Now comments on global variables are no longer destroyed. However some comments appear twice!

src/main/java/dk/camelot64/kickc/model/ControlFlowGraph.java

@@ -9,7 +9,7 @@ import dk.camelot64.kickc.model.values.LabelRef;
 import dk.camelot64.kickc.model.values.ProcedureRef;
 import dk.camelot64.kickc.model.values.ScopeRef;
 import dk.camelot64.kickc.model.values.SymbolRef;
-import dk.camelot64.kickc.passes.Pass2ConstantIdentification;
+import dk.camelot64.kickc.passes.utils.ProcedureUtils;
 
 import java.io.Serializable;
 import java.util.*;
@@ -45,7 +45,7 @@ public class ControlFlowGraph implements Serializable {
    }
 
    public List<ControlFlowBlock> getAllBlocks() {
-         return blocks;
+      return blocks;
    }
 
    public void setAllBlocks(List<ControlFlowBlock> blocks) {
@@ -109,8 +109,8 @@ public class ControlFlowGraph implements Serializable {
    }
 
    public void setSequence(List<LabelRef> sequence) {
-      if(sequence.size()!=blocks.size()) {
-         throw new CompileError("ERROR! Sequence does not contain all blocks from the program. Sequence: "+sequence.size()+" Blocks: "+blocks.size());
+      if(sequence.size() != blocks.size()) {
+         throw new CompileError("ERROR! Sequence does not contain all blocks from the program. Sequence: " + sequence.size() + " Blocks: " + blocks.size());
       }
       this.sequence = sequence;
       ArrayList<ControlFlowBlock> seqBlocks = new ArrayList<>();
@@ -144,29 +144,26 @@ public class ControlFlowGraph implements Serializable {
    /**
     * Get all blocks that are program entry points.
     * This is the start-block and any blocks referenced by the address-off operator (&)
+    *
     * @param program The program
     * @return All entry-point blocks
     */
    public List<ControlFlowBlock> getEntryPointBlocks(Program program) {
       List<ControlFlowBlock> entryPointBlocks = new ArrayList<>();
       for(Procedure procedure : program.getScope().getAllProcedures(true)) {
-         if(Pass2ConstantIdentification.isAddressOfUsed(procedure.getRef(), program) || Procedure.CallingConvention.STACK_CALL.equals(procedure.getCallingConvention())) {
+         if(ProcedureUtils.isEntrypoint(procedure.getRef(), program) || Procedure.CallingConvention.STACK_CALL.equals(procedure.getCallingConvention())) {
             // Address-of is used on the procedure
             Label procedureLabel = procedure.getLabel();
             ControlFlowBlock procedureBlock = getBlock(procedureLabel.getRef());
             entryPointBlocks.add(procedureBlock);
          }
       }
-      final ProcedureRef startProcedure = program.getStartProcedure();
-      final ControlFlowBlock startBlock = getBlock(startProcedure.getLabelRef());
-      if(startBlock != null && !entryPointBlocks.contains(startBlock)) {
-         entryPointBlocks.add(startBlock);
-      }
       return entryPointBlocks;
    }
 
    /**
     * Get all called procedures in the graph
+    *
     * @return All called procedures
     */
    public Set<ProcedureRef> getAllCalledProcedures() {
@@ -219,7 +216,7 @@ public class ControlFlowGraph implements Serializable {
    public Statement getStatementByIndex(int statementIdx) {
       for(ControlFlowBlock block : getAllBlocks()) {
          for(Statement statement : block.getStatements()) {
-            if(statement.getIndex()!=null && statementIdx == statement.getIndex()) {
+            if(statement.getIndex() != null && statementIdx == statement.getIndex()) {
                return statement;
             }
          }
@@ -245,20 +242,22 @@ public class ControlFlowGraph implements Serializable {
 
    /**
     * Get information about the size of the program
+    *
     * @return Size information
     */
    public String getSizeInfo() {
       StringBuilder sizeInfo = new StringBuilder();
-      sizeInfo.append("SIZE blocks "+ getAllBlocks().size()).append("\n");
+      sizeInfo.append("SIZE blocks " + getAllBlocks().size()).append("\n");
       int numStmt = getAllBlocks().stream().mapToInt(block -> block.getStatements().size()).sum();
-      sizeInfo.append("SIZE statements "+ numStmt).append("\n");
-      int numPhiVars =  getAllBlocks().stream().mapToInt(value -> value.getStatements().stream().mapToInt(value1 -> (value1 instanceof StatementPhiBlock)?((StatementPhiBlock) value1).getPhiVariables().size():0).sum()).sum();
-      sizeInfo.append("SIZE phi variables "+ numPhiVars).append("\n");
+      sizeInfo.append("SIZE statements " + numStmt).append("\n");
+      int numPhiVars = getAllBlocks().stream().mapToInt(value -> value.getStatements().stream().mapToInt(value1 -> (value1 instanceof StatementPhiBlock) ? ((StatementPhiBlock) value1).getPhiVariables().size() : 0).sum()).sum();
+      sizeInfo.append("SIZE phi variables " + numPhiVars).append("\n");
       return sizeInfo.toString();
    }
 
    /**
     * Get all statements executed between two statements (none of these are included in the result)
+    *
     * @param from The statement to start at
     * @param to The statement to end at
     * @return All statements executed between the two passed statements
@@ -272,6 +271,7 @@ public class ControlFlowGraph implements Serializable {
 
    /**
     * Fill the between collection with all statements executed between two statements (none of these are included in the result)
+    *
     * @param from The statement to start at
     * @param to The statement to end at
     * @param between The between collection

src/main/java/dk/camelot64/kickc/passes/Pass0GenerateStatementSequence.java

@@ -985,8 +985,11 @@ public class Pass0GenerateStatementSequence extends KickCParserBaseVisitor<Objec
                // Add comments to constant
                variable.setComments(ensureUnusedComments(declComments));
             } else if(!variable.isKindConstant() && !isStructMember) {
-               Statement initStmt = new StatementAssignment(variable.getVariableRef(), initValue, true, statementSource, ensureUnusedComments(declComments));
+               final List<Comment> comments = ensureUnusedComments(declComments);
+               Statement initStmt = new StatementAssignment(variable.getVariableRef(), initValue, true, statementSource, comments);
                addStatement(initStmt);
+               // Add comments to variable
+               variable.setComments(comments);
             }
             if(initializer != null)
                PrePostModifierHandler.addPostModifiers(this, initializer, statementSource);

src/main/java/dk/camelot64/kickc/passes/Pass1EliminateUncalledProcedures.java

@@ -4,6 +4,7 @@ import dk.camelot64.kickc.model.ControlFlowBlock;
 import dk.camelot64.kickc.model.Program;
 import dk.camelot64.kickc.model.symbols.Procedure;
 import dk.camelot64.kickc.model.values.ProcedureRef;
+import dk.camelot64.kickc.passes.utils.ProcedureUtils;
 
 import java.util.Collection;
 import java.util.LinkedHashSet;
@@ -20,17 +21,15 @@ public class Pass1EliminateUncalledProcedures extends Pass1Base {
    @Override
    public boolean step() {
       Set<ProcedureRef> calledProcedures = getGraph().getAllCalledProcedures();
-      calledProcedures.add(getProgram().getStartProcedure());
+      //calledProcedures.add(getProgram().getStartProcedure());
 
       Set<ProcedureRef> unusedProcedures = new LinkedHashSet<>();
       Collection<Procedure> allProcedures = getProgram().getScope().getAllProcedures(true);
-      for(Procedure procedure : allProcedures) {
-         // TODO Also look at kickasm/asm uses! (Maybe also look at some directive like "export" )
-         if(!calledProcedures.contains(procedure.getRef())  && !Pass2ConstantIdentification.isAddressOfUsed(procedure.getRef(), getProgram())) {
-            // The procedure is not used - mark for removal!
-            unusedProcedures.add(procedure.getRef());
-         }
-      }
+      for(Procedure procedure : allProcedures)
+         if(!ProcedureUtils.isEntrypoint(procedure.getRef(), getProgram()))
+            if(!calledProcedures.contains(procedure.getRef()))
+               // The procedure is not used - mark for removal!
+               unusedProcedures.add(procedure.getRef());
 
       for(ProcedureRef unusedProcedure : unusedProcedures) {
          removeProcedure(getProgram(), unusedProcedure);
@@ -41,6 +40,7 @@ public class Pass1EliminateUncalledProcedures extends Pass1Base {
 
    /**
     * Removed a procedure from the program (the symbol in the symbol table and all blocks in the control flow graph)
+    *
     * @param program The program
     * @param procedureRef The procedure to be removed
     */

src/main/java/dk/camelot64/kickc/passes/Pass1GenerateSingleStaticAssignmentForm.java

@@ -258,7 +258,6 @@ public class Pass1GenerateSingleStaticAssignmentForm extends Pass1Base {
       if(symbol instanceof Procedure) {
          Procedure procedure = (Procedure) symbol;
          if(procedure.getInterruptType() != null || Pass2ConstantIdentification.isAddressOfUsed(procedure.getRef(), program)) {
-            final ControlFlowBlock startBlock = program.getGraph().getBlock(program.getStartProcedure().getLabelRef());
             // Find all root-level predecessors to the main block
             ControlFlowBlock mainBlock = program.getGraph().getBlock(new LabelRef(SymbolRef.MAIN_PROC_NAME));
             List<ControlFlowBlock> mainPredecessors = program.getGraph().getPredecessors(mainBlock);

src/main/java/dk/camelot64/kickc/passes/Pass4CodeGeneration.java

@@ -87,7 +87,6 @@ public class Pass4CodeGeneration {
       linkScriptBody = linkScriptBody.replace("%_O", outputFileName.toLowerCase());
       linkScriptBody = linkScriptBody.replace("%^O", outputFileName.toUpperCase());
       String entryName = program.getStartProcedure().getFullName();
-      ;
       linkScriptBody = linkScriptBody.replace("%E", entryName);
       Number programPc = program.getProgramPc();
       if(programPc == null) programPc = 0x080d;

src/main/java/dk/camelot64/kickc/passes/Pass4MemoryCoalesce.java

@@ -7,6 +7,7 @@ import dk.camelot64.kickc.model.values.ProcedureRef;
 import dk.camelot64.kickc.model.values.ScopeRef;
 import dk.camelot64.kickc.model.values.SymbolRef;
 import dk.camelot64.kickc.model.values.VariableRef;
+import dk.camelot64.kickc.passes.utils.ProcedureUtils;
 
 import java.util.ArrayList;
 import java.util.Collection;
@@ -32,22 +33,9 @@ public abstract class Pass4MemoryCoalesce extends Pass2Base {
    public static Collection<ScopeRef> getThreadHeads(Program program) {
       ArrayList<ScopeRef> threadHeads = new ArrayList<>();
       Collection<Procedure> procedures = program.getScope().getAllProcedures(true);
-      for(Procedure procedure : procedures) {
-         if(procedure.getRef().equals(program.getStartProcedure())) {
-            // TODO: Handles main() correctly?
+      for(Procedure procedure : procedures)
+         if(ProcedureUtils.isEntrypoint(procedure.getRef(), program))
             threadHeads.add(procedure.getRef());
-            continue;
-         }
-         if(Pass2ConstantIdentification.isAddressOfUsed(procedure.getRef(), program)) {
-            threadHeads.add(procedure.getRef());
-            continue;
-         }
-         if(procedure.getInterruptType()!=null) {
-            // TODO: Handles Interrupts correctly?
-            threadHeads.add(procedure.getRef());
-            continue;
-         }
-      }
       return threadHeads;
    }
 

src/main/java/dk/camelot64/kickc/passes/PassNEliminateEmptyProcedure.java

@@ -10,6 +10,7 @@ import dk.camelot64.kickc.model.symbols.Scope;
 import dk.camelot64.kickc.model.symbols.Symbol;
 import dk.camelot64.kickc.model.symbols.Variable;
 import dk.camelot64.kickc.model.values.ProcedureRef;
+import dk.camelot64.kickc.model.values.SymbolRef;
 import dk.camelot64.kickc.model.values.SymbolVariableRef;
 
 import java.util.Collection;
@@ -30,7 +31,8 @@ public class PassNEliminateEmptyProcedure extends Pass2SsaOptimization {
       boolean optimized = false;
       for(Procedure procedure : allProcedures) {
          if(hasEmptyBody(procedure.getRef())) {
-            if(!hasExternalUsages(procedure.getRef(), getProgram())) {
+            if(!hasExternalUsages(procedure.getRef(), getProgram()) && !SymbolRef.MAIN_PROC_NAME.equals(procedure.getLabel().getLocalName())) {
+               // TODO: Entry point procedures include isAddressOfUsed!
                // Remove all calls
                removeAllCalls(procedure.getRef());
                // Remove the procedure

src/main/java/dk/camelot64/kickc/passes/calcs/PassNCalcLiveRangesEffectiveCallPaths.java

@@ -11,6 +11,7 @@ import dk.camelot64.kickc.model.symbols.Variable;
 import dk.camelot64.kickc.model.values.*;
 import dk.camelot64.kickc.passes.Pass2AliasElimination;
 import dk.camelot64.kickc.passes.Pass2ConstantIdentification;
+import dk.camelot64.kickc.passes.utils.ProcedureUtils;
 
 import java.util.*;
 
@@ -69,8 +70,7 @@ public class PassNCalcLiveRangesEffectiveCallPaths extends PassNCalcBase<LiveRan
       LiveRangeVariablesEffectiveCallPaths.CallPaths callPaths = procedureCallPaths.get(procedureRef);
       if(callPaths == null) {
          callPaths = new LiveRangeVariablesEffectiveCallPaths.CallPaths(procedureRef);
-
-         if(procedure.getInterruptType()!=null || Pass2ConstantIdentification.isAddressOfUsed(procedureRef, getProgram()) || procedureRef.equals(getProgram().getStartProcedure())) {
+         if(ProcedureUtils.isEntrypoint(procedureRef, getProgram())) {
             // Interrupt is called outside procedure scope - create initial call-path.
             ArrayList<CallGraph.CallBlock.Call> rootPath = new ArrayList<>();
             ArrayList<VariableRef> rootAlive = new ArrayList<>();

src/main/java/dk/camelot64/kickc/passes/utils/ProcedureUtils.java

@@ -0,0 +1,34 @@
+package dk.camelot64.kickc.passes.utils;
+
+import dk.camelot64.kickc.model.Program;
+import dk.camelot64.kickc.model.symbols.Procedure;
+import dk.camelot64.kickc.model.values.ProcedureRef;
+import dk.camelot64.kickc.passes.Pass2ConstantIdentification;
+
+public class ProcedureUtils {
+
+   /**
+    * Determines if a procedure is an entry point.
+    * Entry points are:
+    * - the starting procedure (typically main())
+    * - interrupts
+    * - any procedure where address-of is used in the code.
+    * TODO: Should all stack call procedures be included (probably no!)
+    * TODO: Also look at kickasm/asm uses! (Maybe also look at some directive like "export" )
+    *
+    * @param procedure The procedure to examine
+    * @param program The program
+    * @return true if the procedure is an entry point
+    */
+   public static boolean isEntrypoint(ProcedureRef procedureRef, Program program) {
+      if(procedureRef.equals(program.getStartProcedure()))
+         return true;
+      final Procedure procedure = program.getScope().getProcedure(procedureRef);
+      if(procedure.getInterruptType()!=null)
+         return true;
+      if(Pass2ConstantIdentification.isAddressOfUsed(procedureRef, program))
+         return true;
+      return false;
+   }
+
+}

src/test/java/dk/camelot64/kickc/test/TestPrograms.java

@@ -42,6 +42,21 @@ public class TestPrograms {
    public TestPrograms() {
    }
 
+   @Test
+   public void testEmptyFunction2() throws IOException, URISyntaxException {
+      compileAndCompare("empty-function-2.c");
+   }
+
+   @Test
+   public void testEmptyFunction1() throws IOException, URISyntaxException {
+      compileAndCompare("empty-function-1.c");
+   }
+
+   @Test
+   public void testEmptyFunction0() throws IOException, URISyntaxException {
+      compileAndCompare("empty-function-0.c");
+   }
+
    @Test
    public void testStaticInitCode2() throws IOException, URISyntaxException {
       compileAndCompare("static-init-code-2.c");
@@ -1240,7 +1255,7 @@ public class TestPrograms {
 
    @Test
    public void testInnerIncrementProblem() throws IOException, URISyntaxException {
-      compileAndCompare("inner-increment-problem.c");
+      compileAndCompare("inner-increment-problem.c", log());
    }
 
    @Test
@@ -3045,6 +3060,11 @@ public class TestPrograms {
       compileAndCompare("string-const-consolidation.c");
    }
 
+   @Test
+   public void testCommentsGlobalInit() throws IOException, URISyntaxException {
+      compileAndCompare("test-comments-global.c");
+   }
+
    @Test
    public void testCommentsBlock() throws IOException, URISyntaxException {
       compileAndCompare("test-comments-block.c");

src/test/kc/empty-function-0.c

@@ -0,0 +1,14 @@
+// Test removal of empty function
+
+char v;
+
+char * const SCREEN = 0x0400;
+
+void main() {
+    set();
+    SCREEN[0] = v;
+}
+
+void set() {
+    v = 7;
+}

src/test/kc/empty-function-1.c

@@ -0,0 +1,11 @@
+// Test removal of empty function
+
+char * const SCREEN = 0x0400;
+
+void main() {
+    empty();
+    SCREEN[0] = 'x';
+}
+
+void empty() {
+}

src/test/kc/empty-function-2.c

@@ -0,0 +1,9 @@
+// Test removal of empty function
+// main() should not be removed!
+
+void main() {
+    empty();
+}
+
+void empty() {
+}

src/test/kc/struct-ptr-25.c

@@ -1,10 +1,14 @@
-char *fileCur;
-char *fileTop;
-char *filesEnd;
-char *file;
+char *fileCur = 0x1005;
+char *fileTop = 0x1010;
+char *filesEnd = 0x1010;
+char *file = 0x1000;
+
+char * const SCREEN = 0x0400;
 
 void main(void) {
     if (fileTop == filesEnd) --fileTop;
     if (file <= fileCur) --fileCur;
     if (fileCur < fileTop) fileCur = fileTop;
+
+    SCREEN[0] = *fileCur;
 }

src/test/kc/test-comments-global.c

@@ -0,0 +1,13 @@
+// Tests that global variables with initializer gets their comments
+
+// The screen (should become a var-comment in ASM)
+__ma char * screen = 0x0400;
+
+// The program entry point
+void main() {
+    // Put 'a' in screen
+    *screen = 'a';
+    screen++;
+    // Put another 'a' in screen
+    *screen = 'a';
+}

src/test/ref/address-2.asm

@@ -1,16 +1,17 @@
 // Test that address vars are turned into load/store and located at hardcoded addresses
 // Hard-coded mainmem-page address - global variable
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
   .label i = $2000
-__bbegin:
-  // i = 3
-  lda #3
-  sta i
-  jsr main
-  rts
+_start: {
+    // i = 3
+    lda #3
+    sta i
+    jsr main
+    rts
+}
 main: {
   __b1:
     // while(i<7)

src/test/ref/address-5.asm

@@ -1,16 +1,17 @@
 // Test declaring a variable as at a hard-coded address
 // zero-page hard-coded address parameter
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
   .label idx = 3
-__bbegin:
-  // idx
-  lda #0
-  sta.z idx
-  jsr main
-  rts
+_start: {
+    // idx
+    lda #0
+    sta.z idx
+    jsr main
+    rts
+}
 main: {
     // print('c')
     lda #'c'

src/test/ref/address-6.asm

@@ -1,16 +1,17 @@
 // Test declaring a variable as at a hard-coded address
 // mainmem-page hard-coded address parameter
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
   .label idx = $3000
-__bbegin:
-  // idx
-  lda #0
-  sta idx
-  jsr main
-  rts
+_start: {
+    // idx
+    lda #0
+    sta idx
+    jsr main
+    rts
+}
 main: {
     // print('c')
     lda #'c'

src/test/ref/address-of-2.asm

@@ -1,18 +1,20 @@
 // Test address-of by assigning the affected variable in multiple ways
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label val = 2
-__bbegin:
-  // val = 0
-  lda #0
-  sta.z val
-  jsr main
-  rts
+_start: {
+    // val = 0
+    lda #0
+    sta.z val
+    jsr main
+    rts
+}
 main: {
     .label SCREEN1 = $400
     .label SCREEN2 = SCREEN1+$28
     // Use address-of - hereafter all versions of val must be in the same memory
+    // Use address-of - hereafter all versions of val must be in the same memory
     .label ptr = val
     // SCREEN1[idx] = val
     lda.z val

src/test/ref/arrays-init-kasm-1.asm

@@ -0,0 +1,18 @@
+// Test initializing array using KickAssembler
+// Place array at hardcoded address
+.pc = $801 "Basic"
+:BasicUpstart(main)
+.pc = $80d "Program"
+  .label SCREEN = $400
+main: {
+    // SCREEN[0] = SINTAB[0]
+    lda SINTAB
+    sta SCREEN
+    // }
+    rts
+}
+.pc = $1000 "SINTAB"
+// Sinus table at an absolute address in memory
+SINTAB:
+.fill 256, 128 + 128*sin(i*2*PI/256)
+

src/test/ref/asm-uses-0.asm

@@ -3,12 +3,6 @@
 :BasicUpstart(main)
 .pc = $80d "Program"
   .label BG_COLOR = $d020
-main: {
-    // asm
-    jsr init
-    // }
-    rts
-}
 // Function only used inside the inline asm
 init: {
     // *BG_COLOR = 0
@@ -17,3 +11,9 @@ init: {
     // }
     rts
 }
+main: {
+    // asm
+    jsr init
+    // }
+    rts
+}

src/test/ref/atoi-1.asm

@@ -1,6 +1,6 @@
 // Test atoi()
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .const LIGHT_BLUE = $e
   .const OFFSET_STRUCT_PRINTF_BUFFER_NUMBER_DIGITS = 1
@@ -9,33 +9,34 @@
   .label COLORRAM = $d800
   // Default address of screen character matrix
   .label DEFAULT_SCREEN = $400
-  .label conio_cursor_x = 9
-  .label conio_cursor_y = $a
-  .label conio_line_text = $b
-  .label conio_line_color = $d
-__bbegin:
-  // conio_cursor_x = 0
   // The number of bytes on the screen
   // The current cursor x-position
-  lda #0
-  sta.z conio_cursor_x
-  // conio_cursor_y = 0
+  .label conio_cursor_x = 9
   // The current cursor y-position
-  sta.z conio_cursor_y
-  // conio_line_text = CONIO_SCREEN_TEXT
+  .label conio_cursor_y = $a
   // The current text cursor line start
-  lda #<DEFAULT_SCREEN
-  sta.z conio_line_text
-  lda #>DEFAULT_SCREEN
-  sta.z conio_line_text+1
-  // conio_line_color = CONIO_SCREEN_COLORS
+  .label conio_line_text = $b
   // The current color cursor line start
-  lda #<COLORRAM
-  sta.z conio_line_color
-  lda #>COLORRAM
-  sta.z conio_line_color+1
-  jsr main
-  rts
+  .label conio_line_color = $d
+_start: {
+    // conio_cursor_x = 0
+    lda #0
+    sta.z conio_cursor_x
+    // conio_cursor_y = 0
+    sta.z conio_cursor_y
+    // conio_line_text = CONIO_SCREEN_TEXT
+    lda #<DEFAULT_SCREEN
+    sta.z conio_line_text
+    lda #>DEFAULT_SCREEN
+    sta.z conio_line_text+1
+    // conio_line_color = CONIO_SCREEN_COLORS
+    lda #<COLORRAM
+    sta.z conio_line_color
+    lda #>COLORRAM
+    sta.z conio_line_color+1
+    jsr main
+    rts
+}
 main: {
     // clrscr()
     jsr clrscr

src/test/ref/bitmap-line-anim-1.asm

@@ -12,7 +12,7 @@
   .label VIC_MEMORY = $d018
   .label SCREEN = $400
   .label BITMAP = $2000
-  .label next = 2
+  .label next = 5
 main: {
     // *BORDER_COLOR = 0
     lda #0
@@ -89,13 +89,13 @@ bitmap_line: {
     jsr bitmap_line_xdyi
     rts
 }
-// bitmap_line_xdyi(byte zp(3) x, byte zp(4) y, byte zp($b) x1, byte zp(6) xd)
+// bitmap_line_xdyi(byte zp(2) x, byte zp(3) y, byte zp(6) x1, byte zp($b) xd)
 bitmap_line_xdyi: {
-    .label x1 = $b
-    .label xd = 6
-    .label x = 3
-    .label e = 5
-    .label y = 4
+    .label x1 = 6
+    .label xd = $b
+    .label x = 2
+    .label e = 4
+    .label y = 3
     lda #bitmap_line.y1>>1
     sta.z e
     lda #bitmap_line.y0
@@ -136,6 +136,7 @@ bitmap_line_xdyi: {
 }
 // bitmap_plot(byte register(X) x, byte register(Y) y)
 bitmap_plot: {
+    // Needs unsigned int arrays arranged as two underlying char arrays to allow char* plotter_x = plot_x[x]; - and eventually - char* plotter = plot_x[x] + plot_y[y];
     .label plotter_x = 7
     .label plotter_y = 9
     .label plotter = 7
@@ -166,12 +167,12 @@ bitmap_plot: {
     // }
     rts
 }
-// bitmap_line_ydxi(byte zp(4) y, byte zp(3) x, byte zp(6) xd)
+// bitmap_line_ydxi(byte zp(3) y, byte zp(2) x, byte zp(6) xd)
 bitmap_line_ydxi: {
     .label xd = 6
-    .label e = 5
-    .label y = 4
-    .label x = 3
+    .label e = 4
+    .label y = 3
+    .label x = 2
     // e = xd>>1
     lda.z xd
     lsr
@@ -210,12 +211,12 @@ bitmap_line_ydxi: {
     // }
     rts
 }
-// bitmap_line_xdyd(byte zp(3) x, byte zp(4) y, byte zp(6) xd)
+// bitmap_line_xdyd(byte zp(2) x, byte zp(3) y, byte zp(6) xd)
 bitmap_line_xdyd: {
-    .label x = 3
+    .label x = 2
     .label xd = 6
-    .label e = 5
-    .label y = 4
+    .label e = 4
+    .label y = 3
     lda #bitmap_line.y1>>1
     sta.z e
     lda #bitmap_line.y1
@@ -250,12 +251,12 @@ bitmap_line_xdyd: {
     // }
     rts
 }
-// bitmap_line_ydxd(byte zp(4) y, byte zp(3) x, byte zp(6) xd)
+// bitmap_line_ydxd(byte zp(3) y, byte zp(2) x, byte zp(6) xd)
 bitmap_line_ydxd: {
     .label xd = 6
-    .label e = 5
-    .label y = 4
-    .label x = 3
+    .label e = 4
+    .label y = 3
+    .label x = 2
     // e = xd>>1
     lda.z xd
     lsr
@@ -325,7 +326,7 @@ init_screen: {
 // Clear all graphics on the bitmap
 bitmap_clear: {
     .label bitmap = 7
-    .label y = 6
+    .label y = 5
     // bitmap = (char*) { bitmap_plot_xhi[0], bitmap_plot_xlo[0] }
     lda bitmap_plot_xlo
     sta.z bitmap

src/test/ref/bitmap-line-anim-2.asm

@@ -14,7 +14,7 @@
   .label VIC_MEMORY = $d018
   .label SCREEN = $400
   .label BITMAP = $2000
-  .label next = 2
+  .label next = $a
 main: {
     // *BORDER_COLOR = 0
     lda #0
@@ -55,7 +55,7 @@ main: {
     jmp __b1
 }
 // Draw a line on the bitmap using bresenhams algorithm
-// bitmap_line(word zp(2) x2)
+// bitmap_line(word zp($a) x2)
 bitmap_line: {
     .const x1 = 0
     .const y1 = 0
@@ -64,11 +64,13 @@ bitmap_line: {
     .label dy = 8
     .label sx = $12
     .label sy = 6
+    // X is the driver
     .label e1 = 4
-    .label e = $a
-    .label y = $e
-    .label x = $c
-    .label x2 = 2
+    // Y is the driver
+    .label e = $c
+    .label y = 2
+    .label x = $e
+    .label x2 = $a
     // abs_u16(x2-x1)
     lda.z x2
     sta.z abs_u16.w
@@ -287,11 +289,11 @@ bitmap_line: {
     rts
 }
 // Plot a single dot in the bitmap
-// bitmap_plot(word zp($c) x, byte register(X) y)
+// bitmap_plot(word zp($e) x, byte register(X) y)
 bitmap_plot: {
     .label __0 = $16
     .label plotter = $14
-    .label x = $c
+    .label x = $e
     // plotter = (char*) { bitmap_plot_yhi[y], bitmap_plot_ylo[y] }
     lda bitmap_plot_yhi,x
     sta.z plotter+1

