updated the compiled examples disk

version 5.1
stack tested for most example programs
2026-10-22 20:02:27 +00:00 · 2020-11-22 18:45:40 +01:00 · 2020-11-22 18:36:04 +01:00 · 2020-11-22 18:35:43 +01:00 · 2020-11-22 17:44:55 +01:00 · 2020-11-22 16:54:02 +01:00
105 changed files with 5280 additions and 2588 deletions
@@ -55,13 +55,66 @@ 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
+		stx  P8ZP_SCRATCH_REG
+		jsr  MOVFM
+		jsr  GETADR     ; into Y/A
+		ldx  P8ZP_SCRATCH_REG
+		rts
+		.pend
+
+cast_as_w_into_ay	.proc               ; also used for float 2 b
+		; -- cast float at A/Y to word into A/Y
+		stx  P8ZP_SCRATCH_REG
+		jsr  MOVFM
+		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 +124,7 @@ 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 +135,7 @@ stack_ub2float	.proc
 		tay
 		lda  #0
 		jsr  GIVAYF
-		jmp  push_fac1_as_result
+		jmp  push_fac1._internal
 		.pend

 stack_uw2float	.proc
@@ -92,7 +145,7 @@ 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 +199,6 @@ 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 +235,6 @@ 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 +300,11 @@ 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 +313,7 @@ push_fac1_as_result	.proc
 		jmp  push_float
 		.pend

+
 pow_f		.proc
 		; -- push f1 ** f2 on stack
 		lda  #<fmath_float2
@@ -299,8 +329,7 @@ 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 +339,7 @@ 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 +349,7 @@ 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 +359,7 @@ 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 +369,7 @@ 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 +380,33 @@ 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
+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 +447,40 @@ 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 +542,22 @@ _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 +601,3 @@ 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
@@ -213,5 +213,6 @@ sub  print_f  (float value) {
 }

 %asminclude "library:c64/floats.asm", ""
+%asminclude "library: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
@@ -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 = %00111000
+        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 = %00011000
+        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)
@@ -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
 	}}
@@ -149,5 +149,6 @@ sub  print_f  (float value) {
 }

 %asminclude "library:c64/floats.asm", ""
+%asminclude "library:c64/floats_funcs.asm", ""

 }
@@ -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)     ; TODO 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()
@@ -151,8 +151,7 @@ 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)
@@ -1244,45 +1244,6 @@ mul_word_100	.proc
 ; ----------- 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.

@@ -6,4 +6,5 @@

 prog8_lib {
 	%asminclude "library:prog8_lib.asm", ""
+	%asminclude "library:prog8_funcs.asm", ""
 }
@@ -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
+        }}
+    }
+}
@@ -1 +1 @@
-4.6
+5.1
@@ -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 {
@@ -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 -> "?????"
@@ -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)

@@ -288,7 +280,6 @@ 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 +313,6 @@ 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 +365,6 @@ 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) {}
 }

@@ -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())
 }
@@ -30,13 +30,14 @@ enum class DataType {
                UWORD -> targetType in setOf(UWORD, FLOAT)
                WORD -> targetType in setOf(WORD, FLOAT)
                FLOAT -> targetType == FLOAT
-                STR -> targetType == STR
-                in ArrayDatatypes -> targetType == this
+                STR -> targetType == STR || targetType == UWORD
+                in ArrayDatatypes -> targetType == this || targetType == UWORD
                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 {
@@ -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)
            }

@@ -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()
@@ -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")
                }
            }

@@ -344,11 +344,11 @@ 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)
                }
            }
@@ -378,9 +378,9 @@ 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
@@ -433,7 +433,7 @@ 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) {
@@ -593,8 +593,8 @@ 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)
@@ -769,7 +769,11 @@ 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 +988,8 @@ 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 +1222,7 @@ 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 +1354,7 @@ 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)
        }

@@ -168,7 +168,7 @@ 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 +279,7 @@ 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)
        }
@@ -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
+                                }
                            }
                        }
                    }
@@ -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 +40,10 @@ 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) }
+            if(argITypes.any { !it.isKnown })
+                throw FatalAstException("unknown dt")
+            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 +71,12 @@ 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"
                    }
                }
            }
@@ -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)
@@ -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,
@@ -444,9 +448,9 @@ 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 +641,6 @@ 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 +679,18 @@ 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 = mutableMapOf<String, MutableSet<Statement>>()
+    var usedRegsaveA = false
+    var usedRegsaveX = false
+    var usedRegsaveY = false
+    var usedFloatEvalResultVar = false
+}
+
 // 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 +705,24 @@ 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) {
@@ -975,7 +1006,6 @@ 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
@@ -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) {
@@ -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
    }

@@ -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

@@ -14,14 +14,12 @@ 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
@@ -200,11 +198,7 @@ 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 +225,12 @@ 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 +251,7 @@ 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 +548,20 @@ 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 +572,14 @@ 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 +626,7 @@ 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 +736,8 @@ 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 +745,16 @@ 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 +799,8 @@ internal class AsmGen(private val program: Program,
                out("_prog8_regsaveX     .byte  0")
            if(sub.asmGenInfo.usedRegsaveY)
                out("_prog8_regsaveY     .byte  0")
+            if(sub.asmGenInfo.usedFloatEvalResultVar)
+                out("_prog8_float_eval_result    .byte 0,0,0,0,0")
            vardecls2asm(sub.statements)
            out("  .pend\n")
        }
@@ -912,15 +905,16 @@ 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 +1010,16 @@ $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 +1027,7 @@ $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) {
@@ -1139,7 +1133,9 @@ $counterVar    .byte  0""")
            "%breakpoint" -> {
                val label = "_prog8_breakpoint_${breakpointLabels.size+1}"
                breakpointLabels.add(label)
-                out("$label\tnop")
+                out("""
+                    nop
+$label              nop""")
            }
        }
    }
@@ -1165,7 +1161,39 @@ $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
+                    when (returnvalue) {
+                        is NumericLiteralValue -> throw AssemblyError("float literal should have been changed to auto var")
+                        is IdentifierReference -> {
+                            val asmVar = asmVariableName(returnvalue)
+                            out("  lda  #<${asmVar} |  ldy  #>${asmVar} |  jsr  floats.MOVFM")
+                        }
+                        else -> {
+                            // todo evaluate directly into fac1 instead of via stack intermediate  (add RegisterOrPair.FAC1/FAC2 ??)
+                            translateExpression(returnvalue)
+                            out("  jsr  floats.pop_float_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")
    }

@@ -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()
@@ -1,12 +1,17 @@
 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.ArrayIndex
 import prog8.ast.statements.DirectMemoryWrite
 import prog8.ast.statements.FunctionCallStatement
+import prog8.ast.statements.Subroutine
 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
@@ -17,37 +22,40 @@ 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 +71,110 @@ 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_$functionName")
+            "strlen" -> funcStrlen(fcall, resultToStack)
+            "strcmp" -> funcStrcmp(fcall, func, resultToStack, scope)
+            "memcopy", "memset", "memsetw" -> funcMemSetCopy(fcall, func, scope)
+            "substr", "leftstr", "rightstr" -> {
+                translateArguments(fcall.args, func, scope)
+                asmgen.out("  jsr  prog8_lib.func_${func.name}")
            }
+            "exit" -> {
+                translateArguments(fcall.args, func, scope)
+                asmgen.out("  jmp  prog8_lib.func_exit")
+            }
+            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 +184,30 @@ 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 +223,21 @@ 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 +253,7 @@ 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 +261,7 @@ 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 +275,7 @@ 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 +296,7 @@ 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 +304,11 @@ 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 +321,7 @@ 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 +342,7 @@ 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 +350,7 @@ 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 +364,7 @@ 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 +385,7 @@ 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 +393,11 @@ 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 +410,7 @@ 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 -> {
@@ -346,66 +424,124 @@ 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) {
+    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 $dt")
+            }
+        }
+    }
+
+    private fun funcStrlen(fcall: IFunctionCall, resultToStack: Boolean) {
        val name = asmgen.asmVariableName(fcall.args[0] as IdentifierReference)
        val type = fcall.args[0].inferType(program)
        when {
-            type.istype(DataType.STR) -> asmgen.out("""
-                lda  #<$name
-                ldy  #>$name
-                jsr  prog8_lib.strlen
-                sta  P8ESTACK_LO,x
-                dex""")
-            type.istype(DataType.UWORD) -> asmgen.out("""
-                lda  $name
-                ldy  $name+1
-                jsr  prog8_lib.strlen
-                sta  P8ESTACK_LO,x
-                dex""")
+            type.istype(DataType.STR) -> asmgen.out("  lda  #<$name |  ldy  #>$name")
+            type.istype(DataType.UWORD) -> asmgen.out("  lda  $name |  ldy  $name+1")
            else -> throw AssemblyError("strlen requires str or uword arg")
        }
+        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) {
@@ -444,7 +580,7 @@ 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 +616,10 @@ 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 +641,8 @@ 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 +652,43 @@ 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 +921,52 @@ 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 +974,114 @@ 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.usedFloatEvalResultVar = true
+                    val variable = IdentifierReference(listOf("_prog8_float_eval_result"), 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")
+            }
        }
    }

@@ -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())
 }
@@ -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.asmReturnvaluesRegisters.any {it.statusflag!=null}
+        translateFunctionCall(stmt, preserveStatusRegisterAfterCall)
+        // functioncalls no longer return results on the stack, so simply ignore the results in the registers
+        if(preserveStatusRegisterAfterCall)
+            asmgen.out("  plp\t; 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()
        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.
@@ -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)
@@ -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 -> {
@@ -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

