version 5.4

implemented word bit shifts by variable number of bits.

.idea/misc.xml

@@ -16,7 +16,7 @@
       </list>
     </option>
   </component>
-  <component name="ProjectRootManager" version="2" languageLevel="JDK_1_8" default="false" project-jdk-name="Kotlin SDK" project-jdk-type="KotlinSDK">
+  <component name="ProjectRootManager" version="2" languageLevel="JDK_11" default="false" project-jdk-name="Kotlin SDK" project-jdk-type="KotlinSDK">
     <output url="file://$PROJECT_DIR$/out" />
   </component>
 </project>

README.md

@@ -20,36 +20,36 @@ Full documentation (syntax reference, how to use the language and the compiler,
 https://prog8.readthedocs.io/
 
 
-What use Prog8 provide?
------------------------
+What does Prog8 provide?
+------------------------
 
-- reduction of source code length over raw assembly
+- big reduction of source code length over raw assembly
 - modularity, symbol scoping, subroutines
 - various data types other than just bytes (16-bit words, floats, strings)
 - automatic variable allocations, automatic string and array variables and string sharing
 - subroutines with an input- and output parameter signature
-- constant folding in expressions
+- no stack frame allocations because parameters and local variables are automatically allocated statically
+- constant folding in expressions and other high-level program optimizations
 - conditional branches
-- 'when' statement to provide a concise jump table alternative to if/elseif chains
-- structs to group together sets of variables and manipulate them at once
 - floating point operations  (requires the C64 Basic ROM routines for this)
-- abstracting away low level aspects such as ZeroPage handling, program startup, explicit memory addresses
-- various code optimizations (code structure, logical and numerical expressions, unused code removal...)
-- inline assembly allows you to have full control when every cycle or byte matters
+- 'when' statement to provide a concise jump table alternative to if/elseif chains
 - many built-in functions such as ``sin``, ``cos``, ``rnd``, ``abs``, ``min``, ``max``, ``sqrt``, ``msb``, ``rol``, ``ror``, ``swap``, ``memset``, ``memcopy``, ``sort`` and ``reverse``
+- structs to group together sets of variables and manipulate them at once
+- convenience abstractions for low level aspects such as ZeroPage handling, program startup, explicit memory addresses
+- fast execution speed due to compilation to native assembly code
+- inline assembly allows you to have full control when every cycle or byte matters
 
 *Rapid edit-compile-run-debug cycle:*
 
-- use a modern PC to do the work on
-- very quick compilation times
+- use a modern PC to do the work on, use nice editors and enjoy quick compilation times
 - can automatically run the program in the Vice emulator after succesful compilation
 - breakpoints, that let the Vice emulator drop into the monitor if execution hits them
 - source code labels automatically loaded in Vice emulator so it can show them in disassembly
 
 *Two supported compiler targets* (contributions to improve these or to add support for other machines are welcome!):
 
-- "c64": Commodore-64  (6510 CPU = almost a 6502)  premium support.
-- "cx16": [CommanderX16](https://www.commanderx16.com)  (65c02 CPU)  experimental support.
+- "c64": Commodore-64  (6510 CPU = almost a 6502), the main target.
+- "cx16": [CommanderX16](https://www.commanderx16.com)  (65c02 CPU) .
 - If you only use standard kernel and prog8 library routines, it is possible to compile the *exact same program* for both machines (just change the compiler target flag)!
 
 
@@ -61,7 +61,7 @@ Additional required tools
 A recent .exe version of this tool for Windows can be obtained from my [clone](https://github.com/irmen/64tass/releases) of this project.
 For other platforms it is very easy to compile it yourself (make ; make install).
 
-A **Java runtime (jre or jdk), version 8 or newer**  is required to run a prepackaged version of the compiler.
+A **Java runtime (jre or jdk), version 11 or newer**  is required to run a prepackaged version of the compiler.
 If you want to build it from source, you'll need a Java SDK + Kotlin 1.3.x SDK (or for instance,
 IntelliJ IDEA with the Kotlin plugin).
 

compiler/build.gradle

@@ -1,11 +1,11 @@
 buildscript {
     dependencies {
-        classpath "org.jetbrains.kotlin:kotlin-gradle-plugin:1.4.10"
+        classpath "org.jetbrains.kotlin:kotlin-gradle-plugin:1.4.20"
     }
 }
 
 plugins {
-    // id "org.jetbrains.kotlin.jvm" version "1.4.10"
+    // id "org.jetbrains.kotlin.jvm" version "1.4.20"
     id 'application'
     id 'org.jetbrains.dokka' version "0.9.18"
     id 'com.github.johnrengelman.shadow' version '5.2.0'
@@ -15,8 +15,8 @@ plugins {
 apply plugin: "kotlin"
 apply plugin: "java"
 
-targetCompatibility = 1.8
-sourceCompatibility = 1.8
+targetCompatibility = 11
+sourceCompatibility = 11
 
 repositories {
     mavenLocal()
@@ -45,7 +45,7 @@ dependencies {
 
 compileKotlin {
     kotlinOptions {
-        jvmTarget = "1.8"
+        jvmTarget = "11"
         // verbose = true
         // freeCompilerArgs += "-XXLanguage:+NewInference"
     }
@@ -53,7 +53,7 @@ compileKotlin {
 
 compileTestKotlin {
     kotlinOptions {
-        jvmTarget = "1.8"
+        jvmTarget = "11"
     }
 }
 

compiler/res/prog8lib/c64/floats.asm

@@ -55,13 +55,76 @@ w2float		.proc
 		jmp  ub2float._fac_to_mem
 		.pend
 
+
+cast_from_uw	.proc
+		; -- uword in A/Y into float var at (P8ZP_SCRATCH_W2)
+		stx  P8ZP_SCRATCH_REG
+		jsr  GIVUAYFAY
+		jmp  ub2float._fac_to_mem
+		.pend
+
+
+cast_from_w	.proc
+		; -- word in A/Y into float var at (P8ZP_SCRATCH_W2)
+		stx  P8ZP_SCRATCH_REG
+		jsr  GIVAYFAY
+		jmp  ub2float._fac_to_mem
+		.pend
+
+
+cast_from_ub	.proc
+		; -- ubyte in Y into float var at (P8ZP_SCRATCH_W2)
+		stx  P8ZP_SCRATCH_REG
+		jsr  FREADUY
+		jmp  ub2float._fac_to_mem
+		.pend
+
+
+cast_from_b	.proc
+		; -- byte in A into float var at (P8ZP_SCRATCH_W2)
+		stx  P8ZP_SCRATCH_REG
+		jsr  FREADSA
+		jmp  ub2float._fac_to_mem
+		.pend
+
+cast_as_uw_into_ya	.proc               ; also used for float 2 ub
+		; -- cast float at A/Y to uword into Y/A
+		jsr  MOVFM
+		jmp  cast_FAC1_as_uw_into_ya
+		.pend
+
+cast_as_w_into_ay	.proc               ; also used for float 2 b
+		; -- cast float at A/Y to word into A/Y
+		jsr  MOVFM
+		jmp  cast_FAC1_as_w_into_ay
+		.pend
+
+cast_FAC1_as_uw_into_ya	.proc               ; also used for float 2 ub
+		; -- cast fac1 to uword into Y/A
+		stx  P8ZP_SCRATCH_REG
+		jsr  GETADR     ; into Y/A
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+cast_FAC1_as_w_into_ay	.proc               ; also used for float 2 b
+		; -- cast fac1 to word into A/Y
+		stx  P8ZP_SCRATCH_REG
+		jsr  AYINT
+		ldy  $64
+		lda  $65
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+
 stack_b2float	.proc
 		; -- b2float operating on the stack
 		inx
 		lda  P8ESTACK_LO,x
 		stx  P8ZP_SCRATCH_REG
 		jsr  FREADSA
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend
 
 stack_w2float	.proc
@@ -71,7 +134,7 @@ stack_w2float	.proc
 		lda  P8ESTACK_HI,x
 		stx  P8ZP_SCRATCH_REG
 		jsr  GIVAYF
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend
 
 stack_ub2float	.proc
@@ -82,7 +145,7 @@ stack_ub2float	.proc
 		tay
 		lda  #0
 		jsr  GIVAYF
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend
 
 stack_uw2float	.proc
@@ -92,7 +155,7 @@ stack_uw2float	.proc
 		ldy  P8ESTACK_HI,x
 		stx  P8ZP_SCRATCH_REG
 		jsr  GIVUAYFAY
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend
 
 stack_float2w	.proc               ; also used for float2b
@@ -146,22 +209,6 @@ push_float	.proc
 		rts
 		.pend
 
-func_rndf	.proc
-		; -- put a random floating point value on the stack
-		stx  P8ZP_SCRATCH_REG
-		lda  #1
-		jsr  FREADSA
-		jsr  RND		; rng into fac1
-		ldx  #<_rndf_rnum5
-		ldy  #>_rndf_rnum5
-		jsr  MOVMF	; fac1 to mem X/Y
-		ldx  P8ZP_SCRATCH_REG
-		lda  #<_rndf_rnum5
-		ldy  #>_rndf_rnum5
-		jmp  push_float
-_rndf_rnum5	.byte  0,0,0,0,0
-		.pend
-
 pop_float	.proc
 		; ---- pops mflpt5 from stack to memory A/Y
 		; (frees 3 stack positions = 6 bytes of which 1 is padding)
@@ -198,16 +245,6 @@ pop_float_fac1	.proc
 		jmp  MOVFM
 		.pend
 
-pop_float_fac2	.proc
-		; -- pops float from stack into FAC2
-		lda  #<fmath_float1
-		ldy  #>fmath_float1
-		jsr  pop_float
-		lda  #<fmath_float1
-		ldy  #>fmath_float1
-		jmp  CONUPK
-		.pend
-
 copy_float	.proc
 		; -- copies the 5 bytes of the mflt value pointed to by SCRATCH_ZPWORD1,
 		;    into the 5 bytes pointed to by A/Y.  Clobbers A,Y.
@@ -273,9 +310,11 @@ pop_2_floats_f2_in_fac1	.proc
 fmath_float1	.byte 0,0,0,0,0	; storage for a mflpt5 value
 fmath_float2	.byte 0,0,0,0,0	; storage for a mflpt5 value
 
-push_fac1_as_result	.proc
-		; -- push the float in FAC1 onto the stack, and return from calculation
-		ldx  #<fmath_float1
+
+push_fac1	.proc
+		; -- push the float in FAC1 onto the stack
+		stx  P8ZP_SCRATCH_REG
+_internal	ldx  #<fmath_float1
 		ldy  #>fmath_float1
 		jsr  MOVMF
 		lda  #<fmath_float1
@@ -284,6 +323,7 @@ push_fac1_as_result	.proc
 		jmp  push_float
 		.pend
 
+
 pow_f		.proc
 		; -- push f1 ** f2 on stack
 		lda  #<fmath_float2
@@ -299,8 +339,7 @@ pow_f		.proc
 		lda  #<fmath_float2
 		ldy  #>fmath_float2
 		jsr  FPWR
-		ldx  P8ZP_SCRATCH_REG
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend
 
 div_f		.proc
@@ -310,7 +349,7 @@ div_f		.proc
 		lda  #<fmath_float1
 		ldy  #>fmath_float1
 		jsr  FDIV
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend
 
 add_f		.proc
@@ -320,7 +359,7 @@ add_f		.proc
 		lda  #<fmath_float1
 		ldy  #>fmath_float1
 		jsr  FADD
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend
 
 sub_f		.proc
@@ -330,7 +369,7 @@ sub_f		.proc
 		lda  #<fmath_float1
 		ldy  #>fmath_float1
 		jsr  FSUB
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend
 
 mul_f		.proc
@@ -340,7 +379,7 @@ mul_f		.proc
 		lda  #<fmath_float1
 		ldy  #>fmath_float1
 		jsr  FMULT
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend
 
 neg_f		.proc
@@ -351,11 +390,96 @@ neg_f		.proc
 		rts
 		.pend
 
-abs_f		.proc
-		; -- strip the sign bit on the stack
-		lda  P8ESTACK_HI+3,x
-		and  #$7f
-		sta  P8ESTACK_HI+3,x
+var_fac1_less_f	.proc
+		; -- is the float in FAC1 < the variable AY?
+		stx  P8ZP_SCRATCH_REG
+		jsr  FCOMP
+		ldx  P8ZP_SCRATCH_REG
+		cmp  #255
+		beq  +
+		lda  #0
+		rts
++		lda  #1
+		rts
+		.pend
+
+var_fac1_lesseq_f	.proc
+		; -- is the float in FAC1 <= the variable AY?
+		stx  P8ZP_SCRATCH_REG
+		jsr  FCOMP
+		ldx  P8ZP_SCRATCH_REG
+		cmp  #0
+		beq  +
+		cmp  #255
+		beq  +
+		lda  #0
+		rts
++		lda  #1
+		rts
+		.pend
+
+var_fac1_greater_f	.proc
+		; -- is the float in FAC1 > the variable AY?
+		stx  P8ZP_SCRATCH_REG
+		jsr  FCOMP
+		ldx  P8ZP_SCRATCH_REG
+		cmp  #1
+		beq  +
+		lda  #0
++		rts
+		.pend
+
+var_fac1_greatereq_f	.proc
+		; -- is the float in FAC1 >= the variable AY?
+		stx  P8ZP_SCRATCH_REG
+		jsr  FCOMP
+		ldx  P8ZP_SCRATCH_REG
+		cmp  #0
+		beq  +
+		cmp  #1
+		beq  +
+		lda  #0
+		rts
++		lda  #1
+		rts
+		.pend
+
+var_fac1_notequal_f	.proc
+		; -- are the floats numbers in FAC1 and the variable AY *not* identical?
+		stx  P8ZP_SCRATCH_REG
+		jsr  FCOMP
+		ldx  P8ZP_SCRATCH_REG
+		and  #1
+		rts
+		.pend
+
+vars_equal_f	.proc
+		; -- are the mflpt5 numbers in P8ZP_SCRATCH_W1 and AY identical?
+		sta  P8ZP_SCRATCH_W2
+		sty  P8ZP_SCRATCH_W2+1
+		ldy  #0
+		lda  (P8ZP_SCRATCH_W1),y
+		cmp  (P8ZP_SCRATCH_W2),y
+		bne  _false
+		iny
+		lda  (P8ZP_SCRATCH_W1),y
+		cmp  (P8ZP_SCRATCH_W2),y
+		bne  _false
+		iny
+		lda  (P8ZP_SCRATCH_W1),y
+		cmp  (P8ZP_SCRATCH_W2),y
+		bne  _false
+		iny
+		lda  (P8ZP_SCRATCH_W1),y
+		cmp  (P8ZP_SCRATCH_W2),y
+		bne  _false
+		iny
+		lda  (P8ZP_SCRATCH_W1),y
+		cmp  (P8ZP_SCRATCH_W2),y
+		bne  _false
+		lda  #1
+		rts
+_false		lda  #0
 		rts
 		.pend
 
@@ -396,6 +520,40 @@ notequal_f	.proc
 		rts
 		.pend
 
+vars_less_f	.proc
+		; -- is float in AY < float in P8ZP_SCRATCH_W2 ?
+		jsr  MOVFM
+		lda  P8ZP_SCRATCH_W2
+		ldy  P8ZP_SCRATCH_W2+1
+		stx  P8ZP_SCRATCH_REG
+		jsr  FCOMP
+		ldx  P8ZP_SCRATCH_REG
+		cmp  #255
+		bne  +
+		lda  #1
+		rts
++		lda  #0
+		rts
+		.pend
+
+vars_lesseq_f	.proc
+		; -- is float in AY <= float in P8ZP_SCRATCH_W2 ?
+		jsr  MOVFM
+		lda  P8ZP_SCRATCH_W2
+		ldy  P8ZP_SCRATCH_W2+1
+		stx  P8ZP_SCRATCH_REG
+		jsr  FCOMP
+		ldx  P8ZP_SCRATCH_REG
+		cmp  #255
+		bne  +
+-		lda  #1
+		rts
++		cmp  #0
+		beq  -
+		lda  #0
+		rts
+		.pend
+
 less_f		.proc
 		; -- is f1 < f2?
 		jsr  compare_floats
@@ -457,240 +615,22 @@ _return_true	lda  #1
 		bne  _return_result
 		.pend
 
-func_sin	.proc
-		; -- push sin(f) back onto stack
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		jsr  SIN
-		jmp  push_fac1_as_result
-		.pend
-
-func_cos	.proc
-		; -- push cos(f) back onto stack
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		jsr  COS
-		jmp  push_fac1_as_result
-		.pend
-
-func_tan	.proc
-		; -- push tan(f) back onto stack
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		jsr  TAN
-		jmp  push_fac1_as_result
-		.pend
-
-func_atan	.proc
-		; -- push atan(f) back onto stack
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		jsr  ATN
-		jmp  push_fac1_as_result
-		.pend
-
-func_ln		.proc
-		; -- push ln(f) back onto stack
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		jsr  LOG
-		jmp  push_fac1_as_result
-		.pend
-
-func_log2	.proc
-		; -- push log base 2, ln(f)/ln(2), back onto stack
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		jsr  LOG
-		jsr  MOVEF
-		lda  #<c64.FL_LOG2
-		ldy  #>c64.FL_LOG2
-		jsr  MOVFM
-		jsr  FDIVT
-		jmp  push_fac1_as_result
-		.pend
-
-func_sqrt	.proc
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		jsr  SQR
-		jmp  push_fac1_as_result
-		.pend
-
-func_rad	.proc
-		; -- convert degrees to radians (d * pi / 180)
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		lda  #<_pi_div_180
-		ldy  #>_pi_div_180
-		jsr  FMULT
-		jmp  push_fac1_as_result
-_pi_div_180	.byte 123, 14, 250, 53, 18		; pi / 180
-		.pend
-
-func_deg	.proc
-		; -- convert radians to degrees (d * (1/ pi * 180))
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		lda  #<_one_over_pi_div_180
-		ldy  #>_one_over_pi_div_180
-		jsr  FMULT
-		jmp  push_fac1_as_result
-_one_over_pi_div_180	.byte 134, 101, 46, 224, 211		; 1 / (pi * 180)
-		.pend
-
-func_round	.proc
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		jsr  FADDH
-		jsr  INT
-		jmp  push_fac1_as_result
-		.pend
-
-func_floor	.proc
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		jsr  INT
-		jmp  push_fac1_as_result
-		.pend
-
-func_ceil	.proc
-		; -- ceil: tr = int(f); if tr==f -> return  else return tr+1
-		jsr  pop_float_fac1
-		stx  P8ZP_SCRATCH_REG
-		ldx  #<fmath_float1
-		ldy  #>fmath_float1
-		jsr  MOVMF
-		jsr  INT
-		lda  #<fmath_float1
-		ldy  #>fmath_float1
-		jsr  FCOMP
-		cmp  #0
-		beq  +
-		lda  #<FL_ONE_const
-		ldy  #>FL_ONE_const
-		jsr  FADD
-+		jmp  push_fac1_as_result
-		.pend
-
-func_any_f	.proc
-		inx
-		lda  P8ESTACK_LO,x	; array size
+set_array_float_from_fac1	.proc
+		; -- set the float in FAC1 in the array (index in A, array in P8ZP_SCRATCH_W1)
 		sta  P8ZP_SCRATCH_B1
 		asl  a
 		asl  a
 		clc
-		adc  P8ZP_SCRATCH_B1	; times 5 because of float
-		jmp  prog8_lib.func_any_b._entry
-		.pend
-
-func_all_f	.proc
-		inx
-		jsr  prog8_lib.peek_address
-		lda  P8ESTACK_LO,x	; array size
-		sta  P8ZP_SCRATCH_B1
-		asl  a
-		asl  a
-		clc
-		adc  P8ZP_SCRATCH_B1	; times 5 because of float
-		tay
-		dey
--		lda  (P8ZP_SCRATCH_W1),y
-		clc
-		dey
-		adc  (P8ZP_SCRATCH_W1),y
-		dey
-		adc  (P8ZP_SCRATCH_W1),y
-		dey
-		adc  (P8ZP_SCRATCH_W1),y
-		dey
-		adc  (P8ZP_SCRATCH_W1),y
-		dey
-		cmp  #0
-		beq  +
-		cpy  #255
-        	bne  -
-		lda  #1
-		sta  P8ESTACK_LO+1,x
-		rts
-+		sta  P8ESTACK_LO+1,x
-		rts
-		.pend
-
-func_max_f	.proc
-		lda  #255
-		sta  _minmax_cmp+1
-		lda  #<_largest_neg_float
-		ldy  #>_largest_neg_float
-_minmax_entry	jsr  MOVFM
-		jsr  prog8_lib.pop_array_and_lengthmin1Y
-		stx  floats_store_reg
--		sty  P8ZP_SCRATCH_REG
-		lda  P8ZP_SCRATCH_W1
+		adc  P8ZP_SCRATCH_B1
 		ldy  P8ZP_SCRATCH_W1+1
-		jsr  FCOMP
-_minmax_cmp	cmp  #255			; modified
-		bne  +
-		lda  P8ZP_SCRATCH_W1
-		ldy  P8ZP_SCRATCH_W1+1
-		jsr  MOVFM
-+		lda  P8ZP_SCRATCH_W1
 		clc
-		adc  #5
-		sta  P8ZP_SCRATCH_W1
+		adc  P8ZP_SCRATCH_W1
 		bcc  +
-		inc  P8ZP_SCRATCH_W1+1
-+		ldy  P8ZP_SCRATCH_REG
-		dey
-		cpy  #255
-		bne  -
+		iny
++		stx  floats_store_reg
+		tax
+		jsr  MOVMF
 		ldx  floats_store_reg
-		stx  P8ZP_SCRATCH_REG
-		jmp  push_fac1_as_result
-_largest_neg_float	.byte 255,255,255,255,255		; largest negative float -1.7014118345e+38
-		.pend
-
-func_min_f	.proc
-		lda  #1
-		sta  func_max_f._minmax_cmp+1
-		lda  #<_largest_pos_float
-		ldy  #>_largest_pos_float
-		jmp  func_max_f._minmax_entry
-_largest_pos_float	.byte  255,127,255,255,255		; largest positive float
-		rts
-		.pend
-
-func_sum_f	.proc
-		lda  #<FL_ZERO_const
-		ldy  #>FL_ZERO_const
-		jsr  MOVFM
-		jsr  prog8_lib.pop_array_and_lengthmin1Y
-		stx  floats_store_reg
--		sty  P8ZP_SCRATCH_REG
-		lda  P8ZP_SCRATCH_W1
-		ldy  P8ZP_SCRATCH_W1+1
-		jsr  FADD
-		ldy  P8ZP_SCRATCH_REG
-		dey
-		cpy  #255
-		beq  +
-		lda  P8ZP_SCRATCH_W1
-		clc
-		adc  #5
-		sta  P8ZP_SCRATCH_W1
-		bcc  -
-		inc  P8ZP_SCRATCH_W1+1
-		bne  -
-+		ldx  floats_store_reg
-		stx  P8ZP_SCRATCH_REG
-		jmp  push_fac1_as_result
-		.pend
-
-sign_f		.proc
-		jsr  pop_float_fac1
-		jsr  SIGN
-		sta  P8ESTACK_LO,x
-		dex
 		rts
 		.pend
 
@@ -734,16 +674,3 @@ set_array_float		.proc
 		.pend
 
 
-swap_floats	.proc
-		; -- swap floats pointed to by SCRATCH_ZPWORD1, SCRATCH_ZPWORD2
-		ldy  #4
--               lda  (P8ZP_SCRATCH_W1),y
-		pha
-		lda  (P8ZP_SCRATCH_W2),y
-		sta  (P8ZP_SCRATCH_W1),y
-		pla
-		sta  (P8ZP_SCRATCH_W2),y
-		dey
-		bpl  -
-		rts
-		.pend

compiler/res/prog8lib/c64/floats.p8

@@ -213,5 +213,6 @@ sub  print_f  (float value) {
 }
 
 %asminclude "library:c64/floats.asm", ""
+%asminclude "library:c64/floats_funcs.asm", ""
 
 }

compiler/res/prog8lib/c64/floats_funcs.asm

@@ -0,0 +1,437 @@
+; --- floating point builtin functions
+
+
+abs_f_stack	.proc
+		; -- push abs(AY) on stack
+		jsr  floats.MOVFM
+		jsr  floats.ABS
+		jmp  push_fac1
+		.pend
+
+abs_f_fac1	.proc
+		; -- FAC1 = abs(AY)
+		jsr  floats.MOVFM
+		jmp  floats.ABS
+		.pend
+
+func_atan_stack	.proc
+		jsr  func_atan_fac1
+		jmp  push_fac1
+		.pend
+
+func_atan_fac1	.proc
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		jsr  ATN
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_ceil_stack	.proc
+		jsr  func_ceil_fac1
+		jmp  push_fac1
+		.pend
+
+func_ceil_fac1	.proc
+		; -- ceil: tr = int(f); if tr==f -> return  else return tr+1
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		ldx  #<fmath_float1
+		ldy  #>fmath_float1
+		jsr  MOVMF
+		jsr  INT
+		lda  #<fmath_float1
+		ldy  #>fmath_float1
+		jsr  FCOMP
+		cmp  #0
+		beq  +
+		lda  #<FL_ONE_const
+		ldy  #>FL_ONE_const
+		jsr  FADD
++		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_floor_stack	.proc
+		jsr  func_floor_fac1
+		jmp  push_fac1
+		.pend
+
+func_floor_fac1	.proc
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		jsr  INT
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_round_stack	.proc
+		jsr  func_round_fac1
+		jmp  push_fac1
+		.pend
+
+func_round_fac1	.proc
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		jsr  FADDH
+		jsr  INT
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_sin_stack	.proc
+		jsr  func_sin_fac1
+		jmp  push_fac1
+		.pend
+
+func_sin_fac1	.proc
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		jsr  SIN
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_cos_stack	.proc
+		jsr  func_cos_fac1
+		jmp  push_fac1
+		.pend
+
+func_cos_fac1	.proc
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		jsr  COS
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_tan_stack	.proc
+		jsr  func_tan_fac1
+		jmp  push_fac1
+		.pend
+
+func_tan_fac1	.proc
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		jsr  TAN
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_rad_stack	.proc
+		jsr  func_rad_fac1
+		jmp  push_fac1
+		.pend
+
+func_rad_fac1	.proc
+		; -- convert degrees to radians (d * pi / 180)
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		lda  #<_pi_div_180
+		ldy  #>_pi_div_180
+		jsr  FMULT
+		ldx  P8ZP_SCRATCH_REG
+		rts
+_pi_div_180	.byte 123, 14, 250, 53, 18		; pi / 180
+		.pend
+
+func_deg_stack	.proc
+		jsr  func_deg_fac1
+		jmp  push_fac1
+		.pend
+
+func_deg_fac1	.proc
+		; -- convert radians to degrees (d * (1/ pi * 180))
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		lda  #<_one_over_pi_div_180
+		ldy  #>_one_over_pi_div_180
+		jsr  FMULT
+		ldx  P8ZP_SCRATCH_REG
+		rts
+_one_over_pi_div_180	.byte 134, 101, 46, 224, 211		; 1 / (pi * 180)
+		.pend
+
+func_ln_stack	.proc
+		jsr  func_ln_fac1
+		jmp  push_fac1
+		.pend
+
+func_ln_fac1	.proc
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		jsr  LOG
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_log2_stack	.proc
+		jsr  func_log2_fac1
+		jmp  push_fac1
+		.pend
+
+func_log2_fac1	.proc
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		jsr  LOG
+		jsr  MOVEF
+		lda  #<FL_LOG2
+		ldy  #>FL_LOG2
+		jsr  MOVFM
+		jsr  FDIVT
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_sign_f_stack	.proc
+		jsr  func_sign_f_into_A
+		sta  P8ESTACK_LO,x
+		dex
+		rts
+		.pend
+
+func_sign_f_into_A	.proc
+		jsr  MOVFM
+		jmp  SIGN
+		.pend
+
+func_sqrt_stack	.proc
+		jsr  func_sqrt_fac1
+		jmp  push_fac1
+		.pend
+
+func_sqrt_fac1	.proc
+		jsr  MOVFM
+		stx  P8ZP_SCRATCH_REG
+		jsr  SQR
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_rndf_stack	.proc
+		jsr  func_rndf_fac1
+		jmp  push_fac1
+		.pend
+
+func_rndf_fac1	.proc
+		stx  P8ZP_SCRATCH_REG
+		lda  #1
+		jsr  FREADSA
+		jsr  RND		; rng into fac1
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+func_swap_f	.proc
+		; -- swap floats pointed to by SCRATCH_ZPWORD1, SCRATCH_ZPWORD2
+		ldy  #4
+-               lda  (P8ZP_SCRATCH_W1),y
+		pha
+		lda  (P8ZP_SCRATCH_W2),y
+		sta  (P8ZP_SCRATCH_W1),y
+		pla
+		sta  (P8ZP_SCRATCH_W2),y
+		dey
+		bpl  -
+		rts
+		.pend
+
+func_reverse_f	.proc
+		; --- reverse an array of floats (array in P8ZP_SCRATCH_W1, num elements in A)
+_left_index = P8ZP_SCRATCH_W2
+_right_index = P8ZP_SCRATCH_W2+1
+_loop_count = P8ZP_SCRATCH_REG
+		pha
+		jsr  a_times_5
+		sec
+		sbc  #5
+		sta  _right_index
+		lda  #0
+		sta  _left_index
+		pla
+		lsr  a
+		sta  _loop_count
+_loop		; push the left indexed float on the stack
+		ldy  _left_index
+		lda  (P8ZP_SCRATCH_W1),y
+		pha
+		iny
+		lda  (P8ZP_SCRATCH_W1),y
+		pha
+		iny
+		lda  (P8ZP_SCRATCH_W1),y
+		pha
+		iny
+		lda  (P8ZP_SCRATCH_W1),y
+		pha
+		iny
+		lda  (P8ZP_SCRATCH_W1),y
+		pha
+		; copy right index float to left index float
+		ldy  _right_index
+		lda  (P8ZP_SCRATCH_W1),y
+		ldy  _left_index
+		sta  (P8ZP_SCRATCH_W1),y
+		inc  _left_index
+		inc  _right_index
+		ldy  _right_index
+		lda  (P8ZP_SCRATCH_W1),y
+		ldy  _left_index
+		sta  (P8ZP_SCRATCH_W1),y
+		inc  _left_index
+		inc  _right_index
+		ldy  _right_index
+		lda  (P8ZP_SCRATCH_W1),y
+		ldy  _left_index
+		sta  (P8ZP_SCRATCH_W1),y
+		inc  _left_index
+		inc  _right_index
+		ldy  _right_index
+		lda  (P8ZP_SCRATCH_W1),y
+		ldy  _left_index
+		sta  (P8ZP_SCRATCH_W1),y
+		inc  _left_index
+		inc  _right_index
+		ldy  _right_index
+		lda  (P8ZP_SCRATCH_W1),y
+		ldy  _left_index
+		sta  (P8ZP_SCRATCH_W1),y
+		; pop the float off the stack into the right index float
+		ldy  _right_index
+		pla
+		sta  (P8ZP_SCRATCH_W1),y
+		dey
+		pla
+		sta  (P8ZP_SCRATCH_W1),y
+		dey
+		pla
+		sta  (P8ZP_SCRATCH_W1),y
+		dey
+		pla
+		sta  (P8ZP_SCRATCH_W1),y
+		dey
+		pla
+		sta  (P8ZP_SCRATCH_W1),y
+		inc  _left_index
+		lda  _right_index
+		sec
+		sbc  #9
+		sta  _right_index
+		dec  _loop_count
+		bne  _loop
+		rts
+
+		.pend
+
+
+
+a_times_5	.proc
+		sta  P8ZP_SCRATCH_B1
+		asl  a
+		asl  a
+		clc
+		adc  P8ZP_SCRATCH_B1
+		rts
+		.pend
+
+func_any_f_into_A	.proc
+		jsr  a_times_5
+		jmp  prog8_lib.func_any_b_into_A
+		.pend
+
+func_all_f_into_A	.proc
+		jsr  a_times_5
+		jmp  prog8_lib.func_all_b_into_A
+		.pend
+
+func_any_f_stack	.proc
+		jsr  a_times_5
+		jmp  prog8_lib.func_any_b_stack
+		.pend
+
+func_all_f_stack	.proc
+		jsr  a_times_5
+		jmp  prog8_lib.func_all_b_stack
+		.pend
+
+func_max_f_stack	.proc
+		jsr  func_max_f_fac1
+		jmp  push_fac1
+		.pend
+
+func_max_f_fac1	.proc
+		; -- max(array) -> fac1,  array in P8ZP_SCRATCH_W1, num elts in A
+_loop_count = P8ZP_SCRATCH_REG
+		stx  floats_store_reg
+		sta  _loop_count
+		lda  #255
+		sta  _minmax_cmp+1		; modifying
+		lda  #<_largest_neg_float
+		ldy  #>_largest_neg_float
+_minmax_entry	jsr  MOVFM
+-		lda  P8ZP_SCRATCH_W1
+		ldy  P8ZP_SCRATCH_W1+1
+		jsr  FCOMP
+_minmax_cmp	cmp  #255			; modified
+		bne  +
+		lda  P8ZP_SCRATCH_W1
+		ldy  P8ZP_SCRATCH_W1+1
+		jsr  MOVFM
++		lda  P8ZP_SCRATCH_W1
+		clc
+		adc  #5
+		sta  P8ZP_SCRATCH_W1
+		bcc  +
+		inc  P8ZP_SCRATCH_W1+1
++		dec  _loop_count
+		bne  -
+		ldx  floats_store_reg
+		rts
+_largest_neg_float	.byte 255,255,255,255,255		; largest negative float -1.7014118345e+38
+		.pend
+
+func_min_f_stack	.proc
+		jsr  func_min_f_fac1
+		jmp  push_fac1
+		.pend
+
+func_min_f_fac1	.proc
+		; -- min(array) -> fac1,  array in P8ZP_SCRATCH_W1, num elts in A
+		sta  func_max_f_fac1._loop_count
+		lda  #1
+		sta  func_max_f_fac1._minmax_cmp+1
+		lda  #<_largest_pos_float
+		ldy  #>_largest_pos_float
+		jmp  func_max_f_fac1._minmax_entry
+_largest_pos_float	.byte  255,127,255,255,255		; largest positive float
+		rts
+		.pend
+
+
+func_sum_f_stack	.proc
+		jsr  func_sum_f_fac1
+		jmp  push_fac1
+		.pend
+
+func_sum_f_fac1	.proc
+		; -- sum(array) -> fac1,  array in P8ZP_SCRATCH_W1, num elts in A
+_loop_count = P8ZP_SCRATCH_REG
+		stx  floats_store_reg
+		sta  _loop_count
+		lda  #<FL_ZERO_const
+		ldy  #>FL_ZERO_const
+		jsr  MOVFM
+-		lda  P8ZP_SCRATCH_W1
+		ldy  P8ZP_SCRATCH_W1+1
+		jsr  FADD
+		lda  P8ZP_SCRATCH_W1
+		clc
+		adc  #5
+		sta  P8ZP_SCRATCH_W1
+		bcc  +
+		inc  P8ZP_SCRATCH_W1+1
++		dec  _loop_count
+		bne  -
+		ldx  floats_store_reg
+		rts
+		.pend

compiler/res/prog8lib/c64/graphics.p8

@@ -12,11 +12,20 @@ graphics {
 
     sub enable_bitmap_mode() {
         ; enable bitmap screen, erase it and set colors to black/white.
-        c64.SCROLY |= %00100000
+        c64.SCROLY = %00111011
+        c64.SCROLX = %00001000
         c64.VMCSB = (c64.VMCSB & %11110000) | %00001000   ; $2000-$3fff
         clear_screen(1, 0)
     }
 
+    sub disable_bitmap_mode() {
+        ; enables text mode, erase the text screen, color white
+        c64.SCROLY = %00011011
+        c64.SCROLX = %00001000
+        c64.VMCSB = (c64.VMCSB & %11110000) | %00000100   ; $1000-$2fff
+        txt.fill_screen(' ', 1)
+    }
+
     sub clear_screen(ubyte pixelcolor, ubyte bgcolor) {
         memset(BITMAP_ADDRESS, 320*200/8, 0)
         txt.fill_screen(pixelcolor << 4 | bgcolor, 0)
@@ -149,21 +158,17 @@ graphics {
             ubyte ycenter_plus_xx = ycenter + xx
             ubyte ycenter_min_xx = ycenter - xx
 
