.state32

 .if (R2D_ASM = 1)

 .include "r2d_asm_utils.inc"
 .include "r2d_asm_configs.inc"


    ; Main asm code (it is reflecting the C code of
    ; r2d_blit_lcd_to_lcd)

    .text

 

    .global _r2d_blit_lb,_r2d_blit_la
    .global _r2d_sb_down1,_r2d_sb_down2,_r2d_sa_up1,_r2d_sa_up2
    .global _r2d_scb_down1,_r2d_scb_down2,_r2d_sca_up1,_r2d_sca_up2
    .global _r2d_sdb_down1,_r2d_sdb_down2,_r2d_sda_up1,_r2d_sda_up2
    .global _r2d_skip_color_proc_a
    .global _r2d_end_ynb_color_dst_a
    .global _r2d_skip_color_proc_b
    .global _r2d_end_ynb_color_dst_b
    .global _r2d_skip_color_dst_a
    .global _r2d_skip_color_dst_b 
   

    .global _r2d_sconvcolb_down1,_r2d_econvcolb_down1
    .global _r2d_sconvcolb_down2,_r2d_econvcolb_down2
    .global _r2d_sconvcola_up1,_r2d_econvcola_up1
    .global _r2d_sconvcola_up2,_r2d_econvcola_up2

    .global _r2d_swb_down1,_r2d_swb_down2,_r2d_swa_up1,_r2d_swa_up2


    ; Patch list for write_shift_pixel_down
    .global _r2d_sft_dst_n_b_down1,_r2d_sft_dst_n_b_down2
    .global _r2d_sft_srcb__b_down1,_r2d_sft_srcb__b_down2
    .global _r2d_sft_dstc__b_down1,_r2d_sft_dstc__b_down2

    .global _r2d_sft_src_c_b_down1,_r2d_sft_src_c_b_down2
    .global _r2d_sft_b_c_b_down1,_r2d_sft_b_c_b_down2
    .global _r2d_sft_dstb_c_b_down1,_r2d_sft_dstb_c_b_down2
    .global _r2d_sft_d_c_b_down1,_r2d_sft_d_c_b_down2
    .global _r2d_sft_dstc_c_b_down1,_r2d_sft_dstc_c_b_down2

    ; Patch list for write_shift_pixel_up
    .global _r2d_sft_dst_n_a_up1,_r2d_sft_dst_n_a_up2
    .global _r2d_sft_srcb__a_up1,_r2d_sft_srcb__a_up2
    .global _r2d_sft_dstc__a_up1,_r2d_sft_dstc__a_up2

    .global _r2d_sft_src_c_a_up1,_r2d_sft_src_c_a_up2
    .global _r2d_sft_b_c_a_up1,_r2d_sft_b_c_a_up2

    ; Patch list fro shift_pixel_down
    .global _r2d_sft_dst_n_downa,_r2d_sft_dst_n_downb,_r2d_sft_dst_n_downc
    .global _r2d_sft_srcb_n_downa,_r2d_sft_srcb_n_downb,_r2d_sft_srcb_n_downc

    .global _r2d_sft_src_c_downa,_r2d_sft_src_c_downb,_r2d_sft_src_c_downc

    ; Patch list fro shift_pixel_up
    .global _r2d_sft_dst_n_upa,_r2d_sft_dst_n_upb,_r2d_sft_dst_n_upc
    .global _r2d_sft_srcb_n_upa,_r2d_sft_srcb_n_upb,_r2d_sft_srcb_n_upc

    .global _r2d_sft_src_c_upa,_r2d_sft_src_c_upb,_r2d_sft_src_c_upc
    .global _r2d_enda,_r2d_endb

	.global _r2d_skip_load_src_a_up1,_r2d_skip_load_src_a_up2;
	.global _r2d_skip_load_src_b_down1,_r2d_skip_load_src_b_down2;
	.global _r2d_skip_shift_down_downa,_r2d_skip_shift_down_downb,_r2d_skip_shift_down_downc;
    .global _r2d_skip_shift_up_upa,_r2d_skip_shift_up_upb,_r2d_skip_shift_up_upc;

    .global _r2d_no_null_src_counter_a,_r2d_no_null_src_counter_b
	.global _r2d_end_a_up2,_r2d_end_b_down2,_r2d_end_downa,_r2d_end_downb,_r2d_end_downc;
    .global _r2d_end_upa,_r2d_end_upb,_r2d_end_upc;