@@ -58,7 +55,10 @@ internal class AsmAssignTarget(val kind: TargetStorageKind,

    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)
@@ -101,6 +101,14 @@ 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 +129,30 @@ 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)
                }
            }
        }
@@ -10,6 +10,8 @@ import prog8.compiler.target.CpuType
 import prog8.compiler.target.c64.codegen.AsmGen
 import prog8.compiler.target.c64.codegen.ExpressionsAsmGen
 import prog8.compiler.toHex
+import prog8.functions.BuiltinFunctions
+import prog8.functions.builtinFunctionReturnType


 internal class AssignmentAsmGen(private val program: Program, private val asmgen: AsmGen, private val exprAsmgen: ExpressionsAsmGen) {
@@ -46,7 +48,13 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                val variable = assign.source.asmVarname
                when (assign.target.datatype) {
                    DataType.UBYTE, DataType.BYTE -> assignVariableByte(assign.target, variable)
-                    DataType.UWORD, DataType.WORD -> assignVariableWord(assign.target, variable)
+                    DataType.WORD -> assignVariableWord(assign.target, variable)
+                    DataType.UWORD -> {
+                        if(assign.source.datatype in PassByReferenceDatatypes)
+                            assignAddressOf(assign.target, variable)
+                        else
+                            assignVariableWord(assign.target, variable)
+                    }
                    DataType.FLOAT -> assignVariableFloat(assign.target, variable)
                    DataType.STR -> assignVariableString(assign.target, variable)
                    else -> throw AssemblyError("unsupported assignment target type ${assign.target.datatype}")
@@ -60,12 +68,18 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                    // constant array index value
                    val indexValue = value.indexer.constIndex()!! * elementDt.memorySize()
                    when (elementDt) {
-                        in ByteDatatypes ->
-                            asmgen.out("  lda  $arrayVarName+$indexValue |  sta  P8ESTACK_LO,x |  dex")
-                        in WordDatatypes ->
-                            asmgen.out("  lda  $arrayVarName+$indexValue |  sta  P8ESTACK_LO,x |  lda  $arrayVarName+$indexValue+1 |  sta  P8ESTACK_HI,x | dex")
-                        DataType.FLOAT ->
-                            asmgen.out("  lda  #<$arrayVarName+$indexValue |  ldy  #>$arrayVarName+$indexValue |  jsr  floats.push_float")
+                        in ByteDatatypes -> {
+                            asmgen.out("  lda  $arrayVarName+$indexValue")
+                            assignRegisterByte(assign.target, CpuRegister.A)
+                        }
+                        in WordDatatypes -> {
+                            asmgen.out("  lda  $arrayVarName+$indexValue |  ldy  $arrayVarName+$indexValue+1")
+                            assignRegisterpairWord(assign.target, RegisterOrPair.AY)
+                        }
+                        DataType.FLOAT -> {
+                            asmgen.out("  lda  #<$arrayVarName+$indexValue |  ldy  #>$arrayVarName+$indexValue")
+                            assignFloatFromAY(assign.target)
+                        }
                        else ->
                            throw AssemblyError("weird array type")
                    }
@@ -73,11 +87,13 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                    when (elementDt) {
                        in ByteDatatypes -> {
                            asmgen.loadScaledArrayIndexIntoRegister(value, elementDt, CpuRegister.Y)
-                            asmgen.out("  lda  $arrayVarName,y |  sta  P8ESTACK_LO,x |  dex")
+                            asmgen.out("  lda  $arrayVarName,y")
+                            assignRegisterByte(assign.target, CpuRegister.A)
                        }
                        in WordDatatypes  -> {
                            asmgen.loadScaledArrayIndexIntoRegister(value, elementDt, CpuRegister.Y)
-                            asmgen.out("  lda  $arrayVarName,y |  sta  P8ESTACK_LO,x |  lda  $arrayVarName+1,y |  sta  P8ESTACK_HI,x |  dex")
+                            asmgen.out("  lda  $arrayVarName,y |  pha |  lda  $arrayVarName+1,y |  tay |  pla")
+                            assignRegisterpairWord(assign.target, RegisterOrPair.AY)
                        }
                        DataType.FLOAT -> {
                            asmgen.loadScaledArrayIndexIntoRegister(value, elementDt, CpuRegister.A)
@@ -87,13 +103,13 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                                adc  #<$arrayVarName
                                bcc  +
                                iny
-+                               jsr  floats.push_float""")
+""")
+                            assignFloatFromAY(assign.target)
                        }
                        else ->
                            throw AssemblyError("weird array elt type")
                    }
                }
-                assignStackValue(assign.target)
            }
            SourceStorageKind.MEMORY -> {
                val value = assign.source.memory!!
@@ -106,8 +122,8 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                        assignMemoryByte(assign.target, null, value.addressExpression as IdentifierReference)
                    }
                    else -> {
-                        asmgen.translateExpression(value.addressExpression)
-                        asmgen.out("  jsr  prog8_lib.read_byte_from_address_on_stack |  inx")
+                        asmgen.assignExpressionToVariable(value.addressExpression, asmgen.asmVariableName("P8ZP_SCRATCH_W2"), DataType.UWORD, assign.target.scope)
+                        asmgen.out("  ldy  #0 |  lda  (P8ZP_SCRATCH_W2),y")
                        assignRegisterByte(assign.target, CpuRegister.A)
                    }
                }
@@ -124,32 +140,77 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                    is DirectMemoryRead -> throw AssemblyError("source kind should have been memory")
                    is TypecastExpression -> assignTypeCastedValue(assign.target, value.type, value.expression, assign)
                    is FunctionCall -> {
-                        if(value.target.targetSubroutine(program.namespace)?.isAsmSubroutine==true) {
-                            // handle asmsub functioncalls specifically, without shoving stuff on the estack
-                            val sub = value.target.targetSubroutine(program.namespace)!!
-                            val preserveStatusRegisterAfterCall = sub.asmReturnvaluesRegisters.any { it.statusflag != null }
-                            asmgen.translateFunctionCall(value, preserveStatusRegisterAfterCall)
-                            when((sub.asmReturnvaluesRegisters.single { it.registerOrPair!=null }).registerOrPair) {
-                                RegisterOrPair.A -> assignRegisterByte(assign.target, CpuRegister.A)
-                                RegisterOrPair.X -> assignRegisterByte(assign.target, CpuRegister.X)
-                                RegisterOrPair.Y -> assignRegisterByte(assign.target, CpuRegister.Y)
-                                RegisterOrPair.AX -> assignRegisterpairWord(assign.target, RegisterOrPair.AX)
-                                RegisterOrPair.AY -> assignRegisterpairWord(assign.target, RegisterOrPair.AY)
-                                RegisterOrPair.XY -> assignRegisterpairWord(assign.target, RegisterOrPair.XY)
-                                else -> throw AssemblyError("should be just one register byte result value")
+                        when (val sub = value.target.targetStatement(program.namespace)) {
+                            is Subroutine -> {
+                                val preserveStatusRegisterAfterCall = sub.asmReturnvaluesRegisters.any { it.statusflag != null }
+                                asmgen.translateFunctionCall(value, preserveStatusRegisterAfterCall)
+                                val returnValue = sub.returntypes.zip(sub.asmReturnvaluesRegisters).single { it.second.registerOrPair!=null }
+                                when (returnValue.first) {
+                                    DataType.STR -> {
+                                        when(assign.target.datatype) {
+                                            DataType.UWORD -> {
+                                                // assign the address of the string result value
+                                                assignRegisterpairWord(assign.target, RegisterOrPair.AY)
+                                            }
+                                            DataType.STR, DataType.ARRAY_UB, DataType.ARRAY_B -> {
+                                                // copy the actual string result into the target string variable
+                                                asmgen.out("""
+                                                                        pha
+                                                                        lda  #<${assign.target.asmVarname}
+                                                                        sta  P8ZP_SCRATCH_W1
+                                                                        lda  #>${assign.target.asmVarname}
+                                                                        sta  P8ZP_SCRATCH_W1+1
+                                                                        pla
+                                                                        jsr  prog8_lib.strcpy""")
+                                            }
+                                            else -> throw AssemblyError("weird target dt")
+                                        }
+                                    }
+                                    DataType.FLOAT -> {
+                                        // float result from function sits in FAC1
+                                        assignFAC1float(assign.target)
+                                    }
+                                    else -> {
+                                        when (returnValue.second.registerOrPair) {
+                                            RegisterOrPair.A -> assignRegisterByte(assign.target, CpuRegister.A)
+                                            RegisterOrPair.X -> assignRegisterByte(assign.target, CpuRegister.X)
+                                            RegisterOrPair.Y -> assignRegisterByte(assign.target, CpuRegister.Y)
+                                            RegisterOrPair.AX -> assignRegisterpairWord(assign.target, RegisterOrPair.AX)
+                                            RegisterOrPair.AY -> assignRegisterpairWord(assign.target, RegisterOrPair.AY)
+                                            RegisterOrPair.XY -> assignRegisterpairWord(assign.target, RegisterOrPair.XY)
+                                            else -> throw AssemblyError("should be just one register byte result value")
+                                        }
+                                    }
+                                }
+                                if (preserveStatusRegisterAfterCall)
+                                    asmgen.out("  plp\t; restore status flags from call")
+                            }
+                            is BuiltinFunctionStatementPlaceholder -> {
+                                val signature = BuiltinFunctions.getValue(sub.name)
+                                asmgen.translateBuiltinFunctionCallExpression(value, signature, false)
+                                val returntype = builtinFunctionReturnType(sub.name, value.args, program)
+                                if(!returntype.isKnown)
+                                    throw AssemblyError("unknown dt")
+                                when(returntype.typeOrElse(DataType.STRUCT)) {
+                                    in ByteDatatypes -> assignRegisterByte(assign.target, CpuRegister.A)            // function's byte result is in A
+                                    in WordDatatypes -> assignRegisterpairWord(assign.target, RegisterOrPair.AY)    // function's word result is in AY
+                                    DataType.STR -> throw AssemblyError("missing code for assign string from builtin func => copy string or assign string address")
+                                    DataType.FLOAT -> {
+                                        // float result from function sits in FAC1
+                                        assignFAC1float(assign.target)
+                                    }
+                                    else -> throw AssemblyError("weird result type")
+                                }
+                            }
+                            else -> {
+                                throw AssemblyError("weird func call")
                            }
-                            if(preserveStatusRegisterAfterCall)
-                                asmgen.out("  plp\t; restore status flags from call")
-                        } else {
-                            // regular subroutine, return values are (for now) always done via the stack...  TODO optimize this
-                            asmgen.translateExpression(value)
-                            if(assign.target.datatype in WordDatatypes && assign.source.datatype in ByteDatatypes)
-                                asmgen.signExtendStackLsb(assign.source.datatype)
-                            assignStackValue(assign.target)
                        }
                    }
                    else -> {
-                        // everything else just evaluate via the stack.
+                        // Everything else just evaluate via the stack.
+                        // (we can't use the assignment helper functions to do it via registers here,
+                        // because the code here is the implementation of exactly that...)
                        asmgen.translateExpression(value)
                        if(assign.target.datatype in WordDatatypes && assign.source.datatype in ByteDatatypes)
                            asmgen.signExtendStackLsb(assign.source.datatype)
@@ -167,7 +228,10 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
    }

