+            lda  conv.string_out
+            sta  conv.string_out+1
+            lda  #'-'
+            sta  conv.string_out
+            lda  #<conv.string_out
+            ldy  #>conv.string_out
+            rts
+        }}
+    }
 
 asmsub  str_ubhex  (ubyte value @ A) clobbers(X) -> str @AY {
 	; ---- convert the ubyte in A in hex string form
 	%asm {{
-        jsr  conv.ubyte2hex
+        jsr  internal_ubyte2hex
         sta  string_out
         sty  string_out+1
         lda  #0
@@ -158,11 +140,11 @@ asmsub  str_uwhex  (uword value @ AY) -> str @AY  {
 	%asm {{
         pha
         tya
-        jsr  conv.ubyte2hex
+        jsr  internal_ubyte2hex
         sta  string_out
         sty  string_out+1
         pla
-        jsr  conv.ubyte2hex
+        jsr  internal_ubyte2hex
         sta  string_out+2
         sty  string_out+3
         lda  #0
@@ -176,9 +158,9 @@ asmsub  str_uwhex  (uword value @ AY) -> str @AY  {
 asmsub  str_uw0  (uword value @ AY) clobbers(X) -> str @AY  {
 	; ---- convert the uword in A/Y in decimal string form, with left padding 0s (5 positions total)
 	%asm {{
-	    jsr  conv.uword2decimal
+	    jsr  conv.internal_uword2decimal
 	    ldy  #0
--       lda  conv.uword2decimal.decTenThousands,y
+-       lda  conv.internal_uword2decimal.decTenThousands,y
         sta  string_out,y
         beq  +
         iny
@@ -193,11 +175,11 @@ asmsub  str_uw0  (uword value @ AY) clobbers(X) -> str @AY  {
 asmsub  str_uw  (uword value @ AY) clobbers(X) -> str @AY  {
 	; ---- convert the uword in A/Y in decimal string form, without left padding 0s
 	%asm {{
-	    jsr  conv.uword2decimal
+	    jsr  conv.internal_uword2decimal
 	    ldx  #0
 _output_digits
 	    ldy  #0
--       lda  conv.uword2decimal.decTenThousands,y
+-       lda  internal_uword2decimal.decTenThousands,y
         beq  _allzero
         cmp  #'0'
         bne  _gotdigit
@@ -206,7 +188,7 @@ _output_digits
 _gotdigit   sta  string_out,x
         inx
         iny
-        lda  conv.uword2decimal.decTenThousands,y
+        lda  internal_uword2decimal.decTenThousands,y
         bne  _gotdigit
 _end    lda  #0
         sta  string_out,x
@@ -238,7 +220,7 @@ asmsub  str_w  (word value @ AY) clobbers(X) -> str @AY  {
         adc  #1
         bcc  +
         iny
-+	    jsr  conv.uword2decimal
++	    jsr  conv.internal_uword2decimal
 	    ldx  #1
 	    bne  str_uw._output_digits
 	    rts
@@ -533,17 +515,25 @@ _stop
 
 ; ----- low level number conversions to decimal strings ----
 
-asmsub  ubyte2decimal  (ubyte value @A) -> ubyte @Y, ubyte @A, ubyte @X  {
-	; ---- A to decimal string in Y/A/X  (100s in Y, 10s in A, 1s in X)
-	%asm {{
-		ldy  #uword2decimal.ASCII_0_OFFSET
-		jmp  uword2decimal.hex_try200
-	}}
+asmsub internal_ubyte2decimal(ubyte value @A) -> ubyte @Y, ubyte @X, ubyte @A {
+    %asm {{
+        ldy #'0'-1
+        ldx #'9'+1
+        sec
+-       iny
+        sbc #100
+        bcs -
+-       dex
+        adc #10
+        bmi -
+        adc #'0'-1
+        rts
+    }}
 }
 
-asmsub  uword2decimal  (uword value @AY) -> ubyte @Y, ubyte @A, ubyte @X  {
+asmsub  internal_uword2decimal  (uword value @AY) -> ubyte @Y, ubyte @A, ubyte @X  {
 	;  ---- convert 16 bit uword in A/Y to decimal
-	;  output in uword2decimal.decTenThousands, decThousands, decHundreds, decTens, decOnes
+	;  output in internal_uword2decimal.decTenThousands, decThousands, decHundreds, decTens, decOnes
 	;  (these are terminated by a zero byte so they can be easily printed)
 	;  also returns Y = 100's, A = 10's, X = 1's
 
@@ -716,7 +706,7 @@ decOnes   		.byte  0
     }}
 }
 
-asmsub  byte2decimal  (byte value @A) -> ubyte @Y, ubyte @A, ubyte @X  {
+asmsub  internal_byte2decimal  (byte value @A) -> ubyte @Y, ubyte @A, ubyte @X  {
 	; ---- A (signed byte) to decimal string in Y/A/X  (100s in Y, 10s in A, 1s in X)
 	;      note: if the number is negative, you have to deal with the '-' yourself!
 	%asm {{
@@ -725,11 +715,11 @@ asmsub  byte2decimal  (byte value @A) -> ubyte @Y, ubyte @A, ubyte @X  {
 		eor  #255
 		clc
 		adc  #1
-+		jmp  ubyte2decimal
++		jmp  internal_ubyte2decimal
 	}}
 }
 
-asmsub  ubyte2hex  (ubyte value @A) clobbers(X) -> ubyte @A, ubyte @Y  {
+asmsub  internal_ubyte2hex  (ubyte value @A) clobbers(X) -> ubyte @A, ubyte @Y  {
 	; ---- A to hex petscii string in AY (first hex char in A, second hex char in Y)
 	%asm {{
 		pha
@@ -749,7 +739,7 @@ _hex_digits	.text "0123456789abcdef"	; can probably be reused for other stuff as
 	}}
 }
 
-asmsub  uword2hex  (uword value @AY) clobbers(A,Y)  {
+asmsub  internal_uword2hex  (uword value @AY) clobbers(A,Y)  {
 	; ---- convert 16 bit uword in A/Y into 4-character hexadecimal string 'uword2hex.output' (0-terminated)
 	%asm {{
 		sta  P8ZP_SCRATCH_REG

compiler/res/prog8lib/cx16/bmx.p8

@@ -170,6 +170,7 @@ save_end:
     sub set_bpp(ubyte bpp) {
         ubyte[8] depths = [0,1,1,2,2,2,2,3]
         vera_colordepth = depths[bpp-1]
+        bitsperpixel = bpp
     }
 
     sub set_vera_colordepth(ubyte depth) {
@@ -234,7 +235,7 @@ save_end:
 
     sub read_scanline(uword size) {
         while size!=0 {
-            cx16.r0 = cx16.MACPTR(min(255, size) as ubyte, &cx16.VERA_DATA0, true)
+            void, cx16.r0 = cx16.MACPTR(min(255, size) as ubyte, &cx16.VERA_DATA0, true)
             if_cs {
                 ; no MACPTR support
                 repeat size
@@ -296,7 +297,7 @@ save_end:
                 cx16.r0L = lsb(size)
                 if msb(size)!=0
                     cx16.r0L = 0        ; 256 bytes
-                cx16.r0 = cx16.MCIOUT(cx16.r0L, &cx16.VERA_DATA0, true)
+                void, cx16.r0 = cx16.MCIOUT(cx16.r0L, &cx16.VERA_DATA0, true)
                 if_cs {
                     ; no MCIOUT support
                     repeat size

compiler/res/prog8lib/cx16/diskio.p8

@@ -32,6 +32,19 @@ diskio {
         cbm.CHKOUT(WRITE_IO_CHANNEL)
     }
 
+    sub fastmode(ubyte mode) -> bool {
+        ; -- set fast serial mode (0=none, 1=auto_tx, 2=fast writes, 3=both) for the SD card.
+        ;    Returns success status (fails on emulator host fs for example)
+        list_filename[0] = 'u'
+        list_filename[1] = '0'
+        list_filename[2] = '>'
+        list_filename[3] = 'b'
+        list_filename[4] = mode | $30
+        list_filename[5] = 0
+        send_command(list_filename)
+        return status_code()==0
+    }
+
     sub directory() -> bool {
         ; -- Prints the directory contents to the screen. Returns success.
 
@@ -70,7 +83,8 @@ diskio {
             void cbm.CHRIN()     ; skip 2 bytes
             void cbm.CHRIN()
             status = cbm.READST()
-            if cbm.STOP2()
+            void cbm.STOP()
+            if_z
                 break
         }
         status = cbm.READST()
@@ -319,7 +333,7 @@ close_end:
             readsize = 255
             if num_bytes<readsize
                 readsize = num_bytes
-            readsize = cx16.MACPTR(lsb(readsize), bufferpointer, false)     ; fast block reads
+            void, readsize = cx16.MACPTR(lsb(readsize), bufferpointer, false)     ; fast block reads
             if_cs
                 goto byte_read_loop     ; MACPTR block read not supported, do fallback loop
             list_blocks += readsize
@@ -456,7 +470,7 @@ _end        rts
         if num_bytes!=0 {
             reset_write_channel()
             do {
-                cx16.r0 = cx16.MCIOUT(lsb(num_bytes), bufferpointer, false)     ; fast block writes
+                void, cx16.r0 = cx16.MCIOUT(lsb(num_bytes), bufferpointer, false)     ; fast block writes
                 if_cs
                     goto no_mciout
                 num_bytes -= cx16.r0
@@ -525,6 +539,39 @@ io_error:
         goto done
     }
 
+    ; similar to above, but instead of fetching the entire string, it only fetches the status code and returns it as ubyte
+    ; in case of IO error, returns 255 (CMDR-DOS itself is physically unable to return such a value)
+    sub status_code() -> ubyte {
+        if cbm.READST()==128 {
+            return 255
+        }
+
+        cbm.SETNAM(0, list_filename)
+        cbm.SETLFS(15, drivenumber, 15)
+        void cbm.OPEN()          ; open 15,8,15
+        if_cs
+            goto io_error
+        void cbm.CHKIN(15)        ; use #15 as input channel
+
+        list_filename[0] = cbm.CHRIN()
+        list_filename[1] = cbm.CHRIN()
+        list_filename[2] = 0
+
+        while cbm.READST()==0 {
+            void cbm.CHRIN()
+        }
+
+        cbm.CLRCHN()        ; restore default i/o devices
+        cbm.CLOSE(15)
+        return conv.str2ubyte(list_filename)
+
+io_error:
+        cbm.CLRCHN()
+        cbm.CLOSE(15)
+        return 255
+    }
+
+
 
     ; saves a block of memory to disk, including the default 2 byte prg header.
     sub save(uword filenameptr, uword startaddress, uword savesize) -> bool {

compiler/res/prog8lib/cx16/emudbg.p8

@@ -1,5 +1,5 @@
 ; Emulator debug interface.
-; Docs: https://github.com/X16Community/x16-emulator#emulator-io-registers
+; Docs: https://github.com/X16Community/x16-emulator/tree/d52f118ce893fa24c4ca021a0b8de46cb80e5ccf#emulator-io-registers
 
 emudbg {
     %option ignore_unused
@@ -41,7 +41,7 @@ emudbg {
                 chr = @(isoString)
                 if_z
                     break
-                EMU_CPUCLK_U = chr
+                EMU_CHROUT = chr
                 isoString++
             }
         }

compiler/res/prog8lib/cx16/floats.p8

@@ -4,12 +4,7 @@
 %import shared_floats_functions
 
 floats {
-	; ---- this block contains C-64 compatible floating point related functions ----
-
-        const float  π      = 3.141592653589793
-        const float  PI     = π
-        const float  TWOPI  = 2*π
-
+; ---- this block contains C-64 compatible floating point related functions ----
 
 ; ---- ROM float functions (same as on C128 except base page) ----
 
@@ -19,7 +14,7 @@ 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 100-101 ($64-$65) MSB FIRST. (might throw ILLEGAL QUANTITY)
+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.
 
 ; 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
@@ -43,7 +38,7 @@ 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 ROUND or FADDH first to round instead of trunc
+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
@@ -53,7 +48,7 @@ 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 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
@@ -65,10 +60,10 @@ romsub $fe60 = MOVFRM(uword mflpt @ AY) clobbers(A,X,Y)     ; load mflpt value f
 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)
+romsub $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 rounding- call this before INT
+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!)
@@ -98,8 +93,8 @@ asmsub  FREADSA  (byte value @A) 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  $c4        ; facmo     ($64 on c128)
-	sta  $c5        ; facmo+1   ($65 on c128)
+	sty  $c4        ; facmo
+	sta  $c5        ; facmo+1
 	ldx  #$90
 	sec
 	jmp  FLOATC
@@ -136,14 +131,12 @@ asmsub  FREADUY (ubyte value @Y) {
 }
 
 asmsub parse(str value @AY) -> float @FAC1 {
-    ; -- parse a string value of a number to float in FAC1
-    ;    warning: on older <R47 kernals it uses an internal BASIC routine that is ROM version dependent,
-    ;    ($deb6 is inside the routine for VAL at $deb3)  See basic.sym from x16-rom
-    ;    TODO once ROM v47 is released, all the workarounds here can be removed. But probably keep the kernal VAL_1 existance check
+    ; -- Parse a string value of a number to float in FAC1.
+    ;    Requires kernal R47 or newer! (depends on val_1 working)
     %asm {{
         ldx  VAL_1
         cpx  #$4c       ; is there an implementation in VAL_1? (test for JMP)
-        bne  +          ; no, do it ourselves
+        bne  _borked    ; no, print error message
         pha             ; yes, count the length and call rom VAL_1.
         phy
         jsr  prog8_lib.strlen
@@ -151,15 +144,7 @@ asmsub parse(str value @AY) -> float @FAC1 {
         ply
         plx
         jmp  VAL_1
-+       sta  $a9    ; 'index' variable
-        sty  $aa
-        jsr  prog8_lib.strlen
-        lda  $deb6
-        cmp  #$d0   ; sanity check for kernal routine correct
-        bne  +
-        tya
-        jmp  $deb6   ; kernal version dependent...
-+       ; print error message if routine is borked in kernal, and exit program
+_borked
         ldy  #0
 -       lda  _msg,y
         beq  +
@@ -168,12 +153,12 @@ asmsub parse(str value @AY) -> float @FAC1 {
         bne  -
 +       jmp  sys.exit
 
-_msg    .text 13,"?val kaputt",13,0
+_msg    .text 13,"?rom 47+ required for val1",13,0
     }}
 }
 