-            for internal_plotx in xcenter to xcenter+xx {
+            internal_plotx = xcenter-xx
+            repeat xx*2+1 {
                 internal_plot(ycenter_plus_yy)
                 internal_plot(ycenter_min_yy)
+                internal_plotx++
             }
-            for internal_plotx in xcenter-xx to xcenter-1 {
-                internal_plot(ycenter_plus_yy)
-                internal_plot(ycenter_min_yy)
-            }
-            for internal_plotx in xcenter to xcenter+yy {
-                internal_plot(ycenter_plus_xx)
-                internal_plot(ycenter_min_xx)
-            }
-            for internal_plotx in xcenter-yy to xcenter {
+            internal_plotx = xcenter-yy
+            repeat yy*2+1 {
                 internal_plot(ycenter_plus_xx)
                 internal_plot(ycenter_min_xx)
+                internal_plotx++
             }
             yy++
             if decisionOver2<=0

compiler/res/prog8lib/c64/syslib.p8

@@ -269,6 +269,22 @@ asmsub  reset_system()  {
     }}
 }
 
+sub wait(uword jiffies) {
+    uword current_time = 0
+    c64.SETTIM(0,0,0)
+
+    while current_time < jiffies {
+        ; read clock
+        %asm {{
+            stx  P8ZP_SCRATCH_REG
+            jsr  c64.RDTIM
+            sta  current_time
+            stx  current_time+1
+            ldx  P8ZP_SCRATCH_REG
+        }}
+    }
+}
+
 asmsub  disable_runstop_and_charsetswitch() {
     %asm {{
         lda  #$80

compiler/res/prog8lib/c64/textio.p8

@@ -16,7 +16,7 @@ const ubyte DEFAULT_HEIGHT = 25
 
 
 sub  clear_screen() {
-    clear_screenchars(' ')
+    txt.chrout(147)
 }
 
 asmsub  fill_screen (ubyte char @ A, ubyte color @ Y) clobbers(A)  {
@@ -38,13 +38,13 @@ asmsub  clear_screenchars (ubyte char @ A) clobbers(Y)  {
 	; ---- clear the character screen with the given fill character (leaves colors)
 	;      (assumes screen matrix is at the default address)
 	%asm {{
-		ldy  #0
-_loop		sta  c64.Screen,y
-		sta  c64.Screen+$0100,y
-		sta  c64.Screen+$0200,y
-		sta  c64.Screen+$02e8,y
-		iny
-		bne  _loop
+		ldy  #250
+-		sta  c64.Screen+250*0-1,y
+		sta  c64.Screen+250*1-1,y
+		sta  c64.Screen+250*2-1,y
+		sta  c64.Screen+250*3-1,y
+		dey
+		bne  -
 		rts
         }}
 }
@@ -53,13 +53,13 @@ asmsub  clear_screencolors (ubyte color @ A) clobbers(Y)  {
 	; ---- clear the character screen colors with the given color (leaves characters).
 	;      (assumes color matrix is at the default address)
 	%asm {{
-		ldy  #0
-_loop		sta  c64.Colors,y
-		sta  c64.Colors+$0100,y
-		sta  c64.Colors+$0200,y
-		sta  c64.Colors+$02e8,y
-		iny
-		bne  _loop
+		ldy  #250
+-		sta  c64.Colors+250*0-1,y
+		sta  c64.Colors+250*1-1,y
+		sta  c64.Colors+250*2-1,y
+		sta  c64.Colors+250*3-1,y
+		dey
+		bne  -
 		rts
         }}
 }
@@ -83,29 +83,31 @@ asmsub  scroll_left  (ubyte alsocolors @ Pc) clobbers(A, Y)  {
 
 	%asm {{
 		stx  P8ZP_SCRATCH_REG
-		bcs  +
-		jmp  _scroll_screen
+		bcc _scroll_screen
 
-+               ; scroll the color memory
++               ; scroll the screen and the color memory
 		ldx  #0
 		ldy  #38
 -
-	.for row=0, row<=24, row+=1
-		lda  c64.Colors + 40*row + 1,x
-		sta  c64.Colors + 40*row,x
-	.next
+        .for row=0, row<=24, row+=1
+            lda  c64.Screen + 40*row + 1,x
+            sta  c64.Screen + 40*row + 0,x
+            lda  c64.Colors + 40*row + 1,x
+            sta  c64.Colors + 40*row + 0,x
+        .next
 		inx
 		dey
 		bpl  -
+		rts
 
-_scroll_screen  ; scroll the screen memory
+_scroll_screen  ; scroll only the screen memory
 		ldx  #0
 		ldy  #38
 -
-	.for row=0, row<=24, row+=1
-		lda  c64.Screen + 40*row + 1,x
-		sta  c64.Screen + 40*row,x
-	.next
+        .for row=0, row<=24, row+=1
+            lda  c64.Screen + 40*row + 1,x
+            sta  c64.Screen + 40*row + 0,x
+        .next
 		inx
 		dey
 		bpl  -
@@ -121,26 +123,28 @@ asmsub  scroll_right  (ubyte alsocolors @ Pc) clobbers(A)  {
 	;      Carry flag determines if screen color data must be scrolled too
 	%asm {{
 		stx  P8ZP_SCRATCH_REG
-		bcs  +
-		jmp  _scroll_screen
+		bcc  _scroll_screen
 
-+               ; scroll the color memory
++               ; scroll the screen and the color memory
 		ldx  #38
 -
-	.for row=0, row<=24, row+=1
-		lda  c64.Colors + 40*row + 0,x
-		sta  c64.Colors + 40*row + 1,x
-	.next
+        .for row=0, row<=24, row+=1
+            lda  c64.Screen + 40*row + 0,x
+            sta  c64.Screen + 40*row + 1,x
+            lda  c64.Colors + 40*row + 0,x
+            sta  c64.Colors + 40*row + 1,x
+        .next
 		dex
 		bpl  -
+		rts
 
-_scroll_screen  ; scroll the screen memory
+_scroll_screen  ; scroll only the screen memory
 		ldx  #38
 -
-	.for row=0, row<=24, row+=1
-		lda  c64.Screen + 40*row + 0,x
-		sta  c64.Screen + 40*row + 1,x
-	.next
+        .for row=0, row<=24, row+=1
+            lda  c64.Screen + 40*row + 0,x
+            sta  c64.Screen + 40*row + 1,x
+        .next
 		dex
 		bpl  -
 
@@ -155,26 +159,28 @@ asmsub  scroll_up  (ubyte alsocolors @ Pc) clobbers(A)  {
 	;      Carry flag determines if screen color data must be scrolled too
 	%asm {{
 		stx  P8ZP_SCRATCH_REG
-		bcs  +
-		jmp  _scroll_screen
+		bcc  _scroll_screen
 
-+               ; scroll the color memory
++               ; scroll the screen and the color memory
 		ldx #39
 -
-	.for row=1, row<=24, row+=1
-		lda  c64.Colors + 40*row,x
-		sta  c64.Colors + 40*(row-1),x
-	.next
+        .for row=1, row<=24, row+=1
+            lda  c64.Screen + 40*row,x
+            sta  c64.Screen + 40*(row-1),x
+            lda  c64.Colors + 40*row,x
+            sta  c64.Colors + 40*(row-1),x
+        .next
 		dex
 		bpl  -
+		rts
 
-_scroll_screen  ; scroll the screen memory
+_scroll_screen  ; scroll only the screen memory
 		ldx #39
 -
-	.for row=1, row<=24, row+=1
-		lda  c64.Screen + 40*row,x
-		sta  c64.Screen + 40*(row-1),x
-	.next
+        .for row=1, row<=24, row+=1
+            lda  c64.Screen + 40*row,x
+            sta  c64.Screen + 40*(row-1),x
+        .next
 		dex
 		bpl  -
 
@@ -189,26 +195,28 @@ asmsub  scroll_down  (ubyte alsocolors @ Pc) clobbers(A)  {
 	;      Carry flag determines if screen color data must be scrolled too
 	%asm {{
 		stx  P8ZP_SCRATCH_REG
-		bcs  +
-		jmp  _scroll_screen
+		bcc  _scroll_screen
 
-+               ; scroll the color memory
++               ; scroll the screen and the color memory
 		ldx #39
 -
-	.for row=23, row>=0, row-=1
-		lda  c64.Colors + 40*row,x
-		sta  c64.Colors + 40*(row+1),x
-	.next
+        .for row=23, row>=0, row-=1
+            lda  c64.Colors + 40*row,x
+            sta  c64.Colors + 40*(row+1),x
+            lda  c64.Screen + 40*row,x
+            sta  c64.Screen + 40*(row+1),x
+        .next
 		dex
 		bpl  -
+		rts
 
-_scroll_screen  ; scroll the screen memory
+_scroll_screen  ; scroll only the screen memory
 		ldx #39
 -
-	.for row=23, row>=0, row-=1
-		lda  c64.Screen + 40*row,x
-		sta  c64.Screen + 40*(row+1),x
-	.next
+        .for row=23, row>=0, row-=1
+            lda  c64.Screen + 40*row,x
+            sta  c64.Screen + 40*(row+1),x
+        .next
 		dex
 		bpl  -
 

compiler/res/prog8lib/conv.p8

@@ -254,7 +254,6 @@ asmsub  str2ubyte(str string @ AY) clobbers(Y) -> ubyte @A {
 	; -- returns the unsigned byte value of the string number argument in AY
 	;    the number may NOT be preceded by a + sign and may NOT contain spaces
 	;    (any non-digit character will terminate the number string that is parsed)
-	; TODO implement optimized custom version of this instead of simply reusing str2uword
 	%asm {{
     	jmp  str2uword
 	}}
@@ -264,7 +263,6 @@ asmsub  str2byte(str string @ AY) clobbers(Y) -> ubyte @A {
 	; -- returns the signed byte value of the string number argument in AY
 	;    the number may be preceded by a + or - sign but may NOT contain spaces
 	;    (any non-digit character will terminate the number string that is parsed)
-	; TODO implement optimized custom version of this instead of simply reusing str2word
 	%asm {{
     	jmp  str2word
 	}}

compiler/res/prog8lib/cx16/c64colors.p8

@@ -0,0 +1,87 @@
+%target cx16
+
+c64colors {
+    uword[] colorpalette_dark = [   ; this is a darker palette with more contrast
+        $000,  ; 0 = black
+        $FFF,  ; 1 = white
+        $632,  ; 2 = red
+        $7AB,  ; 3 = cyan
+        $638,  ; 4 = purple
+        $584,  ; 5 = green
+        $327,  ; 6 = blue
+        $BC6,  ; 7 = yellow
+        $642,  ; 8 = orange
+        $430,  ; 9 = brown
+        $965,  ; 10 = light red
+        $444,  ; 11 = dark grey
+        $666,  ; 12 = medium grey
+        $9D8,  ; 13 = light green
+        $65B,  ; 14 = light blue
+        $999  ; 15 = light grey
+    ]
+
+    uword[] colorpalette_pepto = [  ; # this is Pepto's Commodore-64 palette  http://www.pepto.de/projects/colorvic/
+        $000,  ; 0 = black
+        $FFF,  ; 1 = white
+        $833,  ; 2 = red
+        $7cc,  ; 3 = cyan
+        $839,  ; 4 = purple
+        $5a4,  ; 5 = green
+        $229,  ; 6 = blue
+        $ef7,  ; 7 = yellow
+        $852,  ; 8 = orange
+        $530,  ; 9 = brown
+        $c67,  ; 10 = light red
+        $444,  ; 11 = dark grey
+        $777,  ; 12 = medium grey
+        $af9,  ; 13 = light green
+        $76e,  ; 14 = light blue
+        $bbb  ; 15 = light grey
+    ]
+
+    uword[] colorpalette_light = [  ; this is a lighter palette
+        $000,  ; 0 = black
+        $FFF,  ; 1 = white
+        $944,  ; 2 = red
+        $7CC,  ; 3 = cyan
+        $95A,  ; 4 = purple
+        $6A5,  ; 5 = green
+        $549,  ; 6 = blue
+        $CD8,  ; 7 = yellow
+        $963,  ; 8 = orange
+        $650,  ; 9 = brown
+        $C77,  ; 10 = light red
+        $666,  ; 11 = dark grey
+        $888,  ; 12 = medium grey
+        $AE9,  ; 13 = light green
+        $87C,  ; 14 = light blue
+        $AAA   ; 15 = light grey
+    ]
+
+    ubyte color
+
+    sub set_palette_pepto() {
+        for color in 0 to 15 {
+            uword cc = colorpalette_pepto[color]
+            cx16.vpoke(1, $fa00 + color*2, lsb(cc))     ; G, B
+            cx16.vpoke(1, $fa01 + color*2, msb(cc))     ; R
+        }
+    }
+
+    sub set_palette_light() {
+        for color in 0 to 15 {
+            uword cc = colorpalette_light[color]
+            cx16.vpoke(1, $fa00 + color*2, lsb(cc))     ; G, B
+            cx16.vpoke(1, $fa01 + color*2, msb(cc))     ; R
+        }
+    }
+
+    sub set_palette_dark() {
+        for color in 0 to 15 {
+            uword cc = colorpalette_dark[color]
+            cx16.vpoke(1, $fa00 + color*2, lsb(cc))     ; G, B
+            cx16.vpoke(1, $fa01 + color*2, msb(cc))     ; R
+        }
+    }
+
+}

compiler/res/prog8lib/cx16/floats.p8

@@ -64,6 +64,7 @@ romsub $fe7e = FINLOG(byte value @A) clobbers (A, X, Y)           ; fac1 += sign
 romsub $fe81 = FOUT() clobbers(X) -> uword @ AY             ; fac1 -> string, address returned in AY
 romsub $fe8a = SQR() clobbers(A,X,Y)                        ; fac1 = SQRT(fac1)
 romsub $fe8d = FPWRT() clobbers(A,X,Y)                      ; fac1 = fac2 ** fac1
+; note: there is no FPWR() on the Cx16
 romsub $fe93 = NEGOP() clobbers(A)                          ; switch the sign of fac1
 romsub $fe96 = EXP() clobbers(A,X,Y)                        ; fac1 = EXP(fac1)  (e ** fac1)
 romsub $fe9f = RND2(byte value @A) clobbers(A,X,Y)                ; fac1 = RND(A) float random number generator
@@ -149,5 +150,6 @@ sub  print_f  (float value) {
 }
 
 %asminclude "library:c64/floats.asm", ""
+%asminclude "library:c64/floats_funcs.asm", ""
 
 }

compiler/res/prog8lib/cx16/graphics.p8

@@ -1,5 +1,6 @@
 %target cx16
 %import syslib
+%import textio
 
 ; bitmap pixel graphics module for the CommanderX16
 ; wraps the graphics functions that are in ROM.
@@ -17,6 +18,13 @@ graphics {
         clear_screen(1, 0)
     }
 
+    sub disable_bitmap_mode() {
+        ; enables text mode, erase the text screen, color white
+        void cx16.screen_set_mode(2)
+        txt.fill_screen(' ', 1)     ; doesn't seem to fully clear the text screen after returning from gfx mode
+    }
+
+
     sub clear_screen(ubyte pixelcolor, ubyte bgcolor) {
         cx16.GRAPH_set_colors(pixelcolor, pixelcolor, bgcolor)
         cx16.GRAPH_clear()
@@ -35,7 +43,7 @@ graphics {
         ;cx16.r1 = ycenter - radius/2
         ;cx16.r2 = radius*2
         ;cx16.r3 = radius*2
-        ;cx16.GRAPH_draw_oval(false)          ; TODO  currently is not implemented on cx16, does a BRK
+        ;cx16.GRAPH_draw_oval(false)          ; currently this call is not implemented on cx16, does a BRK
 
         ; Midpoint algorithm
         ubyte @zp xx = radius
@@ -86,7 +94,7 @@ graphics {
 ;        cx16.r1 = ycenter - radius/2
 ;        cx16.r2 = radius*2
 ;        cx16.r3 = radius*2
-;        cx16.GRAPH_draw_oval(true)          ; TODO  currently is not implemented on cx16, does a BRK
+;        cx16.GRAPH_draw_oval(true)          ; currently this call is not implemented on cx16, does a BRK
 
         ubyte xx = radius
         ubyte yy = 0
@@ -99,42 +107,30 @@ graphics {
             ubyte ycenter_min_xx = ycenter - xx
             uword @zp plotx
 
-            for plotx in xcenter to xcenter+xx {
-                cx16.r0 = plotx
-                cx16.r1 = ycenter_plus_yy
-                cx16.FB_cursor_position()
+            cx16.r0 = xcenter-xx
+            cx16.r1 = ycenter_plus_yy
+            cx16.FB_cursor_position()
+            repeat xx*2+1
                 cx16.FB_set_pixel(1)
-                cx16.r1 = ycenter_min_yy
-                cx16.FB_cursor_position()
+
+            cx16.r0 = xcenter-xx
+            cx16.r1 = ycenter_min_yy
+            cx16.FB_cursor_position()
+            repeat xx*2+1
                 cx16.FB_set_pixel(1)
-            }
-            for plotx in xcenter-xx to xcenter-1 {
-                cx16.r0 = plotx
-                cx16.r1 = ycenter_plus_yy
-                cx16.FB_cursor_position()
+
+            cx16.r0 = xcenter-yy
+            cx16.r1 = ycenter_plus_xx
+            cx16.FB_cursor_position()
+            repeat yy*2+1
                 cx16.FB_set_pixel(1)
-                cx16.r1 = ycenter_min_yy
-                cx16.FB_cursor_position()
+
+            cx16.r0 = xcenter-yy
+            cx16.r1 = ycenter_min_xx
+            cx16.FB_cursor_position()
+            repeat yy*2+1
                 cx16.FB_set_pixel(1)
-            }
-            for plotx in xcenter to xcenter+yy {
-                cx16.r0 = plotx
-                cx16.r1 = ycenter_plus_xx
-                cx16.FB_cursor_position()
-                cx16.FB_set_pixel(1)
-                cx16.r1 = ycenter_min_xx
-                cx16.FB_cursor_position()
-                cx16.FB_set_pixel(1)
-            }
-            for plotx in xcenter-yy to xcenter {
-                cx16.r0 = plotx
-                cx16.r1 = ycenter_plus_xx
-                cx16.FB_cursor_position()
-                cx16.FB_set_pixel(1)
-                cx16.r1 = ycenter_min_xx
-                cx16.FB_cursor_position()
-                cx16.FB_set_pixel(1)
-            }
+
             yy++
             if decisionOver2<=0
                 decisionOver2 += 2*yy+1

compiler/res/prog8lib/cx16/syslib.p8

@@ -220,7 +220,7 @@ romsub $ff08 = FB_get_pixels()  clobbers(A,X,Y)         ; uses ptr=r0, count=r1
 romsub $ff0b = FB_set_pixel(ubyte color @A)  clobbers(A,X,Y)
 romsub $ff0e = FB_set_pixels()  clobbers(A,X,Y)         ; uses ptr=r0, count=r1
 romsub $ff11 = FB_set_8_pixels(ubyte pattern @A, ubyte color @X)  clobbers(A,X,Y)
-romsub $ff14 = FB_set_8_pixels_opaque(ubyte pattern @A, ubyte color1 @X, ubyte color2 @Y)  clobbers(A,X,Y)  ; also uses mask=r0L
+romsub $ff14 = FB_set_8_pixels_opaque(ubyte mask @A, ubyte color1 @X, ubyte color2 @Y)  clobbers(A,X,Y)  ; also uses r0L=pattern
 romsub $ff17 = FB_fill_pixels(ubyte color @A)  clobbers(A,X,Y)   ; also uses count=r0, step=r1
 romsub $ff1a = FB_filter_pixels()  clobbers(A,X,Y)      ; uses ptr=r0, count=r1
 romsub $ff1d = FB_move_pixels()  clobbers(A,X,Y)        ; uses sx=r0, sy=r1, tx=r2, ty=r3, count=r4
@@ -242,9 +242,96 @@ romsub $fecf = entropy_get() -> ubyte @A, ubyte @X, ubyte @Y
 romsub $fecc = monitor()  clobbers(A,X,Y)
 
 
-
 ; ---- end of kernal routines ----
 
+; ---- utilities -----
+asmsub vpeek(ubyte bank @A, uword address @XY) -> ubyte @A {
+        ; -- get a byte from VERA's video memory
+        ;    note: inefficient when reading multiple sequential bytes!
+        %asm {{
+                stz  cx16.VERA_CTRL
+                and  #1
+                sta  cx16.VERA_ADDR_H
+                sty  cx16.VERA_ADDR_M
+                stx  cx16.VERA_ADDR_L
+                lda  cx16.VERA_DATA0
+                rts
+            }}
+}
+
+sub vpoke(ubyte bank, uword address, ubyte value) {
+        ; -- write a single byte to VERA's video memory
+        ;    note: inefficient when writing multiple sequential bytes!
+        %asm {{
+            stz  cx16.VERA_CTRL
+            lda  bank
+            and  #1
+            sta  cx16.VERA_ADDR_H
+            lda  address
+            sta  cx16.VERA_ADDR_L
+            lda  address+1
+            sta  cx16.VERA_ADDR_M
+            lda  value
+            sta  cx16.VERA_DATA0
+            rts
+        }}
+}
+
+sub FB_set_pixels_from_buf(uword buffer, uword count) {
+    %asm {{
+            ; -- This is replacement code for the normal FB_set_pixels subroutine in ROM
+            ;    However that routine contains a bug in the current v38 ROM that makes it crash when count > 255.
+            ;    So the code below replaces that. Once the ROM is patched this routine is no longer necessary.
+            ;    See https://github.com/commanderx16/x16-rom/issues/179
+            phx
+            lda  buffer
+            ldy  buffer+1
+            sta  P8ZP_SCRATCH_W1
+            sty  P8ZP_SCRATCH_W1+1
+            jsr  _pixels
+            plx
+            rts
+
+_pixels     lda  count+1
+            beq  +
+            ldx  #0
+-           jsr  _loop
+            inc  P8ZP_SCRATCH_W1+1
+            dec  count+1
+            bne  -
+
++           ldx  count
+_loop       ldy  #0
+-           lda  (P8ZP_SCRATCH_W1),y
+            sta  cx16.VERA_DATA0
+            iny
+            dex
+            bne  -
+            rts
+        }}
+}
+
+
+
+sub wait(uword jiffies) {
+    uword current_time = 0
+    c64.SETTIM(0,0,0)
+
+    while current_time < jiffies {
+        ; read clock
+        %asm {{
+            phx
+            jsr  c64.RDTIM
+            sta  current_time
+            stx  current_time+1
+            plx
+        }}
+    }
+}
+
+
+; ---- system stuff -----
+
 asmsub init_system()  {
     ; Initializes the machine to a sane starting state.
     ; Called automatically by the loader program logic.

compiler/res/prog8lib/cx16/textio.p8

@@ -16,7 +16,7 @@ const ubyte DEFAULT_HEIGHT = 60
 
 
 sub  clear_screen() {
-    clear_screenchars(' ')
+    txt.chrout(147)
 }
 
 asmsub  fill_screen (ubyte char @ A, ubyte color @ Y) clobbers(A)  {
@@ -154,10 +154,9 @@ sub uppercase() {
     cx16.screen_set_charset(2, 0)  ; uppercase charset
 }
 
-asmsub  scroll_left  (ubyte dummy @ Pc) clobbers(A, Y)  {
+asmsub  scroll_left() clobbers(A, Y)  {
 	; ---- scroll the whole screen 1 character to the left
 	;      contents of the rightmost column are unchanged, you should clear/refill this yourself
-	;      Carry flag is a dummy on the cx16
 	%asm {{
 	    phx
 	    jsr  c64.SCREEN
@@ -198,10 +197,9 @@ _lx     ldx  #0                ; modified
 	}}
 }
 
-asmsub  scroll_right  (ubyte dummy @ Pc) clobbers(A)  {
+asmsub  scroll_right() clobbers(A)  {
 	; ---- scroll the whole screen 1 character to the right
 	;      contents of the leftmost column are unchanged, you should clear/refill this yourself
-	;      Carry flag is a dummy on the cx16
 	%asm {{
 	    phx
 	    jsr  c64.SCREEN
@@ -250,10 +248,9 @@ _lx     ldx  #0                 ; modified
 	}}
 }
 
-asmsub  scroll_up  (ubyte dummy @ Pc) clobbers(A, Y)  {
+asmsub  scroll_up() clobbers(A, Y)  {
 	; ---- scroll the whole screen 1 character up
 	;      contents of the bottom row are unchanged, you should refill/clear this yourself
-	;      Carry flag is a dummy on the cx16
 	%asm {{
 	    phx
 	    jsr  c64.SCREEN
@@ -300,10 +297,9 @@ _nextline
 	}}
 }
 
-asmsub  scroll_down  (ubyte dummy @ Pc) clobbers(A, Y)  {
+asmsub  scroll_down() clobbers(A, Y)  {
 	; ---- scroll the whole screen 1 character down
 	;      contents of the top row are unchanged, you should refill/clear this yourself
-	;      Carry flag is a dummy on the cx16
 	%asm {{
 	    phx
 	    jsr  c64.SCREEN

compiler/res/prog8lib/diskio.p8

@@ -6,15 +6,15 @@
 diskio {
 
 
-    sub directory(ubyte drivenumber) -> byte {
-        ; -- Shows the directory contents of disk drive 8-11 (provide as argument).
+    sub directory(ubyte drivenumber) -> ubyte {
+        ; -- Prints the directory contents of disk drive 8-11 to the screen. Returns success.
 
         c64.SETNAM(1, "$")
-        c64.SETLFS(1, drivenumber, 0)
-        void c64.OPEN()          ; open 1,8,0,"$"
+        c64.SETLFS(13, drivenumber, 0)
+        void c64.OPEN()          ; open 13,8,0,"$"
         if_cs
             goto io_error
-        void c64.CHKIN(1)        ; use #1 as input channel
+        void c64.CHKIN(13)        ; use #13 as input channel
         if_cs
             goto io_error
 
@@ -25,7 +25,9 @@ diskio {
         ; while not key pressed / EOF encountered, read data.
         ubyte status = c64.READST()
         while not status {
-            txt.print_uw(mkword(c64.CHRIN(), c64.CHRIN()))
+            ubyte low = c64.CHRIN()
+            ubyte high = c64.CHRIN()
+            txt.print_uw(mkword(high, low))
             txt.chrout(' ')
             ubyte @zp char
             do {
@@ -37,14 +39,14 @@ diskio {
             void c64.CHRIN()
             status = c64.READST()
             void c64.STOP()
-            if_nz
+            if_z
                 break
         }
 
 io_error:
         status = c64.READST()
         c64.CLRCHN()        ; restore default i/o devices
-        c64.CLOSE(1)
+        c64.CLOSE(13)
 
         if status and status != 64 {            ; 64=end of file
             txt.print("\ni/o error, status: ")
@@ -57,9 +59,208 @@ io_error:
     }
 
 
-    sub status(ubyte drivenumber) {
-        ; -- display the disk drive's current status message
-        c64.SETNAM(0, $0000)
+    ; internal variables for the iterative file lister / loader
+    ubyte list_suffixmatch
+    ubyte list_pattern_size
+    ubyte list_skip_disk_name
+    uword list_pattern
+    uword list_blocks
+    ubyte iteration_in_progress = false
+    str   list_filename = "?" * 32
+
+
+    ; ----- get a list of files (uses iteration functions internally -----
+
+    sub list_files(ubyte drivenumber, uword pattern, ubyte suffixmatch, uword name_ptrs, ubyte max_names) -> ubyte {
+        ; -- fill the array 'name_ptrs' with (pointers to) the names of the files requested.
+        ubyte[256] names_buffer
+        ubyte[256] names_buffer1        ; to store a bit more names
+        uword buf_ptr = &names_buffer
+        ubyte files_found = 0
+        if lf_start_list(drivenumber, pattern, suffixmatch) {
+            while lf_next_entry() {
+                @(name_ptrs) = lsb(buf_ptr)
+                name_ptrs++
+                @(name_ptrs) = msb(buf_ptr)
+                name_ptrs++
+                buf_ptr += strcopy(diskio.list_filename, buf_ptr) + 1
+                files_found++
+                if buf_ptr - &names_buffer > (len(names_buffer) + len(names_buffer1) - 18)
+                    break
+                if files_found == max_names
+                    break
+            }
+            lf_end_list()
+        }
+        return files_found
+    }
+
+    ; ----- iterative file lister functions -----
+
+    sub lf_start_list(ubyte drivenumber, uword pattern, ubyte suffixmatch) -> ubyte {
+        ; -- start an iterative file listing with optional prefix or suffix name matching.
+        ;    note: only a single iteration loop can be active at a time!
+        lf_end_list()
+        list_pattern = pattern
+        list_suffixmatch = suffixmatch
+        list_skip_disk_name = true
+        iteration_in_progress = true
+        if pattern==0
+            list_pattern_size = 0
+        else
+            list_pattern_size = strlen(pattern)
+
+        c64.SETNAM(1, "$")
+        c64.SETLFS(12, drivenumber, 0)
+        void c64.OPEN()          ; open 12,8,0,"$"
+        if_cs
+            goto io_error
+        void c64.CHKIN(12)        ; use #12 as input channel
+        if_cs
+            goto io_error
+
+        repeat 4 {
+            void c64.CHRIN()     ; skip the 4 prologue bytes
+        }
+
+        if c64.READST()==0
+            return true
+
+io_error:
+        lf_end_list()
+        return false
+    }
+
+    sub lf_next_entry() -> ubyte {
+        ; -- retrieve the next entry from an iterative file listing session.
+        ;    results will be found in list_blocks and list_filename.
+        ;    if it returns false though, there are no more entries (or an error occurred).
+
+        if not iteration_in_progress
+            return false
+
+        repeat {
+            void c64.CHKIN(12)        ; use #12 as input channel again
+
+            uword nameptr = &list_filename
+            ubyte blocks_lsb = c64.CHRIN()
+            ubyte blocks_msb = c64.CHRIN()
+
+            if c64.READST()
+                goto close_end
+
+            list_blocks = mkword(blocks_msb, blocks_lsb)
+
+            ; read until the filename starts after the first "
+            while c64.CHRIN()!='\"'  {
+                if c64.READST()
+                    goto close_end
+            }
+
+            ; read the filename
+            repeat {
+                ubyte char = c64.CHRIN()
+                if_z
+                    break
+                if char=='\"'
+                    break
+                @(nameptr) = char
+                nameptr++
+            }
+
+            @(nameptr) = 0
+
+            while c64.CHRIN() {
+                ; read the rest of the entry until the end
+            }
+
+            void c64.CHRIN()     ; skip 2 bytes
+            void c64.CHRIN()
+
+            if not list_skip_disk_name {
+                if list_pattern_size {
+                    ; do filename matching
+                    if list_suffixmatch
+                        rightstr(list_filename, filename, list_pattern_size)
+                    else
+                        leftstr(list_filename, filename, list_pattern_size)
+                    if strcmp(filename, list_pattern)==0
+                        return true
+                } else
+                    return true
+            }
+            list_skip_disk_name = false
+        }
+
+close_end:
+        lf_end_list()
+        return false
+    }
+
+    sub lf_end_list() {
+        ; -- end an iterative file listing session (close channels).
+        if iteration_in_progress {
+            c64.CLRCHN()
+            c64.CLOSE(12)
+            iteration_in_progress = false
+        }
+    }
+
+
+    ; ----- iterative file loader functions -----
+
+    sub f_open(ubyte drivenumber, uword filenameptr) -> ubyte {
+        ; -- open a file for iterative reading with f_read
+        ;    note: only a single iteration loop can be active at a time!
+        f_close()
+
+        c64.SETNAM(strlen(filenameptr), filenameptr)
+        c64.SETLFS(11, drivenumber, 0)
+        void c64.OPEN()          ; open 11,8,0,"filename"
+        if_cc {
+            iteration_in_progress = true
+            void c64.CHKIN(11)        ; use #11 as input channel
+            if_cc
+                return true
+        }
+        f_close()
+        return false
+    }
+
+    sub f_read(uword bufferpointer, uword buffersize) -> uword {
+        if not iteration_in_progress
+            return 0
+
+        uword actual = 0
+        void c64.CHKIN(11)        ; use #11 as input channel again
+        repeat buffersize {
+            ubyte data = c64.CHRIN()
+            @(bufferpointer) = data
+            bufferpointer++
+            actual++
+            ubyte status = c64.READST()
+            if status==64
+                f_close()       ; end of file, close it
+            if status
+                return actual
+        }
+        return actual
+    }
+
+    sub f_close() {
+        ; -- end an iterative file loading session (close channels).
+        if iteration_in_progress {
+            c64.CLRCHN()
+            c64.CLOSE(11)
+            iteration_in_progress = false
+        }
+    }
+
+
+    sub status(ubyte drivenumber) -> uword {
+        ; -- retrieve the disk drive's current status message
+        uword messageptr = &filename
+        c64.SETNAM(0, filename)
         c64.SETLFS(15, drivenumber, 15)
         void c64.OPEN()          ; open 15,8,15
         if_cs
@@ -68,16 +269,20 @@ io_error:
         if_cs
             goto io_error
 
-        while not c64.READST()
-            txt.chrout(c64.CHRIN())
+        while not c64.READST() {
+            @(messageptr) = c64.CHRIN()
+            messageptr++
+        }
 
 io_error:
+        @(messageptr) = 0
         c64.CLRCHN()        ; restore default i/o devices
         c64.CLOSE(15)
+        return filename
     }
 
 
-    sub save(ubyte drivenumber, uword filenameptr, uword address, uword size) -> byte {
+    sub save(ubyte drivenumber, uword filenameptr, uword address, uword size) -> ubyte {
         c64.SETNAM(strlen(filenameptr), filenameptr)
         c64.SETLFS(1, drivenumber, 0)
         uword end_address = address + size
@@ -97,10 +302,14 @@ io_error:
             plp
         }}
 
+        ubyte result=0
         if_cc
-            return c64.READST()==0
+            result = c64.READST()==0
 
-        return false
+        c64.CLRCHN()
+        c64.CLOSE(1)
+
+        return result
     }
 
     sub load(ubyte drivenumber, uword filenameptr, uword address_override) -> uword {
@@ -122,6 +331,9 @@ io_error:
 +           ldx  P8ZP_SCRATCH_REG
         }}
 
+        c64.CLRCHN()
+        c64.CLOSE(1)
+
         if end_of_load
             return end_of_load - address_override
 
@@ -151,8 +363,7 @@ io_error:
         filename[0] = 'r'
         filename[1] = ':'
         memcopy(newfileptr, &filename+2, flen_new)
-        ubyte fis_ix = flen_new+2       ; TODO is temp var for array indexing
-        filename[fis_ix] = '='
+        filename[flen_new+2] = '='
         memcopy(oldfileptr, &filename+3+flen_new, flen_old+1)
         c64.SETNAM(3+flen_new+flen_old, filename)
         c64.SETLFS(1, drivenumber, 15)

compiler/res/prog8lib/math.asm

@@ -774,6 +774,31 @@ stack_mul_word_100	.proc
 		rts
 		.pend
 
+stack_mul_word_320	.proc
+		; stackW = stackLo * 256 + stackLo * 64	 (stackHi doesn't matter)
+		ldy  P8ESTACK_LO+1,x
+		lda  #0
+		sta  P8ESTACK_HI+1,x
+		tya
+		asl  a
+		rol  P8ESTACK_HI+1,x
+		asl  a
+		rol  P8ESTACK_HI+1,x
+		asl  a
+		rol  P8ESTACK_HI+1,x
+		asl  a
+		rol  P8ESTACK_HI+1,x
+		asl  a
+		rol  P8ESTACK_HI+1,x
+		asl  a
+		rol  P8ESTACK_HI+1,x
+		sta  P8ESTACK_LO+1,x
+		tya
+		clc
+		adc  P8ESTACK_HI+1,x
+		sta  P8ESTACK_HI+1,x
+		rts
+		.pend
 
 ; ----------- optimized multiplications (in-place A (byte) and ?? (word)) : ---------
 mul_byte_3	.proc
@@ -1241,48 +1266,35 @@ mul_word_100	.proc
 		rts
 		.pend
 
+mul_word_320	.proc
+		; AY = A * 256 + A * 64	 (msb doesn't matter)
+		sta  P8ZP_SCRATCH_B1
+		ldy  #0
+		sty  P8ZP_SCRATCH_REG
+		asl  a
+		rol  P8ZP_SCRATCH_REG
+		asl  a
+		rol  P8ZP_SCRATCH_REG
+		asl  a
+		rol  P8ZP_SCRATCH_REG
+		asl  a
+		rol  P8ZP_SCRATCH_REG
+		asl  a
+		rol  P8ZP_SCRATCH_REG
+		asl  a
+		rol  P8ZP_SCRATCH_REG
+		pha
+		clc
+		lda  P8ZP_SCRATCH_B1
+		adc  P8ZP_SCRATCH_REG
+		tay
+		pla
+		rts
+		.pend
+
 ; ----------- end optimized multiplications -----------
 
 
-sign_b		.proc
-		lda  P8ESTACK_LO+1,x
-		beq  _sign_zero
-		bmi  _sign_neg
-_sign_pos	lda  #1
-		sta  P8ESTACK_LO+1,x
-		rts
-_sign_neg	lda  #-1
-_sign_zero	sta  P8ESTACK_LO+1,x
-		rts
-		.pend
-
-sign_ub		.proc
-		lda  P8ESTACK_LO+1,x
-		beq  sign_b._sign_zero
-		bne  sign_b._sign_pos
-		.pend
-
-sign_w		.proc
-		lda  P8ESTACK_HI+1,x
-		bmi  sign_b._sign_neg
-		beq  sign_ub
-		bne  sign_b._sign_pos
-		.pend
-
-sign_uw		.proc
-		lda  P8ESTACK_HI+1,x
-		beq  _sign_possibly_zero
-_sign_pos	lda  #1
-		sta  P8ESTACK_LO+1,x
-		rts
-_sign_possibly_zero	lda  P8ESTACK_LO+1,x
-		bne  _sign_pos
-		sta  P8ESTACK_LO+1,x
-		rts
-		.pend
-
-
-
 ; bit shifts.
 ; anything below 3 is done inline. anything above 7 is done via other optimizations.
 