    private fun assignTypeCastedValue(target: AsmAssignTarget, targetDt: DataType, value: Expression, origAssign: AsmAssignment) {
-        val valueDt = value.inferType(program).typeOrElse(DataType.STRUCT)
+        val valueIDt = value.inferType(program)
+        if(!valueIDt.isKnown)
+            throw AssemblyError("unknown dt")
+        val valueDt = valueIDt.typeOrElse(DataType.STRUCT)
        when(value) {
            is IdentifierReference -> {
                if(targetDt in WordDatatypes) {
@@ -199,6 +263,13 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
        }

        when(value) {
+            is IdentifierReference -> {
+                if(target.kind==TargetStorageKind.VARIABLE) {
+                    val sourceDt = value.inferType(program).typeOrElse(DataType.STRUCT)
+                    if (sourceDt != DataType.STRUCT)
+                        return assignTypeCastedIdentifier(target.asmVarname, targetDt, asmgen.asmVariableName(value), sourceDt, origAssign.source.expression!!)
+                }
+            }
            is PrefixExpression -> {}
            is BinaryExpression -> {}
            is ArrayIndexedExpression -> {}
@@ -208,7 +279,7 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
            else -> {
                // TODO optimize the others further?
                if(this.asmgen.options.slowCodegenWarnings)
-                    println("warning: slow stack evaluation used for typecast: into $targetDt at ${value.position}")
+                    println("warning: slow stack evaluation used for typecast: $value into $targetDt at ${value.position}")
            }
        }

@@ -217,6 +288,125 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
        assignStackValue(target)
    }

+    private fun assignTypeCastedIdentifier(targetAsmVarName: String, targetDt: DataType,
+                                           sourceAsmVarName: String, sourceDt: DataType,
+                                           origExpr: Expression) {
+        if(sourceDt == targetDt)
+            throw AssemblyError("typecast to identical value")
+
+        // also see: ExpressionAsmGen,   fun translateExpression(typecast: TypecastExpression)
+        when(sourceDt) {
+            DataType.UBYTE -> {
+                when(targetDt) {
+                    DataType.UBYTE, DataType.BYTE -> {
+                        asmgen.out("  lda  $sourceAsmVarName |  sta  $targetAsmVarName")
+                    }
+                    DataType.UWORD, DataType.WORD -> {
+                        if(CompilationTarget.instance.machine.cpu==CpuType.CPU65c02)
+                            asmgen.out("  lda  $sourceAsmVarName |  sta  $targetAsmVarName |  stz  $targetAsmVarName+1")
+                        else
+                            asmgen.out("  lda  $sourceAsmVarName |  sta  $targetAsmVarName |  lda  #0  |  sta  $targetAsmVarName+1")
+                    }
+                    DataType.FLOAT -> {
+                        asmgen.out("""
+                            lda  #<$targetAsmVarName
+                            ldy  #>$targetAsmVarName
+                            sta  P8ZP_SCRATCH_W2
+                            sty  P8ZP_SCRATCH_W2+1
+                            ldy  $sourceAsmVarName
+                            jsr  floats.cast_from_ub""")
+                    }
+                    else -> throw AssemblyError("weird type")
+                }
+            }
+            DataType.BYTE -> {
+                when(targetDt) {
+                    DataType.UBYTE, DataType.BYTE -> {
+                        asmgen.out("  lda  $sourceAsmVarName |  sta  $targetAsmVarName")
+                    }
+                    DataType.UWORD, DataType.WORD -> {
+                        asmgen.out("  lda  $sourceAsmVarName |  sta  $targetAsmVarName")
+                        asmgen.signExtendVariableLsb(targetAsmVarName, DataType.BYTE)
+                    }
+                    DataType.FLOAT -> {
+                        asmgen.out("""
+                            lda  #<$targetAsmVarName
+                            ldy  #>$targetAsmVarName
+                            sta  P8ZP_SCRATCH_W2
+                            sty  P8ZP_SCRATCH_W2+1
+                            lda  $sourceAsmVarName
+                            jsr  floats.cast_from_b""")
+                    }
+                    else -> throw AssemblyError("weird type")
+                }
+            }
+            DataType.UWORD -> {
+                when(targetDt) {
+                    DataType.BYTE, DataType.UBYTE -> {
+                        asmgen.out("  lda  $sourceAsmVarName |  sta  $targetAsmVarName")
+                    }
+                    DataType.WORD, DataType.UWORD -> {
+                        asmgen.out("  lda  $sourceAsmVarName |  sta  $targetAsmVarName |  lda  $sourceAsmVarName+1 |  sta  $targetAsmVarName+1")
+                    }
+                    DataType.FLOAT -> {
+                        asmgen.out("""
+                            lda  #<$targetAsmVarName
+                            ldy  #>$targetAsmVarName
+                            sta  P8ZP_SCRATCH_W2
+                            sty  P8ZP_SCRATCH_W2+1
+                            lda  $sourceAsmVarName
+                            ldy  $sourceAsmVarName+1
+                            jsr  floats.cast_from_uw""")
+                    }
+                    else -> throw AssemblyError("weird type")
+                }
+            }
+            DataType.WORD -> {
+                when(targetDt) {
+                    DataType.BYTE, DataType.UBYTE -> {
+                        asmgen.out("  lda  $sourceAsmVarName |  sta  $targetAsmVarName")
+                    }
+                    DataType.WORD, DataType.UWORD -> {
+                        asmgen.out("  lda  $sourceAsmVarName |  sta  $targetAsmVarName |  lda  $sourceAsmVarName+1 |  sta  $targetAsmVarName+1")
+                    }
+                    DataType.FLOAT -> {
+                        asmgen.out("""
+                            lda  #<$targetAsmVarName
+                            ldy  #>$targetAsmVarName
+                            sta  P8ZP_SCRATCH_W2
+                            sty  P8ZP_SCRATCH_W2+1
+                            lda  $sourceAsmVarName
+                            ldy  $sourceAsmVarName+1
+                            jsr  floats.cast_from_w""")
+                    }
+                    else -> throw AssemblyError("weird type")
+                }
+            }
+            DataType.FLOAT -> {
+                asmgen.out("""
+                            lda  #<$targetAsmVarName
+                            ldy  #>$targetAsmVarName
+                            sta  P8ZP_SCRATCH_W2
+                            sty  P8ZP_SCRATCH_W2+1
+                            lda  #<$sourceAsmVarName
+                            ldy  #>$sourceAsmVarName""")
+                when(targetDt) {
+                    DataType.UBYTE -> asmgen.out("  jsr  floats.cast_as_uw_into_ya |  sty  $targetAsmVarName")
+                    DataType.BYTE -> asmgen.out("  jsr  floats.cast_as_w_into_ay |  sta  $targetAsmVarName")
+                    DataType.UWORD -> asmgen.out("  jsr  floats.cast_as_uw_into_ya |  sty  $targetAsmVarName |  sta  $targetAsmVarName+1")
+                    DataType.WORD -> asmgen.out("  jsr  floats.cast_as_w_into_ay |  sta  $targetAsmVarName |  sty  $targetAsmVarName+1")
+                    else -> throw AssemblyError("weird type")
+                }
+            }
+            DataType.STR -> {
+                if (targetDt != DataType.UWORD && targetDt == DataType.STR)
+                    throw AssemblyError("cannot typecast a string into another incompatitble type")
+            }
+            else -> throw AssemblyError("weird type")
+        }
+    }
+
+
    private fun assignStackValue(target: AsmAssignTarget) {
        when(target.kind) {
            TargetStorageKind.VARIABLE -> {
@@ -243,9 +433,9 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                    DataType.STR -> {
                        asmgen.out("""
                            lda  #<${target.asmVarname}
+                            ldy  #>${target.asmVarname}
                            sta  P8ZP_SCRATCH_W1
-                            lda  #>${target.asmVarname}
-                            sta  P8ZP_SCRATCH_W1+1
+                            sty  P8ZP_SCRATCH_W1+1
                            inx
                            lda  P8ESTACK_HI,x
                            tay
@@ -386,17 +576,17 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                    DataType.UWORD -> {
                        asmgen.out("""
                            lda #<$sourceName
+                            ldy #>$sourceName
                            sta ${target.asmVarname}
-                            lda #>$sourceName
-                            sta ${target.asmVarname}+1
+                            sty ${target.asmVarname}+1
                        """)
                    }
                    DataType.STR, DataType.ARRAY_UB, DataType.ARRAY_B -> {
                        asmgen.out("""
                            lda  #<${target.asmVarname}
+                            ldy  #>${target.asmVarname}
                            sta  P8ZP_SCRATCH_W1
-                            lda  #>${target.asmVarname}
-                            sta  P8ZP_SCRATCH_W1+1
+                            sty  P8ZP_SCRATCH_W1+1
                            lda  #<$sourceName
                            ldy  #>$sourceName
                            jsr  prog8_lib.strcpy""")
@@ -407,9 +597,9 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
            TargetStorageKind.STACK -> {
                asmgen.out("""
                    lda  #<$sourceName
+                    ldy  #>$sourceName+1
                    sta  P8ESTACK_LO,x
-                    lda  #>$sourceName+1
-                    sta  P8ESTACK_HI,x
+                    sty  P8ESTACK_HI,x
                    dex""")
            }
            else -> throw AssemblyError("string-assign to weird target")
@@ -503,15 +693,80 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
            }
            TargetStorageKind.STACK -> {
                asmgen.out("""
-                    lda  #$sourceName
+                    lda  $sourceName
                    sta  P8ESTACK_LO,x
-                    lda  #$sourceName+1
+                    lda  $sourceName+1
                    sta  P8ESTACK_HI,x
                    dex""")
            }
        }
    }