 
-&uword AYINT_facmo = $c6      ; $c6/$c7 contain result of AYINT
+&uword AYINT_facmo = $c6      ; $c6/$c7 contain result of AYINT   (See "basic.sym" kernal symbol file)
 
 sub rnd() -> float {
     %asm {{

compiler/res/prog8lib/cx16/gfx2.p8

@@ -55,9 +55,11 @@ gfx2 {
             }
             else -> {
                 ; back to default 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
+                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
+                }
             }
         }
 

compiler/res/prog8lib/cx16/psg.p8

@@ -48,7 +48,7 @@ psg {
     sub freq(ubyte voice_num, uword vera_freq) {
         ; -- Changes the frequency of the voice's sound.
         ;    voice_num = 0-15,  vera_freq = 0-65535  calculate this via the formula given in the Vera's PSG documentation.
-        ;    (https://github.com/x16community/x16-docs/blob/master/VERA%20Programmer's%20Reference.md)
+        ;    (https://github.com/X16Community/x16-docs/blob/101759f3bfa5e6cce4e8c5a0b67cb0f2f1c6341e/X16%20Reference%20-%2009%20-%20VERA%20Programmer's%20Reference.md)
         ;    Write freq MSB first and then LSB to reduce the chance on clicks
         sys.irqsafe_set_irqd()
         cx16.r0 = $f9c1 + voice_num * 4

compiler/res/prog8lib/cx16/sprites.p8

@@ -5,6 +5,7 @@
 
 ; note: sprites z-order will be in front of all layers.
 ; note: collision mask is not supported here yet.
+; note: "palette offset" is counted as 0-15  (vera multiplies the offset by 16 to get at the actual color index)
 
 sprites {
     %option ignore_unused
@@ -21,7 +22,7 @@ sprites {
     sub init(ubyte spritenum,
              ubyte databank, uword dataaddr,
              ubyte width_flag, ubyte height_flag,
-             ubyte colors_flag, ubyte palette_offset_idx) {
+             ubyte colors_flag, ubyte palette_offset) {
         hide(spritenum)
         cx16.VERA_DC_VIDEO |= %01000000             ; enable sprites globally
         dataaddr >>= 5
@@ -30,7 +31,7 @@ sprites {
         cx16.vpoke(1, sprite_reg, lsb(dataaddr))                    ; address 12:5
         cx16.vpoke(1, sprite_reg+1, colors_flag | msb(dataaddr))    ; 4 bpp + address 16:13
         cx16.vpoke(1, sprite_reg+6, %00001100)                      ; z depth %11 = in front of both layers, no flips
-        cx16.vpoke(1, sprite_reg+7, height_flag<<6 | width_flag<<4 | palette_offset_idx>>4) ; 64x64 pixels, palette offset
+        cx16.vpoke(1, sprite_reg+7, height_flag<<6 | width_flag<<4 | palette_offset&15) ; 64x64 pixels, palette offset
     }
 
     sub data(ubyte spritenum, ubyte bank, uword addr) {
@@ -143,7 +144,7 @@ sprites {
     }
 
     sub set_palette_offset(ubyte spritenum, ubyte offset) {
-        cx16.vpoke_mask(1, VERA_SPRITEREGS + 7 + spritenum*$0008, %11110000, offset>>4)
+        cx16.vpoke_mask(1, VERA_SPRITEREGS + 7 + spritenum*$0008, %11110000, offset&15)
     }
 
     sub set_mousepointer_hand() {

compiler/res/prog8lib/cx16/syslib.p8

@@ -19,7 +19,7 @@ romsub $FF8D = VECTOR(uword userptr @ XY, bool dir @ Pc) clobbers(A,Y)     ; rea
 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.   NOTE: as a Cx16 extension, also returns the number of RAM memory banks in register A !  See cx16.numbanks()
+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
@@ -32,7 +32,7 @@ romsub $FFB4 = TALK(ubyte device @ A) clobbers(A)               ; command serial
 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 $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
@@ -42,37 +42,50 @@ romsub $FFD2 = CHROUT(ubyte character @ A)                           ; (via 806
 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 (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)
-romsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A    ; (via 810 ($32A)) get a character
+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                 ; read number of screen rows and columns
-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 $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
 
 ; ---- utility
 
-asmsub STOP2() clobbers(X) -> bool @A  {
-    ; -- check if STOP key was pressed, returns true if so.  More convenient to use than STOP() because that only sets the zero status flag.
+inline asmsub STOP2() clobbers(X,A) -> bool @Pz  {
+    ; -- just like STOP, but omits the special keys result value in A.
+    ;    just for convenience because most of the times you're only interested in the stop pressed or not status.
     %asm {{
         jsr  cbm.STOP
-        beq  +
-        lda  #0
-        rts
-+       lda  #1
-        rts
     }}
 }
 
-asmsub RDTIM16() clobbers(X) -> uword @AY {
-    ; --  like RDTIM() but only returning the lower 16 bits in AY for convenience. Also avoids ram bank issue for irqs.
+inline asmsub GETIN2() clobbers(X,Y) -> ubyte @A {
+    ; -- just like GETIN, but omits the carry flag result value.
+    ;    just for convenience because GETIN is so often used to just read keyboard input,
+    ;    where you don't have to deal with a potential error status
+    %asm {{
+        jsr  cbm.GETIN
+    }}
+}
+
+asmsub RDTIM_safe() -> ubyte @ A, ubyte @ X, ubyte @ Y {
+    ; -- read the software clock (in little endian order: A=lo,X=mid,Y=high)
+    ;    with safeguard for ram bank issue for irqs.
     %asm {{
         php
         sei
         jsr  cbm.RDTIM
         plp
-        cli
+        rts
+    }}
+}
+
+asmsub RDTIM16() clobbers(X) -> uword @AY {
+    ; --  like RDTIM_safe() but only returning the lower 16 bits in AY for convenience.
+    %asm {{
+        jsr  RDTIM_safe
         pha
         txa
         tay
@@ -295,33 +308,33 @@ cx16 {
     &ubyte  VERA_SPI_CTRL       = VERA_BASE + $001F
 
     ; Vera FX registers: (accessing depends on particular DCSEL value set in VERA_CTRL!)
-    &ubyte VERA_FX_CTRL         = VERA_BASE + $0009
-    &ubyte VERA_FX_TILEBASE     = VERA_BASE + $000a
-    &ubyte VERA_FX_MAPBASE      = VERA_BASE + $000b
-    &ubyte VERA_FX_MULT         = VERA_BASE + $000c
-    &ubyte VERA_FX_X_INCR_L     = VERA_BASE + $0009
-    &ubyte VERA_FX_X_INCR_H     = VERA_BASE + $000a
-    &uword VERA_FX_X_INCR       = VERA_BASE + $0009
-    &ubyte VERA_FX_Y_INCR_L     = VERA_BASE + $000b
-    &ubyte VERA_FX_Y_INCR_H     = VERA_BASE + $000c
-    &uword VERA_FX_Y_INCR       = VERA_BASE + $000b
-    &ubyte VERA_FX_X_POS_L      = VERA_BASE + $0009
-    &ubyte VERA_FX_X_POS_H      = VERA_BASE + $000a
-    &uword VERA_FX_X_POS        = VERA_BASE + $0009
-    &ubyte VERA_FX_Y_POS_L      = VERA_BASE + $000b
-    &ubyte VERA_FX_Y_POS_H      = VERA_BASE + $000c
-    &uword VERA_FX_Y_POS        = VERA_BASE + $000b
-    &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
-    &uword VERA_FX_POLY_FILL    = VERA_BASE + $000b
-    &ubyte VERA_FX_CACHE_L      = VERA_BASE + $0009
-    &ubyte VERA_FX_CACHE_M      = VERA_BASE + $000a
-    &ubyte VERA_FX_CACHE_H      = VERA_BASE + $000b
-    &ubyte VERA_FX_CACHE_U      = VERA_BASE + $000c
-    &ubyte VERA_FX_ACCUM        = VERA_BASE + $000a
-    &ubyte VERA_FX_ACCUM_RESET  = VERA_BASE + $0009
+    &ubyte  VERA_FX_CTRL        = VERA_BASE + $0009
+    &ubyte  VERA_FX_TILEBASE    = VERA_BASE + $000a
+    &ubyte  VERA_FX_MAPBASE     = VERA_BASE + $000b
+    &ubyte  VERA_FX_MULT        = VERA_BASE + $000c
+    &ubyte  VERA_FX_X_INCR_L    = VERA_BASE + $0009
+    &ubyte  VERA_FX_X_INCR_H    = VERA_BASE + $000a
+    &uword  VERA_FX_X_INCR      = VERA_BASE + $0009
+    &ubyte  VERA_FX_Y_INCR_L    = VERA_BASE + $000b
+    &ubyte  VERA_FX_Y_INCR_H    = VERA_BASE + $000c
+    &uword  VERA_FX_Y_INCR      = VERA_BASE + $000b
+    &ubyte  VERA_FX_X_POS_L     = VERA_BASE + $0009
+    &ubyte  VERA_FX_X_POS_H     = VERA_BASE + $000a
+    &uword  VERA_FX_X_POS       = VERA_BASE + $0009
+    &ubyte  VERA_FX_Y_POS_L     = VERA_BASE + $000b
+    &ubyte  VERA_FX_Y_POS_H     = VERA_BASE + $000c
+    &uword  VERA_FX_Y_POS       = VERA_BASE + $000b
+    &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
+    &uword  VERA_FX_POLY_FILL   = VERA_BASE + $000b
+    &ubyte  VERA_FX_CACHE_L     = VERA_BASE + $0009
+    &ubyte  VERA_FX_CACHE_M     = VERA_BASE + $000a
+    &ubyte  VERA_FX_CACHE_H     = VERA_BASE + $000b
+    &ubyte  VERA_FX_CACHE_U     = VERA_BASE + $000c
+    &ubyte  VERA_FX_ACCUM       = VERA_BASE + $000a
+    &ubyte  VERA_FX_ACCUM_RESET = VERA_BASE + $0009
 
 
 ; VERA_PSG_BASE     = $1F9C0
@@ -376,20 +389,16 @@ cx16 {
 ; ---- Commander X-16 additions on top of C64 kernal routines ----
 ; spelling of the names is taken from the Commander X-16 rom sources
 
-; supported C128 additions
 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)  clobbers(X, Y) -> ubyte @A, bool @Pc        ; note: X,Y size result is not supported, use SCREEN or get_screen_mode routine for that
-romsub $ff62 = screen_set_charset(ubyte charset @A, uword charsetptr @XY)  clobbers(A,X,Y)      ; incompatible with C128  dlchr()
-; not yet supported: romsub $ff65 = pfkey()  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 bank @X, ubyte index @Y)  clobbers(X) -> ubyte @A
 romsub $ff77 = stash(ubyte data @A, ubyte bank @X, ubyte index @Y)  clobbers(X)
 romsub $ff7d = PRIMM()
 
-; It's not documented what registers are clobbered, so we assume the worst for all following kernal routines...:
-
 ; high level graphics & fonts
 romsub $ff20 = GRAPH_init(uword vectors @R0)  clobbers(A,X,Y)
 romsub $ff23 = GRAPH_clear()  clobbers(A,X,Y)
@@ -422,9 +431,11 @@ romsub $ff1a = FB_filter_pixels(uword pointer @ R0, uword count @R1)  clobbers(A
 romsub $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 $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 $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
+
 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)
@@ -437,10 +448,13 @@ 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)
 
 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()
@@ -448,16 +462,14 @@ romsub $ff50 = clock_get_date_time()  clobbers(A, X, Y)  -> uword @R0, uword @R1
 ; 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 $febd = kbdbuf_peek2() -> uword @AX             ; alternative to above to not have the hassle to deal with multiple return values
 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 zpdataptr @X) -> ubyte @A
+romsub $ff6b = mouse_get(ubyte zpdataptr @X) -> ubyte @A        ;  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) -> ubyte @A, ubyte @X, ubyte @Y
-romsub $ff56 = joystick_get2(ubyte joynr @A) clobbers(Y) -> uword @AX   ; alternative to above to not have the hassle to deal with multiple return values
+romsub $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)
@@ -524,6 +536,27 @@ 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
 
+; extapi call numbers
+const ubyte  EXTAPI_clear_status = $01
+const ubyte  EXTAPI_getlfs = $02
+const ubyte  EXTAPI_mouse_sprite_offset = $03
+const ubyte  EXTAPI_joystick_ps2_keycodes = $04
+const ubyte  EXTAPI_iso_cursor_char = $05
+const ubyte  EXTAPI_ps2kbd_typematic = $06
+const ubyte  EXTAPI_pfkey = $07
+const ubyte  EXTAPI_ps2data_fetch = $08
+const ubyte  EXTAPI_ps2data_raw = $09
+const ubyte  EXTAPI_cursor_blink = $0A
+const ubyte  EXTAPI_led_update = $0B
+const ubyte  EXTAPI_mouse_set_position = $0C
+
+; extapi16 call numbers
+const ubyte  EXTAPI16_test = $00
+const ubyte  EXTAPI16_stack_push = $01
+const ubyte  EXTAPI16_stack_pop = $02
+const ubyte  EXTAPI16_stack_enter_kernal_stack = $03
+const ubyte  EXTAPI16_stack_leave_kernal_stack = $04
+
 
 asmsub set_screen_mode(ubyte mode @A) clobbers(A,X,Y) -> bool @Pc {
     ; -- convenience wrapper for screen_mode() to just set a new mode (and return success)
@@ -533,9 +566,8 @@ asmsub set_screen_mode(ubyte mode @A) clobbers(A,X,Y) -> bool @Pc {
     }}
 }
 
-asmsub get_screen_mode() -> byte @A, byte @X, byte @Y {
-    ; -- convenience wrapper for screen_mode() to just get the current mode in A, and size in characters in X+Y
-    ;    this does need a piece of inlined asm to call it ans store the result values if you call this from prog8 code
+asmsub get_screen_mode() -> ubyte @A, ubyte @X, ubyte @Y {
+    ; -- convenience wrapper for screen_mode() to just get the current mode in A, and size in characters in X (width) and Y (height)
     ;    Note: you can also just do the SEC yourself and simply call screen_mode() directly,
     ;          or use the existing SCREEN kernal routine for just getting the size in characters.
     %asm {{
@@ -555,15 +587,63 @@ asmsub mouse_config2(byte shape @A) clobbers (A, X, Y) {
     }}
 }
 
-asmsub mouse_pos() clobbers(X) -> ubyte @A {
+asmsub mouse_pos() clobbers(X) -> ubyte @A, word @R0, word @R1 {
     ; -- short wrapper around mouse_get() kernal routine:
     ; -- gets the position of the mouse cursor in cx16.r0 and cx16.r1 (x/y coordinate), returns mouse button status in A.
+    ;    Note: mouse pointer needs to be enabled for this to do anything.
     %asm {{
         ldx  #cx16.r0
         jmp  cx16.mouse_get
     }}
 }
 
+; shims for the kernal routines called via the extapi call:
+
+asmsub mouse_set_pos(uword xpos @R0, uword ypos @R1) clobbers(X) {
+    ; -- sets the mouse sprite position
+    ;    Note: mouse pointer needs to be enabled for this to do anything.
+    %asm {{
+        ldx  #cx16.r0L
+        lda  #EXTAPI_mouse_set_position
+        jmp  cx16.extapi
+    }}
+}
+
+asmsub mouse_set_sprite_offset(word xoffset @R0, word yoffset @R1) clobbers(A,X,Y) {
+    %asm {{
+        clc
+        lda  #EXTAPI_mouse_sprite_offset
+        jmp  cx16.extapi
+    }}
+}
+
+asmsub mouse_get_sprite_offset() clobbers(A,X,Y) -> word @R0, word @R1 {
+    %asm {{
+        sec
+        lda  #EXTAPI_mouse_sprite_offset
+        jmp  cx16.extapi
+    }}
+}
+
+asmsub getlfs() -> ubyte @X, ubyte @A, ubyte @Y {
+    ; -- return the result of the last call to SETLFS:  A=logical, X=device, Y=secondary.
+    %asm {{
+        lda  #EXTAPI_mouse_set_position
+        jmp  cx16.extapi
+    }}
+}
+
+asmsub iso_cursor_char(ubyte character @X) clobbers(A,X,Y) {
+    ; -- set the screen code for the cursor character in ISO mode (the default is $9f).
+    %asm {{
+        clc
+        lda  #EXTAPI_iso_cursor_char
+        jmp  cx16.extapi
+    }}
+}
+
+; TODO : implement shims for the remaining extapi calls.
+
 