@@ -1340,6 +1352,33 @@ shift_left_w_3	.proc
 		jmp  shift_left_w_7._shift3
 		.pend
 
+
+shift_left_w	.proc
+		; -- variable number of shifts left
+		inx
+		ldy  P8ESTACK_LO,x
+		bne  _shift
+		rts
+_shift		asl  P8ESTACK_LO+1,x
+		rol  P8ESTACK_HI+1,x
+		dey
+		bne  _shift
+		rts
+		.pend
+
+shift_right_uw	.proc
+		; -- uword variable number of shifts right
+		inx
+		ldy  P8ESTACK_LO,x
+		bne  _shift
+		rts
+_shift		lsr  P8ESTACK_HI+1,x
+		ror  P8ESTACK_LO+1,x
+		dey
+		bne  _shift
+		rts
+		.pend
+
 shift_right_uw_7	.proc
 		lda  P8ESTACK_LO+1,x
 		sta  P8ZP_SCRATCH_B1
@@ -1470,6 +1509,21 @@ shift_right_w_3	.proc
 		.pend
 
 
+shift_right_w	.proc
+		; -- signed word variable number of shifts right
+		inx
+		ldy  P8ESTACK_LO,x
+		bne  _shift
+		rts
+_shift		lda  P8ESTACK_HI+1,x
+		asl  a
+		ror  P8ESTACK_HI+1,x
+		ror  P8ESTACK_LO+1,x
+		dey
+		bne  _shift
+		rts
+		.pend
+
 
 ; support for bit shifting that is too large to be unrolled:
 

compiler/res/prog8lib/prog8_lib.p8

@@ -6,4 +6,5 @@
 
 prog8_lib {
 	%asminclude "library:prog8_lib.asm", ""
+	%asminclude "library:prog8_funcs.asm", ""
 }

compiler/res/prog8lib/test_stack.p8

@@ -0,0 +1,48 @@
+%import textio
+
+test_stack {
+
+    asmsub test() {
+        %asm {{
+            stx  _saveX
+            lda  #13
+            jsr  txt.chrout
+            lda  #'-'
+            ldy  #12
+-           jsr  txt.chrout
+            dey
+            bne  -
+            lda  #13
+            jsr  txt.chrout
+            lda  #'x'
+            jsr  txt.chrout
+            lda  #'='
+            jsr  txt.chrout
+            lda  _saveX
+            jsr  txt.print_ub
+            lda  #' '
+            jsr  txt.chrout
+            lda  #'s'
+            jsr  txt.chrout
+            lda  #'p'
+            jsr  txt.chrout
+            lda  #'='
+            jsr  txt.chrout
+            tsx
+            txa
+            jsr  txt.print_ub
+            lda  #13
+            jsr  txt.chrout
+            lda  #'-'
+            ldy  #12
+-           jsr  txt.chrout
+            dey
+            bne  -
+            lda  #13
+            jsr  txt.chrout
+            ldx  _saveX
+            rts
+_saveX   .byte 0
+        }}
+    }
+}

compiler/res/version.txt

@@ -1 +1 @@
-4.6
+5.4

compiler/src/prog8/CompilerMain.kt

@@ -55,32 +55,42 @@ private fun compileMain(args: Array<String>) {
         exitProcess(1)
     }
 
-    if(watchMode && moduleFiles.size<=1) {
+    if(watchMode) {
         val watchservice = FileSystems.getDefault().newWatchService()
 
         while(true) {
-            val filepath = pathFrom(moduleFiles.single()).normalize()
-            println("Continuous watch mode active. Main module: $filepath")
-
-            try {
+            println("Continuous watch mode active. Modules: $moduleFiles")
+            val results = mutableListOf<CompilationResult>()
+            for(filepathRaw in moduleFiles) {
+                val filepath = pathFrom(filepathRaw).normalize()
                 val compilationResult = compileProgram(filepath, !dontOptimize, !dontWriteAssembly, slowCodegenWarnings, compilationTarget, outputPath)
-                println("Imported files (now watching:)")
-                for (importedFile in compilationResult.importedFiles) {
-                    print("  ")
-                    println(importedFile)
-                    importedFile.parent.register(watchservice, StandardWatchEventKinds.ENTRY_MODIFY)
-                }
-                println("[${LocalDateTime.now().withNano(0)}]  Waiting for file changes.")
+                results.add(compilationResult)
+            }
+
+            val allImportedFiles = results.flatMap { it.importedFiles }
+            println("Imported files (now watching:)")
+            for (importedFile in allImportedFiles) {
+                print("  ")
+                println(importedFile)
+                val watchDir = importedFile.parent ?: Path.of(".")
+                watchDir.register(watchservice, StandardWatchEventKinds.ENTRY_MODIFY)
+            }
+            println("[${LocalDateTime.now().withNano(0)}]  Waiting for file changes.")
+
+            var recompile=false
+            while(!recompile) {
                 val event = watchservice.take()
-                for(changed in event.pollEvents()) {
+                for (changed in event.pollEvents()) {
                     val changedPath = changed.context() as Path
-                    println("  change detected: $changedPath")
+                    if(allImportedFiles.any { it.fileName == changedPath.fileName }) {
+                        println("  change detected: $changedPath")
+                        recompile = true
+                    }
                 }
                 event.reset()
-                println("\u001b[H\u001b[2J")      // clear the screen
-            } catch (x: Exception) {
-                throw x
             }
+
+            println("\u001b[H\u001b[2J")      // clear the screen
         }
 
     } else {

compiler/src/prog8/ast/AstToSourceCode.kt

@@ -143,7 +143,6 @@ class AstToSourceCode(val output: (text: String) -> Unit, val program: Program):
             for(param in subroutine.parameters.zip(subroutine.asmParameterRegisters)) {
                 val reg =
                         when {
-                            param.second.stack -> "stack"
                             param.second.registerOrPair!=null -> param.second.registerOrPair.toString()
                             param.second.statusflag!=null -> param.second.statusflag.toString()
                             else -> "?????"

compiler/src/prog8/ast/AstToplevel.kt

@@ -57,19 +57,11 @@ interface IFunctionCall {
 }
 
 
-class AsmGenInfo {
-    var usedAutoArrayIndexerForStatements = mutableMapOf<String, MutableSet<Statement>>()
-    var usedRegsaveA = false
-    var usedRegsaveX = false
-    var usedRegsaveY = false
-}
-
 interface INameScope {
     val name: String
     val position: Position
     val statements: MutableList<Statement>
     val parent: Node
-    val asmGenInfo: AsmGenInfo
 
     fun linkParents(parent: Node)
 
@@ -224,6 +216,14 @@ interface INameScope {
             null
     }
 
+    fun previousSibling(stmt: Statement): Statement? {
+        val previousIdx = statements.indexOfFirst { it===stmt } - 1
+        return if(previousIdx>=0)
+            statements[previousIdx]
+        else
+            null
+    }
+
     fun indexOfChild(stmt: Statement): Int {
         val idx = statements.indexOfFirst { it===stmt }
         if(idx>=0)
@@ -288,7 +288,6 @@ class Module(override val name: String,
 
     override lateinit var parent: Node
     lateinit var program: Program
-    override val asmGenInfo = AsmGenInfo()
     val importedBy = mutableListOf<Module>()
     val imports = mutableSetOf<Module>()
 
@@ -322,7 +321,6 @@ class GlobalNamespace(val modules: List<Module>): Node, INameScope {
     override val position = Position("<<<global>>>", 0, 0, 0)
     override val statements = mutableListOf<Statement>()        // not used
     override var parent: Node = ParentSentinel
-    override val asmGenInfo = AsmGenInfo()
 
     override fun linkParents(parent: Node) {
         modules.forEach { it.linkParents(this) }
@@ -375,7 +373,6 @@ object BuiltinFunctionScopePlaceholder : INameScope {
     override val position = Position("<<placeholder>>", 0, 0, 0)
     override var statements = mutableListOf<Statement>()
     override var parent: Node = ParentSentinel
-    override val asmGenInfo = AsmGenInfo()
     override fun linkParents(parent: Node) {}
 }
 

compiler/src/prog8/ast/antlr/Antr2Kotlin.kt

@@ -254,8 +254,8 @@ private fun prog8Parser.Asmsub_declContext.toAst(): AsmsubDecl {
     val clobbers = asmsub_clobbers()?.clobber()?.toAst() ?: emptySet()
     val normalParameters = params.map { SubroutineParameter(it.name, it.type, it.position) }
     val normalReturntypes = returns.map { it.type }
-    val paramRegisters = params.map { RegisterOrStatusflag(it.registerOrPair, it.statusflag, false) }
-    val returnRegisters = returns.map { RegisterOrStatusflag(it.registerOrPair, it.statusflag, it.stack) }
+    val paramRegisters = params.map { RegisterOrStatusflag(it.registerOrPair, it.statusflag) }
+    val returnRegisters = returns.map { RegisterOrStatusflag(it.registerOrPair, it.statusflag) }
     return AsmsubDecl(name, normalParameters, normalReturntypes, paramRegisters, returnRegisters, clobbers)
 }
 
@@ -263,13 +263,11 @@ private class AsmSubroutineParameter(name: String,
                                      type: DataType,
                                      val registerOrPair: RegisterOrPair?,
                                      val statusflag: Statusflag?,
-                                     // TODO implement: val stack: Boolean,
                                      position: Position) : SubroutineParameter(name, type, position)
 
 private class AsmSubroutineReturn(val type: DataType,
                                   val registerOrPair: RegisterOrPair?,
                                   val statusflag: Statusflag?,
-                                  val stack: Boolean,
                                   val position: Position)
 
 private fun prog8Parser.Asmsub_returnsContext.toAst(): List<AsmSubroutineReturn>
@@ -288,7 +286,7 @@ private fun prog8Parser.Asmsub_returnsContext.toAst(): List<AsmSubroutineReturn>
                     it.datatype().toAst(),
                     registerorpair,
                     statusregister,
-                    !it.stack?.text.isNullOrEmpty(), toPosition())
+                    toPosition())
         }
 
 private fun prog8Parser.Asmsub_paramsContext.toAst(): List<AsmSubroutineParameter>
@@ -342,14 +340,11 @@ private fun prog8Parser.LabeldefContext.toAst(): Statement =
         Label(children[0].text, toPosition())
 
 private fun prog8Parser.SubroutineContext.toAst() : Subroutine {
+    // non-asm subroutine
+    val returntypes = sub_return_part()?.toAst() ?: emptyList()
     return Subroutine(identifier().text,
             sub_params()?.toAst() ?: emptyList(),
-            sub_return_part()?.toAst() ?: emptyList(),
-            emptyList(),
-            emptyList(),
-            emptySet(),
-            null,
-            false,
+            returntypes,
             statement_block()?.toAst() ?: mutableListOf(),
             toPosition())
 }

compiler/src/prog8/ast/base/Base.kt

@@ -35,8 +35,9 @@ enum class DataType {
                 else -> false
             }
 
-
     infix fun isAssignableTo(targetTypes: Set<DataType>) = targetTypes.any { this isAssignableTo it }
+    infix fun isNotAssignableTo(targetType: DataType) = !this.isAssignableTo(targetType)
+    infix fun isNotAssignableTo(targetTypes: Set<DataType>) = !this.isAssignableTo(targetTypes)
 
     infix fun largerThan(other: DataType) =
             when {
@@ -79,7 +80,9 @@ enum class RegisterOrPair {
     Y,
     AX,
     AY,
-    XY;
+    XY,
+    FAC1,
+    FAC2;
 
     companion object {
         val names by lazy { values().map { it.toString()} }

compiler/src/prog8/ast/expressions/AstExpressions.kt

@@ -17,7 +17,7 @@ import kotlin.math.abs
 
 val associativeOperators = setOf("+", "*", "&", "|", "^", "or", "and", "xor", "==", "!=")
 val comparisonOperators = setOf("==", "!=", "<", ">", "<=", ">=")
-val augmentAssignmentOperators = setOf("+", "-", "/", "*", "**", "&", "|", "^", "<<", ">>")
+val augmentAssignmentOperators = setOf("+", "-", "/", "*", "**", "&", "|", "^", "<<", ">>", "%", "and", "or", "xor")
 
 sealed class Expression: Node {
     abstract fun constValue(program: Program): NumericLiteralValue?
@@ -779,6 +779,9 @@ data class IdentifierReference(val nameInSource: List<String>, override val posi
         // if it's just a regular variable, return null.
         val struct = memberOfStruct(namespace) ?: return null
         val decl = targetVarDecl(namespace)!!
+        if(decl.datatype!=DataType.STRUCT)
+            return null
+
         val firstStructMember = struct.nameOfFirstMember()
         // find the flattened var that belongs to this first struct member
         val firstVarName = listOf(decl.name, firstStructMember)
@@ -821,7 +824,7 @@ class FunctionCall(override var target: IdentifierReference,
                 val exprfunc = func.constExpressionFunc
                 if(exprfunc!=null)
                     resultValue = exprfunc(args, position, program)
-                else if(func.returntype==null)
+                else if(func.known_returntype==null)
                     throw ExpressionError("builtin function ${target.nameInSource[0]} can't be used here because it doesn't return a value", position)
             }
 

compiler/src/prog8/ast/expressions/InferredTypes.kt

@@ -35,6 +35,13 @@ object InferredTypes {
         }
 
         override fun hashCode(): Int = Objects.hash(isVoid, datatype)
+
+        infix fun isAssignableTo(targetDt: InferredType): Boolean =
+                isKnown && targetDt.isKnown && (datatype!! isAssignableTo targetDt.datatype!!)
+        infix fun isAssignableTo(targetDt: DataType): Boolean =
+                isKnown && (datatype!! isAssignableTo targetDt)
+        infix fun isNotAssignableTo(targetDt: InferredType): Boolean = !this.isAssignableTo(targetDt)
+        infix fun isNotAssignableTo(targetDt: DataType): Boolean = !this.isAssignableTo(targetDt)
     }
 
     private val unknownInstance = InferredType.unknown()

compiler/src/prog8/ast/processing/AstChecker.kt

@@ -250,19 +250,19 @@ internal class AstChecker(private val program: Program,
                         err("parameter '${param.first.name}' should be ubyte")
                 }
             }
-            for(ret in subroutine.returntypes.withIndex().zip(subroutine.asmReturnvaluesRegisters)) {
-                if(ret.second.registerOrPair in setOf(RegisterOrPair.A, RegisterOrPair.X, RegisterOrPair.Y)) {
-                    if (ret.first.value != DataType.UBYTE && ret.first.value != DataType.BYTE)
-                        err("return value #${ret.first.index + 1} should be (u)byte")
+            subroutine.returntypes.zip(subroutine.asmReturnvaluesRegisters).forEachIndexed { index, pair ->
+                if(pair.second.registerOrPair in setOf(RegisterOrPair.A, RegisterOrPair.X, RegisterOrPair.Y)) {
+                    if (pair.first != DataType.UBYTE && pair.first != DataType.BYTE)
+                        err("return value #${index + 1} should be (u)byte")
                 }
-                else if(ret.second.registerOrPair in setOf(RegisterOrPair.AX, RegisterOrPair.AY, RegisterOrPair.XY)) {
-                    if (ret.first.value != DataType.UWORD && ret.first.value != DataType.WORD
-                            && ret.first.value != DataType.STR && ret.first.value !in ArrayDatatypes && ret.first.value != DataType.FLOAT)
-                        err("return value #${ret.first.index + 1} should be (u)word/address")
+                else if(pair.second.registerOrPair in setOf(RegisterOrPair.AX, RegisterOrPair.AY, RegisterOrPair.XY)) {
+                    if (pair.first != DataType.UWORD && pair.first != DataType.WORD
+                            && pair.first != DataType.STR && pair.first !in ArrayDatatypes && pair.first != DataType.FLOAT)
+                        err("return value #${index + 1} should be (u)word/address")
                 }
-                else if(ret.second.statusflag!=null) {
-                    if (ret.first.value != DataType.UBYTE)
-                        err("return value #${ret.first.index + 1} should be ubyte")
+                else if(pair.second.statusflag!=null) {
+                    if (pair.first != DataType.UBYTE)
+                        err("return value #${index + 1} should be ubyte")
                 }
             }
 
@@ -288,6 +288,7 @@ internal class AstChecker(private val program: Program,
                             regCounts[CpuRegister.X]=regCounts.getValue(CpuRegister.X)+1
                             regCounts[CpuRegister.Y]=regCounts.getValue(CpuRegister.Y)+1
                         }
+                        RegisterOrPair.FAC1, RegisterOrPair.FAC2 -> { /* no sensible way to count this */ }
                         null ->
                             if(p.statusflag!=null)
                                 statusflagCounts[p.statusflag] = statusflagCounts.getValue(p.statusflag) + 1
@@ -344,11 +345,11 @@ internal class AstChecker(private val program: Program,
         if(assignment.value is FunctionCall) {
             val stmt = (assignment.value as FunctionCall).target.targetStatement(program.namespace)
             if (stmt is Subroutine) {
-                val idt = assignment.target.inferType(program, assignment)
+                val idt = assignment.target.inferType(program)
                 if(!idt.isKnown) {
                      errors.err("return type mismatch", assignment.value.position)
                 }
-                if(stmt.returntypes.size <= 1 && stmt.returntypes.single()!=idt.typeOrElse(DataType.BYTE)) {
+                if(stmt.returntypes.size <= 1 && stmt.returntypes.single() isNotAssignableTo idt.typeOrElse(DataType.BYTE)) {
                     errors.err("return type mismatch", assignment.value.position)
                 }
             }
@@ -378,12 +379,12 @@ internal class AstChecker(private val program: Program,
             }
         }
 
-        val targetDt = assignment.target.inferType(program, assignment)
+        val targetDt = assignment.target.inferType(program)
         val valueDt = assignment.value.inferType(program)
-        if(valueDt.isKnown && valueDt != targetDt) {
+        if(valueDt.isKnown && !(valueDt isAssignableTo targetDt)) {
             if(targetDt.typeOrElse(DataType.STRUCT) in IterableDatatypes)
                 errors.err("cannot assign value to string or array", assignment.value.position)
-            else
+            else if(!(valueDt.istype(DataType.STR) && targetDt.istype(DataType.UWORD)))
                 errors.err("value's type doesn't match target", assignment.value.position)
         }
 
@@ -433,7 +434,7 @@ internal class AstChecker(private val program: Program,
 
         if (assignment is Assignment) {
 
-            val targetDatatype = assignTarget.inferType(program, assignment)
+            val targetDatatype = assignTarget.inferType(program)
             if (targetDatatype.isKnown) {
                 val constVal = assignment.value.constValue(program)
                 if (constVal != null) {
@@ -454,15 +455,12 @@ internal class AstChecker(private val program: Program,
 
     override fun visit(addressOf: AddressOf) {
         val variable=addressOf.identifier.targetVarDecl(program.namespace)
-        if(variable==null)
-            errors.err("pointer-of operand must be the name of a heap variable", addressOf.position)
-        else {
-            if(variable.datatype !in ArrayDatatypes
-                    && variable.type!=VarDeclType.MEMORY
-                    && variable.struct == null
-                    && variable.datatype != DataType.STR && variable.datatype!=DataType.STRUCT)
+        if(variable!=null
+            && variable.datatype !in ArrayDatatypes
+            && variable.type!=VarDeclType.MEMORY
+            && variable.struct == null
+            && variable.datatype != DataType.STR && variable.datatype!=DataType.STRUCT)
                 errors.err("invalid pointer-of operand type", addressOf.position)
-        }
         super.visit(addressOf)
     }
 
@@ -593,8 +591,8 @@ internal class AstChecker(private val program: Program,
         if(declValue!=null && decl.type==VarDeclType.VAR) {
             if(decl.datatype==DataType.STRUCT) {
                 val valueIdt = declValue.inferType(program)
-                if(valueIdt.isUnknown)
-                    throw AstException("invalid value type")
+                if(!valueIdt.isKnown)
+                    throw AstException("unknown dt")
                 val valueDt = valueIdt.typeOrElse(DataType.STRUCT)
                 if(valueDt !in ArrayDatatypes)
                     err("initialisation of struct should be with array value", declValue.position)
@@ -603,23 +601,28 @@ internal class AstChecker(private val program: Program,
             }
         }
 
-        // array length limits
+        // array length limits and constant lenghts
         if(decl.isArray) {
-            val length = decl.arraysize!!.constIndex() ?: 1
-            when (decl.datatype) {
-                DataType.STR, DataType.ARRAY_UB, DataType.ARRAY_B -> {
-                    if(length==0 || length>256)
-                        err("string and byte array length must be 1-256")
+            val length = decl.arraysize!!.constIndex()
+            if(length==null)
+                err("array length must be a constant")
+            else {
+                when (decl.datatype) {
+                    DataType.STR, DataType.ARRAY_UB, DataType.ARRAY_B -> {
+                        if (length == 0 || length > 256)
+                            err("string and byte array length must be 1-256")
+                    }
+                    DataType.ARRAY_UW, DataType.ARRAY_W -> {
+                        if (length == 0 || length > 128)
+                            err("word array length must be 1-128")
+                    }
+                    DataType.ARRAY_F -> {
+                        if (length == 0 || length > 51)
+                            err("float array length must be 1-51")
+                    }
+                    else -> {
+                    }
                 }
-                DataType.ARRAY_UW, DataType.ARRAY_W -> {
-                    if(length==0 || length>128)
-                        err("word array length must be 1-128")
-                }
-                DataType.ARRAY_F -> {
-                    if(length==0 || length>51)
-                        err("float array length must be 1-51")
-                }
-                else -> {}
             }
         }
 
@@ -769,7 +772,11 @@ internal class AstChecker(private val program: Program,
     }
 
     override fun visit(expr: PrefixExpression) {
-        val dt = expr.inferType(program).typeOrElse(DataType.STRUCT)
+        val idt = expr.inferType(program)
+        if(!idt.isKnown)
+            return  // any error should be reported elsewhere
+
+        val dt = idt.typeOrElse(DataType.STRUCT)
         if(expr.operator=="-") {
             if (dt != DataType.BYTE && dt != DataType.WORD && dt != DataType.FLOAT) {
                 errors.err("can only take negative of a signed number type", expr.position)
@@ -984,8 +991,8 @@ internal class AstChecker(private val program: Program,
             if(target.regXasResult())
                 errors.warn("subroutine call return value in X register is discarded and replaced by 0", position)
             if(target.isAsmSubroutine) {
-                for (arg in args.withIndex().zip(target.parameters)) {
-                    val argIDt = arg.first.value.inferType(program)
+                for (arg in args.zip(target.parameters)) {
+                    val argIDt = arg.first.inferType(program)
                     if (!argIDt.isKnown)
                         return
                 }
@@ -1218,7 +1225,7 @@ internal class AstChecker(private val program: Program,
                     for(elt in value.value.zip(struct.statements)) {
                         val vardecl = elt.second as VarDecl
                         val valuetype = elt.first.inferType(program)
-                        if (!valuetype.isKnown || !(valuetype.typeOrElse(DataType.STRUCT) isAssignableTo vardecl.datatype)) {
+                        if (!valuetype.isKnown || valuetype isNotAssignableTo vardecl.datatype) {
                             errors.err("invalid struct member init value type $valuetype, expected ${vardecl.datatype}", elt.first.position)
                             return false
                         }
@@ -1350,9 +1357,7 @@ internal class AstChecker(private val program: Program,
         else if(sourceDatatype== DataType.FLOAT && targetDatatype in IntegerDatatypes)
             errors.err("cannot assign float to ${targetDatatype.name.toLowerCase()}; possible loss of precision. Suggestion: round the value or revert to integer arithmetic", position)
         else {
-            if(targetDatatype==DataType.UWORD && sourceDatatype in PassByReferenceDatatypes)
-                errors.err("cannot assign ${sourceDatatype.name.toLowerCase()} to ${targetDatatype.name.toLowerCase()}, perhaps forgot '&' ?", position)
-            else
+            if(targetDatatype!=DataType.UWORD && sourceDatatype !in PassByReferenceDatatypes)
                 errors.err("cannot assign ${sourceDatatype.name.toLowerCase()} to ${targetDatatype.name.toLowerCase()}", position)
         }
 

compiler/src/prog8/ast/processing/StatementReorderer.kt

@@ -4,6 +4,7 @@ import prog8.ast.*
 import prog8.ast.base.*
 import prog8.ast.expressions.*
 import prog8.ast.statements.*
+import prog8.functions.BuiltinFunctions
 
 
 internal class StatementReorderer(val program: Program, val errors: ErrorReporter) : AstWalker() {
@@ -91,41 +92,84 @@ internal class StatementReorderer(val program: Program, val errors: ErrorReporte
         }
     }
 
+    override fun after(expr: BinaryExpression, parent: Node): Iterable<IAstModification> {
+        // when using a simple bit shift and assigning it to a variable of a different type,
+        // try to make the bit shifting 'wide enough' to fall into the variable's type.
+        // with this, for instance, uword x = 1 << 10  will result in 1024 rather than 0 (the ubyte result).
+        if(expr.operator=="<<" || expr.operator==">>") {
+            val leftDt = expr.left.inferType(program)
+            when (parent) {
+                is Assignment -> {
+                    val targetDt = parent.target.inferType(program)
+                    if(leftDt != targetDt) {
+                        val cast = TypecastExpression(expr.left, targetDt.typeOrElse(DataType.STRUCT), true, parent.position)
+                        return listOf(IAstModification.ReplaceNode(expr.left, cast, expr))
+                    }
+                }
+                is VarDecl -> {
+                    if(!leftDt.istype(parent.datatype)) {
+                        val cast = TypecastExpression(expr.left, parent.datatype, true, parent.position)
+                        return listOf(IAstModification.ReplaceNode(expr.left, cast, expr))
+                    }
+                }
+                is IFunctionCall -> {
+                    val argnum = parent.args.indexOf(expr)
+                    when (val callee = parent.target.targetStatement(program.namespace)) {
+                        is Subroutine -> {
+                            val paramType = callee.parameters[argnum].type
+                            if(leftDt isAssignableTo paramType) {
+                                val cast = TypecastExpression(expr.left, paramType, true, parent.position)
+                                return listOf(IAstModification.ReplaceNode(expr.left, cast, expr))
+                            }
+                        }
+                        is BuiltinFunctionStatementPlaceholder -> {
+                            val func = BuiltinFunctions.getValue(callee.name)
+                            val paramTypes = func.parameters[argnum].possibleDatatypes
+                            for(type in paramTypes) {
+                                if(leftDt isAssignableTo type) {
+                                    val cast = TypecastExpression(expr.left, type, true, parent.position)
+                                    return listOf(IAstModification.ReplaceNode(expr.left, cast, expr))
+                                }
+                            }
+                        }
+                        else -> throw FatalAstException("weird callee")
+                    }
+                }
+                else -> return noModifications
+            }
+        }
+        return noModifications
+    }
+
     private fun getAutoIndexerVarFor(expr: ArrayIndexedExpression): MutableList<IAstModification> {
         val modifications = mutableListOf<IAstModification>()
         val subroutine = expr.definingSubroutine()!!
         val statement = expr.containingStatement()
         val indexerVarPrefix = "prog8_autovar_index_"
-        val indexerVarName: String
         val repo = subroutine.asmGenInfo.usedAutoArrayIndexerForStatements
 
-        val freeVar = repo.filter { statement !in it.value }.map { it.key }.firstOrNull()
-        if(freeVar==null) {
+        // TODO make this even smarter so that an indexerVar can be reused for a different following statement...  requires updating the partOfStatement?
+        var indexerVar = repo.firstOrNull { it.replaces isSameAs expr.indexer }
+        if(indexerVar==null) {
             // add another loop index var to be used for this expression
-            val statementId = expr.hashCode()
-            indexerVarName = "$indexerVarPrefix$statementId"
+            val indexerVarName = "$indexerVarPrefix${expr.indexer.hashCode()}"
+            indexerVar = AsmGenInfo.ArrayIndexerInfo(indexerVarName, expr.indexer, statement)
+            repo.add(indexerVar)
             // create the indexer var at block level scope
             val vardecl = VarDecl(VarDeclType.VAR, DataType.UBYTE, ZeropageWish.PREFER_ZEROPAGE,
                     null, indexerVarName, null, null, isArray = false, autogeneratedDontRemove = true, position = expr.position)
             modifications.add(IAstModification.InsertFirst(vardecl, subroutine))
-            var statements = repo[indexerVarName]
-            if(statements==null) {
-                statements = mutableSetOf()
-                repo[indexerVarName] = statements
-            }
-            statements.add(statement)
-        } else {
-            // reuse an already created indexer autovar
-            repo.getValue(freeVar).add(statement)
-            indexerVarName = freeVar
         }
+        indexerVar.used++       // keep track of how many times it it used, to avoid assigning it multiple times
 
-        // assign the indexing expression to the helper variable, and replace the indexer with just the variable
+        // replace the indexer with just the variable
+        // assign the indexing expression to the helper variable, but only if that hasn't been done already
         val indexerExpression = expr.indexer.origExpression!!
-        val target = AssignTarget(IdentifierReference(listOf(indexerVarName), indexerExpression.position), null, null, indexerExpression.position)
-        val assign = Assignment(target, indexerExpression, indexerExpression.position)
-        val beforeWhat = expr.containingStatement()
-        modifications.add(IAstModification.InsertBefore(beforeWhat, assign, beforeWhat.definingScope()))
+        val target = AssignTarget(IdentifierReference(listOf(indexerVar.name), indexerExpression.position), null, null, indexerExpression.position)
+        if(indexerVar.used==1) {
+            val assign = Assignment(target, indexerExpression, indexerExpression.position)
+            modifications.add(IAstModification.InsertBefore(statement, assign, statement.definingScope()))
+        }
         modifications.add(IAstModification.SetExpression( {
             expr.indexer.indexVar = it as IdentifierReference
             expr.indexer.indexNum = null
@@ -168,7 +212,7 @@ internal class StatementReorderer(val program: Program, val errors: ErrorReporte
 
     override fun before(assignment: Assignment, parent: Node): Iterable<IAstModification> {
         val valueType = assignment.value.inferType(program)
-        val targetType = assignment.target.inferType(program, assignment)
+        val targetType = assignment.target.inferType(program)
         var assignments = emptyList<Assignment>()
 
         if(targetType.istype(DataType.STRUCT) && (valueType.istype(DataType.STRUCT) || valueType.typeOrElse(DataType.STRUCT) in ArrayDatatypes )) {
@@ -279,7 +323,7 @@ internal class StatementReorderer(val program: Program, val errors: ErrorReporte
         }
 
         // TODO use a pointer loop instead of individual assignments
-        return alv.value.withIndex().map { (index, value)->
+        return alv.value.mapIndexed { index, value ->
             val idx = ArrayIndexedExpression(identifier, ArrayIndex(NumericLiteralValue(DataType.UBYTE, index, position), position), position)
             Assignment(AssignTarget(null, idx, null, position), value, value.position)
         }

compiler/src/prog8/ast/processing/TypecastsAdder.kt

@@ -60,12 +60,15 @@ class TypecastsAdder(val program: Program, val errors: ErrorReporter) : AstWalke
     override fun after(assignment: Assignment, parent: Node): Iterable<IAstModification> {
         // see if a typecast is needed to convert the value's type into the proper target type
         val valueItype = assignment.value.inferType(program)
-        val targetItype = assignment.target.inferType(program, assignment)
+        val targetItype = assignment.target.inferType(program)
         if(targetItype.isKnown && valueItype.isKnown) {
             val targettype = targetItype.typeOrElse(DataType.STRUCT)
             val valuetype = valueItype.typeOrElse(DataType.STRUCT)
             if (valuetype != targettype) {
                 if (valuetype isAssignableTo targettype) {
+                    if(valuetype in IterableDatatypes && targettype==DataType.UWORD)
+                        // special case, don't typecast STR/arrays to UWORD, we support those assignments "directly"
+                        return noModifications
                     return listOf(IAstModification.ReplaceNode(
                             assignment.value,
                             TypecastExpression(assignment.value, targettype, true, assignment.value.position),
@@ -116,30 +119,30 @@ class TypecastsAdder(val program: Program, val errors: ErrorReporter) : AstWalke
 
         when(val sub = call.target.targetStatement(scope)) {
             is Subroutine -> {
-                for(arg in sub.parameters.zip(call.args.withIndex())) {
-                    val argItype = arg.second.value.inferType(program)
+                sub.parameters.zip(call.args).forEachIndexed { index, pair ->
+                    val argItype = pair.second.inferType(program)
                     if(argItype.isKnown) {
                         val argtype = argItype.typeOrElse(DataType.STRUCT)
-                        val requiredType = arg.first.type
+                        val requiredType = pair.first.type
                         if (requiredType != argtype) {
                             if (argtype isAssignableTo requiredType) {
                                 modifications += IAstModification.ReplaceNode(
-                                        call.args[arg.second.index],
-                                        TypecastExpression(arg.second.value, requiredType, true, arg.second.value.position),
+                                        call.args[index],
+                                        TypecastExpression(pair.second, requiredType, true, pair.second.position),
                                         call as Node)
                             } else if(requiredType == DataType.UWORD && argtype in PassByReferenceDatatypes) {
                                 // we allow STR/ARRAY values in place of UWORD parameters. Take their address instead.
-                                if(arg.second.value is IdentifierReference) {
+                                if(pair.second is IdentifierReference) {
                                     modifications += IAstModification.ReplaceNode(
-                                            call.args[arg.second.index],
-                                            AddressOf(arg.second.value as IdentifierReference, arg.second.value.position),
+                                            call.args[index],
+                                            AddressOf(pair.second as IdentifierReference, pair.second.position),
                                             call as Node)
                                 }
-                            } else if(arg.second.value is NumericLiteralValue) {
-                                val cast = (arg.second.value as NumericLiteralValue).cast(requiredType)
+                            } else if(pair.second is NumericLiteralValue) {
+                                val cast = (pair.second as NumericLiteralValue).cast(requiredType)
                                 if(cast.isValid)
                                     modifications += IAstModification.ReplaceNode(
-                                            call.args[arg.second.index],
+                                            call.args[index],
                                             cast.valueOrZero(),
                                             call as Node)
                             }
@@ -149,19 +152,19 @@ class TypecastsAdder(val program: Program, val errors: ErrorReporter) : AstWalke
             }
             is BuiltinFunctionStatementPlaceholder -> {
                 val func = BuiltinFunctions.getValue(sub.name)
-                for (arg in func.parameters.zip(call.args.withIndex())) {
-                    val argItype = arg.second.value.inferType(program)
+                func.parameters.zip(call.args).forEachIndexed { index, pair ->
+                    val argItype = pair.second.inferType(program)
                     if (argItype.isKnown) {
                         val argtype = argItype.typeOrElse(DataType.STRUCT)
-                        if (arg.first.possibleDatatypes.any { argtype == it })
-                            continue
-                        for (possibleType in arg.first.possibleDatatypes) {
-                            if (argtype isAssignableTo possibleType) {
-                                modifications += IAstModification.ReplaceNode(
-                                        call.args[arg.second.index],
-                                        TypecastExpression(arg.second.value, possibleType, true, arg.second.value.position),
-                                        call as Node)
-                                break
+                        if (pair.first.possibleDatatypes.all { argtype != it }) {
+                            for (possibleType in pair.first.possibleDatatypes) {
+                                if (argtype isAssignableTo possibleType) {
+                                    modifications += IAstModification.ReplaceNode(
+                                            call.args[index],
+                                            TypecastExpression(pair.second, possibleType, true, pair.second.position),
+                                            call as Node)
+                                    break
+                                }
                             }
                         }
                     }

compiler/src/prog8/ast/processing/VerifyFunctionArgTypes.kt

@@ -4,6 +4,7 @@ import prog8.ast.IFunctionCall
 import prog8.ast.INameScope
 import prog8.ast.Program
 import prog8.ast.base.DataType
+import prog8.ast.base.FatalAstException
 import prog8.ast.expressions.Expression
 import prog8.ast.expressions.FunctionCall
 import prog8.ast.statements.*
@@ -39,7 +40,11 @@ class VerifyFunctionArgTypes(val program: Program) : IAstVisitor {
         }
 
         fun checkTypes(call: IFunctionCall, scope: INameScope, program: Program): String? {
-            val argtypes = call.args.map { it.inferType(program).typeOrElse(DataType.STRUCT) }
+            val argITypes = call.args.map { it.inferType(program) }
+            val firstUnknownDt = argITypes.indexOfFirst { it.isUnknown }
+            if(firstUnknownDt>=0)
+                return "argument ${firstUnknownDt+1} invalid argument type"
+            val argtypes = argITypes.map { it.typeOrElse(DataType.STRUCT) }
             val target = call.target.targetStatement(scope)
             if (target is Subroutine) {
                 if(call.args.size != target.parameters.size)
@@ -67,12 +72,12 @@ class VerifyFunctionArgTypes(val program: Program) : IAstVisitor {
                 if(call.args.size != func.parameters.size)
                     return "invalid number of arguments"
                 val paramtypes = func.parameters.map { it.possibleDatatypes }
-                for (x in argtypes.zip(paramtypes).withIndex()) {
-                    val anyCompatible = x.value.second.any { argTypeCompatible(x.value.first, it) }
+                argtypes.zip(paramtypes).forEachIndexed { index, pair ->
+                    val anyCompatible = pair.second.any { argTypeCompatible(pair.first, it) }
                     if (!anyCompatible) {
-                        val actual = x.value.first.toString()
-                        val expected = x.value.second.toString()
-                        return "argument ${x.index + 1} type mismatch, was: $actual expected: $expected"
+                        val actual = pair.first.toString()
+                        val expected = pair.second.toString()
+                        return "argument ${index + 1} type mismatch, was: $actual expected: $expected"
                     }
                 }
             }

compiler/src/prog8/ast/statements/AstStatements.kt

@@ -44,7 +44,7 @@ class BuiltinFunctionStatementPlaceholder(val name: String, override val positio
     }
 }
 