src/test/ref/bitmap-plot-0.asm

@@ -4,7 +4,7 @@
 // The MOS 6526 Complex Interface Adapter (CIA)
 // http://archive.6502.org/datasheets/mos_6526_cia_recreated.pdf
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   // Value that disables all CIA interrupts when stored to the CIA Interrupt registers
   .const CIA_INTERRUPT_CLEAR = $7f
@@ -40,14 +40,15 @@
   .label HARDWARE_IRQ = $fffe
   .label BITMAP = $2000
   .label SCREEN = $400
-  .label frame_cnt = 8
-__bbegin:
-  // frame_cnt = 1
   // Counts frames - updated by the IRQ
-  lda #1
-  sta.z frame_cnt
-  jsr main
-  rts
+  .label frame_cnt = 8
+_start: {
+    // frame_cnt = 1
+    lda #1
+    sta.z frame_cnt
+    jsr main
+    rts
+}
 main: {
     .const toD0181_return = (>(SCREEN&$3fff)*4)|(>BITMAP)/4&$f
     .label x = 2

src/test/ref/bitmap-plot-1.asm

@@ -4,7 +4,7 @@
 // The MOS 6526 Complex Interface Adapter (CIA)
 // http://archive.6502.org/datasheets/mos_6526_cia_recreated.pdf
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   // Value that disables all CIA interrupts when stored to the CIA Interrupt registers
   .const CIA_INTERRUPT_CLEAR = $7f
@@ -47,16 +47,17 @@
   .label HARDWARE_IRQ = $fffe
   .label BITMAP = $2000
   .label SCREEN = $400
+  // Counts frames - updated by the IRQ
   .label frame_cnt = $16
   // Remainder after unsigned 16-bit division
-  .label rem16u = $14
-__bbegin:
-  // frame_cnt = 1
-  // Counts frames - updated by the IRQ
-  lda #1
-  sta.z frame_cnt
-  jsr main
-  rts
+  .label rem16u = $2e
+_start: {
+    // frame_cnt = 1
+    lda #1
+    sta.z frame_cnt
+    jsr main
+    rts
+}
 main: {
     .const toD0181_return = (>(SCREEN&$3fff)*4)|(>BITMAP)/4&$f
     .label __6 = 8
@@ -596,10 +597,14 @@ sin16s_gen2: {
     .label wavelength = $200
     .label __6 = 8
     .label __8 = $24
+    // ampl is always positive so shifting left does not alter the sign
+    // u[4.28] step = PI*2/wavelength
     .label step = $20
     .label sintab = $19
     // u[4.28]
     // Iterate over the table
+    // u[4.28]
+    // Iterate over the table
     .label x = $c
     .label i = $17
     // div32u16u(PI2_u4f28, wavelength)
@@ -700,14 +705,23 @@ sin16s: {
     .label __4 = $26
     .label x = $10
     .label return = $1b
+    // sinx = x - x^3/6 + x5/128;
     .label x1 = $2a
+    // u[1.15]
     .label x2 = $14
+    // u[2.14] x^2
     .label x3 = $14
+    // u[2.14] x^3
     .label x3_6 = $2c
+    // u[1.15] x^3/6;
     .label usinx = $1b
+    // u[1.15] x - x^3/6
     .label x4 = $14
+    // u[3.13] x^4
     .label x5 = $2c
+    // u[4.12] x^5
     .label x5_128 = $2c
+    // u[1.15] (first bit is always zero)
     .label sinx = $1b
     // if(x >= PI_u4f28 )
     lda.z x+3
@@ -962,6 +976,10 @@ div32u16u: {
     lda.z divr16u.return+1
     sta.z quotient_hi+1
     // divr16u(<dividend, divisor, rem16u)
+    lda.z rem16u
+    sta.z divr16u.rem
+    lda.z rem16u+1
+    sta.z divr16u.rem+1
     lda #<PI2_u4f28&$ffff
     sta.z divr16u.dividend
     lda #>PI2_u4f28&$ffff
@@ -1045,6 +1063,10 @@ divr16u: {
     cpx #$10
     bne __b1
     // rem16u = rem
+    lda.z rem
+    sta.z rem16u
+    lda.z rem+1
+    sta.z rem16u+1
     // }
     rts
 }

src/test/ref/bitmap-plot-2.asm

@@ -4,7 +4,7 @@
 // The MOS 6526 Complex Interface Adapter (CIA)
 // http://archive.6502.org/datasheets/mos_6526_cia_recreated.pdf
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   // Value that disables all CIA interrupts when stored to the CIA Interrupt registers
   .const CIA_INTERRUPT_CLEAR = $7f
@@ -48,30 +48,31 @@
   .label HARDWARE_IRQ = $fffe
   .label BITMAP = $2000
   .label SCREEN = $400
-  .label frame_cnt = $1b
-  // Remainder after unsigned 16-bit division
-  .label rem16u = $19
-__bbegin:
-  // frame_cnt = 1
   // Counts frames - updated by the IRQ
-  lda #1
-  sta.z frame_cnt
-  jsr main
-  rts
+  .label frame_cnt = $19
+  // Remainder after unsigned 16-bit division
+  .label rem16u = $2f
+_start: {
+    // frame_cnt = 1
+    lda #1
+    sta.z frame_cnt
+    jsr main
+    rts
+}
 main: {
     .const toD0181_return = (>(SCREEN&$3fff)*4)|(>BITMAP)/4&$f
-    .label __7 = $1c
-    .label __11 = $1e
+    .label __7 = $1a
+    .label __11 = $1c
     .label __26 = 9
     .label __27 = 9
-    .label __28 = $1c
-    .label __29 = $1e
+    .label __28 = $1a
+    .label __29 = $1c
     .label cos_x = 9
     .label xpos = $b
-    .label x = $1c
+    .label x = $1a
     .label sin_y = 9
     .label ypos = $b
-    .label y = $1e
+    .label y = $1c
     .label idx_x = 2
     .label idx_y = 6
     .label r = 4
@@ -283,11 +284,11 @@ main: {
     jmp __b7
 }
 // Plot a single dot in the bitmap
-// bitmap_plot(word zp($1c) x, byte register(A) y)
+// bitmap_plot(word zp($1a) x, byte register(A) y)
 bitmap_plot: {
-    .label __0 = $2b
+    .label __0 = $1e
     .label plotter = $25
-    .label x = $1c
+    .label x = $1a
     // plotter = (char*) { bitmap_plot_yhi[y], bitmap_plot_ylo[y] }
     tay
     lda bitmap_plot_yhi,y
@@ -323,9 +324,9 @@ bitmap_plot: {
 // Fixes offsets introduced by using unsigned multiplication
 // mul16s(signed word zp(4) a, signed word zp(9) b)
 mul16s: {
-    .label __6 = $2b
+    .label __6 = $1e
     .label __9 = $25
-    .label __11 = $2b
+    .label __11 = $1e
     .label __12 = $25
     .label m = $b
     .label return = $b
@@ -390,12 +391,12 @@ mul16s: {
     rts
 }
 // Perform binary multiplication of two unsigned 16-bit unsigned ints into a 32-bit unsigned long
-// mul16u(word zp($25) a, word zp($1e) b)
+// mul16u(word zp($25) a, word zp($1c) b)
 mul16u: {
     .label mb = $15
     .label a = $25
     .label res = $b
-    .label b = $1e
+    .label b = $1c
     .label return = $b
     // mb = b
     lda.z b
@@ -520,11 +521,11 @@ bitmap_clear: {
     rts
 }
 // Copies the character c (an unsigned char) to the first num characters of the object pointed to by the argument str.
-// memset(void* zp($25) str, byte register(X) c, word zp($1e) num)
+// memset(void* zp($25) str, byte register(X) c, word zp($1c) num)
 memset: {
-    .label end = $1e
+    .label end = $1c
     .label dst = $25
-    .label num = $1e
+    .label num = $1c
     .label str = $25
     // if(num>0)
     lda.z num
@@ -625,7 +626,7 @@ bitmap_init: {
 // Generate signed int sinus table - with values in the range min-max.
 // sintab - the table to generate into
 // wavelength - the number of sinus points in a total sinus wavelength (the size of the table)
-// sin16s_gen2(signed word* zp($1c) sintab)
+// sin16s_gen2(signed word* zp($1a) sintab)
 sin16s_gen2: {
     .const min = -$1001
     .const max = $1001
@@ -633,8 +634,12 @@ sin16s_gen2: {
     .label wavelength = $200
     .label __6 = $b
     .label __8 = $25
+    // ampl is always positive so shifting left does not alter the sign
+    // u[4.28] step = PI*2/wavelength
     .label step = $21
-    .label sintab = $1c
+    .label sintab = $1a
+    // u[4.28]
+    // Iterate over the table
     // u[4.28]
     // Iterate over the table
     .label x = $11
@@ -737,14 +742,23 @@ sin16s: {
     .label __4 = $27
     .label x = $15
     .label return = 4
+    // sinx = x - x^3/6 + x5/128;
     .label x1 = $2d
-    .label x2 = $19
-    .label x3 = $19
+    // u[1.15]
+    .label x2 = $1e
+    // u[2.14] x^2
+    .label x3 = $1e
+    // u[2.14] x^3
     .label x3_6 = $2b
+    // u[1.15] x^3/6;
     .label usinx = 4
-    .label x4 = $19
+    // u[1.15] x - x^3/6
+    .label x4 = $1e
+    // u[3.13] x^4
     .label x5 = $2b
+    // u[4.12] x^5
     .label x5_128 = $2b
+    // u[1.15] (first bit is always zero)
     .label sinx = 4
     // if(x >= PI_u4f28 )
     lda.z x+3
@@ -943,14 +957,14 @@ sin16s: {
 }
 // Calculate val*val for two unsigned int values - the result is 16 selected bits of the 32-bit result.
 // The select parameter indicates how many of the highest bits of the 32-bit result to skip
-// mulu16_sel(word zp($19) v1, word zp($1e) v2, byte register(X) select)
+// mulu16_sel(word zp($1e) v1, word zp($1c) v2, byte register(X) select)
 mulu16_sel: {
     .label __0 = $b
     .label __1 = $b
-    .label v1 = $19
-    .label v2 = $1e
+    .label v1 = $1e
+    .label v2 = $1c
     .label return = $2b
-    .label return_1 = $19
+    .label return_1 = $1e
     // mul16u(v1, v2)
     lda.z v1
     sta.z mul16u.a
@@ -999,6 +1013,10 @@ div32u16u: {
     lda.z divr16u.return+1
     sta.z quotient_hi+1
     // divr16u(<dividend, divisor, rem16u)
+    lda.z rem16u
+    sta.z divr16u.rem
+    lda.z rem16u+1
+    sta.z divr16u.rem+1
     lda #<PI2_u4f28&$ffff
     sta.z divr16u.dividend
     lda #>PI2_u4f28&$ffff
@@ -1022,10 +1040,10 @@ div32u16u: {
 // Returns the quotient dividend/divisor.
 // The final remainder will be set into the global variable rem16u
 // Implemented using simple binary division
-// divr16u(word zp($1e) dividend, word zp($19) rem)
+// divr16u(word zp($1c) dividend, word zp($1e) rem)
 divr16u: {
-    .label rem = $19
-    .label dividend = $1e
+    .label rem = $1e
+    .label dividend = $1c
     .label quotient = $25
     .label return = $25
     ldx #0
@@ -1082,6 +1100,10 @@ divr16u: {
     cpx #$10
     bne __b1
     // rem16u = rem
+    lda.z rem
+    sta.z rem16u
+    lda.z rem+1
+    sta.z rem16u+1
     // }
     rts
 }

src/test/ref/bitmap-plot-3.asm

@@ -97,7 +97,9 @@ bitmap_line: {
     .label dy = 8
     .label sx = $16
     .label sy = 6
+    // X is the driver
     .label e1 = $e
+    // Y is the driver
     .label e = $a
     .label y = $c
     .label x = 4

src/test/ref/c-types.asm

@@ -2,8 +2,9 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
-  .label print_line_cursor = 2
-  .label print_char_cursor = 8
+  .label print_screen = $400
+  .label print_char_cursor = 6
+  .label print_line_cursor = $a
 main: {
     // print_cls()
     jsr print_cls
@@ -99,9 +100,9 @@ print_ln: {
     rts
 }
 // Print a signed long as HEX
-// print_slong(signed dword zp(4) dw)
+// print_slong(signed dword zp(2) dw)
 print_slong: {
-    .label dw = 4
+    .label dw = 2
     // if(dw<0)
     lda.z dw+3
     bmi __b1
@@ -152,9 +153,9 @@ print_char: {
     rts
 }
 // Print a unsigned long as HEX
-// print_ulong(dword zp(4) dw)
+// print_ulong(dword zp(2) dw)
 print_ulong: {
-    .label dw = 4
+    .label dw = 2
     // print_uint(>dw)
     lda.z dw+2
     sta.z print_uint.w
@@ -171,9 +172,9 @@ print_ulong: {
     rts
 }
 // Print a unsigned int as HEX
-// print_uint(word zp($a) w)
+// print_uint(word zp(8) w)
 print_uint: {
-    .label w = $a
+    .label w = 8
     // print_uchar(>w)
     ldx.z w+1
     jsr print_uchar
@@ -207,9 +208,9 @@ print_uchar: {
     rts
 }
 // Print a zero-terminated string
-// print_str(byte* zp($a) str)
+// print_str(byte* zp(8) str)
 print_str: {
-    .label str = $a
+    .label str = 8
   __b1:
     // while(*str)
     ldy #0
@@ -276,9 +277,9 @@ testInt: {
     .byte 0
 }
 // Print a signed int as HEX
-// print_sint(signed word zp($a) w)
+// print_sint(signed word zp(8) w)
 print_sint: {
-    .label w = $a
+    .label w = 8
     // if(w<0)
     lda.z w+1
     bmi __b1
@@ -354,9 +355,9 @@ testChar: {
     .const n = $e
     .label s = -$e
     // print_str("char: ")
-    lda #<$400
+    lda #<print_screen
     sta.z print_char_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_char_cursor+1
     lda #<str
     sta.z print_str.str
@@ -378,9 +379,9 @@ testChar: {
     // print_schar(s)
     jsr print_schar
     // print_ln()
-    lda #<$400
+    lda #<print_screen
     sta.z print_line_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_line_cursor+1
     jsr print_ln
     // }
@@ -411,7 +412,7 @@ print_cls: {
 memset: {
     .const c = ' '
     .const num = $3e8
-    .label str = $400
+    .label str = print_screen
     .label end = str+num
     .label dst = $a
     lda #<str

src/test/ref/c64dtv-8bppcharstretch.asm

@@ -263,6 +263,7 @@ gfx_init: {
 }
 // Initialize Plane with 8bpp charset
 gfx_init_plane_charset8: {
+    // 8bpp cells for Plane B (charset) - ROM charset with 256 colors
     // 8bpp cells for Plane B (charset) - ROM charset with 256 colors
     .const gfxbCpuBank = $ff&CHARSET8/$4000
     .label bits = 2
@@ -354,6 +355,7 @@ gfx_init_plane_charset8: {
 // The actual memory addressed will be $4000*cpuSegmentIdx
 // dtvSetCpuBankSegment1(byte register(A) cpuBankIdx)
 dtvSetCpuBankSegment1: {
+    // Move CPU BANK 1 SEGMENT ($4000-$7fff)
     // Move CPU BANK 1 SEGMENT ($4000-$7fff)
     .label cpuBank = $ff
     // *cpuBank = cpuBankIdx

src/test/ref/c64dtv-8bppchunkystretch.asm

@@ -309,6 +309,7 @@ gfx_init_chunky: {
 // The actual memory addressed will be $4000*cpuSegmentIdx
 // dtvSetCpuBankSegment1(byte register(A) cpuBankIdx)
 dtvSetCpuBankSegment1: {
+    // Move CPU BANK 1 SEGMENT ($4000-$7fff)
     // Move CPU BANK 1 SEGMENT ($4000-$7fff)
     .label cpuBank = $ff
     // *cpuBank = cpuBankIdx

src/test/ref/c64dtv-blitter-box.asm

@@ -70,6 +70,7 @@
   .label DTV_BLITTER_CONTROL2 = $d33f
   .label SCREEN = $400
   // Controls the ALU operation
+  // Controls the ALU operation
   .label DTV_BLITTER_ALU = $d33e
 main: {
     // *DTV_FEATURE = DTV_FEATURE_ENABLE

src/test/ref/c64dtv-blittermin.asm

@@ -75,6 +75,7 @@
   .label DTV_BLITTER_CONTROL2 = $d33f
   .label SCREEN = $400
   // Controls the ALU operation
+  // Controls the ALU operation
   .label DTV_BLITTER_ALU = $d33e
 main: {
     // *DTV_FEATURE = DTV_FEATURE_ENABLE

src/test/ref/c64dtv-gfxexplorer.asm

@@ -75,6 +75,7 @@
   .const OFFSET_STRUCT_MOS6569_VICII_CONTROL2 = $16
   .const OFFSET_STRUCT_MOS6569_VICII_MEMORY = $18
   // Number of form fields
+  // Number of form fields
   .const form_fields_cnt = $24
   .label VIC_CONTROL = $d011
   .label VIC_CONTROL2 = $d016
@@ -169,15 +170,15 @@
   .label form_vic_bg2_lo = form_fields_val+$21
   .label form_vic_bg3_hi = form_fields_val+$22
   .label form_vic_bg3_lo = form_fields_val+$23
-  .label print_line_cursor = 7
+  .label print_char_cursor = $d
+  .label print_line_cursor = $10
   // Keyboard event buffer size. The number of events currently in the event buffer
-  .label keyboard_events_size = 2
+  .label keyboard_events_size = 9
   // Counts down to blink for form cursor (it is inversed in the lower half)
   // Always blink cursor in new field
-  .label form_cursor_count = 3
+  .label form_cursor_count = $b
   // Current selected field in the form
-  .label form_field_idx = 4
-  .label print_char_cursor = 5
+  .label form_field_idx = $c
 main: {
     // asm
     sei
@@ -213,23 +214,24 @@ main: {
 }
 // Change graphics mode to show the selected graphics mode
 gfx_mode: {
-    .label __20 = 9
-    .label __24 = $14
-    .label __26 = $1d
-    .label __34 = 9
-    .label __38 = $1f
-    .label __40 = $1b
-    .label __47 = $10
-    .label __48 = $10
-    .label __50 = 7
+    .label __20 = 2
+    .label __24 = $1c
+    .label __26 = $1e
+    .label __34 = 2
+    .label __38 = $14
+    .label __40 = $1a
+    .label __47 = 6
+    .label __48 = 6
+    .label __50 = $10
     .label __52 = $13
-    .label __82 = $10
-    .label __83 = 7
-    .label plane_a = 9
-    .label plane_b = 9
-    .label vic_colors = $10
-    .label col = 5
-    .label cy = $d
+    .label __82 = 6
+    .label __83 = $10
+    .label plane_a = 2
+    .label plane_b = 2
+    // VIC Colors
+    .label vic_colors = 6
+    .label col = $d
+    .label cy = $f
     // if(*form_ctrl_line!=0)
     lda form_ctrl_line
     cmp #0
@@ -663,9 +665,9 @@ keyboard_event_get: {
 // Handles debounce and only generates events when the status of a key changes.
 // Also stores current status of modifiers in keyboard_modifiers.
 keyboard_event_scan: {
-    .label row_scan = $16
-    .label keycode = $22
-    .label row = $e
+    .label row_scan = $20
+    .label keycode = 8
+    .label row = $22
     lda #0
     sta.z keycode
     sta.z row
@@ -800,10 +802,10 @@ keyboard_event_scan: {
 }
 // Determine if a specific key is currently pressed based on the last keyboard_event_scan()
 // Returns 0 is not pressed and non-0 if pressed
-// keyboard_event_pressed(byte zp($f) keycode)
+// keyboard_event_pressed(byte zp($a) keycode)
 keyboard_event_pressed: {
     .label row_bits = $21
-    .label keycode = $f
+    .label keycode = $a
     // keycode>>3
     lda.z keycode
     lsr
@@ -842,7 +844,7 @@ keyboard_matrix_read: {
 // Get the VIC screen address from the screen index
 // get_vic_screen(byte register(A) idx)
 get_vic_screen: {
-    .label return = $10
+    .label return = 6
     // if(idx==0)
     cmp #0
     beq __b1
@@ -892,7 +894,7 @@ get_vic_screen: {
 // Get the VIC charset/bitmap address from the index
 // get_vic_charset(byte register(A) idx)
 get_vic_charset: {
-    .label return = 7
+    .label return = $10
     // if(idx==0)
     cmp #0
     beq __b1
@@ -915,7 +917,7 @@ get_vic_charset: {
 // Get plane address from a plane index (from the form)
 // get_plane(byte register(A) idx)
 get_plane: {
-    .label return = 9
+    .label return = 2
     // if(idx==0)
     cmp #0
     beq __b1
@@ -1117,7 +1119,7 @@ get_plane: {
 }
 // Show the form - and let the user change values
 form_mode: {
-    .label preset_current = $d
+    .label preset_current = $f
     // print_set_screen(COLS)
   // Form Colors
     lda #<COLS
@@ -1256,7 +1258,7 @@ form_mode: {
 // idx is the ID of the preset
 // render_preset_name(byte register(A) idx)
 render_preset_name: {
-    .label name = $10
+    .label name = 6
     // if(idx==0)
     cmp #0
     beq __b3
@@ -1385,10 +1387,10 @@ render_preset_name: {
     .byte 0
 }
 // Print a string at a specific screen position
-// print_str_at(byte* zp($10) str, byte* zp(5) at)
+// print_str_at(byte* zp(6) str, byte* zp($d) at)
 print_str_at: {
-    .label at = 5
-    .label str = $10
+    .label at = $d
+    .label str = 6
     lda #<FORM_SCREEN+$28*2+$a
     sta.z at
     lda #>FORM_SCREEN+$28*2+$a
@@ -1442,9 +1444,9 @@ form_render_values: {
 // field_idx is the index of the field to get the screen address for
 // form_field_ptr(byte register(X) field_idx)
 form_field_ptr: {
-    .label line = $14
-    .label x = $16
-    .label return = $1d
+    .label line = $1c
+    .label x = $13
+    .label return = $1e
     // y = form_fields_y[field_idx]
     ldy form_fields_y,x
     // line = (byte*) { form_line_hi[y], form_line_lo[y] }
@@ -1469,7 +1471,7 @@ form_field_ptr: {
 // idx is the ID of the preset
 // apply_preset(byte register(A) idx)
 apply_preset: {
-    .label preset = 7
+    .label preset = $10
     // if(idx==0)
     cmp #0
     beq __b3
@@ -1588,7 +1590,7 @@ apply_preset: {
 // Reads keyboard and allows the user to navigate and change the fields of the form
 // Returns 0 if space is not pressed, non-0 if space is pressed
 form_control: {
-    .label field = $1d
+    .label field = $1e
     // form_field_ptr(form_field_idx)
     ldx.z form_field_idx
     jsr form_field_ptr
@@ -1724,7 +1726,8 @@ form_control: {
 // Set the screen to use for the form.
 // screen is the start address of the screen to use
 form_set_screen: {
-    .label line = $10
+    // Calculate the field line table
+    .label line = 6
     ldx #0
     lda #<FORM_SCREEN
     sta.z line
@@ -1756,9 +1759,9 @@ form_set_screen: {
 }
 // Print a number of zero-terminated strings, each followed by a newline.
 // The sequence of lines is terminated by another zero.
-// print_str_lines(byte* zp($10) str)
+// print_str_lines(byte* zp(6) str)
 print_str_lines: {
-    .label str = $10
+    .label str = 6
     lda.z print_set_screen.screen
     sta.z print_char_cursor
     lda.z print_set_screen.screen+1
@@ -1845,13 +1848,13 @@ print_cls: {
     rts
 }
 // Copies the character c (an unsigned char) to the first num characters of the object pointed to by the argument str.
-// memset(void* zp($10) str)
+// memset(void* zp($d) str)
 memset: {
     .const c = ' '
     .const num = $3e8
-    .label end = $1f
-    .label dst = $10
-    .label str = $10
+    .label end = $14
+    .label dst = $d
+    .label str = $d
     // end = (char*)str + num
     lda.z str
     clc
@@ -1883,9 +1886,9 @@ memset: {
     jmp __b2
 }
 // Set the screen to print on. Also resets current line/char cursor.
-// print_set_screen(byte* zp(7) screen)
+// print_set_screen(byte* zp($10) screen)
 print_set_screen: {
-    .label screen = 7
+    .label screen = $10
     // }
     rts
 }
@@ -1942,16 +1945,16 @@ gfx_init_plane_full: {
     rts
 }
 // Initialize 320*200 1bpp pixel ($2000) plane with identical bytes
-// gfx_init_plane_fill(dword zp(9) plane_addr, byte zp($e) fill)
+// gfx_init_plane_fill(dword zp(2) plane_addr, byte zp($22) fill)
 gfx_init_plane_fill: {
-    .label __0 = $17
-    .label __1 = $1b
+    .label __0 = $16
+    .label __1 = $1a
     .label __4 = $10
     .label __5 = $10
     .label gfxb = $10
-    .label by = $22
-    .label plane_addr = 9
-    .label fill = $e
+    .label by = 8
+    .label plane_addr = 2
+    .label fill = $22
     // plane_addr*4
     lda.z plane_addr
     asl
@@ -2033,6 +2036,7 @@ gfx_init_plane_fill: {
 // The actual memory addressed will be $4000*cpuSegmentIdx
 // dtvSetCpuBankSegment1(byte register(A) cpuBankIdx)
 dtvSetCpuBankSegment1: {
+    // Move CPU BANK 1 SEGMENT ($4000-$7fff)
     // Move CPU BANK 1 SEGMENT ($4000-$7fff)
     .label cpuBank = $ff
     // *cpuBank = cpuBankIdx
@@ -2081,8 +2085,8 @@ gfx_init_plane_vertical2: {
 // Initialize Plane with Horizontal Stripes every 2 pixels
 gfx_init_plane_horisontal2: {
     .const gfxbCpuBank = PLANE_HORISONTAL2/$4000
-    .label gfxa = $10
-    .label ay = $f
+    .label gfxa = 6
+    .label ay = 9
     // dtvSetCpuBankSegment1(gfxbCpuBank++)
     lda #gfxbCpuBank
     jsr dtvSetCpuBankSegment1
@@ -2130,8 +2134,8 @@ gfx_init_plane_horisontal2: {
 // Initialize Plane with Vertical Stripes
 gfx_init_plane_vertical: {
     .const gfxbCpuBank = PLANE_VERTICAL/$4000
-    .label gfxb = $10
-    .label by = $d
+    .label gfxb = 6
+    .label by = $a
     // dtvSetCpuBankSegment1(gfxbCpuBank++)
     lda #gfxbCpuBank
     jsr dtvSetCpuBankSegment1
@@ -2172,8 +2176,8 @@ gfx_init_plane_vertical: {
 // Initialize Plane with Horizontal Stripes
 gfx_init_plane_horisontal: {
     .const gfxbCpuBank = PLANE_HORISONTAL/$4000
-    .label gfxa = $10
-    .label ay = $d
+    .label gfxa = 6
+    .label ay = $b
     // dtvSetCpuBankSegment1(gfxbCpuBank++)
     lda #gfxbCpuBank
     jsr dtvSetCpuBankSegment1
@@ -2231,14 +2235,15 @@ gfx_init_plane_horisontal: {
 }
 // Initialize Plane with 8bpp charset
 gfx_init_plane_charset8: {
+    // 8bpp cells for Plane B (charset) - ROM charset with 256 colors
     // 8bpp cells for Plane B (charset) - ROM charset with 256 colors
     .const gfxbCpuBank = PLANE_CHARSET8/$4000
-    .label bits = $f
-    .label chargen = $10
-    .label gfxa = $14
-    .label col = $12
-    .label cr = $22
-    .label ch = $e
+    .label bits = $22
+    .label chargen = 6
+    .label gfxa = $d
+    .label col = 8
+    .label cr = $f
+    .label ch = $c
     // dtvSetCpuBankSegment1(gfxbCpuBank++)
     lda #gfxbCpuBank
     jsr dtvSetCpuBankSegment1
@@ -2319,10 +2324,10 @@ gfx_init_plane_charset8: {
 }
 // Initialize 8BPP Chunky Bitmap (contains 8bpp pixels)
 gfx_init_plane_8bppchunky: {
-    .label __5 = $1d
+    .label __5 = $1c
     .label gfxb = $10
-    .label x = $14
-    .label y = $e
+    .label x = $d
+    .label y = $f
     // dtvSetCpuBankSegment1(gfxbCpuBank++)
     lda #PLANE_8BPP_CHUNKY/$4000
     jsr dtvSetCpuBankSegment1
@@ -2432,12 +2437,12 @@ gfx_init_vic_bitmap: {
     lines_y: .byte 0, 0, $c7, $c7, 0, 0, $64, $c7, $64, 0
 }
 // Draw a line on the bitmap
-// bitmap_line(byte zp($16) x0, byte zp($21) x1, byte register(X) y0, byte zp($13) y1)
+// bitmap_line(byte zp($13) x0, byte zp($f) x1, byte register(X) y0, byte zp($a) y1)
 bitmap_line: {
     .label xd = $12
-    .label x0 = $16
-    .label x1 = $21
-    .label y1 = $13
+    .label x0 = $13
+    .label x1 = $f
+    .label y1 = $a
     // if(x0<x1)
     lda.z x0
     cmp.z x1
@@ -2558,13 +2563,13 @@ bitmap_line: {
     jsr bitmap_line_xdyi
     rts
 }
-// bitmap_line_xdyi(byte zp($d) x, byte register(X) y, byte zp($16) x1, byte zp($12) xd, byte zp($f) yd)
+// bitmap_line_xdyi(byte zp(9) x, byte register(X) y, byte zp($13) x1, byte zp($12) xd, byte zp(8) yd)
 bitmap_line_xdyi: {
-    .label x = $d
-    .label x1 = $16
+    .label x = 9
+    .label x1 = $13
     .label xd = $12
-    .label yd = $f
-    .label e = $13
+    .label yd = 8
+    .label e = $a
     // e = yd>>1
     lda.z yd
     lsr
@@ -2604,9 +2609,10 @@ bitmap_line_xdyi: {
 }
 // bitmap_plot(byte register(Y) x, byte register(X) y)
 bitmap_plot: {
-    .label plotter_x = $1d
-    .label plotter_y = $1f
-    .label plotter = $1d
+    // Needs unsigned int arrays arranged as two underlying char arrays to allow char* plotter_x = plot_x[x]; - and eventually - char* plotter = plot_x[x] + plot_y[y];
+    .label plotter_x = $1c
+    .label plotter_y = $1e
+    .label plotter = $1c
     // plotter_x = { bitmap_plot_xhi[x], bitmap_plot_xlo[x] }
     lda bitmap_plot_xhi,y
     sta.z plotter_x+1
@@ -2634,14 +2640,14 @@ bitmap_plot: {
     // }
     rts
 }
-// bitmap_line_ydxi(byte zp($e) y, byte zp($21) x, byte zp($13) y1, byte zp($d) yd, byte zp($12) xd)
+// bitmap_line_ydxi(byte zp($b) y, byte zp($f) x, byte zp($a) y1, byte zp(9) yd, byte zp($12) xd)
 bitmap_line_ydxi: {
-    .label y = $e
-    .label x = $21
-    .label y1 = $13
-    .label yd = $d
+    .label y = $b
+    .label x = $f
+    .label y1 = $a
+    .label yd = 9
     .label xd = $12
-    .label e = $f
+    .label e = $c
     // e = xd>>1
     lda.z xd
     lsr
@@ -2679,13 +2685,13 @@ bitmap_line_ydxi: {
     // }
     rts
 }
-// bitmap_line_xdyd(byte zp($f) x, byte register(X) y, byte zp($16) x1, byte zp($12) xd, byte zp($e) yd)
+// bitmap_line_xdyd(byte zp($c) x, byte register(X) y, byte zp($13) x1, byte zp($12) xd, byte zp($b) yd)
 bitmap_line_xdyd: {
-    .label x = $f
-    .label x1 = $16
+    .label x = $c
+    .label x1 = $13
     .label xd = $12
-    .label yd = $e
-    .label e = $d
+    .label yd = $b
+    .label e = $f
     // e = yd>>1
     lda.z yd
     lsr
@@ -2723,14 +2729,14 @@ bitmap_line_xdyd: {
     // }
     rts
 }
-// bitmap_line_ydxd(byte zp($e) y, byte zp($16) x, byte zp($13) y1, byte zp($d) yd, byte zp($12) xd)
+// bitmap_line_ydxd(byte zp($20) y, byte zp($13) x, byte zp($a) y1, byte zp($f) yd, byte zp($12) xd)
 bitmap_line_ydxd: {
-    .label y = $e
-    .label x = $16
-    .label y1 = $13
-    .label yd = $d
+    .label y = $20
+    .label x = $13
+    .label y1 = $a
+    .label yd = $f
     .label xd = $12
-    .label e = $f
+    .label e = $21
     // e = xd>>1
     lda.z xd
     lsr
@@ -2771,7 +2777,7 @@ bitmap_line_ydxd: {
 // Clear all graphics on the bitmap
 bitmap_clear: {
     .label bitmap = $10
-    .label y = $16
+    .label y = $12
     // bitmap = (char*) { bitmap_plot_xhi[0], bitmap_plot_xlo[0] }
     lda bitmap_plot_xlo
     sta.z bitmap
@@ -2805,8 +2811,8 @@ bitmap_clear: {
 }
 // Initialize the bitmap plotter tables for a specific bitmap
 bitmap_init: {
-    .label __10 = $21
-    .label yoffs = $14
+    .label __10 = $20
+    .label yoffs = $d
     ldy #$80
     ldx #0
   __b1:
@@ -2874,8 +2880,8 @@ bitmap_init: {
 }
 gfx_init_charset: {
     .label charset = $10
-    .label chargen = $14
-    .label c = $21
+    .label chargen = $d
+    .label c = $13
     // *PROCPORT = $32
     lda #$32
     sta PROCPORT
@@ -2923,7 +2929,7 @@ gfx_init_charset: {
 // Initialize VIC screen 4 - all chars are 00
 gfx_init_screen4: {
     .label ch = $10
-    .label cy = $e
+    .label cy = $f
     lda #0
     sta.z cy
     lda #<VIC_SCREEN4
@@ -2956,9 +2962,9 @@ gfx_init_screen4: {
 }
 // Initialize VIC screen 3 ( value is %00xx00yy where xx is xpos and yy is ypos
 gfx_init_screen3: {
-    .label __1 = $22
-    .label ch = $10
-    .label cy = $f
+    .label __1 = $21
+    .label ch = $14
+    .label cy = $20
     lda #<VIC_SCREEN3
     sta.z ch
     lda #>VIC_SCREEN3
@@ -3005,8 +3011,8 @@ gfx_init_screen3: {
 // Initialize VIC screen 2 ( value is %ccccrrrr where cccc is (x+y mod $f) and rrrr is %1111-%cccc)
 gfx_init_screen2: {
     .label col2 = $22
-    .label ch = $10
-    .label cy = $12
+    .label ch = $14
+    .label cy = $21
     lda #<VIC_SCREEN2
     sta.z ch
     lda #>VIC_SCREEN2
@@ -3100,7 +3106,7 @@ gfx_init_screen1: {
 gfx_init_screen0: {
     .label __1 = $22
     .label ch = $14
-    .label cy = $13
+    .label cy = $12
     lda #<VIC_SCREEN0
     sta.z ch
     lda #>VIC_SCREEN0