	.global _r2d_always_read_a_up1,_r2d_always_read_a_up2
	.global _r2d_always_read_b_down1,_r2d_always_read_b_down2
  
    .ref $r2d_blit_lcd_to_lcd

 .endif

write_shift_pixel_down .macro label
       ; Create free space for result
	   .if (R2D_PIXEL_DEPTH != 32)
_r2d_sft_dst_n_:label:       MOV dstvalue,dstvalue, LSR #R2D_PIXEL_DEPTH 
       .else
_r2d_sft_dst_n_:label:       MOV dstvalue,#0
	   .endif

       ; Mask to extract pixel from src_current
_r2d_sft_src_c_:label:       MOV calca,src_current,LSL #((1<<R2D_MEMORY_WORD) - R2D_PIXEL_DEPTH)
_r2d_always_read_:label:  MOV src_current,src_current
_r2d_sft_b_c_:label:       MOV calca,calca,LSR #((1<<R2D_MEMORY_WORD) - R2D_PIXEL_DEPTH)

       ; Mask to extract pixel from dst_current
_r2d_sft_dstb_c_:label:       MOV calcb,dst_current,LSL #((1<<R2D_MEMORY_WORD) - R2D_PIXEL_DEPTH)
_r2d_sft_d_c_:label:       MOV calcb,calcb,LSR #((1<<R2D_MEMORY_WORD) - R2D_PIXEL_DEPTH)
   
       ; Do foreground detection
_r2d_sd:label:
       MOV calca,calca
       MOV calca,calca
       MOV calca,calca
_r2d_ed:label:

       ; Get foreground/background colors if needed
_r2d_sc:label:
       MOV calca,calca
       MOV calca,calca
       MOV calca,calca
_r2d_ec:label:

_r2d_sconvcol:label:
       MOV calca,calca
       MOV calca,calca
       MOV calca,calca
_r2d_econvcol:label:

       ; Compute drawing mode
_r2d_s:label:
       MOV calca,calca
       MOV calcb,calca
       MOV calcb,calcb
_r2d_e:label:
_r2d_sft_dstc_c_:label:       ORR dstvalue,dstvalue,calcb, LSL #((1<<R2D_MEMORY_WORD) - R2D_PIXEL_DEPTH)

_r2d_sw:label:
       MOV dstvalue,dstvalue
_r2d_ew:label:

       ; Read new data if needed
	   .if (R2D_PIXEL_DEPTH != 32)
_r2d_sft_srcb__:label:       MOV src_current,src_current, LSR #R2D_PIXEL_DEPTH
_r2d_sft_dstc__:label:       MOV dst_current,dst_current, LSR #R2D_PIXEL_DEPTH
       .else
_r2d_sft_srcb__:label:       MOV src_current,src_current,LSR #0
_r2d_sft_dstc__:label:       MOV dst_current,dst_current,LSR #0
	   .endif

       SUBS dstcounter,dstcounter,#1
       MOVEQ dstcounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
_r2d_skip_load_src_:label: MOV src_current,src_current
       SUBS srccounter,srccounter,#1
       BNE M1END?
       MOV srccounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
       LDR src_current,[p_src_current,#4]!
M1END?
  .endm

write_shift_pixel_up .macro label
     .if (R2D_PIXEL_DEPTH != 32)
_r2d_sft_dst_n_:label:       MOV dstvalue,dstvalue, LSL #R2D_PIXEL_DEPTH 
     .else
_r2d_sft_dst_n_:label:       MOV dstvalue,#0
	 .endif

_r2d_sft_src_c_:label:      MOV calca,src_current,LSR #((1<<R2D_MEMORY_WORD) - R2D_PIXEL_DEPTH)
_r2d_always_read_:label:  MOV src_current,src_current
_r2d_sft_b_c_:label:      MOV calcb,dst_current,LSR #((1<<R2D_MEMORY_WORD) - R2D_PIXEL_DEPTH)

       ; Do foreground detection
_r2d_sd:label:
       MOV calca,calca
       MOV calca,calca
       MOV calca,calca
_r2d_ed:label:

       ; Get foreground/background colors if needed
_r2d_sc:label:
       MOV calca,calca
       MOV calca,calca
       MOV calca,calca
_r2d_ec:label:

_r2d_sconvcol:label:
       MOV calca,calca
       MOV calca,calca
       MOV calca,calca
_r2d_econvcol:label:

       ; Compute drawing mode
_r2d_s:label:
       MOV calca,calca
       MOV calcb,calca
       MOV calcb,calcb
_r2d_e:label:
       ORR dstvalue,dstvalue,calcb

_r2d_sw:label:
       MOV dstvalue,dstvalue
_r2d_ew:label:

       .if (R2D_PIXEL_DEPTH != 32)
_r2d_sft_srcb__:label:       MOV src_current,src_current, LSL #R2D_PIXEL_DEPTH
_r2d_sft_dstc__:label:       MOV dst_current,dst_current, LSL #R2D_PIXEL_DEPTH
       .else
_r2d_sft_srcb__:label:       MOV src_current,src_current,LSL #0
_r2d_sft_dstc__:label:       MOV dst_current,dst_current,LSL #0
	   .endif

       SUBS dstcounter,dstcounter,#1
       MOVEQ dstcounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
_r2d_skip_load_src_:label: MOV src_current,src_current
       SUBS srccounter,srccounter,#1
       BNE M1END?
       MOV srccounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
       LDR src_current,[p_src_current,#-4]!
M1END?
  .endm

shift_pixel_down .macro arg,reg,label
_r2d_skip_shift_down_:label:  MOV arg,arg

     .if (R2D_PIXEL_DEPTH != 32)
_r2d_sft_dst_n_:label:       MOV arg,arg, LSR #R2D_PIXEL_DEPTH 
     .else
_r2d_sft_dst_n_:label:       MOV arg,#0
	 .endif

_r2d_sft_src_c_:label:       ORR arg,arg,:reg:_current, LSL #((1<<R2D_MEMORY_WORD) - R2D_PIXEL_DEPTH)

     .if (R2D_PIXEL_DEPTH != 32)
_r2d_sft_srcb_n_:label:       MOV :reg:_current,:reg:_current, LSR #R2D_PIXEL_DEPTH
     .else
_r2d_sft_srcb_n_:label:       MOV :reg:_current,#0
	 .endif

       SUBS :reg:counter,:reg:counter,#1
       BNE M1END?
       MOV :reg:counter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
       LDR :reg:_current,[p_:reg:_current,#4]!
M1END?
  .endm

shift_pixel_up .macro arg,reg,label
_r2d_skip_shift_up_:label:  MOV arg,arg
     .if (R2D_PIXEL_DEPTH != 32)
_r2d_sft_dst_n_:label:       MOV arg,arg, LSL #R2D_PIXEL_DEPTH 
     .else
_r2d_sft_dst_n_:label:       MOV arg,#0
	 .endif

_r2d_sft_src_c_:label:       ORR arg,arg,:reg:_current, LSR #((1<<R2D_MEMORY_WORD) - R2D_PIXEL_DEPTH)

      .if (R2D_PIXEL_DEPTH != 32)
_r2d_sft_srcb_n_:label:       MOV :reg:_current,:reg:_current, LSL #R2D_PIXEL_DEPTH
      .else
_r2d_sft_srcb_n_:label:       MOV :reg:_current,#0
	  .endif


       SUBS :reg:counter,:reg:counter,#1
       BNE M1END?
       MOV :reg:counter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
       LDR :reg:_current,[p_:reg:_current,#-4]!
M1END?
  .endm

  .if (R2D_ASM = 1)

;--------------------------------------
;
; v_direction == -1
;

_r2d_blit_la
	STMFD SP!, {R4,R5,R6,R7,R8,R9,R10,R11,R12,LR}
    LDR dstvalue, [R0,#S_FRAME.dstvalue_ind]
    LDR dst_current, [R0,#S_FRAME.dst_current_ind]
    LDRH dstcounter, [R0,#S_FRAME.dstcounter_ind]
    LDR src_current, [R0,#S_FRAME.src_current_ind]
    LDRH ynb, [R0,#S_FRAME.ynb_ind]
    LDR p_src_current, [R0,#S_FRAME.p_src_current_ind]
    LDR p_dst_current, [R0,#S_FRAME.p_dst_current_ind]
    LDR x, [R0,#S_FRAME.x_ind]
    LDR y, [R0,#S_FRAME.y_ind]
    LDRH current_dst_dy, [R0,#S_FRAME.current_dst_dy_ind]
    LDRH current_src_dy, [R0,#S_FRAME.current_src_dy_ind]
    LDRH dst_dy, [R0,#S_FRAME.dst_dy_ind]