-data class RegisterOrStatusflag(val registerOrPair: RegisterOrPair?, val statusflag: Statusflag?, val stack: Boolean)
+data class RegisterOrStatusflag(val registerOrPair: RegisterOrPair?, val statusflag: Statusflag?)
 
 class Block(override val name: String,
             val address: Int?,
@@ -52,7 +52,6 @@ class Block(override val name: String,
             val isInLibrary: Boolean,
             override val position: Position) : Statement(), INameScope {
     override lateinit var parent: Node
-    override val asmGenInfo = AsmGenInfo()
 
     override fun linkParents(parent: Node) {
         this.parent = parent
@@ -187,7 +186,12 @@ open class VarDecl(val type: VarDeclType,
 
         fun createAuto(array: ArrayLiteralValue): VarDecl {
             val autoVarName = "auto_heap_value_${++autoHeapValueSequenceNumber}"
-            val declaredType = ArrayElementTypes.getValue(array.type.typeOrElse(DataType.STRUCT))
+            val arrayDt =
+                if(!array.type.isKnown)
+                    throw FatalAstException("unknown dt")
+                else
+                    array.type.typeOrElse(DataType.STRUCT)
+            val declaredType = ArrayElementTypes.getValue(arrayDt)
             val arraysize = ArrayIndex.forArray(array)
             return VarDecl(VarDeclType.VAR, declaredType, ZeropageWish.NOT_IN_ZEROPAGE, arraysize, autoVarName, null, array,
                     isArray = true, autogeneratedDontRemove = true, position = array.position)
@@ -196,7 +200,7 @@ open class VarDecl(val type: VarDeclType,
         fun defaultZero(dt: DataType, position: Position) = when(dt) {
             DataType.UBYTE -> NumericLiteralValue(DataType.UBYTE, 0,  position)
             DataType.BYTE -> NumericLiteralValue(DataType.BYTE, 0,  position)
-            DataType.UWORD -> NumericLiteralValue(DataType.UWORD, 0, position)
+            DataType.UWORD, DataType.STR -> NumericLiteralValue(DataType.UWORD, 0, position)
             DataType.WORD -> NumericLiteralValue(DataType.WORD, 0, position)
             DataType.FLOAT -> NumericLiteralValue(DataType.FLOAT, 0.0, position)
             else -> throw FatalAstException("can only determine default zero value for a numeric type")
@@ -252,9 +256,9 @@ open class VarDecl(val type: VarDeclType,
     }
 
     fun flattenStructMembers(): MutableList<Statement> {
-        val result = struct!!.statements.withIndex().map {
-            val member = it.value as VarDecl
-            val initvalue = if(value!=null) (value as ArrayLiteralValue).value[it.index] else null
+        val result = struct!!.statements.mapIndexed { index, statement ->
+            val member = statement as VarDecl
+            val initvalue = if(value!=null) (value as ArrayLiteralValue).value[index] else null
             VarDecl(
                     VarDeclType.VAR,
                     member.datatype,
@@ -309,6 +313,7 @@ class ArrayIndex(var origExpression: Expression?,            // will be replaced
                     }
                     is IdentifierReference -> {
                         indexVar = replacement
+                        indexNum = null
                     }
                     else -> {
                         throw FatalAstException("invalid replace")
@@ -343,7 +348,12 @@ class ArrayIndex(var origExpression: Expression?,            // will be replaced
 
     fun constIndex() = indexNum?.number?.toInt()
 
-    infix fun isSameAs(other: ArrayIndex) = indexNum==other.indexNum && indexVar == other.indexVar
+    infix fun isSameAs(other: ArrayIndex): Boolean {
+        return if(indexNum!=null || indexVar!=null)
+            indexNum==other.indexNum && indexVar == other.indexVar
+        else
+            other.origExpression!=null && origExpression!! isSameAs other.origExpression!!
+    }
 }
 
 open class Assignment(var target: AssignTarget, var value: Expression, override val position: Position) : Statement() {
@@ -444,9 +454,9 @@ data class AssignTarget(var identifier: IdentifierReference?,
     fun accept(visitor: IAstVisitor) = visitor.visit(this)
     fun accept(visitor: AstWalker, parent: Node) = visitor.visit(this, parent)
 
-    fun inferType(program: Program, stmt: Statement): InferredTypes.InferredType {  // TODO why does this have the extra 'stmt' scope parameter???
+    fun inferType(program: Program): InferredTypes.InferredType {
         if (identifier != null) {
-            val symbol = program.namespace.lookup(identifier!!.nameInSource, stmt) ?: return InferredTypes.unknown()
+            val symbol = program.namespace.lookup(identifier!!.nameInSource, this) ?: return InferredTypes.unknown()
             if (symbol is VarDecl) return InferredTypes.knownFor(symbol.datatype)
         }
 
@@ -637,7 +647,6 @@ class AnonymousScope(override var statements: MutableList<Statement>,
                      override val position: Position) : INameScope, Statement() {
     override val name: String
     override lateinit var parent: Node
-    override val asmGenInfo = AsmGenInfo()
 
     companion object {
         private var sequenceNumber = 1
@@ -676,6 +685,22 @@ class NopStatement(override val position: Position): Statement() {
     override fun accept(visitor: AstWalker, parent: Node) = visitor.visit(this, parent)
 }
 
+
+class AsmGenInfo {
+    // This class contains various attributes that influence the assembly code generator.
+    // Conceptually it should be part of any INameScope.
+    // But because the resulting code only creates "real" scopes on a subroutine level,
+    // it's more consistent to only define these attributes on a Subroutine node.
+    var usedAutoArrayIndexerForStatements = mutableListOf<ArrayIndexerInfo>()
+    var usedRegsaveA = false
+    var usedRegsaveX = false
+    var usedRegsaveY = false
+    var usedFloatEvalResultVar1 = false
+    var usedFloatEvalResultVar2 = false
+
+    class ArrayIndexerInfo(val name: String, val replaces: ArrayIndex, val partOfStatement: Statement, var used: Int=0)
+}
+
 // the subroutine class covers both the normal user-defined subroutines,
 // and also the predefined/ROM/register-based subroutines.
 // (multiple return types can only occur for the latter type)
@@ -690,8 +715,24 @@ class Subroutine(override val name: String,
                  override var statements: MutableList<Statement>,
                  override val position: Position) : Statement(), INameScope {
 
+    constructor(name: String, parameters: List<SubroutineParameter>, returntypes: List<DataType>, statements: MutableList<Statement>, position: Position)
+            : this(name, parameters, returntypes, emptyList(), determineReturnRegisters(returntypes), emptySet(), null, false, statements, position)
+
+    companion object {
+        private fun determineReturnRegisters(returntypes: List<DataType>): List<RegisterOrStatusflag> {
+            // for non-asm subroutines, determine the return registers based on the type of the return value
+            return when(returntypes.singleOrNull()) {
+                in ByteDatatypes -> listOf(RegisterOrStatusflag(RegisterOrPair.A, null))
+                in WordDatatypes -> listOf(RegisterOrStatusflag(RegisterOrPair.AY, null))
+                DataType.FLOAT -> listOf(RegisterOrStatusflag(RegisterOrPair.AY, null))
+                null -> emptyList()
+                else -> listOf(RegisterOrStatusflag(RegisterOrPair.AY, null))
+            }
+        }
+    }
+
     override lateinit var parent: Node
-    override val asmGenInfo = AsmGenInfo()
+    val asmGenInfo = AsmGenInfo()
     val scopedname: String by lazy { makeScopedName(name) }
 
     override fun linkParents(parent: Node) {
@@ -716,6 +757,8 @@ class Subroutine(override val name: String,
 
     fun regXasResult() = asmReturnvaluesRegisters.any { it.registerOrPair in setOf(RegisterOrPair.X, RegisterOrPair.AX, RegisterOrPair.XY) }
     fun regXasParam() = asmParameterRegisters.any { it.registerOrPair in setOf(RegisterOrPair.X, RegisterOrPair.AX, RegisterOrPair.XY) }
+    fun shouldPreserveStatusRegisterAfterCall() = asmReturnvaluesRegisters.any { it.statusflag != null } || shouldSaveX()
+    fun shouldSaveX() = CpuRegister.X in asmClobbers || regXasResult() || regXasParam()
 
     fun amountOfRtsInAsm(): Int = statements
             .asSequence()
@@ -975,7 +1018,6 @@ class StructDecl(override val name: String,
                  override val position: Position): Statement(), INameScope {
 
     override lateinit var parent: Node
-    override val asmGenInfo = AsmGenInfo()
 
     override fun linkParents(parent: Node) {
         this.parent = parent

compiler/src/prog8/compiler/BeforeAsmGenerationAstChanger.kt

@@ -15,7 +15,7 @@ internal class BeforeAsmGenerationAstChanger(val program: Program, val errors: E
     private val noModifications = emptyList<IAstModification>()
 
     override fun after(decl: VarDecl, parent: Node): Iterable<IAstModification> {
-        subroutineVariables.add(Pair(decl.name, decl))
+        subroutineVariables.add(decl.name to decl)
         if (decl.value == null && !decl.autogeneratedDontRemove && decl.type == VarDeclType.VAR && decl.datatype in NumericDatatypes) {
             // a numeric vardecl without an initial value is initialized with zero,
             // unless there's already an assignment below, that initializes the value
@@ -74,7 +74,7 @@ internal class BeforeAsmGenerationAstChanger(val program: Program, val errors: E
 
     override fun after(scope: AnonymousScope, parent: Node): Iterable<IAstModification> {
         val decls = scope.statements.filterIsInstance<VarDecl>()
-        subroutineVariables.addAll(decls.map { Pair(it.name, it) })
+        subroutineVariables.addAll(decls.map { it.name to it })
 
         val sub = scope.definingSubroutine()
         if (sub != null) {
@@ -170,6 +170,12 @@ internal class BeforeAsmGenerationAstChanger(val program: Program, val errors: E
                             AddressOf(typecast.expression as IdentifierReference, typecast.position),
                             parent
                     ))
+                } else if(typecast.expression is IFunctionCall) {
+                    return listOf(IAstModification.ReplaceNode(
+                            typecast,
+                            typecast.expression,
+                            parent
+                    ))
                 }
             } else {
                 errors.err("cannot cast pass-by-reference value to type ${typecast.type} (only to UWORD)", typecast.position)

compiler/src/prog8/compiler/Zeropage.kt

@@ -64,7 +64,7 @@ abstract class Zeropage(protected val options: CompilationOptions) {
 
     private fun makeAllocation(address: Int, size: Int, datatype: DataType, name: String?): Int {
         free.removeAll(address until address+size)
-        allocations[address] = Pair(name ?: "<unnamed>", datatype)
+        allocations[address] = (name ?: "<unnamed>") to datatype
         return address
     }
 

compiler/src/prog8/compiler/target/IAssemblyGenerator.kt

@@ -7,6 +7,8 @@ internal interface IAssemblyGenerator {
 }
 
 internal const val generatedLabelPrefix = "_prog8_label_"
+internal const val subroutineFloatEvalResultVar1 = "_prog8_float_eval_result1"
+internal const val subroutineFloatEvalResultVar2 = "_prog8_float_eval_result2"
 
 internal interface IAssemblyProgram {
     val name: String

compiler/src/prog8/compiler/target/c64/C64MachineDefinition.kt

@@ -110,7 +110,7 @@ internal object C64MachineDefinition: IMachineDefinition {
     internal class C64Zeropage(options: CompilationOptions) : Zeropage(options) {
 
         override val SCRATCH_B1 = 0x02      // temp storage for a single byte
-        override val SCRATCH_REG = 0x03     // temp storage for a register
+        override val SCRATCH_REG = 0x03     // temp storage for a register, must be B1+1
         override val SCRATCH_W1 = 0xfb      // temp storage 1 for a word  $fb+$fc
         override val SCRATCH_W2 = 0xfd      // temp storage 2 for a word  $fb+$fc
 

compiler/src/prog8/compiler/target/c64/codegen/AsmGen.kt

@@ -8,20 +8,19 @@ import prog8.ast.base.*
 import prog8.ast.expressions.*
 import prog8.ast.statements.*
 import prog8.compiler.*
+import prog8.compiler.target.*
 import prog8.compiler.target.CompilationTarget
 import prog8.compiler.target.CpuType
 import prog8.compiler.target.IAssemblyGenerator
 import prog8.compiler.target.IAssemblyProgram
 import prog8.compiler.target.c64.AssemblyProgram
 import prog8.compiler.target.c64.Petscii
-import prog8.compiler.target.c64.codegen.assignment.AsmAssignSource
-import prog8.compiler.target.c64.codegen.assignment.AsmAssignTarget
 import prog8.compiler.target.c64.codegen.assignment.AsmAssignment
 import prog8.compiler.target.c64.codegen.assignment.AssignmentAsmGen
-import prog8.compiler.target.c64.codegen.assignment.TargetStorageKind
 import prog8.compiler.target.generatedLabelPrefix
 import prog8.functions.BuiltinFunctions
 import prog8.functions.FSignature
+import java.io.CharConversionException
 import java.nio.file.Path
 import java.time.LocalDate
 import java.time.LocalDateTime
@@ -37,7 +36,7 @@ internal class AsmGen(private val program: Program,
 
     // for expressions and augmented assignments:
     val optimizedByteMultiplications = setOf(3,5,6,7,9,10,11,12,13,14,15,20,25,40,50,80,100)
-    val optimizedWordMultiplications = setOf(3,5,6,7,9,10,12,15,20,25,40,50,80,100)
+    val optimizedWordMultiplications = setOf(3,5,6,7,9,10,12,15,20,25,40,50,80,100,320)
 
     private val assemblyLines = mutableListOf<String>()
     private val globalFloatConsts = mutableMapOf<Double, String>()     // all float values in the entire program (value -> varname)
@@ -161,6 +160,7 @@ internal class AsmGen(private val program: Program,
             val floatvalue = flt.key
             out("${flt.value}\t.byte  $floatFill  ; float $floatvalue")
         }
+        out("prog8_program_end\t; end of program label for progend()")
     }
 
     private fun block2asm(block: Block) {
@@ -200,11 +200,7 @@ internal class AsmGen(private val program: Program,
 
     private fun assignInitialValueToVar(decl: VarDecl, variableName: List<String>) {
         val asmName = asmVariableName(variableName)
-        val asgn = AsmAssignment(
-                AsmAssignSource.fromAstSource(decl.value!!, program, this),
-                AsmAssignTarget(TargetStorageKind.VARIABLE, program, this, decl.datatype, decl.definingSubroutine(), variableAsmName = asmName),
-                false, decl.position)
-        assignmentAsmGen.translateNormalAssignment(asgn)
+        assignmentAsmGen.assignExpressionToVariable(decl.value!!, asmName, decl.datatype, decl.definingSubroutine())
     }
 
     private var generatedLabelSequenceNumber: Int = 0
@@ -231,8 +227,12 @@ internal class AsmGen(private val program: Program,
     }
 
     private fun encode(str: String, altEncoding: Boolean): List<Short> {
-         val bytes = if(altEncoding) Petscii.encodeScreencode(str, true) else Petscii.encodePetscii(str, true)
-         return bytes.plus(0)
+        try {
+            val bytes = if (altEncoding) Petscii.encodeScreencode(str, true) else Petscii.encodePetscii(str, true)
+            return bytes.plus(0)
+        } catch(x: CharConversionException) {
+            throw AssemblyError("There was a problem converting a string to the target machine's char encoding: ${x.message}")
+        }
     }
 
     private fun zeropagevars2asm(statements: List<Statement>) {
@@ -253,7 +253,7 @@ internal class AsmGen(private val program: Program,
                         errors.handle()
                         out("${variable.name} = $address\t; auto zp ${variable.datatype}")
                         // make sure we add the var to the set of zpvars for this block
-                        allocatedZeropageVariables[fullName] = Pair(address, variable.datatype)
+                        allocatedZeropageVariables[fullName] = address to variable.datatype
                     } catch (x: ZeropageDepletedError) {
                         // leave it as it is.
                     }
@@ -550,23 +550,20 @@ internal class AsmGen(private val program: Program,
 
     private  fun fixNameSymbols(name: String) = name.replace("<", "prog8_").replace(">", "")     // take care of the autogenerated invalid (anon) label names
 
-    internal fun saveRegister(register: CpuRegister, dontUseStack: Boolean, scope: Subroutine?) {
+    internal fun saveRegister(register: CpuRegister, dontUseStack: Boolean, scope: Subroutine) {
         if(dontUseStack) {
             when (register) {
                 CpuRegister.A -> {
                     out("  sta  _prog8_regsaveA")
-                    if (scope != null)
-                        scope.asmGenInfo.usedRegsaveA = true
+                    scope.asmGenInfo.usedRegsaveA = true
                 }
                 CpuRegister.X -> {
                     out("  stx  _prog8_regsaveX")
-                    if (scope != null)
-                        scope.asmGenInfo.usedRegsaveX = true
+                    scope.asmGenInfo.usedRegsaveX = true
                 }
                 CpuRegister.Y -> {
                     out("  sty  _prog8_regsaveY")
-                    if (scope != null)
-                        scope.asmGenInfo.usedRegsaveY = true
+                    scope.asmGenInfo.usedRegsaveY = true
                 }
             }
 
@@ -577,16 +574,14 @@ internal class AsmGen(private val program: Program,
                     if (CompilationTarget.instance.machine.cpu == CpuType.CPU65c02) out("  phx")
                     else {
                         out("  stx  _prog8_regsaveX")
-                        if (scope != null)
-                            scope.asmGenInfo.usedRegsaveX = true
+                        scope.asmGenInfo.usedRegsaveX = true
                     }
                 }
                 CpuRegister.Y -> {
                     if (CompilationTarget.instance.machine.cpu == CpuType.CPU65c02) out("  phy")
                     else {
                         out("  sty  _prog8_regsaveY")
-                        if (scope != null)
-                            scope.asmGenInfo.usedRegsaveY = true
+                        scope.asmGenInfo.usedRegsaveY = true
                     }
                 }
             }
@@ -633,19 +628,7 @@ internal class AsmGen(private val program: Program,
                 if (builtinFunc != null) {
                     builtinFunctionsAsmGen.translateFunctioncallStatement(stmt, builtinFunc)
                 } else {
-                    val sub = stmt.target.targetSubroutine(program.namespace)!!
-                    val preserveStatusRegisterAfterCall = sub.asmReturnvaluesRegisters.any {it.statusflag!=null}
-                    functioncallAsmGen.translateFunctionCall(stmt, preserveStatusRegisterAfterCall)
-                    // discard any results from the stack:
-                    val returns = sub.returntypes.zip(sub.asmReturnvaluesRegisters)
-                    for ((t, reg) in returns) {
-                        if (reg.stack) {
-                            if (t in IntegerDatatypes || t in PassByReferenceDatatypes) out("  inx")
-                            else if (t == DataType.FLOAT) out("  inx |  inx |  inx")
-                        }
-                    }
-                    if(preserveStatusRegisterAfterCall)
-                        out("  plp\t; restore status flags from call")
+                    functioncallAsmGen.translateFunctionCallStatement(stmt)
                 }
             }
             is Assignment -> assignmentAsmGen.translate(stmt)
@@ -755,8 +738,8 @@ internal class AsmGen(private val program: Program,
     internal fun translateExpression(indexer: ArrayIndex) =
             expressionsAsmGen.translateExpression(indexer)
 
-    internal fun translateFunctioncallExpression(functionCall: FunctionCall, signature: FSignature) =
-            builtinFunctionsAsmGen.translateFunctioncallExpression(functionCall, signature)
+    internal fun translateBuiltinFunctionCallExpression(functionCall: FunctionCall, signature: FSignature, resultToStack: Boolean) =
+            builtinFunctionsAsmGen.translateFunctioncallExpression(functionCall, signature, resultToStack)
 
     internal fun translateFunctionCall(functionCall: FunctionCall, preserveStatusRegisterAfterCall: Boolean) =
             functioncallAsmGen.translateFunctionCall(functionCall, preserveStatusRegisterAfterCall)
@@ -764,6 +747,16 @@ internal class AsmGen(private val program: Program,
     internal fun translateNormalAssignment(assign: AsmAssignment) =
             assignmentAsmGen.translateNormalAssignment(assign)
 
+    internal fun assignExpressionToRegister(expr: Expression, register: RegisterOrPair) =
+            assignmentAsmGen.assignExpressionToRegister(expr, register)
+
+    internal fun assignExpressionToVariable(expr: Expression, asmVarName: String, dt: DataType, scope: Subroutine?) =
+            assignmentAsmGen.assignExpressionToVariable(expr, asmVarName, dt, scope)
+
+    internal fun assignVariableToRegister(asmVarName: String, register: RegisterOrPair) =
+            assignmentAsmGen.assignVariableToRegister(asmVarName, register)
+
+
     private fun translateSubroutine(sub: Subroutine) {
         out("")
         outputSourceLine(sub)
@@ -808,6 +801,10 @@ internal class AsmGen(private val program: Program,
                 out("_prog8_regsaveX     .byte  0")
             if(sub.asmGenInfo.usedRegsaveY)
                 out("_prog8_regsaveY     .byte  0")
+            if(sub.asmGenInfo.usedFloatEvalResultVar1)
+                out("$subroutineFloatEvalResultVar1    .byte 0,0,0,0,0")
+            if(sub.asmGenInfo.usedFloatEvalResultVar2)
+                out("$subroutineFloatEvalResultVar2    .byte 0,0,0,0,0")
             vardecls2asm(sub.statements)
             out("  .pend\n")
         }
@@ -818,12 +815,12 @@ internal class AsmGen(private val program: Program,
                 when (condition) {
                     BranchCondition.CS -> "bcc"
                     BranchCondition.CC -> "bcs"
-                    BranchCondition.EQ, BranchCondition.Z -> "beq"
-                    BranchCondition.NE, BranchCondition.NZ -> "bne"
+                    BranchCondition.EQ, BranchCondition.Z -> "bne"
+                    BranchCondition.NE, BranchCondition.NZ -> "beq"
                     BranchCondition.VS -> "bvc"
                     BranchCondition.VC -> "bvs"
-                    BranchCondition.MI, BranchCondition.NEG -> "bmi"
-                    BranchCondition.PL, BranchCondition.POS -> "bpl"
+                    BranchCondition.MI, BranchCondition.NEG -> "bpl"
+                    BranchCondition.PL, BranchCondition.POS -> "bmi"
                 }
             } else {
                 when (condition) {
@@ -912,15 +909,16 @@ internal class AsmGen(private val program: Program,
                 }
             }
             else -> {
-                translateExpression(stmt.iterations!!)
-                val dt = stmt.iterations!!.inferType(program).typeOrElse(DataType.STRUCT)
-                when (dt) {
+                val dt = stmt.iterations!!.inferType(program)
+                if(!dt.isKnown)
+                    throw AssemblyError("unknown dt")
+                when (dt.typeOrElse(DataType.STRUCT)) {
                     in ByteDatatypes -> {
-                        out("  inx |  lda  P8ESTACK_LO,x")
+                        assignExpressionToRegister(stmt.iterations!!, RegisterOrPair.A)
                         repeatByteCountInA(null, repeatLabel, endLabel, stmt.body)
                     }
                     in WordDatatypes -> {
-                        out("  inx |  lda  P8ESTACK_LO,x |  ldy  P8ESTACK_HI,x")
+                        assignExpressionToRegister(stmt.iterations!!, RegisterOrPair.AY)
                         repeatWordCountInAY(null, repeatLabel, endLabel, stmt.body)
                     }
                     else -> throw AssemblyError("invalid loop expression datatype $dt")
@@ -1016,16 +1014,16 @@ $counterVar    .byte  0""")
     }
 
     private fun translate(stmt: WhenStatement) {
-        expressionsAsmGen.translateExpression(stmt.condition)
         val endLabel = makeLabel("choice_end")
         val choiceBlocks = mutableListOf<Pair<String, AnonymousScope>>()
         val conditionDt = stmt.condition.inferType(program)
         if(!conditionDt.isKnown)
             throw AssemblyError("unknown condition dt")
         if(conditionDt.typeOrElse(DataType.BYTE) in ByteDatatypes)
-            out("  inx |  lda  P8ESTACK_LO,x")
+            assignExpressionToRegister(stmt.condition, RegisterOrPair.A)
         else
-            out("  inx |  lda  P8ESTACK_LO,x |  ldy  P8ESTACK_HI,x")
+            assignExpressionToRegister(stmt.condition, RegisterOrPair.AY)
+
         for(choice in stmt.choices) {
             val choiceLabel = makeLabel("choice")
             if(choice.values==null) {
@@ -1033,7 +1031,7 @@ $counterVar    .byte  0""")
                 translate(choice.statements)
                 out("  jmp  $endLabel")
             } else {
-                choiceBlocks.add(Pair(choiceLabel, choice.statements))
+                choiceBlocks.add(choiceLabel to choice.statements)
                 for (cv in choice.values!!) {
                     val value = (cv as NumericLiteralValue).number.toInt()
                     if(conditionDt.typeOrElse(DataType.BYTE) in ByteDatatypes) {
@@ -1074,26 +1072,48 @@ $counterVar    .byte  0""")
 
         val jump = stmt.truepart.statements.first() as? Jump
         if(jump!=null) {
-            // branch with only a jump
+            // branch with only a jump (goto)
             val instruction = branchInstruction(stmt.condition, false)
             out("  $instruction  ${getJumpTarget(jump)}")
             translate(stmt.elsepart)
         } else {
+            val truePartIsJustBreak = stmt.truepart.statements.firstOrNull() is Break
+            val elsePartIsJustBreak = stmt.elsepart.statements.firstOrNull() is Break
             if(stmt.elsepart.containsNoCodeNorVars()) {
+                if(truePartIsJustBreak) {
+                    // branch with just a break (jump out of loop)
+                    val instruction = branchInstruction(stmt.condition, false)
+                    val loopEndLabel = loopEndLabels.peek()
+                    out("  $instruction  $loopEndLabel")
+                } else {
+                    val instruction = branchInstruction(stmt.condition, true)
+                    val elseLabel = makeLabel("branch_else")
+                    out("  $instruction  $elseLabel")
+                    translate(stmt.truepart)
+                    out(elseLabel)
+                }
+            }
+            else if(truePartIsJustBreak) {
+                // branch with just a break (jump out of loop)
+                val instruction = branchInstruction(stmt.condition, false)
+                val loopEndLabel = loopEndLabels.peek()
+                out("  $instruction  $loopEndLabel")
+                translate(stmt.elsepart)
+            } else if(elsePartIsJustBreak) {
+                // branch with just a break (jump out of loop) but true/false inverted
+                val instruction = branchInstruction(stmt.condition, true)
+                val loopEndLabel = loopEndLabels.peek()
+                out("  $instruction  $loopEndLabel")
+                translate(stmt.truepart)
+            } else {
                 val instruction = branchInstruction(stmt.condition, true)
                 val elseLabel = makeLabel("branch_else")
+                val endLabel = makeLabel("branch_end")
                 out("  $instruction  $elseLabel")
                 translate(stmt.truepart)
-                out(elseLabel)
-            } else {
-                val instruction = branchInstruction(stmt.condition, false)
-                val trueLabel = makeLabel("branch_true")
-                val endLabel = makeLabel("branch_end")
-                out("  $instruction  $trueLabel")
-                translate(stmt.elsepart)
                 out("  jmp  $endLabel")
-                out(trueLabel)
-                translate(stmt.truepart)
+                out(elseLabel)
+                translate(stmt.elsepart)
                 out(endLabel)
             }
         }
@@ -1139,7 +1159,9 @@ $counterVar    .byte  0""")
             "%breakpoint" -> {
                 val label = "_prog8_breakpoint_${breakpointLabels.size+1}"
                 breakpointLabels.add(label)
-                out("$label\tnop")
+                out("""
+                    nop
+$label              nop""")
             }
         }
     }
@@ -1165,7 +1187,28 @@ $counterVar    .byte  0""")
     }
 
     private fun translate(ret: Return) {
-        ret.value?.let { expressionsAsmGen.translateExpression(it) }
+        ret.value?.let { returnvalue ->
+            val sub = ret.definingSubroutine()!!
+            val returnType = sub.returntypes.single()
+            val returnReg = sub.asmReturnvaluesRegisters.single()
+            if(returnReg.registerOrPair==null)
+                throw AssemblyError("normal subroutines can't return value in status register directly")
+
+            when (returnType) {
+                in IntegerDatatypes -> {
+                    assignmentAsmGen.assignExpressionToRegister(returnvalue, returnReg.registerOrPair)
+                }
+                DataType.FLOAT -> {
+                    // return the float value via FAC1
+                    assignExpressionToRegister(returnvalue, RegisterOrPair.FAC1)
+                }
+                else -> {
+                    // all else take its address and assign that also to AY register pair
+                    val addrofValue = AddressOf(returnvalue as IdentifierReference, returnvalue.position)
+                    assignmentAsmGen.assignExpressionToRegister(addrofValue, returnReg.registerOrPair)
+                }
+            }
+        }
         out("  rts")
     }
 
@@ -1174,33 +1217,41 @@ $counterVar    .byte  0""")
         assemblyLines.add(assembly)
     }
 
+    internal fun signExtendAYlsb(valueDt: DataType) {
+        // sign extend signed byte in A to full word in AY
+        when(valueDt) {
+            DataType.UBYTE -> out("  ldy  #0")
+            DataType.BYTE -> out("  jsr  prog8_lib.sign_extend_AY_byte")
+            else -> throw AssemblyError("need byte type")
+        }
+    }
+
     internal fun signExtendStackLsb(valueDt: DataType) {
-        // sign extend signed byte on stack to signed word
+        // sign extend signed byte on stack to signed word on stack
         when(valueDt) {
             DataType.UBYTE -> {
-                out("  lda  #0 |  sta  P8ESTACK_HI+1,x")
-            }
-            DataType.BYTE -> {
-                out("""
-                    lda  P8ESTACK_LO+1,x
-                    ora  #$7f
-                    bmi  +
-                    lda  #0
-+                   sta  P8ESTACK_HI+1,x""")
+                if(CompilationTarget.instance.machine.cpu == CpuType.CPU65c02)
+                    out("  stz  P8ESTACK_HI+1,x")
+                else
+                    out("  lda  #0 |  sta  P8ESTACK_HI+1,x")
             }
+            DataType.BYTE -> out("  jsr  prog8_lib.sign_extend_stack_byte")
             else -> throw AssemblyError("need byte type")
         }
     }
 
     internal fun signExtendVariableLsb(asmvar: String, valueDt: DataType) {
-        // sign extend signed byte in a word variable
+        // sign extend signed byte in a var to a full word in that variable
         when(valueDt) {
             DataType.UBYTE -> {
-                out("  lda  #0 |  sta  $asmvar+1")
+                if(CompilationTarget.instance.machine.cpu == CpuType.CPU65c02)
+                    out("  stz  $asmvar+1")
+                else
+                    out("  lda  #0 |  sta  $asmvar+1")
             }
             DataType.BYTE -> {
                 out("""
-                    lda  $asmvar+1
+                    lda  $asmvar
                     ora  #$7f
                     bmi  +
                     lda  #0

compiler/src/prog8/compiler/target/c64/codegen/AsmOptimizer.kt

@@ -75,11 +75,11 @@ private fun getLinesBy(lines: MutableList<String>, windowSize: Int) =
 
 private fun optimizeCmpSequence(linesByFour: List<List<IndexedValue<String>>>): List<Modification> {
     // the when statement (on bytes) generates a sequence of:
-    //	 lda $ce01,x
-    //	 cmp #$20
+    //	 lda  $ce01,x
+    //	 cmp  #$20
     //	 beq  check_prog8_s72choice_32
-    //	 lda $ce01,x
-    //	 cmp #$21
+    //	 lda  $ce01,x
+    //	 cmp  #$21
     //	 beq  check_prog8_s73choice_33
     // the repeated lda can be removed
     val mods = mutableListOf<Modification>()
@@ -179,8 +179,8 @@ private fun optimizeSameAssignments(linesByFourteen: List<List<IndexedValue<Stri
 }
 
 private fun optimizeStoreLoadSame(linesByFour: List<List<IndexedValue<String>>>): List<Modification> {
-    // TODO not sure if this is correct in all situations....:
     // sta X + lda X,  sty X + ldy X,   stx X + ldx X  -> the second instruction can be eliminated
+    // TODO this is not true if X is not a regular RAM memory address (but instead mapped I/O or ROM)
     val mods = mutableListOf<Modification>()
     for (pair in linesByFour) {
         val first = pair[0].value.trimStart()

compiler/src/prog8/compiler/target/c64/codegen/BuiltinFunctionsAsmGen.kt

@@ -1,53 +1,59 @@
 package prog8.compiler.target.c64.codegen
 
 import prog8.ast.IFunctionCall
+import prog8.ast.Node
 import prog8.ast.Program
 import prog8.ast.base.*
 import prog8.ast.expressions.*
-import prog8.ast.statements.DirectMemoryWrite
-import prog8.ast.statements.FunctionCallStatement
+import prog8.ast.statements.*
 import prog8.compiler.AssemblyError
+import prog8.compiler.target.CompilationTarget
+import prog8.compiler.target.Cx16Target
 import prog8.compiler.target.c64.codegen.assignment.AsmAssignSource
 import prog8.compiler.target.c64.codegen.assignment.AsmAssignTarget
 import prog8.compiler.target.c64.codegen.assignment.AsmAssignment
 import prog8.compiler.target.c64.codegen.assignment.SourceStorageKind
 import prog8.compiler.target.c64.codegen.assignment.TargetStorageKind
+import prog8.compiler.target.subroutineFloatEvalResultVar2
 import prog8.compiler.toHex
 import prog8.functions.FSignature
 
 internal class BuiltinFunctionsAsmGen(private val program: Program, private val asmgen: AsmGen) {
 
-    internal fun translateFunctioncallExpression(fcall: FunctionCall, func: FSignature) {
-        translateFunctioncall(fcall, func, false)
+    internal fun translateFunctioncallExpression(fcall: FunctionCall, func: FSignature, resultToStack: Boolean) {
+        translateFunctioncall(fcall, func, discardResult = false, resultToStack = resultToStack)
     }
 
     internal fun translateFunctioncallStatement(fcall: FunctionCallStatement, func: FSignature) {
-        translateFunctioncall(fcall, func, true)
+        translateFunctioncall(fcall, func, discardResult = true, resultToStack = false)
     }
 
-    private fun translateFunctioncall(fcall: IFunctionCall, func: FSignature, discardResult: Boolean) {
-        val functionName = fcall.target.nameInSource.last()
-        if (discardResult) {
-            if (func.pure)
-                return  // can just ignore the whole function call altogether
-            else if (func.returntype != null)
-                throw AssemblyError("discarding result of non-pure function $fcall")
-        }
+    private fun translateFunctioncall(fcall: IFunctionCall, func: FSignature, discardResult: Boolean, resultToStack: Boolean) {
+        if (discardResult && func.pure)
+            return  // can just ignore the whole function call altogether
 