src/test/ref/c64dtv-gfxmodes.asm

@@ -99,7 +99,7 @@
   // Memory address of VIC Graphics is GraphicsBank*$10000
   .label DTV_GRAPHICS_VIC_BANK = $d03d
   .label print_char_cursor = 6
-  .label print_line_cursor = 8
+  .label print_line_cursor = $b
 main: {
     // asm
     sei
@@ -192,9 +192,6 @@ menu: {
     sta BG_COLOR
     // *BORDER_COLOR = 0
     sta BORDER_COLOR
-    // print_set_screen(SCREEN)
-  // Display menu Text
-    jsr print_set_screen
     // print_cls()
     jsr print_cls
     // print_str_lines(MENU_TEXT)
@@ -355,8 +352,8 @@ menu: {
 mode_8bppchunkybmm: {
     // 8BPP Chunky Bitmap (contains 8bpp pixels)
     .const PLANEB = $20000
-    .label __7 = $d
-    .label gfxb = 8
+    .label __7 = $b
+    .label gfxb = $d
     .label x = 6
     .label y = 2
     // *DTV_CONTROL = DTV_HIGHCOLOR | DTV_LINEAR | DTV_CHUNKY | DTV_COLORRAM_OFF
@@ -478,6 +475,7 @@ mode_8bppchunkybmm: {
 }
 // Allow the user to control the DTV graphics using different keys
 mode_ctrl: {
+    // Read the current control byte
     // DTV Graphics Mode - Reset
     .label ctrl = 3
   __b1:
@@ -499,6 +497,7 @@ mode_ctrl: {
   __b4:
     // ctrl = dtv_control
     // Read the current control byte
+    // DTV Graphics Mode - Reset
     stx.z ctrl
     // keyboard_key_pressed(KEY_L)
     ldy #KEY_L
@@ -601,7 +600,7 @@ mode_ctrl: {
 // Returns zero if the key is not pressed and a non-zero value if the key is currently pressed
 // keyboard_key_pressed(byte register(Y) key)
 keyboard_key_pressed: {
-    .label colidx = $13
+    .label colidx = $11
     // colidx = key&7
     tya
     and #7
@@ -641,6 +640,7 @@ keyboard_matrix_read: {
 // The actual memory addressed will be $4000*cpuSegmentIdx
 // dtvSetCpuBankSegment1(byte register(A) cpuBankIdx)
 dtvSetCpuBankSegment1: {
+    // Move CPU BANK 1 SEGMENT ($4000-$7fff)
     // Move CPU BANK 1 SEGMENT ($4000-$7fff)
     .label cpuBank = $ff
     // *cpuBank = cpuBankIdx
@@ -667,15 +667,16 @@ mode_8bpppixelcell: {
     // 8BPP Pixel Cell Charset (contains 256 64 byte chars)
     .label PLANEB = $4000
     .label CHARGEN = $d000
-    .label __3 = $a
+    .label __3 = 8
+    // Screen Chars for Plane A (screen) - 16x16 repeating
     // Screen Chars for Plane A (screen) - 16x16 repeating
     .label gfxa = 6
     .label ay = 2
-    .label bits = $11
-    .label chargen = 8
+    .label bits = $f
+    .label chargen = $d
     .label gfxb = 4
-    .label col = $13
-    .label cr = $c
+    .label col = $11
+    .label cr = $a
     .label ch = 3
     // *DTV_CONTROL = DTV_HIGHCOLOR | DTV_LINEAR | DTV_CHUNKY
     lda #DTV_HIGHCOLOR|DTV_LINEAR|DTV_CHUNKY
@@ -860,14 +861,17 @@ mode_sixsfred: {
     .label PLANEB = $6000
     .label COLORS = $8000
     // Colors for high 4 bits of 8bpp
+    // Colors for high 4 bits of 8bpp
     .label col = 4
-    .label cy = $c
+    .label cy = $a
+    // Graphics for Plane A () - horizontal stripes every 2 pixels
     // Graphics for Plane A () - horizontal stripes every 2 pixels
     .label gfxa = 6
-    .label ay = $11
+    .label ay = $f
     // Graphics for Plane B - vertical stripes every 2 pixels
-    .label gfxb = 8
-    .label by = $13
+    // Graphics for Plane B - vertical stripes every 2 pixels
+    .label gfxb = $d
+    .label by = $11
     // *DTV_CONTROL = DTV_HIGHCOLOR | DTV_LINEAR
     lda #DTV_HIGHCOLOR|DTV_LINEAR
     sta DTV_CONTROL
@@ -1047,17 +1051,21 @@ mode_twoplanebitmap: {
     .label PLANEA = $4000
     .label PLANEB = $6000
     .label COLORS = $8000
-    .label __3 = $a
+    .label __3 = 8
+    // Color for bits 11
+    // Colors for bits 01 / 10
     // Color for bits 11
     // Colors for bits 01 / 10
     .label col = 6
     .label cy = 2
     // Graphics for Plane A - horizontal stripes
-    .label gfxa = 8
+    // Graphics for Plane A - horizontal stripes
+    .label gfxa = $d
     .label ay = 3
     // Graphics for Plane B - vertical stripes
+    // Graphics for Plane B - vertical stripes
     .label gfxb = 4
-    .label by = $c
+    .label by = $a
     // *DTV_CONTROL = DTV_HIGHCOLOR | DTV_LINEAR
     lda #DTV_HIGHCOLOR|DTV_LINEAR
     sta DTV_CONTROL
@@ -1260,16 +1268,19 @@ mode_sixsfred2: {
     .label PLANEA = $4000
     .label PLANEB = $6000
     .label COLORS = $8000
-    .label __3 = $b
+    .label __3 = 9
+    // Colors for high 4 bits of 8bpp
     // Colors for high 4 bits of 8bpp
     .label col = 4
     .label cy = 2
     // Graphics for Plane A () - horizontal stripes every 2 pixels
+    // Graphics for Plane A () - horizontal stripes every 2 pixels
     .label gfxa = 6
     .label ay = 3
     // Graphics for Plane B - vertical stripes every 2 pixels
-    .label gfxb = 8
-    .label by = $c
+    // Graphics for Plane B - vertical stripes every 2 pixels
+    .label gfxb = $d
+    .label by = $a
     // *DTV_CONTROL = DTV_LINEAR
     lda #DTV_LINEAR
     sta DTV_CONTROL
@@ -1461,11 +1472,12 @@ mode_hicolmcchar: {
     .label CHARSET = $9000
     // Charset ROM
     .label COLORS = $8400
-    .label __3 = $b
+    .label __3 = 9
+    // Char Colors and screen chars
     // Char Colors and screen chars
     .label col = 6
-    .label ch = 8
-    .label cy = $11
+    .label ch = $d
+    .label cy = $f
     // *DTV_GRAPHICS_VIC_BANK = (byte)((dword)CHARSET/$10000)
     // DTV Graphics Bank
     lda #0
@@ -1598,11 +1610,12 @@ mode_hicolecmchar: {
     .label CHARSET = $9000
     // Charset ROM
     .label COLORS = $8400
-    .label __3 = $c
+    .label __3 = $a
+    // Char Colors and screen chars
     // Char Colors and screen chars
     .label col = 6
-    .label ch = 8
-    .label cy = $11
+    .label ch = $d
+    .label cy = $f
     // *DTV_GRAPHICS_VIC_BANK = (byte)((dword)CHARSET/$10000)
     // DTV Graphics Bank
     lda #0
@@ -1734,11 +1747,12 @@ mode_hicolstdchar: {
     .label CHARSET = $9000
     // Charset ROM
     .label COLORS = $8400
-    .label __3 = $c
+    .label __3 = $a
+    // Char Colors and screen chars
     // Char Colors and screen chars
     .label col = 6
-    .label ch = 8
-    .label cy = $13
+    .label ch = $d
+    .label cy = $11
     // *DTV_GRAPHICS_VIC_BANK = (byte)((dword)CHARSET/$10000)
     // DTV Graphics Bank
     lda #0
@@ -1859,11 +1873,12 @@ mode_stdbitmap: {
     .const lines_cnt = 9
     .label SCREEN = $4000
     .label BITMAP = $6000
-    .label col2 = $11
+    .label col2 = $f
+    // Bitmap Colors
     // Bitmap Colors
     .label ch = 4
-    .label cy = $13
-    .label l = $c
+    .label cy = $11
+    .label l = $a
     // *DTV_GRAPHICS_VIC_BANK = (byte)((dword)BITMAP/$10000)
     // DTV Graphics Bank
     lda #0
@@ -1989,12 +2004,12 @@ mode_stdbitmap: {
     lines_y: .byte 0, 0, $c7, $c7, 0, 0, $64, $c7, $64, 0
 }
 // Draw a line on the bitmap
-// bitmap_line(byte zp($11) x0, byte zp($12) x1, byte register(X) y0, byte zp($a) y1)
+// bitmap_line(byte zp($f) x0, byte zp($10) x1, byte register(X) y0, byte zp(8) y1)
 bitmap_line: {
-    .label xd = $b
-    .label x0 = $11
-    .label x1 = $12
-    .label y1 = $a
+    .label xd = 9
+    .label x0 = $f
+    .label x1 = $10
+    .label y1 = 8
     // if(x0<x1)
     lda.z x0
     cmp.z x1
@@ -2115,13 +2130,13 @@ bitmap_line: {
     jsr bitmap_line_xdyi
     rts
 }
-// bitmap_line_xdyi(byte zp(3) x, byte register(X) y, byte zp($11) x1, byte zp($b) xd, byte zp(2) yd)
+// bitmap_line_xdyi(byte zp(3) x, byte register(X) y, byte zp($f) x1, byte zp(9) xd, byte zp(2) yd)
 bitmap_line_xdyi: {
     .label x = 3
-    .label x1 = $11
-    .label xd = $b
+    .label x1 = $f
+    .label xd = 9
     .label yd = 2
-    .label e = $a
+    .label e = 8
     // e = yd>>1
     lda.z yd
     lsr
@@ -2161,9 +2176,10 @@ bitmap_line_xdyi: {
 }
 // bitmap_plot(byte register(Y) x, byte register(X) y)
 bitmap_plot: {
-    .label plotter_x = $d
-    .label plotter_y = $f
-    .label plotter = $d
+    // Needs unsigned int arrays arranged as two underlying char arrays to allow char* plotter_x = plot_x[x]; - and eventually - char* plotter = plot_x[x] + plot_y[y];
+    .label plotter_x = $b
+    .label plotter_y = $d
+    .label plotter = $b
     // plotter_x = { bitmap_plot_xhi[x], bitmap_plot_xlo[x] }
     lda bitmap_plot_xhi,y
     sta.z plotter_x+1
@@ -2191,14 +2207,14 @@ bitmap_plot: {
     // }
     rts
 }
-// bitmap_line_ydxi(byte zp(3) y, byte zp($12) x, byte zp($a) y1, byte zp(2) yd, byte zp($b) xd)
+// bitmap_line_ydxi(byte zp(3) y, byte zp($10) x, byte zp(8) y1, byte zp(2) yd, byte zp(9) xd)
 bitmap_line_ydxi: {
     .label y = 3
-    .label x = $12
-    .label y1 = $a
+    .label x = $10
+    .label y1 = 8
     .label yd = 2
-    .label xd = $b
-    .label e = $13
+    .label xd = 9
+    .label e = $11
     // e = xd>>1
     lda.z xd
     lsr
@@ -2236,12 +2252,12 @@ bitmap_line_ydxi: {
     // }
     rts
 }
-// bitmap_line_xdyd(byte zp(2) x, byte register(X) y, byte zp($11) x1, byte zp($b) xd, byte zp($13) yd)
+// bitmap_line_xdyd(byte zp(2) x, byte register(X) y, byte zp($f) x1, byte zp(9) xd, byte zp($11) yd)
 bitmap_line_xdyd: {
     .label x = 2
-    .label x1 = $11
-    .label xd = $b
-    .label yd = $13
+    .label x1 = $f
+    .label xd = 9
+    .label yd = $11
     .label e = 3
     // e = yd>>1
     lda.z yd
@@ -2280,14 +2296,14 @@ bitmap_line_xdyd: {
     // }
     rts
 }
-// bitmap_line_ydxd(byte zp(3) y, byte zp($11) x, byte zp($a) y1, byte zp(2) yd, byte zp($b) xd)
+// bitmap_line_ydxd(byte zp(3) y, byte zp($f) x, byte zp(8) y1, byte zp(2) yd, byte zp(9) xd)
 bitmap_line_ydxd: {
     .label y = 3
-    .label x = $11
-    .label y1 = $a
+    .label x = $f
+    .label y1 = 8
     .label yd = 2
-    .label xd = $b
-    .label e = $13
+    .label xd = 9
+    .label e = $11
     // e = xd>>1
     lda.z xd
     lsr
@@ -2328,7 +2344,7 @@ bitmap_line_ydxd: {
 // Clear all graphics on the bitmap
 bitmap_clear: {
     .label bitmap = 4
-    .label y = $b
+    .label y = 9
     // bitmap = (char*) { bitmap_plot_xhi[0], bitmap_plot_xlo[0] }
     lda bitmap_plot_xlo
     sta.z bitmap
@@ -2362,7 +2378,7 @@ bitmap_clear: {
 }
 // Initialize the bitmap plotter tables for a specific bitmap
 bitmap_init: {
-    .label __10 = $12
+    .label __10 = $10
     .label yoffs = 4
     ldy #$80
     ldx #0
@@ -2446,11 +2462,12 @@ mode_mcchar: {
     .label CHARSET = $9000
     // Charset ROM
     .label COLORS = $d800
-    .label __5 = $11
+    .label __5 = $f
+    // Char Colors and screen chars
     // Char Colors and screen chars
     .label col = 4
     .label ch = $d
-    .label cy = $c
+    .label cy = $a
     // *DTV_GRAPHICS_VIC_BANK = (byte)((dword)CHARSET/$10000)
     // DTV Graphics Bank
     lda #0
@@ -2588,11 +2605,12 @@ mode_ecmchar: {
     .label CHARSET = $9000
     // Charset ROM
     .label COLORS = $d800
-    .label __5 = $12
+    .label __5 = $10
+    // Char Colors and screen chars
     // Char Colors and screen chars
     .label col = $d
     .label ch = 6
-    .label cy = $11
+    .label cy = $f
     // *DTV_GRAPHICS_VIC_BANK = (byte)((dword)CHARSET/$10000)
     // DTV Graphics Bank
     lda #0
@@ -2728,11 +2746,12 @@ mode_stdchar: {
     .label CHARSET = $9000
     // Charset ROM
     .label COLORS = $d800
-    .label __5 = $13
+    .label __5 = $11
+    // Char Colors and screen chars
     // Char Colors and screen chars
     .label col = 6
     .label ch = $d
-    .label cy = $12
+    .label cy = $10
     // *DTV_GRAPHICS_VIC_BANK = (byte)((dword)CHARSET/$10000)
     // DTV Graphics Bank
     lda #0
@@ -2947,7 +2966,7 @@ memset: {
     .const num = $3e8
     .label str = menu.SCREEN
     .label end = str+num
-    .label dst = $d
+    .label dst = $b
     lda #<str
     sta.z dst
     lda #>str
@@ -2973,11 +2992,6 @@ memset: {
     inc.z dst+1
   !:
     jmp __b1
-}
-// Set the screen to print on. Also resets current line/char cursor.
-print_set_screen: {
-    // }
-    rts
 }
   // Default vallues for the palette
   DTV_PALETTE_DEFAULT: .byte 0, $f, $36, $be, $58, $db, $86, $ff, $29, $26, $3b, 5, 7, $df, $9a, $a

src/test/ref/cia-timer-cyclecount.asm

@@ -18,6 +18,7 @@
   .label SCREEN = $400
 main: {
     .label __1 = 9
+    // Calculate the cycle count - 0x12 is the base usage of start/read
     .label cyclecount = 9
   __b1:
     // clock_start()

src/test/ref/circlechars.asm

@@ -2,7 +2,7 @@
 // C standard library string.h
 // Functions to manipulate C strings and arrays.
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .const LIGHT_BLUE = $e
   .const OFFSET_STRUCT_PRINTF_BUFFER_NUMBER_DIGITS = 1
@@ -12,37 +12,40 @@
   // Default address of screen character matrix
   .label DEFAULT_SCREEN = $400
   .label SCREEN = $400
-  .label conio_cursor_x = 9
-  .label conio_cursor_y = $a
-  .label conio_line_text = $b
-  .label conio_line_color = $d
-__bbegin:
-  // conio_cursor_x = 0
   // The number of bytes on the screen
   // The current cursor x-position
-  lda #0
-  sta.z conio_cursor_x
-  // conio_cursor_y = 0
+  .label conio_cursor_x = 9
   // The current cursor y-position
-  sta.z conio_cursor_y
-  // conio_line_text = CONIO_SCREEN_TEXT
+  .label conio_cursor_y = $a
   // The current text cursor line start
-  lda #<DEFAULT_SCREEN
-  sta.z conio_line_text
-  lda #>DEFAULT_SCREEN
-  sta.z conio_line_text+1
-  // conio_line_color = CONIO_SCREEN_COLORS
+  .label conio_line_text = $b
   // The current color cursor line start
-  lda #<COLORRAM
-  sta.z conio_line_color
-  lda #>COLORRAM
-  sta.z conio_line_color+1
-  jsr main
-  rts
+  .label conio_line_color = $d
+_start: {
+    // conio_cursor_x = 0
+    lda #0
+    sta.z conio_cursor_x
+    // conio_cursor_y = 0
+    sta.z conio_cursor_y
+    // conio_line_text = CONIO_SCREEN_TEXT
+    lda #<DEFAULT_SCREEN
+    sta.z conio_line_text
+    lda #>DEFAULT_SCREEN
+    sta.z conio_line_text+1
+    // conio_line_color = CONIO_SCREEN_COLORS
+    lda #<COLORRAM
+    sta.z conio_line_color
+    lda #>COLORRAM
+    sta.z conio_line_color+1
+    jsr main
+    rts
+}
 main: {
     .label __9 = $10
     .label __10 = $d
+    // yd = y-12 
     .label yd = $f
+    // dist_sq = xd*xd + yd*yd
     .label dist_sq = $10
     .label y = 9
     .label sc = $b

src/test/ref/clobber-a-problem.asm

@@ -1,5 +1,5 @@
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .const DARK_GREY = $b
   .const BLACK = 0
@@ -7,12 +7,13 @@
   .label BORDER_COLOR = $d020
   .label RASTER = $d012
   .label irq_raster_next = 2
-__bbegin:
-  // irq_raster_next = 0
-  lda #0
-  sta.z irq_raster_next
-  jsr main
-  rts
+_start: {
+    // irq_raster_next = 0
+    lda #0
+    sta.z irq_raster_next
+    jsr main
+    rts
+}
 main: {
     // *KERNEL_IRQ = &irq
     lda #<irq