+    private fun assignFAC1float(target: AsmAssignTarget) {
+        when(target.kind) {
+            TargetStorageKind.VARIABLE -> {
+                asmgen.out("""
+                    stx  P8ZP_SCRATCH_REG
+                    ldx  #<${target.asmVarname}
+                    ldy  #>${target.asmVarname}
+                    jsr  floats.MOVMF
+                    ldx  P8ZP_SCRATCH_REG
+                """)
+            }
+            TargetStorageKind.ARRAY -> {
+                asmgen.out("""
+                    lda  #<${target.asmVarname} 
+                    ldy  #>${target.asmVarname}
+                    sta  P8ZP_SCRATCH_W1
+                    sty  P8ZP_SCRATCH_W1+1""")
+                if(target.array!!.indexer.indexNum!=null) {
+                    val index = target.array.indexer.constIndex()!!
+                    asmgen.out(" lda  #$index")
+                } else {
+                    val asmvarname = asmgen.asmVariableName(target.array.indexer.indexVar!!)
+                    asmgen.out(" lda  $asmvarname")
+                }
+                asmgen.out("  jsr  floats.set_array_float_from_fac1")
+            }
+            TargetStorageKind.MEMORY -> throw AssemblyError("can't assign float to mem byte")
+            TargetStorageKind.REGISTER -> throw AssemblyError("can't assign float to register")
+            TargetStorageKind.STACK -> asmgen.out("  jsr  floats.push_fac1")
+        }
+    }
+
+    private fun assignFloatFromAY(target: AsmAssignTarget) {
+        when(target.kind) {
+            TargetStorageKind.VARIABLE -> {
+                asmgen.out("""
+                    sta  P8ZP_SCRATCH_W1                    
+                    sty  P8ZP_SCRATCH_W1+1
+                    lda  #<${target.asmVarname} 
+                    ldy  #>${target.asmVarname}
+                    jsr  floats.copy_float""")
+            }
+            TargetStorageKind.ARRAY -> {
+                asmgen.out("""
+                    sta  P8ZP_SCRATCH_W1
+                    sty  P8ZP_SCRATCH_W1+1
+                    lda  #<${target.asmVarname} 
+                    ldy  #>${target.asmVarname}
+                    sta  P8ZP_SCRATCH_W2
+                    sty  P8ZP_SCRATCH_W2+1""")
+                if(target.array!!.indexer.indexNum!=null) {
+                    val index = target.array.indexer.constIndex()!!
+                    asmgen.out(" lda  #$index")
+                } else {
+                    val asmvarname = asmgen.asmVariableName(target.array.indexer.indexVar!!)
+                    asmgen.out(" lda  $asmvarname")
+                }
+                asmgen.out(" jsr  floats.set_array_float")
+            }
+            TargetStorageKind.MEMORY -> throw AssemblyError("can't assign float to mem byte")
+            TargetStorageKind.REGISTER -> throw AssemblyError("can't assign float to register")
+            TargetStorageKind.STACK -> asmgen.out("  jsr  floats.push_float")
+        }
+    }
+
    private fun assignVariableFloat(target: AsmAssignTarget, sourceName: String) {
        when(target.kind) {
            TargetStorageKind.VARIABLE -> {
@@ -586,7 +841,7 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
            }
            TargetStorageKind.STACK -> {
                asmgen.out("""
-                    lda  #$sourceName
+                    lda  $sourceName
                    sta  P8ESTACK_LO,x
                    dex""")
            }
@@ -763,7 +1018,10 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
    }

    private fun assignRegisterpairWord(target: AsmAssignTarget, regs: RegisterOrPair) {
-        require(target.datatype in WordDatatypes)
+        require(target.datatype in NumericDatatypes)
+        if(target.datatype==DataType.FLOAT)
+            throw AssemblyError("float value should be from FAC1 not from registerpair memory pointer")
+
        when(target.kind) {
            TargetStorageKind.VARIABLE -> {
                when(regs) {
@@ -774,7 +1032,20 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                }
            }
            TargetStorageKind.ARRAY -> {
-                TODO("store register pair $regs into word-array ${target.array}")
+                // TODO can be a bit more optimized if the array indexer is a number
+                when(regs) {
+                    RegisterOrPair.AX -> asmgen.out("  pha |  txa |  pha")
+                    RegisterOrPair.AY -> asmgen.out("  pha |  tya |  pha")
+                    RegisterOrPair.XY -> asmgen.out("  txa |  pha |  tya |  pha")
+                    else -> throw AssemblyError("expected reg pair")
+                }
+                asmgen.loadScaledArrayIndexIntoRegister(target.array!!, DataType.UWORD, CpuRegister.Y, true)
+                asmgen.out("""
+                    pla
+                    sta  ${target.asmVarname},y
+                    dey
+                    pla
+                    sta  ${target.asmVarname},y""")
            }
            TargetStorageKind.REGISTER -> {
                when(regs) {
@@ -1138,17 +1409,8 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                asmgen.storeByteIntoPointer(addressExpr, ldaInstructionArg)
            }
            else -> {
-                asmgen.out(" lda  $ldaInstructionArg |  pha")
-                asmgen.translateExpression(addressExpr)
-                asmgen.out("""
-                    inx
-                    lda  P8ESTACK_LO,x
-                    sta  P8ZP_SCRATCH_W2
-                    lda  P8ESTACK_HI,x
-                    sta  P8ZP_SCRATCH_W2+1
-                    ldy  #0
-                    pla
-                    sta  (P8ZP_SCRATCH_W2),y""")
+                asmgen.assignExpressionToVariable(addressExpr, asmgen.asmVariableName("P8ZP_SCRATCH_W2"), DataType.UWORD, null)
+                asmgen.out("  ldy  #0 |  lda  $ldaInstructionArg |  sta  (P8ZP_SCRATCH_W2),y")
            }
        }
    }
@@ -1171,18 +1433,32 @@ internal class AssignmentAsmGen(private val program: Program, private val asmgen
                asmgen.storeByteIntoPointer(addressExpr, null)
            }
            else -> {
-                asmgen.saveRegister(register, false, memoryAddress.definingSubroutine())
-                asmgen.translateExpression(addressExpr)
+                asmgen.saveRegister(register, false, memoryAddress.definingSubroutine()!!)
+                asmgen.assignExpressionToVariable(addressExpr, asmgen.asmVariableName("P8ZP_SCRATCH_W2"), DataType.UWORD, null)
                asmgen.restoreRegister(CpuRegister.A, false)
-                asmgen.out("""
-                    inx
-                    ldy  P8ESTACK_LO,x
-                    sty  P8ZP_SCRATCH_W2
-                    ldy  P8ESTACK_HI,x
-                    sty  P8ZP_SCRATCH_W2+1
-                    ldy  #0
-                    sta  (P8ZP_SCRATCH_W2),y""")
+                asmgen.out("  ldy  #0 |  sta  (P8ZP_SCRATCH_W2),y")
            }
        }
    }
+
+    internal fun assignExpressionToRegister(expr: Expression, register: RegisterOrPair) {
+        val src = AsmAssignSource.fromAstSource(expr, program, asmgen)
+        val tgt = AsmAssignTarget.fromRegisters(register, null, program, asmgen)
+        val assign = AsmAssignment(src, tgt, false, expr.position)
+        translateNormalAssignment(assign)
+    }
+
+    internal fun assignExpressionToVariable(expr: Expression, asmVarName: String, dt: DataType, scope: Subroutine?) {
+        val src = AsmAssignSource.fromAstSource(expr, program, asmgen)
+        val tgt = AsmAssignTarget(TargetStorageKind.VARIABLE, program, asmgen, dt, scope, variableAsmName = asmVarName)
+        val assign = AsmAssignment(src, tgt, false, expr.position)
+        translateNormalAssignment(assign)
+    }
+
+    internal fun assignVariableToRegister(asmVarName: String, register: RegisterOrPair) {
+        val tgt = AsmAssignTarget.fromRegisters(register, null, program, asmgen)
+        val src = AsmAssignSource(SourceStorageKind.VARIABLE, program, asmgen, tgt.datatype, variableAsmName = asmVarName)
+        val assign = AsmAssignment(src, tgt, false, Position.DUMMY)
+        translateNormalAssignment(assign)
+    }
 }
@@ -22,7 +22,10 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
        when (val value = assign.source.expression!!) {
            is PrefixExpression -> {
                // A = -A , A = +A, A = ~A, A = not A
-                val type = value.inferType(program).typeOrElse(DataType.STRUCT)
+                val itype = value.inferType(program)
+                if(!itype.isKnown)
+                    throw AssemblyError("unknown dt")
+                val type = itype.typeOrElse(DataType.STRUCT)
                when (value.operator) {
                    "+" -> {}
                    "-" -> inplaceNegate(assign.target, type)
@@ -136,13 +139,13 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
                    }
                    DataType.FLOAT -> {
                        when {
-                            valueLv != null -> inplaceModification_float_litval_to_variable(target.asmVarname, operator, valueLv.toDouble(), target.scope)
-                            ident != null -> inplaceModification_float_variable_to_variable(target.asmVarname, operator, ident, target.scope)
+                            valueLv != null -> inplaceModification_float_litval_to_variable(target.asmVarname, operator, valueLv.toDouble(), target.scope!!)
+                            ident != null -> inplaceModification_float_variable_to_variable(target.asmVarname, operator, ident, target.scope!!)
                            value is TypecastExpression -> {
                                if (tryRemoveRedundantCast(value, target, operator)) return
-                                inplaceModification_float_value_to_variable(target.asmVarname, operator, value, target.scope)
+                                inplaceModification_float_value_to_variable(target.asmVarname, operator, value, target.scope!!)
                            }
-                            else -> inplaceModification_float_value_to_variable(target.asmVarname, operator, value, target.scope)
+                            else -> inplaceModification_float_value_to_variable(target.asmVarname, operator, value, target.scope!!)
                        }
                    }
                    else -> throw AssemblyError("weird type to do in-place modification on ${target.datatype}")
@@ -168,20 +171,20 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
                    is IdentifierReference -> {
                        val pointer = memory.addressExpression as IdentifierReference
                        when {
-                            valueLv != null -> inplaceModification_byte_litval_to_memory(pointer, operator, valueLv.toInt())
-                            ident != null -> inplaceModification_byte_variable_to_memory(pointer, operator, ident)
+                            valueLv != null -> inplaceModification_byte_litval_to_pointer(pointer, operator, valueLv.toInt())
+                            ident != null -> inplaceModification_byte_variable_to_pointer(pointer, operator, ident)
                            // TODO more specialized code for types such as memory read etc.
                            value is TypecastExpression -> {
                                if (tryRemoveRedundantCast(value, target, operator)) return
-                                inplaceModification_byte_value_to_memory(pointer, operator, value)
+                                inplaceModification_byte_value_to_pointer(pointer, operator, value)
                            }
-                            else -> inplaceModification_byte_value_to_memory(pointer, operator, value)
+                            else -> inplaceModification_byte_value_to_pointer(pointer, operator, value)
                        }
                    }
                    else -> {
                        if(asmgen.options.slowCodegenWarnings)
                            println("warning: slow stack evaluation used (1): ${memory.addressExpression::class.simpleName} at ${memory.addressExpression.position}") // TODO optimize...
-                        asmgen.translateExpression(memory.addressExpression)
+                        asmgen.translateExpression(memory.addressExpression)  // TODO directly into P8ZP_SCRATCH_W2?
                        asmgen.out("  jsr  prog8_lib.read_byte_from_address_on_stack |  sta  P8ZP_SCRATCH_B1")
                        val zp = CompilationTarget.instance.machine.zeropage
                        when {
@@ -210,9 +213,13 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,