-        when (functionName) {
-            "msb" -> funcMsb(fcall)
-            "lsb" -> funcLsb(fcall)
-            "mkword" -> funcMkword(fcall, func)
-            "abs" -> funcAbs(fcall, func)
+        if(discardResult && resultToStack)
+            throw AssemblyError("cannot both discard the result AND put it onto stack")
+
+        val scope = (fcall as Node).definingSubroutine()!!
+        when (func.name) {
+            "msb" -> funcMsb(fcall, resultToStack)
+            "lsb" -> funcLsb(fcall, resultToStack)
+            "mkword" -> funcMkword(fcall, resultToStack)
+            "abs" -> funcAbs(fcall, func, resultToStack, scope)
             "swap" -> funcSwap(fcall)
-            "strlen" -> funcStrlen(fcall)
-            "min", "max", "sum" -> funcMinMaxSum(fcall, functionName)
-            "any", "all" -> funcAnyAll(fcall, functionName)
-            "sgn" -> funcSgn(fcall, func)
+            "min", "max" -> funcMinMax(fcall, func, resultToStack)
+            "sum" -> funcSum(fcall, resultToStack)
+            "any", "all" -> funcAnyAll(fcall, func, resultToStack)
+            "sin8", "sin8u", "sin16", "sin16u",
+            "cos8", "cos8u", "cos16", "cos16u" -> funcSinCosInt(fcall, func, resultToStack, scope)
+            "sgn" -> funcSgn(fcall, func, resultToStack, scope)
             "sin", "cos", "tan", "atan",
             "ln", "log2", "sqrt", "rad",
             "deg", "round", "floor", "ceil",
-            "rdnf" -> funcVariousFloatFuncs(fcall, func, functionName)
+            "rndf" -> funcVariousFloatFuncs(fcall, func, resultToStack, scope)
+            "rnd", "rndw" -> funcRnd(func, resultToStack)
+            "sqrt16" -> funcSqrt16(fcall, func, resultToStack, scope)
             "rol" -> funcRol(fcall)
             "rol2" -> funcRol2(fcall)
             "ror" -> funcRor(fcall)
@@ -63,17 +69,128 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
                 // restore all registers and cpu status flag
                 asmgen.out(" pla |  tay |  pla |  tax |  pla |  plp")
             }
+            "read_flags" -> {
+                if(resultToStack)
+                    asmgen.out("  jsr  prog8_lib.func_read_flags_stack")
+                else
+                    asmgen.out("  php |  pla")
+            }
             "clear_carry" -> asmgen.out("  clc")
             "set_carry" -> asmgen.out("  sec")
             "clear_irqd" -> asmgen.out("  cli")
             "set_irqd" -> asmgen.out("  sei")
-            else -> {
-                translateFunctionArguments(fcall.args, func)
-                asmgen.out("  jsr  prog8_lib.func_$functionName")
+            "strlen" -> funcStrlen(fcall, resultToStack)
+            "strcmp" -> funcStrcmp(fcall, func, resultToStack, scope)
+            "strcopy" -> {
+                translateArguments(fcall.args, func, scope)
+                if(resultToStack)
+                    asmgen.out("  jsr  prog8_lib.func_strcopy_to_stack")
+                else
+                    asmgen.out("  jsr  prog8_lib.func_strcopy")
             }
+            "memcopy", "memset", "memsetw" -> funcMemSetCopy(fcall, func, scope)
+            "substr", "leftstr", "rightstr" -> {
+                translateArguments(fcall.args, func, scope)
+                asmgen.out("  jsr  prog8_lib.func_${func.name}")
+            }
+            "exit" -> {
+                translateArguments(fcall.args, func, scope)
+                asmgen.out("  jmp  prog8_lib.func_exit")
+            }
+            "progend" -> {
+                if(resultToStack)
+                    asmgen.out("""
+                        lda  #<prog8_program_end
+                        sta  P8ESTACK_LO,x
+                        lda  #>prog8_program_end
+                        sta  P8ESTACK_HI,x
+                        dex""")
+                else
+                    asmgen.out("  lda  #<prog8_program_end |  ldy  #>prog8_program_end")
+            }
+            else -> TODO("missing asmgen for builtin func ${func.name}")
         }
     }
 
+    private fun funcMemSetCopy(fcall: IFunctionCall, func: FSignature, scope: Subroutine) {
+        if(CompilationTarget.instance is Cx16Target) {
+            when(func.name) {
+                "memset" -> {
+                    // use the ROM function of the Cx16
+                    asmgen.assignExpressionToVariable(fcall.args[0], "cx16.r0", DataType.UWORD, scope)
+                    asmgen.assignExpressionToVariable(fcall.args[1], "cx16.r1", DataType.UWORD, scope)
+                    asmgen.assignExpressionToRegister(fcall.args[2], RegisterOrPair.A)
+                    val sub = (fcall as FunctionCallStatement).definingSubroutine()!!
+                    asmgen.saveRegister(CpuRegister.X, false, sub)
+                    asmgen.out("  jsr  cx16.memory_fill")
+                    asmgen.restoreRegister(CpuRegister.X, false)
+                }
+                "memcopy" -> {
+                    val count = fcall.args[2].constValue(program)?.number?.toInt()
+                    val countDt = fcall.args[2].inferType(program)
+                    if((count!=null && count <= 255) || countDt.istype(DataType.UBYTE) || countDt.istype(DataType.BYTE)) {
+                        // fast memcopy of up to 255
+                        translateArguments(fcall.args, func, scope)
+                        asmgen.out("  jsr  prog8_lib.func_memcopy255")
+                        return
+                    }
+
+                    // use the ROM function of the Cx16
+                    asmgen.assignExpressionToVariable(fcall.args[0], "cx16.r0", DataType.UWORD, scope)
+                    asmgen.assignExpressionToVariable(fcall.args[1], "cx16.r1", DataType.UWORD, scope)
+                    asmgen.assignExpressionToVariable(fcall.args[2], "cx16.r2", DataType.UWORD, scope)
+                    val sub = (fcall as FunctionCallStatement).definingSubroutine()!!
+                    asmgen.saveRegister(CpuRegister.X, false, sub)
+                    asmgen.out("  jsr  cx16.memory_copy")
+                    asmgen.restoreRegister(CpuRegister.X, false)
+                }
+                "memsetw" -> {
+                    translateArguments(fcall.args, func, scope)
+                    asmgen.out("  jsr  prog8_lib.func_memsetw")
+                }
+            }
+        } else {
+            if(func.name=="memcopy") {
+                val count = fcall.args[2].constValue(program)?.number?.toInt()
+                val countDt = fcall.args[2].inferType(program)
+                if((count!=null && count <= 255) || countDt.istype(DataType.UBYTE) || countDt.istype(DataType.BYTE)) {
+                    translateArguments(fcall.args, func, scope)
+                    asmgen.out("  jsr  prog8_lib.func_memcopy255")
+                    return
+                }
+            }
+            translateArguments(fcall.args, func, scope)
+            asmgen.out("  jsr  prog8_lib.func_${func.name}")
+        }
+    }
+
+    private fun funcStrcmp(fcall: IFunctionCall, func: FSignature, resultToStack: Boolean, scope: Subroutine) {
+        translateArguments(fcall.args, func, scope)
+        if(resultToStack)
+            asmgen.out("  jsr  prog8_lib.func_strcmp_stack")
+        else
+            asmgen.out("  jsr  prog8_lib.func_strcmp")
+    }
+
+    private fun funcSqrt16(fcall: IFunctionCall, func: FSignature, resultToStack: Boolean, scope: Subroutine) {
+        translateArguments(fcall.args, func, scope)
+        if(resultToStack)
+            asmgen.out("  jsr  prog8_lib.func_sqrt16_stack")
+        else
+            asmgen.out("  jsr  prog8_lib.func_sqrt16_into_A")
+    }
+
+    private fun funcSinCosInt(fcall: IFunctionCall, func: FSignature, resultToStack: Boolean, scope: Subroutine) {
+        translateArguments(fcall.args, func, scope)
+        if(resultToStack)
+            asmgen.out("  jsr  prog8_lib.func_${func.name}_stack")
+        else
+            when(func.name) {
+                "sin8", "sin8u", "cos8", "cos8u" -> asmgen.out("  jsr  prog8_lib.func_${func.name}_into_A")
+                "sin16", "sin16u", "cos16", "cos16u" -> asmgen.out("  jsr  prog8_lib.func_${func.name}_into_AY")
+            }
+    }
+
     private fun funcReverse(fcall: IFunctionCall) {
         val variable = fcall.args.single()
         if (variable is IdentifierReference) {
@@ -83,33 +200,30 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             when (decl.datatype) {
                 DataType.ARRAY_UB, DataType.ARRAY_B -> {
                     asmgen.out("""
-                                    lda  #<$varName
-                                    ldy  #>$varName
-                                    sta  P8ZP_SCRATCH_W1
-                                    sty  P8ZP_SCRATCH_W1+1
-                                    lda  #$numElements
-                                    jsr  prog8_lib.reverse_b
-                                """)
+                        lda  #<$varName
+                        ldy  #>$varName
+                        sta  P8ZP_SCRATCH_W1
+                        sty  P8ZP_SCRATCH_W1+1
+                        lda  #$numElements
+                        jsr  prog8_lib.func_reverse_b""")
                 }
                 DataType.ARRAY_UW, DataType.ARRAY_W -> {
                     asmgen.out("""
-                                    lda  #<$varName
-                                    ldy  #>$varName
-                                    sta  P8ZP_SCRATCH_W1
-                                    sty  P8ZP_SCRATCH_W1+1
-                                    lda  #$numElements
-                                    jsr  prog8_lib.reverse_w
-                                """)
+                        lda  #<$varName
+                        ldy  #>$varName
+                        sta  P8ZP_SCRATCH_W1
+                        sty  P8ZP_SCRATCH_W1+1
+                        lda  #$numElements
+                        jsr  prog8_lib.func_reverse_w""")
                 }
                 DataType.ARRAY_F -> {
                     asmgen.out("""
-                                    lda  #<$varName
-                                    ldy  #>$varName
-                                    sta  P8ZP_SCRATCH_W1
-                                    sty  P8ZP_SCRATCH_W1+1
-                                    lda  #$numElements
-                                    jsr  prog8_lib.reverse_f
-                                """)
+                        lda  #<$varName
+                        ldy  #>$varName
+                        sta  P8ZP_SCRATCH_W1
+                        sty  P8ZP_SCRATCH_W1+1
+                        lda  #$numElements
+                        jsr  floats.func_reverse_f""")
                 }
                 else -> throw AssemblyError("weird type")
             }
@@ -125,25 +239,21 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             when (decl.datatype) {
                 DataType.ARRAY_UB, DataType.ARRAY_B -> {
                     asmgen.out("""
-                                    lda  #<$varName
-                                    ldy  #>$varName
-                                    sta  P8ZP_SCRATCH_W1
-                                    sty  P8ZP_SCRATCH_W1+1
-                                    lda  #$numElements
-                                    sta  P8ZP_SCRATCH_B1
-                                """)
-                    asmgen.out(if (decl.datatype == DataType.ARRAY_UB) "  jsr  prog8_lib.sort_ub" else "  jsr  prog8_lib.sort_b")
+                        lda  #<$varName
+                        ldy  #>$varName
+                        sta  P8ZP_SCRATCH_W1
+                        sty  P8ZP_SCRATCH_W1+1
+                        lda  #$numElements""")
+                    asmgen.out(if (decl.datatype == DataType.ARRAY_UB) "  jsr  prog8_lib.func_sort_ub" else "  jsr  prog8_lib.func_sort_b")
                 }
                 DataType.ARRAY_UW, DataType.ARRAY_W -> {
                     asmgen.out("""
-                                    lda  #<$varName
-                                    ldy  #>$varName
-                                    sta  P8ZP_SCRATCH_W1
-                                    sty  P8ZP_SCRATCH_W1+1
-                                    lda  #$numElements
-                                    sta  P8ZP_SCRATCH_B1
-                                """)
-                    asmgen.out(if (decl.datatype == DataType.ARRAY_UW) "  jsr  prog8_lib.sort_uw" else "  jsr  prog8_lib.sort_w")
+                        lda  #<$varName
+                        ldy  #>$varName
+                        sta  P8ZP_SCRATCH_W1
+                        sty  P8ZP_SCRATCH_W1+1
+                        lda  #$numElements""")
+                    asmgen.out(if (decl.datatype == DataType.ARRAY_UW) "  jsr  prog8_lib.func_sort_uw" else "  jsr  prog8_lib.func_sort_w")
                 }
                 DataType.ARRAY_F -> throw AssemblyError("sorting of floating point array is not supported")
                 else -> throw AssemblyError("weird type")
@@ -159,8 +269,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             DataType.UBYTE -> {
                 when (what) {
                     is ArrayIndexedExpression -> {
-                        asmgen.translateExpression(what.arrayvar)
-                        asmgen.translateExpression(what.indexer)
+                        translateRolRorArrayArgs(what.arrayvar, what.indexer, "ror2", 'b')
                         asmgen.out("  jsr  prog8_lib.ror2_array_ub")
                     }
                     is DirectMemoryRead -> {
@@ -168,7 +277,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
                             val number = (what.addressExpression as NumericLiteralValue).number
                             asmgen.out("  lda  ${number.toHex()} |  lsr  a |  bcc  + |  ora  #\$80 |+  |  sta  ${number.toHex()}")
                         } else {
-                            asmgen.translateExpression(what.addressExpression)
+                            asmgen.assignExpressionToRegister(what.addressExpression, RegisterOrPair.AY)
                             asmgen.out("  jsr  prog8_lib.ror2_mem_ub")
                         }
                     }
@@ -182,8 +291,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             DataType.UWORD -> {
                 when (what) {
                     is ArrayIndexedExpression -> {
-                        asmgen.translateExpression(what.arrayvar)
-                        asmgen.translateExpression(what.indexer)
+                        translateRolRorArrayArgs(what.arrayvar, what.indexer, "ror2", 'w')
                         asmgen.out("  jsr  prog8_lib.ror2_array_uw")
                     }
                     is IdentifierReference -> {
@@ -204,8 +312,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             DataType.UBYTE -> {
                 when (what) {
                     is ArrayIndexedExpression -> {
-                        asmgen.translateExpression(what.arrayvar)
-                        asmgen.translateExpression(what.indexer)
+                        translateRolRorArrayArgs(what.arrayvar, what.indexer, "ror", 'b')
                         asmgen.out("  jsr  prog8_lib.ror_array_ub")
                     }
                     is DirectMemoryRead -> {
@@ -213,15 +320,11 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
                             val number = (what.addressExpression as NumericLiteralValue).number
                             asmgen.out("  ror  ${number.toHex()}")
                         } else {
-                            asmgen.translateExpression(what.addressExpression)
+                            asmgen.assignExpressionToRegister(what.addressExpression, RegisterOrPair.AY)
                             asmgen.out("""
-                        inx
-                        lda  P8ESTACK_LO,x
-                        sta  (+) + 1
-                        lda  P8ESTACK_HI,x
-                        sta  (+) + 2
-    +                   ror  ${'$'}ffff            ; modified                    
-                                        """)
+                                sta  (+) + 1
+                                sty  (+) + 2
++                               ror  ${'$'}ffff            ; modified""")
                         }
                     }
                     is IdentifierReference -> {
@@ -234,8 +337,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             DataType.UWORD -> {
                 when (what) {
                     is ArrayIndexedExpression -> {
-                        asmgen.translateExpression(what.arrayvar)
-                        asmgen.translateExpression(what.indexer)
+                        translateRolRorArrayArgs(what.arrayvar, what.indexer, "ror", 'w')
                         asmgen.out("  jsr  prog8_lib.ror_array_uw")
                     }
                     is IdentifierReference -> {
@@ -256,8 +358,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             DataType.UBYTE -> {
                 when (what) {
                     is ArrayIndexedExpression -> {
-                        asmgen.translateExpression(what.arrayvar)
-                        asmgen.translateExpression(what.indexer)
+                        translateRolRorArrayArgs(what.arrayvar, what.indexer, "rol2", 'b')
                         asmgen.out("  jsr  prog8_lib.rol2_array_ub")
                     }
                     is DirectMemoryRead -> {
@@ -265,7 +366,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
                             val number = (what.addressExpression as NumericLiteralValue).number
                             asmgen.out("  lda  ${number.toHex()} |  cmp  #\$80 |  rol  a |  sta  ${number.toHex()}")
                         } else {
-                            asmgen.translateExpression(what.addressExpression)
+                            asmgen.assignExpressionToRegister(what.addressExpression, RegisterOrPair.AY)
                             asmgen.out("  jsr  prog8_lib.rol2_mem_ub")
                         }
                     }
@@ -279,8 +380,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             DataType.UWORD -> {
                 when (what) {
                     is ArrayIndexedExpression -> {
-                        asmgen.translateExpression(what.arrayvar)
-                        asmgen.translateExpression(what.indexer)
+                        translateRolRorArrayArgs(what.arrayvar, what.indexer, "rol2", 'w')
                         asmgen.out("  jsr  prog8_lib.rol2_array_uw")
                     }
                     is IdentifierReference -> {
@@ -301,8 +401,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             DataType.UBYTE -> {
                 when (what) {
                     is ArrayIndexedExpression -> {
-                        asmgen.translateExpression(what.arrayvar)
-                        asmgen.translateExpression(what.indexer)
+                        translateRolRorArrayArgs(what.arrayvar, what.indexer, "rol", 'b')
                         asmgen.out("  jsr  prog8_lib.rol_array_ub")
                     }
                     is DirectMemoryRead -> {
@@ -310,15 +409,11 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
                             val number = (what.addressExpression as NumericLiteralValue).number
                             asmgen.out("  rol  ${number.toHex()}")
                         } else {
-                            asmgen.translateExpression(what.addressExpression)
+                            asmgen.assignExpressionToRegister(what.addressExpression, RegisterOrPair.AY)
                             asmgen.out("""
-                        inx
-                        lda  P8ESTACK_LO,x
-                        sta  (+) + 1
-                        lda  P8ESTACK_HI,x
-                        sta  (+) + 2
-    +                   rol  ${'$'}ffff            ; modified                    
-                                        """)
+                                sta  (+) + 1
+                                sty  (+) + 2
++                               rol  ${'$'}ffff            ; modified""")
                         }
                     }
                     is IdentifierReference -> {
@@ -331,8 +426,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             DataType.UWORD -> {
                 when (what) {
                     is ArrayIndexedExpression -> {
-                        asmgen.translateExpression(what.arrayvar)
-                        asmgen.translateExpression(what.indexer)
+                        translateRolRorArrayArgs(what.arrayvar, what.indexer, "rol", 'w')
                         asmgen.out("  jsr  prog8_lib.rol_array_uw")
                     }
                     is IdentifierReference -> {
@@ -346,68 +440,133 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
         }
     }
 
-    private fun funcVariousFloatFuncs(fcall: IFunctionCall, func: FSignature, functionName: String) {
-        translateFunctionArguments(fcall.args, func)
-        asmgen.out("  jsr  floats.func_$functionName")
+    private fun translateRolRorArrayArgs(arrayvar: IdentifierReference, indexer: ArrayIndex, operation: String, dt: Char) {
+        asmgen.assignExpressionToVariable(AddressOf(arrayvar, arrayvar.position), "prog8_lib.${operation}_array_u${dt}._arg_target", DataType.UWORD, null)
+        val indexerExpr = if(indexer.indexVar!=null) indexer.indexVar!! else indexer.indexNum!!
+        asmgen.assignExpressionToVariable(indexerExpr, "prog8_lib.${operation}_array_u${dt}._arg_index", DataType.UBYTE, null)
     }
 
-    private fun funcSgn(fcall: IFunctionCall, func: FSignature) {
-        translateFunctionArguments(fcall.args, func)
+    private fun funcVariousFloatFuncs(fcall: IFunctionCall, func: FSignature, resultToStack: Boolean, scope: Subroutine) {
+        translateArguments(fcall.args, func, scope)
+        if(resultToStack)
+            asmgen.out("  jsr  floats.func_${func.name}_stack")
+        else
+            asmgen.out("  jsr  floats.func_${func.name}_fac1")
+    }
+
+    private fun funcSgn(fcall: IFunctionCall, func: FSignature, resultToStack: Boolean, scope: Subroutine) {
+        translateArguments(fcall.args, func, scope)
         val dt = fcall.args.single().inferType(program)
-        when (dt.typeOrElse(DataType.STRUCT)) {
-            DataType.UBYTE -> asmgen.out("  jsr  math.sign_ub")
-            DataType.BYTE -> asmgen.out("  jsr  math.sign_b")
-            DataType.UWORD -> asmgen.out("  jsr  math.sign_uw")
-            DataType.WORD -> asmgen.out("  jsr  math.sign_w")
-            DataType.FLOAT -> asmgen.out("  jsr  floats.sign_f")
-            else -> throw AssemblyError("weird type $dt")
+        if(resultToStack) {
+            when (dt.typeOrElse(DataType.STRUCT)) {
+                DataType.UBYTE -> asmgen.out("  jsr  prog8_lib.func_sign_ub_stack")
+                DataType.BYTE -> asmgen.out("  jsr  prog8_lib.func_sign_b_stack")
+                DataType.UWORD -> asmgen.out("  jsr  prog8_lib.func_sign_uw_stack")
+                DataType.WORD -> asmgen.out("  jsr  prog8_lib.func_sign_w_stack")
+                DataType.FLOAT -> asmgen.out("  jsr  floats.func_sign_f_stack")
+                else -> throw AssemblyError("weird type $dt")
+            }
+        } else {
+            when (dt.typeOrElse(DataType.STRUCT)) {
+                DataType.UBYTE -> asmgen.out("  jsr  prog8_lib.func_sign_ub_into_A")
+                DataType.BYTE -> asmgen.out("  jsr  prog8_lib.func_sign_b_into_A")
+                DataType.UWORD -> asmgen.out("  jsr  prog8_lib.func_sign_uw_into_A")
+                DataType.WORD -> asmgen.out("  jsr  prog8_lib.func_sign_w_into_A")
+                DataType.FLOAT -> asmgen.out("  jsr  floats.func_sign_f_into_A")
+                else -> throw AssemblyError("weird type $dt")
+            }
         }
     }
 
-    private fun funcAnyAll(fcall: IFunctionCall, functionName: String) {
-        outputPushAddressAndLenghtOfArray(fcall.args[0])
+    private fun funcAnyAll(fcall: IFunctionCall, function: FSignature, resultToStack: Boolean) {
+        outputAddressAndLenghtOfArray(fcall.args[0])
         val dt = fcall.args.single().inferType(program)
-        when (dt.typeOrElse(DataType.STRUCT)) {
-            DataType.ARRAY_B, DataType.ARRAY_UB, DataType.STR -> asmgen.out("  jsr  prog8_lib.func_${functionName}_b")
-            DataType.ARRAY_UW, DataType.ARRAY_W -> asmgen.out("  jsr  prog8_lib.func_${functionName}_w")
-            DataType.ARRAY_F -> asmgen.out("  jsr  floats.func_${functionName}_f")
-            else -> throw AssemblyError("weird type $dt")
+        if(resultToStack) {
+            when (dt.typeOrElse(DataType.STRUCT)) {
+                DataType.ARRAY_B, DataType.ARRAY_UB, DataType.STR -> asmgen.out("  jsr  prog8_lib.func_${function.name}_b_stack")
+                DataType.ARRAY_UW, DataType.ARRAY_W -> asmgen.out("  jsr  prog8_lib.func_${function.name}_w_stack")
+                DataType.ARRAY_F -> asmgen.out("  jsr  floats.func_${function.name}_f_stack")
+                else -> throw AssemblyError("weird type $dt")
+            }
+        } else {
+            when (dt.typeOrElse(DataType.STRUCT)) {
+                DataType.ARRAY_B, DataType.ARRAY_UB, DataType.STR -> asmgen.out("  jsr  prog8_lib.func_${function.name}_b_into_A")
+                DataType.ARRAY_UW, DataType.ARRAY_W -> asmgen.out("  jsr  prog8_lib.func_${function.name}_w_into_A")
+                DataType.ARRAY_F -> asmgen.out("  jsr  floats.func_${function.name}_f_into_A")
+                else -> throw AssemblyError("weird type $dt")
+            }
         }
     }
 
-    private fun funcMinMaxSum(fcall: IFunctionCall, functionName: String) {
-        outputPushAddressAndLenghtOfArray(fcall.args[0])
+    private fun funcMinMax(fcall: IFunctionCall, function: FSignature, resultToStack: Boolean) {
+        outputAddressAndLenghtOfArray(fcall.args[0])
         val dt = fcall.args.single().inferType(program)
-        when (dt.typeOrElse(DataType.STRUCT)) {
-            DataType.ARRAY_UB, DataType.STR -> asmgen.out("  jsr  prog8_lib.func_${functionName}_ub")
-            DataType.ARRAY_B -> asmgen.out("  jsr  prog8_lib.func_${functionName}_b")
-            DataType.ARRAY_UW -> asmgen.out("  jsr  prog8_lib.func_${functionName}_uw")
-            DataType.ARRAY_W -> asmgen.out("  jsr  prog8_lib.func_${functionName}_w")
-            DataType.ARRAY_F -> asmgen.out("  jsr  floats.func_${functionName}_f")
-            else -> throw AssemblyError("weird type $dt")
+        if(resultToStack) {
+            when (dt.typeOrElse(DataType.STRUCT)) {
+                DataType.ARRAY_UB, DataType.STR -> asmgen.out("  jsr  prog8_lib.func_${function.name}_ub_stack")
+                DataType.ARRAY_B -> asmgen.out("  jsr  prog8_lib.func_${function.name}_b_stack")
+                DataType.ARRAY_UW -> asmgen.out("  jsr  prog8_lib.func_${function.name}_uw_stack")
+                DataType.ARRAY_W -> asmgen.out("  jsr  prog8_lib.func_${function.name}_w_stack")
+                DataType.ARRAY_F -> asmgen.out("  jsr  floats.func_${function.name}_f_stack")
+                else -> throw AssemblyError("weird type $dt")
+            }
+        } else {
+            when (dt.typeOrElse(DataType.STRUCT)) {
+                DataType.ARRAY_UB, DataType.STR -> asmgen.out("  jsr  prog8_lib.func_${function.name}_ub_into_A")
+                DataType.ARRAY_B -> asmgen.out("  jsr  prog8_lib.func_${function.name}_b_into_A")
+                DataType.ARRAY_UW -> asmgen.out("  jsr  prog8_lib.func_${function.name}_uw_into_AY")
+                DataType.ARRAY_W -> asmgen.out("  jsr  prog8_lib.func_${function.name}_w_into_AY")
+                DataType.ARRAY_F -> asmgen.out("  jsr  floats.func_${function.name}_f_fac1")
+                else -> throw AssemblyError("weird type $dt")
+            }
         }
     }
 
-    private fun funcStrlen(fcall: IFunctionCall) {
-        val name = asmgen.asmVariableName(fcall.args[0] as IdentifierReference)
-        val type = fcall.args[0].inferType(program)
-        when {
-            type.istype(DataType.STR) -> asmgen.out("""
-                lda  #<$name
-                ldy  #>$name
-                jsr  prog8_lib.strlen
-                sta  P8ESTACK_LO,x
-                dex""")
-            type.istype(DataType.UWORD) -> asmgen.out("""
-                lda  $name
-                ldy  $name+1
-                jsr  prog8_lib.strlen
-                sta  P8ESTACK_LO,x
-                dex""")
-            else -> throw AssemblyError("strlen requires str or uword arg")
+    private fun funcSum(fcall: IFunctionCall, resultToStack: Boolean) {
+        outputAddressAndLenghtOfArray(fcall.args[0])
+        val dt = fcall.args.single().inferType(program)
+        if(resultToStack) {
+            when (dt.typeOrElse(DataType.STRUCT)) {
+                DataType.ARRAY_UB, DataType.STR -> asmgen.out("  jsr  prog8_lib.func_sum_ub_stack")
+                DataType.ARRAY_B -> asmgen.out("  jsr  prog8_lib.func_sum_b_stack")
+                DataType.ARRAY_UW -> asmgen.out("  jsr  prog8_lib.func_sum_uw_stack")
+                DataType.ARRAY_W -> asmgen.out("  jsr  prog8_lib.func_sum_w_stack")
+                DataType.ARRAY_F -> asmgen.out("  jsr  floats.func_sum_f_stack")
+                else -> throw AssemblyError("weird type $dt")
+            }
+        } else {
+            when (dt.typeOrElse(DataType.STRUCT)) {
+                DataType.ARRAY_UB, DataType.STR -> asmgen.out("  jsr  prog8_lib.func_sum_ub_into_AY")
+                DataType.ARRAY_B -> asmgen.out("  jsr  prog8_lib.func_sum_b_into_AY")
+                DataType.ARRAY_UW -> asmgen.out("  jsr  prog8_lib.func_sum_uw_into_AY")
+                DataType.ARRAY_W -> asmgen.out("  jsr  prog8_lib.func_sum_w_into_AY")
+                DataType.ARRAY_F -> asmgen.out("  jsr  floats.func_sum_f_fac1")
+                else -> throw AssemblyError("weird type $dt")
+            }
         }
     }
 
+    private fun funcStrlen(fcall: IFunctionCall, resultToStack: Boolean) {
+        if (fcall.args[0] is IdentifierReference) {
+            // use the address of the variable
+            val name = asmgen.asmVariableName(fcall.args[0] as IdentifierReference)
+            val type = fcall.args[0].inferType(program)
+            when {
+                type.istype(DataType.STR) -> asmgen.out("  lda  #<$name |  ldy  #>$name")
+                type.istype(DataType.UWORD) -> asmgen.out("  lda  $name |  ldy  $name+1")
+                else -> throw AssemblyError("strlen requires str or uword arg")
+            }
+        }
+        else {
+            // use the expression value as address of the string
+            asmgen.assignExpressionToRegister(fcall.args[0], RegisterOrPair.AY)
+        }
+        if (resultToStack)
+            asmgen.out("  jsr  prog8_lib.func_strlen_stack")
+        else
+            asmgen.out("  jsr  prog8_lib.func_strlen_into_A")
+    }
+
     private fun funcSwap(fcall: IFunctionCall) {
         val first = fcall.args[0]
         val second = fcall.args[1]
@@ -444,7 +603,7 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
                     sta  P8ZP_SCRATCH_W2
                     lda  #>$secondName
                     sta  P8ZP_SCRATCH_W2+1
-                    jsr  floats.swap_floats
+                    jsr  floats.func_swap_f
                 """)
                 return
             }
@@ -480,7 +639,10 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
         if(first is ArrayIndexedExpression && second is ArrayIndexedExpression) {
             val arrayVarName1 = asmgen.asmVariableName(first.arrayvar)
             val arrayVarName2 = asmgen.asmVariableName(second.arrayvar)
-            val elementDt = first.inferType(program).typeOrElse(DataType.STRUCT)
+            val elementIDt = first.inferType(program)
+            if(!elementIDt.isKnown)
+                throw AssemblyError("unknown dt")
+            val elementDt = elementIDt.typeOrElse(DataType.STRUCT)
 
             val firstNum = first.indexer.indexNum
             val firstVar = first.indexer.indexVar
@@ -502,7 +664,8 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             }
         }
 
-        // all other types of swap() calls are done via the evaluation stack
+        // all other types of swap() calls are done via a temporary variable
+
         fun targetFromExpr(expr: Expression, datatype: DataType): AsmAssignTarget {
             return when (expr) {
                 is IdentifierReference -> AsmAssignTarget(TargetStorageKind.VARIABLE, program, asmgen, datatype, expr.definingSubroutine(), variableAsmName = asmgen.asmVariableName(expr))
@@ -512,21 +675,43 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             }
         }
 
-        asmgen.translateExpression(first)
-        asmgen.translateExpression(second)
         val datatype = first.inferType(program).typeOrElse(DataType.STRUCT)
-        val assignFirst = AsmAssignment(
-                AsmAssignSource(SourceStorageKind.STACK, program, asmgen, datatype),
-                targetFromExpr(first, datatype),
-                false, first.position
-        )
-        val assignSecond = AsmAssignment(
-                AsmAssignSource(SourceStorageKind.STACK, program, asmgen, datatype),
-                targetFromExpr(second, datatype),
-                false, second.position
-        )
-        asmgen.translateNormalAssignment(assignFirst)
-        asmgen.translateNormalAssignment(assignSecond)
+        when(datatype) {
+            in ByteDatatypes, in WordDatatypes -> {
+                asmgen.assignExpressionToVariable(first, "P8ZP_SCRATCH_W1", datatype, null)
+                asmgen.assignExpressionToVariable(second, "P8ZP_SCRATCH_W2", datatype, null)
+                val assignFirst = AsmAssignment(
+                        AsmAssignSource(SourceStorageKind.VARIABLE, program, asmgen, datatype, variableAsmName = "P8ZP_SCRATCH_W2"),
+                        targetFromExpr(first, datatype),
+                        false, first.position
+                )
+                val assignSecond = AsmAssignment(
+                        AsmAssignSource(SourceStorageKind.VARIABLE, program, asmgen, datatype, variableAsmName = "P8ZP_SCRATCH_W1"),
+                        targetFromExpr(second, datatype),
+                        false, second.position
+                )
+                asmgen.translateNormalAssignment(assignFirst)
+                asmgen.translateNormalAssignment(assignSecond)
+            }
+            DataType.FLOAT -> {
+                // via evaluation stack
+                asmgen.translateExpression(first)
+                asmgen.translateExpression(second)
+                val assignFirst = AsmAssignment(
+                        AsmAssignSource(SourceStorageKind.STACK, program, asmgen, DataType.FLOAT),
+                        targetFromExpr(first, datatype),
+                        false, first.position
+                )
+                val assignSecond = AsmAssignment(
+                        AsmAssignSource(SourceStorageKind.STACK, program, asmgen, DataType.FLOAT),
+                        targetFromExpr(second, datatype),
+                        false, second.position
+                )
+                asmgen.translateNormalAssignment(assignFirst)
+                asmgen.translateNormalAssignment(assignSecond)
+            }
+            else -> throw AssemblyError("weird swap dt")
+        }
     }
 
     private fun swapArrayValues(elementDt: DataType, arrayVarName1: String, indexValue1: NumericLiteralValue, arrayVarName2: String, indexValue2: NumericLiteralValue) {
@@ -759,25 +944,52 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
         }
     }
 
-    private fun funcAbs(fcall: IFunctionCall, func: FSignature) {
-        translateFunctionArguments(fcall.args, func)
-        val dt = fcall.args.single().inferType(program)
-        when (dt.typeOrElse(DataType.STRUCT)) {
-            in ByteDatatypes -> asmgen.out("  jsr  prog8_lib.abs_b")
-            in WordDatatypes -> asmgen.out("  jsr  prog8_lib.abs_w")
-            DataType.FLOAT -> asmgen.out("  jsr  floats.abs_f")
-            else -> throw AssemblyError("weird type")
+    private fun funcAbs(fcall: IFunctionCall, func: FSignature, resultToStack: Boolean, scope: Subroutine) {
+        translateArguments(fcall.args, func, scope)
+        val dt = fcall.args.single().inferType(program).typeOrElse(DataType.STRUCT)
+        if(resultToStack) {
+            when (dt) {
+                in ByteDatatypes -> asmgen.out("  jsr  prog8_lib.abs_b_stack")
+                in WordDatatypes -> asmgen.out("  jsr  prog8_lib.abs_w_stack")
+                DataType.FLOAT -> asmgen.out("  jsr  floats.abs_f_stack")
+                else -> throw AssemblyError("weird type")
+            }
+        } else {
+            when (dt) {
+                in ByteDatatypes -> asmgen.out("  jsr  prog8_lib.abs_b_into_A")
+                in WordDatatypes -> asmgen.out("  jsr  prog8_lib.abs_w_into_AY")
+                DataType.FLOAT -> asmgen.out("  jsr  floats.abs_f_fac1")
+                else -> throw AssemblyError("weird type")
+            }
         }
     }
 
-    private fun funcMkword(fcall: IFunctionCall, func: FSignature) {
-        // trick: push the args in reverse order (msb first, lsb second) this saves some instructions
-        asmgen.translateExpression(fcall.args[1])
-        asmgen.translateExpression(fcall.args[0])
-        asmgen.out("  inx | lda  P8ESTACK_LO,x  | sta  P8ESTACK_HI+1,x")
+    private fun funcRnd(func: FSignature, resultToStack: Boolean) {
+        when(func.name) {
+            "rnd" -> {
+                if(resultToStack)
+                    asmgen.out("  jsr  prog8_lib.func_rnd_stack")
+                else
+                    asmgen.out("  jsr  math.randbyte")
+            }
+            "rndw" -> {
+                if(resultToStack)
+                    asmgen.out("  jsr  prog8_lib.func_rndw_stack")
+                else
+                    asmgen.out("  jsr  math.randword")
+            }
+            else -> throw AssemblyError("wrong func")
+        }
     }
 
-    private fun funcMsb(fcall: IFunctionCall) {
+    private fun funcMkword(fcall: IFunctionCall, resultToStack: Boolean) {
+        asmgen.assignExpressionToRegister(fcall.args[0], RegisterOrPair.Y)      // msb
+        asmgen.assignExpressionToRegister(fcall.args[1], RegisterOrPair.A)      // lsb
+        if(resultToStack)
+            asmgen.out("  sta  P8ESTACK_LO,x |  tya |  sta  P8ESTACK_HI,x |  dex")
+    }
+
+    private fun funcMsb(fcall: IFunctionCall, resultToStack: Boolean) {
         val arg = fcall.args.single()
         if (arg.inferType(program).typeOrElse(DataType.STRUCT) !in WordDatatypes)
             throw AssemblyError("msb required word argument")
@@ -785,47 +997,114 @@ internal class BuiltinFunctionsAsmGen(private val program: Program, private val
             throw AssemblyError("msb(const) should have been const-folded away")
         if (arg is IdentifierReference) {
             val sourceName = asmgen.asmVariableName(arg)
-            asmgen.out("  lda  $sourceName+1 |  sta  P8ESTACK_LO,x |  dex")
+            asmgen.out("  lda  $sourceName+1")
+            if (resultToStack)
+                asmgen.out("  sta  P8ESTACK_LO,x |  dex")
         } else {
-            asmgen.translateExpression(arg)
-            asmgen.out("  lda  P8ESTACK_HI+1,x |  sta  P8ESTACK_LO+1,x")
+            asmgen.assignExpressionToRegister(fcall.args.single(), RegisterOrPair.AY)
+            if (resultToStack)
+                asmgen.out("  tya |  sta  P8ESTACK_LO,x |  dex")
+            else
+                asmgen.out("  tya")
         }
     }
 
-    private fun funcLsb(fcall: IFunctionCall) {
+    private fun funcLsb(fcall: IFunctionCall, resultToStack: Boolean) {
         val arg = fcall.args.single()
         if (arg.inferType(program).typeOrElse(DataType.STRUCT) !in WordDatatypes)
             throw AssemblyError("lsb required word argument")
         if (arg is NumericLiteralValue)
             throw AssemblyError("lsb(const) should have been const-folded away")
+
         if (arg is IdentifierReference) {
             val sourceName = asmgen.asmVariableName(arg)
-            asmgen.out("  lda  $sourceName |  sta  P8ESTACK_LO,x |  dex")
+            asmgen.out("  lda  $sourceName")
+            if (resultToStack)
+                asmgen.out("  sta  P8ESTACK_LO,x |  dex")
         } else {
-            asmgen.translateExpression(arg)
-            // just ignore any high-byte
+            asmgen.assignExpressionToRegister(fcall.args.single(), RegisterOrPair.AY)
+            if (resultToStack)
+                asmgen.out("  sta  P8ESTACK_LO,x |  dex")
         }
     }
 