 ; ---- end of kernal routines ----
 
@@ -918,7 +998,7 @@ asmsub restore_vera_context() clobbers(A) {
 
     asmsub set_chrin_keyhandler(ubyte handlerbank @A, uword handler @XY) clobbers(A) {
         ; Install a custom CHRIN (BASIN) key handler in a safe manner. Call this before each line you want to read.
-        ; See https://github.com/X16Community/x16-docs/blob/master/X16%20Reference%20-%2002%20-%20Editor.md#custom-basin-petscii-code-override-handler
+        ; See https://github.com/X16Community/x16-docs/blob/101759f3bfa5e6cce4e8c5a0b67cb0f2f1c6341e/X16%20Reference%20-%2003%20-%20Editor.md#custom-basin-petscii-code-override-handler
         %asm {{
             sei
             sta  P8ZP_SCRATCH_REG
@@ -1158,7 +1238,7 @@ sub search_x16edit() -> ubyte {
 }
 
     asmsub cpu_is_65816() -> bool @A {
-        ; Returns true when you have a 65816 cpu, false when it's a 6502.
+        ; -- Returns true when you have a 65816 cpu, false when it's a 6502.
         %asm {{
 			php
 			clv
@@ -1174,9 +1254,9 @@ sub search_x16edit() -> ubyte {
     }
 
     sub set_program_args(uword args_ptr, ubyte args_size) {
-        ; Set the inter-program arguments.
+        ; -- Set the inter-program arguments.
         ; standardized way to pass arguments between programs is in ram bank 0, address $bf00-$bfff.
-        ; see https://github.com/X16Community/x16-docs/blob/master/X16%20Reference%20-%2007%20-%20Memory%20Map.md#bank-0
+        ; see https://github.com/X16Community/x16-docs/blob/101759f3bfa5e6cce4e8c5a0b67cb0f2f1c6341e/X16%20Reference%20-%2008%20-%20Memory%20Map.md#bank-0
         sys.push(getrambank())
         rambank(0)
         sys.memcopy(args_ptr, $bf00, args_size)
@@ -1186,9 +1266,9 @@ sub search_x16edit() -> ubyte {
     }
 
     asmsub get_program_args(uword buffer @R0, ubyte buf_size @R1, bool binary @Pc) {
-        ; Retrieve the inter-program arguments. If binary=false, it treats them as a string (stops copying at first zero).
+        ; -- Retrieve the inter-program arguments. If binary=false, it treats them as a string (stops copying at first zero).
         ; standardized way to pass arguments between programs is in ram bank 0, address $bf00-$bfff.
-        ; see https://github.com/X16Community/x16-docs/blob/master/X16%20Reference%20-%2007%20-%20Memory%20Map.md#bank-0
+        ; see https://github.com/X16Community/x16-docs/blob/101759f3bfa5e6cce4e8c5a0b67cb0f2f1c6341e/X16%20Reference%20-%2008%20-%20Memory%20Map.md#bank-0
         %asm {{
             lda  #0
             rol  a
@@ -1215,17 +1295,21 @@ _continue   iny
     }
 
     sub reset_system() {
-        ; Soft-reset the system back to initial power-on Basic prompt.
+        ; -- Soft-reset the system back to initial power-on Basic prompt.
         sys.reset_system()
     }
 
     sub poweroff_system() {
-        ; use the SMC to shutdown the computer
+        ; -- use the SMC to shutdown the computer
         void cx16.i2c_write_byte($42, $01, $00)
     }
 
-    sub set_led_brightness(ubyte brightness) {
-        void cx16.i2c_write_byte($42, $05, brightness)
+    sub set_led_state(bool on) {
+        ; -- sets the computer's activity led on/off
+        cx16.r0L = 0
+        if on
+            cx16.r0 = 255
+        void cx16.i2c_write_byte($42, $05, cx16.r0L)
     }
 
 }
@@ -1442,11 +1526,6 @@ asmsub  set_rasterline(uword line @AY) {
         void cx16.i2c_write_byte($42, $01, $00)
     }
 
-    sub set_leds_brightness(ubyte activity, ubyte power) {
-        void cx16.i2c_write_byte($42, $04, power)
-        void cx16.i2c_write_byte($42, $05, activity)
-    }
-
     asmsub wait(uword jiffies @AY) clobbers(X) {
         ; --- wait approximately the given number of jiffies (1/60th seconds) (N or N+1)
         ;     note: the system irq handler has to be active for this to work as it depends on the system jiffy clock

compiler/res/prog8lib/cx16/textio.p8

@@ -4,7 +4,7 @@
 
 %import syslib
 %import conv
-%import shared_textio_functions
+%import shared_cbm_textio_functions
 
 txt {
 
@@ -16,11 +16,17 @@ 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.
+
 
 sub clear_screen() {
     chrout(147)
 }
 
+sub cls() {
+    chrout(147)
+}
+
 sub home() {
     chrout(19)
 }
@@ -223,8 +229,8 @@ sub uppercase() {
 }
 
 sub iso() {
+    ; -- switch to iso-8859-15 character set
     cbm.CHROUT($0f)
-    ; This doesn't enable it completely: cx16.screen_set_charset(1, 0)  ; iso charset
 }
 
 sub iso_off() {
@@ -232,6 +238,29 @@ sub iso_off() {
     cbm.CHROUT($8f)
 }
 
+sub cp437() {
+    ; -- switch to CP-437 (ibm PC) character set
+    cbm.CHROUT($0f)                 ; iso mode
+    cx16.screen_set_charset(7, 0)   ; charset
+    %asm {{
+        clc
+        ldx  #95        ; underscore
+        lda  #cx16.EXTAPI_iso_cursor_char
+        jsr  cx16.extapi
+    }}
+}
+
+sub iso5() {
+    ; -- switch to iso-8859-5 character set (Cyrillic)
+    cbm.CHROUT($0f)                 ; iso mode
+    cx16.screen_set_charset(8, 0)   ; charset
+}
+
+sub iso16() {
+    ; -- switch to iso-8859-16 character set (Eastern Europe)
+    cbm.CHROUT($0f)                 ; iso mode
+    cx16.screen_set_charset(10, 0)  ; charset
+}
 
 asmsub  scroll_left() clobbers(A, X, Y)  {
 	; ---- scroll the whole screen 1 character to the left
@@ -431,215 +460,6 @@ _nextline
 	}}
 }
 
-romsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
-
-asmsub  print (str text @ AY) clobbers(A,Y)  {
-	; ---- print null terminated string, in PETSCII encoding, from A/Y
-	; note: the compiler contains an optimization that will replace
-	;       a call to this subroutine with a string argument of just one char,
-	;       by just one call to cbm.CHROUT of that single char.
-	%asm {{
-		sta  P8ZP_SCRATCH_B1
-		sty  P8ZP_SCRATCH_REG
-		ldy  #0
--		lda  (P8ZP_SCRATCH_B1),y
-		beq  +
-		jsr  cbm.CHROUT
-		iny
-		bne  -
-+		rts
-	}}
-}
-
-asmsub  print_ub0  (ubyte value @ A) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in decimal form, with left padding 0s (3 positions total)
-	%asm {{
-		jsr  conv.ubyte2decimal
-		pha
-		tya
-		jsr  cbm.CHROUT
-		pla
-		jsr  cbm.CHROUT
-		txa
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_ub  (ubyte value @ A) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in decimal form, without left padding 0s
-	%asm {{
-		jsr  conv.ubyte2decimal
-_print_byte_digits
-		pha
-		cpy  #'0'
-		beq  +
-		tya
-		jsr  cbm.CHROUT
-		pla
-		jsr  cbm.CHROUT
-		bra  _ones
-+       pla
-        cmp  #'0'
-        beq  _ones
-        jsr  cbm.CHROUT
-_ones   txa
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_b  (byte value @ A) clobbers(A,X,Y)  {
-	; ---- print the byte in A in decimal form, without left padding 0s
-	%asm {{
-		pha
-		cmp  #0
-		bpl  +
-		lda  #'-'
-		jsr  cbm.CHROUT
-+		pla
-		jsr  conv.byte2decimal
-		bra  print_ub._print_byte_digits
-	}}
-}
-
-asmsub  print_ubhex  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in hex form (if Carry is set, a radix prefix '$' is printed as well)
-	%asm {{
-		bcc  +
-		pha
-		lda  #'$'
-		jsr  cbm.CHROUT
-		pla
-+		jsr  conv.ubyte2hex
-		jsr  cbm.CHROUT
-		tya
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_ubbin  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in binary form (if Carry is set, a radix prefix '%' is printed as well)
-	%asm {{
-		sta  P8ZP_SCRATCH_B1
-		bcc  +
-		lda  #'%'
-		jsr  cbm.CHROUT
-+		ldy  #8
--		lda  #'0'
-		asl  P8ZP_SCRATCH_B1
-		bcc  +
-		lda  #'1'
-+		jsr  cbm.CHROUT
-		dey
-		bne  -
-		rts
-	}}
-}
-
-asmsub  print_uwbin  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in binary form (if Carry is set, a radix prefix '%' is printed as well)
-	%asm {{
-		pha
-		tya
-		jsr  print_ubbin
-		pla
-		clc
-		bra  print_ubbin
-	}}
-}
-
-asmsub  print_uwhex  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in hexadecimal form (4 digits)
-	;      (if Carry is set, a radix prefix '$' is printed as well)
-	%asm {{
-		pha
-		tya
-		jsr  print_ubhex
-		pla
-		clc
-		bra  print_ubhex
-	}}
-}
-
-asmsub  print_uw0  (uword value @ AY) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in decimal form, with left padding 0s (5 positions total)
-	%asm {{
-		jsr  conv.uword2decimal
-		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
-        beq  +
-		jsr  cbm.CHROUT
-		iny
-		bne  -
-+		rts
-	}}
-}
-
-asmsub  print_uw  (uword value @ AY) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in decimal form, without left padding 0s
-	%asm {{
-		jsr  conv.uword2decimal
-		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
-		beq  _allzero
-		cmp  #'0'
-		bne  _gotdigit
-		iny
-		bne  -
-
-_gotdigit
-		jsr  cbm.CHROUT
-		iny
-		lda  conv.uword2decimal.decTenThousands,y
-		bne  _gotdigit
-		rts
-_allzero
-        lda  #'0'
-        jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_w  (word value @ AY) clobbers(A,X,Y)  {
-	; ---- print the (signed) word in A/Y in decimal form, without left padding 0's
-	%asm {{
-		cpy  #0
-		bpl  +
-		pha
-		lda  #'-'
-		jsr  cbm.CHROUT
-		tya
-		eor  #255
-		tay
-		pla
-		eor  #255
-		ina
-		bne +
-		iny
-+		bra  print_uw
-	}}
-}
-
-asmsub  input_chars  (uword buffer @ AY) clobbers(A) -> ubyte @ Y  {
-	; ---- Input a string (max. 80 chars) from the keyboard, in PETSCII encoding.
-	;      Returns length in Y. (string is terminated with a 0 byte as well)
-	;      It assumes the keyboard is selected as I/O channel!
-
-	%asm {{
-		sta  P8ZP_SCRATCH_W1
-		sty  P8ZP_SCRATCH_W1+1
-		ldy  #0				; char counter = 0
--		jsr  cbm.CHRIN
-		cmp  #$0d			; return (ascii 13) pressed?
-		beq  +				; yes, end.
-		sta  (P8ZP_SCRATCH_W1),y	; else store char in buffer
-		iny
-		bne  -
-+		lda  #0
-		sta  (P8ZP_SCRATCH_W1),y	; finish string with 0 byte
-		rts
-
-	}}
-}
-
 asmsub  setchr  (ubyte col @X, ubyte row @Y, ubyte character @A) clobbers(A)  {
 	; ---- sets the character in the screen matrix at the given position
 	%asm {{
@@ -811,4 +631,35 @@ asmsub waitkey() -> ubyte @A {
         rts
     }}
 }
+
+asmsub chrout_lit(ubyte character @A) {
+    ; -- print the character always as a literal character, not as possible control code.
+    %asm {{
+        tax
+        lda  #128
+        jsr  cbm.CHROUT
+        txa
+        jmp  cbm.CHROUT
+    }}
+}
+
+asmsub print_lit(str text @ AY) clobbers(A,Y)  {
+    ; -- print zero terminated string, from A/Y, as all literal characters (no control codes)
+    %asm {{
+        sta  P8ZP_SCRATCH_W2
+        sty  P8ZP_SCRATCH_W2+1
+        ldy  #0
+-       lda  (P8ZP_SCRATCH_W2),y
+        beq  +
+        tax
+        lda  #128
+        jsr  cbm.CHROUT
+        txa
+        jsr  cbm.CHROUT
+        iny
+        bne  -
++       rts
+    }}
+}
+
 }

compiler/res/prog8lib/cx16/verafx.p8

@@ -1,6 +1,6 @@
 ; Somewhat experimental Vera FX support.
 ; Docs:
-; https://github.com/X16Community/x16-docs/blob/master/X16%20Reference%20-%2010%20-%20VERA%20FX%20Reference.md
+; https://github.com/X16Community/x16-docs/blob/101759f3bfa5e6cce4e8c5a0b67cb0f2f1c6341e/X16%20Reference%20-%2010%20-%20VERA%20FX%20Reference.md
 ; https://docs.google.com/document/d/1q34uWOiM3Be2pnaHRVgSdHySI-qsiQWPTo_gfE54PTg
 
 verafx {
@@ -111,17 +111,15 @@ verafx {
 
     ; unsigned multiplication just passes the values as signed to muls
     ; if you do this yourself in your call to muls, it will save a few instructions.
-    sub mult(uword value1, uword value2) -> uword {
-        ; Returns the lower 16 bits of the 32 bits result,
-        ; the upper 16 bits are stored in cx16.r0 so you can access those separately.
-        ; It's not part of the subroutine's signature to avoid awkward use of multiple returnvalues.
-        return muls(value1 as word, value2 as word) as uword
+    inline asmsub mult(uword value1 @R0, uword value2 @R1) clobbers(X) -> uword @AY, uword @R0 {
+        ; Returns the 32 bits unsigned result in AY and R0  (lower word, upper word).
+        %asm {{
+            jsr  verafx.muls
+        }}
     }
 
-    asmsub muls(word value1 @R0, word value2 @R1) clobbers(X) -> word @AY {
-        ; Returns the lower 16 bits of the 32 bits result in AY,
-        ; the upper 16 bits are stored in cx16.r0 so you can access those separately.
-        ; It's not part of the subroutine's signature to avoid awkward use of multiple returnvalues.
+    asmsub muls(word value1 @R0, word value2 @R1) clobbers(X) -> word @AY, word @R0 {
+        ; Returns the 32 bits signed result in AY and R0  (lower word, upper word).
         %asm {{
             lda  #(2 << 1)
             sta  cx16.VERA_CTRL        ; $9F25
@@ -131,13 +129,13 @@ verafx {
             lda  #(6 << 1)
             sta  cx16.VERA_CTRL        ; $9F25
             lda  cx16.r0
-            ldy  cx16.r0+1
             sta  cx16.VERA_FX_CACHE_L  ; $9F29
-            sty  cx16.VERA_FX_CACHE_M  ; $9F2A
+            lda  cx16.r0+1
+            sta  cx16.VERA_FX_CACHE_M  ; $9F2A
             lda  cx16.r1
-            ldy  cx16.r1+1
             sta  cx16.VERA_FX_CACHE_H  ; $9F2B
-            sty  cx16.VERA_FX_CACHE_U  ; $9F2C
+            lda  cx16.r1+1
+            sta  cx16.VERA_FX_CACHE_U  ; $9F2C
             lda  cx16.VERA_FX_ACCUM_RESET   ; $9F29 (DCSEL=6)
 
             ; Set the ADDR0 pointer to $1f9bc and write our multiplication result there
@@ -155,12 +153,12 @@ verafx {
             stz  cx16.VERA_DATA0      ; multiply and write out result
             lda  #%00010001           ; $01 with Increment 1
             sta  cx16.VERA_ADDR_H     ; so we can read out the result
-            lda  cx16.VERA_DATA0
+            lda  cx16.VERA_DATA0      ; store the lower 16 bits of the result in AY
             ldy  cx16.VERA_DATA0
-            ldx  cx16.VERA_DATA0      ; store the upper 16 bits of the result in r0
-            stx  cx16.r0
+            ldx  cx16.VERA_DATA0      ; store the upper 16 bits of the result in R0
+            stx  cx16.r0s
             ldx  cx16.VERA_DATA0
-            stx  cx16.r0+1
+            stx  cx16.r0s+1
             stz  cx16.VERA_FX_CTRL    ; Cache write disable
             stz  cx16.VERA_CTRL       ; reset DCSEL
             rts

compiler/res/prog8lib/diskio.p8

@@ -59,7 +59,8 @@ diskio {
             void cbm.CHRIN()     ; skip 2 bytes
             void cbm.CHRIN()
             status = cbm.READST()
-            if cbm.STOP2()
+            void cbm.STOP()
+            if_z
                 break
         }
         status = cbm.READST()
@@ -460,6 +461,38 @@ io_error:
         goto done
     }
 