;-------------------------------------
;
; First while
;
; current_dst_dy != dst_dy
current_dst_a: CMP current_dst_dy,dst_dy
  BEQ end_current_dst_a

  shift_pixel_up dstvalue,dst,upa

_r2d_end_upa:  SUB current_dst_dy,current_dst_dy,#1
  SUB y,y,#1
  B current_dst_a

end_current_dst_a:

  ; Load srcounter(and overwrite dst_dy)
  LDRH srccounter, [R0,#S_FRAME.srccounter_ind]

  LDRH temp,[R0,#S_FRAME.src_dy_ind]
  SUB current_src_dy,current_src_dy,temp

;-------------------------------------
;
; Second while
;
; current_src_dy - src_dy != 0
current_src_a: CMP current_src_dy,#0
  BEQ end_current_src_a

  shift_pixel_up temp,src,upb

_r2d_end_upb:  SUB current_src_dy,current_src_dy,#1
  B current_src_a

end_current_src_a:

;-------------------------------------
;
; Third while
;
; (ynb!=0)
_r2d_ynb_loop_a: TEQ ynb,#0
       BEQ end_ynb_loop_a

      write_shift_pixel_up a_up1 ; was a_up1

_r2d_no_null_src_counter_a:
       SUB y,y,#1 

       ; When dst is color framebuffer, following instruction is replaced by
       ; BL no_end_dst_counter_a so, to ynb_loop_a
_r2d_skip_color_proc_a: MOV y,y

       CMP dstcounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
       BNE no_end_dst_counter_a
       
       SUB ynb,ynb,#1
       STR dstvalue,[p_dst_current],#-4
       
       LDR dst_current,[p_dst_current]
       MOV dstvalue,#0
       MOV current_dst_dy,#((1<<R2D_PIXELS_PER_MEMORY_WORD)-1)

no_end_dst_counter_a:       B _r2d_ynb_loop_a
_r2d_end_ynb_color_dst_a:SUB ynb,ynb,#1
   B _r2d_ynb_loop_a

end_ynb_loop_a:
_r2d_skip_color_dst_a: MOV y,y

    ADD current_dst_dy,current_dst_dy,#1
    LDRH temp,[R0,#S_FRAME.end_dst_dy_ind]
    SUB current_dst_dy,current_dst_dy,temp

;-------------------------------------
;
; Fourth while
;
; current_dst_dy - end_dst_dy != 0
end_dst_loop_a: TEQ current_dst_dy,#0
    BEQ end_end_dst_loop_a
    
    write_shift_pixel_up a_up2

_r2d_end_a_up2:   SUB current_dst_dy,current_dst_dy,#1
    SUB y,y,#1
    B end_dst_loop_a
end_end_dst_loop_a:
    MOV temp,#0
    CMP dstcounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
    MOVEQ temp,#1

;-------------------------------------
;
; Last while
;
; dstcounter != (1<<R2D_PIXELS_PER_MEMORY_WORD)
dst_counter_loop_a: CMP dstcounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
  BEQ end_dst_counter_loop_a
  shift_pixel_up dstvalue,dst,upc
_r2d_end_upc:  SUB y,y,#1
  B dst_counter_loop_a
end_dst_counter_loop_a:
    CMP temp,#0
    STREQ dstvalue,[p_dst_current,#4]!