-    private fun outputPushAddressAndLenghtOfArray(arg: Expression) {
+    private fun outputAddressAndLenghtOfArray(arg: Expression) {
+        // address in P8ZP_SCRATCH_W1,  number of elements in A
         arg as IdentifierReference
         val identifierName = asmgen.asmVariableName(arg)
         val size = arg.targetVarDecl(program.namespace)!!.arraysize!!.constIndex()!!
         asmgen.out("""
                     lda  #<$identifierName
-                    sta  P8ESTACK_LO,x
-                    lda  #>$identifierName
-                    sta  P8ESTACK_HI,x
-                    dex
+                    ldy  #>$identifierName
+                    sta  P8ZP_SCRATCH_W1
+                    sty  P8ZP_SCRATCH_W1+1
                     lda  #$size
-                    sta  P8ESTACK_LO,x
-                    dex
                     """)
     }
 
-    private fun translateFunctionArguments(args: MutableList<Expression>, signature: FSignature) {
-        args.forEach {
-            asmgen.translateExpression(it)
+    private fun translateArguments(args: MutableList<Expression>, signature: FSignature, scope: Subroutine) {
+        val callConv = signature.callConvention(args.map { it.inferType(program).typeOrElse(DataType.STRUCT) })
+
+        fun getSourceForFloat(value: Expression): AsmAssignSource {
+            return when (value) {
+                is IdentifierReference -> {
+                    val addr = AddressOf(value, value.position)
+                    AsmAssignSource.fromAstSource(addr, program, asmgen)
+                }
+                is NumericLiteralValue -> {
+                    throw AssemblyError("float literals should have been converted into autovar")
+                }
+                else -> {
+                    scope.asmGenInfo.usedFloatEvalResultVar2 = true
+                    val variable = IdentifierReference(listOf(subroutineFloatEvalResultVar2), value.position)
+                    val addr = AddressOf(variable, value.position)
+                    addr.linkParents(value)
+                    asmgen.assignExpressionToVariable(value, asmgen.asmVariableName(variable), DataType.FLOAT, scope)
+                    AsmAssignSource.fromAstSource(addr, program, asmgen)
+                }
+            }
+        }
+
+        args.zip(callConv.params).zip(signature.parameters).forEach {
+            val paramName = it.second.name
+            val conv = it.first.second
+            val value = it.first.first
+            when {
+                conv.variable -> {
+                    val varname = "prog8_lib.func_${signature.name}._arg_${paramName}"
+                    val src = when (conv.dt) {
+                        DataType.FLOAT -> getSourceForFloat(value)
+                        in PassByReferenceDatatypes -> {
+                            // put the address of the argument in AY
+                            val addr = AddressOf(value as IdentifierReference, value.position)
+                            AsmAssignSource.fromAstSource(addr, program, asmgen)
+                        }
+                        else -> {
+                            AsmAssignSource.fromAstSource(value, program, asmgen)
+                        }
+                    }
+                    val tgt = AsmAssignTarget(TargetStorageKind.VARIABLE, program, asmgen, conv.dt, null, variableAsmName = varname)
+                    val assign = AsmAssignment(src, tgt, false, value.position)
+                    asmgen.translateNormalAssignment(assign)
+                }
+                conv.reg != null -> {
+                    val src = when (conv.dt) {
+                        DataType.FLOAT -> getSourceForFloat(value)
+                        in PassByReferenceDatatypes -> {
+                            // put the address of the argument in AY
+                            val addr = AddressOf(value as IdentifierReference, value.position)
+                            AsmAssignSource.fromAstSource(addr, program, asmgen)
+                        }
+                        else -> {
+                            AsmAssignSource.fromAstSource(value, program, asmgen)
+                        }
+                    }
+                    val tgt = AsmAssignTarget.fromRegisters(conv.reg, null, program, asmgen)
+                    val assign = AsmAssignment(src, tgt, false, value.position)
+                    asmgen.translateNormalAssignment(assign)
+                }
+                else -> throw AssemblyError("callconv")
+            }
         }
     }
 

compiler/src/prog8/compiler/target/c64/codegen/ForLoopsAsmGen.kt

@@ -1,15 +1,13 @@
 package prog8.compiler.target.c64.codegen
 
 import prog8.ast.Program
+import prog8.ast.base.ArrayElementTypes
 import prog8.ast.base.DataType
+import prog8.ast.base.RegisterOrPair
 import prog8.ast.expressions.IdentifierReference
 import prog8.ast.expressions.RangeExpr
 import prog8.ast.statements.ForLoop
 import prog8.compiler.AssemblyError
-import prog8.compiler.target.c64.codegen.assignment.AsmAssignSource
-import prog8.compiler.target.c64.codegen.assignment.AsmAssignTarget
-import prog8.compiler.target.c64.codegen.assignment.AsmAssignment
-import prog8.compiler.target.c64.codegen.assignment.TargetStorageKind
 import prog8.compiler.toHex
 import kotlin.math.absoluteValue
 
@@ -18,7 +16,7 @@ internal class ForLoopsAsmGen(private val program: Program, private val asmgen:
     internal fun translate(stmt: ForLoop) {
         val iterableDt = stmt.iterable.inferType(program)
         if(!iterableDt.isKnown)
-            throw AssemblyError("can't determine iterable dt")
+            throw AssemblyError("unknown dt")
         when(stmt.iterable) {
             is RangeExpr -> {
                 val range = (stmt.iterable as RangeExpr).toConstantIntegerRange()
@@ -51,23 +49,17 @@ internal class ForLoopsAsmGen(private val program: Program, private val asmgen:
                     val incdec = if(stepsize==1) "inc" else "dec"
                     // loop over byte range via loopvar
                     val varname = asmgen.asmVariableName(stmt.loopVar)
-                    asmgen.translateExpression(range.to)
-                    asmgen.translateExpression(range.from)
+                    asmgen.assignExpressionToVariable(range.from, varname, ArrayElementTypes.getValue(iterableDt), null)
+                    asmgen.assignExpressionToVariable(range.to, "$modifiedLabel+1", ArrayElementTypes.getValue(iterableDt), null)
+                    asmgen.out(loopLabel)
+                    asmgen.translate(stmt.body)
                     asmgen.out("""
-                inx
-                lda  P8ESTACK_LO,x
-                sta  $varname
-                lda  P8ESTACK_LO+1,x
-                sta  $modifiedLabel+1
-$loopLabel""")
-                        asmgen.translate(stmt.body)
-                        asmgen.out("""
-                lda  $varname
-$modifiedLabel  cmp  #0         ; modified 
-                beq  $endLabel
-                $incdec  $varname
-                jmp  $loopLabel
-$endLabel       inx""")
+                        lda  $varname
+$modifiedLabel          cmp  #0         ; modified 
+                        beq  $endLabel
+                        $incdec  $varname
+                        jmp  $loopLabel
+$endLabel""")
 
                 } else {
 
@@ -75,36 +67,29 @@ $endLabel       inx""")
 
                     // loop over byte range via loopvar
                     val varname = asmgen.asmVariableName(stmt.loopVar)
-                    asmgen.translateExpression(range.to)
-                    asmgen.translateExpression(range.from)
-                    asmgen.out("""
-                inx
-                lda  P8ESTACK_LO,x
-                sta  $varname
-                lda  P8ESTACK_LO+1,x
-                sta  $modifiedLabel+1
-$loopLabel""")
+                    asmgen.assignExpressionToVariable(range.from, varname, ArrayElementTypes.getValue(iterableDt), null)
+                    asmgen.assignExpressionToVariable(range.to, "$modifiedLabel+1", ArrayElementTypes.getValue(iterableDt), null)
+                    asmgen.out(loopLabel)
                     asmgen.translate(stmt.body)
-                    asmgen.out("""
-                lda  $varname""")
                     if(stepsize>0) {
                         asmgen.out("""
-                clc
-                adc  #$stepsize
-                sta  $varname
-$modifiedLabel  cmp  #0    ; modified
-                bcc  $loopLabel
-                beq  $loopLabel""")
+                            lda  $varname
+                            clc
+                            adc  #$stepsize
+                            sta  $varname
+$modifiedLabel              cmp  #0    ; modified
+                            bmi  $loopLabel
+                            beq  $loopLabel""")
                     } else {
                         asmgen.out("""
-                sec
-                sbc  #${stepsize.absoluteValue}
-                sta  $varname
-$modifiedLabel  cmp  #0     ; modified 
-                bcs  $loopLabel""")
+                            lda  $varname
+                            sec
+                            sbc  #${stepsize.absoluteValue}
+                            sta  $varname
+$modifiedLabel              cmp  #0     ; modified
+                            bpl  $loopLabel""")
                     }
-                    asmgen.out("""
-$endLabel       inx""")
+                    asmgen.out(endLabel)
                 }
             }
             DataType.ARRAY_W, DataType.ARRAY_UW -> {
@@ -113,13 +98,11 @@ $endLabel       inx""")
                     // words, step 1 or -1
 
                     stepsize == 1 || stepsize == -1 -> {
-                        asmgen.translateExpression(range.to)
-                        assignLoopvar(stmt, range)
                         val varname = asmgen.asmVariableName(stmt.loopVar)
+                        assignLoopvar(stmt, range)
+                        asmgen.assignExpressionToRegister(range.to, RegisterOrPair.AY)
                         asmgen.out("""
-                            lda  P8ESTACK_HI+1,x
-                            sta  $modifiedLabel+1
-                            lda  P8ESTACK_LO+1,x
+                            sty  $modifiedLabel+1
                             sta  $modifiedLabel2+1
 $loopLabel""")
                         asmgen.translate(stmt.body)
@@ -146,21 +129,17 @@ $modifiedLabel2 cmp  #0    ; modified
                 jmp  $loopLabel""")
                         }
                         asmgen.out(endLabel)
-                        asmgen.out(" inx")
                     }
                     stepsize > 0 -> {
 
                         // (u)words, step >= 2
-                        asmgen.translateExpression(range.to)
-                        asmgen.out("""
-                            lda  P8ESTACK_HI+1,x
-                            sta  $modifiedLabel+1
-                            lda  P8ESTACK_LO+1,x
-                            sta  $modifiedLabel2+1
-                        """)
-                        assignLoopvar(stmt, range)
                         val varname = asmgen.asmVariableName(stmt.loopVar)
-                        asmgen.out(loopLabel)
+                        assignLoopvar(stmt, range)
+                        asmgen.assignExpressionToRegister(range.to, RegisterOrPair.AY)
+                        asmgen.out("""
+                            sty  $modifiedLabel+1
+                            sta  $modifiedLabel2+1
+$loopLabel""")
                         asmgen.translate(stmt.body)
 
                         if (iterableDt == DataType.ARRAY_UW) {
@@ -179,7 +158,7 @@ $modifiedLabel2 lda  #0     ; modified
                 cmp  $varname
                 bcc  $endLabel
                 bcs  $loopLabel
-$endLabel       inx""")
+$endLabel""")
                         } else {
                             asmgen.out("""
                 lda  $varname
@@ -196,22 +175,19 @@ $modifiedLabel  lda  #0   ; modified
                 bvc  +
                 eor  #$80
 +               bpl  $loopLabel                
-$endLabel       inx""")
+$endLabel""")
                         }
                     }
                     else -> {
 
                         // (u)words, step <= -2
-                        asmgen.translateExpression(range.to)
-                        asmgen.out("""
-                            lda  P8ESTACK_HI+1,x
-                            sta  $modifiedLabel+1
-                            lda  P8ESTACK_LO+1,x
-                            sta  $modifiedLabel2+1
-                        """)
-                        assignLoopvar(stmt, range)
                         val varname = asmgen.asmVariableName(stmt.loopVar)
-                        asmgen.out(loopLabel)
+                        assignLoopvar(stmt, range)
+                        asmgen.assignExpressionToRegister(range.to, RegisterOrPair.AY)
+                        asmgen.out("""
+                            sty  $modifiedLabel+1
+                            sta  $modifiedLabel2+1
+$loopLabel""")
                         asmgen.translate(stmt.body)
 
                         if(iterableDt==DataType.ARRAY_UW) {
@@ -229,7 +205,7 @@ $modifiedLabel  cmp  #0    ; modified
                 lda  $varname
 $modifiedLabel2 cmp  #0    ; modified 
                 bcs  $loopLabel
-$endLabel       inx""")
+$endLabel""")
                         } else {
                             asmgen.out("""
                 lda  $varname
@@ -247,7 +223,7 @@ $modifiedLabel  sbc  #0    ; modified
                 bvc  +
                 eor  #$80
 +               bpl  $loopLabel                
-$endLabel       inx""")
+$endLabel""")
                         }
                     }
                 }
@@ -609,10 +585,6 @@ $endLabel""")
         asmgen.loopEndLabels.pop()
     }
 
-    private fun assignLoopvar(stmt: ForLoop, range: RangeExpr) {
-        val target = AsmAssignTarget(TargetStorageKind.VARIABLE, program, asmgen, stmt.loopVarDt(program).typeOrElse(DataType.STRUCT), stmt.definingSubroutine(), variableAsmName=asmgen.asmVariableName(stmt.loopVar))
-        val src = AsmAssignSource.fromAstSource(range.from, program, asmgen).adjustSignedUnsigned(target)
-        val assign = AsmAssignment(src, target, false, range.position)
-        asmgen.translateNormalAssignment(assign)
-    }
+    private fun assignLoopvar(stmt: ForLoop, range: RangeExpr) =
+        asmgen.assignExpressionToVariable(range.from, asmgen.asmVariableName(stmt.loopVar), stmt.loopVarDt(program).typeOrElse(DataType.STRUCT), stmt.definingSubroutine())
 }

compiler/src/prog8/compiler/target/c64/codegen/FunctionCallAsmGen.kt

@@ -14,13 +14,23 @@ import prog8.compiler.target.c64.codegen.assignment.*
 
 internal class FunctionCallAsmGen(private val program: Program, private val asmgen: AsmGen) {
 
+    internal fun translateFunctionCallStatement(stmt: IFunctionCall) {
+        val sub = stmt.target.targetSubroutine(program.namespace)!!
+        val preserveStatusRegisterAfterCall = sub.shouldPreserveStatusRegisterAfterCall()
+        translateFunctionCall(stmt, preserveStatusRegisterAfterCall)
+        // functioncalls no longer return results on the stack, so simply ignore the results in the registers
+        if(preserveStatusRegisterAfterCall)
+            asmgen.out("  plp")     // restore status flags from call
+    }
+
+
     internal fun translateFunctionCall(stmt: IFunctionCall, preserveStatusRegisterAfterCall: Boolean) {
         // output the code to setup the parameters and perform the actual call
         // does NOT output the code to deal with the result values!
         val sub = stmt.target.targetSubroutine(program.namespace) ?: throw AssemblyError("undefined subroutine ${stmt.target}")
-        val saveX = CpuRegister.X in sub.asmClobbers || sub.regXasResult() || sub.regXasParam()
+        val saveX = sub.shouldSaveX()
         if(saveX)
-            asmgen.saveRegister(CpuRegister.X, preserveStatusRegisterAfterCall, (stmt as Node).definingSubroutine())
+            asmgen.saveRegister(CpuRegister.X, preserveStatusRegisterAfterCall, (stmt as Node).definingSubroutine()!!)
 
         val subName = asmgen.asmSymbolName(stmt.target)
         if(stmt.args.isNotEmpty()) {
@@ -33,7 +43,7 @@ internal class FunctionCallAsmGen(private val program: Program, private val asmg
                 // via registers
                 if(sub.parameters.size==1) {
                     // just a single parameter, no risk of clobbering registers
-                    argumentViaRegister(sub, sub.parameters.withIndex().single(), stmt.args[0])
+                    argumentViaRegister(sub, IndexedValue(0, sub.parameters.single()), stmt.args[0])
                 } else {
                     // multiple register arguments, risk of register clobbering.
                     // evaluate arguments onto the stack, and load the registers from the evaluated values on the stack.
@@ -59,7 +69,7 @@ internal class FunctionCallAsmGen(private val program: Program, private val asmg
         asmgen.out("  jsr  $subName")
 
         if(preserveStatusRegisterAfterCall) {
-            asmgen.out("  php\t; save status flags from call")
+            asmgen.out("  php")    // save status flags from call
             // note: the containing statement (such as the FunctionCallStatement or the Assignment or the Expression)
             //       must take care of popping this value again at the end!
         }
@@ -87,7 +97,6 @@ internal class FunctionCallAsmGen(private val program: Program, private val asmg
 
         for(argi in stmt.args.zip(sub.asmParameterRegisters).withIndex()) {
             when {
-                argi.value.second.stack -> TODO("asmsub @stack parameter")
                 argi.value.second.statusflag == Statusflag.Pc -> {
                     require(argForCarry == null)
                     argForCarry = argi
@@ -150,23 +159,20 @@ internal class FunctionCallAsmGen(private val program: Program, private val asmg
         // pass parameter via a regular variable (not via registers)
         val valueIDt = value.inferType(program)
         if(!valueIDt.isKnown)
-            throw AssemblyError("arg type unknown")
+            throw AssemblyError("unknown dt")
         val valueDt = valueIDt.typeOrElse(DataType.STRUCT)
         if(!isArgumentTypeCompatible(valueDt, parameter.value.type))
             throw AssemblyError("argument type incompatible")
 
         val varName = asmgen.asmVariableName(sub.scopedname+"."+parameter.value.name)
-        val tgt = AsmAssignTarget(TargetStorageKind.VARIABLE, program, asmgen, parameter.value.type, sub, variableAsmName = varName)
-        val source = AsmAssignSource.fromAstSource(value, program, asmgen).adjustSignedUnsigned(tgt)
-        val asgn = AsmAssignment(source, tgt, false, Position.DUMMY)
-        asmgen.translateNormalAssignment(asgn)
+        asmgen.assignExpressionToVariable(value, varName, parameter.value.type, sub)
     }
 
     private fun argumentViaRegister(sub: Subroutine, parameter: IndexedValue<SubroutineParameter>, value: Expression) {
         // pass argument via a register parameter
         val valueIDt = value.inferType(program)
         if(!valueIDt.isKnown)
-            throw AssemblyError("arg type unknown")
+            throw AssemblyError("unknown dt")
         val valueDt = valueIDt.typeOrElse(DataType.STRUCT)
         if(!isArgumentTypeCompatible(valueDt, parameter.value.type))
             throw AssemblyError("argument type incompatible")
@@ -174,20 +180,12 @@ internal class FunctionCallAsmGen(private val program: Program, private val asmg
         val paramRegister = sub.asmParameterRegisters[parameter.index]
         val statusflag = paramRegister.statusflag
         val register = paramRegister.registerOrPair
-        val stack = paramRegister.stack
         val requiredDt = parameter.value.type
         if(requiredDt!=valueDt) {
             if(valueDt largerThan requiredDt)
                 throw AssemblyError("can only convert byte values to word param types")
         }
         when {
-            stack -> {
-                // push arg onto the stack
-                // note: argument order is reversed (first argument will be deepest on the stack)
-                asmgen.translateExpression(value)
-                if(requiredDt!=valueDt)
-                    asmgen.signExtendStackLsb(valueDt)
-            }
             statusflag!=null -> {
                 if(requiredDt!=valueDt)
                     throw AssemblyError("for statusflag, byte value is required")
@@ -212,17 +210,13 @@ internal class FunctionCallAsmGen(private val program: Program, private val asmg
 """)
                         }
                         else -> {
-                            asmgen.translateExpression(value)
+                            asmgen.assignExpressionToRegister(value, RegisterOrPair.A)
                             asmgen.out("""
-            inx
-            pha
-            lda  P8ESTACK_LO,x
-            beq  +
-            sec  
-            bcs  ++
-+           clc
-+           pla
-""")
+                                beq  +
+                                sec
+                                bcs  ++
++                               clc
++""")
                         }
                     }
                 }
@@ -230,18 +224,16 @@ internal class FunctionCallAsmGen(private val program: Program, private val asmg
             }
             else -> {
                 // via register or register pair
-                val target = AsmAssignTarget.fromRegisters(register!!, sub, program, asmgen)
+                register!!
                 if(requiredDt largerThan valueDt) {
                     // we need to sign extend the source, do this via temporary word variable
                     val scratchVar = asmgen.asmVariableName("P8ZP_SCRATCH_W1")
-                    val scratchTarget = AsmAssignTarget(TargetStorageKind.VARIABLE, program, asmgen, DataType.UBYTE, sub, scratchVar)
-                    val source = AsmAssignSource.fromAstSource(value, program, asmgen)
-                    asmgen.translateNormalAssignment(AsmAssignment(source, scratchTarget, false, value.position))
+                    asmgen.assignExpressionToVariable(value, scratchVar, DataType.UBYTE, sub)
                     asmgen.signExtendVariableLsb(scratchVar, valueDt)
-                    val src = AsmAssignSource(SourceStorageKind.VARIABLE, program, asmgen, DataType.UWORD, scratchVar)
-                    asmgen.translateNormalAssignment(AsmAssignment(src, target, false, Position.DUMMY))
+                    asmgen.assignVariableToRegister(scratchVar, register)
                 }
                 else {
+                    val target = AsmAssignTarget.fromRegisters(register, sub, program, asmgen)
                     val src = if(valueDt in PassByReferenceDatatypes) {
                         if(value is IdentifierReference) {
                             val addr = AddressOf(value, Position.DUMMY)

compiler/src/prog8/compiler/target/c64/codegen/PostIncrDecrAsmGen.kt

@@ -19,7 +19,7 @@ internal class PostIncrDecrAsmGen(private val program: Program, private val asmg
         when {
             targetIdent!=null -> {
                 val what = asmgen.asmVariableName(targetIdent)
-                when (stmt.target.inferType(program, stmt).typeOrElse(DataType.STRUCT)) {
+                when (stmt.target.inferType(program).typeOrElse(DataType.STRUCT)) {
                     in ByteDatatypes -> asmgen.out(if (incr) "  inc  $what" else "  dec  $what")
                     in WordDatatypes -> {
                         if(incr)
@@ -54,14 +54,8 @@ internal class PostIncrDecrAsmGen(private val program: Program, private val asmg
                             asmgen.out("+\tdec  ${'$'}ffff\t; modified")
                     }
                     else -> {
-                        asmgen.translateExpression(addressExpr)
-                        asmgen.out("""
-                            inx
-                            lda  P8ESTACK_LO,x
-                            sta  (+) + 1
-                            lda  P8ESTACK_HI,x
-                            sta  (+) + 2
-                        """)
+                        asmgen.assignExpressionToRegister(addressExpr, RegisterOrPair.AY)
+                        asmgen.out("  sta  (+) + 1 |  sty  (+) + 2")
                         if(incr)
                             asmgen.out("+\tinc  ${'$'}ffff\t; modified")
                         else
@@ -97,7 +91,7 @@ internal class PostIncrDecrAsmGen(private val program: Program, private val asmg
                 else
                 {
                     asmgen.loadScaledArrayIndexIntoRegister(targetArrayIdx, elementDt, CpuRegister.A)
-                    asmgen.saveRegister(CpuRegister.X, false, scope)
+                    asmgen.saveRegister(CpuRegister.X, false, scope!!)
                     asmgen.out("  tax")
                     when(elementDt) {
                         in ByteDatatypes -> {

compiler/src/prog8/compiler/target/c64/codegen/assignment/AsmAssignment.kt

@@ -3,10 +3,7 @@ package prog8.compiler.target.c64.codegen.assignment
 import prog8.ast.Program
 import prog8.ast.base.*
 import prog8.ast.expressions.*
-import prog8.ast.statements.AssignTarget
-import prog8.ast.statements.Assignment
-import prog8.ast.statements.DirectMemoryWrite
-import prog8.ast.statements.Subroutine
+import prog8.ast.statements.*
 import prog8.compiler.AssemblyError
 import prog8.compiler.target.c64.codegen.AsmGen
 
@@ -52,13 +49,16 @@ internal class AsmAssignTarget(val kind: TargetStorageKind,
     lateinit var origAssign: AsmAssignment
 
     init {
-        if(register!=null && datatype !in IntegerDatatypes)
-            throw AssemblyError("register must be integer type")
+        if(register!=null && datatype !in NumericDatatypes)
+            throw AssemblyError("register must be integer or float type")
     }
 
     companion object {
         fun fromAstAssignment(assign: Assignment, program: Program, asmgen: AsmGen): AsmAssignTarget = with(assign.target) {
-            val dt = inferType(program, assign).typeOrElse(DataType.STRUCT)
+            val idt = inferType(program)
+            if(!idt.isKnown)
+                throw AssemblyError("unknown dt")
+            val dt = idt.typeOrElse(DataType.STRUCT)
             when {
                 identifier != null -> AsmAssignTarget(TargetStorageKind.VARIABLE, program, asmgen, dt, assign.definingSubroutine(), variableAsmName = asmgen.asmVariableName(identifier!!), origAstTarget =  this)
                 arrayindexed != null -> AsmAssignTarget(TargetStorageKind.ARRAY, program, asmgen, dt, assign.definingSubroutine(), array = arrayindexed, origAstTarget =  this)
@@ -75,6 +75,8 @@ internal class AsmAssignTarget(val kind: TargetStorageKind,
                     RegisterOrPair.AX,
                     RegisterOrPair.AY,
                     RegisterOrPair.XY -> AsmAssignTarget(TargetStorageKind.REGISTER, program, asmgen, DataType.UWORD, scope, register = registers)
+                    RegisterOrPair.FAC1,
+                    RegisterOrPair.FAC2 -> AsmAssignTarget(TargetStorageKind.REGISTER, program, asmgen, DataType.FLOAT, scope, register = registers)
                 }
     }
 }
@@ -101,6 +103,14 @@ internal class AsmAssignSource(val kind: SourceStorageKind,
             asmgen.asmVariableName(array.arrayvar)
 
     companion object {
+        fun fromAstSource(indexer: ArrayIndex, program: Program, asmgen: AsmGen): AsmAssignSource {
+            return when {
+                indexer.indexNum!=null -> fromAstSource(indexer.indexNum!!, program, asmgen)
+                indexer.indexVar!=null -> fromAstSource(indexer.indexVar!!, program, asmgen)
+                else -> throw AssemblyError("weird indexer")
+            }
+        }
+
         fun fromAstSource(value: Expression, program: Program, asmgen: AsmGen): AsmAssignSource {
             val cv = value.constValue(program)
             if(cv!=null)
@@ -121,33 +131,30 @@ internal class AsmAssignSource(val kind: SourceStorageKind,
                     val dt = value.inferType(program).typeOrElse(DataType.STRUCT)
                     AsmAssignSource(SourceStorageKind.ARRAY, program, asmgen, dt, array = value)
                 }
-                else -> {
-                    if(value is FunctionCall) {
-                        // functioncall.
-                        val asmSub = value.target.targetStatement(program.namespace)
-                        if(asmSub is Subroutine && asmSub.isAsmSubroutine) {
-                            when (asmSub.asmReturnvaluesRegisters.count { rr -> rr.registerOrPair!=null }) {
-                                0 -> throw AssemblyError("can't translate zero return values in assignment")
-                                1 -> {
-                                    // assignment generation itself must make sure the status register is correct after the subroutine call, if status register is involved!
-                                    val reg = asmSub.asmReturnvaluesRegisters.single { rr->rr.registerOrPair!=null }.registerOrPair!!
-                                    val dt = when(reg) {
-                                        RegisterOrPair.A,
-                                        RegisterOrPair.X,
-                                        RegisterOrPair.Y -> DataType.UBYTE
-                                        RegisterOrPair.AX,
-                                        RegisterOrPair.AY,
-                                        RegisterOrPair.XY -> DataType.UWORD
-                                    }
-                                    return AsmAssignSource(SourceStorageKind.EXPRESSION, program, asmgen, dt, expression = value)
-                                }
-                                else -> throw AssemblyError("can't translate multiple return values in assignment")
-                            }
+                is FunctionCall -> {
+                    when (val sub = value.target.targetStatement(program.namespace)) {
+                        is Subroutine -> {
+                            val returnType = sub.returntypes.zip(sub.asmReturnvaluesRegisters).firstOrNull { rr -> rr.second.registerOrPair != null }?.first
+                                    ?: throw AssemblyError("can't translate zero return values in assignment")
+
+                            AsmAssignSource(SourceStorageKind.EXPRESSION, program, asmgen, returnType, expression = value)
+                        }
+                        is BuiltinFunctionStatementPlaceholder -> {
+                            val returnType = value.inferType(program)
+                            if(!returnType.isKnown)
+                                throw AssemblyError("unknown dt")
+                            AsmAssignSource(SourceStorageKind.EXPRESSION, program, asmgen, returnType.typeOrElse(DataType.STRUCT), expression = value)
+                        }
+                        else -> {
+                            throw AssemblyError("weird call")
                         }
                     }
-
-                    val dt = value.inferType(program).typeOrElse(DataType.STRUCT)
-                    return AsmAssignSource(SourceStorageKind.EXPRESSION, program, asmgen, dt, expression = value)
+                }
+                else -> {
+                    val dt = value.inferType(program)
+                    if(!dt.isKnown)
+                        throw AssemblyError("unknown dt")
+                    AsmAssignSource(SourceStorageKind.EXPRESSION, program, asmgen, dt.typeOrElse(DataType.STRUCT), expression = value)
                 }
             }
         }

compiler/src/prog8/compiler/target/cx16/CX16MachineDefinition.kt

@@ -38,8 +38,7 @@ internal object CX16MachineDefinition: IMachineDefinition {
     override fun launchEmulator(programName: String) {
         for(emulator in listOf("x16emu")) {
             println("\nStarting Commander X16 emulator $emulator...")
-            val cmdline = listOf(emulator, "-rom", "/usr/share/x16-rom/rom.bin", "-scale", "2",
-                    "-run", "-prg", programName + ".prg")
+            val cmdline = listOf(emulator, "-scale", "2", "-run", "-prg", "$programName.prg")
             val processb = ProcessBuilder(cmdline).inheritIO()
             val process: Process
             try {
@@ -75,7 +74,7 @@ internal object CX16MachineDefinition: IMachineDefinition {
     internal class CX16Zeropage(options: CompilationOptions) : Zeropage(options) {
 
         override val SCRATCH_B1 = 0x79      // temp storage for a single byte
-        override val SCRATCH_REG = 0x7a     // temp storage for a register
+        override val SCRATCH_REG = 0x7a     // temp storage for a register, must be B1+1
         override val SCRATCH_W1 = 0x7c      // temp storage 1 for a word  $7c+$7d
         override val SCRATCH_W2 = 0x7e      // temp storage 2 for a word  $7e+$7f
 

compiler/src/prog8/functions/BuiltinFunctions.kt

@@ -15,96 +15,167 @@ class FParam(val name: String, val possibleDatatypes: Set<DataType>)
 typealias ConstExpressionCaller = (args: List<Expression>, position: Position, program: Program) -> NumericLiteralValue
 
 
-class FSignature(val pure: Boolean,      // does it have side effects?
+class ReturnConvention(val dt: DataType, val reg: RegisterOrPair?, val floatFac1: Boolean)
+class ParamConvention(val dt: DataType, val reg: RegisterOrPair?, val variable: Boolean)
+class CallConvention(val params: List<ParamConvention>, val returns: ReturnConvention) {
+    override fun toString(): String {
+        val paramConvs =  params.mapIndexed { index, it ->
+            when {
+                it.reg!=null -> "$index:${it.reg}"
+                it.variable -> "$index:variable"
+                else -> "$index:???"
+            }
+        }
+        val returnConv =
+                when {
+                    returns.reg!=null -> returns.reg.toString()
+                    returns.floatFac1 -> "floatFAC1"
+                    else -> "<no returnvalue>"
+                }
+        return "CallConvention[" + paramConvs.joinToString() + " ; returns: $returnConv]"
+    }
+}
+
+
+class FSignature(val name: String,
+                 val pure: Boolean,      // does it have side effects?
                  val parameters: List<FParam>,
-                 val returntype: DataType?,
-                 val constExpressionFunc: ConstExpressionCaller? = null)
+                 val known_returntype: DataType?,     // specify return type if fixed, otherwise null if it depends on the arguments
+                 val constExpressionFunc: ConstExpressionCaller? = null) {
 
+    fun callConvention(actualParamTypes: List<DataType>): CallConvention {
+        val returns = when(known_returntype) {
+            DataType.UBYTE, DataType.BYTE -> ReturnConvention(known_returntype, RegisterOrPair.A, false)
+            DataType.UWORD, DataType.WORD -> ReturnConvention(known_returntype, RegisterOrPair.AY, false)
+            DataType.FLOAT -> ReturnConvention(known_returntype, null, true)
+            in PassByReferenceDatatypes -> ReturnConvention(known_returntype!!, RegisterOrPair.AY, false)
+            else -> {
+                val paramType = actualParamTypes.first()
+                if(pure)
+                    // return type depends on arg type
+                    when(paramType) {
+                        DataType.UBYTE, DataType.BYTE -> ReturnConvention(paramType, RegisterOrPair.A, false)
+                        DataType.UWORD, DataType.WORD -> ReturnConvention(paramType, RegisterOrPair.AY, false)
+                        DataType.FLOAT -> ReturnConvention(paramType, null, true)
+                        in PassByReferenceDatatypes -> ReturnConvention(paramType, RegisterOrPair.AY, false)
+                        else -> ReturnConvention(paramType, null, false)
+                    }
+                else {
+                    ReturnConvention(paramType, null, false)
+                }
+            }
+        }
+        return when {
+            actualParamTypes.isEmpty() -> CallConvention(emptyList(), returns)
+            actualParamTypes.size==1 -> {
+                // one parameter? via register/registerpair
+                val paramConv = when(val paramType = actualParamTypes[0]) {
+                    DataType.UBYTE, DataType.BYTE -> ParamConvention(paramType, RegisterOrPair.A, false)
+                    DataType.UWORD, DataType.WORD -> ParamConvention(paramType, RegisterOrPair.AY, false)
+                    DataType.FLOAT -> ParamConvention(paramType, RegisterOrPair.AY, false)
+                    in PassByReferenceDatatypes -> ParamConvention(paramType, RegisterOrPair.AY, false)
+                    else -> ParamConvention(paramType, null, false)
+                }
+                CallConvention(listOf(paramConv), returns)
+            }
+            else -> {
+                // multiple parameters? via variables
+                val paramConvs = actualParamTypes.map { ParamConvention(it, null, true) }
+                CallConvention(paramConvs, returns)
+            }
+        }
+    }
+}
 