+    ; similar to above, but instead of fetching the entire string, it only fetches the status code and returns it as ubyte
+    ; in case of IO error, returns 255 (CBM-DOS itself is physically unable to return such a value)
+    sub status_code() -> ubyte {
+        if cbm.READST()==128 {
+            return 255
+        }
+
+        cbm.SETNAM(0, list_filename)
+        cbm.SETLFS(15, drivenumber, 15)
+        void cbm.OPEN()          ; open 15,8,15
+        if_cs
+            goto io_error
+        void cbm.CHKIN(15)        ; use #15 as input channel
+
+        list_filename[0] = cbm.CHRIN()
+        list_filename[1] = cbm.CHRIN()
+        list_filename[2] = 0
+
+        while cbm.READST()==0 {
+            void cbm.CHRIN()
+        }
+
+        cbm.CLRCHN()        ; restore default i/o devices
+        cbm.CLOSE(15)
+        return conv.str2ubyte(list_filename)
+
+io_error:
+        cbm.CLRCHN()
+        cbm.CLOSE(15)
+        return 255
+    }
+
     sub save(uword filenameptr, uword start_address, uword savesize) -> bool {
         cbm.SETNAM(string.length(filenameptr), filenameptr)
         cbm.SETLFS(1, drivenumber, 0)

compiler/res/prog8lib/pet32/syslib.p8

@@ -29,20 +29,26 @@ romsub $FFC9 = CHKOUT(ubyte logical @ X) clobbers(A,X)          ; define an outp
 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)
-romsub $FFE4 = GETIN() clobbers(X,Y) -> bool @Pc, ubyte @ A     ; get 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
 
-asmsub STOP2() clobbers(X) -> bool @A  {
-    ; -- check if STOP key was pressed, returns true if so.  More convenient to use than STOP() because that only sets the carry status flag.
+
+inline asmsub STOP2() clobbers(X,A) -> bool @Pz  {
+    ; -- just like STOP, but omits the special keys result value in A.
+    ;    just for convenience because most of the times you're only interested in the stop pressed or not status.
     %asm {{
         jsr  cbm.STOP
-        beq  +
-        lda  #0
-        rts
-+       lda  #1
-        rts
+    }}
+}
+
+inline asmsub GETIN2() clobbers(X,Y) -> ubyte @A {
+    ; -- just like GETIN, but omits the carry flag result value.
+    ;    just for convenience because GETIN is so often used to just read keyboard input,
+    ;    where you don't have to deal with a potential error status
+    %asm {{
+        jsr  cbm.GETIN
     }}
 }
 

compiler/res/prog8lib/pet32/textio.p8

@@ -4,7 +4,7 @@
 
 %import syslib
 %import conv
-%import shared_textio_functions
+%import shared_cbm_textio_functions
 
 txt {
     %option no_symbol_prefixing, ignore_unused
@@ -12,11 +12,17 @@ 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.
+
 
 sub  clear_screen() {
     chrout(147)
 }
 
+sub cls() {
+    chrout(147)
+}
+
 sub home() {
     chrout(19)
 }
@@ -145,216 +151,6 @@ asmsub  scroll_down  () clobbers(A,X)  {
 	}}
 }
 
-romsub $FFD2 = chrout(ubyte character @ A)    ; for consistency. You can also use cbm.CHROUT directly ofcourse. Note: takes a PETSCII encoded character.
-
-asmsub  print (str text @ AY) clobbers(A,Y)  {
-	; ---- print null terminated string, in PETSCII encoding, from A/Y
-	; note: the compiler contains an optimization that will replace
-	;       a call to this subroutine with a string argument of just one char,
-	;       by just one call to cbm.CHROUT of that single char.
-	%asm {{
-		sta  P8ZP_SCRATCH_B1
-		sty  P8ZP_SCRATCH_REG
-		ldy  #0
--		lda  (P8ZP_SCRATCH_B1),y
-		beq  +
-		jsr  cbm.CHROUT
-		iny
-		bne  -
-+		rts
-	}}
-}
-
-asmsub  print_ub0  (ubyte value @ A) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in decimal form, with left padding 0s (3 positions total)
-	%asm {{
-		jsr  conv.ubyte2decimal
-		pha
-		tya
-		jsr  cbm.CHROUT
-		pla
-		jsr  cbm.CHROUT
-		txa
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_ub  (ubyte value @ A) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in decimal form, without left padding 0s
-	%asm {{
-		jsr  conv.ubyte2decimal
-_print_byte_digits
-		pha
-		cpy  #'0'
-		beq  +
-		tya
-		jsr  cbm.CHROUT
-		pla
-		jsr  cbm.CHROUT
-		jmp  _ones
-+       pla
-        cmp  #'0'
-        beq  _ones
-        jsr  cbm.CHROUT
-_ones   txa
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_b  (byte value @ A) clobbers(A,X,Y)  {
-	; ---- print the byte in A in decimal form, without left padding 0s
-	%asm {{
-		pha
-		cmp  #0
-		bpl  +
-		lda  #'-'
-		jsr  cbm.CHROUT
-+		pla
-		jsr  conv.byte2decimal
-		jmp  print_ub._print_byte_digits
-	}}
-}
-
-asmsub  print_ubhex  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in hex form (if Carry is set, a radix prefix '$' is printed as well)
-	%asm {{
-		bcc  +
-		pha
-		lda  #'$'
-		jsr  cbm.CHROUT
-		pla
-+		jsr  conv.ubyte2hex
-		jsr  cbm.CHROUT
-		tya
-		jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_ubbin  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the ubyte in A in binary form (if Carry is set, a radix prefix '%' is printed as well)
-	%asm {{
-		sta  P8ZP_SCRATCH_B1
-		bcc  +
-		lda  #'%'
-		jsr  cbm.CHROUT
-+		ldy  #8
--		lda  #'0'
-		asl  P8ZP_SCRATCH_B1
-		bcc  +
-		lda  #'1'
-+		jsr  cbm.CHROUT
-		dey
-		bne  -
-		rts
-	}}
-}
-
-asmsub  print_uwbin  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in binary form (if Carry is set, a radix prefix '%' is printed as well)
-	%asm {{
-		pha
-		tya
-		jsr  print_ubbin
-		pla
-		clc
-		jmp  print_ubbin
-	}}
-}
-
-asmsub  print_uwhex  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in hexadecimal form (4 digits)
-	;      (if Carry is set, a radix prefix '$' is printed as well)
-	%asm {{
-		pha
-		tya
-		jsr  print_ubhex
-		pla
-		clc
-		jmp  print_ubhex
-	}}
-}
-
-asmsub  print_uw0  (uword value @ AY) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in decimal form, with left padding 0s (5 positions total)
-	%asm {{
-		jsr  conv.uword2decimal
-		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
-        beq  +
-		jsr  cbm.CHROUT
-		iny
-		bne  -
-+		rts
-	}}
-}
-
-asmsub  print_uw  (uword value @ AY) clobbers(A,X,Y)  {
-	; ---- print the uword in A/Y in decimal form, without left padding 0s
-	%asm {{
-		jsr  conv.uword2decimal
-		ldy  #0
--		lda  conv.uword2decimal.decTenThousands,y
-		beq  _allzero
-		cmp  #'0'
-		bne  _gotdigit
-		iny
-		bne  -
-
-_gotdigit
-		jsr  cbm.CHROUT
-		iny
-		lda  conv.uword2decimal.decTenThousands,y
-		bne  _gotdigit
-		rts
-_allzero
-        lda  #'0'
-        jmp  cbm.CHROUT
-	}}
-}
-
-asmsub  print_w  (word value @ AY) clobbers(A,X,Y)  {
-	; ---- print the (signed) word in A/Y in decimal form, without left padding 0's
-	%asm {{
-		cpy  #0
-		bpl  +
-		pha
-		lda  #'-'
-		jsr  cbm.CHROUT
-		tya
-		eor  #255
-		tay
-		pla
-		eor  #255
-		clc
-		adc  #1
-		bcc  +
-		iny
-+		jmp  print_uw
-	}}
-}
-
-asmsub  input_chars  (uword buffer @ AY) clobbers(A) -> ubyte @ Y  {
-	; ---- Input a string (max. 80 chars) from the keyboard, in PETSCII encoding.
-	;      Returns length in Y. (string is terminated with a 0 byte as well)
-	;      It assumes the keyboard is selected as I/O channel!
-
-	%asm {{
-		sta  P8ZP_SCRATCH_W1
-		sty  P8ZP_SCRATCH_W1+1
-		ldy  #0				; char counter = 0
--		jsr  cbm.CHRIN
-		cmp  #$0d			; return (ascii 13) pressed?
-		beq  +				; yes, end.
-		sta  (P8ZP_SCRATCH_W1),y	; else store char in buffer
-		iny
-		bne  -
-+		lda  #0
-		sta  (P8ZP_SCRATCH_W1),y	; finish string with 0 byte
-		rts
-
-	}}
-}
-
 asmsub  setchr  (ubyte col @X, ubyte row @Y, ubyte character @A) clobbers(A, Y)  {
 	; ---- sets the character in the screen matrix at the given position
 	%asm {{

compiler/res/prog8lib/shared_cbm_textio_functions.p8

@@ -0,0 +1,251 @@
+txt {
+    ; the textio functions common across CBM-compatible compiler targets (c64, c128, pet32 and cx16)
+
+    %option merge, no_symbol_prefixing, ignore_unused
+
+    asmsub  print (str text @ AY) clobbers(A,Y)  {
+    	; ---- print zero terminated string, in PETSCII encoding, from A/Y
+    	; note: the compiler contains an optimization that will replace
+    	;       a call to this subroutine with a string argument of just one char,
+    	;       by just one call to cbm.CHROUT of that single char.
+    	%asm {{
+    		sta  P8ZP_SCRATCH_W2
+    		sty  P8ZP_SCRATCH_W2+1
+    		ldy  #0
+-	    	lda  (P8ZP_SCRATCH_W2),y
+    		beq  +
+    		jsr  cbm.CHROUT
+    		iny
+    		bne  -
++	    	rts
+    	}}
+    }
+
+    asmsub  print_ubhex  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
+        ; ---- print the ubyte in A in hex form (if Carry is set, a radix prefix '$' is printed as well)
+        %asm {{
+            bcc  +
+            pha
+            lda  #'$'
+            jsr  cbm.CHROUT
+            pla
++           jsr  conv.internal_ubyte2hex
+            jsr  cbm.CHROUT
+            tya
+            jmp  cbm.CHROUT
+        }}
+    }
+
+    asmsub  print_ubbin  (ubyte value @ A, bool prefix @ Pc) clobbers(A,X,Y)  {
+        ; ---- print the ubyte in A in binary form (if Carry is set, a radix prefix '%' is printed as well)
+        %asm {{
+            sta  P8ZP_SCRATCH_B1
+            bcc  +
+            lda  #'%'
+            jsr  cbm.CHROUT
++           ldy  #8
+-           lda  #'0'
+            asl  P8ZP_SCRATCH_B1
+            bcc  +
+            lda  #'1'
++           jsr  cbm.CHROUT
+            dey
+            bne  -
+            rts
+        }}
+    }
+
+    asmsub  print_uwbin  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
+        ; ---- print the uword in A/Y in binary form (if Carry is set, a radix prefix '%' is printed as well)
+        %asm {{
+            pha
+            tya
+            jsr  print_ubbin
+            pla
+            clc
+            jmp  print_ubbin
+        }}
+    }
+
+    asmsub  print_uwhex  (uword value @ AY, bool prefix @ Pc) clobbers(A,X,Y)  {
+        ; ---- print the uword in A/Y in hexadecimal form (4 digits)
+        ;      (if Carry is set, a radix prefix '$' is printed as well)
+        %asm {{
+            pha
+            tya
+            jsr  print_ubhex
+            pla
+            clc
+            jmp  print_ubhex
+        }}
+    }
+
+    asmsub  print_uw0  (uword value @ AY) clobbers(A,X,Y)  {
+        ; ---- print the uword in A/Y in decimal form, with left padding 0s (5 positions total)
+        %asm {{
+            jsr  conv.internal_uword2decimal
+            ldy  #0
+-           lda  conv.internal_uword2decimal.decTenThousands,y
+            beq  +
+            jsr  cbm.CHROUT
+            iny
+            bne  -
++           rts
+        }}
+    }
+
+    asmsub  print_uw  (uword value @ AY) clobbers(A,X,Y)  {
+        ; ---- print the uword in A/Y in decimal form, without left padding 0s
+        %asm {{
+            jsr  conv.internal_uword2decimal
+            ldy  #0
+-           lda  conv.internal_uword2decimal.decTenThousands,y
+            beq  _allzero
+            cmp  #'0'
+            bne  _gotdigit
+            iny
+            bne  -
+_gotdigit   jsr  cbm.CHROUT
+            iny
+            lda  conv.internal_uword2decimal.decTenThousands,y
+            bne  _gotdigit
+            rts
+_allzero    lda  #'0'
+            jmp  cbm.CHROUT
+        }}
+    }
+
+    asmsub  print_w  (word value @ AY) clobbers(A,X,Y)  {
+        ; ---- print the (signed) word in A/Y in decimal form, without left padding 0's
+        %asm {{
+            cpy  #0
+            bpl  +
+            pha
+            lda  #'-'
+            jsr  cbm.CHROUT
+            tya
+            eor  #255
+            tay
+            pla
+            eor  #255
+            clc
+            adc  #1
+            bcc  +
+            iny
++           jmp  print_uw
+        }}
+    }
+
+    asmsub  input_chars  (uword buffer @ AY) clobbers(A) -> ubyte @ Y  {
+        ; ---- Input a string (max. 80 chars) from the keyboard, in PETSCII encoding.
+        ;      Returns length in Y. (string is terminated with a 0 byte as well)
+        ;      It assumes the keyboard is selected as I/O channel!
+
+        %asm {{
+            sta  P8ZP_SCRATCH_W1
+            sty  P8ZP_SCRATCH_W1+1
+            ldy  #0				; char counter = 0
+-           jsr  cbm.CHRIN
+            cmp  #$0d			; return (ascii 13) pressed?
+            beq  +				; yes, end.
+            sta  (P8ZP_SCRATCH_W1),y	; else store char in buffer
+            iny
+            bne  -
++           lda  #0
+            sta  (P8ZP_SCRATCH_W1),y	; finish string with 0 byte
+            rts
+        }}
+    }
+
+    asmsub petscii2scr(ubyte petscii_char @A) -> ubyte @A {
+        ; -- convert petscii character to screencode
+        %asm {{
+            sta  P8ZP_SCRATCH_REG
+            lsr  a
+            lsr  a
+            lsr  a
+            lsr  a
+            lsr  a
+            tax
+            lda  _offsets,x
+            eor  P8ZP_SCRATCH_REG
+            rts
+_offsets    .byte  128, 0, 64, 32, 64, 192, 128, 128
+        }}
+    }
+
+    asmsub petscii2scr_str(str petscii_string @AY) {
+        ; -- convert petscii string to screencodes string
+        %asm {{
+            sta  P8ZP_SCRATCH_W1
+            sty  P8ZP_SCRATCH_W1+1
+            ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            beq  +
+            jsr  petscii2scr
+            sta  (P8ZP_SCRATCH_W1),y
+            iny
+            bne  -
++           rts
+        }}
+}
+
+
+    sub print_bool(bool value) {
+        if value
+            txt.print("true")
+        else
+            txt.print("false")
+    }
+
+    asmsub  print_ub0  (ubyte value @ A) clobbers(A,X,Y)  {
+        ; ---- print the ubyte in A in decimal form, with left padding 0s (3 positions total)
+        %asm {{
+            jsr  conv.internal_ubyte2decimal
+            pha
+            tya
+            jsr  cbm.CHROUT
+            txa
+            jsr  cbm.CHROUT
+            pla
+            jmp  cbm.CHROUT
+        }}
+    }
+
+    asmsub  print_ub  (ubyte value @ A) clobbers(A,X,Y)  {
+        ; ---- print the ubyte in A in decimal form, without left padding 0s
+        %asm {{
+            jsr  conv.internal_ubyte2decimal
+_print_byte_digits
+            pha
+            cpy  #'0'
+            beq  +
+            tya
+            jsr  cbm.CHROUT
+            txa
+            jsr  cbm.CHROUT
+            jmp  _ones
++           cpx  #'0'
+            beq  _ones
+            txa
+            jsr  cbm.CHROUT
+_ones       pla
+            jmp  cbm.CHROUT
+        }}
+    }
+
+    asmsub  print_b  (byte value @ A) clobbers(A,X,Y)  {
+        ; ---- print the byte in A in decimal form, without left padding 0s
+        %asm {{
+            pha
+            cmp  #0
+            bpl  +
+            lda  #'-'
+            jsr  cbm.CHROUT
++		    pla
+            jsr  conv.internal_byte2decimal
+            jmp  print_ub._print_byte_digits
+        }}
+    }
+
+}

compiler/res/prog8lib/shared_floats_functions.p8

@@ -2,6 +2,11 @@ floats {
     ; the floating point functions shared across compiler targets
     %option merge, no_symbol_prefixing, ignore_unused
 