    private fun tryRemoveRedundantCast(value: TypecastExpression, target: AsmAssignTarget, operator: String): Boolean {
        if (target.datatype == value.type) {
-            val childDt = value.expression.inferType(program).typeOrElse(DataType.STRUCT)
-            if (value.type.equalsSize(childDt) || value.type.largerThan(childDt)) {
+            val childIDt = value.expression.inferType(program)
+            if(!childIDt.isKnown)
+                throw AssemblyError("unknown dt")
+            val childDt = childIDt.typeOrElse(DataType.STRUCT)
+            if (value.type!=DataType.FLOAT && (value.type.equalsSize(childDt) || value.type.largerThan(childDt))) {
                // this typecast is redundant here; the rest of the code knows how to deal with the uncasted value.
+                // (works for integer types, not for float.)
                inplaceModification(target, operator, value.expression)
                return true
            }
@@ -220,15 +227,15 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
        return false
    }

-    private fun inplaceModification_byte_value_to_memory(pointervar: IdentifierReference, operator: String, value: Expression) {
+    private fun inplaceModification_byte_value_to_pointer(pointervar: IdentifierReference, operator: String, value: Expression) {
        if(asmgen.options.slowCodegenWarnings)
            println("warning: slow stack evaluation used (3):  @(${pointervar.nameInSource.last()}) $operator= ${value::class.simpleName} at ${value.position}") // TODO
        asmgen.translateExpression(value)
        val (ptrOnZp, sourceName) = asmgen.loadByteFromPointerIntoA(pointervar)
        when (operator) {
            // note: ** (power) operator requires floats.
-            "+" -> asmgen.out(" clc |  adc  P8ESTACK_LO+1,x")
-            "-" -> asmgen.out(" sec |  sbc  P8ESTACK_LO+1,x")
+            "+" -> asmgen.out("  clc |  adc  P8ESTACK_LO+1,x")
+            "-" -> asmgen.out("  sec |  sbc  P8ESTACK_LO+1,x")
            "*" -> asmgen.out("  pha |  lda  P8ESTACK_LO+1,x |  tay |  pla |  jsr  math.multiply_bytes |  ldy  #0")
            "/" -> asmgen.out("  pha |  lda  P8ESTACK_LO+1,x |  tay |  pla |  jsr  math.divmod_ub_asm |  tya |  ldy  #0")
            "%" -> asmgen.out("  pha |  lda  P8ESTACK_LO+1,x |  tay |  pla |  jsr  math.divmod_ub_asm |  ldy  #0")
@@ -272,14 +279,14 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
        asmgen.out(" inx")
    }

-    private fun inplaceModification_byte_variable_to_memory(pointervar: IdentifierReference, operator: String, value: IdentifierReference) {
+    private fun inplaceModification_byte_variable_to_pointer(pointervar: IdentifierReference, operator: String, value: IdentifierReference) {
        val otherName = asmgen.asmVariableName(value)
        val (ptrOnZp, sourceName) = asmgen.loadByteFromPointerIntoA(pointervar)

        when (operator) {
            // note: ** (power) operator requires floats.
-            "+" -> asmgen.out(" clc |  adc  $otherName")
-            "-" -> asmgen.out(" sec |  sbc  $otherName")
+            "+" -> asmgen.out("  clc |  adc  $otherName")
+            "-" -> asmgen.out("  sec |  sbc  $otherName")
            "*" -> asmgen.out("  ldy  $otherName |  jsr  math.multiply_bytes |  ldy  #0")
            "/" -> asmgen.out("  ldy  $otherName |  jsr  math.divmod_ub_asm |  tya |  ldy  #0")
            "%" -> asmgen.out("  ldy  $otherName |  jsr  math.divmod_ub_asm |  ldy  #0")
@@ -312,7 +319,7 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
            asmgen.out("  sta  (P8ZP_SCRATCH_W1),y")
    }

-    private fun inplaceModification_byte_litval_to_memory(pointervar: IdentifierReference, operator: String, value: Int) {
+    private fun inplaceModification_byte_litval_to_pointer(pointervar: IdentifierReference, operator: String, value: Int) {
        when (operator) {
            // note: ** (power) operator requires floats.
            "+" -> {
@@ -413,30 +420,36 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
    private fun inplaceModification_byte_value_to_variable(name: String, dt: DataType, operator: String, value: Expression) {
        // this should be the last resort for code generation for this,
        // because the value is evaluated onto the eval stack (=slow).
-        if(asmgen.options.slowCodegenWarnings)
-            println("warning: slow stack evaluation used (5):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
-        asmgen.translateExpression(value)
        when (operator) {
            // note: ** (power) operator requires floats.
-            "+" -> asmgen.out(" lda  $name |  clc |  adc  P8ESTACK_LO+1,x |  sta  $name")
-            "-" -> asmgen.out(" lda  $name |  sec |  sbc  P8ESTACK_LO+1,x |  sta  $name")
-            "*" -> asmgen.out("  lda  P8ESTACK_LO+1,x |  ldy  $name |  jsr  math.multiply_bytes |  sta  $name")
+            "+" -> {
+                asmgen.assignExpressionToRegister(value, RegisterOrPair.A)
+                asmgen.out("  clc |  adc  $name |  sta  $name")
+            }
+            "-" -> {
+                asmgen.assignExpressionToVariable(value, "P8ZP_SCRATCH_B1", dt, null)
+                asmgen.out(" lda  $name |  sec |  sbc  P8ZP_SCRATCH_B1 |  sta  $name")
+            }
+            "*" -> {
+                asmgen.assignExpressionToRegister(value, RegisterOrPair.A)
+                asmgen.out("  ldy  $name |  jsr  math.multiply_bytes |  sta  $name")
+            }
            "/" -> {
+                asmgen.assignExpressionToRegister(value, RegisterOrPair.Y)
                if(dt==DataType.UBYTE)
-                    asmgen.out("  lda  P8ESTACK_LO+1,x |  tay |  lda  $name |  jsr  math.divmod_ub_asm |  sty  $name")
+                    asmgen.out("  lda  $name |  jsr  math.divmod_ub_asm |  sty  $name")
                else
-                    asmgen.out("  lda  P8ESTACK_LO+1,x |  tay |  lda  $name |  jsr  math.divmod_b_asm |  sty  $name")
+                    asmgen.out("  lda  $name |  jsr  math.divmod_b_asm |  sty  $name")
            }
            "%" -> {
                if(dt==DataType.BYTE)
                    throw AssemblyError("remainder of signed integers is not properly defined/implemented, use unsigned instead")
-                asmgen.out("  lda  P8ESTACK_LO+1,x |  tay |  lda  $name |  jsr  math.divmod_ub_asm |  sta  $name")
+                asmgen.assignExpressionToRegister(value, RegisterOrPair.Y)
+                asmgen.out("  lda  $name |  jsr  math.divmod_ub_asm |  sta  $name")
            }
            "<<" -> {
-                asmgen.translateExpression(value)
+                asmgen.assignExpressionToRegister(value, RegisterOrPair.Y)
                asmgen.out("""
-                    inx
-                    ldy  P8ESTACK_LO,x
                    beq  +
 -                   asl  $name
                    dey
@@ -444,11 +457,9 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
 +""")
            }
            ">>" -> {
-                asmgen.translateExpression(value)
+                asmgen.assignExpressionToRegister(value, RegisterOrPair.Y)
                if(dt==DataType.UBYTE) {
                    asmgen.out("""
-                        inx
-                        ldy  P8ESTACK_LO,x
                        beq  +
 -                       lsr  $name
                        dey
@@ -456,8 +467,6 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
 +""")
                } else {
                    asmgen.out("""
-                        inx
-                        ldy  P8ESTACK_LO,x
                        beq  +
 -                       lda  $name
                        asl  a
@@ -467,12 +476,20 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
 +""")
                }
            }
-            "&" -> asmgen.out(" lda  $name |  and  P8ESTACK_LO+1,x |  sta  $name")
-            "^" -> asmgen.out(" lda  $name |  eor  P8ESTACK_LO+1,x |  sta  $name")
-            "|" -> asmgen.out(" lda  $name |  ora  P8ESTACK_LO+1,x |  sta  $name")
+            "&" -> {
+                asmgen.assignExpressionToRegister(value, RegisterOrPair.A)
+                asmgen.out("  and  $name |  sta  $name")
+            }
+            "^" -> {
+                asmgen.assignExpressionToRegister(value, RegisterOrPair.A)
+                asmgen.out("  eor  $name |  sta  $name")
+            }
+            "|" -> {
+                asmgen.assignExpressionToRegister(value, RegisterOrPair.A)
+                asmgen.out("  ora  $name |  sta  $name")
+            }
            else -> throw AssemblyError("invalid operator for in-place modification $operator")
        }
-        asmgen.out(" inx")
    }

    private fun inplaceModification_byte_variable_to_variable(name: String, dt: DataType, operator: String, ident: IdentifierReference) {
@@ -1093,11 +1110,13 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
    private fun inplaceModification_word_value_to_variable(name: String, dt: DataType, operator: String, value: Expression) {
        // this should be the last resort for code generation for this,
        // because the value is evaluated onto the eval stack (=slow).
-        if(asmgen.options.slowCodegenWarnings)
-            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
-        asmgen.translateExpression(value)
-        val valueDt = value.inferType(program).typeOrElse(DataType.STRUCT)

