1.1 --- a/src/swapbuf.asm	Tue Apr 15 20:52:05 2014 +0300
     1.2 +++ b/src/swapbuf.asm	Fri May 02 14:32:58 2014 +0300
     1.3 @@ -1,46 +1,46 @@
     1.4 -; vim:set ft=nasm:
     1.5 -	segment code use32
     1.6 -
     1.7 -	; void swap_buffers_asm(void *dest, void *src, int xsz, int ysz, int bpp)
     1.8 -	; dest	-> eax
     1.9 -	; src	-> edx
    1.10 -	; xsz	-> ebx
    1.11 -	; ysz	-> ecx
    1.12 -	; bpp	-> [ebp + 8] (after pushing ebp)
    1.13 -	global swap_buffers_asm_
    1.14 -swap_buffers_asm_:
    1.15 -	push ebp
    1.16 -	mov ebp, esp
    1.17 -
    1.18 -	mov edi, eax	; let's hold dest ptr in edi, frees up eax
    1.19 -	mov esi, edx	; let's hold src ptr in esi, frees up edx
    1.20 -	; calculate pixel count -> ecx, frees up ebx
    1.21 -	mov eax, ebx
    1.22 -	mul ecx
    1.23 -	mov ecx, eax	; now ecx = xsz * ysz	
    1.24 -	
    1.25 -	mov eax, [ebp + 8]	; eax <- bpp
    1.26 -	cmp eax, 32
    1.27 -	je .bpp32
    1.28 -	cmp eax, 24
    1.29 -	je .bpp24
    1.30 -	cmp eax, 16
    1.31 -	je .bpp16
    1.32 -	; invalid bpp, ignore
    1.33 -	jmp .done
    1.34 -
    1.35 -.bpp32:				; 32bit block transfer, no conversion
    1.36 -	rep movsd		; esi, edi, and ecx already loaded, just go...
    1.37 -	jmp .done
    1.38 -	
    1.39 -.bpp24:				; 32bpp -> 24bpp conversion (LSB-first), 1 byte overrun!
    1.40 -	movsd			; transfer a full 32bit chunk and inc esi,edi by 4
    1.41 -	dec edi			; backtrack dest one byte after last transfer
    1.42 -	dec ecx
    1.43 -	jnz .bpp24
    1.44 -	jmp .done
    1.45 -
    1.46 -.bpp16:				; fuck 16bpp for now (TODO)
    1.47 -.done:
    1.48 -	pop ebp
    1.49 -	ret	
    1.50 +; vim:set ft=nasm:
    1.51 +	segment code use32
    1.52 +
    1.53 +	; void swap_buffers_asm(void *dest, void *src, int xsz, int ysz, int bpp)
    1.54 +	; dest	-> eax
    1.55 +	; src	-> edx
    1.56 +	; xsz	-> ebx
    1.57 +	; ysz	-> ecx
    1.58 +	; bpp	-> [ebp + 8] (after pushing ebp)
    1.59 +	global swap_buffers_asm_
    1.60 +swap_buffers_asm_:
    1.61 +	push ebp
    1.62 +	mov ebp, esp
    1.63 +
    1.64 +	mov edi, eax	; let's hold dest ptr in edi, frees up eax
    1.65 +	mov esi, edx	; let's hold src ptr in esi, frees up edx
    1.66 +	; calculate pixel count -> ecx, frees up ebx
    1.67 +	mov eax, ebx
    1.68 +	mul ecx
    1.69 +	mov ecx, eax	; now ecx = xsz * ysz	
    1.70 +	
    1.71 +	mov eax, [ebp + 8]	; eax <- bpp
    1.72 +	cmp eax, 32
    1.73 +	je .bpp32
    1.74 +	cmp eax, 24
    1.75 +	je .bpp24
    1.76 +	cmp eax, 16
    1.77 +	je .bpp16
    1.78 +	; invalid bpp, ignore
    1.79 +	jmp .done
    1.80 +
    1.81 +.bpp32:				; 32bit block transfer, no conversion
    1.82 +	rep movsd		; esi, edi, and ecx already loaded, just go...
    1.83 +	jmp .done
    1.84 +	
    1.85 +.bpp24:				; 32bpp -> 24bpp conversion (LSB-first), 1 byte overrun!
    1.86 +	movsd			; transfer a full 32bit chunk and inc esi,edi by 4
    1.87 +	dec edi			; backtrack dest one byte after last transfer
    1.88 +	dec ecx
    1.89 +	jnz .bpp24
    1.90 +	jmp .done
    1.91 +
    1.92 +.bpp16:				; fuck 16bpp for now (TODO)
    1.93 +.done:
    1.94 +	pop ebp
    1.95 +	ret