+    const float  π      = 3.141592653589793
+    const float  PI     = π
+    const float  TWOPI  = 2*π
+
+
 asmsub print(float value @FAC1) clobbers(A,X,Y) {
 	; ---- prints the floating point value (without a newline). No leading space (unlike BASIC)!
 	%asm {{
@@ -88,6 +93,20 @@ sub atan(float value) -> float {
     }}
 }
 
+; two-argument arctangent that returns an angle in the correct quadrant
+; for the signs of x and y, normalized to the range [0, 2π]
+sub atan2(float y, float x) -> float {
+    float atn = atan(y / x)
+    if x < 0 atn += π
+    if atn < 0 atn += 2*π
+    return atn
+}
+
+; reciprocal functions
+sub secant(float value) -> float { return 1.0 / cos(value) }
+sub csc(float value)    -> float { return 1.0 / sin(value) }
+sub cot(float value)    -> float { return 1.0 / tan(value) }
+
 sub ln(float value) -> float {
     %asm {{
         lda  #<value

compiler/res/prog8lib/shared_string_functions.p8

@@ -76,4 +76,44 @@ string {
         return s+cx16.r0L
     }
 
+    sub startswith(str st, str prefix) -> bool {
+        ubyte prefix_len = string.length(prefix)
+        ubyte str_len = string.length(st)
+        if prefix_len > str_len
+            return false
+        cx16.r9L = st[prefix_len]
+        st[prefix_len] = 0
+        cx16.r9H = string.compare(st, prefix) as ubyte
+        st[prefix_len] = cx16.r9L
+        return cx16.r9H==0
+    }
+
+    sub endswith(str st, str suffix) -> bool {
+        ubyte suffix_len = string.length(suffix)
+        ubyte str_len = string.length(st)
+        if suffix_len > str_len
+            return false
+        return string.compare(st + str_len - suffix_len, suffix) == 0
+    }
+
+    sub findstr(str haystack, str needle) -> ubyte {
+        ; searches for needle in haystack.
+        ; returns index in haystack where it first occurs, and Carry set,
+        ; or if needle doesn't occur in haystack it returns Carry clear and 255 (an invalid index.)
+        cx16.r2L = string.length(haystack)
+        cx16.r3L = string.length(needle)
+        if cx16.r3L <= cx16.r2L {
+            cx16.r2L = cx16.r2L-cx16.r3L+1
+            cx16.r3 = haystack
+            repeat cx16.r2L {
+                if string.startswith(cx16.r3, needle) {
+                    sys.set_carry()
+                    return cx16.r3-haystack as ubyte
+                }
+                cx16.r3++
+            }
+        }
+        sys.clear_carry()
+        return 255
+    }
 }

compiler/res/prog8lib/shared_textio_functions.p8

@@ -1,46 +0,0 @@
-txt {
-    ; the textio functions shared across compiler targets
-    %option merge, no_symbol_prefixing, ignore_unused
-
-    asmsub petscii2scr(ubyte petscii_char @A) -> ubyte @A {
-        ; -- convert petscii character to screencode
-        %asm {{
-            sta  P8ZP_SCRATCH_REG
-            lsr  a
-            lsr  a
-            lsr  a
-            lsr  a
-            lsr  a
-            tax
-            lda  _offsets,x
-            eor  P8ZP_SCRATCH_REG
-            rts
-_offsets    .byte  128, 0, 64, 32, 64, 192, 128, 128
-        }}
-    }
-
-    asmsub petscii2scr_str(str petscii_string @AY) {
-        ; -- convert petscii string to screencodes string
-        %asm {{
-            sta  P8ZP_SCRATCH_W1
-            sty  P8ZP_SCRATCH_W1+1
-            ldy  #0
--           lda  (P8ZP_SCRATCH_W1),y
-            beq  +
-            jsr  petscii2scr
-            sta  (P8ZP_SCRATCH_W1),y
-            iny
-            bne  -
-+           rts
-        }}
-}
-
-
-    sub print_bool(bool value) {
-        if value
-            txt.print("true")
-        else
-            txt.print("false")
-    }
-
-}

compiler/res/prog8lib/string.p8

@@ -129,7 +129,7 @@ _startloop	dey
 
     asmsub find(uword string @AY, ubyte character @X) -> ubyte @A, bool @Pc {
         ; Locates the first position of the given character in the string,
-        ; returns Carry set if found + index in A, or A=0 + Carry clear if not found.
+        ; returns Carry set if found + index in A, or Carry clear if not found (and A will be 255, an invalid index).
         %asm {{
 		; need to copy the the cx16 virtual registers to zeropage to make this run on C64...
 		sta  P8ZP_SCRATCH_W1
@@ -142,7 +142,7 @@ _startloop	dey
 		beq  _found
 		iny
 		bne  -
-_notfound	lda  #0
+_notfound	lda  #255
         clc
 		rts
 _found	tya
@@ -277,26 +277,6 @@ _done       rts
         }}
     }
 
-    sub startswith(str st, str prefix) -> bool {
-        ubyte prefix_len = length(prefix)
-        ubyte str_len = length(st)
-        if prefix_len > str_len
-            return false
-        cx16.r9L = st[prefix_len]
-        st[prefix_len] = 0
-        cx16.r9H = compare(st, prefix) as ubyte
-        st[prefix_len] = cx16.r9L
-        return cx16.r9H==0
-    }
-
-    sub endswith(str st, str suffix) -> bool {
-        ubyte suffix_len = length(suffix)
-        ubyte str_len = length(st)
-        if suffix_len > str_len
-            return false
-        return compare(st + str_len - suffix_len, suffix) == 0
-    }
-
     asmsub pattern_match(str string @AY, str pattern @R0) clobbers(Y) -> bool @A {
 		%asm {{
 ; pattern matching of a string.
@@ -366,7 +346,6 @@ fail    clc             ; yes, no match found, return with c=0
 		}}
 	}
 
-
     asmsub hash(str string @R0) -> ubyte @A {
         ; experimental 8 bit hashing function.
         ; hash(-1)=179;  clear carry;  hash(i) = ROL hash(i-1)  XOR  string[i]

compiler/res/prog8lib/virtual/diskio.p8

@@ -145,6 +145,12 @@ diskio {
         return "unknown"
     }
 
+    sub status_code() -> ubyte {
+        ; -- return status code instead of whole CBM-DOS status string. (in this case always 255, which means 'unable to return sensible value')
+        return 255
+    }
+
+
     sub save(uword filenameptr, uword start_address, uword savesize) -> bool {
         %ir {{
             load.b r65532,0

compiler/res/prog8lib/virtual/emudbg.p8

@@ -1,5 +1,5 @@
 ; Emulator debug interface. (reflecting the Commander X16 emudbg library module)
-; Docs: https://github.com/X16Community/x16-emulator#emulator-io-registers
+; Docs: https://github.com/X16Community/x16-emulator/tree/d52f118ce893fa24c4ca021a0b8de46cb80e5ccf#emulator-io-registers
 
 %import textio
 

compiler/res/prog8lib/virtual/floats.p8

@@ -80,6 +80,20 @@ sub atan(float value) -> float {
     }}
 }
 
+; two-argument arctangent that returns an angle in the correct quadrant
+; for the signs of x and y, normalized to the range [0, 2π]
+sub atan2(float y, float x) -> float {
+    float atn = atan(y / x)
+    if x < 0 atn += π
+    if atn < 0 atn += 2*π
+    return atn
+}
+
+; reciprocal functions
+sub secant(float value) -> float { return 1.0 / cos(value) }
+sub csc(float value)    -> float { return 1.0 / sin(value) }
+sub cot(float value)    -> float { return 1.0 / tan(value) }
+
 sub ln(float value) -> float {
     %ir {{
         loadm.f fr0,floats.ln.value

compiler/res/prog8lib/virtual/string.p8

@@ -55,7 +55,7 @@ string {
 
     sub find(str st, ubyte character) -> ubyte {
         ; Locates the first position of the given character in the string,
-        ; returns Carry set if found + index in A, or Carry clear if not found.
+        ; returns Carry set if found + index in A, or Carry clear if not found (and A will be 255, an invalid index).
         ubyte ix
         for ix in 0 to length(st)-1 {
             if st[ix]==character {
@@ -64,7 +64,7 @@ string {
             }
         }
         sys.clear_carry()
-        return 0
+        return 255
     }
 
     sub contains(str st, ubyte character) -> bool {
@@ -147,26 +147,6 @@ string {
         return char
     }
 
-    sub startswith(str st, str prefix) -> bool {
-        ubyte prefix_len = length(prefix)
-        ubyte str_len = length(st)
-        if prefix_len > str_len
-            return false
-        cx16.r9L = st[prefix_len]
-        st[prefix_len] = 0
-        cx16.r9H = compare(st, prefix) as ubyte
-        st[prefix_len] = cx16.r9L
-        return cx16.r9H==0
-    }
-
-    sub endswith(str st, str suffix) -> bool {
-        ubyte suffix_len = length(suffix)
-        ubyte str_len = length(st)
-        if suffix_len > str_len
-            return false
-        return compare(st + str_len - suffix_len, suffix) == 0
-    }
-
     sub hash(str st) -> ubyte {
         ; experimental 8 bit hashing function.
         ; hash(-1)=179;  hash(i) = ROL hash(i-1)  XOR  string[i]

compiler/res/prog8lib/virtual/textio.p8

@@ -28,6 +28,10 @@ sub  clear_screen() {
     }}
 }
 
+sub cls() {
+    clear_screen()
+}
+
 sub nl() {
     chrout('\n')
 }

compiler/src/prog8/CompilerMain.kt

@@ -43,6 +43,7 @@ private fun compileMain(args: Array<String>): Boolean {
     val startEmulator2 by cli.option(ArgType.Boolean, fullName = "emu2", description = "auto-start alternative emulator after successful compilation")
     val experimentalCodegen by cli.option(ArgType.Boolean, fullName = "expericodegen", description = "use experimental/alternative codegen")
     val dumpVariables by cli.option(ArgType.Boolean, fullName = "dumpvars", description = "print a dump of the variables in the program")
+    val dumpSymbols by cli.option(ArgType.Boolean, fullName = "dumpsymbols", description = "print a dump of the variable declarations and subroutine signatures")
     val dontWriteAssembly by cli.option(ArgType.Boolean, fullName = "noasm", description="don't create assembly code")
     val noStrictBool by cli.option(ArgType.Boolean, fullName = "nostrictbool", description = "allow implicit conversions between bool and bytes")
     val dontOptimize by cli.option(ArgType.Boolean, fullName = "noopt", description = "don't perform code optimizations")
@@ -56,6 +57,8 @@ private fun compileMain(args: Array<String>): Boolean {
     val printAst2 by cli.option(ArgType.Boolean, fullName = "printast2", description = "print out the intermediate AST that is used for code generation")
     val breakpointCpuInstruction by cli.option(ArgType.Choice(listOf("brk", "stp"), { it }), fullName = "breakinstr", description = "the CPU instruction to use as well for %breakpoint")
     val compilationTarget by cli.option(ArgType.String, fullName = "target", description = "target output of the compiler (one of '${C64Target.NAME}', '${C128Target.NAME}', '${Cx16Target.NAME}', '${AtariTarget.NAME}', '${PETTarget.NAME}', '${VMTarget.NAME}') (required)")
+    val bytes2float by cli.option(ArgType.String, fullName = "bytes2float", description = "convert a comma separated list of bytes from the target system to a float value. NOTE: you need to supply a target option too, and also still have to supply a dummy module file name as well!")
+    val float2bytes by cli.option(ArgType.String, fullName = "float2bytes", description = "convert floating point number to a list of bytes for the target system. NOTE: you need to supply a target option too, and also still have to supply a dummy module file name as well!")
     val startVm by cli.option(ArgType.Boolean, fullName = "vm", description = "load and run a .p8ir IR source file in the VM")
     val watchMode by cli.option(ArgType.Boolean, fullName = "watch", description = "continuous compilation mode (watch for file changes)")
     val varsGolden by cli.option(ArgType.Boolean, fullName = "varsgolden", description = "put uninitialized variables in 'golden ram' memory area instead of at the end of the program. On the cx16 target this is $0400-07ff. This is unavailable on other systems.")
@@ -99,6 +102,11 @@ private fun compileMain(args: Array<String>): Boolean {
         }
     }
 
+    if(bytes2float!=null)
+        return convertBytesToFloat(bytes2float!!, compilationTarget!!)
+    if(float2bytes!=null)
+        return convertFloatToBytes(float2bytes!!, compilationTarget!!)
+
     if(varsHighBank==0 && compilationTarget==Cx16Target.NAME) {
         System.err.println("On the Commander X16, HiRAM bank 0 is used by the kernal and can't be used.")
         return false
@@ -155,6 +163,7 @@ private fun compileMain(args: Array<String>): Boolean {
                     includeSourcelines == true,
                     experimentalCodegen == true,
                     dumpVariables == true,
+                    dumpSymbols == true,
                     varsHighBank,
                     varsGolden == true,
                     slabsHighBank,
@@ -235,6 +244,7 @@ private fun compileMain(args: Array<String>): Boolean {
                     includeSourcelines == true,
                     experimentalCodegen == true,
                     dumpVariables == true,
+                    dumpSymbols==true,
                     varsHighBank,
                     varsGolden == true,
                     slabsHighBank,
@@ -287,6 +297,23 @@ private fun compileMain(args: Array<String>): Boolean {
     return true
 }
 
+fun convertFloatToBytes(number: String, target: String): Boolean {
+    val tgt = getCompilationTargetByName(target)
+    val dbl = number.toDouble()
+    val bytes = tgt.machine.convertFloatToBytes(dbl)
+    print("$dbl in bytes on '$target': ")
+    println(bytes.joinToString(","))
+    return true
+}
+
+fun convertBytesToFloat(bytelist: String, target: String): Boolean {
+    val tgt = getCompilationTargetByName(target)
+    val bytes = bytelist.split(',').map { it.trim().toUByte() }
+    val number = tgt.machine.convertBytesToFloat(bytes)
+    println("floating point value on '$target': $number")
+    return true
+}
+
 private fun processSymbolDefs(symbolDefs: List<String>): Map<String, String>? {
     val result = mutableMapOf<String, String>()
     val defPattern = """(.+)\s*=\s*(.+)""".toRegex()

compiler/src/prog8/compiler/Compiler.kt

@@ -7,13 +7,17 @@ import prog8.ast.base.AstException
 import prog8.ast.expressions.Expression
 import prog8.ast.expressions.NumericLiteral
 import prog8.ast.printProgram
+import prog8.ast.printSymbols
 import prog8.ast.statements.Directive
 import prog8.code.SymbolTableMaker
 import prog8.code.ast.PtProgram
 import prog8.code.ast.printAst
 import prog8.code.core.*
 import prog8.code.optimize.optimizeIntermediateAst
-import prog8.code.target.*
+import prog8.code.target.AtariTarget
+import prog8.code.target.Cx16Target
+import prog8.code.target.VMTarget
+import prog8.code.target.getCompilationTargetByName
 import prog8.codegen.vm.VmCodeGen
 import prog8.compiler.astprocessing.*
 import prog8.optimizer.*
@@ -22,6 +26,7 @@ import java.nio.file.Path
 import kotlin.io.path.Path
 import kotlin.io.path.nameWithoutExtension
 import kotlin.math.round
+import kotlin.system.exitProcess
 import kotlin.system.measureTimeMillis
 