    CMP temp,#0
    STRNE dstvalue,[p_dst_current]
_r2d_enda:
	LDMFD SP!, {R4,R5,R6,R7,R8,R9,R10,R11,R12,LR}
	MOV PC,LR


;--------------------------------------
;
; v_direction != -1
;
_r2d_blit_lb
	STMFD SP!, {R4,R5,R6,R7,R8,R9,R10,R11,R12,LR}
    LDR dstvalue, [R0,#S_FRAME.dstvalue_ind]
    LDR dst_current, [R0,#S_FRAME.dst_current_ind]
    LDRH dstcounter, [R0,#S_FRAME.dstcounter_ind]
    LDR src_current, [R0,#S_FRAME.src_current_ind]
    LDRH ynb, [R0,#S_FRAME.ynb_ind]
    LDR p_src_current, [R0,#S_FRAME.p_src_current_ind]
    LDR p_dst_current, [R0,#S_FRAME.p_dst_current_ind]
    LDR x, [R0,#S_FRAME.x_ind]
    LDR y, [R0,#S_FRAME.y_ind]
    LDRH current_dst_dy, [R0,#S_FRAME.current_dst_dy_ind]
    LDRH current_src_dy, [R0,#S_FRAME.current_src_dy_ind]
    LDRH dst_dy, [R0,#S_FRAME.dst_dy_ind]


;-------------------------------------
;
; First while
;
; current_dst_dy != dst_dy
current_dst_b: CMP current_dst_dy,dst_dy
  BEQ end_current_dst_b

  shift_pixel_down dstvalue,dst,downa

_r2d_end_downa:  ADD current_dst_dy,current_dst_dy,#1
  ADD y,y,#1
  B current_dst_b

end_current_dst_b:

  ; Load srcounter(and overwrite dst_dy)
  LDRH srccounter, [R0,#S_FRAME.srccounter_ind]

  LDRH temp,[R0,#S_FRAME.src_dy_ind]
  SUB current_src_dy,current_src_dy,temp

;-------------------------------------
;
; Second while
;
; current_src_dy - src_dy != 0
current_src_b: CMP current_src_dy,#0
  BEQ end_current_src_b

  shift_pixel_down temp,src,downb

_r2d_end_downb:  ADD current_src_dy,current_src_dy,#1
  B current_src_b

end_current_src_b:

;-------------------------------------
;
; Third while
;
; (ynb!=0)
_r2d_ynb_loop_b: TEQ ynb,#0
       BEQ end_ynb_loop_b

      write_shift_pixel_down b_down1

_r2d_no_null_src_counter_b:
       ADD y,y,#1 

       ; When dst is color framebuffer, following instruction is replaced by
       ; BL no_end_dst_counter_a so, to ynb_loop_a
_r2d_skip_color_proc_b: MOV y,y

       CMP dstcounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
       BNE no_end_dst_counter_b
       
       SUB ynb,ynb,#1
       STR dstvalue,[p_dst_current],#4
       
       LDR dst_current,[p_dst_current]
       MOV dstvalue,#0
       MOV current_dst_dy,#0

no_end_dst_counter_b:       B _r2d_ynb_loop_b
_r2d_end_ynb_color_dst_b:    SUB ynb,ynb,#1
   B _r2d_ynb_loop_b
end_ynb_loop_b:
_r2d_skip_color_dst_b: MOV y,y

    SUB current_dst_dy,current_dst_dy,#1
    LDRH temp,[R0,#S_FRAME.end_dst_dy_ind]
    SUB current_dst_dy,current_dst_dy,temp

;-------------------------------------
;
; Fourth while
;
; current_dst_dy - end_dst_dy != 0
end_dst_loop_b: TEQ current_dst_dy,#0
    BEQ end_end_dst_loop_b
    
    write_shift_pixel_down b_down2

_r2d_end_b_down2:    ADD current_dst_dy,current_dst_dy,#1
    ADD y,y,#1
    B end_dst_loop_b
end_end_dst_loop_b:
    MOV temp,#0
    CMP dstcounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
    MOVEQ temp,#1

;-------------------------------------
;
; Last while
;
; dstcounter != (1<<R2D_PIXELS_PER_MEMORY_WORD)
dst_counter_loop_b: CMP dstcounter,#(1<<R2D_PIXELS_PER_MEMORY_WORD)
  BEQ end_dst_counter_loop_b
  shift_pixel_down dstvalue,dst,downc
_r2d_end_downc:  ADD y,y,#1
  B dst_counter_loop_b
end_dst_counter_loop_b:
    CMP temp,#0
    STREQ dstvalue,[p_dst_current,#-4]!

    CMP temp,#0
    STRNE dstvalue,[p_dst_current]
_r2d_endb:
	LDMFD SP!, {R4,R5,R6,R7,R8,R9,R10,R11,R12,LR}
	MOV PC,LR

 .endif

    .end