-val BuiltinFunctions = mapOf(
+private val functionSignatures: List<FSignature> = listOf(
         // this set of function have no return value and operate in-place:
-    "rol"         to FSignature(false, listOf(FParam("item", setOf(DataType.UBYTE, DataType.UWORD))), null),
-    "ror"         to FSignature(false, listOf(FParam("item", setOf(DataType.UBYTE, DataType.UWORD))), null),
-    "rol2"        to FSignature(false, listOf(FParam("item", setOf(DataType.UBYTE, DataType.UWORD))), null),
-    "ror2"        to FSignature(false, listOf(FParam("item", setOf(DataType.UBYTE, DataType.UWORD))), null),
-    "sort"        to FSignature(false, listOf(FParam("array", ArrayDatatypes)), null),
-    "reverse"     to FSignature(false, listOf(FParam("array", ArrayDatatypes)), null),
+    FSignature("rol"         , false, listOf(FParam("item", setOf(DataType.UBYTE, DataType.UWORD))), null),
+    FSignature("ror"         , false, listOf(FParam("item", setOf(DataType.UBYTE, DataType.UWORD))), null),
+    FSignature("rol2"        , false, listOf(FParam("item", setOf(DataType.UBYTE, DataType.UWORD))), null),
+    FSignature("ror2"        , false, listOf(FParam("item", setOf(DataType.UBYTE, DataType.UWORD))), null),
+    FSignature("sort"        , false, listOf(FParam("array", ArrayDatatypes)), null),
+    FSignature("reverse"     , false, listOf(FParam("array", ArrayDatatypes)), null),
         // these few have a return value depending on the argument(s):
-    "max"         to FSignature(true, listOf(FParam("values", ArrayDatatypes)), null) { a, p, prg -> collectionArg(a, p, prg, ::builtinMax) },    // type depends on args
-    "min"         to FSignature(true, listOf(FParam("values", ArrayDatatypes)), null) { a, p, prg -> collectionArg(a, p, prg, ::builtinMin) },    // type depends on args
-    "sum"         to FSignature(true, listOf(FParam("values", ArrayDatatypes)), null) { a, p, prg -> collectionArg(a, p, prg, ::builtinSum) },      // type depends on args
-    "abs"         to FSignature(true, listOf(FParam("value", NumericDatatypes)), null, ::builtinAbs),      // type depends on argument
-    "len"         to FSignature(true, listOf(FParam("values", IterableDatatypes)), null, ::builtinLen),    // type is UBYTE or UWORD depending on actual length
-    "sizeof"      to FSignature(true, listOf(FParam("object", DataType.values().toSet())), DataType.UBYTE, ::builtinSizeof),
+    FSignature("max"         , true, listOf(FParam("values", ArrayDatatypes)), null) { a, p, prg -> collectionArg(a, p, prg, ::builtinMax) },    // type depends on args
+    FSignature("min"         , true, listOf(FParam("values", ArrayDatatypes)), null) { a, p, prg -> collectionArg(a, p, prg, ::builtinMin) },    // type depends on args
+    FSignature("sum"         , true, listOf(FParam("values", ArrayDatatypes)), null) { a, p, prg -> collectionArg(a, p, prg, ::builtinSum) },      // type depends on args
+    FSignature("abs"         , true, listOf(FParam("value", NumericDatatypes)), null, ::builtinAbs),      // type depends on argument
+    FSignature("len"         , true, listOf(FParam("values", IterableDatatypes)), null, ::builtinLen),    // type is UBYTE or UWORD depending on actual length
+    FSignature("sizeof"      , true, listOf(FParam("object", DataType.values().toSet())), DataType.UBYTE, ::builtinSizeof),
         // normal functions follow:
-    "sgn"         to FSignature(true, listOf(FParam("value", NumericDatatypes)), DataType.BYTE, ::builtinSgn ),
-    "sin"         to FSignature(true, listOf(FParam("rads", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::sin) },
-    "sin8"        to FSignature(true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.BYTE, ::builtinSin8 ),
-    "sin8u"       to FSignature(true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.UBYTE, ::builtinSin8u ),
-    "sin16"       to FSignature(true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.WORD, ::builtinSin16 ),
-    "sin16u"      to FSignature(true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.UWORD, ::builtinSin16u ),
-    "cos"         to FSignature(true, listOf(FParam("rads", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::cos) },
-    "cos8"        to FSignature(true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.BYTE, ::builtinCos8 ),
-    "cos8u"       to FSignature(true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.UBYTE, ::builtinCos8u ),
-    "cos16"       to FSignature(true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.WORD, ::builtinCos16 ),
-    "cos16u"      to FSignature(true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.UWORD, ::builtinCos16u ),
-    "tan"         to FSignature(true, listOf(FParam("rads", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::tan) },
-    "atan"        to FSignature(true, listOf(FParam("rads", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::atan) },
-    "ln"          to FSignature(true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::log) },
-    "log2"        to FSignature(true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, ::log2) },
-    "sqrt16"      to FSignature(true, listOf(FParam("value", setOf(DataType.UWORD))), DataType.UBYTE) { a, p, prg -> oneIntArgOutputInt(a, p, prg) { sqrt(it.toDouble()).toInt() } },
-    "sqrt"        to FSignature(true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::sqrt) },
-    "rad"         to FSignature(true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::toRadians) },
-    "deg"         to FSignature(true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::toDegrees) },
-    "round"       to FSignature(true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArgOutputWord(a, p, prg, Math::round) },
-    "floor"       to FSignature(true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArgOutputWord(a, p, prg, Math::floor) },
-    "ceil"        to FSignature(true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArgOutputWord(a, p, prg, Math::ceil) },
-    "any"         to FSignature(true, listOf(FParam("values", ArrayDatatypes)), DataType.UBYTE) { a, p, prg -> collectionArg(a, p, prg, ::builtinAny) },
-    "all"         to FSignature(true, listOf(FParam("values", ArrayDatatypes)), DataType.UBYTE) { a, p, prg -> collectionArg(a, p, prg, ::builtinAll) },
-    "lsb"         to FSignature(true, listOf(FParam("value", setOf(DataType.UWORD, DataType.WORD))), DataType.UBYTE) { a, p, prg -> oneIntArgOutputInt(a, p, prg) { x: Int -> x and 255 }},
-    "msb"         to FSignature(true, listOf(FParam("value", setOf(DataType.UWORD, DataType.WORD))), DataType.UBYTE) { a, p, prg -> oneIntArgOutputInt(a, p, prg) { x: Int -> x ushr 8 and 255}},
-    "mkword"      to FSignature(true, listOf(FParam("msb", setOf(DataType.UBYTE)), FParam("lsb", setOf(DataType.UBYTE))), DataType.UWORD, ::builtinMkword),
-    "rnd"         to FSignature(true, emptyList(), DataType.UBYTE),
-    "rndw"        to FSignature(true, emptyList(), DataType.UWORD),
-    "rndf"        to FSignature(true, emptyList(), DataType.FLOAT),
-    "exit"        to FSignature(false, listOf(FParam("returnvalue", setOf(DataType.UBYTE))), null),
-    "rsave"       to FSignature(false, emptyList(), null),
-    "rrestore"    to FSignature(false, emptyList(), null),
-    "set_carry"   to FSignature(false, emptyList(), null),
-    "clear_carry" to FSignature(false, emptyList(), null),
-    "set_irqd"    to FSignature(false, emptyList(), null),
-    "clear_irqd"  to FSignature(false, emptyList(), null),
-    "read_flags"  to FSignature(false, emptyList(), DataType.UBYTE),
-    "swap"        to FSignature(false, listOf(FParam("first", NumericDatatypes), FParam("second", NumericDatatypes)), null),
-    "memcopy"     to FSignature(false, listOf(
+    FSignature("sgn"         , true, listOf(FParam("value", NumericDatatypes)), DataType.BYTE, ::builtinSgn ),
+    FSignature("sin"         , true, listOf(FParam("rads", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::sin) },
+    FSignature("sin8"        , true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.BYTE, ::builtinSin8 ),
+    FSignature("sin8u"       , true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.UBYTE, ::builtinSin8u ),
+    FSignature("sin16"       , true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.WORD, ::builtinSin16 ),
+    FSignature("sin16u"      , true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.UWORD, ::builtinSin16u ),
+    FSignature("cos"         , true, listOf(FParam("rads", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::cos) },
+    FSignature("cos8"        , true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.BYTE, ::builtinCos8 ),
+    FSignature("cos8u"       , true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.UBYTE, ::builtinCos8u ),
+    FSignature("cos16"       , true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.WORD, ::builtinCos16 ),
+    FSignature("cos16u"      , true, listOf(FParam("angle8", setOf(DataType.UBYTE))), DataType.UWORD, ::builtinCos16u ),
+    FSignature("tan"         , true, listOf(FParam("rads", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::tan) },
+    FSignature("atan"        , true, listOf(FParam("rads", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::atan) },
+    FSignature("ln"          , true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::log) },
+    FSignature("log2"        , true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, ::log2) },
+    FSignature("sqrt16"      , true, listOf(FParam("value", setOf(DataType.UWORD))), DataType.UBYTE) { a, p, prg -> oneIntArgOutputInt(a, p, prg) { sqrt(it.toDouble()).toInt() } },
+    FSignature("sqrt"        , true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::sqrt) },
+    FSignature("rad"         , true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::toRadians) },
+    FSignature("deg"         , true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArg(a, p, prg, Math::toDegrees) },
+    FSignature("round"       , true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArgOutputWord(a, p, prg, Math::round) },
+    FSignature("floor"       , true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArgOutputWord(a, p, prg, Math::floor) },
+    FSignature("ceil"        , true, listOf(FParam("value", setOf(DataType.FLOAT))), DataType.FLOAT) { a, p, prg -> oneDoubleArgOutputWord(a, p, prg, Math::ceil) },
+    FSignature("any"         , true, listOf(FParam("values", ArrayDatatypes)), DataType.UBYTE) { a, p, prg -> collectionArg(a, p, prg, ::builtinAny) },
+    FSignature("all"         , true, listOf(FParam("values", ArrayDatatypes)), DataType.UBYTE) { a, p, prg -> collectionArg(a, p, prg, ::builtinAll) },
+    FSignature("lsb"         , true, listOf(FParam("value", setOf(DataType.UWORD, DataType.WORD))), DataType.UBYTE) { a, p, prg -> oneIntArgOutputInt(a, p, prg) { x: Int -> x and 255 }},
+    FSignature("msb"         , true, listOf(FParam("value", setOf(DataType.UWORD, DataType.WORD))), DataType.UBYTE) { a, p, prg -> oneIntArgOutputInt(a, p, prg) { x: Int -> x ushr 8 and 255}},
+    FSignature("mkword"      , true, listOf(FParam("msb", setOf(DataType.UBYTE)), FParam("lsb", setOf(DataType.UBYTE))), DataType.UWORD, ::builtinMkword),
+    FSignature("rnd"         , false, emptyList(), DataType.UBYTE),
+    FSignature("rndw"        , false, emptyList(), DataType.UWORD),
+    FSignature("rndf"        , false, emptyList(), DataType.FLOAT),
+    FSignature("exit"        , false, listOf(FParam("returnvalue", setOf(DataType.UBYTE))), null),
+    FSignature("rsave"       , false, emptyList(), null),
+    FSignature("rrestore"    , false, emptyList(), null),
+    FSignature("set_carry"   , false, emptyList(), null),
+    FSignature("clear_carry" , false, emptyList(), null),
+    FSignature("set_irqd"    , false, emptyList(), null),
+    FSignature("clear_irqd"  , false, emptyList(), null),
+    FSignature("read_flags"  , false, emptyList(), DataType.UBYTE),
+    FSignature("progend"     , true, emptyList(), DataType.UWORD),
+    FSignature("swap"        , false, listOf(FParam("first", NumericDatatypes), FParam("second", NumericDatatypes)), null),
+    FSignature("memcopy"     , false, listOf(
                             FParam("from", IterableDatatypes + DataType.UWORD),
                             FParam("to", IterableDatatypes + DataType.UWORD),
-                            FParam("numbytes", setOf(DataType.UBYTE))), null),
-    "memset"      to FSignature(false, listOf(
+                            FParam("numbytes", setOf(DataType.UBYTE, DataType.UWORD))), null),
+    FSignature("memset"      , false, listOf(
                             FParam("address", IterableDatatypes + DataType.UWORD),
                             FParam("numbytes", setOf(DataType.UWORD)),
                             FParam("bytevalue", ByteDatatypes)), null),
-    "memsetw"     to FSignature(false, listOf(
+    FSignature("memsetw"     , false, listOf(
                             FParam("address", IterableDatatypes + DataType.UWORD),
                             FParam("numwords", setOf(DataType.UWORD)),
                             FParam("wordvalue", setOf(DataType.UWORD, DataType.WORD))), null),
-    "strlen"      to FSignature(true, listOf(FParam("string", setOf(DataType.STR))), DataType.UBYTE, ::builtinStrlen),
-    "substr"      to FSignature(false, listOf(
+    FSignature("strlen"      , true, listOf(FParam("string", setOf(DataType.STR))), DataType.UBYTE, ::builtinStrlen),
+    FSignature("strcopy"     , false, listOf(FParam("from", IterableDatatypes + DataType.UWORD), FParam("to", IterableDatatypes + DataType.UWORD)), DataType.UBYTE),
+    FSignature("substr"      , false, listOf(
             FParam("source", IterableDatatypes + DataType.UWORD),
             FParam("target", IterableDatatypes + DataType.UWORD),
             FParam("start", setOf(DataType.UBYTE)),
             FParam("length", setOf(DataType.UBYTE))), null),
-    "leftstr"      to FSignature(false, listOf(
+    FSignature("leftstr"     , false, listOf(
             FParam("source", IterableDatatypes + DataType.UWORD),
             FParam("target", IterableDatatypes + DataType.UWORD),
             FParam("length", setOf(DataType.UBYTE))), null),
-    "rightstr"      to FSignature(false, listOf(
+    FSignature("rightstr"    , false, listOf(
             FParam("source", IterableDatatypes + DataType.UWORD),
             FParam("target", IterableDatatypes + DataType.UWORD),
             FParam("length", setOf(DataType.UBYTE))), null),
-    "strcmp"      to FSignature(false, listOf(FParam("s1", IterableDatatypes + DataType.UWORD), FParam("s2", IterableDatatypes + DataType.UWORD)), DataType.BYTE, null)
+    FSignature("strcmp"      , false, listOf(FParam("s1", IterableDatatypes + DataType.UWORD), FParam("s2", IterableDatatypes + DataType.UWORD)), DataType.BYTE, null)
 )
 
+val BuiltinFunctions = functionSignatures.associateBy { it.name }
+
+
 fun builtinMax(array: List<Number>): Number = array.maxByOrNull { it.toDouble() }!!
 
 fun builtinMin(array: List<Number>): Number = array.minByOrNull { it.toDouble() }!!
@@ -144,17 +215,17 @@ fun builtinFunctionReturnType(function: String, args: List<Expression>, program:
     }
 
     val func = BuiltinFunctions.getValue(function)
-    if(func.returntype!=null)
-        return InferredTypes.knownFor(func.returntype)
-    // function has return values, but the return type depends on the arguments
+    if(func.known_returntype!=null)
+        return InferredTypes.knownFor(func.known_returntype)
 
+    // function has return values, but the return type depends on the arguments
     return when (function) {
         "abs" -> {
             val dt = args.single().inferType(program)
-            if(dt.typeOrElse(DataType.STRUCT) in NumericDatatypes)
-                return dt
+            return if(dt.typeOrElse(DataType.STRUCT) in NumericDatatypes)
+                dt
             else
-                throw FatalAstException("weird datatype passed to abs $dt")
+                InferredTypes.InferredType.unknown()
         }
         "max", "min" -> {
             when(val dt = datatypeFromIterableArg(args.single())) {
@@ -284,7 +355,7 @@ private fun builtinStrlen(args: List<Expression>, position: Position, program: P
     val argument=args[0]
     if(argument is StringLiteralValue)
         return NumericLiteralValue.optimalInteger(argument.value.length, argument.position)
-    val vardecl = (argument as IdentifierReference).targetVarDecl(program.namespace)
+    val vardecl = (argument as? IdentifierReference)?.targetVarDecl(program.namespace)
     if(vardecl!=null) {
         if(vardecl.datatype!=DataType.STR && vardecl.datatype!=DataType.UWORD)
             throw SyntaxError("strlen must have string argument", position)

compiler/src/prog8/optimizer/BinExprSplitter.kt

@@ -38,22 +38,26 @@ internal class BinExprSplitter(private val program: Program) : AstWalker() {
         if (binExpr != null) {
 /*
 
-reduce the complexity of a (binary) expression that has to be evaluated on the eval stack,
-by attempting to splitting it up into individual simple steps:
+Reduce the complexity of a (binary) expression that has to be evaluated on the eval stack,
+by attempting to splitting it up into individual simple steps.
+We only consider a binary expression *one* level deep (so the operands must not be a combined expression)
 
 
-X =      BinExpr                                   X   =   LeftExpr
-        <operator>                                   followed by
-          /   \              IF 'X' not used       X   =   BinExpr
-         /     \               IN LEFTEXPR ==>             <operator>
+X =      BinExpr                                    X   =   LeftExpr
+        <operator>                                     followed by
+          /   \             IF 'X' not used         X   =   BinExpr
+         /     \             IN expression ==>             <operator>
         /       \                                           /   \
     LeftExpr.  RightExpr.                                  /     \
-      /  \       /  \                                     X     RightExpr.
-    ..  ..      ..  ..
+                                                          X     RightExpr.
+
 
  */
             if(binExpr.operator in augmentAssignmentOperators && isSimpleTarget(assignment.target, program.namespace)) {
-                if (!assignment.isAugmentable) {
+                if(assignment.target isSameAs binExpr.left || assignment.target isSameAs binExpr.right)
+                    return noModifications
+
+                if(isSimpleExpression(binExpr.right) && !assignment.isAugmentable) {
                     val firstAssign = Assignment(assignment.target, binExpr.left, binExpr.left.position)
                     val targetExpr = assignment.target.toExpression()
                     val augExpr = BinaryExpression(targetExpr, binExpr.operator, binExpr.right, binExpr.right.position)
@@ -71,8 +75,11 @@ X =      BinExpr                                   X   =   LeftExpr
         return noModifications
     }
 
+    private fun isSimpleExpression(expr: Expression) =
+            expr is IdentifierReference || expr is NumericLiteralValue || expr is AddressOf || expr is DirectMemoryRead || expr is StringLiteralValue || expr is ArrayLiteralValue || expr is RangeExpr
+
     private fun isSimpleTarget(target: AssignTarget, namespace: INameScope) =
-            if (target.identifier!=null || target.memoryAddress!=null || target.arrayindexed!=null)
+            if (target.identifier!=null || target.memoryAddress!=null)
                 target.isInRegularRAM(namespace)
             else
                 false

compiler/src/prog8/optimizer/ConstantFoldingOptimizer.kt

@@ -7,6 +7,7 @@ import prog8.ast.expressions.*
 import prog8.ast.processing.AstWalker
 import prog8.ast.processing.IAstModification
 import prog8.ast.statements.*
+import kotlin.math.pow
 
 
 internal class ConstantFoldingOptimizer(private val program: Program) : AstWalker() {
@@ -97,6 +98,43 @@ internal class ConstantFoldingOptimizer(private val program: Program) : AstWalke
     override fun after(expr: BinaryExpression, parent: Node): Iterable<IAstModification> {
         val leftconst = expr.left.constValue(program)
         val rightconst = expr.right.constValue(program)
+        val modifications = mutableListOf<IAstModification>()
+
+        if(expr.operator == "**" && leftconst!=null) {
+            // optimize various simple cases of ** :
+            //  optimize away 1 ** x into just 1 and 0 ** x into just 0
+            //  optimize 2 ** x into (1<<x)  if both operands are integer.
+            val leftDt = leftconst.inferType(program).typeOrElse(DataType.STRUCT)
+            when (leftconst.number.toDouble()) {
+                0.0 -> {
+                    val value = NumericLiteralValue(leftDt, 0, expr.position)
+                    modifications += IAstModification.ReplaceNode(expr, value, parent)
+                }
+                1.0 -> {
+                    val value = NumericLiteralValue(leftDt, 1, expr.position)
+                    modifications += IAstModification.ReplaceNode(expr, value, parent)
+                }
+                2.0 -> {
+                    if(rightconst!=null) {
+                        val value = NumericLiteralValue(leftDt, 2.0.pow(rightconst.number.toDouble()), expr.position)
+                        modifications += IAstModification.ReplaceNode(expr, value, parent)
+                    } else {
+                        val rightDt = expr.right.inferType(program).typeOrElse(DataType.STRUCT)
+                        if(leftDt in IntegerDatatypes && rightDt in IntegerDatatypes) {
+                            val targetDt =
+                                when (parent) {
+                                    is Assignment -> parent.target.inferType(program).typeOrElse(DataType.STRUCT)
+                                    is VarDecl -> parent.datatype
+                                    else -> leftDt
+                                }
+                            val one = NumericLiteralValue(targetDt, 1, expr.position)
+                            val shift = BinaryExpression(one, "<<", expr.right, expr.position)
+                            modifications += IAstModification.ReplaceNode(expr, shift, parent)
+                        }
+                    }
+                }
+            }
+        }
 
         val subExpr: BinaryExpression? = when {
             leftconst!=null -> expr.right as? BinaryExpression
@@ -111,7 +149,8 @@ internal class ConstantFoldingOptimizer(private val program: Program) : AstWalke
                 val change = groupTwoConstsTogether(expr, subExpr,
                         leftconst != null, rightconst != null,
                         subleftconst != null, subrightconst != null)
-                return change?.let { listOf(it) } ?: noModifications
+                if(change!=null)
+                    modifications += change
             }
         }
 
@@ -119,10 +158,10 @@ internal class ConstantFoldingOptimizer(private val program: Program) : AstWalke
         if(leftconst != null && rightconst != null) {
             val evaluator = ConstExprEvaluator()
             val result = evaluator.evaluate(leftconst, expr.operator, rightconst)
-            return listOf(IAstModification.ReplaceNode(expr, result, parent))
+            modifications += IAstModification.ReplaceNode(expr, result, parent)
         }
 
-        return noModifications
+        return modifications
     }
 
     override fun after(array: ArrayLiteralValue, parent: Node): Iterable<IAstModification> {

compiler/src/prog8/optimizer/ExpressionSimplifier.kt

@@ -6,6 +6,7 @@ import prog8.ast.base.*
 import prog8.ast.expressions.*
 import prog8.ast.processing.AstWalker
 import prog8.ast.processing.IAstModification
+import prog8.ast.statements.Assignment
 import kotlin.math.abs
 import kotlin.math.log2
 import kotlin.math.pow
@@ -98,6 +99,12 @@ internal class ExpressionSimplifier(private val program: Program) : AstWalker()
         if (leftVal != null && expr.operator in associativeOperators && rightVal == null)
             return listOf(IAstModification.SwapOperands(expr))
 
+        // NonBinaryExpression  <associativeoperator>  BinaryExpression  -->  BinaryExpression  <associativeoperator>  NonBinaryExpression
+        if (expr.operator in associativeOperators && expr.left !is BinaryExpression && expr.right is BinaryExpression) {
+            if(parent !is Assignment || !(expr.left isSameAs parent.target))
+                return listOf(IAstModification.SwapOperands(expr))
+        }
+
         // X + (-A)  -->  X - A
         if (expr.operator == "+" && (expr.right as? PrefixExpression)?.operator == "-") {
             return listOf(IAstModification.ReplaceNode(
@@ -338,7 +345,7 @@ internal class ExpressionSimplifier(private val program: Program) : AstWalker()
         if (leftVal == null && rightVal == null)
             return null
 
-        val (expr2, _, rightVal2) = reorderAssociative(expr, leftVal)
+        val (expr2, _, rightVal2) = reorderAssociativeWithConstant(expr, leftVal)
         if (rightVal2 != null) {
             // right value is a constant, see if we can optimize
             val rightConst: NumericLiteralValue = rightVal2
@@ -477,7 +484,10 @@ internal class ExpressionSimplifier(private val program: Program) : AstWalker()
         when (expr.operator) {
             "%" -> {
                 if (cv == 1.0) {
-                    return NumericLiteralValue(expr.inferType(program).typeOrElse(DataType.STRUCT), 0, expr.position)
+                    val idt = expr.inferType(program)
+                    if(!idt.isKnown)
+                        throw FatalAstException("unknown dt")
+                    return NumericLiteralValue(idt.typeOrElse(DataType.STRUCT), 0, expr.position)
                 } else if (cv == 2.0) {
                     expr.operator = "&"
                     expr.right = NumericLiteralValue.optimalInteger(1, expr.position)
@@ -559,7 +569,7 @@ internal class ExpressionSimplifier(private val program: Program) : AstWalker()
         if (leftVal == null && rightVal == null)
             return null
 
-        val (expr2, _, rightVal2) = reorderAssociative(expr, leftVal)
+        val (expr2, _, rightVal2) = reorderAssociativeWithConstant(expr, leftVal)
         if (rightVal2 != null) {
             // right value is a constant, see if we can optimize
             val leftValue: Expression = expr2.left
@@ -606,8 +616,10 @@ internal class ExpressionSimplifier(private val program: Program) : AstWalker()
         if (amount == 0) {
             return expr.left
         }
-        val targetDt = expr.left.inferType(program).typeOrElse(DataType.STRUCT)
-        when (targetDt) {
+        val targetIDt = expr.left.inferType(program)
+        if(!targetIDt.isKnown)
+            throw FatalAstException("unknown dt")
+        when (val targetDt = targetIDt.typeOrElse(DataType.STRUCT)) {
             DataType.UBYTE, DataType.BYTE -> {
                 if (amount >= 8) {
                     return NumericLiteralValue(targetDt, 0, expr.position)
@@ -617,8 +629,7 @@ internal class ExpressionSimplifier(private val program: Program) : AstWalker()
                 if (amount >= 16) {
                     return NumericLiteralValue(targetDt, 0, expr.position)
                 } else if (amount >= 8) {
-                    // TODO is this correct???
-                    val lsb = TypecastExpression(expr.left, DataType.UBYTE, true, expr.position)
+                    val lsb = FunctionCall(IdentifierReference(listOf("lsb"), expr.position), mutableListOf(expr.left), expr.position)
                     if (amount == 8) {
                         return FunctionCall(IdentifierReference(listOf("mkword"), expr.position), mutableListOf(lsb, NumericLiteralValue.optimalInteger(0, expr.position)), expr.position)
                     }
@@ -640,7 +651,10 @@ internal class ExpressionSimplifier(private val program: Program) : AstWalker()
         if (amount == 0) {
             return expr.left
         }
-        when (expr.left.inferType(program).typeOrElse(DataType.STRUCT)) {
+        val idt = expr.left.inferType(program)
+        if(!idt.isKnown)
+            throw FatalAstException("unknown dt")
+        when (idt.typeOrElse(DataType.STRUCT)) {
             DataType.UBYTE -> {
                 if (amount >= 8) {
                     return NumericLiteralValue.optimalInteger(0, expr.position)
@@ -675,17 +689,17 @@ internal class ExpressionSimplifier(private val program: Program) : AstWalker()
         return null
     }
 
-    private fun reorderAssociative(expr: BinaryExpression, leftVal: NumericLiteralValue?): ReorderedAssociativeBinaryExpr {
+    private fun reorderAssociativeWithConstant(expr: BinaryExpression, leftVal: NumericLiteralValue?): BinExprWithConstants {
         if (expr.operator in associativeOperators && leftVal != null) {
             // swap left and right so that right is always the constant
             val tmp = expr.left
             expr.left = expr.right
             expr.right = tmp
-            return ReorderedAssociativeBinaryExpr(expr, expr.right.constValue(program), leftVal)
+            return BinExprWithConstants(expr, expr.right.constValue(program), leftVal)
         }
-        return ReorderedAssociativeBinaryExpr(expr, leftVal, expr.right.constValue(program))
+        return BinExprWithConstants(expr, leftVal, expr.right.constValue(program))
     }
 
-    private data class ReorderedAssociativeBinaryExpr(val expr: BinaryExpression, val leftVal: NumericLiteralValue?, val rightVal: NumericLiteralValue?)
+    private data class BinExprWithConstants(val expr: BinaryExpression, val leftVal: NumericLiteralValue?, val rightVal: NumericLiteralValue?)
 
 }

compiler/src/prog8/optimizer/StatementOptimizer.kt

@@ -323,7 +323,7 @@ internal class StatementOptimizer(private val program: Program,
                     val op1 = binExpr.operator
                     val op2 = rExpr.operator
 
-                    if(rExpr.left is NumericLiteralValue && op2 in setOf("+", "*", "&", "|")) {
+                    if(rExpr.left is NumericLiteralValue && op2 in associativeOperators) {
                         // associative operator, make sure the constant numeric value is second (right)
                         return listOf(IAstModification.SwapOperands(rExpr))
                     }
@@ -377,7 +377,7 @@ internal class StatementOptimizer(private val program: Program,
             return listOf(IAstModification.Remove(assignment, assignment.definingScope()))
         }
 
-        val targetIDt = assignment.target.inferType(program, assignment)
+        val targetIDt = assignment.target.inferType(program)
         if(!targetIDt.isKnown)
             throw FatalAstException("can't infer type of assignment target")
 

compiler/src/prog8/parser/ModuleParsing.kt

@@ -92,7 +92,7 @@ internal class ModuleImporter {
         // accept additional imports
         val lines = moduleAst.statements.toMutableList()
         lines.asSequence()
-                .mapIndexed { i, it -> Pair(i, it) }
+                .mapIndexed { i, it -> i to it }
                 .filter { (it.second as? Directive)?.directive == "%import" }
                 .forEach { executeImportDirective(program, it.second as Directive, modulePath) }
 
@@ -102,7 +102,7 @@ internal class ModuleImporter {
 
     private fun discoverImportedModuleFile(name: String, source: Path, position: Position?): Path {
         val fileName = "$name.p8"
-        val locations = if(source.toString().isEmpty()) mutableListOf<Path>() else mutableListOf(source.parent)
+        val locations = if(source.toString().isEmpty()) mutableListOf<Path>() else mutableListOf(source.parent ?: Path.of("."))
 
         val propPath = System.getProperty("prog8.libdir")
         if(propPath!=null)

compiler/test/UnitTests.kt

@@ -5,8 +5,6 @@ import org.hamcrest.Matchers.closeTo
 import org.hamcrest.Matchers.equalTo
 import org.junit.jupiter.api.Test
 import org.junit.jupiter.api.TestInstance
-import prog8.ast.Module
-import prog8.ast.Program
 import prog8.ast.base.*
 import prog8.ast.expressions.*
 import prog8.ast.statements.*
@@ -18,8 +16,8 @@ import prog8.compiler.target.c64.C64MachineDefinition.FLOAT_MAX_NEGATIVE
 import prog8.compiler.target.c64.C64MachineDefinition.FLOAT_MAX_POSITIVE
 import prog8.compiler.target.c64.C64MachineDefinition.Mflpt5
 import prog8.compiler.target.c64.Petscii
+import prog8.compiler.target.cx16.CX16MachineDefinition
 import java.io.CharConversionException
-import java.nio.file.Path
 import kotlin.test.*
 
 @TestInstance(TestInstance.Lifecycle.PER_CLASS)
@@ -288,6 +286,22 @@ class TestC64Zeropage {
         assertEquals(0xf9, zp.allocate("", DataType.UBYTE, null, errors))
         assertEquals(0, zp.available())
     }
+
+    @Test
+    fun testReservedLocations() {
+        val zp = C64Zeropage(CompilationOptions(OutputType.RAW, LauncherType.NONE, ZeropageType.BASICSAFE, emptyList(), false, false))
+        assertEquals(zp.SCRATCH_REG, zp.SCRATCH_B1+1, "zp _B1 and _REG must be next to each other to create a word")
+    }
+}
+
+
+@TestInstance(TestInstance.Lifecycle.PER_CLASS)
+class TestCx16Zeropage {
+    @Test
+    fun testReservedLocations() {
+        val zp = CX16MachineDefinition.CX16Zeropage(CompilationOptions(OutputType.RAW, LauncherType.NONE, ZeropageType.BASICSAFE, emptyList(), false, false))
+        assertEquals(zp.SCRATCH_REG, zp.SCRATCH_B1+1, "zp _B1 and _REG must be next to each other to create a word")
+    }
 }
 
 

docs/source/building.rst

@@ -93,6 +93,8 @@ Almost instant compilation times (less than a second) can be achieved when using
 Start the compiler with the ``-watch`` argument to enable this.
 It will compile your program and then instead of exiting, it waits for any changes in the module source files.
 As soon as a change happens, the program gets compiled again.
+It is possible to use the watch mode with multiple modules as well, but it will
+recompile everything in that list even if only of the files got updated.
 
 Other options
 ^^^^^^^^^^^^^

docs/source/index.rst

@@ -38,6 +38,27 @@ This software is licensed under the GNU GPL 3.0, see https://www.gnu.org/license
     :alt: Fully playable tetris clone
 
 
+Language features
+-----------------
+
+- It is a cross-compiler running on modern machines (Linux, MacOS, Windows, ...)
+  The generated output is a machine code program runnable on actual 8-bit 6502 hardware.
+- Provide a very convenient edit/compile/run cycle by being able to directly launch
+  the compiled program in an emulator and provide debugging information to this emulator.
+- Based on simple and familiar imperative structured programming (it looks like a mix of C and Python)
+- Modular programming and scoping via modules, code blocks, and subroutines.
+- Provide high level programming constructs but at the same time stay close to the metal;
+  still able to directly use memory addresses and ROM subroutines,
+  and inline assembly to have full control when every register, cycle or byte matters
+- Arbitrary number of subroutine parameters, Complex nested expressions are possible
+- No stack frame allocations because parameters and local variables are automatically allocated statically
+- Nested subroutines can access variables from outer scopes to avoids the overhead to pass everything via parameters
+- Variable data types include signed and unsigned bytes and words, arrays, strings and floats.
+- High-level code optimizations, such as const-folding, expression and statement simplifications/rewriting.
+- Many built-in functions, such as ``sin``, ``cos``, ``rnd``, ``abs``, ``min``, ``max``, ``sqrt``, ``msb``, ``rol``, ``ror``, ``swap``, ``memset``, ``memcopy``, ``substr``, ``sort`` and ``reverse`` (and others)
+- If you only use standard kernel and prog8 library routines, it is possible to compile the *exact same program* for both machines (just change the compiler target flag)!
+
+
 Code example
 ------------
 
@@ -104,37 +125,6 @@ when the exact same program is compiled for the Commander X16 target, and run on
 
 
 
-Design principles and features
-------------------------------
-
-- It is a cross-compiler running on modern machines (Linux, MacOS, Windows, ...)
-  The generated output is a machine code program runnable on actual 8-bit 6502 hardware.
-- Based on simple and familiar imperative structured programming (it looks like a mix of C and Python)
-- 'One statement per line' code, resulting in clear readable programs.
-- Modular programming and scoping via modules, code blocks, and subroutines.
-- Provide high level programming constructs but at the same time stay close to the metal;
-  still able to directly use memory addresses and ROM subroutines,
-  and inline assembly to have full control when every register, cycle or byte matters
-- Arbitrary number of subroutine parameters
-- Complex nested expressions are possible
-- Nested subroutines can access variables from outer scopes to avoids the overhead to pass everything via parameters
-- Values are typed. Available data types include signed and unsigned bytes and words, arrays, strings and floats.
-- No dynamic memory allocation or sizing! All variables stay fixed size as determined at compile time.
-- Provide various quality of life language features and library subroutines specifically for the target platform.
-- Provide a very convenient edit/compile/run cycle by being able to directly launch
-  the compiled program in an emulator and provide debugging information to this emulator.
-- Arbitrary control flow jumps and branches are possible,
-  and will usually translate directly into the appropriate single 6502 jump/branch instruction.
-- There are no complicated built-in error handling or overflow checks, you'll have to take care
-  of this yourself if required. This keeps the language and code simple and efficient.
-- The compiler tries to optimize the program and generated code a bit, but hand-tuning of the
-  performance or space-critical parts will likely still be required. This is supported by
-  the ability to easily write embedded assembly code directly in the program source code.
-- There are many built-in functions, such as ``sin``, ``cos``, ``rnd``, ``abs``, ``min``, ``max``, ``sqrt``, ``msb``, ``rol``, ``ror``, ``swap``, ``memset``, ``memcopy``, ``substr``, ``sort`` and ``reverse`` (and others)
-- Assembling the generated code into a program wil be done by an external cross-assembler tool.
-- If you only use standard kernel and prog8 library routines, it is possible to compile the *exact same program* for both machines (just change the compiler target flag)!
-
-
 .. _requirements:
 
 Required tools
@@ -143,14 +133,17 @@ Required tools
 `64tass <https://sourceforge.net/projects/tass64/>`_ - cross assembler. Install this on your shell path.
 It's very easy to compile yourself.
 A recent precompiled .exe for Windows can be obtained from my `clone <https://github.com/irmen/64tass/releases>`_ of this project.
+*You need at least version 1.55.2257 of this assembler to correctly use the breakpoints feature.*
+It's possible to use older versions, but it is very likely that the automatic Vice breakpoints won't work with them.
 
-A **Java runtime (jre or jdk), version 8 or newer**  is required to run the prog8 compiler itself.
-If you're scared of Oracle's licensing terms, most Linux distributions ship OpenJDK instead.
-Fnd for Windows it's possible to get that as well. Check out `AdoptOpenJDK <https://adoptopenjdk.net/>`_ .
+A **Java runtime (jre or jdk), version 11 or newer**  is required to run the prog8 compiler itself.
+If you're scared of Oracle's licensing terms, most Linux distributions ship OpenJDK in their packages repository instead.
+For Windows it's possible to get that as well; check out `AdoptOpenJDK <https://adoptopenjdk.net/>`_ .
+For MacOS you can use the Homebrew system to install a recent version of OpenJDK.
 
-Finally: a **C-64 emulator** (or a real C-64 ofcourse) can be nice to test and run your programs on.
-The compiler assumes the presence of the `Vice emulator <http://vice-emu.sourceforge.net/>`_.
-If you're targeting the CommanderX16, there's the `x16emu <https://github.com/commanderx16/x16-emulator>`_.
+Finally: an **emulator** (or a real machine ofcourse) to test and run your programs on.
+In C64 mode, thhe compiler assumes the presence of the `Vice emulator <http://vice-emu.sourceforge.net/>`_.
+If you're targeting the CommanderX16 instead, there's the `x16emu <https://github.com/commanderx16/x16-emulator>`_.
 
 .. important::
     **Building the compiler itself:** (*Only needed if you have not downloaded a pre-built 'fat-jar'*)

docs/source/libraries.rst

@@ -54,7 +54,8 @@ diskio
 ------
 Provides several routines that deal with disk drive I/O, such as:
 
-- show directory
+- list files on disk, optionally filtering by prefix or suffix
+- show disk directory as-is
 - display disk drive status
 - load and save data from and to the disk
 - delete and rename files on the disk
@@ -88,6 +89,15 @@ A 'fun' module that contains the Commander X16 logo and that allows you
 to print it anywhere on the screen.
 
 
+c64colors
+---------
+Available for the CommanderX16 target, a module that contains a few better
+color palettes for how the colors of the VIC-II looked on the Commodore-64.
+There are subroutines to activate one of the several palettes of your liking.
+The Commander X16's default colors for this (the first 16 colors) are too saturated
+and are quite different than how a C-64 looked.
+
+
 prog8_lib
 ---------
 Low level language support. You should not normally have to bother with this directly.

docs/source/programming.rst

@@ -309,18 +309,21 @@ read the syntax reference on strings.
 
 
 .. hint::
-    Strings and uwords (=memory address) can often be interchanged.
+    Strings/arrays and uwords (=memory address) can often be interchanged.
     An array of strings is actually an array of uwords where every element is the memory
     address of the string. You can pass a memory address to assembly functions
     that require a string as an argument.
+    For regular assignments you still need to use an explicit ``&`` (address-of) to take
+    the address of the string or array.
 