 
@@ -39,6 +44,7 @@ class CompilerArguments(val filepath: Path,
                         val includeSourcelines: Boolean,
                         val experimentalCodegen: Boolean,
                         val dumpVariables: Boolean,
+                        val dumpSymbols: Boolean,
                         val varsHighBank: Int?,
                         val varsGolden: Boolean,
                         val slabsHighBank: Int?,
@@ -57,21 +63,11 @@ class CompilerArguments(val filepath: Path,
 
 fun compileProgram(args: CompilerArguments): CompilationResult? {
 
-    val compTarget =
-        when(args.compilationTarget) {
-            C64Target.NAME -> C64Target()
-            C128Target.NAME -> C128Target()
-            Cx16Target.NAME -> Cx16Target()
-            PETTarget.NAME -> PETTarget()
-            AtariTarget.NAME -> AtariTarget()
-            VMTarget.NAME -> VMTarget()
-            else -> throw IllegalArgumentException("invalid compilation target")
-        }
-
+    val compTarget = getCompilationTargetByName(args.compilationTarget)
     var compilationOptions: CompilationOptions
     var ast: PtProgram? = null
     var resultingProgram: Program? = null
-    var importedFiles: List<Path> = emptyList()
+    var importedFiles: List<Path>
 
     try {
         val totalTime = measureTimeMillis {
@@ -86,6 +82,7 @@ fun compileProgram(args: CompilerArguments): CompilationResult? {
                 includeSourcelines = args.includeSourcelines
                 experimentalCodegen = args.experimentalCodegen
                 dumpVariables = args.dumpVariables
+                dumpSymbols = args.dumpSymbols
                 breakpointCpuInstruction = args.breakpointCpuInstruction
                 varsHighBank = args.varsHighBank
                 varsGolden = args.varsGolden
@@ -134,8 +131,12 @@ fun compileProgram(args: CompilerArguments): CompilationResult? {
                 }
 
                 val intermediateAst = IntermediateAstMaker(program, args.errors).transform()
-                optimizeIntermediateAst(intermediateAst, compilationOptions, args.errors)
-                args.errors.report()
+                val stMaker = SymbolTableMaker(intermediateAst, compilationOptions)
+                val symbolTable = stMaker.make()
+                if(compilationOptions.optimize) {
+                    optimizeIntermediateAst(intermediateAst, compilationOptions, symbolTable, args.errors)
+                    args.errors.report()
+                }
 
                 if(args.printAst2) {
                     println("\n*********** INTERMEDIATE AST *************")
@@ -398,6 +399,12 @@ fun determineCompilationOptions(program: Program, compTarget: ICompilationTarget
 
 private fun processAst(program: Program, errors: IErrorReporter, compilerOptions: CompilationOptions) {
     program.preprocessAst(errors, compilerOptions)
+
+    if(compilerOptions.dumpSymbols) {
+        printSymbols(program)
+        exitProcess(0)
+    }
+
     program.checkIdentifiers(errors, compilerOptions)
     errors.report()
     program.charLiteralsToUByteLiterals(compilerOptions.compTarget, errors)
@@ -444,8 +451,13 @@ private fun optimizeAst(program: Program, compilerOptions: CompilationOptions, e
             break
     }
     removeUnusedCode(program, errors, compilerOptions)
-    if(errors.noErrors())
+    if(errors.noErrors()) {
+        // last round of optimizations because constFold may have enabled more...
+        program.simplifyExpressions(errors, compilerOptions)
+        program.optimizeStatements(errors, functions, compilerOptions)
         program.constantFold(errors, compilerOptions) // because simplified statements and expressions can result in more constants that can be folded away
+    }
+
     errors.report()
 }
 
@@ -477,8 +489,10 @@ private fun createAssemblyAndAssemble(program: PtProgram,
     else
         throw NotImplementedError("no code generator for cpu ${compilerOptions.compTarget.machine.cpu}")
 
+    // need to make a new symboltable here to capture possible changes made by optimization steps performed earlier!
     val stMaker = SymbolTableMaker(program, compilerOptions)
     val symbolTable = stMaker.make()
+
     val assembly = asmgen.generate(program, symbolTable, compilerOptions, errors)
     errors.report()
 

compiler/src/prog8/compiler/astprocessing/AstChecker.kt

@@ -52,6 +52,8 @@ internal class AstChecker(private val program: Program,
 
     override fun visit(module: Module) {
         super.visit(module)
+        if(module.name.startsWith('_'))
+            errors.err("identifiers cannot start with an underscore", module.position)
         val directives = module.statements.filterIsInstance<Directive>().groupBy { it.directive }
         directives.filter { it.value.size > 1 }.forEach{ entry ->
             when(entry.key) {
@@ -62,6 +64,18 @@ internal class AstChecker(private val program: Program,
     }
 
     override fun visit(identifier: IdentifierReference) {
+        if(identifier.nameInSource.any { it.startsWith('_') }) {
+            errors.err("identifiers cannot start with an underscore", identifier.position)
+        }
+        if(identifier.nameInSource.any { it=="void" }) {
+            // 'void' as "identifier" is only allowed as part of a multi-assignment expression
+            if (!(identifier.nameInSource == listOf("void") && (identifier.parent as? AssignTarget)?.multi?.isNotEmpty() == true
+                || identifier.parent is AssignTarget && (identifier.parent.parent as? AssignTarget)?.multi?.isNotEmpty() == true)
+            ) {
+                errors.err("identifiers cannot contain the 'void' keyword", identifier.position)
+            }
+        }
+
         checkLongType(identifier)
         val stmt = identifier.targetStatement(program)
         if(stmt==null)
@@ -116,9 +130,7 @@ internal class AstChecker(private val program: Program,
         }
         if(expectedReturnValues.size==1 && returnStmt.value!=null) {
             val valueDt = returnStmt.value!!.inferType(program)
-            if(!valueDt.isKnown) {
-                errors.err("return value type mismatch or unknown symbol", returnStmt.value!!.position)
-            } else {
+            if(valueDt.isKnown) {
                 if (expectedReturnValues[0] != valueDt.getOr(DataType.UNDEFINED)) {
                     if(valueDt istype DataType.BOOL && expectedReturnValues[0] == DataType.UBYTE) {
                         // if the return value is a bool and the return type is ubyte, allow this. But give a warning.
@@ -266,6 +278,9 @@ internal class AstChecker(private val program: Program,
     }
 
     override fun visit(block: Block) {
+        if(block.name.startsWith('_'))
+            errors.err("identifiers cannot start with an underscore", block.position)
+
         val addr = block.address
         if(addr!=null && addr>65535u) {
             errors.err("block memory address must be valid integer 0..\$ffff", block.position)
@@ -295,6 +310,9 @@ internal class AstChecker(private val program: Program,
     }
 
     override fun visit(label: Label) {
+        if(label.name.startsWith('_'))
+            errors.err("identifiers cannot start with an underscore", label.position)
+
         // scope check
         if(label.parent !is Block && label.parent !is Subroutine && label.parent !is AnonymousScope) {
             errors.err("Labels can only be defined in the scope of a block, a loop body, or within another subroutine", label.position)
@@ -336,6 +354,9 @@ internal class AstChecker(private val program: Program,
     override fun visit(subroutine: Subroutine) {
         fun err(msg: String) = errors.err(msg, subroutine.position)
 
+        if(subroutine.name.startsWith('_'))
+            errors.err("identifiers cannot start with an underscore", subroutine.position)
+
         if(subroutine.name in BuiltinFunctions)
             err("cannot redefine a built-in function")
 
@@ -474,6 +495,9 @@ internal class AstChecker(private val program: Program,
         // Non-string and non-ubytearray Pass-by-reference datatypes can not occur as parameters to a subroutine directly
         // Instead, their reference (address) should be passed (as an UWORD).
         for(p in subroutine.parameters) {
+            if(p.name.startsWith('_'))
+                errors.err("identifiers cannot start with an underscore", p.position)
+
             if(p.type in PassByReferenceDatatypes && p.type !in listOf(DataType.STR, DataType.ARRAY_UB)) {
                 errors.err("this pass-by-reference type can't be used as a parameter type. Instead, use an uword to receive the address, or access the variable from the outer scope directly.", p.position)
             }
@@ -517,32 +541,71 @@ internal class AstChecker(private val program: Program,
     }
 
     override fun visit(assignment: Assignment) {
-        val targetDt = assignment.target.inferType(program)
-        val valueDt = assignment.value.inferType(program)
-        if(valueDt.isKnown && !(valueDt isAssignableTo targetDt)) {
-            if(targetDt.isIterable)
-                errors.err("cannot assign value to string or array", assignment.value.position)
-            else if(!(valueDt istype DataType.STR && targetDt istype DataType.UWORD)) {
-                if(targetDt.isUnknown) {
-                    if(assignment.target.identifier?.targetStatement(program)!=null)
-                        errors.err("target datatype is unknown", assignment.target.position)
-                    // otherwise, another error about missing symbol is already reported.
+        fun checkType(target: AssignTarget, value: Expression, augmentable: Boolean) {
+            val targetDt = target.inferType(program)
+            val valueDt = value.inferType(program)
+            if(valueDt.isKnown && !(valueDt isAssignableTo targetDt)) {
+                if(targetDt.isIterable)
+                    errors.err("cannot assign value to string or array", value.position)
+                else if(!(valueDt istype DataType.STR && targetDt istype DataType.UWORD)) {
+                    if(targetDt.isUnknown) {
+                        if(target.identifier?.targetStatement(program)!=null)
+                            errors.err("target datatype is unknown", target.position)
+                        // otherwise, another error about missing symbol is already reported.
+                    }
                 }
             }
+
+            if(value is TypecastExpression) {
+                if(augmentable && targetDt istype DataType.FLOAT)
+                    errors.err("typecasting a float value in-place makes no sense", value.position)
+            }
+
+            val numvalue = value.constValue(program)
+            if(numvalue!=null && targetDt.isKnown)
+                checkValueTypeAndRange(targetDt.getOr(DataType.UNDEFINED), numvalue)
+        }
+
+        if(assignment.target.multi==null) {
+            checkType(assignment.target, assignment.value, assignment.isAugmentable)
+        }
+
+
+        val fcall = assignment.value as? IFunctionCall
+        val fcallTarget = fcall?.target?.targetSubroutine(program)
+        if(assignment.target.multi!=null) {
+            checkMultiAssignment(assignment, fcall, fcallTarget)
+        } else if(fcallTarget!=null) {
+            if(fcallTarget.returntypes.size!=1) {
+                errors.err("number of assignment targets doesn't match number of return values from the subroutine", fcall.position)
+                return
+            }
         }
 
-        if(assignment.value is TypecastExpression) {
-            if(assignment.isAugmentable && targetDt istype DataType.FLOAT)
-                errors.err("typecasting a float value in-place makes no sense", assignment.value.position)
-        }
-
-        val numvalue = assignment.value.constValue(program)
-        if(numvalue!=null && targetDt.isKnown)
-            checkValueTypeAndRange(targetDt.getOr(DataType.UNDEFINED), numvalue)
-
         super.visit(assignment)
     }
 
+    private fun checkMultiAssignment(assignment: Assignment, fcall: IFunctionCall?, fcallTarget: Subroutine?) {
+        // multi-assign: check the number of assign targets vs. the number of return values of the subroutine
+        // also check the types of the variables vs the types of each return value
+        if(fcall==null || fcallTarget==null) {
+            errors.err("expected a function call with multiple return values", assignment.value.position)
+            return
+        }
+        val targets = assignment.target.multi!!
+        if(fcallTarget.returntypes.size!=targets.size) {
+            errors.err("number of assignment targets doesn't match number of return values from the subroutine", fcall.position)
+            return
+        }
+        fcallTarget.returntypes.zip(targets).withIndex().forEach { (index, p) ->
+            val (returnType, target) = p
+            val targetDt = target.inferType(program).getOr(DataType.UNDEFINED)
+            if (!target.void && !(returnType isAssignableTo targetDt))
+                errors.err("can't assign returnvalue #${index + 1} to corresponding target; $returnType vs $targetDt", target.position)
+        }
+    }
+
+
     override fun visit(assignTarget: AssignTarget) {
         super.visit(assignTarget)
 
@@ -552,6 +615,9 @@ internal class AstChecker(private val program: Program,
                 errors.err("address out of range", assignTarget.position)
         }
 
+        if(assignTarget.parent is AssignTarget)
+            return      // sub-target of a multi-assign is tested elsewhere
+
         val assignment = assignTarget.parent as Statement
         val targetIdentifier = assignTarget.identifier
         if (targetIdentifier != null) {
@@ -825,7 +891,7 @@ internal class AstChecker(private val program: Program,
                     err("invalid import directive, cannot import itself")
             }
             "%breakpoint" -> {
-                if(directive.parent !is INameScope || directive.parent is Module)
+                if(directive.parent !is INameScope && directive.parent !is AnonymousScope || directive.parent is Module)
                     err("this directive can't be used here")
                 if(directive.args.isNotEmpty())
                     err("invalid breakpoint directive, expected no arguments")
@@ -1174,25 +1240,6 @@ internal class AstChecker(private val program: Program,
         if(error!=null)
             errors.err(error.first, error.second)
 
-        // check the functions that return multiple returnvalues.
-        val stmt = functionCallExpr.target.targetStatement(program)
-        if (stmt is Subroutine) {
-            if (stmt.returntypes.size > 1) {
-                // Currently, it's only possible to handle ONE (or zero) return values from a subroutine.
-                // asmsub routines can have multiple return values, for instance in 2 different registers.
-                // It's not (yet) possible to handle these multiple return values because assignments
-                // are only to a single unique target at the same time.
-                //   EXCEPTION:
-                // if the asmsub returns multiple values and one of them is via a status register bit (such as carry),
-                // it *is* possible to handle them by just actually assigning the register value and
-                // dealing with the status bit as just being that, the status bit after the call.
-                val (returnRegisters, _) = stmt.asmReturnvaluesRegisters.partition { rr -> rr.registerOrPair != null }
-                if (returnRegisters.size>1) {
-                    errors.err("It's not possible to store the multiple result values of this asmsub call; you should use a small block of custom inline assembly for this.", functionCallExpr.position)
-                }
-            }
-        }
-
         // functions that don't return a value, can't be used in an expression or assignment
         if(targetStatement is Subroutine) {
             if(targetStatement.returntypes.isEmpty()) {
@@ -1264,7 +1311,14 @@ internal class AstChecker(private val program: Program,
                 if(funcName[0]=="setlsb" || funcName[0]=="setmsb") {
                     val firstArg = functionCallStatement.args[0]
                     if(firstArg !is IdentifierReference && firstArg !is ArrayIndexedExpression)
-                        errors.err("invalid argument to a in-place modifying function", functionCallStatement.args.first().position)
+                        errors.err("invalid argument to a in-place modifying function", firstArg.position)
+                } else if(funcName[0]=="divmod" || funcName[0].startsWith("divmod__")) {
+                    val thirdArg = functionCallStatement.args[2]
+                    val fourthArg = functionCallStatement.args[3]
+                    if(thirdArg !is IdentifierReference && thirdArg !is ArrayIndexedExpression)
+                        errors.err("invalid argument to a in-place modifying function", thirdArg.position)
+                    if(fourthArg !is IdentifierReference && fourthArg !is ArrayIndexedExpression)
+                        errors.err("invalid argument to a in-place modifying function", fourthArg.position)
                 } else {
                     if(functionCallStatement.args.any { it !is IdentifierReference && it !is ArrayIndexedExpression && it !is DirectMemoryRead })
                         errors.err("invalid argument to a in-place modifying function", functionCallStatement.args.first().position)

compiler/src/prog8/compiler/astprocessing/AstPreprocessor.kt

@@ -121,7 +121,7 @@ class AstPreprocessor(val program: Program,
                         // we need to handle multi-decl here too, the desugarer maybe has not processed it here yet...
                         if(decl.value!=null) {
                             decl.names.forEach { name ->
-                                val target = AssignTarget(IdentifierReference(listOf(name), decl.position), null, null, decl.position)
+                                val target = AssignTarget(IdentifierReference(listOf(name), decl.position), null, null, null, false, decl.position)
                                 val assign = Assignment(target.copy(), decl.value!!.copy(), AssignmentOrigin.VARINIT, decl.position)
                                 replacements.add(IAstModification.InsertAfter(decl, assign, scope))
                             }
@@ -137,7 +137,7 @@ class AstPreprocessor(val program: Program,
                     } else {
                         // handle declaration of a single variable
                         if(decl.value!=null && (decl.datatype in NumericDatatypes || decl.datatype==DataType.BOOL)) {
-                            val target = AssignTarget(IdentifierReference(listOf(decl.name), decl.position), null, null, decl.position)
+                            val target = AssignTarget(IdentifierReference(listOf(decl.name), decl.position), null, null, null, false, decl.position)
                             val assign = Assignment(target, decl.value!!, AssignmentOrigin.VARINIT, decl.position)
                             replacements.add(IAstModification.ReplaceNode(decl, assign, scope))
                             decl.value = null