src/test/ref/complex/clearscreen/clearscreen.asm

@@ -4,7 +4,7 @@
 // C standard library string.h
 // Functions to manipulate C strings and arrays.
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   // The number of iterations performed during 16-bit CORDIC atan2 calculation
   .const CORDIC_ITERATIONS_16 = $f
@@ -80,30 +80,38 @@
   // Sprite data for the animating sprites
   .label SPRITE_DATA = $2000
   // Top of the heap used by malloc()
+  // Top of the heap used by malloc()
   .label HEAP_TOP = $a000
   // Head of the heap. Moved backward each malloc()
   .label heap_head = 8
+  // Copy of the screen used for finding chars to process
   .label SCREEN_COPY = $c
+  // Screen containing bytes representing the distance to the center
   .label SCREEN_DIST = $e
-__bbegin:
-  // malloc(1000)
-  lda #<HEAP_TOP
-  sta.z heap_head
-  lda #>HEAP_TOP
-  sta.z heap_head+1
-  jsr malloc
-  // malloc(1000)
-  lda.z malloc.mem
-  sta.z SCREEN_COPY
-  lda.z malloc.mem+1
-  sta.z SCREEN_COPY+1
-  jsr malloc
-  // malloc(1000)
-  jsr main
-  rts
+_start: {
+    // malloc(1000)
+    lda #<HEAP_TOP
+    sta.z heap_head
+    lda #>HEAP_TOP
+    sta.z heap_head+1
+    jsr malloc
+    // malloc(1000)
+    lda.z malloc.mem
+    sta.z SCREEN_COPY
+    lda.z malloc.mem+1
+    sta.z SCREEN_COPY+1
+    jsr malloc
+    // malloc(1000)
+    lda.z malloc.mem
+    sta.z SCREEN_DIST
+    lda.z malloc.mem+1
+    sta.z SCREEN_DIST+1
+    jsr main
+    rts
+}
 main: {
     .label dst = 2
-    .label src = 8
+    .label src = $22
     .label i = 4
     .label center_y = $10
     // init_angle_screen(SCREEN_DIST)
@@ -147,7 +155,7 @@ main: {
     tax
     ldy #0
   !:
-    lda __2,y
+    lda __0,y
     sta PROCESSING,x
     inx
     iny
@@ -205,36 +213,38 @@ main: {
     jmp __b1
 }
 // Start processing a char - by inserting it into the PROCESSING array
-// startProcessing(byte zp($23) center_x, byte zp($10) center_y)
+// startProcessing(byte zp($21) center_x, byte zp($10) center_y)
 startProcessing: {
-    .label __0 = $11
-    .label __4 = 8
+    .label __0 = $22
+    .label __4 = $1f
     .label __6 = 6
-    .label __8 = $18
-    .label __9 = $18
-    .label __11 = $1a
-    .label __12 = $1a
-    .label __23 = $11
-    .label __24 = 8
+    .label __8 = $16
+    .label __9 = $16
+    .label __11 = $18
+    .label __12 = $18
+    .label __23 = $22
+    .label __24 = $1f
     .label __25 = 6
-    .label __26 = $18
-    .label __27 = $1a
-    .label center_x = $23
+    .label __26 = $16
+    .label __27 = $18
+    .label center_x = $21
     .label center_y = $10
     .label i = 5
-    .label offset = $11
-    .label colPtr = $15
-    .label spriteCol = $17
-    .label screenPtr = $11
-    .label spriteData = 8
+    // Copy char into sprite
+    .label offset = $22
+    .label colPtr = $13
+    .label spriteCol = $15
+    .label screenPtr = $22
+    .label spriteData = $1f
     .label chargenData = 6
-    .label spriteX = $18
-    .label spriteY = $1a
-    .label spritePtr = $1c
+    .label spriteX = $16
+    .label spriteY = $18
+    .label spritePtr = $1a
+    // Busy-wait while finding an empty slot in the PROCESSING array
     // Busy-wait while finding an empty slot in the PROCESSING array
     .label freeIdx = 5
-    .label __33 = $13
-    .label __34 = $11
+    .label __33 = $11
+    .label __34 = $22
     ldx #$ff
   __b1:
     lda #0
@@ -518,19 +528,19 @@ startProcessing: {
 // Find the non-space char closest to the center of the screen
 // If no non-space char is found the distance will be 0xffff
 getCharToProcess: {
-    .label __8 = $1d
-    .label __9 = $1d
-    .label __11 = $1d
+    .label __8 = $1b
+    .label __9 = $1b
+    .label __11 = $1b
     .label screen_line = 6
-    .label dist_line = 8
+    .label dist_line = $1f
     .label y = 5
-    .label return_x = $17
-    .label return_y = $1c
+    .label return_x = $15
+    .label return_y = $1a
     .label closest_dist = $10
-    .label closest_x = $17
-    .label closest_y = $1c
-    .label __12 = $1f
-    .label __13 = $1d
+    .label closest_x = $15
+    .label closest_y = $1a
+    .label __12 = $1d
+    .label __13 = $1b
     // screen_line = SCREEN_COPY
     lda.z SCREEN_COPY
     sta.z screen_line
@@ -738,18 +748,18 @@ initSprites: {
 }
 // Populates 1000 chars (a screen) with values representing the angle to the center.
 // Utilizes symmetry around the  center
-// init_angle_screen(byte* zp($11) screen)
+// init_angle_screen(byte* zp($22) screen)
 init_angle_screen: {
-    .label __7 = $18
-    .label screen = $11
+    .label __7 = $16
+    .label screen = $22
     .label screen_topline = 6
-    .label screen_bottomline = $11
-    .label xw = $1f
-    .label yw = $21
-    .label angle_w = $18
-    .label ang_w = $23
-    .label x = $17
-    .label xb = $1c
+    .label screen_bottomline = $22
+    .label xw = $1d
+    .label yw = $1f
+    .label angle_w = $16
+    .label ang_w = $21
+    .label x = $15
+    .label xb = $1a
     .label y = $10
     // screen_topline = screen+40*12
     lda.z screen
@@ -864,18 +874,19 @@ init_angle_screen: {
 // Find the atan2(x, y) - which is the angle of the line from (0,0) to (x,y)
 // Finding the angle requires a binary search using CORDIC_ITERATIONS_16
 // Returns the angle in hex-degrees (0=0, 0x8000=PI, 0x10000=2*PI)
-// atan2_16(signed word zp($1f) x, signed word zp($21) y)
+// atan2_16(signed word zp($1d) x, signed word zp($1f) y)
 atan2_16: {
-    .label __2 = $13
-    .label __7 = $15
-    .label yi = $13
-    .label xi = $15
-    .label angle = $18
-    .label xd = $1d
-    .label yd = $1a
-    .label return = $18
-    .label x = $1f
-    .label y = $21
+    .label __2 = $11
+    .label __7 = $13
+    .label yi = $11
+    .label xi = $13
+    .label angle = $16
+    // Optimized shift of 2 values: xd=xi>>i; yd=yi>>i
+    .label xd = $1b
+    .label yd = $18
+    .label return = $16
+    .label x = $1d
+    .label y = $1f
     // (y>=0)?y:-y
     lda.z y+1
     bmi !__b1+
@@ -1079,7 +1090,7 @@ atan2_16: {
 // Allocates a block of size chars of memory, returning a pointer to the beginning of the block.
 // The content of the newly allocated block of memory is not initialized, remaining with indeterminate values.
 malloc: {
-    .label mem = $e
+    .label mem = $22
     // mem = heap_head-size
     lda.z heap_head
     sec
@@ -1486,6 +1497,6 @@ VYSIN:
 
   // Sprites currently being processed in the interrupt
   PROCESSING: .fill $e*NUM_PROCESSING, 0
-  __2: .word 0, 0, 0, 0
+  __0: .word 0, 0, 0, 0
   .byte 0, 0, 0, STATUS_FREE
   .word 0

src/test/ref/complex/polygon/polygon.asm

@@ -5,7 +5,7 @@
 // - Up-to-down EOR filling 
 // - Double buffering
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   // Value that disables all CIA interrupts when stored to the CIA Interrupt registers
   .const CIA_INTERRUPT_CLEAR = $7f
@@ -31,7 +31,6 @@
   .const OFFSET_STRUCT_MOS6569_VICII_IRQ_ENABLE = $1a
   .const OFFSET_STRUCT_MOS6569_VICII_MEMORY = $18
   .const OFFSET_STRUCT_MOS6569_VICII_IRQ_STATUS = $19
-  .const toD0181_return = (>(SCREEN&$3fff)*4)|(>CANVAS2)/4&$f
   // The VIC-II MOS 6567/6569
   .label VICII = $d000
   // Color Ram
@@ -56,27 +55,31 @@
   // The default charset address
   .label PETSCII = $1000
   .label COSTAB = SINTAB+$40
-  .label canvas_show_memory = $11
-  .label canvas_show_flag = $12
-__bbegin:
-  // canvas_show_memory = toD018(SCREEN, CANVAS2)
   // The current canvas being rendered to the screen - in D018 format.
-  lda #toD0181_return
-  sta.z canvas_show_memory
-  // canvas_show_flag = 0
+  .label canvas_show_memory = $11
   // Flag signalling that the canvas on screen needs to be updated.
   // Set to 1 by the renderer when a new canvas is ready for showing, and to 0 by the raster when the canvas is shown on screen.
-  lda #0
-  sta.z canvas_show_flag
-  jsr main
-  rts
+  .label canvas_show_flag = $12
+_start: {
+    .const _init1_toD0181_return = (>(SCREEN&$3fff)*4)|(>CANVAS2)/4&$f
+    // canvas_show_memory = toD018(SCREEN, CANVAS2)
+    lda #_init1_toD0181_return
+    sta.z canvas_show_memory
+    // canvas_show_flag = 0
+    lda #0
+    sta.z canvas_show_flag
+    jsr main
+    rts
+}
 main: {
     .const toD0181_return = (>(SCREEN&$3fff)*4)|(>CANVAS1)/4&$f
     .const toD0182_return = (>(SCREEN&$3fff)*4)|(>CANVAS2)/4&$f
     .label cols = 3
     // Setup 16x16 canvas for rendering
+    // Setup 16x16 canvas for rendering
     .label screen = 5
     .label y = 2
+    // Plot in line buffer
     .label x0 = $13
     .label y0 = $14
     .label x1 = $c
@@ -87,6 +90,7 @@ main: {
     .label p1_idx = 8
     .label p2_idx = 9
     // The current canvas being rendered to
+    // The current canvas being rendered to
     .label canvas = $a
     // memset(CONSOLE, ' ', 40*25)
   // Clear the console
@@ -148,9 +152,6 @@ main: {
     // setup_irq()
     // Set-up the raster IRQ
     jsr setup_irq
-    // textcolor(WHITE)
-  // Set text color
-    jsr textcolor
     lda #<CANVAS1
     sta.z canvas
     lda #>CANVAS1
@@ -266,8 +267,6 @@ main: {
     // Set flag used to signal when the canvas has been shown
     lda #1
     sta.z canvas_show_flag
-    // clock()
-    jsr clock
     jmp __b8
   __b2:
     ldx.z y
@@ -309,12 +308,6 @@ main: {
     iny
     jmp __b4
 }
-// Returns the processor clock time used since the beginning of an implementation defined era (normally the beginning of the program).
-// This uses CIA #2 Timer A+B on the C64, and must be initialized using clock_start()
-clock: {
-    // }
-    rts
-}
 // EOR fill from the line buffer onto the canvas
 // eorfill(byte* zp($1f) canvas)
 eorfill: {
@@ -387,13 +380,19 @@ line: {
     .label dy = $18
     .label sx = $19
     .label sy = $1a
+    // Find the canvas column
     .label plot1_column = $21
     .label plot2_y = $1b
+    // Find the canvas column
     .label plot2_column = $1c
+    // Find the canvas column
     .label plot3_column = $1f
     .label e1 = $e
+    // Find the canvas column
     .label plot4_column = $23
+    // Find the canvas column
     .label plot5_column = $25
+    // Find the canvas column
     .label plot6_column = $28
     // abs_u8(x2-x1)
     lda.z x2
@@ -803,10 +802,6 @@ clock_start: {
     // }
     rts
 }
-// Set the color for text output. The old color setting is returned.
-textcolor: {
-    rts
-}
 // Setup raster IRQ to change charset at different lines
 setup_irq: {
     // asm

src/test/ref/complex/prebob/grid-bobs.asm

@@ -33,17 +33,18 @@
   // The BOB charset
   .label BOB_CHARSET = $2000
   // BOB charset ID of the next glyph to be added
-  .label bob_charset_next_id = $11
+  .label bob_charset_next_id = $15
   // Current index within the progress cursor (0-7)
-  .label progress_idx = $10
+  .label progress_idx = $e
   // Current position of the progress cursor
-  .label progress_cursor = $e
+  .label progress_cursor = $c
   // Pointer to the next clean-up to add
   // Prepare for next clean-up
-  .label renderBobCleanupNext = $c
+  .label renderBobCleanupNext = $f
 main: {
     .const origY = $a00
     // Row and column offset vectors
+    // Row and column offset vectors
     .const rowOffsetX = $c00
     .const colOffsetX = $100
     .const colOffsetY = $1800
@@ -57,9 +58,10 @@ main: {
     .label y_1 = $a
     .label rowX = 4
     .label rowY = 6
-    .label col = $10
+    .label col = $e
     // Origin point
-    .label origX = $e
+    // Origin point
+    .label origX = $c
     .label rowOffsetY = 2
     // mulf_init()
     jsr mulf_init
@@ -255,16 +257,16 @@ keyboard_matrix_read: {
 // Render a single BOB at a given x/y-position
 // X-position is 0-151. Each x-position is 2 pixels wide.
 // Y-position is 0-183. Each y-position is 1 pixel high.
-// renderBob(byte zp($16) xpos, byte zp($17) ypos)
+// renderBob(byte zp($13) xpos, byte zp($14) ypos)
 renderBob: {
-    .label __2 = $19
-    .label __5 = $1b
-    .label xpos = $16
-    .label ypos = $17
-    .label x_char_offset = $18
-    .label y_offset = $19
-    .label screen = $19
-    .label bob_table_idx = $1b
+    .label __2 = $16
+    .label __5 = $18
+    .label xpos = $13
+    .label ypos = $14
+    .label x_char_offset = $15
+    .label y_offset = $16
+    .label screen = $16
+    .label bob_table_idx = $18
     // x_char_offset = xpos/BOB_SHIFTS_X
     lda.z xpos
     lsr
@@ -377,7 +379,7 @@ renderBob: {
 }
 // Clean Up the rendered BOB's
 renderBobCleanup: {
-    .label screen = $1c
+    .label screen = $19
     ldx #0
   __b1:
     // screen = RENDERBOB_CLEANUP[i]
@@ -429,7 +431,7 @@ memset: {
     .const num = $3e8
     .label str = BOB_SCREEN
     .label end = str+num
-    .label dst = $19
+    .label dst = $f
     lda #<str
     sta.z dst
     lda #>str
@@ -458,10 +460,10 @@ memset: {
 }
 // Initialize the tables used by renderBob()
 renderBobInit: {
-    .label __0 = $1e
-    .label __5 = $1e
-    .label __6 = $20
-    .label __7 = $1e
+    .label __0 = $1b
+    .label __5 = $1b
+    .label __6 = $1d
+    .label __7 = $1b
     ldx #0
   __b1:
     // ((unsigned int)y)*40
@@ -522,15 +524,14 @@ renderBobInit: {
 // Creates the pre-shifted bobs into BOB_CHARSET and populates the BOB_TABLES
 // Modifies PROTO_BOB by shifting it around
 prepareBobs: {
-    .label bob_table = $20
-    .label shift_y = $12
+    .label bob_table = $1b
+    .label shift_y = $11
     // Populate charset and tables
-    .label bob_glyph = $19
-    .label cell = $17
-    .label bob_table_idx = $15
-    .label shift_x = $16
-    // progress_init(BASIC_SCREEN)
-    jsr progress_init
+    // Populate charset and tables
+    .label bob_glyph = $f
+    .label cell = $14
+    .label bob_table_idx = $12
+    .label shift_x = $13
     // charsetFindOrAddGlyph(PROTO_BOB+48, BOB_CHARSET)
   // Ensure that glyph #0 is empty
     lda #<PROTO_BOB+$30
@@ -665,10 +666,10 @@ progress_inc: {
 }
 // Looks through a charset to find a glyph if present. If not present it is added.
 // Returns the glyph ID
-// charsetFindOrAddGlyph(byte* zp($19) glyph)
+// charsetFindOrAddGlyph(byte* zp($f) glyph)
 charsetFindOrAddGlyph: {
-    .label glyph = $19
-    .label glyph_cursor = $13
+    .label glyph = $f
+    .label glyph_cursor = $1d
     lda #<BOB_CHARSET
     sta.z glyph_cursor
     lda #>BOB_CHARSET
@@ -733,7 +734,7 @@ charsetFindOrAddGlyph: {
 }
 // Shift PROTO_BOB right one X pixel
 protoBobShiftRight: {
-    .label carry = $1b
+    .label carry = $14
     .label i = $18
     ldy #0
     ldx #0
@@ -819,25 +820,25 @@ protoBobShiftDown: {
     dex
     jmp __b1
 }
-// Initialize the PETSCII progress bar
-progress_init: {
-    // }
-    rts
-}
 // Initialize the mulf_sqr multiplication tables with f(x)=int(x*x/4)
 mulf_init: {
     // x/2
-    .label c = $12
+    // x/2
+    .label c = $11
     // Counter used for determining x%2==0
-    .label sqr1_hi = $13
+    // Counter used for determining x%2==0
+    .label sqr1_hi = $1b
     // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
-    .label sqr = $1e
-    .label sqr1_lo = $20
+    // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
+    .label sqr = $19
+    .label sqr1_lo = $f
     // Decrease or increase x_255 - initially we decrease
-    .label sqr2_hi = $1c
-    .label sqr2_lo = $19
+    // Decrease or increase x_255 - initially we decrease
+    .label sqr2_hi = $16
+    .label sqr2_lo = $1d
     //Start with g(0)=f(255)
-    .label dir = $15
+    //Start with g(0)=f(255)
+    .label dir = $12
     ldx #0
     lda #<mulf_sqr1_hi+1
     sta.z sqr1_hi

src/test/ref/complex/prebob/vogel-bobs.asm

@@ -34,29 +34,30 @@
   .label BOB_CHARSET = $2000
   .label COS = SIN+$40
   // BOB charset ID of the next glyph to be added
-  .label bob_charset_next_id = 9
+  .label bob_charset_next_id = 7
   // Current index within the progress cursor (0-7)
-  .label progress_idx = 8
+  .label progress_idx = 6
   // Current position of the progress cursor
-  .label progress_cursor = 6
+  .label progress_cursor = 4
   // Pointer to the next clean-up to add
   // Prepare for next clean-up
-  .label renderBobCleanupNext = 3
+  .label renderBobCleanupNext = 8
 main: {
     .const vicSelectGfxBank1_toDd001_return = 3
     .const vicSelectGfxBank2_toDd001_return = 3
     .const toD0181_return = (>(BOB_SCREEN&$3fff)*4)|(>BOB_CHARSET)/4&$f
     .const toD0182_return = (>(SCREEN_BASIC&$3fff)*4)|(>CHARSET_BASIC)/4&$f
-    .label __10 = 6
-    .label __12 = 6
-    .label __13 = 6
+    .label __10 = 4
+    .label __12 = 4
+    .label __13 = 4
     .label x = $c
-    .label y = 6
+    .label y = 4
     .label a = 2
-    .label r = 9
-    .label i = 5
+    .label r = 7
+    .label i = 3
     // Render Rotated BOBs
-    .label angle = 8
+    // Render Rotated BOBs
+    .label angle = 6
     // mulf_init()
     jsr mulf_init
     // prepareBobs()
@@ -344,7 +345,7 @@ renderBob: {
 // Fast multiply two signed chars to a unsigned int result
 // mulf8s(signed byte register(A) a, signed byte register(X) b)
 mulf8s: {
-    .label return = 6
+    .label return = 4
     // mulf8u_prepare((char)a)
     jsr mulf8u_prepare
     // mulf8s_prepared(b)
@@ -358,7 +359,7 @@ mulf8s: {
 // mulf8s_prepared(signed byte zp($13) b)
 mulf8s_prepared: {
     .label memA = $fd
-    .label m = 6
+    .label m = 4
     .label b = $13
     // mulf8u_prepared((char) b)
     lda.z b
@@ -395,7 +396,7 @@ mulf8s_prepared: {
 mulf8u_prepared: {
     .label resL = $fe
     .label memB = $ff
-    .label return = 6
+    .label return = 4
     // *memB = b
     sta memB
     // asm
@@ -488,7 +489,7 @@ memset: {
     .const num = $3e8
     .label str = BOB_SCREEN
     .label end = str+num
-    .label dst = $c
+    .label dst = 8
     lda #<str
     sta.z dst
     lda #>str
@@ -517,10 +518,10 @@ memset: {
 }
 // Initialize the tables used by renderBob()
 renderBobInit: {
-    .label __0 = $16
-    .label __5 = $16
-    .label __6 = $18
-    .label __7 = $16
+    .label __0 = $14
+    .label __5 = $14
+    .label __6 = $16
+    .label __7 = $14
     ldx #0
   __b1:
     // ((unsigned int)y)*40
@@ -581,15 +582,14 @@ renderBobInit: {
 // Creates the pre-shifted bobs into BOB_CHARSET and populates the BOB_TABLES
 // Modifies PROTO_BOB by shifting it around
 prepareBobs: {
-    .label bob_table = $16
+    .label bob_table = $14
     .label shift_y = $a
     // Populate charset and tables
-    .label bob_glyph = $c
+    // Populate charset and tables
+    .label bob_glyph = 8
     .label cell = $f
     .label bob_table_idx = $b
     .label shift_x = $e
-    // progress_init(SCREEN_BASIC)
-    jsr progress_init
     // bobCharsetFindOrAddGlyph(PROTO_BOB+48)
   // Ensure that glyph #0 is empty
     lda #<PROTO_BOB+$30
@@ -725,10 +725,10 @@ progress_inc: {
 // Looks through BOB_CHARSET to find the passed bob glyph if present.
 // If not present it is added
 // Returns the glyph ID
-// bobCharsetFindOrAddGlyph(byte* zp($c) bob_glyph)
+// bobCharsetFindOrAddGlyph(byte* zp(8) bob_glyph)
 bobCharsetFindOrAddGlyph: {
-    .label bob_glyph = $c
-    .label glyph_cursor = $18
+    .label bob_glyph = 8
+    .label glyph_cursor = $16
     lda #<BOB_CHARSET
     sta.z glyph_cursor
     lda #>BOB_CHARSET
@@ -879,23 +879,23 @@ shiftProtoBobDown: {
     dex
     jmp __b1
 }
-// Initialize the PETSCII progress bar
-progress_init: {
-    // }
-    rts
-}
 // Initialize the mulf_sqr multiplication tables with f(x)=int(x*x/4)
 mulf_init: {
+    // x/2
     // x/2
     .label c = $a
     // Counter used for determining x%2==0
-    .label sqr1_hi = $18
+    // Counter used for determining x%2==0
+    .label sqr1_hi = $14
     // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
-    .label sqr = $14
-    .label sqr1_lo = $16
+    // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
+    .label sqr = $10
+    .label sqr1_lo = 8
     // Decrease or increase x_255 - initially we decrease
-    .label sqr2_hi = $10
-    .label sqr2_lo = $c
+    // Decrease or increase x_255 - initially we decrease
+    .label sqr2_hi = $c
+    .label sqr2_lo = $16
+    //Start with g(0)=f(255)
     //Start with g(0)=f(255)
     .label dir = $b
     ldx #0

src/test/ref/complex/prebob/vogel-sprites.asm

@@ -3,7 +3,7 @@
 // The MOS 6526 Complex Interface Adapter (CIA)
 // http://archive.6502.org/datasheets/mos_6526_cia_recreated.pdf
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .const VIC_RST8 = $80
   .const VIC_DEN = $10
@@ -32,28 +32,29 @@
   .label COS = SIN+$40
   // The address of the sprite pointers on the current screen (screen+0x3f8).
   .label PLEX_SCREEN_PTR = SCREEN+$3f8
-  .label plex_show_idx = $11
-  .label plex_sprite_idx = $12
-  .label plex_sprite_msb = $13
-  .label plex_free_next = $14
-__bbegin:
-  // plex_show_idx=0
   // The index in the PLEX tables of the next sprite to show
-  lda #0
-  sta.z plex_show_idx
-  // plex_sprite_idx=0
+  .label plex_show_idx = $11
   // The index the next sprite to use for showing (sprites are used round-robin)
-  sta.z plex_sprite_idx
-  // plex_sprite_msb=1
+  .label plex_sprite_idx = $12
   // The MSB bit of the next sprite to use for showing
-  lda #1
-  sta.z plex_sprite_msb
-  // plex_free_next = 0
+  .label plex_sprite_msb = $13
   // The index of the sprite that is free next. Since sprites are used round-robin this moves forward each time a sprite is shown.
-  lda #0
-  sta.z plex_free_next
-  jsr main
-  rts
+  .label plex_free_next = $14
+_start: {
+    // plex_show_idx=0
+    lda #0
+    sta.z plex_show_idx
+    // plex_sprite_idx=0
+    sta.z plex_sprite_idx
+    // plex_sprite_msb=1
+    lda #1
+    sta.z plex_sprite_msb
+    // plex_free_next = 0
+    lda #0
+    sta.z plex_free_next
+    jsr main
+    rts
+}
 main: {
     // asm
     sei
@@ -122,6 +123,7 @@ loop: {
     .label r = 7
     .label i = 2
     // Render Rotated BOBs
+    // Render Rotated BOBs
     .label angle = 4
     .label plexFreeNextYpos1_return = $15
     .label i1 = 3
@@ -623,17 +625,22 @@ memset: {
 }
 // Initialize the mulf_sqr multiplication tables with f(x)=int(x*x/4)
 mulf_init: {
+    // x/2
     // x/2
     .label c = 7
     // Counter used for determining x%2==0
+    // Counter used for determining x%2==0
     .label sqr1_hi = 8
     // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
+    // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
     .label sqr = $f
     .label sqr1_lo = 5
     // Decrease or increase x_255 - initially we decrease
+    // Decrease or increase x_255 - initially we decrease
     .label sqr2_hi = $c
     .label sqr2_lo = $a
     //Start with g(0)=f(255)
+    //Start with g(0)=f(255)
     .label dir = $e
     ldx #0
     lda #<mulf_sqr1_hi+1

src/test/ref/complex/splines/truetype-splines.asm

@@ -311,7 +311,9 @@ bitmap_line: {
     .label dy = $1a
     .label sx = $20
     .label sy = $1e
+    // X is the driver
     .label e1 = 9
+    // Y is the driver
     .label e = 7
     .label x1 = $f
     .label y1 = $11
@@ -958,8 +960,11 @@ rotate: {
     .label return_x = 3
     .label return_y = 5
     .label cos_a = $18
+    // signed fixed[0.7]
     .label xr = $1a
+    // signed fixed[8.8]
     .label yr = $1c
+    // signed fixed[8.8]
     .label sin_a = $18
     // cos_a = (signed int) COS[angle]
     lda COS,y
@@ -1404,17 +1409,22 @@ bitmap_init: {
 }
 // Initialize the mulf_sqr multiplication tables with f(x)=int(x*x/4)
 mulf_init: {
+    // x/2
     // x/2
     .label c = $26
     // Counter used for determining x%2==0
+    // Counter used for determining x%2==0
     .label sqr1_hi = $16
     // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
+    // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
     .label sqr = $18
     .label sqr1_lo = $14
     // Decrease or increase x_255 - initially we decrease
+    // Decrease or increase x_255 - initially we decrease
     .label sqr2_hi = $20
     .label sqr2_lo = $1c
     //Start with g(0)=f(255)
+    //Start with g(0)=f(255)
     .label dir = $13
     ldx #0
     lda #<mulf_sqr1_hi+1