 .. caution::
-    It's probably best to avoid changing strings after they've been created. This
-    includes changing certain letters by index, or by assigning a new value, or by
+    It's probably best to avoid changing the contents in strings and treat them as static.
+    This includes changing certain letters by index, or by assigning a new value, or by
     modifying the string via other means for example ``substr`` function and its cousins.
-    This is because if your program exits and is restarted (without loading it again),
-    it will then start working with the changed strings instead of the original ones!
-    The same is true for arrays.
+    This is because the changes persist in memory. If your program exits and is restarted
+    (without reloading it from disk), it will then start working with the modified strings
+    instead of the original ones!
+    The same is true for arrays! So be careful to (re)initialize them if needed.
 
 
 Structs
@@ -765,18 +768,18 @@ sort(array)
 Strings and memory blocks
 ^^^^^^^^^^^^^^^^^^^^^^^^^
 memcopy(from, to, numbytes)
-    Efficiently copy a number of bytes (1 - 256) from a memory location to another.
+    Efficiently copy a number of bytes from a memory location to another.
     NOTE: 'to' must NOT overlap with 'from', unless it is *before* 'from'.
     Because this function imposes some overhead to handle the parameters,
     it is only faster if the number of bytes is larger than a certain threshold.
     Compare the generated code to see if it was beneficial or not.
-    The most efficient will always be to write a specialized copy routine in assembly yourself!
+    The most efficient will often be to write a specialized copy routine in assembly yourself!
 
 memset(address, numbytes, bytevalue)
     Efficiently set a part of memory to the given (u)byte value.
     But the most efficient will always be to write a specialized fill routine in assembly yourself!
-    Note that for clearing the character screen, very fast specialized subroutines are
-    available in the ``txt`` block (part of the ``textio`` module)
+    Note that for clearing the screen, very fast specialized subroutines are
+    available in the ``textio`` and ``graphics`` library modules.
 
 memsetw(address, numwords, wordvalue)
     Efficiently set a part of memory to the given (u)word value.
@@ -785,11 +788,13 @@ memsetw(address, numwords, wordvalue)
 leftstr(source, target, length)
     Copies the left side of the source string of the given length to target string.
     It is assumed the target string buffer is large enough to contain the result.
+    Also, you have to make sure yourself that length is smaller or equal to the length of the source string.
     Modifies in-place, doesn't return a value (so can't be used in an expression).
 
 rightstr(source, target, length)
     Copies the right side of the source string of the given length to target string.
     It is assumed the target string buffer is large enough to contain the result.
+    Also, you have to make sure yourself that length is smaller or equal to the length of the source string.
     Modifies in-place, doesn't return a value (so can't be used in an expression).
 
 strlen(str)
@@ -806,8 +811,15 @@ substr(source, target, start, length)
     Copies a segment from the source string, starting at the given index,
     and of the given length to target string.
     It is assumed the target string buffer is large enough to contain the result.
+    Also, you have to make sure yourself that start and length are within bounds of the strings.
     Modifies in-place, doesn't return a value (so can't be used in an expression).
 
+strcopy(from, to)
+    Copy a string to another, overwriting that one. Returns the length of the string that was copied.
+    Often you don't have to call this explicitly and can just write ``string1 = string2``
+    but this function is useful if you're dealing with addresses for instance.
+
+
 Miscellaneous
 ^^^^^^^^^^^^^
 exit(returncode)
@@ -824,6 +836,12 @@ mkword(msb, lsb)
     Efficiently create a word value from two bytes (the msb and the lsb). Avoids multiplication and shifting.
     So mkword($80, $22) results in $8022.
 
+    .. note::
+        The arguments to the mkword() function are in 'natural' order that is first the msb then the lsb.
+        Don't get confused by how the system actually stores this 16-bit word value in memory (which is
+        in little-endian format, so lsb first then msb)
+
+
 rnd()
     returns a pseudo-random byte from 0..255
 
@@ -888,6 +906,10 @@ set_irqd()  /  clear_irqd()
 swap(x, y)
     Swap the values of numerical variables (or memory locations) x and y in a fast way.
 
+progend()
+    Returns the last address of the program in memory + 1.
+    Can be used to load dynamic data after the program, instead of hardcoding something.
+
 
 Library routines
 ----------------

docs/source/syntaxreference.rst

@@ -122,7 +122,8 @@ Directives
 
     - For a module option, there is ``enable_floats``, which will tell the compiler
       to deal with floating point numbers (by using various subroutines from the Commodore-64 kernal).
-      Otherwise, floating point support is not enabled.
+      Otherwise, floating point support is not enabled. Normally you don't have to use this yourself as
+      importing the ``floats`` library is required anyway and that will enable it for you automatically.
     - There's also ``no_sysinit`` which cause the resulting program to *not* include
       the system re-initialization logic of clearing the screen, resetting I/O config etc. You'll have to
       take care of that yourself. The program will just start running from whatever state the machine is in when the

docs/source/todo.rst

@@ -2,14 +2,16 @@
 TODO
 ====
 
-- get rid of all other TODO's in the code ;-)
-- implement @stack for asmsub parameters
+- see if we can group some errors together for instance the (now single) errors about unidentified symbols
+- Cx16 target: support full-screen 640x480 and 320x240 graphics? That requires our own custom graphics routines though to draw lines.
+- hoist all variable declarations up to the subroutine scope *before* even the constant folding takes place (to avoid undefined symbol errors when referring to a variable from another nested scope in the subroutine)
 - make it possible to use cpu opcodes such as 'nop' as variable names by prefixing all asm vars with something such as '_'
 - option to load the built-in library files from a directory instead of the embedded ones (for easier library development/debugging)
-- see if we can group some errors together for instance the (now single) errors about unidentified symbols
-- use VIC banking to move up the graphics bitmap memory location. Don't move it under the ROM though as that would require IRQ disabling and memory bank swapping for every bitmap manipulation
-- add some primitives/subroutines/examples for using custom char sets, copying the default charset.
-- recursive subroutines? via %option recursive, allocate all params and local vars on estack, don't allow nested subroutines, can begin by first not allowing any local variables just fixing the parameters
+- use VIC banking to move up the graphics bitmap memory location. Move it to $e000 under the kernal rom?
+- some support for recursive subroutines?
+    - via %option recursive?: allocate all params and local vars on estack, don't allow nested subroutines, can begin by first not allowing any local variables just fixing the parameters
+    - Or via a special recursive call operation that copies the current values of all local vars (including arguments) to the stack, replaces the arguments, jsr subroutine, and after returning copy the stack back to the local variables
+- get rid of all other TODO's in the code ;-)
 
 More optimizations
 ^^^^^^^^^^^^^^^^^^
@@ -18,8 +20,7 @@ Add more compiler optimizations to the existing ones.
 
 - further optimize assignment codegeneration, such as the following:
 - binexpr splitting (beware self-referencing expressions and asm code ballooning though)
-- subroutine calling convention? like: 1 byte arg -> pass in A, 2 bytes -> pass in A+Y, return value likewise.  Especially for built-in functions!
-- can such parameter passing to subroutines be optimized to avoid copying?
+- detect var->var argument passing to subroutines and avoid the second variable and copying of the value
 - more optimizations on the language AST level
 - more optimizations on the final assembly source level
 - note: subroutine inlining is abandoned because of problems referencing non-local stuff. Can't move everything around.

examples/arithmetic/builtins.p8

@@ -0,0 +1,840 @@
+%import textio
+%import floats
+%import syslib
+%import test_stack
+%zeropage basicsafe
+
+main {
+
+    sub start() {
+        rotations()
+        strings()
+        integers()
+        floatingpoint()
+
+        test_stack.test()
+    }
+
+    sub rotations() {
+        ubyte[] ubarr = [%11000111]
+        uword[] uwarr = [%1100111110101010]
+
+        repeat(10) {
+            txt.chrout('\n')
+        }
+
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        rol(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        rol(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        rol(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        rol(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        uwarr[0] = %1100111110101010
+        ror(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        ror(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        ror(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        ror(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        ror(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        clear_carry()
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        rol2(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        rol2(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        rol2(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        rol2(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        ror2(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        ror2(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        ror2(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        ror2(uwarr[0])
+        txt.print_uwbin(uwarr[0], true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        rol(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        rol(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        rol(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        rol(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        set_carry()
+        ror(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        ror(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        ror(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        ror(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        set_carry()
+        ror(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        rol2(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        rol2(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        rol2(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        rol2(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        ror2(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        ror2(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        ror2(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        ror2(ubarr[0])
+        txt.print_ubbin(ubarr[0], true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        &ubyte  membyte = $c000
+        uword addr = $c000
+
+        @(addr) = %10110101
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        set_carry()
+        rol(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        set_carry()
+        rol(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        set_carry()
+        rol(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        set_carry()
+        rol(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        @(addr) = %10110101
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        set_carry()
+        ror(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        set_carry()
+        ror(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        set_carry()
+        ror(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        set_carry()
+        ror(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        @(addr) = %10110101
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        rol2(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        rol2(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        rol2(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        rol2(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        @(addr) = %10110101
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        ror2(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        ror2(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        ror2(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        ror2(@(addr))
+        txt.print_ubbin(@(addr), true)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        test_stack.test()
+
+    }
+
+    sub strings() {
+        const uword ADDR = $8400
+        const uword ADDR2 = $8000
+
+        memset(ADDR2, 40*25, '*')
+        memset(ADDR2, 40, '1')
+        memset(ADDR2+24*40, 39, '2')
+        memsetw(ADDR2, 40*25/2, $3132)
+        memsetw(ADDR2, 20, $4142)
+        memsetw(ADDR2+24*40, 19, $4241)
+        memcopy(ADDR2, ADDR, 200)
+
+        str result = "?" *10
+        str s1 = "irmen"
+        str s2 = "hello"
+        str dots = "....."
+
+        ubyte ub
+        byte bb
+        ubyte zero=0
+
+        bb = strcmp(s1, s2)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = strcmp(s2, s1)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        txt.print_ub(s1==s2)
+        txt.chrout('\n')
+        txt.print_ub(s1<s2)
+        txt.chrout('\n')
+        txt.print_ub(s1>s2)
+        txt.chrout('\n')
+        bb = zero+strcmp(s1,s2)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = zero+strcmp(s2,s1)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        ub = strlen(s1)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+strlen(s1)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        leftstr(s1, result, 3)
+        txt.print(result)
+        txt.chrout('\n')
+        leftstr(s1, result, len(s1))
+        txt.print(result)
+        txt.chrout('\n')
+        txt.chrout('\n')
+
+        result = "x"*8
+        rightstr(s2, result, 3)
+        txt.print(result)
+        txt.chrout('\n')
+        rightstr(s2, result, len(s1))
+        txt.print(result)
+        txt.chrout('\n')
+
+        result = "y"*10
+        substr(s2, result, 1, 3)
+        txt.print(result)
+        txt.chrout('\n')
+
+        void strcopy(s2, s1)
+        txt.print_ub(99+strcopy(s2,s1))
+        txt.chrout('\n')
+
+        test_stack.test()
+
+    }
+
+    sub integers() {
+        ubyte[]  ubarr = [1,2,3,4,5,0,4,3,2,1, 255, 255, 255]
+        byte[]  barr = [1,2,3,4,5,-4,0,-3,2,1, -128, -128, -127]
+        uword[]  uwarr = [100,200,300,400,0,500,400,300,200,100]
+        word[] warr = [100,200,300,400,500,0,-400,-300,200,100,-99, -4096]
+
+        ubyte zero=0
+        ubyte ub
+        ubyte ub2
+        byte bb
+        uword uw
+        word ww
+
+        repeat(20) {
+            txt.chrout('\n')
+        }
+
+        ub = read_flags()
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+read_flags()*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+
+        ub = rnd()
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+rnd()*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        uw = rndw()
+        txt.print_uw(uw)
+        txt.chrout('\n')
+        uw = zero+rndw()*1+zero
+        txt.print_uw(uw)
+        txt.chrout('\n')
+
+
+        uw = 50000
+        ub = sqrt16(uw)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+sqrt16(uw)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        bb = -100
+        bb = sgn(bb)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = -100
+        bb = zero+sgn(bb)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        ub = 100
+        bb = sgn(ub)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        ub = 100
+        bb = zero+sgn(ub)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        ww = -1000
+        bb = sgn(ww)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = zero+sgn(ww)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        uw = 1000
+        bb = sgn(uw)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = zero+sgn(uw)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        ub = 0
+        uw = sin16u(ub)
+        txt.print_uw(uw)
+        txt.chrout('\n')
+        uw = zero+sin16u(ub)*1+zero
+        txt.print_uw(uw)
+        txt.chrout('\n')
+
+        ub = 0
+        uw = cos16u(ub)
+        txt.print_uw(uw)
+        txt.chrout('\n')
+        uw = zero+cos16u(ub)*1+zero
+        txt.print_uw(uw)
+        txt.chrout('\n')
+
+        ub = 0
+        ww = sin16(ub)
+        txt.print_w(ww)
+        txt.chrout('\n')
+        ww = zero+sin16(ub)*1+zero
+        txt.print_w(ww)
+        txt.chrout('\n')
+
+        ub = 0
+        ww = cos16(ub)
+        txt.print_w(ww)
+        txt.chrout('\n')
+        uw = 0
+        ww = zero+cos16(ub)*1+zero
+        txt.print_w(ww)
+        txt.chrout('\n')
+
+        ub2 = 0
+        ub = sin8u(ub2)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+sin8u(ub2)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        ub2 = 0
+        ub = cos8u(ub2)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+cos8u(ub2)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        ub2 = 0
+        bb = sin8(ub2)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = zero+sin8(ub2)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        ub2 = 0
+        bb = cos8(ub2)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = zero+cos8(ub2)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        bb = -100
+        bb = abs(bb)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = -100
+        bb = zero+abs(bb)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        ww = -1000
+        ww = abs(ww)
+        txt.print_w(ww)
+        txt.chrout('\n')
+        ww = -1000
+        ww = zero+abs(ww)*1+zero
+        txt.print_w(ww)
+        txt.chrout('\n')
+
+        ub = min(ubarr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+min(ubarr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        bb = min(barr)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = zero+min(barr)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        uw = min(uwarr)
+        txt.print_uw(uw)
+        txt.chrout('\n')
+        uw = zero+min(uwarr)*1+zero
+        txt.print_uw(uw)
+        txt.chrout('\n')
+
+        ww = min(warr)
+        txt.print_w(ww)
+        txt.chrout('\n')
+        ww = zero+min(warr)*1+zero
+        txt.print_w(ww)
+        txt.chrout('\n')
+
+        ub = max(ubarr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+max(ubarr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        bb = max(barr)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        bb = zero+max(barr)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        uw = max(uwarr)
+        txt.print_uw(uw)
+        txt.chrout('\n')
+        uw = zero+max(uwarr)*1+zero
+        txt.print_uw(uw)
+        txt.chrout('\n')
+
+        ww = max(warr)
+        txt.print_w(ww)
+        txt.chrout('\n')
+        ww = zero+max(warr)*1+zero
+        txt.print_w(ww)
+        txt.chrout('\n')
+
+        ub = any(ubarr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+any(ubarr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        ub = any(barr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+any(barr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        ub = any(uwarr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+any(uwarr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        ub = any(warr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+any(warr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        ub = all(ubarr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+all(ubarr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        ub = all(barr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+all(barr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        ub = all(uwarr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+all(uwarr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        ub = all(warr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+all(warr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+
+        uw = sum(ubarr)
+        txt.print_uw(uw)
+        txt.chrout('\n')
+        uw = zero+sum(ubarr)*1+zero
+        txt.print_uw(uw)
+        txt.chrout('\n')
+
+        ww = sum(barr)
+        txt.print_w(ww)
+        txt.chrout('\n')
+        ww = zero+sum(barr)*1+zero
+        txt.print_w(ww)
+        txt.chrout('\n')
+
+        uw = sum(uwarr)
+        txt.print_uw(uw)
+        txt.chrout('\n')
+        uw = zero+sum(uwarr)*1+zero
+        txt.print_uw(uw)
+        txt.chrout('\n')
+
+        ww = sum(warr)
+        txt.print_w(ww)
+        txt.chrout('\n')
+        ww = zero+sum(warr)*1+zero
+        txt.print_w(ww)
+        txt.chrout('\n')
+
+        sort(ubarr)
+        sort(barr)
+        sort(uwarr)
+        sort(warr)
+        reverse(ubarr)
+        reverse(barr)
+        reverse(uwarr)
+        reverse(warr)
+
+        test_stack.test()
+    }
+
+    sub floatingpoint() {
+        ubyte[]  barr = [1,2,3,4,5,0,4,3,2,1]
+        float[] flarr = [1.1, 2.2, 3.3, 0.0, -9.9, 5.5, 4.4]
+
+        ubyte zero=0
+        ubyte ub
+        byte bb
+        uword uw
+        float fl
+        float fzero=0.0
+
+        fl = -9.9
+        fl = abs(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = -9.9
+        fl = fzero+abs(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = 9.9
+        fl = atan(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = 9.9
+        fl = fzero+atan(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = -9.9
+        fl = ceil(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = -9.9
+        fl = fzero+ceil(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = -9.9
+        fl = cos(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = -9.9
+        fl = fzero+cos(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = -9.9
+        fl = sin(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = -9.9
+        fl = fzero+sin(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = 9.9
+        fl = tan(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = 9.9
+        fl = fzero+tan(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = 3.1415927
+        fl = deg(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = 3.1415927
+        fl = fzero+deg(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = 90
+        fl = rad(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = 90
+        fl = fzero+rad(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = -9.9
+        fl = floor(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = -9.9
+        fl = fzero+floor(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = 3.1415927
+        fl = ln(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = 3.1415927
+        fl = fzero+ln(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = 3.1415927
+        fl = log2(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = 3.1415927
+        fl = fzero+log2(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = -9.9
+        fl = round(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = -9.9
+        fl = fzero+round(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = -9.9
+        bb = sgn(fl)
+        txt.print_b(bb)
+        txt.chrout('\n')
+        fl = -9.9
+        bb = zero+sgn(fl)*1+zero
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        fl = 3.1415927
+        fl = sqrt(fl)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = 3.1415927
+        fl = fzero+sqrt(fl)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        fl = rndf()
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = fzero+rndf()*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        swap(fl, fzero)
+        swap(fzero, fl)
+
+        ub = any(flarr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+any(flarr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = all(flarr)
+        txt.print_ub(ub)
+        txt.chrout('\n')
+        ub = zero+all(flarr)*1+zero
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        reverse(flarr)
+        for ub in 0 to len(flarr)-1 {
+            floats.print_f(flarr[ub])
+            txt.chrout(',')
+        }
+        txt.chrout('\n')
+        fl = max(flarr)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = fzero+max(flarr)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = min(flarr)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = fzero+min(flarr)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = sum(flarr)
+        floats.print_f(fl)
+        txt.chrout('\n')
+        fl = fzero+sum(flarr)*1.0+fzero
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        test_stack.test()
+    }
+}

examples/arithmetic/minus.p8

@@ -1,5 +1,6 @@
 %import floats
 %import textio
+%import test_stack
 %zeropage basicsafe
 
 main {
@@ -31,6 +32,8 @@ main {
         minus_float(0,0,0)
         minus_float(2.5,1.5,1.0)
         minus_float(-1.5,3.5,-5.0)
+
+        test_stack.test()
     }
 
     sub minus_ubyte(ubyte a1, ubyte a2, ubyte c) {
@@ -97,10 +100,11 @@ main {
         float r = a1-a2
         if abs(r-c)<0.00001
             txt.print(" ok  ")
-        else
+        else {
             txt.print("err! ")
+        }
 
-        txt.print("float ")
+        txt.print("  float ")
         floats.print_f(a1)
         txt.print(" - ")
         floats.print_f(a2)

examples/balloonflight.p8

@@ -1,8 +1,15 @@
 %target c64
 %import syslib
 %import textio
+%import test_stack
 %zeropage basicsafe
 
+; note: The flickering in the scrolling is caused by the CPU requiring
+;       too long to scroll the characters + the colors in course scroll.
+;       This takes nearly a full frame to accomplish, and causes tearing.
+;       It's very difficult to remove this flicker: it requires double buffering
+;       and splitting the coarse character scrolling on multiple phases...
+
 main {
 
     ubyte perform_scroll = false
@@ -16,13 +23,16 @@ main {
 
         c64.SCROLX &= %11110111     ; 38 column mode
 
-        c64.set_rasterirq(1)     ; enable animation
+        c64.set_rasterirq(200)     ; enable animation
 
         ubyte target_height = 10
         ubyte active_height = 24
         ubyte upwards = true
 
         repeat {
+            ;txt.plot(0,0)
+            ;test_stack.test()
+
             ubyte mountain = 223        ; slope upwards
             if active_height < target_height {
                 active_height++
@@ -119,10 +129,10 @@ spritedata $0f00 {
 
 
 irq {
-    ubyte smoothx=7
+    ubyte smoothx=0
     sub irq() {
         smoothx = (smoothx-1) & 7
-        main.perform_scroll = smoothx==0
+        main.perform_scroll = smoothx==7
         c64.SCROLX = (c64.SCROLX & %11111000) | smoothx
     }
 }

examples/balls.p8

@@ -0,0 +1,103 @@
+%import textio
+%import test_stack
+
+%zeropage basicsafe
+
+; Note: this program is compatible with C64 and CX16.
+
+main {
+
+    sub start() {
+        str  input = ".........."
+        ubyte ballCount
+        ubyte[255] BX
+        ubyte[255] BY
+        ubyte[255] BC
+        ubyte[255] DX
+        ubyte[255] DY
+
+        txt.print("number of balls (1-255)? ")
+        void txt.input_chars(input)
+        ballCount = conv.str2ubyte(input)
+        txt.fill_screen(81, 0)
+
+        ; Setup Starting Ball Positions
+        ubyte lp
+        for lp in 0 to ballCount-1 {
+            BX[lp] = rnd() % txt.DEFAULT_WIDTH
+            BY[lp] = rnd() % txt.DEFAULT_HEIGHT
+            BC[lp] = rnd() & 15
+            DX[lp] = rnd() & 1
+            DY[lp] = rnd() & 1
+            void rnd()
+        }
+
+        ; start clock
+        c64.SETTIM(0,0,0)
+
+        ; display balls
+        uword frame
+        for frame in 0 to 999 {
+            ; Loop though all balls clearing current spot and setting new spot
+            for lp in 0 to ballCount-1 {
+
+                ; Clear existing Location the ball is at
+                txt.setclr(BX[lp], BY[lp], 0)
+
+                if DX[lp] == 0 {
+                    if (BX[lp] == 0)
+                    {
+                        DX[lp] = 1
+                    } else {
+                        BX[lp]=BX[lp]-1
+                    }
+                } else if DX[lp] == 1 {
+                    if (BX[lp] == txt.DEFAULT_WIDTH-1)
+                    {
+                        BX[lp] = txt.DEFAULT_WIDTH-2
+                        DX[lp] = 0
+                    } else {
+                        BX[lp]=BX[lp]+1
+                    }
+                }
+                if DY[lp] == 0 {
+                    if (BY[lp] == 0)
+                    {
+                        DY[lp] = 1
+                    } else {
+                        BY[lp]=BY[lp]-1
+                    }
+                } else if DY[lp] == 1 {
+                    if (BY[lp] == txt.DEFAULT_HEIGHT-1)
+                    {
+                        BY[lp] = txt.DEFAULT_HEIGHT-2
+                        DY[lp] = 0
+                    } else {
+                        BY[lp]=BY[lp]+1
+                    }
+                }
+
+                ; Put the new ball possition
+                txt.setclr(BX[lp], BY[lp], BC[lp])
+            }
+
+            ;txt.plot(0,0)
+            ;txt.print_uw(frame)
+        }
+
+        ; read clock
+        uword jiffies
+        %asm {{
+            stx  P8ZP_SCRATCH_REG
+            jsr  c64.RDTIM
+            sta  jiffies
+            stx  jiffies+1
+            ldx  P8ZP_SCRATCH_REG
+        }}
+        txt.print("\nbenchmark: ")
+        txt.print_uw(jiffies)
+        txt.print(" jiffies for 1000 frames.\n")
+
+        ; test_stack.test()
+    }
+}

examples/charset.p8

@@ -0,0 +1,85 @@
+%target c64
+%import syslib
+%import textio
+%zeropage basicsafe
+%option no_sysinit
+
+;  Create a custom character set on the C64.
+
+main {
+
+    sub start() {
+        txt.color(1)
+        txt.print("creating charset...\n")
+        charset.make_custom_charset()
+
+        ; activate the new charset in RAM
+        ubyte block = c64.CIA2PRA
+        const ubyte PAGE1 = ((c64.Screen >> 6) & $F0) | ((charset.CHARSET >> 10) & $0E)
+
+        c64.CIA2PRA = (block & $FC) | (lsb(c64.Screen >> 14) ^ $03)
+        c64.VMCSB = PAGE1
+
+        txt.print("\n @ @ @ @\n")
+    }
+}
+
+charset {
+    const uword CHARSET = $2000
+
+    sub copy_rom_charset() {
+        ; copies the charset from ROM to RAM so we can modify it
+
+        set_irqd()
+        ubyte bank = @($0001)
+        @($0001) = bank & %11111011     ; enable CHAREN, so the character rom accessible at $d000
+        memcopy($d000, CHARSET, 256*8*2)  ; copy the charset to RAM
+
+        @($0001) = bank         ; reset previous memory banking
+        clear_irqd()
+    }
+
+    sub make_custom_charset() {
+        copy_rom_charset()
+
+        ; make all characters italic
+        ubyte c
+        for c in 0 to 255 {
+            uword ptr = CHARSET + c*$0008
+            @(ptr) >>= 2
+            @(ptr+1) >>= 2
+            @(ptr+2) >>= 1
+            @(ptr+3) >>= 1
+            ;@(ptr+4) >>= 0
+            ;@(ptr+5) >>= 0
+            @(ptr+6) <<= 1
+            @(ptr+7) <<= 1
+
+            ptr = CHARSET + 256*8 + c*$0008
+            @(ptr) >>= 2
+            @(ptr+1) >>= 2
+            @(ptr+2) >>= 1
+            @(ptr+3) >>= 1
+            ;@(ptr+4) >>= 0
+            ;@(ptr+5) >>= 0
+            @(ptr+6) <<= 1
+            @(ptr+7) <<= 1
+        }
+
+        ; add a smiley over the '@'
+
+        ubyte[] smiley = [
+            %00111100,
+            %01000010,
+            %10100101,
+            %10000001,
+            %10100101,
+            %10011001,
+            %01000010,
+            %00111100
+        ]
+
+        memcopy(smiley, CHARSET, len(smiley))
+
+    }
+}

examples/cmp/word_comps.p8

@@ -0,0 +1,913 @@
+%import textio
+%import floats
+%zeropage basicsafe
+%import test_stack
+
+main {
+
+    sub start() {
+        word_less()
+        word_lessequal()
+        word_greaterequal()
+        word_greater()
+        uword_lessequal()
+    }
+
+    sub uword_lessequal() {
+        uword lessvar
+        uword comparevar
+
+        txt.print("uword <=\n")
+
+        txt.print_uw(65535)
+        txt.chrout('\n')
+        check_lesseq_uw(0, 65535)
+        txt.print_uw(0)
+        txt.chrout('\n')
+        check_not_lesseq_uw(65535, 0)
+
+        comparevar = 65535
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto 0  {
+            check_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 65535-2
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto 0  {
+            check_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 65535-254
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto 0  {
+            check_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 65535-255
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto 0  {
+            check_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 65535-256
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto 0  {
+            check_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 65535-5000
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto 0  {
+            check_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 32769
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto 0  {
+            check_lesseq_uw(lessvar, comparevar)
+        }
+
+
+        comparevar = 32768
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in 65535 downto comparevar+1  {
+            check_not_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 1
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in 65535 downto comparevar+1  {
+            check_not_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 0
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in 65535 downto comparevar+1  {
+            check_not_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 11111
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in 65535 downto comparevar+1  {
+            check_not_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 255
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in 65535 downto comparevar+1  {
+            check_not_lesseq_uw(lessvar, comparevar)
+        }
+
+        comparevar = 256
+        txt.print_uw(comparevar)
+        txt.chrout('\n')
+        for lessvar in 65535 downto comparevar+1  {
+            check_not_lesseq_uw(lessvar, comparevar)
+        }
+
+
+        test_stack.test()
+        return
+
+        sub check_lesseq_uw(uword w1, uword w2) {
+            uword zero = 0
+            ubyte error=0
+
+            ubyte ub = w1<=w2
+            if not ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1<=(w2+zero) {
+                zero = 0 ; dummy
+            } else {
+                error++
+                txt.print("c!")
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_uw(w1)
+                txt.print(" <= ")
+                txt.print_uw(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+
+        sub check_not_lesseq_uw(uword w1, uword w2) {
+            uword zero = 0
+            ubyte error=0
+
+            ubyte ub = w1<=w2
+            if ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1<=(w2+zero) {
+                error++
+                txt.print("c!")
+            } else {
+                zero = 0 ; dummy
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_uw(w1)
+                txt.print(" not <= ")
+                txt.print_uw(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+    }
+
+    sub word_greater() {
+        word biggervar
+        word comparevar
+
+        txt.print("word >\n")
+
+        txt.print_w(-32767)
+        txt.chrout('\n')
+        check_greater_w(32767, -32767)
+        txt.print_w(32766)
+        txt.chrout('\n')
+        check_not_greater_w(-32766, 32766)
+
+        comparevar = 32765
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in comparevar downto -32768  {
+            check_not_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = -1
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in comparevar downto -32768  {
+            check_not_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = 0
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in comparevar downto -32768  {
+            check_not_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = 11111
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in comparevar downto -32768  {
+            check_not_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = 0
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = -2
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = -254
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = -255
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = -256
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = -5000
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = 1
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = 255
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = 256
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = 257
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+        comparevar = 32760
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar+1  {
+            check_greater_w(biggervar, comparevar)
+        }
+
+
+        test_stack.test()
+        return
+
+        sub check_greater_w(word w1, word w2) {
+            word zero = 0
+            ubyte error=0
+
+            ubyte ub = w1>(w2+zero)
+            if not ub {
+                error++
+                txt.print("ubz!")
+            }
+
+            ub = w1>w2
+            if not ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1>(w2+zero) {
+                zero = 0 ; dummy
+            } else {
+                error++
+                txt.print("c!")
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_w(w1)
+                txt.print(" > ")
+                txt.print_w(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+
+        sub check_not_greater_w(word w1, word w2) {
+            word zero = 0
+            ubyte error=0
+
+            ubyte ub = w1>w2
+            if ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1>(w2+zero) {
+                error++
+                txt.print("c!")
+            } else {
+                zero = 0 ; dummy
+            }
+
+            if w1>w2 {
+                error++
+                txt.print("c2!")
+            } else {
+                zero = 0 ; dummy
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_w(w1)
+                txt.print(" not > ")
+                txt.print_w(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+    }
+
+    sub word_greaterequal() {
+        word biggervar
+        word comparevar
+
+        txt.print("word >=\n")
+
+        txt.print_w(-32767)
+        txt.chrout('\n')
+        check_greatereq_w(32767, -32767)
+        txt.print_w(32766)
+        txt.chrout('\n')
+        check_not_greatereq_w(-32766, 32766)
+
+        comparevar = 32765
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in comparevar-1 downto -32768  {
+            check_not_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = -1
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in comparevar-1 downto -32768  {
+            check_not_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = 0
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in comparevar-1 downto -32768  {
+            check_not_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = 11111
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in comparevar-1 downto -32768  {
+            check_not_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = 0
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = -2
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = -254
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = -255
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = -256
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = -5000
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = 1
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = 255
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = 256
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = 257
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+        comparevar = 32767
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for biggervar in 32767 downto comparevar  {
+            check_greatereq_w(biggervar, comparevar)
+        }
+
+
+        test_stack.test()
+        return
+
+        sub check_greatereq_w(word w1, word w2) {
+            word zero = 0
+            ubyte error=0
+
+            ubyte ub = w1>=(w2+zero)
+            if not ub {
+                error++
+                txt.print("ubz!")
+            }
+
+            ub = w1>=w2
+            if not ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1>=(w2+zero) {
+                zero = 0 ; dummy
+            } else {
+                error++
+                txt.print("c!")
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_w(w1)
+                txt.print(" >= ")
+                txt.print_w(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+
+        sub check_not_greatereq_w(word w1, word w2) {
+            word zero = 0
+            ubyte error=0
+
+            ubyte ub = w1>=w2
+            if ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1>=(w2+zero) {
+                error++
+                txt.print("c!")
+            } else {
+                zero = 0 ; dummy
+            }
+
+            if w1>=w2 {
+                error++
+                txt.print("c2!")
+            } else {
+                zero = 0 ; dummy
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_w(w1)
+                txt.print(" not >= ")
+                txt.print_w(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+    }
+
+    sub word_lessequal() {
+        word lessvar
+        word comparevar
+
+        txt.print("word <=\n")
+
+        txt.print_w(32767)
+        txt.chrout('\n')
+        check_lesseq_w(-32767, 32767)
+        txt.print_w(-32767)
+        txt.chrout('\n')
+        check_not_lesseq_w(32767, -32767)
+
+        comparevar = 0
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = -2
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = -254
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = -255
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = -256
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = -5000
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = 1
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = 255
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = 256
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = 257
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = 32767
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in comparevar downto -32768  {
+            check_lesseq_w(lessvar, comparevar)
+        }
+
+
+        comparevar = -32768
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 32766 downto comparevar+1  {
+            check_not_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = -1
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 32766 downto comparevar+1  {
+            check_not_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = 0
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 32766 downto comparevar+1  {
+            check_not_lesseq_w(lessvar, comparevar)
+        }
+
+        comparevar = 11111
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 32766 downto comparevar+1  {
+            check_not_lesseq_w(lessvar, comparevar)
+        }
+
+        test_stack.test()
+        return
+
+        sub check_lesseq_w(word w1, word w2) {
+            word zero = 0
+            ubyte error=0
+
+            ubyte ub = w1<=w2
+            if not ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1<=(w2+zero) {
+                zero = 0 ; dummy
+            } else {
+                error++
+                txt.print("c!")
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_w(w1)
+                txt.print(" <= ")
+                txt.print_w(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+
+        sub check_not_lesseq_w(word w1, word w2) {
+            word zero = 0
+            ubyte error=0
+
+            ubyte ub = w1<=w2
+            if ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1<=(w2+zero) {
+                error++
+                txt.print("c!")
+            } else {
+                zero = 0 ; dummy
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_w(w1)
+                txt.print(" not <= ")
+                txt.print_w(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+    }
+
+    sub word_less() {
+        word lessvar
+        word comparevar
+
+        txt.print("word <\n")
+
+        txt.print_w(32767)
+        txt.chrout('\n')
+        check_less_w(-32767, 32767)
+        txt.print_w(-32767)
+        txt.chrout('\n')
+        check_not_less_w(32767, -32767)
+
+        comparevar = 0
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in -1 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = -2
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in -3 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = -254
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in -255 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = -255
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in -256 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = -256
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in -257 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = -5000
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in -5001 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = 1
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 0 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = 255
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 254 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = 256
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 255 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = 257
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 256 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = 32767
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 32766 downto -32768  {
+            check_less_w(lessvar, comparevar)
+        }
+
+        comparevar = -32768
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 32766 downto -32768  {
+            check_not_less_w(lessvar, comparevar)
+        }
+
+        comparevar = -1
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 32766 downto -1  {
+            check_not_less_w(lessvar, comparevar)
+        }
+
+        comparevar = 0
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 32766 downto 0  {
+            check_not_less_w(lessvar, comparevar)
+        }
+
+        comparevar = 11111
+        txt.print_w(comparevar)
+        txt.chrout('\n')
+        for lessvar in 32766 downto 11111  {
+            check_not_less_w(lessvar, comparevar)
+        }
+
+        test_stack.test()
+        return
+
+        sub check_less_w(word w1, word w2) {
+            word zero = 0
+            ubyte error=0
+
+            ubyte ub = w1<w2
+            if not ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1<(w2+zero) {
+                zero = 0 ; dummy
+            } else {
+                error++
+                txt.print("c!")
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_w(w1)
+                txt.print(" < ")
+                txt.print_w(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+
+        sub check_not_less_w(word w1, word w2) {
+            word zero = 0
+            ubyte error=0
+
+            ubyte ub = w1<w2
+            if ub {
+                error++
+                txt.print("ub!")
+            }
+
+            if w1<(w2+zero) {
+                error++
+                txt.print("c!")
+            } else {
+                zero = 0 ; dummy
+            }
+
+            if error {
+                txt.print("  ")
+                txt.print_w(w1)
+                txt.print(" not < ")
+                txt.print_w(w2)
+                txt.chrout('\n')
+                exit(1)
+            }
+        }
+    }
+
+}

examples/cube3d-float.p8

@@ -1,5 +1,6 @@
 %import floats
 %import textio
+%import test_stack
 %zeropage basicsafe
 
 ; Note: this program is compatible with C64 and CX16.
@@ -41,6 +42,8 @@ main {
             txt.print(" jiffies/fr = ")
             txt.print_ub(60/timer_jiffies)
             txt.print(" fps")
+
+            ;test_stack.test()
         }
     }