+        val valueiDt = value.inferType(program)
+        if(!valueiDt.isKnown)
+            throw AssemblyError("unknown dt")
+        val valueDt = valueiDt.typeOrElse(DataType.STRUCT)
+
+        // TODO can use registers instead of stack value?
        fun multiplyWord() {
            asmgen.out("""
                lda  P8ESTACK_LO+1,x
@@ -1114,6 +1133,7 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
            """)
        }

+        // TODO can use registers instead of stack value?
        fun divideWord() {
            if (dt == DataType.WORD) {
                asmgen.out("""
@@ -1142,6 +1162,7 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
            }
        }

+        // TODO can use registers instead of stack value?
        fun remainderWord() {
            if(dt==DataType.WORD)
                throw AssemblyError("remainder of signed integers is not properly defined/implemented, use unsigned instead")
@@ -1166,6 +1187,9 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
                when (operator) {
                    // note: ** (power) operator requires floats.
                    "+" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
                        if(valueDt==DataType.UBYTE)
                            asmgen.out("""
                                lda  $name
@@ -1187,8 +1211,12 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
                                tya
                                adc  $name+1
                                sta  $name+1""")
+                        asmgen.out(" inx")
                    }
                    "-" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
                        if(valueDt==DataType.UBYTE)
                            asmgen.out("""
                                lda  $name
@@ -1212,25 +1240,38 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
                                lda  $name+1
                                sbc  P8ZP_SCRATCH_B1
                                sta  $name+1""")
+                        asmgen.out(" inx")
                    }
                    "*" -> {
                        // stack contains (u) byte value, sign extend that and proceed with regular 16 bit operation
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
                        asmgen.signExtendStackLsb(valueDt)
                        multiplyWord()
+                        asmgen.out(" inx")
                    }
                    "/" -> {
                        // stack contains (u) byte value, sign extend that and proceed with regular 16 bit operation
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
                        asmgen.signExtendStackLsb(valueDt)
                        divideWord()
+                        asmgen.out(" inx")
                    }
                    "%" -> {
                        // stack contains (u) byte value, sign extend that and proceed with regular 16 bit operation
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
                        asmgen.signExtendStackLsb(valueDt)
                        remainderWord()
+                        asmgen.out(" inx")
                    }
                    "<<" -> {
+                        asmgen.assignExpressionToRegister(value, RegisterOrPair.Y)
                        asmgen.out("""
-                            ldy  P8ESTACK_LO+1,x
                            beq  +
 -                   	    asl  $name
                            rol  $name+1
@@ -1239,18 +1280,17 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
 +""")
                    }
                    ">>" -> {
-                        if(dt==DataType.UWORD) {
+                        asmgen.assignExpressionToRegister(value, RegisterOrPair.Y)
+                        if(dt==DataType.UWORD)
                            asmgen.out("""
-                            ldy  P8ESTACK_LO+1,x
                            beq  +
 -                           lsr  $name+1
                            ror  $name
                            dey
                            bne  -
-+""") }
-                        else {
+""")
+                        else
                            asmgen.out("""
-                            ldy  P8ESTACK_LO+1,x
                            beq  +
 -                           lda  $name+1
                            asl  a
@@ -1259,15 +1299,21 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
                            dey
                            bne  -
 +""")
-                        }
                    }
                    "&" -> {
-                        asmgen.out("  lda  P8ESTACK_LO+1,x |  and  $name |  sta  $name")
+                        asmgen.assignExpressionToRegister(value, RegisterOrPair.A)
+                        asmgen.out("  and  $name |  sta  $name")
                        if(dt in WordDatatypes)
                            asmgen.out(" lda  #0 |  sta  $name+1")
                    }
-                    "^" -> asmgen.out("  lda  P8ESTACK_LO+1,x |  eor  $name |  sta  $name")
-                    "|" -> asmgen.out("  lda  P8ESTACK_LO+1,x |  ora  $name |  sta  $name")
+                    "^" -> {
+                        asmgen.assignExpressionToRegister(value, RegisterOrPair.A)
+                        asmgen.out("  eor  $name |  sta  $name")
+                    }
+                    "|" -> {
+                        asmgen.assignExpressionToRegister(value, RegisterOrPair.A)
+                        asmgen.out("  ora  $name |  sta  $name")
+                    }
                    else -> throw AssemblyError("invalid operator for in-place modification $operator")
                }
            }
@@ -1275,27 +1321,66 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
                // the value is a proper 16-bit word, so use both bytes of it.
                when (operator) {
                    // note: ** (power) operator requires floats.
-                    "+" -> asmgen.out(" lda  $name |  clc |  adc  P8ESTACK_LO+1,x |  sta  $name |  lda  $name+1 |  adc  P8ESTACK_HI+1,x |  sta  $name+1")
-                    "-" -> asmgen.out(" lda  $name |  sec |  sbc  P8ESTACK_LO+1,x |  sta  $name |  lda  $name+1 |  sbc  P8ESTACK_HI+1,x |  sta  $name+1")
-                    "*" -> multiplyWord()
-                    "/" -> divideWord()
-                    "%" -> remainderWord()
+                    "+" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
+                        asmgen.out(" lda  $name |  clc |  adc  P8ESTACK_LO+1,x |  sta  $name |  lda  $name+1 |  adc  P8ESTACK_HI+1,x |  sta  $name+1 |  inx")
+                    }
+                    "-" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
+                        asmgen.out(" lda  $name |  sec |  sbc  P8ESTACK_LO+1,x |  sta  $name |  lda  $name+1 |  sbc  P8ESTACK_HI+1,x |  sta  $name+1 |  inx")
+                    }
+                    "*" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
+                        multiplyWord()
+                        asmgen.out(" inx")
+                    }
+                    "/" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
+                        divideWord()
+                        asmgen.out(" inx")
+                    }
+                    "%" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
+                        remainderWord()
+                        asmgen.out(" inx")
+                    }
                    "<<", ">>" -> throw AssemblyError("shift by a word value not supported, max is a byte")
-                    "&" -> asmgen.out(" lda  $name |  and  P8ESTACK_LO+1,x |  sta  $name | lda  $name+1 |  and  P8ESTACK_HI+1,x  |  sta  $name+1")
-                    "^" -> asmgen.out(" lda  $name |  eor  P8ESTACK_LO+1,x |  sta  $name | lda  $name+1 |  eor  P8ESTACK_HI+1,x  |  sta  $name+1")
-                    "|" -> asmgen.out(" lda  $name |  ora  P8ESTACK_LO+1,x |  sta  $name | lda  $name+1 |  ora  P8ESTACK_HI+1,x  |  sta  $name+1")
+                    "&" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
+                        asmgen.out(" lda  $name |  and  P8ESTACK_LO+1,x |  sta  $name | lda  $name+1 |  and  P8ESTACK_HI+1,x  |  sta  $name+1 |  inx")
+                    }
+                    "^" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
+                        asmgen.out(" lda  $name |  eor  P8ESTACK_LO+1,x |  sta  $name | lda  $name+1 |  eor  P8ESTACK_HI+1,x  |  sta  $name+1 |  inx")
+                    }
+                    "|" -> {
+                        if(asmgen.options.slowCodegenWarnings)
+                            println("warning: slow stack evaluation used (4):  $name $operator= ${value::class.simpleName} at ${value.position}") // TODO
+                        asmgen.translateExpression(value)
+                        asmgen.out(" lda  $name |  ora  P8ESTACK_LO+1,x |  sta  $name | lda  $name+1 |  ora  P8ESTACK_HI+1,x  |  sta  $name+1 |  inx")
+                    }
                    else -> throw AssemblyError("invalid operator for in-place modification $operator")
                }
            }
-            else -> {
-                throw AssemblyError("can only use integer datatypes here")
-            }
+            else -> throw AssemblyError("can only use integer datatypes here")
        }
-
-        asmgen.out(" inx")
    }

-    private fun inplaceModification_float_value_to_variable(name: String, operator: String, value: Expression, scope: Subroutine?) {
+    private fun inplaceModification_float_value_to_variable(name: String, operator: String, value: Expression, scope: Subroutine) {
        // this should be the last resort for code generation for this,
        // because the value is evaluated onto the eval stack (=slow).
        if(asmgen.options.slowCodegenWarnings)
@@ -1350,7 +1435,7 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
        asmgen.restoreRegister(CpuRegister.X, false)
    }

-    private fun inplaceModification_float_variable_to_variable(name: String, operator: String, ident: IdentifierReference, scope: Subroutine?) {
+    private fun inplaceModification_float_variable_to_variable(name: String, operator: String, ident: IdentifierReference, scope: Subroutine) {
        val valueDt = ident.targetVarDecl(program.namespace)!!.datatype
        if(valueDt != DataType.FLOAT)
            throw AssemblyError("float variable expected")
@@ -1419,7 +1504,7 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
        asmgen.restoreRegister(CpuRegister.X, false)
    }