src/test/ref/complex/spritescroller/spritescroller.asm

@@ -3,7 +3,7 @@
 // The MOS 6526 Complex Interface Adapter (CIA)
 // http://archive.6502.org/datasheets/mos_6526_cia_recreated.pdf
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   // Value that disables all CIA interrupts when stored to the CIA Interrupt registers
   .const CIA_INTERRUPT_CLEAR = $7f
@@ -17,7 +17,8 @@
   // The number of sprites in the multiplexer
   .const PLEX_COUNT = $20
   .const OFFSET_STRUCT_MOS6526_CIA_INTERRUPT = $d
-  .const toSpritePtr1_return = SPRITES/$40
+  // Sprite pointer for sprite 0
+  .const SPRITE_0 = SPRITES/$40
   .label SPRITES_XPOS = $d000
   .label SPRITES_YPOS = $d001
   .label SPRITES_XMSB = $d010
@@ -43,44 +44,45 @@
   .label SCREEN = $400
   // The high-value table
   .label XMOVEMENT_HI = XMOVEMENT+$100
-  // The address of the sprite pointers on the current screen (screen+0x3f8).
-  .label PLEX_SCREEN_PTR = $400+$3f8
-  .label plex_show_idx = $d
-  .label plex_sprite_idx = $e
-  .label plex_sprite_msb = $f
-  .label plex_free_next = $10
-  .label frame_done = $11
-  // The next char to use from the scroll text
-  .label scroll_text_next = 4
-  // Y-sine index
-  .label y_sin_idx = 2
-  // X-movement index
-  .label x_movement_idx = 3
-__bbegin:
-  // plex_show_idx=0
   // The index in the PLEX tables of the next sprite to show
-  lda #0
-  sta.z plex_show_idx
-  // plex_sprite_idx=0
+  .label plex_show_idx = $10
   // The index the next sprite to use for showing (sprites are used round-robin)
-  sta.z plex_sprite_idx
-  // plex_sprite_msb=1
+  .label plex_sprite_idx = $11
   // The MSB bit of the next sprite to use for showing
-  lda #1
-  sta.z plex_sprite_msb
-  // plex_free_next = 0
+  .label plex_sprite_msb = $12
   // The index of the sprite that is free next. Since sprites are used round-robin this moves forward each time a sprite is shown.
-  lda #0
-  sta.z plex_free_next
-  // frame_done = false
+  .label plex_free_next = $13
   // Signal used between IRQ and main loop. Set to true when the IRQ is done showing the sprites.
-  sta.z frame_done
-  jsr main
-  rts
+  .label frame_done = $14
+  // The next char to use from the scroll text
+  .label scroll_text_next = 5
+  // Y-sine index
+  .label y_sin_idx = 3
+  // X-movement index
+  .label x_movement_idx = 4
+  // The address of the sprite pointers on the current screen (screen+0x3f8).
+  .label PLEX_SCREEN_PTR = $e
+_start: {
+    // plex_show_idx=0
+    lda #0
+    sta.z plex_show_idx
+    // plex_sprite_idx=0
+    sta.z plex_sprite_idx
+    // plex_sprite_msb=1
+    lda #1
+    sta.z plex_sprite_msb
+    // plex_free_next = 0
+    lda #0
+    sta.z plex_free_next
+    // frame_done = false
+    sta.z frame_done
+    jsr main
+    rts
+}
 main: {
     .const toD0181_return = (>(SCREEN&$3fff)*4)|(>CHARSET_DEFAULT)/4&$f
     .label s = 2
-    .label __13 = $12
+    .label __13 = $15
     // asm
     // Create 2x2 font from CHARGEN
     sei
@@ -189,7 +191,7 @@ main: {
     jmp __b6
   __b2:
     // PLEX_PTR[s] = SPRITE_0+' '
-    lda #toSpritePtr1_return+' '
+    lda #SPRITE_0+' '
     ldy.z s
     sta PLEX_PTR,y
     // PLEX_XPOS[s] = { XMOVEMENT_HI[x], XMOVEMENT[x] }
@@ -220,9 +222,9 @@ main: {
 //     elements before the marker are shifted right one at a time until encountering one smaller than the current one.
 //      It is then inserted at the spot. Now the marker can move forward.
 plexSort: {
-    .label nxt_idx = $14
-    .label nxt_y = $15
-    .label m = $c
+    .label nxt_idx = $17
+    .label nxt_y = $18
+    .label m = $d
     lda #0
     sta.z m
   __b1:
@@ -289,10 +291,10 @@ plexSort: {
 }
 // Move the plex sprites in an Y-sine and scroll them to the left.
 plex_move: {
-    .label y_idx = $c
-    .label x_idx = $15
-    .label s = $14
-    .label __7 = $16
+    .label y_idx = $d
+    .label x_idx = $18
+    .label s = $17
+    .label __7 = $19
     // y_idx = y_sin_idx
     lda.z y_sin_idx
     sta.z y_idx
@@ -339,7 +341,7 @@ plex_move: {
     sta.z scroll_text_next+1
   __b3:
     // SPRITE_0+*scroll_text_next++
-    lda #toSpritePtr1_return
+    lda #SPRITE_0
     clc
     ldy #0
     adc (scroll_text_next),y
@@ -391,13 +393,14 @@ plexInit: {
 // - num_chars The number of chars to convert
 font_2x2_to_sprites: {
     .const num_chars = $40
-    .label __3 = $18
-    .label char_right = $a
-    .label sprite_idx = $15
-    .label char_left = 8
-    .label char_current = $16
-    .label sprite = 6
-    .label c = $14
+    .label __3 = $1b
+    .label char_right = $b
+    .label sprite_idx = $18
+    // Upper char
+    .label char_left = 9
+    .label char_current = $19
+    .label sprite = 7
+    .label c = $17
     lda #<SPRITES
     sta.z sprite
     lda #>SPRITES
@@ -519,18 +522,18 @@ font_2x2_to_sprites: {
 // - 0x80 - 0xbf Lower left glyphs
 // - 0xc0 - 0xff Lower right glyphs
 font_2x2: {
-    .label __5 = $12
-    .label __7 = $12
-    .label next_2x2_left = $16
-    .label next_2x2_right = $a
-    .label glyph_bits = $c
-    .label glyph_bits_2x2 = $12
-    .label l2 = $18
-    .label l = $15
-    .label next_2x2_left_1 = 8
-    .label next_2x2 = $16
-    .label next_original = 6
-    .label c = $14
+    .label __5 = $15
+    .label __7 = $15
+    .label next_2x2_left = $19
+    .label next_2x2_right = $b
+    .label glyph_bits = $d
+    .label glyph_bits_2x2 = $15
+    .label l2 = $1b
+    .label l = $18
+    .label next_2x2_left_1 = 9
+    .label next_2x2 = $19
+    .label next_original = 7
+    .label c = $17
     lda #0
     sta.z c
     lda #<CHARGEN
@@ -679,7 +682,7 @@ font_2x2: {
 }
 // Show sprites from the multiplexer, rescheduling the IRQ as many times as needed
 plex_irq: {
-    .label __4 = $19
+    .label __4 = $1c
     // asm
     sei
   __b3:
@@ -731,7 +734,7 @@ plex_irq: {
 // Show the next sprite.
 // plexSort() prepares showing the sprites
 plexShowSprite: {
-    .label plex_sprite_idx2 = $1a
+    .label plex_sprite_idx2 = $1d
     // plex_sprite_idx2 = plex_sprite_idx*2
     lda.z plex_sprite_idx
     asl
@@ -760,8 +763,8 @@ plexShowSprite: {
     // PLEX_SCREEN_PTR[plex_sprite_idx] = PLEX_PTR[PLEX_SORTED_IDX[plex_show_idx]]
     ldy PLEX_SORTED_IDX,x
     lda PLEX_PTR,y
-    ldx.z plex_sprite_idx
-    sta PLEX_SCREEN_PTR,x
+    ldy.z plex_sprite_idx
+    sta (PLEX_SCREEN_PTR),y
     // xpos_idx = PLEX_SORTED_IDX[plex_show_idx]
     ldy.z plex_show_idx
     lda PLEX_SORTED_IDX,y

src/test/ref/complex/tetris/test-sprites.asm

@@ -4,7 +4,7 @@
 // The MOS 6526 Complex Interface Adapter (CIA)
 // http://archive.6502.org/datasheets/mos_6526_cia_recreated.pdf
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   // Value that disables all CIA interrupts when stored to the CIA Interrupt registers
   .const CIA_INTERRUPT_CLEAR = $7f
@@ -24,7 +24,6 @@
   .const IRQ_RASTER_FIRST = SPRITES_FIRST_YPOS+$13
   .const OFFSET_STRUCT_MOS6526_CIA_PORT_A_DDR = 2
   .const OFFSET_STRUCT_MOS6526_CIA_INTERRUPT = $d
-  .const toSpritePtr1_return = PLAYFIELD_SPRITES/$40
   .label SPRITES_XPOS = $d000
   .label SPRITES_YPOS = $d001
   .label SPRITES_COLOR = $d027
@@ -59,48 +58,43 @@
   .label PLAYFIELD_SPRITE_PTRS_1 = PLAYFIELD_SCREEN_1+SPRITE_PTRS
   // Screen Sprite pointers on screen 2
   .label PLAYFIELD_SPRITE_PTRS_2 = PLAYFIELD_SCREEN_2+SPRITE_PTRS
-  // Address of the sprites covering the playfield
-  .label PLAYFIELD_SPRITES = $3000
-  // Address of the charset
-  .label PLAYFIELD_CHARSET = $2800
-  .label SIN_SPRITE = $2800
-  .label render_screen_showing = 5
-  .label irq_raster_next = 6
-  .label irq_sprite_ypos = 7
-  .label irq_sprite_ptr = 8
-  .label irq_cnt = 9
-  .label sin_idx = 3
-__bbegin:
-  // render_screen_showing = 0
   // The screen currently being showed to the user. 0x00 for screen 1 / 0x20 for screen 2.
-  lda #0
-  sta.z render_screen_showing
-  // kickasm
-  // irq_raster_next = IRQ_RASTER_FIRST
+  .label render_screen_showing = 6
   // The raster line of the next IRQ
-  lda #IRQ_RASTER_FIRST
-  sta.z irq_raster_next
-  // irq_sprite_ypos = SPRITES_FIRST_YPOS + 21
+  .label irq_raster_next = 7
   // Y-pos of the sprites on the next IRQ
-  lda #SPRITES_FIRST_YPOS+$15
-  sta.z irq_sprite_ypos
-  // irq_sprite_ptr = toSpritePtr(PLAYFIELD_SPRITES) + 3
+  .label irq_sprite_ypos = 8
   // Index of the sprites to show on the next IRQ
-  lda #toSpritePtr1_return+3
-  sta.z irq_sprite_ptr
-  // irq_cnt = 0
+  .label irq_sprite_ptr = 9
   // Counting the 10 IRQs
-  lda #0
-  sta.z irq_cnt
-  // kickasm
-  jsr main
-  rts
+  .label irq_cnt = $a
+  .label sin_idx = 4
+_start: {
+    .const _init1_toSpritePtr1_return = $ff&PLAYFIELD_SPRITES/$40
+    // render_screen_showing = 0
+    lda #0
+    sta.z render_screen_showing
+    // irq_raster_next = IRQ_RASTER_FIRST
+    lda #IRQ_RASTER_FIRST
+    sta.z irq_raster_next
+    // irq_sprite_ypos = SPRITES_FIRST_YPOS + 21
+    lda #SPRITES_FIRST_YPOS+$15
+    sta.z irq_sprite_ypos
+    // irq_sprite_ptr = toSpritePtr(PLAYFIELD_SPRITES) + 3
+    lda #_init1_toSpritePtr1_return+3
+    sta.z irq_sprite_ptr
+    // irq_cnt = 0
+    lda #0
+    sta.z irq_cnt
+    jsr main
+    rts
+}
 main: {
-    .const toSpritePtr1_return = SIN_SPRITE/$40
+    .const toSpritePtr1_return = $ff&SIN_SPRITE/$40
     .const vicSelectGfxBank1_toDd001_return = 3
     .const toD0181_return = (>(PLAYFIELD_SCREEN_1&$3fff)*4)|(>PLAYFIELD_CHARSET)/4&$f
-    .label xpos = 3
-    .label ypos = 2
+    .label xpos = 2
+    .label ypos = 3
     // CIA2->PORT_A_DDR = %00000011
     lda #3
     sta CIA2+OFFSET_STRUCT_MOS6526_CIA_PORT_A_DDR
@@ -160,7 +154,7 @@ main: {
     rts
 }
 loop: {
-    .label s = 4
+    .label s = 5
     lda #0
     sta.z sin_idx
   __b2:
@@ -238,7 +232,7 @@ sprites_irq_init: {
 }
 // Setup the sprites
 sprites_init: {
-    .label xpos = 4
+    .label xpos = 5
     // *SPRITES_ENABLE = %00001111
     lda #$f
     sta SPRITES_ENABLE
@@ -278,8 +272,9 @@ sprites_init: {
 // Repeats 10 timers every 2 lines from line IRQ_RASTER_FIRST
 // Utilizes duplicated gfx in the sprites to allow for some leeway in updating the sprite pointers
 sprites_irq: {
-    .const toSpritePtr1_return = PLAYFIELD_SPRITES/$40
-    .label raster_sprite_gfx_modify = $a
+    .const toSpritePtr1_return = $ff&PLAYFIELD_SPRITES/$40
+    // Wait for the y-position before changing sprite pointers
+    .label raster_sprite_gfx_modify = $b
     sta rega+1
     stx regx+1
     // asm
@@ -416,9 +411,11 @@ SIN:
   	  .byte 51+AMPL/2+sin(toRadians([i*360]/256))*AMPL/2
     }
 
-.pc = PLAYFIELD_SPRITES "PLAYFIELD_SPRITES"
-  .var sprites = LoadPicture("playfield-sprites.png", List().add($010101, $000000))
-	// Put the sprites into memory 
+.pc = $3000 "PLAYFIELD_SPRITES"
+// Sprites covering the playfield
+PLAYFIELD_SPRITES:
+.var sprites = LoadPicture("playfield-sprites.png", List().add($010101, $000000))
+	// Put the sprites into memory
 	.for(var sy=0;sy<10;sy++) {
 	    .var sprite_gfx_y = sy*20
 		.for(var sx=0;sx<3;sx++) {
@@ -432,6 +429,13 @@ SIN:
 	  	}
 	}
 
-.pc = SIN_SPRITE "SIN_SPRITE"
-  .fill $40, $ff
+.pc = $2800 "PLAYFIELD_CHARSET"
+// Address of the charset
+PLAYFIELD_CHARSET:
+.fill 8,$00 // Place a filled char at the start of the charset
+    .import binary "playfield-screen.imap"
+
+.pc = $3800 "SIN_SPRITE"
+SIN_SPRITE:
+.fill $40, $ff
 

src/test/ref/complex/tetris/tetris.asm

@@ -2,7 +2,7 @@
 // The tetris game tries to match NES tetris gameplay pretty closely
 // Source: https://meatfighter.com/nintendotetrisai/
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   // Value that disables all CIA interrupts when stored to the CIA Interrupt registers
   .const CIA_INTERRUPT_CLEAR = $7f
@@ -66,7 +66,6 @@
   .const OFFSET_STRUCT_MOS6581_SID_CH3_OSC = $1b
   .const OFFSET_STRUCT_MOS6526_CIA_PORT_B = 1
   .const OFFSET_STRUCT_MOS6526_CIA_INTERRUPT = $d
-  .const toSpritePtr1_return = PLAYFIELD_SPRITES/$40
   .label SPRITES_XPOS = $d000
   .label SPRITES_YPOS = $d001
   .label SPRITES_COLOR = $d027
@@ -111,23 +110,25 @@
   .label PLAYFIELD_SPRITE_PTRS_1 = PLAYFIELD_SCREEN_1+SPRITE_PTRS
   // Screen Sprite pointers on screen 2
   .label PLAYFIELD_SPRITE_PTRS_2 = PLAYFIELD_SCREEN_2+SPRITE_PTRS
-  // Address of the sprites covering the playfield
-  .label PLAYFIELD_SPRITES = $3000
-  // Address of the charset
-  .label PLAYFIELD_CHARSET = $2800
-  .label render_screen_showing = $26
-  .label score_bcd = $27
-  .label irq_raster_next = $2b
-  .label irq_sprite_ypos = $2c
-  .label irq_sprite_ptr = $2d
-  .label irq_cnt = $2e
+  // The screen currently being showed to the user. 0x00 for screen 1 / 0x20 for screen 2.
+  .label render_screen_showing = $22
+  // Current score in BCD-format
+  .label score_bcd = $23
+  // The raster line of the next IRQ
+  .label irq_raster_next = $27
+  // Y-pos of the sprites on the next IRQ
+  .label irq_sprite_ypos = $28
+  // Index of the sprites to show on the next IRQ
+  .label irq_sprite_ptr = $29
+  // Counting the 10 IRQs
+  .label irq_cnt = $2a
   // Keyboard event buffer size. The number of events currently in the event buffer
-  .label keyboard_events_size = $20
+  .label keyboard_events_size = $21
   // The rate of moving down the current piece (number of frames between moves if movedown is  not forced)
   .label current_movedown_slow = $11
-  .label current_ypos = $d
   // Position of top left corner of current moving piece on the playfield
   .label current_xpos = $19
+  .label current_ypos = $d
   // The curent piece orientation - each piece have 4 orientations (00/0x10/0x20/0x30).
   // The orientation chooses one of the 4 sub-graphics of the piece.
   .label current_orientation = $16
@@ -137,14 +138,6 @@
   .label current_piece_char = $15
   // Current level BCD-format
   .label level_bcd = $12
-  // The current moving piece. Points to the start of the piece definition.
-  .label current_piece = $13
-  // Is the game over?
-  .label game_over = $1b
-  // The index of the next moving piece. (0-6)
-  .label next_piece_idx = $1a
-  // Current level in normal (non-BCD) format
-  .label level = $10
   // The screen currently being rendered to. 0x00 for screen 1 / 0x20 for screen 2.
   .label render_screen_render = 3
   // The screen currently to show next to the user. 0x00 for screen 1 / 0x20 for screen 2.
@@ -154,6 +147,14 @@
   .label current_movedown_counter = 4
   // Current number of cleared lines in BCD-format
   .label lines_bcd = $e
+  // Current level in normal (non-BCD) format
+  .label level = $10
+  // The current moving piece. Points to the start of the piece definition.
+  .label current_piece = $13
+  // Is the game over?
+  .label game_over = $1b
+  // The index of the next moving piece. (0-6)
+  .label next_piece_idx = $1a
   // The current moving piece. Points to the start of the piece definition.
   .label current_piece_1 = $a
   // The screen currently being rendered to. 0x00 for screen 1 / 0x20 for screen 2.
@@ -164,40 +165,33 @@
   .label current_piece_gfx_1 = 7
   // The char of the current piece
   .label current_piece_char_1 = 9
-__bbegin:
-  // render_screen_showing = 0
-  // The screen currently being showed to the user. 0x00 for screen 1 / 0x20 for screen 2.
-  lda #0
-  sta.z render_screen_showing
-  // score_bcd = 0
-  // Current score in BCD-format
-  sta.z score_bcd
-  sta.z score_bcd+1
-  lda #<0>>$10
-  sta.z score_bcd+2
-  lda #>0>>$10
-  sta.z score_bcd+3
-  // kickasm
-// Tetris Game for the Commodore 64
-// A sprite multiplexer covering the playfield with a black layer to allow for 3 single-pixel colors
-  // irq_raster_next = IRQ_RASTER_FIRST
-  // The raster line of the next IRQ
-  lda #IRQ_RASTER_FIRST
-  sta.z irq_raster_next
-  // irq_sprite_ypos = SPRITES_FIRST_YPOS + 21
-  // Y-pos of the sprites on the next IRQ
-  lda #SPRITES_FIRST_YPOS+$15
-  sta.z irq_sprite_ypos
-  // irq_sprite_ptr = toSpritePtr(PLAYFIELD_SPRITES) + 3
-  // Index of the sprites to show on the next IRQ
-  lda #toSpritePtr1_return+3
-  sta.z irq_sprite_ptr
-  // irq_cnt = 0
-  // Counting the 10 IRQs
-  lda #0
-  sta.z irq_cnt
-  jsr main
-  rts
+_start: {
+    .const _init1_toSpritePtr1_return = $ff&PLAYFIELD_SPRITES/$40
+    // render_screen_showing = 0
+    lda #0
+    sta.z render_screen_showing
+    // score_bcd = 0
+    sta.z score_bcd
+    sta.z score_bcd+1
+    lda #<0>>$10
+    sta.z score_bcd+2
+    lda #>0>>$10
+    sta.z score_bcd+3
+    // irq_raster_next = IRQ_RASTER_FIRST
+    lda #IRQ_RASTER_FIRST
+    sta.z irq_raster_next
+    // irq_sprite_ypos = SPRITES_FIRST_YPOS + 21
+    lda #SPRITES_FIRST_YPOS+$15
+    sta.z irq_sprite_ypos
+    // irq_sprite_ptr = toSpritePtr(PLAYFIELD_SPRITES) + 3
+    lda #_init1_toSpritePtr1_return+3
+    sta.z irq_sprite_ptr
+    // irq_cnt = 0
+    lda #0
+    sta.z irq_cnt
+    jsr main
+    rts
+}
 main: {
     // SID->CH3_FREQ = 0xffff
     lda #<$ffff
@@ -347,7 +341,7 @@ render_score: {
     .const lines_offset = $28*1+$16
     .const level_offset = $28*$13+$1f
     .label score_bytes = score_bcd
-    .label screen = 7
+    .label screen = $2c
     // if(render_screen_render==0)
     lda.z render_screen_render
     cmp #0
@@ -419,12 +413,12 @@ render_score: {
 // - offset: offset on the screen
 // - bcd: The BCD-value to render
 // - only_low: if non-zero only renders the low digit
-// render_bcd(byte* zp(7) screen, word zp($a) offset, byte register(X) bcd, byte register(Y) only_low)
+// render_bcd(byte* zp($2c) screen, word zp($2e) offset, byte register(X) bcd, byte register(Y) only_low)
 render_bcd: {
     .const ZERO_CHAR = $35
-    .label screen = 7
-    .label screen_pos = $a
-    .label offset = $a
+    .label screen = $2c
+    .label screen_pos = $2e
+    .label offset = $2e
     // screen_pos = screen+offset
     lda.z screen_pos
     clc
@@ -468,12 +462,14 @@ render_bcd: {
 }
 // Render the next tetromino in the "next" area
 render_next: {
+    // Find the screen area
     // Find the screen area
     .const next_area_offset = $28*$c+$18+4
-    .label next_piece_char = $3b
-    .label next_piece_gfx = 7
-    .label screen_next_area = $a
-    .label l = 5
+    .label next_piece_char = $32
+    // Render the next piece
+    .label next_piece_gfx = $2c
+    .label screen_next_area = $2e
+    .label l = $c
     // if(render_screen_render==0)
     cpx #0
     beq __b1
@@ -554,8 +550,8 @@ render_next: {
 // Ignores cases where parts of the tetromino is outside the playfield (sides/bottom) since the movement collision routine prevents this.
 render_moving: {
     .label ypos = $c
-    .label screen_line = $30
-    .label xpos = $21
+    .label screen_line = $2c
+    .label xpos = $1e
     .label i = $1d
     .label l = $1c
     // ypos = current_ypos
@@ -621,11 +617,12 @@ render_moving: {
 }
 // Render the static playfield on the screen (all pieces already locked into place)
 render_playfield: {
-    .label screen_line = $a
+    .label screen_line = $2c
     // Do not render the top 2 lines.
-    .label i = $1c
-    .label c = $1d
-    .label l = $c
+    // Do not render the top 2 lines.
+    .label i = $1d
+    .label c = $1e
+    .label l = $1c
     lda #PLAYFIELD_COLS*2
     sta.z i
     lda #2
@@ -672,11 +669,11 @@ render_playfield: {
 // Perform any movement of the current piece
 // key_event is the next keyboard_event() og 0xff if no keyboard event is pending
 // Returns a byte signaling whether rendering is needed. (0 no render, >0 render needed)
-// play_movement(byte zp($2f) key_event)
+// play_movement(byte zp($2b) key_event)
 play_movement: {
-    .label render = $21
-    .label return = $21
-    .label key_event = $2f
+    .label render = $1f
+    .label return = $1f
+    .label key_event = $2b
     // play_move_down(key_event)
     lda.z key_event
     jsr play_move_down
@@ -710,6 +707,7 @@ play_movement: {
 // Return non-zero if a render is needed
 // play_move_rotate(byte register(A) key_event)
 play_move_rotate: {
+    // Handle keyboard events
     // Handle keyboard events
     .label orientation = $c
     // if(key_event==KEY_Z)
@@ -774,21 +772,21 @@ play_move_rotate: {
 play_collision: {
     .label xpos = $1c
     .label ypos = $1d
-    .label piece_gfx = $a
+    .label piece_gfx = $2e
     .label yp = $1d
-    .label playfield_line = $32
-    .label i = $3b
-    .label xp = $3c
+    .label playfield_line = $30
+    .label i = $32
+    .label xp = $3b
     .label l = $1e
-    .label i_1 = $1f
+    .label i_1 = $20
     // piece_gfx = current_piece + orientation
     txa
     clc
-    adc.z piece_gfx
+    adc.z current_piece_1
     sta.z piece_gfx
-    bcc !+
-    inc.z piece_gfx+1
-  !:
+    lda #0
+    adc.z current_piece_1+1
+    sta.z piece_gfx+1
     lda #0
     sta.z l
     sta.z i_1
@@ -1030,7 +1028,9 @@ play_move_down: {
 // Spawn a new piece
 // Moves the next piece into the current and spawns a new next piece
 play_spawn_current: {
-    .label __7 = $34
+    .label __7 = $33
+    // Spawn a new next piece
+    // Pick a random piece (0-6)
     // Spawn a new next piece
     // Pick a random piece (0-6)
     .label piece_idx = $1a
@@ -1090,8 +1090,8 @@ play_spawn_current: {
 // Update the score based on the number of lines removed
 // play_update_score(byte register(X) removed)
 play_update_score: {
-    .label lines_before = $34
-    .label add_bcd = $35
+    .label lines_before = $33
+    .label add_bcd = $34
     // if(removed!=0)
     cpx #0
     beq __breturn
@@ -1221,11 +1221,11 @@ play_increase_level: {
 // Whenever a full line is detected the writing cursor is instructed to write to the same line once more.
 // Returns the number of lines removed
 play_remove_lines: {
-    .label c = $3b
+    .label c = $3a
     .label x = $1e
     .label y = $1c
     .label removed = $1d
-    .label full = $1f
+    .label full = $20
     lda #0
     sta.z removed
     sta.z y
@@ -1290,13 +1290,15 @@ play_remove_lines: {
 }
 // Lock the current piece onto the playfield
 play_lock_current: {
-    .label yp = $d
-    .label playfield_line = $39
-    .label xp = $1f
-    .label i = $3b
-    .label l = $3c
-    .label i_1 = $1e
+    .label yp = $3b
+    .label playfield_line = $38
+    .label xp = $33
+    .label i = $3a
+    .label l = $20
+    .label i_1 = $32
     // yp = current_ypos
+    lda.z current_ypos
+    sta.z yp
     lda #0
     sta.z l
     sta.z i_1
@@ -1396,9 +1398,9 @@ keyboard_event_get: {
 // Handles debounce and only generates events when the status of a key changes.
 // Also stores current status of modifiers in keyboard_modifiers.
 keyboard_event_scan: {
-    .label row_scan = $3c
-    .label keycode = $1f
-    .label row = $1e
+    .label row_scan = $3b
+    .label keycode = $20
+    .label row = $1f
     lda #0
     sta.z keycode
     sta.z row
@@ -1549,9 +1551,10 @@ render_show: {
 }
 // Initialize play data tables
 play_init: {
-    .label pli = $22
+    .label pli = $2c
     // Initialize the playfield line pointers;
-    .label idx = $21
+    // Initialize the playfield line pointers;
+    .label idx = $32
     lda #0
     sta.z idx
     lda #<playfield
@@ -1663,7 +1666,7 @@ sprites_irq_init: {
 }
 // Setup the sprites
 sprites_init: {
-    .label xpos = $21
+    .label xpos = $33
     // *SPRITES_ENABLE = %00001111
     lda #$f
     sta SPRITES_ENABLE
@@ -1701,10 +1704,11 @@ sprites_init: {
 }
 // Initialize rendering
 render_init: {
-    .const vicSelectGfxBank1_toDd001_return = 3
+    .const vicSelectGfxBank1_toDd001_return = 3^(>PLAYFIELD_CHARSET)/$40
     // Initialize the screen line pointers;
-    .label li_1 = $22
-    .label li_2 = $30
+    // Initialize the screen line pointers;
+    .label li_1 = $2c
+    .label li_2 = $2e
     // CIA2->PORT_A_DDR = %00000011
     lda #3
     sta CIA2+OFFSET_STRUCT_MOS6526_CIA_PORT_A_DDR
@@ -1790,14 +1794,14 @@ render_init: {
 }
 // Copy the original screen data to the passed screen
 // Also copies colors to 0xd800
-// render_screen_original(byte* zp($30) screen)
+// render_screen_original(byte* zp($2e) screen)
 render_screen_original: {
     .const SPACE = 0
-    .label screen = $30
-    .label cols = $24
-    .label oscr = $32
-    .label ocols = $39
-    .label y = $2f
+    .label screen = $2e
+    .label cols = $2c
+    .label oscr = $30
+    .label ocols = $38
+    .label y = $2b
     lda #0
     sta.z y
     lda #<PLAYFIELD_COLORS_ORIGINAL+$20*2
@@ -1903,8 +1907,9 @@ render_screen_original: {
 // Repeats 10 timers every 2 lines from line IRQ_RASTER_FIRST
 // Utilizes duplicated gfx in the sprites to allow for some leeway in updating the sprite pointers
 sprites_irq: {
-    .const toSpritePtr1_return = PLAYFIELD_SPRITES/$40
-    .label raster_sprite_gfx_modify = $3d
+    .const toSpritePtr1_return = $ff&PLAYFIELD_SPRITES/$40
+    // Wait for the y-position before changing sprite pointers
+    .label raster_sprite_gfx_modify = $3c
     sta rega+1
     stx regx+1
     // asm
@@ -2113,13 +2118,11 @@ PLAYFIELD_COLORS_ORIGINAL:
   // Score values for removing 0-4 lines (in BCD)
   // These values are updated based on the players level and the base values from SCORE_BASE_BCD
   score_add_bcd: .fill 4*5, 0
-.pc = PLAYFIELD_CHARSET "PLAYFIELD_CHARSET"
-  .fill 8,$00 // Place a filled char at the start of the charset
-    .import binary "playfield-screen.imap"
-
-.pc = PLAYFIELD_SPRITES "PLAYFIELD_SPRITES"
-  .var sprites = LoadPicture("playfield-sprites.png", List().add($010101, $000000))
-	// Put the sprites into memory 
+.pc = $3000 "PLAYFIELD_SPRITES"
+// Sprites covering the playfield
+PLAYFIELD_SPRITES:
+.var sprites = LoadPicture("playfield-sprites.png", List().add($010101, $000000))
+	// Put the sprites into memory
 	.for(var sy=0;sy<10;sy++) {
 	    .var sprite_gfx_y = sy*20
 		.for(var sx=0;sx<3;sx++) {
@@ -2133,3 +2136,9 @@ PLAYFIELD_COLORS_ORIGINAL:
 	  	}
 	}
 