compiler/src/prog8/compiler/astprocessing/CodeDesugarer.kt

@@ -183,7 +183,7 @@ _after:
         }
         if(functionCall.target.nameInSource==listOf("poke")) {
             // poke(a, v) is synonymous with @(a) = v
-            val tgt = AssignTarget(null, null, DirectMemoryWrite(functionCall.args[0], position), position)
+            val tgt = AssignTarget(null, null, DirectMemoryWrite(functionCall.args[0], position), null, false, position)
             val assign = Assignment(tgt, functionCall.args[1], AssignmentOrigin.OPTIMIZER, position)
             return listOf(IAstModification.ReplaceNode(functionCall as Node, assign, parent))
         }
@@ -222,7 +222,7 @@ _after:
             return if(parent is AssignTarget) {
                 // assignment to array
                 val memwrite = DirectMemoryWrite(address, arrayIndexedExpression.position)
-                val newtarget = AssignTarget(null, null, memwrite, arrayIndexedExpression.position)
+                val newtarget = AssignTarget(null, null, memwrite, null, false, arrayIndexedExpression.position)
                 listOf(IAstModification.ReplaceNode(parent, newtarget, parent.parent))
             } else {
                 // read from array

compiler/src/prog8/compiler/astprocessing/IntermediateAstMaker.kt

@@ -90,6 +90,7 @@ class IntermediateAstMaker(private val program: Program, private val errors: IEr
 
     private fun transform(srcAssign: Assignment): PtNode {
         if(srcAssign.isAugmentable) {
+            require(srcAssign.target.multi==null)
             val srcExpr = srcAssign.value
             val (operator: String, augmentedValue: Expression?) = when(srcExpr) {
                 is BinaryExpression -> {
@@ -136,20 +137,25 @@ class IntermediateAstMaker(private val program: Program, private val errors: IEr
         }
 
         val assign = PtAssignment(srcAssign.position)
-        assign.add(transform(srcAssign.target))
+        val multi = srcAssign.target.multi
+        if(multi==null) {
+            assign.add(transform(srcAssign.target))
+        } else {
+            multi.forEach { target -> assign.add(transform(target)) }
+        }
         assign.add(transformExpression(srcAssign.value))
         return assign
     }
 
     private fun transform(srcTarget: AssignTarget): PtAssignTarget {
-        val target = PtAssignTarget(srcTarget.position)
+        val target = PtAssignTarget(srcTarget.void, srcTarget.position)
         if(srcTarget.identifier!=null)
             target.add(transform(srcTarget.identifier!!))
         else if(srcTarget.arrayindexed!=null)
             target.add(transform(srcTarget.arrayindexed!!))
         else if(srcTarget.memoryAddress!=null)
             target.add(transform(srcTarget.memoryAddress!!))
-        else
+        else if(!srcTarget.void)
             throw FatalAstException("invalid AssignTarget")
         return target
     }
@@ -264,7 +270,7 @@ class IntermediateAstMaker(private val program: Program, private val errors: IEr
 
     private fun transform(srcCall: FunctionCallStatement): PtFunctionCall {
         val (target, type) = srcCall.target.targetNameAndType(program)
-        val call = PtFunctionCall(target,true, type, srcCall.position)
+        val call = PtFunctionCall(target, true, type, srcCall.position)
         for (arg in srcCall.args)
             call.add(transformExpression(arg))
         return call
@@ -272,11 +278,8 @@ class IntermediateAstMaker(private val program: Program, private val errors: IEr
 
     private fun transform(srcCall: FunctionCallExpression): PtFunctionCall {
         val (target, _) = srcCall.target.targetNameAndType(program)
-        val type = srcCall.inferType(program).getOrElse {
-            throw FatalAstException("unknown dt $srcCall")
-        }
-        val isVoid = type==DataType.UNDEFINED
-        val call = PtFunctionCall(target, isVoid, type, srcCall.position)
+        val iType = srcCall.inferType(program)
+        val call = PtFunctionCall(target, iType.isUnknown && srcCall.parent !is Assignment, iType.getOrElse { DataType.UNDEFINED }, srcCall.position)
         for (arg in srcCall.args)
             call.add(transformExpression(arg))
         return call

compiler/src/prog8/compiler/astprocessing/LiteralsToAutoVars.kt

@@ -74,6 +74,13 @@ internal class LiteralsToAutoVars(private val program: Program, private val erro
 
     override fun after(decl: VarDecl, parent: Node): Iterable<IAstModification> {
         if(decl.names.size>1) {
+
+            val fcallTarget = (decl.value as? IFunctionCall)?.target?.targetSubroutine(program)
+            if(fcallTarget!=null && fcallTarget.returntypes.size>1) {
+                errors.err("ambiguous multi-variable initialization. Use separate variable declaration and assignment(s) instead.", decl.value!!.position)
+                return noModifications
+            }
+
             // note: the desugaring of a multi-variable vardecl has to be done here
             // and not in CodeDesugarer, that one is too late (identifiers can't be found otherwise)
             if(decl.datatype !in NumericDatatypesWithBoolean)

compiler/src/prog8/compiler/astprocessing/StatementReorderer.kt

@@ -60,7 +60,8 @@ internal class StatementReorderer(
                                 // Add assignment to initialize with zero
                                 // Note: for block-level vars, this will introduce assignments in the block scope. These have to be dealt with correctly later.
                                 val identifier = IdentifierReference(listOf(decl.name), decl.position)
-                                val assignzero = Assignment(AssignTarget(identifier, null, null, decl.position), decl.zeroElementValue(), AssignmentOrigin.VARINIT, decl.position)
+                                val assignzero = Assignment(AssignTarget(identifier, null, null, null, false, decl.position),
+                                    decl.zeroElementValue(), AssignmentOrigin.VARINIT, decl.position)
                                 return listOf(IAstModification.InsertAfter(
                                     decl, assignzero, parent as IStatementContainer
                                 ))
@@ -72,7 +73,8 @@ internal class StatementReorderer(
                         // So basically consider 'ubyte xx=99' as a short form for 'ubyte xx; xx=99'
                         val pos = decl.value!!.position
                         val identifier = IdentifierReference(listOf(decl.name), pos)
-                        val assign = Assignment(AssignTarget(identifier, null, null, pos), decl.value!!, AssignmentOrigin.VARINIT, pos)
+                        val assign = Assignment(AssignTarget(identifier, null, null, null, false, pos),
+                            decl.value!!, AssignmentOrigin.VARINIT, pos)
                         decl.value = null
                         return listOf(IAstModification.InsertAfter(
                             decl, assign, parent as IStatementContainer
@@ -90,7 +92,8 @@ internal class StatementReorderer(
                     if(target!=null && target.isArray) {
                         val pos = decl.value!!.position
                         val identifier = IdentifierReference(listOf(decl.name), pos)
-                        val assign = Assignment(AssignTarget(identifier, null, null, pos), decl.value!!, AssignmentOrigin.VARINIT, pos)
+                        val assign = Assignment(AssignTarget(identifier, null, null, null, false, pos),
+                            decl.value!!, AssignmentOrigin.VARINIT, pos)
                         decl.value = null
                         return listOf(IAstModification.InsertAfter(
                             decl, assign, parent as IStatementContainer
@@ -214,8 +217,12 @@ internal class StatementReorderer(
             }
         }
 
-        if(valueType.isString && (targetType istype DataType.STR || targetType istype DataType.ARRAY_B || targetType istype DataType.ARRAY_UB))
-            return copyStringValue(assignment)
+        if(!assignment.isAugmentable) {
+            if (valueType.isString && (targetType istype DataType.STR || targetType istype DataType.ARRAY_B || targetType istype DataType.ARRAY_UB)) {
+                // replace string assignment by a call to stringcopy
+                return copyStringValue(assignment)
+            }
+        }
 
         return noModifications
     }

compiler/src/prog8/compiler/astprocessing/TypecastsAdder.kt

@@ -463,14 +463,16 @@ class TypecastsAdder(val program: Program, val options: CompilationOptions, val
 
     override fun after(whenChoice: WhenChoice, parent: Node): Iterable<IAstModification> {
         val conditionDt = (whenChoice.parent as When).condition.inferType(program)
-        val values = whenChoice.values
-        values?.toTypedArray()?.withIndex()?.forEach { (index, value) ->
-            val valueDt = value.inferType(program)
-            if(valueDt!=conditionDt) {
-                val castedValue = value.typecastTo(conditionDt.getOr(DataType.UNDEFINED), valueDt.getOr(DataType.UNDEFINED), true)
-                if(castedValue.first) {
-                    castedValue.second.linkParents(whenChoice)
-                    values[index] = castedValue.second
+        if(conditionDt.isKnown) {
+            val values = whenChoice.values
+            values?.toTypedArray()?.withIndex()?.forEach { (index, value) ->
+                val valueDt = value.inferType(program)
+                if(valueDt!=conditionDt) {
+                    val castedValue = value.typecastTo(conditionDt.getOr(DataType.UNDEFINED), valueDt.getOr(DataType.UNDEFINED), true)
+                    if(castedValue.first) {
+                        castedValue.second.linkParents(whenChoice)
+                        values[index] = castedValue.second
+                    }
                 }
             }
         }

compiler/src/prog8/compiler/astprocessing/VariousCleanups.kt

@@ -156,28 +156,6 @@ internal class VariousCleanups(val program: Program, val errors: IErrorReporter,
             }
         }
 
-/* TODO: is this really no longer needed in boolean branch?
-        if(expr.operator in LogicalOperators) {
-            // remove redundant !=0 comparisons from logical expressions such as:   a!=0 xor b --> a xor b  (only for byte operands!)
-            val leftExpr = expr.left as? BinaryExpression
-            if(leftExpr != null &&
-                leftExpr.operator == "!=" &&
-                !leftExpr.left.isSimple &&
-                leftExpr.left.inferType(program).isBytes &&
-                leftExpr.right.constValue(program)?.number == 0.0) {
-                return listOf(IAstModification.ReplaceNode(leftExpr, leftExpr.left, expr))
-            }
-            val rightExpr = expr.right as? BinaryExpression
-            if(rightExpr != null &&
-                rightExpr.operator == "!=" &&
-                !rightExpr.left.isSimple &&
-                rightExpr.left.inferType(program).isBytes &&
-                rightExpr.right.constValue(program)?.number == 0.0) {
-                return listOf(IAstModification.ReplaceNode(rightExpr, rightExpr.left, expr))
-            }
-        }
-*/
-
         return noModifications
     }
 

compiler/src/prog8/compiler/astprocessing/VerifyFunctionArgTypes.kt

@@ -114,12 +114,7 @@ internal class VerifyFunctionArgTypes(val program: Program, val options: Compila
                 }
                 if(target.isAsmSubroutine) {
                     if(target.asmReturnvaluesRegisters.size>1) {
-                        // multiple return values will NOT work inside an expression.
-                        // they MIGHT work in a regular assignment or just a function call statement.
-                        //   EXCEPTION:
-                        // if the asmsub returns multiple values and one of them is via a status register bit (such as carry),
-                        // it *is* possible to handle them by just actually assigning the register value and
-                        // dealing with the status bit as just being that, the status bit after the call.
+                        // multiple return values will NOT work inside an expression. Use an assignment first.
                         val parent = if(call is Statement) call.parent else if(call is Expression) call.parent else null
                         if (call !is FunctionCallStatement) {
                             val checkParent =
@@ -128,8 +123,7 @@ internal class VerifyFunctionArgTypes(val program: Program, val options: Compila
                                 else
                                     parent
                             if (checkParent !is Assignment && checkParent !is VarDecl) {
-                                val (returnRegisters, _) = target.asmReturnvaluesRegisters.partition { rr -> rr.registerOrPair != null }
-                                if (returnRegisters.size>1) {
+                                if (target.asmReturnvaluesRegisters.size>1) {
                                     return Pair("can't use subroutine call that returns multiple return values here", call.position)
                                 }
                             }

compiler/test/TestBuiltinFunctions.kt

@@ -2,6 +2,7 @@ package prog8tests
 
 import io.kotest.core.spec.style.FunSpec
 import io.kotest.matchers.shouldBe
+import io.kotest.matchers.shouldNotBe
 import prog8.ast.expressions.NumericLiteral
 import prog8.ast.statements.Assignment
 import prog8.ast.statements.FunctionCallStatement
@@ -98,5 +99,19 @@ main {
         (a4.value as NumericLiteral).number shouldBe 200*256+100
         (a5.args[0] as NumericLiteral).number shouldBe 6.0
     }
+
+    test("divmod target args should be treated as variables that are written") {
+        val src="""
+main {
+    ubyte c
+    ubyte l
+
+    sub start() {
+        divmod(99, 10, c, l)
+    }
+}"""
+
+        compileText(Cx16Target(), true, src, writeAssembly = true) shouldNotBe null
+    }
 })
 

compiler/test/TestCompilerOnExamples.kt

@@ -33,6 +33,7 @@ private fun compileTheThing(filepath: Path, optimize: Boolean, target: ICompilat
         includeSourcelines = false,
         experimentalCodegen = false,
         dumpVariables = false,
+        dumpSymbols = false,
         varsHighBank = null,
         varsGolden = false,
         slabsHighBank = null,
@@ -107,6 +108,7 @@ class TestCompilerOnExamplesCx16: FunSpec({
             "vtui/testvtui",
             "pcmaudio/play-adpcm",
             "pcmaudio/stream-wav",
+            "pcmaudio/vumeter",
             "sprites/dragon",
             "sprites/dragons",
             "amiga",
@@ -115,6 +117,7 @@ class TestCompilerOnExamplesCx16: FunSpec({
             "bdmusic",
             "bobs",
             "bubbleuniverse",
+            "charsets",
             "circles",
             "cobramk3-gfx",
             "colorbars",

compiler/test/TestCompilerOptionLibdirs.kt

@@ -31,6 +31,7 @@ class TestCompilerOptionSourcedirs: FunSpec({
             includeSourcelines = false,
             experimentalCodegen = false,
             dumpVariables = false,
+            dumpSymbols = false,
             varsHighBank = null,
             varsGolden = false,
             slabsHighBank = null,

compiler/test/TestImportedModulesOrderAndOptions.kt

@@ -42,7 +42,7 @@ main {
                 "textio",
                 "syslib",
                 "conv",
-                "shared_textio_functions",
+                "shared_cbm_textio_functions",
                 "floats",
                 "shared_floats_functions",
                 "math",
@@ -101,7 +101,7 @@ main {
                 listOf(
                     internedStringsModuleName,
                     filenameBase,
-                    "textio", "syslib", "conv", "shared_textio_functions", "floats", "shared_floats_functions", "math", "prog8_lib"
+                    "textio", "syslib", "conv", "shared_cbm_textio_functions", "floats", "shared_floats_functions", "math", "prog8_lib"
                 )
         }
         options.floats shouldBe true

compiler/test/TestMemory.kt

@@ -36,49 +36,49 @@ class TestMemory: FunSpec({
     test("assignment target not in mapped IO space C64") {
 
         var memexpr = NumericLiteral.optimalInteger(0x0002, Position.DUMMY)
-        var target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        var target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         var assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe false
 
         memexpr = NumericLiteral.optimalInteger(0x1000, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe false
 
         memexpr = NumericLiteral.optimalInteger(0x9fff, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe false
 
         memexpr = NumericLiteral.optimalInteger(0xa000, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe false
 
         memexpr = NumericLiteral.optimalInteger(0xc000, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe false
 
         memexpr = NumericLiteral.optimalInteger(0xcfff, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe false
 
         memexpr = NumericLiteral.optimalInteger(0xeeee, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe false
 
         memexpr = NumericLiteral.optimalInteger(0xffff, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe false
@@ -87,25 +87,25 @@ class TestMemory: FunSpec({
     test("assign target in mapped IO space C64") {
 
         var memexpr = NumericLiteral.optimalInteger(0x0000, Position.DUMMY)
-        var target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        var target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         var assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe true
 