-    private fun inplaceModification_float_litval_to_variable(name: String, operator: String, value: Double, scope: Subroutine?) {
+    private fun inplaceModification_float_litval_to_variable(name: String, operator: String, value: Double, scope: Subroutine) {
        val constValueName = asmgen.getFloatAsmConst(value)
        asmgen.saveRegister(CpuRegister.X, false, scope)
        when (operator) {
@@ -1581,16 +1666,14 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
                                    asmgen.out("  sta  (P8ZP_SCRATCH_W1),y")
                            }
                            else -> {
-                                if(asmgen.options.slowCodegenWarnings)
-                                    println("warning: slow stack evaluation used (6): ${mem.addressExpression::class.simpleName} at ${mem.addressExpression.position}") // TODO
-                                asmgen.translateExpression(mem.addressExpression)
+                                asmgen.assignExpressionToVariable(mem.addressExpression, asmgen.asmVariableName("P8ZP_SCRATCH_W2"), DataType.UWORD, target.scope)
                                asmgen.out("""
-                                    jsr  prog8_lib.read_byte_from_address_on_stack
+                                    ldy  #0
+                                    lda  (P8ZP_SCRATCH_W2),y
                                    beq  +
                                    lda  #1
 +                                   eor  #1                                    
-                                    jsr  prog8_lib.write_byte_to_address_on_stack
-                                    inx""")
+                                    sta  (P8ZP_SCRATCH_W2),y""")
                            }
                        }
                    }
@@ -1651,14 +1734,12 @@ internal class AugmentableAssignmentAsmGen(private val program: Program,
                                    asmgen.out("  sta  (P8ZP_SCRATCH_W1),y")
                            }
                            else -> {
-                                if(asmgen.options.slowCodegenWarnings)
-                                    println("warning: slow stack evaluation used (7): ${memory.addressExpression::class.simpleName} at ${memory.addressExpression.position}") // TODO
-                                asmgen.translateExpression(memory.addressExpression)
+                                asmgen.assignExpressionToVariable(memory.addressExpression, asmgen.asmVariableName("P8ZP_SCRATCH_W2"), DataType.UWORD, target.scope)
                                asmgen.out("""
-                                    jsr  prog8_lib.read_byte_from_address_on_stack
+                                    ldy  #0
+                                    lda  (P8ZP_SCRATCH_W2),y
                                    eor  #255
-                                    jsr  prog8_lib.write_byte_to_address_on_stack
-                                    inx""")
+                                    sta  (P8ZP_SCRATCH_W2),y""")
                            }
                        }
                    }
@@ -75,7 +75,7 @@ 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

@@ -15,96 +15,165 @@ 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("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("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 +213,17 @@ 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())) {
@@ -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
@@ -477,7 +477,10 @@ 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)
@@ -606,8 +609,10 @@ 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 +622,7 @@ 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 +644,10 @@ 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)
@@ -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")

@@ -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) }

@@ -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")
+    }
 }


@@ -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
 ^^^^^^^^^^^^^
@@ -143,14 +143,17 @@ 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/>`_ .
+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'*)
@@ -765,18 +765,18 @@ 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 +785,13 @@ 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)
@@ -806,6 +808,7 @@ 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).

 Miscellaneous
@@ -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
@@ -2,14 +2,14 @@
 TODO
 ====

- get rid of all other TODO's in the code ;-)
- implement @stack for asmsub parameters
 - 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 +18,7 @@ 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.
@@ -0,0 +1,836 @@
+%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')
+
+        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()
+    }
+}
@@ -1,6 +1,7 @@
 %target c64
 %import syslib
 %import textio
+%import test_stack
 %zeropage basicsafe

 main {
@@ -23,6 +24,9 @@ main {
        ubyte upwards = true

        repeat {
+            ;txt.plot(0,0)
+            ;test_stack.test()
+
            ubyte mountain = 223        ; slope upwards
            if active_height < target_height {
                active_height++
@@ -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()
+    }
+}
@@ -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))
+
+    }
+}
@@ -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()
        }
    }

@@ -1,5 +1,6 @@
 %import syslib
 %import graphics
+%import test_stack

 ; Note: this program is compatible with C64 and CX16.

@@ -37,6 +38,8 @@ main {
            anglez+=452

            wait_a_little_bit()
+
+            ; test_stack.test()
        }
    }

@@ -1,6 +1,7 @@
 %target c64
 %import syslib
 %import textio
+%import test_stack

 spritedata $2000 {
    ; this memory block contains the sprite data
@@ -95,6 +96,8 @@ main {
            txt.print(" jiffies/fr = ")
            txt.print_ub(60/c64.TIME_LO)
            txt.print(" fps")
+
+            ; test_stack.test()
        }
    }

@@ -1,5 +1,6 @@
 %target c64
 %import syslib
+%import test_stack
 %import textio

 main {
@@ -33,6 +34,8 @@ main {
            txt.print_ub(60/c64.TIME_LO)
            txt.print(" fps")
            c64.TIME_LO=0
+
+            ; test_stack.test()
        }
    }

@@ -1,5 +1,6 @@
 %target cx16
 %import syslib
+%import test_stack
 %import conv

 ; TODO add all other Elite's ships, show their name, advance to next ship on keypress
@@ -36,11 +37,13 @@ main {

            cx16.GRAPH_set_colors(1, 0, 0)
            draw_lines_hiddenremoval()
-            ;draw_lines()
+            ; draw_lines()

            anglex += 217
            angley -= 505
            anglez += 452
+
+            ; test_stack.test()
        }
    }

@@ -191,6 +194,7 @@ main {

    sub facing_away(ubyte edgePointsIdx) -> ubyte {
        ; simplistic visibility determination by checking the Z component of the surface normal
+        ; TODO: actually take the line of sight vector into account
        ubyte p1 = shipdata.facesPoints[edgePointsIdx]
        edgePointsIdx++
        ubyte p2 = shipdata.facesPoints[edgePointsIdx]
@@ -223,7 +227,7 @@ main {

 shipdata {

-    ; Ship model data converted from BBC Elite's Cobra MK 3
+        ; Ship model data converted from BBC Elite's Cobra MK 3
    ; downloaded from http://www.elitehomepage.org/archive/index.htm

 const ubyte totalNumberOfEdges = 51
@@ -1,4 +1,5 @@
 %import graphics
+%import test_stack

 ; Note: this program is compatible with C64 and CX16.

@@ -6,8 +7,13 @@ main {

    sub start() {
        graphics.enable_bitmap_mode()
+
        draw_lines()
        draw_circles()
+
+        ; graphics.disable_bitmap_mode()
+        ; test_stack.test()
+
        repeat {
        }
    }
@@ -1,5 +1,6 @@
 %import textio
 %import syslib
+%import test_stack
 %zeropage basicsafe

 ; Note: this program is compatible with C64 and CX16.
@@ -40,6 +41,8 @@ main {
        rect(10, 10, 10, 10, false)
        rect(6, 0, 16, 20, true)

+        ; test_stack.test()
+

        sub rect(ubyte x1, ubyte y1, ubyte x2, ubyte y2, ubyte fill) {
            ubyte x
@@ -1,5 +1,6 @@
 %import graphics
 %import floats
+%import test_stack
 %zeropage floatsafe

 ; Draw a mandelbrot in graphics mode (the image will be 256 x 200 pixels).
@@ -44,6 +45,8 @@ main {
            }
        }

+        ; test_stack.test()
+
        repeat {
        }
    }
@@ -1,5 +1,6 @@
 %import textio
 %import floats
+%import test_stack
 %zeropage basicsafe

 ; Note: this program is compatible with C64 and CX16.
@@ -57,5 +58,7 @@ main {
        txt.print("finished in ")
        floats.print_f(duration)
        txt.print(" seconds!\n")
+
+        ; test_stack.test()
    }
 }
@@ -1,5 +1,6 @@
 %import textio
 %import conv
+%import test_stack
 %zeropage basicsafe

 ; The classic number guessing game.
@@ -59,6 +60,8 @@ main {
            txt.print("Thanks for playing, ")
            txt.print(name)
            txt.print(".\n")
+
+            ; test_stack.test()
        }
    }
 }
@@ -1,5 +1,6 @@
 %target c64
 %import syslib
+%import test_stack
 %import textio

 ;  converted from plasma test program for cc65.
@@ -26,7 +27,7 @@ main {
        ; ubyte v = c64.VMCSB
        c64.CIA2PRA = (block & $FC) | (lsb(SCREEN1 >> 14) ^ $03)

-        repeat {
+        repeat 100 {
            doplasma(SCREEN1)
            c64.VMCSB = PAGE1
            doplasma(SCREEN2)
@@ -37,6 +38,9 @@ main {
        ;c64.VMCSB = v
        ;c64.CIA2PRA = block
        ;txt.print("done!\n")
+        ;test_stack.test()
+        ;repeat {
+        ;}
    }

    ; several variables outside of doplasma to make them retain their value
@@ -1,4 +1,5 @@
 %import textio
+%import test_stack
 %zeropage basicsafe

 ; Note: this program is compatible with C64 and CX16.
@@ -26,6 +27,8 @@ main {
        txt.print("number of primes (expected 54): ")
        txt.print_ub(amount)
        txt.chrout('\n')
+
+        ; test_stack.test()
    }


@@ -1,4 +1,5 @@
 %import textio
+%import test_stack
 %zeropage basicsafe

 ; Note: this program is compatible with C64 and CX16.
@@ -31,6 +32,7 @@ main {
        txt.print("reversed\n")
        print_arrays()

+        ;test_stack.test()
        return


@@ -10,6 +10,7 @@
 %target c64
 %import syslib
 %import textio
+%import test_stack


 main {
@@ -67,6 +68,9 @@ waitkey:
            }

            drawScore()
+
+            ; txt.plot(0,0)
+            ; test_stack.test()
        }

        ubyte key=c64.GETIN()
@@ -1,38 +1,48 @@
 %import textio
-%import conv
 %import floats
 %zeropage basicsafe
+%import test_stack

 main {

    sub start() {

-        uword[]  array = [1, 2, 3]
-        ubyte ii = 0
-        ubyte ii2 = ii+2
-        array[ii+1] = array[ii2]           ; TODO fix overwriting the single array index autovar
+        ubyte[]  ubarray = [1,2,3,4]
+        byte[]  barray = [1,2,3,4]
+        uword[]  uwarray = [1,2,3,4]
+        word[]  warray = [1,2,3,4]
+        float[]  farray = [1.1,2.2,3.3,4.4]

-        uword xx
-        for xx in array {
-            txt.print_uw(xx)
-            txt.chrout('\n')
-        }
+        ubyte ub
+        byte bb
+        uword uw
+        word ww
+        float fl