+.pc = $2800 "PLAYFIELD_CHARSET"
+// Address of the charset
+PLAYFIELD_CHARSET:
+.fill 8,$00 // Place a filled char at the start of the charset
+    .import binary "playfield-screen.imap"
+

src/test/ref/complex/xmega65/xmega65.asm

@@ -18,7 +18,22 @@
   .label BG_COLOR = $d021
   .label COLS = $d800
 .segment Code
+syscall2: {
+    // *(SCREEN+78) = '<'
+    lda #'<'
+    sta SCREEN+$4e
+    // }
+    rts
+}
+syscall1: {
+    // *(SCREEN+79) = '>'
+    lda #'>'
+    sta SCREEN+$4f
+    // }
+    rts
+}
 main: {
+    // Print message
     // Print message
     .label sc = 4
     .label msg = 2
@@ -124,20 +139,6 @@ memset: {
   !:
     jmp __b2
 }
-syscall2: {
-    // *(SCREEN+78) = '<'
-    lda #'<'
-    sta SCREEN+$4e
-    // }
-    rts
-}
-syscall1: {
-    // *(SCREEN+79) = '>'
-    lda #'>'
-    sta SCREEN+$4f
-    // }
-    rts
-}
 .segment Data
   MESSAGE: .text "hello world!"
   .byte 0

src/test/ref/complex/xmega65/xmega65logo.asm

@@ -3,6 +3,7 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
+  // Create pointers to the palette RGBs in the logo (assumes dimensions are 128x128)
   // Create pointers to the palette RGBs in the logo (assumes dimensions are 128x128)
   .label LOGO256_RED = LOGO256+$80*$80
   .label LOGO256_GREEN = LOGO256_RED+$100

src/test/ref/const-early-identification.asm

@@ -1,16 +1,17 @@
 // Tests that constants are identified early
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
-  .label A = 2
-__bbegin:
-  // A = 'a'
   // Not an early constant (address-of is used)
-  lda #'a'
-  sta.z A
-  jsr main
-  rts
+  .label A = 2
+_start: {
+    // A = 'a'
+    lda #'a'
+    sta.z A
+    jsr main
+    rts
+}
 main: {
     .const B = 'b'
     .label addrA = A

src/test/ref/constants.asm

@@ -6,6 +6,7 @@
   .const GREEN = 5
   .const RED = 2
   .label BG_COLOR = $d021
+  .label print_screen = $400
   .label print_char_cursor = 2
   .label print_line_cursor = 4
 main: {
@@ -192,16 +193,16 @@ test_bytes: {
     .const bc = 2
     .const bd = bc-4
     // assert_byte("0=0", bb, 0)
-    lda #<$400
+    lda #<print_screen
     sta.z print_line_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_line_cursor+1
     lda #0
     sta.z assert_byte.c
     ldx #bb
-    lda #<$400
+    lda #<print_screen
     sta.z print_char_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_char_cursor+1
     lda #<msg
     sta.z assert_byte.msg
@@ -288,7 +289,7 @@ print_cls: {
 memset: {
     .const c = ' '
     .const num = $3e8
-    .label str = $400
+    .label str = print_screen
     .label end = str+num
     .label dst = 7
     lda #<str

src/test/ref/cordic-atan2-16-ref.asm

@@ -10,11 +10,16 @@
   .label COLS = $d800
   .label CHARSET = $2000
   .label SCREEN = $2800
+  .label print_screen = $400
   .label print_char_cursor = 8
 main: {
     .const toD0181_return = (>(SCREEN&$3fff)*4)|(>CHARSET)/4&$f
     .label col00 = COLS+$c*$28+$13
     .label __4 = $a
+    //byte angle_b = atan2_8(x, y);
+    //diff_sum += diff(angle_b, *screen_ref);
+    //*screen = angle_b - *screen_ref++;
+    //*screen = angle_b;
     .label xw = $17
     .label yw = $19
     .label angle_w = $a
@@ -120,9 +125,9 @@ print_uint: {
     .label w = 4
     // print_uchar(>w)
     ldx.z w+1
-    lda #<$400
+    lda #<print_screen
     sta.z print_char_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_char_cursor+1
     jsr print_uchar
     // print_uchar(<w)
@@ -197,6 +202,7 @@ atan2_16: {
     .label yi = $e
     .label xi = $11
     .label angle = $a
+    // Optimized shift of 2 values: xd=xi>>i; yd=yi>>i
     .label xd = $c
     .label yd = $13
     .label return = $a

src/test/ref/cordic-atan2-16.asm

@@ -90,6 +90,7 @@ atan2_16: {
     .label yi = $a
     .label xi = $d
     .label angle = 6
+    // Optimized shift of 2 values: xd=xi>>i; yd=yi>>i
     .label xd = 8
     .label yd = $f
     .label return = 6

src/test/ref/cordic-atan2-clear.asm

@@ -13,6 +13,7 @@
 main: {
     .const toD0181_return = (>(SCREEN&$3fff)*4)|(>CHARSET)/4&$f
     // Clear the screen by modifying the charset
+    // Clear the screen by modifying the charset
     .label clear_char = 2
     // init_font_hex(CHARSET)
     jsr init_font_hex
@@ -177,6 +178,7 @@ atan2_16: {
     .label yi = $11
     .label xi = 4
     .label angle = 6
+    // Optimized shift of 2 values: xd=xi>>i; yd=yi>>i
     .label xd = $a
     .label yd = 8
     .label return = 6

src/test/ref/declared-memory-var-7.asm

@@ -1,16 +1,17 @@
 // Test declaring a variable as "memory", meaning it will be stored in memory and accessed through an implicit pointer (using load/store)
 // Test a zeropage notregister variable
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
   .label idx = 2
-__bbegin:
-  // idx
-  lda #0
-  sta.z idx
-  jsr main
-  rts
+_start: {
+    // idx
+    lda #0
+    sta.z idx
+    jsr main
+    rts
+}
 main: {
     ldx #0
   __b1:

src/test/ref/declared-memory-var-8.asm

@@ -1,16 +1,17 @@
 // Test declaring a variable as "memory", meaning it will be stored in main memory
 // Test a fixed main memory address __notssa variable
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
   .label idx = $1000
-__bbegin:
-  // idx
-  lda #0
-  sta idx
-  jsr main
-  rts
+_start: {
+    // idx
+    lda #0
+    sta idx
+    jsr main
+    rts
+}
 main: {
     ldx #0
   __b1:

src/test/ref/declared-ssa-var-0.asm

@@ -1,18 +1,19 @@
 // Tests declaring variables as __ssa / __notssa
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN1 = $400
   .label SCREEN2 = $400+$28
   .label SCREEN3 = $400+$50
   .label SCREEN4 = $400+$78
   .label idx_nssa_g = 2
-__bbegin:
-  // idx_nssa_g
-  lda #0
-  sta.z idx_nssa_g
-  jsr main
-  rts
+_start: {
+    // idx_nssa_g
+    lda #0
+    sta.z idx_nssa_g
+    jsr main
+    rts
+}
 main: {
     .label idx_nssa_l = 3
     // idx_nssa_l

src/test/ref/deep-nesting.asm

@@ -700,11 +700,5 @@ f98: {
 }
 f99: {
     .label return = 1
-    // f100(x)
-    jsr f100
-    // }
-    rts
-}
-f100: {
     rts
 }

src/test/ref/empty-function-0.asm

@@ -0,0 +1,13 @@
+// Test removal of empty function
+.pc = $801 "Basic"
+:BasicUpstart(main)
+.pc = $80d "Program"
+  .const v = 7
+  .label SCREEN = $400
+main: {
+    // SCREEN[0] = v
+    lda #v
+    sta SCREEN
+    // }
+    rts
+}

src/test/ref/empty-function-1.asm

@@ -0,0 +1,12 @@
+// Test removal of empty function
+.pc = $801 "Basic"
+:BasicUpstart(main)
+.pc = $80d "Program"
+  .label SCREEN = $400
+main: {
+    // SCREEN[0] = 'x'
+    lda #'x'
+    sta SCREEN
+    // }
+    rts
+}

src/test/ref/empty-function-2.asm

@@ -0,0 +1,9 @@
+// Test removal of empty function
+// main() should not be removed!
+.pc = $801 "Basic"
+:BasicUpstart(main)
+.pc = $80d "Program"
+main: {
+    // }
+    rts
+}

src/test/ref/euclid-3.asm

@@ -6,21 +6,22 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
-  .label print_line_cursor = 4
-  .label print_char_cursor = 6
+  .label print_screen = $400
+  .label print_line_cursor = 6
+  .label print_char_cursor = 4
 main: {
     // print_cls()
     jsr print_cls
     // print_euclid(128,2)
-    lda #<$400
+    lda #<print_screen
     sta.z print_line_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_line_cursor+1
     lda #2
     sta.z print_euclid.b
-    lda #<$400
+    lda #<print_screen
     sta.z print_char_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_char_cursor+1
     lda #$80
     sta.z print_euclid.a
@@ -206,9 +207,9 @@ print_cls: {
 memset: {
     .const c = ' '
     .const num = $3e8
-    .label str = $400
+    .label str = print_screen
     .label end = str+num
-    .label dst = 8
+    .label dst = 6
     lda #<str
     sta.z dst
     lda #>str

src/test/ref/examples/3d/3d.asm

@@ -21,25 +21,25 @@
   // The VIC-II MOS 6567/6569
   .label VICII = $d000
   // The rotated point - updated by calling rotate_matrix()
+  // The rotated point - updated by calling rotate_matrix()
   .label xr = $f0
   .label yr = $f1
   .label zr = $f2
   // The rotated point with perspective
+  // The rotated point with perspective
   .label pp = $f3
   .label xp = $f4
   .label yp = $f5
   // Pointers used to multiply perspective (d/z0-z) onto x- & y-coordinates. Points into mulf_sqr1 / mulf_sqr2  
+  // Pointers used to multiply perspective (d/z0-z) onto x- & y-coordinates. Points into mulf_sqr1 / mulf_sqr2  
   .label psp1 = $f6
   .label psp2 = $f8
   .label SCREEN = $400
-  // A single sprite
-  .label SPRITE = $3000
   .label COSH = SINH+$40
   .label COSQ = SINQ+$40
   .label print_screen = $400
-  .label sx = 2
-  .label sy = 3
-  // kickasm
+  .label sx = 3
+  .label sy = 4
 main: {
     // asm
     sei
@@ -65,7 +65,7 @@ main: {
     rts
 }
 anim: {
-    .label i = 4
+    .label i = 2
     lda #0
     sta.z sy
     sta.z sx
@@ -188,8 +188,8 @@ debug_print: {
     .const print_schar_pos12_row = 6
     .const print_schar_pos12_col = $25
     .label at_line = SCREEN+$13*$28
-    .label c = 5
-    .label i = 6
+    .label c = 6
+    .label i = 7
     // print_schar_pos(sx, 0, 37)
     lda.z sx
     // print_schar_at(sb, print_screen+row*40+col)
@@ -389,9 +389,9 @@ debug_print: {
     rts
 }
 // Print a signed char as hex at a specific screen position
-// print_schar_at(signed byte zp(7) b, byte* zp($e) at)
+// print_schar_at(signed byte zp(5) b, byte* zp($e) at)
 print_schar_at: {
-    .label b = 7
+    .label b = 5
     .label at = $e
     // if(b<0)
     lda.z b
@@ -432,9 +432,9 @@ print_char_at: {
     rts
 }
 // Print a char as HEX at a specific position
-// print_uchar_at(byte zp(7) b, byte* zp($e) at)
+// print_uchar_at(byte zp(5) b, byte* zp($e) at)
 print_uchar_at: {
-    .label b = 7
+    .label b = 5
     .label at = $e
     // b>>4
     lda.z b
@@ -615,17 +615,21 @@ store_matrix: {
 // Prepare the 3x3 rotation matrix into rotation_matrix[]
 // Angles sx, sy, sz are based on 2*PI=$100 
 // Method described in C= Hacking Magazine Issue 8. http://www.ffd2.com/fridge/chacking/c=hacking8.txt
-// calculate_matrix(signed byte register(X) sx, signed byte zp(3) sy)
+// calculate_matrix(signed byte register(X) sx, signed byte zp(4) sy)
 calculate_matrix: {
-    .label sy = 3
+    .label sy = 4
     .label t1 = 5
     .label t2 = 6
     .label t3 = 7
     .label t4 = 8
     .label t5 = 9
+    // = sx+sy+sz
     .label t6 = $a
+    // = sx-sy+sz
     .label t7 = $b
+    // = sx+sy-sz
     .label t8 = $c
+    // = sy+sz-sx
     .label t9 = $d
     // t1 = sy-sz
     lda.z sy
@@ -818,8 +822,8 @@ debug_print_init: {
     .label __59 = $1a
     .label __62 = $1c
     .label __65 = $1e
-    .label c = 4
-    .label i = 5
+    .label c = 3
+    .label i = 4
     // print_cls()
     jsr print_cls
     // print_str_at("sx", SCREEN+40*0+34)
@@ -1210,7 +1214,7 @@ sprites_init: {
     ldx #0
   __b1:
     // sprites_ptr[i] = (char)(SPRITE/$40)
-    lda #SPRITE/$40
+    lda #$ff&SPRITE/$40
     sta sprites_ptr,x
     // SPRITES_COLOR[i] = GREEN
     lda #GREEN
@@ -1301,8 +1305,10 @@ SINQ:
     }
     }
 
-.pc = SPRITE "SPRITE"
-  .var pic = LoadPicture("balloon.png", List().add($000000, $ffffff))
+.pc = $3000 "SPRITE"
+// A single sprite
+SPRITE:
+.var pic = LoadPicture("balloon.png", List().add($000000, $ffffff))
     .for (var y=0; y<21; y++)
         .for (var x=0;x<3; x++)
             .byte pic.getSinglecolorByte(x,y)

src/test/ref/examples/3d/perspective.asm

@@ -8,13 +8,16 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
+  // The rotated point - updated by calling rotate()
   // The rotated point - updated by calling rotate()
   .label xr = $f0
   .label yr = $f1
   .label zr = $f2
   // Pointers used to multiply perspective (d/z0-z) onto x- & y-coordinates. Points into mulf_sqr1 / mulf_sqr2.  
+  // Pointers used to multiply perspective (d/z0-z) onto x- & y-coordinates. Points into mulf_sqr1 / mulf_sqr2.  
   .label psp1 = $f3
   .label psp2 = $f5
+  .label print_screen = $400
   .label print_char_cursor = 4
   .label print_line_cursor = 2
 main: {
@@ -44,9 +47,9 @@ do_perspective: {
     .label y = -$47
     .label z = $36
     // print_str("(")
-    lda #<$400
+    lda #<print_screen
     sta.z print_char_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_char_cursor+1
     lda #<str
     sta.z print_str.str
@@ -115,9 +118,9 @@ do_perspective: {
 }
 // Print a newline
 print_ln: {
-    lda #<$400
+    lda #<print_screen
     sta.z print_line_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_line_cursor+1
   __b1:
     // print_line_cursor + $28
@@ -141,9 +144,9 @@ print_ln: {
     rts
 }
 // Print a zero-terminated string
-// print_str(byte* zp(6) str)
+// print_str(byte* zp(2) str)
 print_str: {
-    .label str = 6
+    .label str = 2
   __b1:
     // while(*str)
     ldy #0
@@ -272,9 +275,9 @@ print_cls: {
 memset: {
     .const c = ' '
     .const num = $3e8
-    .label str = $400
+    .label str = print_screen
     .label end = str+num
-    .label dst = 6
+    .label dst = 4
     lda #<str
     sta.z dst
     lda #>str
@@ -303,9 +306,9 @@ memset: {
 }
 // Initialize the mulf_sqr multiplication tables with f(x)=int(x*x) and g(x) = f(1-x) 
 mulf_init: {
-    .label val = $a
-    .label sqr = 6
-    .label add = 8
+    .label val = 6
+    .label sqr = 2
+    .label add = 4
     lda #<1
     sta.z add
     lda #>1

src/test/ref/examples/atari2600/atari2600-sprites.asm

@@ -8,9 +8,9 @@
 .segmentdef Data [startAfter="Code",max=$fff9]
 .segmentdef Vectors [start=$fffa,max=$ffff]
 .segment Vectors
-.word __bbegin // NMI
-.word __bbegin // RESET
-.word __bbegin // IRQ
+.word _start // NMI
+.word _start // RESET
+.word _start // IRQ
 .segment Code
 
   // The number of CPU cycles per scanline
@@ -34,20 +34,21 @@
   .label TIA = 0
   // Atari RIOT registers
   .label RIOT = $280
+  // Player 0 X position
   .label p0_xpos = $82
   // Counts frames
   .label idx = $80
   // Player 0 Y position
   .label p0_ypos = $83
   .label idx2 = $81
-__bbegin:
-  // p0_xpos
-  // Player 0 X position
-  lda #0
-  sta.z p0_xpos
-  jsr main
-  rts
 .segment Code
+_start: {
+    // p0_xpos
+    lda #0
+    sta.z p0_xpos
+    jsr main
+    rts
+}
 main: {
     // asm
     cld

src/test/ref/examples/bresenham/bitmap-bresenham.asm

@@ -236,6 +236,7 @@ bitmap_line_xdyi: {
 }
 // bitmap_plot(byte register(X) x, byte register(Y) y)
 bitmap_plot: {
+    // Needs unsigned int arrays arranged as two underlying char arrays to allow char* plotter_x = plot_x[x]; - and eventually - char* plotter = plot_x[x] + plot_y[y];
     .label plotter_x = 9
     .label plotter_y = $b
     .label plotter = 9

src/test/ref/examples/chargen/chargen-analysis.asm

@@ -73,8 +73,10 @@ main: {
     .label i = 4
     .label ch = 7
     // Which char canvas to use
+    // Which char canvas to use
     .label cur_pos = 5
     // Is shift pressed
+    // Is shift pressed
     .label shift = 6
     lda #<SCREEN
     sta.z sc

src/test/ref/examples/conio/nacht-screen.asm

@@ -2,7 +2,7 @@
 // From CC65 sample "Eine kleine Nachtmusik" by Ullrich von Bassewitz
 // https://github.com/cc65/cc65/blob/master/samples/nachtm.c
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .const LIGHT_BLUE = $e
   // The horizontal line character
@@ -28,51 +28,52 @@
   .label COLORRAM = $d800
   // Default address of screen character matrix
   .label DEFAULT_SCREEN = $400
+  // The number of bytes on the screen
+  // The current cursor x-position
   .label conio_cursor_x = $b
+  // The current cursor y-position
   .label conio_cursor_y = $c
+  // The current text cursor line start
   .label conio_line_text = $d
+  // The current color cursor line start
   .label conio_line_color = $f
+  // The current text color
   .label conio_textcolor = $11
+  // Is scrolling enabled when outputting beyond the end of the screen (1: yes, 0: no).
+  // If disabled the cursor just moves back to (0,0) instead
   .label conio_scroll_enable = $12
   .label XSize = $13
   .label YSize = $14
-__bbegin:
-  // conio_cursor_x = 0
-  // The number of bytes on the screen
-  // The current cursor x-position
-  lda #0
-  sta.z conio_cursor_x
-  // conio_cursor_y = 0
-  // The current cursor y-position
-  sta.z conio_cursor_y
-  // conio_line_text = CONIO_SCREEN_TEXT
-  // The current text cursor line start
-  lda #<DEFAULT_SCREEN
-  sta.z conio_line_text
-  lda #>DEFAULT_SCREEN
-  sta.z conio_line_text+1
-  // conio_line_color = CONIO_SCREEN_COLORS
-  // The current color cursor line start
-  lda #<COLORRAM
-  sta.z conio_line_color
-  lda #>COLORRAM
-  sta.z conio_line_color+1
-  // conio_textcolor = CONIO_TEXTCOLOR_DEFAULT
-  // The current text color
-  lda #LIGHT_BLUE
-  sta.z conio_textcolor
-  // conio_scroll_enable = 1
-  // Is scrolling enabled when outputting beyond the end of the screen (1: yes, 0: no).
-  // If disabled the cursor just moves back to (0,0) instead
-  lda #1
-  sta.z conio_scroll_enable
-  // XSize
-  lda #0
-  sta.z XSize
-  // YSize
-  sta.z YSize
-  jsr main
-  rts
+_start: {
+    // conio_cursor_x = 0
+    lda #0
+    sta.z conio_cursor_x
+    // conio_cursor_y = 0
+    sta.z conio_cursor_y
+    // conio_line_text = CONIO_SCREEN_TEXT
+    lda #<DEFAULT_SCREEN
+    sta.z conio_line_text
+    lda #>DEFAULT_SCREEN
+    sta.z conio_line_text+1
+    // conio_line_color = CONIO_SCREEN_COLORS
+    lda #<COLORRAM
+    sta.z conio_line_color
+    lda #>COLORRAM
+    sta.z conio_line_color+1
+    // conio_textcolor = CONIO_TEXTCOLOR_DEFAULT
+    lda #LIGHT_BLUE
+    sta.z conio_textcolor
+    // conio_scroll_enable = 1
+    lda #1
+    sta.z conio_scroll_enable
+    // XSize
+    lda #0
+    sta.z XSize
+    // YSize
+    sta.z YSize
+    jsr main
+    rts
+}
 main: {
     // *VIC_MEMORY = 0x17
     lda #$17
@@ -193,8 +194,6 @@ MakeNiceScreen: {
     jsr bgcolor
     // clrscr ()
     jsr clrscr
-    // cursor (0)
-    jsr cursor
     // cputcxy (0, 0, CH_ULCORNER)
   /* Top line */
     ldy #CH_ULCORNER
@@ -782,13 +781,6 @@ cvline: {
     inc.z i
     jmp __b1
 }
-// If onoff is 1, a cursor is displayed when waiting for keyboard input.
-// If onoff is 0, the cursor is hidden when waiting for keyboard input.
-// The function returns the old cursor setting.
-cursor: {
-    // }
-    rts
-}
 // Set the color for the background. The old color setting is returned.
 bgcolor: {
     // The background color register address

src/test/ref/examples/eightqueens/eightqueens-recursive.asm

@@ -4,7 +4,7 @@
 //
 // This is a recursive solution
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .const LIGHT_BLUE = $e
   .const OFFSET_STRUCT_PRINTF_BUFFER_NUMBER_DIGITS = 1
@@ -25,43 +25,44 @@
   .label DEFAULT_SCREEN = $400
   // The CIA#1: keyboard matrix, joystick #1/#2
   .label CIA1 = $dc00
-  .label conio_cursor_x = $14
-  .label conio_cursor_y = $15
-  .label conio_line_text = $16
-  .label conio_line_color = $18
-  .label count = 2
-__bbegin:
-  // conio_cursor_x = 0
   // The number of bytes on the screen
   // The current cursor x-position
-  lda #0
-  sta.z conio_cursor_x
-  // conio_cursor_y = 0
+  .label conio_cursor_x = $14
   // The current cursor y-position
-  sta.z conio_cursor_y
-  // conio_line_text = CONIO_SCREEN_TEXT
+  .label conio_cursor_y = $15
   // The current text cursor line start
-  lda #<DEFAULT_SCREEN
-  sta.z conio_line_text
-  lda #>DEFAULT_SCREEN
-  sta.z conio_line_text+1
-  // conio_line_color = CONIO_SCREEN_COLORS
+  .label conio_line_text = $16
   // The current color cursor line start
-  lda #<COLORRAM
-  sta.z conio_line_color
-  lda #>COLORRAM
-  sta.z conio_line_color+1
-  // count = 0
+  .label conio_line_color = $18
   // The number of found solutions
-  lda #<0
-  sta.z count
-  sta.z count+1
-  lda #<0>>$10
-  sta.z count+2
-  lda #>0>>$10
-  sta.z count+3
-  jsr main
-  rts
+  .label count = 2
+_start: {
+    // conio_cursor_x = 0
+    lda #0
+    sta.z conio_cursor_x
+    // conio_cursor_y = 0
+    sta.z conio_cursor_y
+    // conio_line_text = CONIO_SCREEN_TEXT
+    lda #<DEFAULT_SCREEN
+    sta.z conio_line_text
+    lda #>DEFAULT_SCREEN
+    sta.z conio_line_text+1
+    // conio_line_color = CONIO_SCREEN_COLORS
+    lda #<COLORRAM
+    sta.z conio_line_color
+    lda #>COLORRAM
+    sta.z conio_line_color+1
+    // count = 0
+    lda #<0
+    sta.z count
+    sta.z count+1
+    lda #<0>>$10
+    sta.z count+2
+    lda #>0>>$10
+    sta.z count+3
+    jsr main
+    rts
+}
 main: {
     // clrscr()
     jsr clrscr
@@ -411,6 +412,7 @@ printf_ulong: {
 printf_number_buffer: {
     .label __19 = $a
     .label buffer_sign = $1a
+    // There is a minimum length - work out the padding
     .label len = $11
     .label padding = $10
     .label format_min_length = $10

src/test/ref/examples/eightqueens/eightqueens.asm

@@ -4,7 +4,7 @@
 //
 // This is an iterative solution.
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .const LIGHT_BLUE = $e
   .const OFFSET_STRUCT_PRINTF_BUFFER_NUMBER_DIGITS = 1
@@ -13,35 +13,36 @@
   .label COLORRAM = $d800
   // Default address of screen character matrix
   .label DEFAULT_SCREEN = $400
+  // The number of bytes on the screen
+  // The current cursor x-position
   .label conio_cursor_x = $19
+  // The current cursor y-position
   .label conio_cursor_y = $1a
+  // The current text cursor line start
   .label conio_line_text = $1b
+  // The current color cursor line start
   .label conio_line_color = $1d
   // The number of found solutions
   .label count = 2
-__bbegin:
-  // conio_cursor_x = 0
-  // The number of bytes on the screen
-  // The current cursor x-position
-  lda #0
-  sta.z conio_cursor_x
-  // conio_cursor_y = 0
-  // The current cursor y-position
-  sta.z conio_cursor_y
-  // conio_line_text = CONIO_SCREEN_TEXT
-  // The current text cursor line start
-  lda #<DEFAULT_SCREEN
-  sta.z conio_line_text
-  lda #>DEFAULT_SCREEN
-  sta.z conio_line_text+1
-  // conio_line_color = CONIO_SCREEN_COLORS
-  // The current color cursor line start
-  lda #<COLORRAM
-  sta.z conio_line_color
-  lda #>COLORRAM
-  sta.z conio_line_color+1
-  jsr main
-  rts
+_start: {
+    // conio_cursor_x = 0
+    lda #0
+    sta.z conio_cursor_x
+    // conio_cursor_y = 0
+    sta.z conio_cursor_y
+    // conio_line_text = CONIO_SCREEN_TEXT
+    lda #<DEFAULT_SCREEN
+    sta.z conio_line_text
+    lda #>DEFAULT_SCREEN
+    sta.z conio_line_text+1
+    // conio_line_color = CONIO_SCREEN_COLORS
+    lda #<COLORRAM
+    sta.z conio_line_color
+    lda #>COLORRAM
+    sta.z conio_line_color+1
+    jsr main
+    rts
+}
 main: {
     // clrscr()
     jsr clrscr
@@ -710,6 +711,7 @@ ultoa_append: {
 // The solution uses the board itself as a "cursor" moving through all possibilities
 // When all columns on a row is exhausted move back down to the lower level and move forward one position until we are done with the last position on the first row
 queens: {
+    // The current row where the queen is moving
     // The current row where the queen is moving
     .label row = $d
     lda #<0

src/test/ref/examples/fastmultiply/fastmultiply8.asm

@@ -8,6 +8,7 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
+  // Pointers to a, b and c=a*b
   // Pointers to a, b and c=a*b
   .label ap = $fd
   .label bp = $fe

src/test/ref/examples/fire/fire.asm

@@ -98,7 +98,9 @@ main: {
 }
 // Animate the fire on the passed screen. Uses BUFFER to store the current values.
 fire: {
+    // Average characters from below the current character (24 lines)
     .label screen = 2
+    // Average characters from below the current character (24 lines)
     .label screen_1 = $b
     .label buffer = 4
     .label buffer_1 = 9

src/test/ref/examples/helloworld/helloworld.asm

@@ -1,39 +1,40 @@
 // Functions for performing input and output.
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .const LIGHT_BLUE = $e
   // Color Ram
   .label COLORRAM = $d800
   // Default address of screen character matrix
   .label DEFAULT_SCREEN = $400
-  .label conio_cursor_x = 6
-  .label conio_cursor_y = 7
-  .label conio_line_text = 8
-  .label conio_line_color = $a
-__bbegin:
-  // conio_cursor_x = 0
   // The number of bytes on the screen
   // The current cursor x-position
-  lda #0
-  sta.z conio_cursor_x
-  // conio_cursor_y = 0
+  .label conio_cursor_x = 6
   // The current cursor y-position
-  sta.z conio_cursor_y
-  // conio_line_text = CONIO_SCREEN_TEXT
+  .label conio_cursor_y = 7
   // The current text cursor line start
-  lda #<DEFAULT_SCREEN
-  sta.z conio_line_text
-  lda #>DEFAULT_SCREEN
-  sta.z conio_line_text+1
-  // conio_line_color = CONIO_SCREEN_COLORS
+  .label conio_line_text = 8
   // The current color cursor line start
-  lda #<COLORRAM
-  sta.z conio_line_color
-  lda #>COLORRAM
-  sta.z conio_line_color+1
-  jsr main
-  rts
+  .label conio_line_color = $a
+_start: {
+    // conio_cursor_x = 0
+    lda #0
+    sta.z conio_cursor_x
+    // conio_cursor_y = 0
+    sta.z conio_cursor_y
+    // conio_line_text = CONIO_SCREEN_TEXT
+    lda #<DEFAULT_SCREEN
+    sta.z conio_line_text
+    lda #>DEFAULT_SCREEN
+    sta.z conio_line_text+1
+    // conio_line_color = CONIO_SCREEN_COLORS
+    lda #<COLORRAM
+    sta.z conio_line_color
+    lda #>COLORRAM
+    sta.z conio_line_color+1
+    jsr main
+    rts
+}
 main: {
     // printf("hello world!\n")
     jsr cputs