         memexpr = NumericLiteral.optimalInteger(0x0001, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe true
 
         memexpr = NumericLiteral.optimalInteger(0xd000, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe true
 
         memexpr = NumericLiteral.optimalInteger(0xdfff, Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe true
@@ -114,7 +114,7 @@ class TestMemory: FunSpec({
     fun createTestProgramForMemoryRefViaVar(address: UInt, vartype: VarDeclType): AssignTarget {
         val decl = VarDecl(vartype, VarDeclOrigin.USERCODE, DataType.BYTE, ZeropageWish.DONTCARE, null, "address", emptyList(), NumericLiteral.optimalInteger(address, Position.DUMMY), false, false, Position.DUMMY)
         val memexpr = IdentifierReference(listOf("address"), Position.DUMMY)
-        val target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        val target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         val assignment = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(decl, assignment))
         return target
@@ -137,13 +137,13 @@ class TestMemory: FunSpec({
 
     test("memory expression mapped to IO memory on C64") {
         var memexpr = PrefixExpression("+", NumericLiteral.optimalInteger(0x1000, Position.DUMMY), Position.DUMMY)
-        var target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        var target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         var assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe false
 
         memexpr = PrefixExpression("+", NumericLiteral.optimalInteger(0xd020, Position.DUMMY), Position.DUMMY)
-        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), Position.DUMMY)
+        target = AssignTarget(null, null, DirectMemoryWrite(memexpr, Position.DUMMY), null, false, Position.DUMMY)
         assign = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         wrapWithProgram(listOf(assign))
         target.isIOAddress(c64target.machine) shouldBe true
@@ -151,7 +151,7 @@ class TestMemory: FunSpec({
 
     test("regular variable not in mapped IO ram on C64") {
         val decl = VarDecl(VarDeclType.VAR, VarDeclOrigin.USERCODE, DataType.BYTE, ZeropageWish.DONTCARE, null, "address", emptyList(), null, false, false, Position.DUMMY)
-        val target = AssignTarget(IdentifierReference(listOf("address"), Position.DUMMY), null, null, Position.DUMMY)
+        val target = AssignTarget(IdentifierReference(listOf("address"), Position.DUMMY), null, null, null, false, Position.DUMMY)
         val assignment = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         val subroutine = Subroutine("test", mutableListOf(), mutableListOf(), emptyList(), emptyList(), emptySet(), null, false, false, false, mutableListOf(decl, assignment), Position.DUMMY)
         val module = Module(mutableListOf(subroutine), Position.DUMMY, SourceCode.Generated("test"))
@@ -163,7 +163,7 @@ class TestMemory: FunSpec({
     test("memory mapped variable not in mapped IO ram on C64") {
         val address = 0x1000u
         val decl = VarDecl(VarDeclType.MEMORY, VarDeclOrigin.USERCODE, DataType.UBYTE, ZeropageWish.DONTCARE, null, "address", emptyList(), NumericLiteral.optimalInteger(address, Position.DUMMY), false, false, Position.DUMMY)
-        val target = AssignTarget(IdentifierReference(listOf("address"), Position.DUMMY), null, null, Position.DUMMY)
+        val target = AssignTarget(IdentifierReference(listOf("address"), Position.DUMMY), null, null, null, false, Position.DUMMY)
         val assignment = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         val subroutine = Subroutine("test", mutableListOf(), mutableListOf(), emptyList(), emptyList(), emptySet(), null, false, false, false, mutableListOf(decl, assignment), Position.DUMMY)
         val module = Module(mutableListOf(subroutine), Position.DUMMY, SourceCode.Generated("test"))
@@ -175,7 +175,7 @@ class TestMemory: FunSpec({
     test("memory mapped variable in mapped IO ram on C64") {
         val address = 0xd020u
         val decl = VarDecl(VarDeclType.MEMORY, VarDeclOrigin.USERCODE, DataType.UBYTE, ZeropageWish.DONTCARE, null, "address", emptyList(), NumericLiteral.optimalInteger(address, Position.DUMMY), false, false, Position.DUMMY)
-        val target = AssignTarget(IdentifierReference(listOf("address"), Position.DUMMY), null, null, Position.DUMMY)
+        val target = AssignTarget(IdentifierReference(listOf("address"), Position.DUMMY), null, null, null, false, Position.DUMMY)
         val assignment = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         val subroutine = Subroutine("test", mutableListOf(), mutableListOf(), emptyList(), emptyList(), emptySet(), null, false, false, false, mutableListOf(decl, assignment), Position.DUMMY)
         val module = Module(mutableListOf(subroutine), Position.DUMMY, SourceCode.Generated("test"))
@@ -187,7 +187,7 @@ class TestMemory: FunSpec({
     test("array not in mapped IO ram") {
         val decl = VarDecl(VarDeclType.VAR, VarDeclOrigin.USERCODE, DataType.ARRAY_UB, ZeropageWish.DONTCARE, null, "address", emptyList(), null, false, false, Position.DUMMY)
         val arrayindexed = ArrayIndexedExpression(IdentifierReference(listOf("address"), Position.DUMMY), ArrayIndex(NumericLiteral.optimalInteger(1, Position.DUMMY), Position.DUMMY), Position.DUMMY)
-        val target = AssignTarget(null, arrayindexed, null, Position.DUMMY)
+        val target = AssignTarget(null, arrayindexed, null, null, false, Position.DUMMY)
         val assignment = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         val subroutine = Subroutine("test", mutableListOf(), mutableListOf(), emptyList(), emptyList(), emptySet(), null, false, false, false, mutableListOf(decl, assignment), Position.DUMMY)
         val module = Module(mutableListOf(subroutine), Position.DUMMY, SourceCode.Generated("test"))
@@ -200,7 +200,7 @@ class TestMemory: FunSpec({
         val address = 0x1000u
         val decl = VarDecl(VarDeclType.MEMORY, VarDeclOrigin.USERCODE, DataType.ARRAY_UB, ZeropageWish.DONTCARE, null, "address", emptyList(), NumericLiteral.optimalInteger(address, Position.DUMMY), false, false, Position.DUMMY)
         val arrayindexed = ArrayIndexedExpression(IdentifierReference(listOf("address"), Position.DUMMY), ArrayIndex(NumericLiteral.optimalInteger(1, Position.DUMMY), Position.DUMMY), Position.DUMMY)
-        val target = AssignTarget(null, arrayindexed, null, Position.DUMMY)
+        val target = AssignTarget(null, arrayindexed, null, null, false, Position.DUMMY)
         val assignment = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         val subroutine = Subroutine("test", mutableListOf(), mutableListOf(), emptyList(), emptyList(), emptySet(), null, false, false, false, mutableListOf(decl, assignment), Position.DUMMY)
         val module = Module(mutableListOf(subroutine), Position.DUMMY, SourceCode.Generated("test"))
@@ -213,7 +213,7 @@ class TestMemory: FunSpec({
         val address = 0xd800u
         val decl = VarDecl(VarDeclType.MEMORY, VarDeclOrigin.USERCODE, DataType.ARRAY_UB, ZeropageWish.DONTCARE, null, "address", emptyList(), NumericLiteral.optimalInteger(address, Position.DUMMY), false, false, Position.DUMMY)
         val arrayindexed = ArrayIndexedExpression(IdentifierReference(listOf("address"), Position.DUMMY), ArrayIndex(NumericLiteral.optimalInteger(1, Position.DUMMY), Position.DUMMY), Position.DUMMY)
-        val target = AssignTarget(null, arrayindexed, null, Position.DUMMY)
+        val target = AssignTarget(null, arrayindexed, null, null, false, Position.DUMMY)
         val assignment = Assignment(target, NumericLiteral.optimalInteger(0, Position.DUMMY), AssignmentOrigin.USERCODE, Position.DUMMY)
         val subroutine = Subroutine("test", mutableListOf(), mutableListOf(), emptyList(), emptyList(), emptySet(), null, false, false, false, mutableListOf(decl, assignment), Position.DUMMY)
         val module = Module(mutableListOf(subroutine), Position.DUMMY, SourceCode.Generated("test"))

compiler/test/TestOptimization.kt

@@ -12,6 +12,9 @@ import prog8.ast.ParentSentinel
 import prog8.ast.Program
 import prog8.ast.expressions.*
 import prog8.ast.statements.*
+import prog8.code.ast.PtAssignTarget
+import prog8.code.ast.PtAssignment
+import prog8.code.ast.PtFunctionCall
 import prog8.code.core.DataType
 import prog8.code.core.Position
 import prog8.code.target.C64Target
@@ -790,6 +793,18 @@ main {
         (expr.right as? NumericLiteral)?.number shouldBe 10.0
     }
 
+    test("var to const inside typecast") {
+        val src="""
+main {
+    uword variable  ; will be made a const
+
+    sub start() {
+        cx16.r0 = msb(variable) as uword
+    }
+}"""
+        compileText(VMTarget(), true, src, writeAssembly = false) shouldNotBe null
+    }
+
     test("De Morgan's laws") {
         val src="""
 main {
@@ -935,4 +950,113 @@ main {
         funcarg3.target.nameInSource shouldBe listOf("msb")
         funcarg3.args.single() shouldBe instanceOf<BinaryExpression>()
     }
+
+    test("no operand swap on logical expressions with shortcircuit evaluation") {
+        val src="""
+%import diskio
+%zeropage basicsafe
+%option no_sysinit
+
+main {
+    str scanline_buf = "?"* 20
+
+    sub start() {
+        if diskio.f_open("test.prg") and diskio.f_read(scanline_buf, 2)==2
+            cx16.r0++
+
+        if diskio.f_open("test.prg") or diskio.f_read(scanline_buf, 2)==2
+            cx16.r0++
+
+        if diskio.f_open("test.prg") xor diskio.f_read(scanline_buf, 2)==2
+            cx16.r0++
+    }
+}"""
+        val result = compileText(Cx16Target(), true, src, writeAssembly = false)!!
+        val st = result.compilerAst.entrypoint.statements
+        st.size shouldBe 3
+        val ifCond1 = (st[0] as IfElse).condition as BinaryExpression
+        val ifCond2 = (st[1] as IfElse).condition as BinaryExpression
+        val ifCond3 = (st[2] as IfElse).condition as BinaryExpression
+        (ifCond1.left as FunctionCallExpression).target.nameInSource shouldBe listOf("diskio", "f_open")
+        (ifCond2.left as FunctionCallExpression).target.nameInSource shouldBe listOf("diskio", "f_open")
+        (ifCond3.left as FunctionCallExpression).target.nameInSource shouldBe listOf("diskio", "f_open")
+        val right1 = ifCond1.right as BinaryExpression
+        val right2 = ifCond2.right as BinaryExpression
+        val right3 = ifCond3.right as BinaryExpression
+        (right1.left as FunctionCallExpression).target.nameInSource shouldBe listOf("diskio", "f_read")
+        (right2.left as FunctionCallExpression).target.nameInSource shouldBe listOf("diskio", "f_read")
+        (right3.left as FunctionCallExpression).target.nameInSource shouldBe listOf("diskio", "f_read")
+    }
+
+    test("eliminate same target register assignments") {
+        val src="""
+%zeropage basicsafe
+%option no_sysinit
+
+main {
+    romsub ${'$'}2000 = func1() clobbers(X) -> ubyte @A, word @R0, byte @R1
+    romsub ${'$'}3000 = func2() clobbers(X) -> ubyte @A, uword @R0, uword @R1
+    romsub ${'$'}4000 = func3() clobbers(X) -> ubyte @R0
+
+    sub start() {
+        bool flag
+        void cbm.GETIN()
+        flag, cx16.r1L = cbm.GETIN()
+        void, cx16.r0s, cx16.r1sL = func1()
+        void, cx16.r2, cx16.r1 = func2()
+        cx16.r0L = func3()
+        cx16.r0H = func3()
+    }
+}"""
+        val result = compileText(C64Target(), true, src, writeAssembly = true)!!
+        val st = result.codegenAst!!.entrypoint()!!.children
+        st.size shouldBe 9
+        (st[2] as PtFunctionCall).name shouldBe "cbm.GETIN"
+        (st[2] as PtFunctionCall).void shouldBe true
+        val a1 = st[3] as PtAssignment
+        (a1.value as PtFunctionCall).name shouldBe "cbm.GETIN"
+        a1.multiTarget shouldBe true
+        a1.children.size shouldBe 3
+        (a1.children[0] as PtAssignTarget).void shouldBe false
+        (a1.children[0] as PtAssignTarget).identifier!!.name shouldBe "p8b_main.p8s_start.p8v_flag"
+        (a1.children[1] as PtAssignTarget).void shouldBe false
+        (a1.children[1] as PtAssignTarget).identifier!!.name shouldBe "cx16.r1L"
+
+        (st[4] as PtFunctionCall).name shouldBe "p8b_main.p8s_func1"
+        (st[4] as PtFunctionCall).void shouldBe true
+        val a2 = st[5] as PtAssignment
+        (a2.value as PtFunctionCall).name shouldBe "p8b_main.p8s_func2"
+        a2.multiTarget shouldBe true
+        a2.children.size shouldBe 4
+        (a2.children[0] as PtAssignTarget).void shouldBe true
+        (a2.children[1] as PtAssignTarget).void shouldBe false
+        (a2.children[1] as PtAssignTarget).identifier!!.name shouldBe "cx16.r2"
+        (a2.children[2] as PtAssignTarget).void shouldBe true
+
+        (st[6] as PtFunctionCall).name shouldBe "p8b_main.p8s_func3"
+        (st[6] as PtFunctionCall).void shouldBe true
+        val a3 = st[7] as PtAssignment
+        (a3.value as PtFunctionCall).name shouldBe "p8b_main.p8s_func3"
+        a3.multiTarget shouldBe false
+        a3.children.size shouldBe 2
+        (a3.children[0] as PtAssignTarget).void shouldBe false
+        (a3.children[0] as PtAssignTarget).identifier!!.name shouldBe "cx16.r0H"
+    }
+
+    test("symbol table correct in asmgen after earlier optimization steps") {
+        val src="""
+main {
+    sub start() {
+        uword @shared bulletRef, enemyRef
+        const ubyte BD_Y = 10
+        const ubyte EN_Y = 11
+
+        if bulletRef[BD_Y] == enemyRef[EN_Y] or bulletRef[BD_Y] == enemyRef[EN_Y] + 1 {
+            cx16.r0++
+        }
+    }
+}"""
+        compileText(VMTarget(), true, src, writeAssembly = true) shouldNotBe null
+        compileText(C64Target(), true, src, writeAssembly = true) shouldNotBe null
+    }
 })

compiler/test/TestStringEncodings.kt

@@ -16,9 +16,9 @@ import prog8.code.core.unescape
 import prog8.code.target.C64Target
 import prog8.code.target.Cx16Target
 import prog8.code.target.Encoder
-import prog8.code.target.cbm.AtasciiEncoding
-import prog8.code.target.cbm.IsoEncoding
-import prog8.code.target.cbm.PetsciiEncoding
+import prog8.code.target.encodings.AtasciiEncoding
+import prog8.code.target.encodings.IsoEncoding
+import prog8.code.target.encodings.PetsciiEncoding
 import prog8tests.helpers.ErrorReporterForTests
 import prog8tests.helpers.compileText
 

compiler/test/TestTypecasts.kt

@@ -11,7 +11,6 @@ import prog8.ast.expressions.*
 import prog8.ast.printProgram
 import prog8.ast.statements.Assignment
 import prog8.ast.statements.IfElse
-import prog8.ast.statements.VarDecl
 import prog8.code.ast.PtAsmSub
 import prog8.code.ast.PtSub
 import prog8.code.core.DataType
@@ -826,4 +825,18 @@ main {
 }"""
         compileText(C64Target(), true, src, writeAssembly = true) shouldNotBe null
     }
+
+    test("bool to word cast") {
+        val src="""
+main {
+    sub start() {
+        bool @shared flag, flag2
+        cx16.r0L = (flag and flag2) as ubyte
+        cx16.r0 = (flag and flag2) as uword
+    }
+}"""
+
+        compileText(VMTarget(), false, src, writeAssembly = true) shouldNotBe null
+        compileText(C64Target(), false, src, writeAssembly = true) shouldNotBe null
+    }
 })