-        testX()
-    }
+        const ubyte i = 2
+        const ubyte j = 3

-    asmsub testX() {
-        %asm {{
-            stx  _saveX
-            lda  #13
-            jsr  txt.chrout
-            lda  _saveX
-            jsr  txt.print_ub
-            lda  #13
-            jsr  txt.chrout
-            ldx  _saveX
-            rts
-_saveX   .byte 0
-        }}
+        ub = ubarray[i]
+        txt.print_ub(ub)
+        txt.chrout('\n')
+
+        bb = barray[i]
+        txt.print_b(bb)
+        txt.chrout('\n')
+
+        uw = uwarray[i]
+        txt.print_uw(uw)
+        txt.chrout('\n')
+
+        ww = warray[i]
+        txt.print_w(ww)
+        txt.chrout('\n')
+
+        fl = farray[i]
+        floats.print_f(fl)
+        txt.chrout('\n')
+
+        test_stack.test()
+        txt.chrout('\n')
    }
 }
Author	SHA1	Message	Date
Irmen de Jong	d9a8cfed8c	updated the compiled examples disk	2020-11-22 18:45:40 +01:00
Irmen de Jong	122796fbba	version 5.1	2020-11-22 18:36:04 +01:00
Irmen de Jong	510ca042c9	stack tested for most example programs	2020-11-22 18:35:43 +01:00
Irmen de Jong	125f6205f2	optimizing assigning an array value to a var	2020-11-22 17:44:55 +01:00
Irmen de Jong	8136f3df5c	float-const comparison optimizations	2020-11-22 16:54:02 +01:00
Irmen de Jong	38d06a7e94	optimized float var comparison without translateExpression()	2020-11-22 15:05:45 +01:00
Irmen de Jong	49db10539a	optimized float var equality comparison without translateExpression()	2020-11-22 14:33:03 +01:00
Irmen de Jong	8efe4c6267	Fixed compiler watch to work with multiple compilation modules	2020-11-22 13:11:33 +01:00
Irmen de Jong	04d8bb8fbf	Fixed compiler watch flag crashing when not used on a subdirectory. Fixes #20	2020-11-22 12:07:14 +01:00
Irmen de Jong	08aa13c90c	rnd() functions marked as having (internal) side effect	2020-11-22 02:09:32 +01:00
Irmen de Jong	d1febc0208	all in-place byte assignments now without translateExpression()	2020-11-22 01:38:53 +01:00
Irmen de Jong	5980e58ac6	word comparison jumps now without translateExpression()	2020-11-22 01:15:05 +01:00
Irmen de Jong	e1dc283d4b	byte comparison jumps now without translateExpression()	2020-11-21 23:31:26 +01:00
Irmen de Jong	8be234973c	rollback failed optimization of memory expressions (code size got too large)	2020-11-21 19:09:02 +01:00
Irmen de Jong	7def8ff2cd	beginning to optimize comparisons more	2020-11-21 18:44:17 +01:00
Irmen de Jong	340b1c2e42	added balls demo/benchmark	2020-11-21 18:03:57 +01:00
Irmen de Jong	7e0f7ba438	todos	2020-11-20 23:46:14 +01:00
Irmen de Jong	fefd9b52a8	fix for loop with signed byte loopvar over non-const	2020-11-20 22:54:24 +01:00
Irmen de Jong	afd155ac4f	optimize for loops over non const range, without translateExpression()	2020-11-20 22:44:16 +01:00
Irmen de Jong	ee724eb4f1	float variable casts without translateExpression()	2020-11-19 22:58:38 +01:00
Irmen de Jong	2f1f20ea11	rename	2020-11-19 00:28:49 +01:00
Irmen de Jong	063bcf17d8	various inplace modification for word vars now without translateExpression()	2020-11-19 00:08:10 +01:00
Irmen de Jong	72509eef44	inplace modification for memory now without translateExpression()	2020-11-18 23:23:06 +01:00
Irmen de Jong	2da28864e9	inplace not and invert for memory now without translateExpression()	2020-11-18 23:13:07 +01:00
Irmen de Jong	4278f64682	fixed invalid value push for memreads with expression	2020-11-18 22:45:04 +01:00
Irmen de Jong	59ae3c3fcd	<< and >> for byte values slightly optimized, no longer use translateExpression(). preparing for more operator optimizations.	2020-11-18 01:27:02 +01:00
Irmen de Jong	7fa21fbdff	@(...) in an expression is now more efficient, without translateExpression()	2020-11-18 00:58:04 +01:00
Irmen de Jong	e95af7498e	comparing function call result to 0 now more efficient, without translateExpression()	2020-11-18 00:05:48 +01:00
Irmen de Jong	79c75adac1	repeat and when without translateExpression()	2020-11-17 23:52:13 +01:00
Irmen de Jong	d212f69d89	++/-- and @Pc without translateExpression()	2020-11-17 23:40:42 +01:00
Irmen de Jong	edf5e69d39	optimized swap()	2020-11-15 18:04:54 +01:00
Irmen de Jong	574eb0d174	refactoring asmassignment code blocks into utility functions	2020-11-15 17:44:47 +01:00
Irmen de Jong	8bd4914e2f	fix stack error for float casts	2020-11-15 17:34:27 +01:00
Irmen de Jong	5ebaaff64b	refactoring asmassignment code blocks into utility functions	2020-11-15 15:07:55 +01:00
Irmen de Jong	5c9e0c9f51	emit extra nop for breakpoints so vice label list works again (requires 64tass 1.55.2257 or newer!)	2020-11-15 14:31:06 +01:00
Irmen de Jong	8132edbb08	updated some compiled example	2020-11-10 22:51:01 +01:00
Irmen de Jong	d29ce78c86	todos and version	2020-11-10 22:44:48 +01:00
Irmen de Jong	94bc9d7a69	string compare in expression no longer via stack args	2020-11-10 21:48:28 +01:00
Irmen de Jong	e8faec0932	re-introduced more aggressive binexpr splitting optimization	2020-11-10 21:17:33 +01:00
Irmen de Jong	69ca4fe304	cleanup	2020-11-10 21:02:12 +01:00
Irmen de Jong	cd99fe46fd	finished call convention change for builtin functions now no longer via stack	2020-11-10 00:43:45 +01:00
Irmen de Jong	4825b4dc68	fix passing address of pass-by-reference assignment to a UWORD	2020-11-10 00:35:24 +01:00
Irmen de Jong	8d0607ef58	fix missing float casts	2020-11-09 23:57:50 +01:00
Irmen de Jong	225295a7d8	fix float casts	2020-11-09 01:18:22 +01:00
Irmen de Jong	4cd74daf53	float eval result var added, but some examples are broken	2020-11-08 18:54:02 +01:00
Irmen de Jong	6eb9118197	example	2020-11-07 01:08:56 +01:00
Irmen de Jong	d0bd2f522c	rol and ror	2020-11-07 00:56:54 +01:00
Irmen de Jong	661c757236	fix string compare in expressions	2020-11-06 22:59:56 +01:00
Irmen de Jong	aaa20093ef	cleaning up and correcting cc for builtin functions	2020-11-06 00:56:26 +01:00
Irmen de Jong	1eecdd6fa3	fix error when taking address of struct var	2020-11-05 02:39:04 +01:00
Irmen de Jong	800b5b2a43	cleaning up and correcting cc for builtin functions	2020-11-05 02:29:33 +01:00
Irmen de Jong	9d17421c66	implemented the arithmetic functions with new cc. fixed sgn().	2020-11-04 02:27:29 +01:00
Irmen de Jong	0edd50e956	implemented cc for abs()	2020-11-03 23:01:23 +01:00
Irmen de Jong	288d4f08b3	implemented cc for integer sin and cos variants	2020-11-03 22:42:59 +01:00
Irmen de Jong	526e4b8bdc	fix faulty binexpr splitting	2020-11-03 21:31:08 +01:00
Irmen de Jong	e0c5ccc16b	begun with converting builtin functions to new call convention	2020-11-02 23:00:20 +01:00
Irmen de Jong	ebc2c614d7	use non-stack call conv for builtin functions	2020-11-02 19:59:27 +01:00
Irmen de Jong	29f5a85158	callconv	2020-11-01 19:25:23 +01:00
Irmen de Jong	8af2380a47	pair	2020-11-01 18:00:20 +01:00
Irmen de Jong	431f2a2088	optimized memset and memcopy on CX16, memcopy can deal with any size now	2020-11-01 08:00:32 +01:00
Irmen de Jong	e05ea887f6	implement proper returning of float values via FAC1	2020-11-01 06:27:17 +01:00
Irmen de Jong	95c0425151	improved sqrt16	2020-11-01 05:45:49 +01:00
Irmen de Jong	47cbc7b1f9	added a custom-charset example for the c64	2020-10-31 02:26:59 +01:00
Irmen de Jong	e7b75d591c	assigning float results from functions (from FAC1)	2020-10-31 01:22:19 +01:00
Irmen de Jong	99f7d469f4	assigning string result from subroutine	2020-10-30 22:22:06 +01:00
Irmen de Jong	8a6ef17fbf	option	2020-10-30 21:51:15 +01:00
Irmen de Jong	5f337a0bd9	fix typecheck of multiple returnvalues	2020-10-30 21:45:37 +01:00
Irmen de Jong	87862f772a	better handling of inferred type errors	2020-10-30 21:24:49 +01:00
Irmen de Jong	3ab641aa21	removed @stack in subroutine args and returnvalues, can only use variables or registers now	2020-10-30 15:02:42 +01:00
Irmen de Jong	3efa8da8e0	made versions of various builtin funcs returning value in registers	2020-10-30 14:35:20 +01:00
Irmen de Jong	3e28ed4fe4	mader versions of abs() and sgn() returning value in register	2020-10-28 22:56:13 +01:00
Irmen de Jong	44949460ed	change for subroutine return values via registers instead of stack	2020-10-28 00:29:34 +01:00
Irmen de Jong	83cc19ad6f	preparing for subroutine return values via registers instead of stack	2020-10-23 20:56:10 +02:00
Irmen de Jong	66bb98c479	fixed bugs in code assigning values from eval stack	2020-10-23 03:45:09 +02:00
Irmen de Jong	ff3f985658	refactoring	2020-10-22 23:41:16 +02:00
@@ -1 +1 @@
 .6
 .1