src/test/ref/examples/irq/irq-hyperscreen.asm

@@ -25,39 +25,6 @@
   // The vector used when the KERNAL serves IRQ interrupts
   .label KERNEL_IRQ = $314
   .label GHOST_BYTE = $3fff
-main: {
-    // *GHOST_BYTE = 0
-    lda #0
-    sta GHOST_BYTE
-    // asm
-    sei
-    // CIA1->INTERRUPT = CIA_INTERRUPT_CLEAR
-    // Disable CIA 1 Timer IRQ
-    lda #CIA_INTERRUPT_CLEAR
-    sta CIA1+OFFSET_STRUCT_MOS6526_CIA_INTERRUPT
-    // VICII->CONTROL1 &=$7f
-    // Set raster line to $fa
-    lda #$7f
-    and VICII+OFFSET_STRUCT_MOS6569_VICII_CONTROL1
-    sta VICII+OFFSET_STRUCT_MOS6569_VICII_CONTROL1
-    // VICII->RASTER = $fa
-    lda #$fa
-    sta VICII+OFFSET_STRUCT_MOS6569_VICII_RASTER
-    // VICII->IRQ_ENABLE = IRQ_RASTER
-    // Enable Raster Interrupt
-    lda #IRQ_RASTER
-    sta VICII+OFFSET_STRUCT_MOS6569_VICII_IRQ_ENABLE
-    // *KERNEL_IRQ = &irq_bottom_1
-    // Set the IRQ routine
-    lda #<irq_bottom_1
-    sta KERNEL_IRQ
-    lda #>irq_bottom_1
-    sta KERNEL_IRQ+1
-    // asm
-    cli
-    // }
-    rts
-}
 // Interrupt Routine 2
 irq_bottom_2: {
     // VICII->BORDER_COLOR = WHITE
@@ -116,3 +83,36 @@ irq_bottom_1: {
     // }
     jmp $ea81
 }
+main: {
+    // *GHOST_BYTE = 0
+    lda #0
+    sta GHOST_BYTE
+    // asm
+    sei
+    // CIA1->INTERRUPT = CIA_INTERRUPT_CLEAR
+    // Disable CIA 1 Timer IRQ
+    lda #CIA_INTERRUPT_CLEAR
+    sta CIA1+OFFSET_STRUCT_MOS6526_CIA_INTERRUPT
+    // VICII->CONTROL1 &=$7f
+    // Set raster line to $fa
+    lda #$7f
+    and VICII+OFFSET_STRUCT_MOS6569_VICII_CONTROL1
+    sta VICII+OFFSET_STRUCT_MOS6569_VICII_CONTROL1
+    // VICII->RASTER = $fa
+    lda #$fa
+    sta VICII+OFFSET_STRUCT_MOS6569_VICII_RASTER
+    // VICII->IRQ_ENABLE = IRQ_RASTER
+    // Enable Raster Interrupt
+    lda #IRQ_RASTER
+    sta VICII+OFFSET_STRUCT_MOS6569_VICII_IRQ_ENABLE
+    // *KERNEL_IRQ = &irq_bottom_1
+    // Set the IRQ routine
+    lda #<irq_bottom_1
+    sta KERNEL_IRQ
+    lda #>irq_bottom_1
+    sta KERNEL_IRQ+1
+    // asm
+    cli
+    // }
+    rts
+}

src/test/ref/examples/multiplexer/simple-multiplexer.asm

@@ -3,7 +3,7 @@
 // The MOS 6526 Complex Interface Adapter (CIA)
 // http://archive.6502.org/datasheets/mos_6526_cia_recreated.pdf
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .const VIC_RST8 = $80
   .const VIC_DEN = $10
@@ -24,33 +24,33 @@
   // The VIC-II MOS 6567/6569
   .label VICII = $d000
   // Location of screen & sprites
+  // Location of screen & sprites
   .label SCREEN = $400
-  .label SPRITE = $2000
   // The address of the sprite pointers on the current screen (screen+0x3f8).
   .label PLEX_SCREEN_PTR = SCREEN+$3f8
-  .label plex_show_idx = 6
-  .label plex_sprite_idx = 7
-  .label plex_sprite_msb = 8
-  .label plex_free_next = 9
-__bbegin:
-  // plex_show_idx=0
   // The index in the PLEX tables of the next sprite to show
-  lda #0
-  sta.z plex_show_idx
-  // plex_sprite_idx=0
+  .label plex_show_idx = 6
   // The index the next sprite to use for showing (sprites are used round-robin)
-  sta.z plex_sprite_idx
-  // plex_sprite_msb=1
+  .label plex_sprite_idx = 7
   // The MSB bit of the next sprite to use for showing
-  lda #1
-  sta.z plex_sprite_msb
-  // plex_free_next = 0
+  .label plex_sprite_msb = 8
   // The index of the sprite that is free next. Since sprites are used round-robin this moves forward each time a sprite is shown.
-  lda #0
-  sta.z plex_free_next
-  // kickasm
-  jsr main
-  rts
+  .label plex_free_next = 9
+_start: {
+    // plex_show_idx=0
+    lda #0
+    sta.z plex_show_idx
+    // plex_sprite_idx=0
+    sta.z plex_sprite_idx
+    // plex_sprite_msb=1
+    lda #1
+    sta.z plex_sprite_msb
+    // plex_free_next = 0
+    lda #0
+    sta.z plex_free_next
+    jsr main
+    rts
+}
 main: {
     // asm
     sei
@@ -63,6 +63,7 @@ main: {
 }
 // The raster loop
 loop: {
+    // The current index into the y-sinus
     // The current index into the y-sinus
     .label sin_idx = 2
     .label plexFreeNextYpos1_return = $a
@@ -300,6 +301,7 @@ plexSort: {
 }
 // Initialize the program
 init: {
+    // Set the x-positions & pointers
     // Set the x-positions & pointers
     .label xp = 4
     // *D011 = VIC_DEN | VIC_RSEL | 3
@@ -315,7 +317,7 @@ init: {
     ldx #0
   __b1:
     // PLEX_PTR[sx] = (char)(SPRITE/$40)
-    lda #SPRITE/$40
+    lda #$ff&SPRITE/$40
     sta PLEX_PTR,x
     // PLEX_XPOS[sx] = xp
     txa
@@ -385,8 +387,9 @@ YSIN:
     .for(var i=0;i<256;i++)
         .byte round(min+(ampl/2)+(ampl/2)*sin(toRadians(360*i/256)))
 
-.pc = SPRITE "SPRITE"
-  .var pic = LoadPicture("balloon.png", List().add($000000, $ffffff))
+.pc = $2000 "SPRITE"
+SPRITE:
+.var pic = LoadPicture("balloon.png", List().add($000000, $ffffff))
     .for (var y=0; y<21; y++)
         .for (var x=0;x<3; x++)
             .byte pic.getSinglecolorByte(x,y)

src/test/ref/examples/music/music.asm

@@ -9,17 +9,17 @@
   .const OFFSET_STRUCT_MOS6569_VICII_BORDER_COLOR = $20
   // The VIC-II MOS 6567/6569
   .label VICII = $d000
-  .label MUSIC = $1000
-  // kickasm
-  // Load the SID
-  .const music = LoadSid("toiletrensdyr.sid")
-
-// Place the SID into memory
+  // Pointer to the music init routine
+  // Pointer to the music init routine
+  .label musicInit = MUSIC
+  // Pointer to the music play routine
+  // Pointer to the music play routine
+  .label musicPlay = MUSIC+3
 // Play the music 
 main: {
-    // asm
+    // (*musicInit)()
     // Initialize the music
-    jsr music.init
+    jsr musicInit
   // Wait for the RASTER
   __b1:
     // while (VICII->RASTER != $fd)
@@ -28,13 +28,16 @@ main: {
     bne __b1
     // (VICII->BORDER_COLOR)++;
     inc VICII+OFFSET_STRUCT_MOS6569_VICII_BORDER_COLOR
-    // asm
+    // (*musicPlay)()
     // Play the music
-    jsr music.play
+    jsr musicPlay
     // (VICII->BORDER_COLOR)--;
     dec VICII+OFFSET_STRUCT_MOS6569_VICII_BORDER_COLOR
     jmp __b1
 }
-.pc = MUSIC "MUSIC"
-  .fill music.size, music.getData(i)
+.pc = $1000 "MUSIC"
+// SID tune at an absolute address
+MUSIC:
+.const music = LoadSid("toiletrensdyr.sid")
+    .fill music.size, music.getData(i)
 

src/test/ref/examples/music/music_irq.asm

@@ -21,17 +21,34 @@
   .label CIA1 = $dc00
   // The vector used when the KERNAL serves IRQ interrupts
   .label KERNEL_IRQ = $314
-  .label MUSIC = $1000
-  // kickasm
-  // Load the SID
-  .const music = LoadSid("toiletrensdyr.sid")
-
-// Place the SID into memory
+  // Pointer to the music init routine
+  // Pointer to the music init routine
+  .label musicInit = MUSIC
+  // Pointer to the music play routine
+  // Pointer to the music play routine
+  .label musicPlay = MUSIC+3
+// Raster IRQ Routine playing music
+irq_play: {
+    // (VICII->BORDER_COLOR)++;
+    inc VICII+OFFSET_STRUCT_MOS6569_VICII_BORDER_COLOR
+    // (*musicPlay)()
+    // Play SID
+    jsr musicPlay
+    // VICII->IRQ_STATUS = IRQ_RASTER
+    // Acknowledge the IRQ
+    lda #IRQ_RASTER
+    sta VICII+OFFSET_STRUCT_MOS6569_VICII_IRQ_STATUS
+    // (VICII->BORDER_COLOR)--;
+    dec VICII+OFFSET_STRUCT_MOS6569_VICII_BORDER_COLOR
+    // }
+    jmp $ea31
+}
 // Setup Raster IRQ and initialize SID player
 main: {
     // asm
     sei
-    jsr music.init
+    // (*musicInit)()
+    jsr musicInit
     // CIA1->INTERRUPT = CIA_INTERRUPT_CLEAR
     // Disable CIA 1 Timer IRQ
     lda #CIA_INTERRUPT_CLEAR
@@ -59,22 +76,9 @@ main: {
     // }
     rts
 }
-// Raster IRQ Routine playing music
-irq_play: {
-    // (VICII->BORDER_COLOR)++;
-    inc VICII+OFFSET_STRUCT_MOS6569_VICII_BORDER_COLOR
-    // asm
-    // Play SID
-    jsr music.play
-    // VICII->IRQ_STATUS = IRQ_RASTER
-    // Acknowledge the IRQ
-    lda #IRQ_RASTER
-    sta VICII+OFFSET_STRUCT_MOS6569_VICII_IRQ_STATUS
-    // (VICII->BORDER_COLOR)--;
-    dec VICII+OFFSET_STRUCT_MOS6569_VICII_BORDER_COLOR
-    // }
-    jmp $ea31
-}
-.pc = MUSIC "MUSIC"
-  .fill music.size, music.getData(i)
+.pc = $1000 "MUSIC"
+// SID tune at an absolute address
+MUSIC:
+.const music = LoadSid("toiletrensdyr.sid")
+    .fill music.size, music.getData(i)
 

src/test/ref/examples/nes-conio/nes-conio.asm

@@ -81,36 +81,39 @@
   // Pointer to the start of RAM memory
   .label MEMORY = 0
   // NES Picture Processing Unit (PPU)
+  // NES Picture Processing Unit (PPU)
   .label PPU = $2000
   // NES CPU and audion processing unit (APU)
+  // NES CPU and audion processing unit (APU)
   .label APU = $4000
+  // The current cursor x-position
   .label conio_cursor_x = $11
+  // The current cursor y-position
   .label conio_cursor_y = $12
+  // The current text cursor line start
   .label conio_line_text = $13
   .label x_scroll = $15
   .label y_scroll = $16
-__bbegin:
-  // conio_cursor_x = 0
-  // The current cursor x-position
-  lda #0
-  sta.z conio_cursor_x
-  // conio_cursor_y = 0
-  // The current cursor y-position
-  sta.z conio_cursor_y
-  // conio_line_text = CONIO_SCREEN_TEXT
-  // The current text cursor line start
-  lda #<PPU_NAME_TABLE_0
-  sta.z conio_line_text
-  lda #>PPU_NAME_TABLE_0
-  sta.z conio_line_text+1
-  // x_scroll
-  lda #0
-  sta.z x_scroll
-  // y_scroll
-  sta.z y_scroll
-  jsr main
-  rts
 .segment Code
+_start: {
+    // conio_cursor_x = 0
+    lda #0
+    sta.z conio_cursor_x
+    // conio_cursor_y = 0
+    sta.z conio_cursor_y
+    // conio_line_text = CONIO_SCREEN_TEXT
+    lda #<PPU_NAME_TABLE_0
+    sta.z conio_line_text
+    lda #>PPU_NAME_TABLE_0
+    sta.z conio_line_text+1
+    // x_scroll
+    lda #0
+    sta.z x_scroll
+    // y_scroll
+    sta.z y_scroll
+    jsr main
+    rts
+}
 // RESET Called when the NES is reset, including when it is turned on.
 main: {
     .const screensizex1_return = $20
@@ -363,6 +366,7 @@ cputln: {
 }
 // Scroll the entire screen if the cursor is beyond the last line
 cscroll: {
+    // Scroll lines up
     // Scroll lines up
     .label line1 = $17
     .label line2 = $f
@@ -456,6 +460,7 @@ cscroll: {
 // ppuDataTransfer(void* zp(5) ppuData, void* zp(7) cpuData, word zp(9) size)
 ppuDataTransfer: {
     .label ppuDataPrepare1_ppuData = 5
+    // Transfer to PPU
     .label cpuSrc = 7
     .label i = $b
     .label ppuData = 5
@@ -511,6 +516,7 @@ ppuDataFetch: {
     .const size = $20
     .label cpuData = conio_cscroll_buffer
     // Fetch from PPU to CPU
+    // Fetch from PPU to CPU
     .label cpuDst = 7
     .label i = 5
     .label ppuData = $19
@@ -626,9 +632,9 @@ ppuDataSet: {
 // Set the cursor to the specified position
 // gotoxy(byte register(X) x, byte register(A) y)
 gotoxy: {
-    .label __5 = $13
-    .label __6 = $13
-    .label line_offset = $13
+    .label __5 = $1b
+    .label __6 = $1b
+    .label line_offset = $1b
     // if(y>CONIO_HEIGHT)
     cmp #$1e+1
     bcc __b1
@@ -667,6 +673,10 @@ gotoxy: {
     adc #>PPU_NAME_TABLE_0
     sta.z __5+1
     // conio_line_text = CONIO_SCREEN_TEXT + line_offset
+    lda.z __5
+    sta.z conio_line_text
+    lda.z __5+1
+    sta.z conio_line_text+1
     // }
     rts
 }
@@ -682,9 +692,9 @@ cputsxy: {
     rts
 }
 // Output a NUL-terminated string at the current cursor position
-// cputs(byte* zp($d) s)
+// cputs(byte* zp($1b) s)
 cputs: {
-    .label s = $d
+    .label s = $1b
     lda #<num_buffer
     sta.z s
     lda #>num_buffer
@@ -716,10 +726,10 @@ cputs: {
 // uctoa(byte register(X) value, byte* zp($f) buffer)
 uctoa: {
     .const max_digits = 2
-    .label digit_value = $1b
+    .label digit_value = $1d
     .label buffer = $f
-    .label digit = $11
-    .label started = $12
+    .label digit = $d
+    .label started = $e
     lda #<num_buffer
     sta.z buffer
     lda #>num_buffer
@@ -784,10 +794,10 @@ uctoa: {
 // - sub : the value of a '1' in the digit. Subtracted continually while the digit is increased.
 //        (For decimal the subs used are 10000, 1000, 100, 10, 1)
 // returns : the value reduced by sub * digit so that it is less than sub.
-// uctoa_append(byte* zp($f) buffer, byte register(X) value, byte zp($1b) sub)
+// uctoa_append(byte* zp($f) buffer, byte register(X) value, byte zp($1d) sub)
 uctoa_append: {
     .label buffer = $f
-    .label sub = $1b
+    .label sub = $1d
     ldy #0
   __b1:
     // while (value >= sub)
@@ -919,7 +929,7 @@ vblank: {
 // - bit 6: B
 // - bit 7: A
 readJoy1: {
-    .label __1 = $1c
+    .label __1 = $1e
     // APU->JOY1 = 1
     // Latch the controller buttons
     lda #1

src/test/ref/examples/nes-demo/nes-demo.asm

@@ -82,10 +82,143 @@
   // Pointer to the start of RAM memory
   .label MEMORY = 0
   // NES Picture Processing Unit (PPU)
+  // NES Picture Processing Unit (PPU)
   .label PPU = $2000
   // NES CPU and audion processing unit (APU)
+  // NES CPU and audion processing unit (APU)
   .label APU = $4000
 .segment Code
+// NMI Called when the PPU refreshes the screen (also known as the V-Blank period)
+vblank: {
+    pha
+    txa
+    pha
+    tya
+    pha
+    // readJoy1()
+    jsr readJoy1
+    // joy = readJoy1()
+    tax
+    // joy&JOY_DOWN
+    txa
+    and #JOY_DOWN
+    // if(joy&JOY_DOWN)
+    cmp #0
+    beq __b1
+    // SPRITE_BUFFER[0].y++;
+    inc SPRITE_BUFFER
+    // SPRITE_BUFFER[1].y++;
+    inc SPRITE_BUFFER+1*SIZEOF_STRUCT_SPRITEDATA
+    // SPRITE_BUFFER[2].y++;
+    inc SPRITE_BUFFER+2*SIZEOF_STRUCT_SPRITEDATA
+    // SPRITE_BUFFER[3].y++;
+    inc SPRITE_BUFFER+3*SIZEOF_STRUCT_SPRITEDATA
+  __b1:
+    // joy&JOY_UP
+    txa
+    and #JOY_UP
+    // if(joy&JOY_UP)
+    cmp #0
+    beq __b2
+    // SPRITE_BUFFER[0].y--;
+    dec SPRITE_BUFFER
+    // SPRITE_BUFFER[1].y--;
+    dec SPRITE_BUFFER+1*SIZEOF_STRUCT_SPRITEDATA
+    // SPRITE_BUFFER[2].y--;
+    dec SPRITE_BUFFER+2*SIZEOF_STRUCT_SPRITEDATA
+    // SPRITE_BUFFER[3].y--;
+    dec SPRITE_BUFFER+3*SIZEOF_STRUCT_SPRITEDATA
+  __b2:
+    // joy&JOY_LEFT
+    txa
+    and #JOY_LEFT
+    // if(joy&JOY_LEFT)
+    cmp #0
+    beq __b3
+    // SPRITE_BUFFER[0].x--;
+    dec SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X
+    // SPRITE_BUFFER[1].x--;
+    dec SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+1*SIZEOF_STRUCT_SPRITEDATA
+    // SPRITE_BUFFER[2].x--;
+    dec SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+2*SIZEOF_STRUCT_SPRITEDATA
+    // SPRITE_BUFFER[3].x--;
+    dec SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+3*SIZEOF_STRUCT_SPRITEDATA
+  __b3:
+    // joy&JOY_RIGHT
+    txa
+    and #JOY_RIGHT
+    // if(joy&JOY_RIGHT)
+    cmp #0
+    beq ppuSpriteBufferDmaTransfer1
+    // SPRITE_BUFFER[0].x++;
+    inc SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X
+    // SPRITE_BUFFER[1].x++;
+    inc SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+1*SIZEOF_STRUCT_SPRITEDATA
+    // SPRITE_BUFFER[2].x++;
+    inc SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+2*SIZEOF_STRUCT_SPRITEDATA
+    // SPRITE_BUFFER[3].x++;
+    inc SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+3*SIZEOF_STRUCT_SPRITEDATA
+  ppuSpriteBufferDmaTransfer1:
+    // PPU->OAMADDR = 0
+    lda #0
+    sta PPU+OFFSET_STRUCT_RICOH_2C02_OAMADDR
+    // APU->OAMDMA = >spriteBuffer
+    lda #>SPRITE_BUFFER
+    sta APU+OFFSET_STRUCT_RICOH_2A03_OAMDMA
+    // PPU->PPUSCROLL = 0
+    // Set scroll
+    lda #0
+    sta PPU+OFFSET_STRUCT_RICOH_2C02_PPUSCROLL
+    // PPU->PPUSCROLL = -8
+    lda #-8
+    sta PPU+OFFSET_STRUCT_RICOH_2C02_PPUSCROLL
+    // }
+    pla
+    tay
+    pla
+    tax
+    pla
+    rti
+}
+// Read Standard Controller #1
+// Returns a byte representing the pushed buttons
+// - bit 0: right
+// - bit 1: left
+// - bit 2: down
+// - bit 3: up
+// - bit 4: start
+// - bit 5: select
+// - bit 6: B
+// - bit 7: A
+readJoy1: {
+    .label __1 = 8
+    // APU->JOY1 = 1
+    // Latch the controller buttons
+    lda #1
+    sta APU+OFFSET_STRUCT_RICOH_2A03_JOY1
+    // APU->JOY1 = 0
+    lda #0
+    sta APU+OFFSET_STRUCT_RICOH_2A03_JOY1
+    tax
+  __b1:
+    // for(char i=0;i<8;i++)
+    cpx #8
+    bcc __b2
+    // }
+    rts
+  __b2:
+    // joy<<1
+    asl
+    sta.z __1
+    // APU->JOY1&1
+    lda #1
+    and APU+OFFSET_STRUCT_RICOH_2A03_JOY1
+    // joy = joy<<1 | APU->JOY1&1
+    ora.z __1
+    // for(char i=0;i<8;i++)
+    inx
+    jmp __b1
+}
 // RESET Called when the NES is reset, including when it is turned on.
 main: {
     // asm
@@ -368,6 +501,7 @@ ppuDataTransfer: {
     .label ppuData = PPU_PALETTE
     .label cpuData = PALETTE
     // Transfer to PPU
+    // Transfer to PPU
     .label cpuSrc = 6
     .label i = 4
     // PPU->PPUADDR = >ppuData
@@ -413,137 +547,6 @@ ppuDataTransfer: {
   !:
     jmp __b1
 }
-// NMI Called when the PPU refreshes the screen (also known as the V-Blank period)
-vblank: {
-    pha
-    txa
-    pha
-    tya
-    pha
-    // readJoy1()
-    jsr readJoy1
-    // joy = readJoy1()
-    tax
-    // joy&JOY_DOWN
-    txa
-    and #JOY_DOWN
-    // if(joy&JOY_DOWN)
-    cmp #0
-    beq __b1
-    // SPRITE_BUFFER[0].y++;
-    inc SPRITE_BUFFER
-    // SPRITE_BUFFER[1].y++;
-    inc SPRITE_BUFFER+1*SIZEOF_STRUCT_SPRITEDATA
-    // SPRITE_BUFFER[2].y++;
-    inc SPRITE_BUFFER+2*SIZEOF_STRUCT_SPRITEDATA
-    // SPRITE_BUFFER[3].y++;
-    inc SPRITE_BUFFER+3*SIZEOF_STRUCT_SPRITEDATA
-  __b1:
-    // joy&JOY_UP
-    txa
-    and #JOY_UP
-    // if(joy&JOY_UP)
-    cmp #0
-    beq __b2
-    // SPRITE_BUFFER[0].y--;
-    dec SPRITE_BUFFER
-    // SPRITE_BUFFER[1].y--;
-    dec SPRITE_BUFFER+1*SIZEOF_STRUCT_SPRITEDATA
-    // SPRITE_BUFFER[2].y--;
-    dec SPRITE_BUFFER+2*SIZEOF_STRUCT_SPRITEDATA
-    // SPRITE_BUFFER[3].y--;
-    dec SPRITE_BUFFER+3*SIZEOF_STRUCT_SPRITEDATA
-  __b2:
-    // joy&JOY_LEFT
-    txa
-    and #JOY_LEFT
-    // if(joy&JOY_LEFT)
-    cmp #0
-    beq __b3
-    // SPRITE_BUFFER[0].x--;
-    dec SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X
-    // SPRITE_BUFFER[1].x--;
-    dec SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+1*SIZEOF_STRUCT_SPRITEDATA
-    // SPRITE_BUFFER[2].x--;
-    dec SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+2*SIZEOF_STRUCT_SPRITEDATA
-    // SPRITE_BUFFER[3].x--;
-    dec SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+3*SIZEOF_STRUCT_SPRITEDATA
-  __b3:
-    // joy&JOY_RIGHT
-    txa
-    and #JOY_RIGHT
-    // if(joy&JOY_RIGHT)
-    cmp #0
-    beq ppuSpriteBufferDmaTransfer1
-    // SPRITE_BUFFER[0].x++;
-    inc SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X
-    // SPRITE_BUFFER[1].x++;
-    inc SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+1*SIZEOF_STRUCT_SPRITEDATA
-    // SPRITE_BUFFER[2].x++;
-    inc SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+2*SIZEOF_STRUCT_SPRITEDATA
-    // SPRITE_BUFFER[3].x++;
-    inc SPRITE_BUFFER+OFFSET_STRUCT_SPRITEDATA_X+3*SIZEOF_STRUCT_SPRITEDATA
-  ppuSpriteBufferDmaTransfer1:
-    // PPU->OAMADDR = 0
-    lda #0
-    sta PPU+OFFSET_STRUCT_RICOH_2C02_OAMADDR
-    // APU->OAMDMA = >spriteBuffer
-    lda #>SPRITE_BUFFER
-    sta APU+OFFSET_STRUCT_RICOH_2A03_OAMDMA
-    // PPU->PPUSCROLL = 0
-    // Set scroll
-    lda #0
-    sta PPU+OFFSET_STRUCT_RICOH_2C02_PPUSCROLL
-    // PPU->PPUSCROLL = -8
-    lda #-8
-    sta PPU+OFFSET_STRUCT_RICOH_2C02_PPUSCROLL
-    // }
-    pla
-    tay
-    pla
-    tax
-    pla
-    rti
-}
-// Read Standard Controller #1
-// Returns a byte representing the pushed buttons
-// - bit 0: right
-// - bit 1: left
-// - bit 2: down
-// - bit 3: up
-// - bit 4: start
-// - bit 5: select
-// - bit 6: B
-// - bit 7: A
-readJoy1: {
-    .label __1 = 8
-    // APU->JOY1 = 1
-    // Latch the controller buttons
-    lda #1
-    sta APU+OFFSET_STRUCT_RICOH_2A03_JOY1
-    // APU->JOY1 = 0
-    lda #0
-    sta APU+OFFSET_STRUCT_RICOH_2A03_JOY1
-    tax
-  __b1:
-    // for(char i=0;i<8;i++)
-    cpx #8
-    bcc __b2
-    // }
-    rts
-  __b2:
-    // joy<<1
-    asl
-    sta.z __1
-    // APU->JOY1&1
-    lda #1
-    and APU+OFFSET_STRUCT_RICOH_2A03_JOY1
-    // joy = joy<<1 | APU->JOY1&1
-    ora.z __1
-    // for(char i=0;i<8;i++)
-    inx
-    jmp __b1
-}
 .segment Data
   // Flag tile
   FLAG: .byte $54, $55, $56, $57

src/test/ref/examples/nes-dxycp/nes-dxycp.asm

@@ -74,28 +74,31 @@
   // Pointer to the start of RAM memory
   .label MEMORY = 0
   // NES Picture Processing Unit (PPU)
+  // NES Picture Processing Unit (PPU)
   .label PPU = $2000
   // NES CPU and audion processing unit (APU)
+  // NES CPU and audion processing unit (APU)
   .label APU = $4000
-  .label y_sin_idx = $c
-  .label x_sin_idx = $d
-  .label x_sin_idx_2 = $e
-__bbegin:
-  // y_sin_idx = 0
   // Index into the Y sine
-  lda #0
-  sta.z y_sin_idx
-  // x_sin_idx = 73
+  .label y_sin_idx = $c
   // Index into the X sine
-  lda #$49
-  sta.z x_sin_idx
-  // x_sin_idx_2 = 82
+  .label x_sin_idx = $d
   // Index into the small X sine
-  lda #$52
-  sta.z x_sin_idx_2
-  jsr main
-  rts
+  .label x_sin_idx_2 = $e
 .segment Code
+_start: {
+    // y_sin_idx = 0
+    lda #0
+    sta.z y_sin_idx
+    // x_sin_idx = 73
+    lda #$49
+    sta.z x_sin_idx
+    // x_sin_idx_2 = 82
+    lda #$52
+    sta.z x_sin_idx_2
+    jsr main
+    rts
+}
 // RESET Called when the NES is reset, including when it is turned on.
 main: {
     // asm
@@ -263,6 +266,7 @@ ppuDataTransfer: {
     .label ppuData = PPU_PALETTE
     .label cpuData = PALETTE
     // Transfer to PPU
+    // Transfer to PPU
     .label cpuSrc = 6
     .label i = 4
     // PPU->PPUADDR = >ppuData
@@ -315,6 +319,7 @@ vblank: {
     .label __23 = $b
     .label __25 = $b
     .label __28 = $f
+    // Update sprite positions
     .label y_idx = 9
     .label x_idx = $a
     .label x_idx_2 = $b

src/test/ref/examples/nmisamples/nmisamples.asm

@@ -2,7 +2,7 @@
 // Code by Scan of Desire (Richard-William Loerakker)
 // Sample from ART OF NOISE: MOMENTS IN LOVE
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   // Value that disables all CIA interrupts when stored to the CIA Interrupt registers
   .const CIA_INTERRUPT_CLEAR = $7f
@@ -23,14 +23,15 @@
   // The vector used when the KERNAL serves NMI interrupts
   .label KERNEL_NMI = $318
   .label sample = 2
-__bbegin:
-  // sample = SAMPLE
-  lda #<SAMPLE
-  sta.z sample
-  lda #>SAMPLE
-  sta.z sample+1
-  jsr main
-  rts
+_start: {
+    // sample = SAMPLE
+    lda #<SAMPLE
+    sta.z sample
+    lda #>SAMPLE
+    sta.z sample+1
+    jsr main
+    rts
+}
 main: {
     // asm
     // Boosting 8580 Digis

src/test/ref/examples/plasma/plasma-unroll.asm

@@ -27,13 +27,13 @@
   .label COLS = $d800
   .label SCREEN1 = $2800
   .label CHARSET = $2000
-  .label print_line_cursor = $400
+  .label print_screen = $400
   .label print_char_cursor = $b
   // Plasma state variables
-  .label c1A = 2
-  .label c1B = 3
-  .label c2A = 4
-  .label c2B = 5
+  .label c1A = $d
+  .label c1B = $f
+  .label c2A = $12
+  .label c2B = 2
 main: {
     .const toD0181_return = (>(SCREEN1&$3fff)*4)|(>CHARSET)/4&$f
     .label col = $b
@@ -82,10 +82,10 @@ main: {
 }
 // Render plasma to the passed screen
 doplasma: {
-    .label c1a = $f
-    .label c1b = $12
+    .label c1a = 4
+    .label c1b = 5
     .label yval = $e
-    .label i = $d
+    .label i = 3
     .label c2a = 7
     .label c2b = 8
     .label i1 = 6
@@ -348,9 +348,9 @@ makecharset: {
     sta SID+OFFSET_STRUCT_MOS6581_SID_CH3_CONTROL
     // print_cls()
     jsr print_cls
-    lda #<print_line_cursor
+    lda #<print_screen
     sta.z print_char_cursor
-    lda #>print_line_cursor
+    lda #>print_screen
     sta.z print_char_cursor+1
     lda #<0
     sta.z c
@@ -481,7 +481,7 @@ print_cls: {
 memset: {
     .const c = ' '
     .const num = $3e8
-    .label str = print_line_cursor
+    .label str = print_screen
     .label end = str+num
     .label dst = $10
     lda #<str

src/test/ref/examples/plasma/plasma.asm

@@ -26,13 +26,13 @@
   .label SCREEN1 = $2800
   .label SCREEN2 = $2c00
   .label CHARSET = $2000
-  .label print_line_cursor = $400
+  .label print_screen = $400
   .label print_char_cursor = $b
   // Plasma state variables
-  .label c1A = 2
-  .label c1B = 3
-  .label c2A = 4
-  .label c2B = 5
+  .label c1A = $d
+  .label c1B = $e
+  .label c2A = $11
+  .label c2B = 2
 main: {
     .const toD0181_return = (>(SCREEN1&$3fff)*4)|(>CHARSET)/4&$f
     .const toD0182_return = (>(SCREEN2&$3fff)*4)|(>CHARSET)/4&$f
@@ -96,9 +96,9 @@ main: {
 // Render plasma to the passed screen
 // doplasma(byte* zp(9) screen)
 doplasma: {
-    .label c1a = $e
-    .label c1b = $11
-    .label i = $d
+    .label c1a = 4
+    .label c1b = 5
+    .label i = 3
     .label c2a = 7
     .label c2b = 8
     .label i1 = 6
@@ -244,9 +244,9 @@ makecharset: {
     sta SID+OFFSET_STRUCT_MOS6581_SID_CH3_CONTROL
     // print_cls()
     jsr print_cls
-    lda #<print_line_cursor
+    lda #<print_screen
     sta.z print_char_cursor
-    lda #>print_line_cursor
+    lda #>print_screen
     sta.z print_char_cursor+1
     lda #<0
     sta.z c
@@ -377,7 +377,7 @@ print_cls: {
 memset: {
     .const c = ' '
     .const num = $3e8
-    .label str = print_line_cursor
+    .label str = print_screen
     .label end = str+num
     .label dst = $f
     lda #<str

src/test/ref/examples/plus4walk/plus4walk.asm

@@ -262,11 +262,11 @@ rand: {
     rts
 }
 // Copies the character c (an unsigned char) to the first num characters of the object pointed to by the argument str.
-// memset(void* zp(5) str, byte register(X) c)
+// memset(void* zp(3) str, byte register(X) c)
 memset: {
     .label end = $13
-    .label dst = 5
-    .label str = 5
+    .label dst = 3
+    .label str = 3
     // end = (char*)str + num
     lda.z str
     clc

src/test/ref/examples/rotate/rotate.asm

@@ -48,14 +48,19 @@ anim: {
     .label __12 = 3
     .label __26 = $13
     .label x = $b
+    // signed fixed[7.0]
     .label y = $c
     .label xr = $d
+    // signed fixed[8.8]
     .label yr = $f
+    // signed fixed[8.8]
     .label xpos = $11
     // signed fixed[0.7]
+    // signed fixed[0.7]
     .label sprite_msb = 2
     .label i = $a
     .label angle = 7
+    // Calculate the cycle count - 0x12 is the base usage of start/read
     .label cyclecount = $13
     lda #0
     sta.z angle
@@ -470,17 +475,22 @@ init: {
 }
 // Initialize the mulf_sqr multiplication tables with f(x)=int(x*x/4)
 mulf_init: {
+    // x/2
     // x/2
     .label c = 7
     // Counter used for determining x%2==0
+    // Counter used for determining x%2==0
     .label sqr1_hi = 8
     // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
+    // Fill mulf_sqr1 = f(x) = int(x*x/4): If f(x) = x*x/4 then f(x+1) = f(x) + x/2 + 1/4
     .label sqr = $11
     .label sqr1_lo = 5
     // Decrease or increase x_255 - initially we decrease
+    // Decrease or increase x_255 - initially we decrease
     .label sqr2_hi = $f
     .label sqr2_lo = $d
     //Start with g(0)=f(255)
+    //Start with g(0)=f(255)
     .label dir = $a
     ldx #0
     lda #<mulf_sqr1_hi+1

src/test/ref/examples/scrollbig/scrollbig.asm

@@ -17,7 +17,7 @@
   .label SCREEN = $400
   .label current_bit = 2
   // Scroll the next bit from the current char onto the screen - trigger next char if needed
-  .label current_chargen = 3
+  .label current_chargen = 7
   .label nxt = 5
 main: {
     // fillscreen(SCREEN, $20)
@@ -67,9 +67,9 @@ scroll_soft: {
     rts
 }
 scroll_bit: {
-    .label __7 = 3
-    .label c = 3
-    .label sc = 7
+    .label __7 = 7
+    .label c = 7
+    .label sc = 3
     // current_bit = current_bit/2
     lsr.z current_bit
     // if(current_bit==0)

src/test/ref/examples/scrolllogo/scrolllogo.asm

@@ -29,6 +29,7 @@
   // Color Ram
   .label COLS = $d800
   .label SCREEN = $400
+  // Remainder after unsigned 16-bit division
   .label rem16u = $1d
   .label xsin_idx = $27
 main: {
@@ -322,10 +323,14 @@ sin16s_gen2: {
     .label ampl = max-min
     .label __6 = 8
     .label __8 = $1d
+    // ampl is always positive so shifting left does not alter the sign
+    // u[4.28] step = PI*2/wavelength
     .label step = $19
     .label sintab = $14
     // u[4.28]
     // Iterate over the table
+    // u[4.28]
+    // Iterate over the table
     .label x = 4
     .label i = $27
     // div32u16u(PI2_u4f28, wavelength)
@@ -529,14 +534,23 @@ sin16s: {
     .label __4 = $1f
     .label x = $c
     .label return = $12
+    // sinx = x - x^3/6 + x5/128;
     .label x1 = $23
+    // u[1.15]
     .label x2 = $16
+    // u[2.14] x^2
     .label x3 = $16
+    // u[2.14] x^3
     .label x3_6 = $25
+    // u[1.15] x^3/6;
     .label usinx = $12
+    // u[1.15] x - x^3/6
     .label x4 = $16
+    // u[3.13] x^4
     .label x5 = $25
+    // u[4.12] x^5
     .label x5_128 = $25
+    // u[1.15] (first bit is always zero)
     .label sinx = $12
     // if(x >= PI_u4f28 )
     lda.z x+3

src/test/ref/examples/sinplotter/sine-plotter.asm

@@ -35,7 +35,7 @@
   .label SCREEN = $400
   .label BITMAP = $2000
   // Remainder after unsigned 16-bit division
-  .label rem16u = $12
+  .label rem16u = $16
 main: {
     .const vicSelectGfxBank1_toDd001_return = 3
     .const toD0181_return = (>(SCREEN&$3fff)*4)|(>BITMAP)/4&$f
@@ -282,10 +282,14 @@ sin16s_gen2: {
     .label ampl = max-min
     .label __6 = 8
     .label __8 = $1c
+    // ampl is always positive so shifting left does not alter the sign
+    // u[4.28] step = PI*2/wavelength
     .label step = $18
     .label sintab = 6
     // u[4.28]
     // Iterate over the table
+    // u[4.28]
+    // Iterate over the table
     .label x = 2
     .label i = $14
     // div32u16u(PI2_u4f28, wavelength)
@@ -375,13 +379,13 @@ sin16s_gen2: {
 }
 // Multiply of two signed ints to a signed long
 // Fixes offsets introduced by using unsigned multiplication
-// mul16s(signed word zp($16) a)
+// mul16s(signed word zp($1c) a)
 mul16s: {
     .label __6 = $22
     .label __11 = $22
     .label m = 8
     .label return = 8
-    .label a = $16
+    .label a = $1c
     // mul16u((unsigned int)a, (unsigned int) b)
     lda.z a
     sta.z mul16u.a
@@ -420,10 +424,10 @@ mul16s: {
     rts
 }
 // Perform binary multiplication of two unsigned 16-bit unsigned ints into a 32-bit unsigned long
-// mul16u(word zp($12) a, word zp($c) b)
+// mul16u(word zp($16) a, word zp($c) b)
 mul16u: {
     .label mb = $e
-    .label a = $12
+    .label a = $16
     .label res = 8
     .label return = 8
     .label b = $c
@@ -488,16 +492,25 @@ mul16u: {
 sin16s: {
     .label __4 = $1e
     .label x = $e
-    .label return = $16
+    .label return = $1c
+    // sinx = x - x^3/6 + x5/128;
     .label x1 = $22
-    .label x2 = $1c
-    .label x3 = $1c
+    // u[1.15]
+    .label x2 = $12
+    // u[2.14] x^2
+    .label x3 = $12
+    // u[2.14] x^3
     .label x3_6 = $24
-    .label usinx = $16
-    .label x4 = $1c
+    // u[1.15] x^3/6;
+    .label usinx = $1c
+    // u[1.15] x - x^3/6
+    .label x4 = $12
+    // u[3.13] x^4
     .label x5 = $24
+    // u[4.12] x^5
     .label x5_128 = $24
-    .label sinx = $16
+    // u[1.15] (first bit is always zero)
+    .label sinx = $1c
     // if(x >= PI_u4f28 )
     lda.z x+3
     cmp #>PI_u4f28>>$10
@@ -695,14 +708,14 @@ sin16s: {
 }
 // Calculate val*val for two unsigned int values - the result is 16 selected bits of the 32-bit result.
 // The select parameter indicates how many of the highest bits of the 32-bit result to skip
-// mulu16_sel(word zp($1c) v1, word zp($c) v2, byte register(X) select)
+// mulu16_sel(word zp($12) v1, word zp($c) v2, byte register(X) select)
 mulu16_sel: {
     .label __0 = 8
     .label __1 = 8
-    .label v1 = $1c
+    .label v1 = $12
     .label v2 = $c
     .label return = $24
-    .label return_1 = $1c
+    .label return_1 = $12
     // mul16u(v1, v2)
     lda.z v1
     sta.z mul16u.a
@@ -733,7 +746,7 @@ mulu16_sel: {
 // The 16-bit unsigned int remainder can be found in rem16u after the division
 div32u16u: {
     .label quotient_hi = $24
-    .label quotient_lo = $1c
+    .label quotient_lo = $12
     .label return = $18
     // divr16u(>dividend, divisor, 0)
     lda #<PI2_u4f28>>$10
@@ -774,12 +787,12 @@ div32u16u: {
 // Returns the quotient dividend/divisor.
 // The final remainder will be set into the global variable rem16u
 // Implemented using simple binary division
-// divr16u(word zp($16) dividend, word zp($12) rem)
+// divr16u(word zp($1c) dividend, word zp($16) rem)
 divr16u: {
-    .label rem = $12
-    .label dividend = $16
-    .label quotient = $1c
-    .label return = $1c
+    .label rem = $16
+    .label dividend = $1c
+    .label quotient = $12
+    .label return = $12
     ldx #0
     txa
     sta.z quotient

src/test/ref/examples/sinsprites/sinus-sprites.asm

@@ -29,11 +29,11 @@
   .label sprites = $2000
   .label SCREEN = $400
   // Current index within the progress cursor (0-7)
-  .label progress_idx = 4
+  .label progress_idx = $a
   // Current position of the progress cursor
-  .label progress_cursor = 5
-  .label sin_idx_x = 2
-  .label sin_idx_y = 3
+  .label progress_cursor = 2
+  .label sin_idx_x = 6
+  .label sin_idx_y = 8
 main: {
     // init()
     jsr init
@@ -53,10 +53,10 @@ anim: {
     .label __7 = $a
     .label xidx = 9
     .label yidx = 4
-    .label x = $10
+    .label x = $f
     .label x_msb = $a
     .label j2 = $b
-    .label j = 8
+    .label j = 5
     // (VICII->BORDER_COLOR)++;
     inc VICII+OFFSET_STRUCT_MOS6569_VICII_BORDER_COLOR
     // xidx = sin_idx_x
@@ -214,7 +214,7 @@ init: {
     rts
 }
 clear_screen: {
-    .label sc = $10
+    .label sc = $f
     lda #<SCREEN
     sta.z sc
     lda #>SCREEN
@@ -248,15 +248,16 @@ clear_screen: {
 // - length is the length of the sine table
 // - min is the minimum value of the generated sinus
 // - max is the maximum value of the generated sinus
-// gen_sintab(byte* zp($10) sintab, byte zp($a) length, byte zp(9) min, byte register(X) max)
+// gen_sintab(byte* zp($f) sintab, byte zp(8) length, byte zp(6) min, byte register(X) max)
 gen_sintab: {
+    // amplitude/2
     // amplitude/2
     .label f_2pi = $e2e5
     .label __20 = $13
-    .label i = $b
-    .label min = 9
-    .label length = $a
-    .label sintab = $10
+    .label i = 9
+    .label min = 6
+    .label length = 8
+    .label sintab = $f
     // setFAC((unsigned int)max)
     txa
     sta.z setFAC.w
@@ -574,14 +575,14 @@ setARGtoFAC: {
     rts
 }
 // Initialize the PETSCII progress bar
-// progress_init(byte* zp(5) line)
+// progress_init(byte* zp(2) line)
 progress_init: {
-    .label line = 5
+    .label line = 2
     rts
 }
 gen_sprites: {
-    .label spr = $10
-    .label i = 8
+    .label spr = 2
+    .label i = $b
     lda #<sprites
     sta.z spr
     lda #>sprites
@@ -617,19 +618,22 @@ gen_sprites: {
 // Generate a sprite from a C64 CHARGEN character (by making each pixel 3x3 pixels large)
 // - c is the character to generate
 // - sprite is a pointer to the position of the sprite to generate
-// gen_chargen_sprite(byte register(X) ch, byte* zp($13) sprite)
+// gen_chargen_sprite(byte register(X) ch, byte* zp($f) sprite)
 gen_chargen_sprite: {
-    .label __0 = $15
-    .label __14 = $15
-    .label sprite = $13
-    .label chargen = $15
-    .label bits = $a
+    .label __0 = $13
+    .label __14 = $13
+    .label sprite = $f
+    .label chargen = $13
+    // current chargen line
+    .label bits = 5
     // current sprite char
-    .label s_gen = $f
-    .label x = $b
-    .label y = 9
+    // current sprite char
+    .label s_gen = 9
+    .label x = 6
+    .label y = 4
     // Find the current chargen pixel (c)
-    .label c = $c
+    // Find the current chargen pixel (c)
+    .label c = 8
     // ((unsigned int)ch)*8
     txa
     sta.z __14
@@ -746,11 +750,11 @@ gen_chargen_sprite: {
 }
 place_sprites: {
     .label sprites_ptr = SCREEN+$3f8
-    .label spr_id = 9
-    .label spr_x = $b
-    .label col = $f
-    .label j2 = $c
-    .label j = $a
+    .label spr_id = 6
+    .label spr_x = 9
+    .label col = $b
+    .label j2 = $a
+    .label j = 8
     // VICII->SPRITES_ENABLE = %01111111
     lda #$7f
     sta VICII+OFFSET_STRUCT_MOS6569_VICII_SPRITES_ENABLE

src/test/ref/examples/zpcode/zpcode.asm

@@ -14,8 +14,22 @@
 
   .label RASTER = $d012
   .label BG_COLOR = $d020
+.segment ZpCode
+zpLoop: {
+    ldx #0
+  __b1:
+    // (*BG_COLOR)++;
+    inc BG_COLOR
+    // for(char i:0..100)
+    inx
+    cpx #$65
+    bne __b1
+    // }
+    rts
+}
 .segment Code
 main: {
+    // Transfer ZP-code to zeropage
     // Transfer ZP-code to zeropage
     .label zpCode = zpLoop
     // asm
@@ -48,20 +62,6 @@ main: {
     inx
     jmp __b1
 }
-.segment ZpCode
-zpLoop: {
-    ldx #0
-  __b1:
-    // (*BG_COLOR)++;
-    inc BG_COLOR
-    // for(char i:0..100)
-    inx
-    cpx #$65
-    bne __b1
-    // }
-    rts
-}
-.segment Code
 // Code in "normal" memory
 loop: {
     ldx #0

src/test/ref/fastmultiply-127.asm

@@ -6,15 +6,16 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
-  .label print_line_cursor = 3
-  .label print_char_cursor = 5
+  .label print_screen = $400
+  .label print_char_cursor = 8
+  .label print_line_cursor = 6
 main: {
     // print_cls()
     jsr print_cls
     // print_str("unsigned")
-    lda #<$400
+    lda #<print_screen
     sta.z print_char_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_char_cursor+1
     lda #<str
     sta.z print_str.str
@@ -22,9 +23,9 @@ main: {
     sta.z print_str.str+1
     jsr print_str
     // print_ln()
-    lda #<$400
+    lda #<print_screen
     sta.z print_line_cursor
-    lda #>$400
+    lda #>print_screen
     sta.z print_line_cursor+1
     jsr print_ln
     // print_mulf8u127(0,0)
@@ -141,10 +142,10 @@ main: {
     str1: .text "signed"
     .byte 0
 }
-// print_mulf8s127(signed byte zp(9) a, signed byte zp(2) b)
+// print_mulf8s127(signed byte zp(5) a, signed byte zp(2) b)
 print_mulf8s127: {
-    .label c = 7
-    .label a = 9
+    .label c = 3
+    .label a = 5
     .label b = 2
     // mulf8s127(a,b)
     ldy.z b
@@ -198,9 +199,9 @@ print_ln: {
     rts
 }
 // Print a signed int as HEX
-// print_sint(signed word zp(7) w)
+// print_sint(signed word zp(3) w)
 print_sint: {
-    .label w = 7
+    .label w = 3
     // if(w<0)
     lda.z w+1
     bmi __b1
@@ -241,9 +242,9 @@ print_char: {
     rts
 }
 // Print a unsigned int as HEX
-// print_uint(word zp(7) w)
+// print_uint(word zp(3) w)
 print_uint: {
-    .label w = 7
+    .label w = 3
     // print_uchar(>w)
     ldx.z w+1
     jsr print_uchar
@@ -302,15 +303,15 @@ print_schar: {
     tax
     jmp __b2
 }
-// mulf8s127(signed byte zp(9) a, signed byte register(Y) b)
+// mulf8s127(signed byte zp(5) a, signed byte register(Y) b)
 mulf8s127: {
-    .label __9 = $a
-    .label __10 = $c
-    .label __11 = $a
-    .label __12 = $c
-    .label a = 9
-    .label return = 7
-    .label c = 7
+    .label __9 = 8
+    .label __10 = $a
+    .label __11 = 8
+    .label __12 = $a
+    .label a = 5
+    .label return = 3
+    .label c = 3
     // mulf8u127((unsigned char)a, (unsigned char)b)
     ldx.z a
     tya
@@ -387,7 +388,7 @@ mulf8u127: {
     .label res = $fe
     .label resL = $fe
     .label resH = $ff
-    .label return = 7
+    .label return = 3
     // *memA = a
     stx memA
     // *memB = b
@@ -420,9 +421,9 @@ mulf8u127: {
     rts
 }
 // Print a zero-terminated string
-// print_str(byte* zp(7) str)
+// print_str(byte* zp(3) str)
 print_str: {
-    .label str = 7
+    .label str = 3
   __b1:
     // while(*str)
     ldy #0
@@ -443,10 +444,10 @@ print_str: {
   !:
     jmp __b1
 }
-// print_mulf8u127(byte register(Y) a, byte zp(9) b)
+// print_mulf8u127(byte register(Y) a, byte zp(5) b)
 print_mulf8u127: {
-    .label c = 7
-    .label b = 9
+    .label c = 3
+    .label b = 5
     // mulf8u127(a,b)
     tya
     tax
@@ -490,9 +491,9 @@ print_cls: {
 memset: {
     .const c = ' '
     .const num = $3e8
-    .label str = $400
+    .label str = print_screen
     .label end = str+num
-    .label dst = $a
+    .label dst = 6
     lda #<str
     sta.z dst
     lda #>str

src/test/ref/function-pointer-noarg-3.asm

@@ -2,10 +2,20 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
-  // kickasm
-  ff:
-    jmp (main.f)
-
+fn2: {
+    .label BG_COLOR = $d021
+    // (*BG_COLOR)++;
+    inc BG_COLOR
+    // }
+    rts
+}
+fn1: {
+    .label BORDER_COLOR = $d020
+    // (*BORDER_COLOR)++;
+    inc BORDER_COLOR
+    // }
+    rts
+}
 main: {
     .label f = 2
     ldx #0
@@ -34,17 +44,7 @@ main: {
         
     jmp __b2
 }
-fn2: {
-    .label BG_COLOR = $d021
-    // (*BG_COLOR)++;
-    inc BG_COLOR
-    // }
-    rts
-}
-fn1: {
-    .label BORDER_COLOR = $d020
-    // (*BORDER_COLOR)++;
-    inc BORDER_COLOR
-    // }
-    rts
-}
+// Inline KickAsm function
+ff:
+jmp (main.f)
+

src/test/ref/function-pointer-noarg-call-10.asm

@@ -1,15 +1,16 @@
 // Tests calling into different function pointers which call a common sub-method
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
   .label idx = 7
-__bbegin:
-  // idx = 0
-  lda #0
-  sta.z idx
-  jsr main
-  rts
+_start: {
+    // idx = 0
+    lda #0
+    sta.z idx
+    jsr main
+    rts
+}
 main: {
     // do10(&hello)
     lda #<hello

src/test/ref/function-pointer-noarg-call-2.asm

@@ -2,6 +2,20 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
+fn2: {
+    .label BG_COLOR = $d021
+    // (*BG_COLOR)++;
+    inc BG_COLOR
+    // }
+    rts
+}
+fn1: {
+    .label BORDER_COLOR = $d020
+    // (*BORDER_COLOR)++;
+    inc BORDER_COLOR
+    // }
+    rts
+}
 main: {
     .label i = 2
     .label f = 3
@@ -33,17 +47,3 @@ main: {
   bi_f:
     jmp (f)
 }
-fn2: {
-    .label BG_COLOR = $d021
-    // (*BG_COLOR)++;
-    inc BG_COLOR
-    // }
-    rts
-}
-fn1: {
-    .label BORDER_COLOR = $d020
-    // (*BORDER_COLOR)++;
-    inc BORDER_COLOR
-    // }
-    rts
-}

src/test/ref/function-pointer-noarg-call-3.asm

@@ -2,6 +2,20 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
+fn2: {
+    .label BG_COLOR = $d021
+    // (*BG_COLOR)++;
+    inc BG_COLOR
+    // }
+    rts
+}
+fn1: {
+    .label BORDER_COLOR = $d020
+    // (*BORDER_COLOR)++;
+    inc BORDER_COLOR
+    // }
+    rts
+}
 main: {
     .label __1 = 3
     .label i = 2
@@ -40,17 +54,3 @@ getfn: {
     // }
     rts
 }
-fn2: {
-    .label BG_COLOR = $d021
-    // (*BG_COLOR)++;
-    inc BG_COLOR
-    // }
-    rts
-}
-fn1: {
-    .label BORDER_COLOR = $d020
-    // (*BORDER_COLOR)++;
-    inc BORDER_COLOR
-    // }
-    rts
-}

src/test/ref/function-pointer-noarg-call-4.asm

@@ -2,17 +2,6 @@
 .pc = $801 "Basic"
 :BasicUpstart(main)
 .pc = $80d "Program"
-main: {
-    ldx #0
-  __b2:
-    // (*getfn(++i))();
-    inx
-    // getfn(++i)
-    jsr getfn
-    // (*getfn(++i))()
-    jsr fn1
-    jmp __b2
-}
 fn1: {
     .label BORDER_COLOR = $d020
     // (*BORDER_COLOR)++;
@@ -20,6 +9,12 @@ fn1: {
     // }
     rts
 }
-getfn: {
-    rts
+main: {
+    ldx #0
+  __b2:
+    // (*getfn(++i))();
+    inx
+    // (*getfn(++i))()
+    jsr fn1
+    jmp __b2
 }

src/test/ref/function-pointer-noarg-call-7.asm

@@ -1,15 +1,16 @@
 // Tests calling into a function pointer with local variables
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
   .label idx = 3
-__bbegin:
-  // idx = 0
-  lda #0
-  sta.z idx
-  jsr main
-  rts
+_start: {
+    // idx = 0
+    lda #0
+    sta.z idx
+    jsr main
+    rts
+}
 main: {
     // do10(f)
     jsr do10

src/test/ref/function-pointer-noarg-call-8.asm

@@ -1,19 +1,20 @@
 // Tests calling into a function pointer with local variables
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
   .label msg = 3
   .label idx = 5
-__bbegin:
-  // msg
-  lda #<0
-  sta.z msg
-  sta.z msg+1
-  // idx = 0
-  sta.z idx
-  jsr main
-  rts
+_start: {
+    // msg
+    lda #<0
+    sta.z msg
+    sta.z msg+1
+    // idx = 0
+    sta.z idx
+    jsr main
+    rts
+}
 main: {
     // msg = msg1
     lda #<msg1

src/test/ref/function-pointer-noarg-call-9.asm

@@ -1,15 +1,16 @@
 // Tests calling into a function pointer which modifies global volatile
 .pc = $801 "Basic"
-:BasicUpstart(__bbegin)
+:BasicUpstart(_start)
 .pc = $80d "Program"
   .label SCREEN = $400
   .label idx = 2
-__bbegin:
-  // idx = 0
-  lda #0
-  sta.z idx
-  jsr main
-  rts
+_start: {
+    // idx = 0
+    lda #0
+    sta.z idx
+    jsr main
+    rts
+}
 main: {
     // (*f)()
     jsr fn1