Fix libimg to execute emms after mmx code

Also, remove an emms hack from Dino game.
Fix #98.

programs/develop/libraries/libs-dev/libimg/blend_sse.asm

@@ -17,6 +17,55 @@
 ;;                                                                                                ;;
 ;;================================================================================================;;
 
+;;============================================================================;;
+proc img.blend uses ebx esi edi, _bottom, _top, _xbottom, _ybottom, \ ;///////;;
+                                 _xtop, _ytop, _width, _height ;//////////////;;
+;;----------------------------------------------------------------------------;;
+;? Alpha blend _top image to _bottom one (both must be of type Image.bpp32)   ;;
+;;----------------------------------------------------------------------------;;
+;> _bottom = pointer to bottom image (will be changed)                        ;;
+;> _top = pointer to top image (unchanged)                                    ;;
+;> _xbottom = x coord inside _bottom image where to put _top image            ;;
+;> _ybottom = y coord inside _bottom image where to put _top image            ;;
+;> _xtop = x coord inside _top image to start from                            ;;
+;> _ytop = y coord inside _top image to start from                            ;;
+;> _width = width of _top image area to put to _bottom image                  ;;
+;> _height = height of _top image area to put to _bottom image                ;;
+;;----------------------------------------------------------------------------;;
+;< eax = 0 (fail) / _bottom (ok)                                              ;;
+;;============================================================================;;
+        mov     esi, [_top]
+        mov     edi, [_bottom]
+
+        mov     eax, [esi+Image.Width]
+        sub     eax, [_width]
+        shl     eax, 2
+        push    eax
+
+        mov     eax, [edi+Image.Width]
+        sub     eax, [_width]
+        shl     eax, 2
+        push    eax
+
+        mov     eax, [_ytop]
+        imul    eax, [esi+Image.Width]
+        add     eax, [_xtop]
+        shl     eax, 2
+        mov     esi, [esi+Image.Data]
+        add     esi, eax
+
+        mov     eax, [_ybottom]
+        imul    eax, [edi+Image.Width]
+        add     eax, [_xbottom]
+        shl     eax, 2
+        mov     edi, [edi+Image.Data]
+        add     edi, eax
+	stdcall	img._.composite_rgb_00, [_width], [_height]
+        mov     eax, [_bottom]
+        ret
+endp
+
+
 proc	xcf._.blend_rgb
 
 	push		eax ebx
@@ -248,7 +297,7 @@ proc	xcf._.merge_8a _copy_width, _copy_height, _img_total_bpl, _bottom_total_bpl
 endp
 
 
-proc	xcf._.composite_rgb_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
+proc	img._.composite_rgb_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
 
 	pxor		xmm0, xmm0
 
@@ -412,7 +461,7 @@ proc	xcf._.composite_rgb_00 _copy_width, _copy_height, _bottom_total_bpl, _img_t
 endp
 
 
-proc	xcf._.composite_gray_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
+proc	img._.composite_gray_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
 
   .line:
 	mov		ecx, [_copy_width]
@@ -470,7 +519,7 @@ proc	xcf._.composite_gray_00 _copy_width, _copy_height, _bottom_total_bpl, _img_
 endp
 
 
-proc	xcf._.composite_indexed_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
+proc	img._.composite_indexed_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
 
   .line:
 	mov		ecx, [_copy_width]
@@ -495,7 +544,7 @@ proc	xcf._.composite_indexed_00 _copy_width, _copy_height, _bottom_total_bpl, _i
 endp
 
 
-proc	xcf._.composite_rgb_01 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
+proc	img._.composite_rgb_01 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
 	pushad
 
 	pxor		xmm4, xmm4
@@ -542,7 +591,7 @@ proc	xcf._.composite_rgb_01 _copy_width, _copy_height, _bottom_total_bpl, _img_t
 endp
 
 
-proc	xcf._.composite_gray_01 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
+proc	img._.composite_gray_01 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
 	pushad
 
 	pxor		xmm4, xmm4
@@ -587,7 +636,7 @@ proc	xcf._.composite_gray_01 _copy_width, _copy_height, _bottom_total_bpl, _img_
 endp
 
 
-proc	xcf._.composite_rgb_03			; Multiply
+proc	img._.composite_rgb_03			; Multiply
 
 	punpcklbw	xmm2, xmm0
 	punpcklbw	xmm3, xmm0
@@ -598,7 +647,7 @@ proc	xcf._.composite_rgb_03			; Multiply
 endp
 
 
-proc	xcf._.composite_rgb_04			; Screen
+proc	img._.composite_rgb_04			; Screen
 
 	punpcklbw	xmm2, xmm0
 	punpcklbw	xmm3, xmm0
@@ -614,7 +663,7 @@ proc	xcf._.composite_rgb_04			; Screen
 endp
 
 
-proc	xcf._.composite_rgb_05			; Overlay
+proc	img._.composite_rgb_05			; Overlay
 
 	punpcklbw	xmm2, xmm0
 	punpcklbw	xmm3, xmm0
@@ -630,7 +679,7 @@ proc	xcf._.composite_rgb_05			; Overlay
 endp
 
 
-proc	xcf._.composite_rgb_06			; Difference
+proc	img._.composite_rgb_06			; Difference
 
 	movdqa		xmm4, xmm3
 	pminub		xmm4, xmm2
@@ -643,7 +692,7 @@ proc	xcf._.composite_rgb_06			; Difference
 endp
 
 
-proc	xcf._.composite_rgb_07			; Addition
+proc	img._.composite_rgb_07			; Addition
 
 	paddusb		xmm3, xmm2
 	punpcklbw	xmm2, xmm0
@@ -653,7 +702,7 @@ proc	xcf._.composite_rgb_07			; Addition
 endp
 
 
-proc	xcf._.composite_rgb_08			; Subtract
+proc	img._.composite_rgb_08			; Subtract
 
 	movdqa		xmm4, xmm2
 	psubusb		xmm4, xmm3
@@ -665,7 +714,7 @@ proc	xcf._.composite_rgb_08			; Subtract
 endp
 
 
-proc	xcf._.composite_rgb_09			; Darken Only
+proc	img._.composite_rgb_09			; Darken Only
 
 	pminub		xmm3, xmm2
 	punpcklbw	xmm2, xmm0
@@ -675,7 +724,7 @@ proc	xcf._.composite_rgb_09			; Darken Only
 endp
 
 
-proc	xcf._.composite_rgb_10			; Lighten Only
+proc	img._.composite_rgb_10			; Lighten Only
 
 	pmaxub		xmm3, xmm2
 	punpcklbw	xmm2, xmm0
@@ -685,7 +734,7 @@ proc	xcf._.composite_rgb_10			; Lighten Only
 endp
 
 
-proc	xcf._.composite_rgb_11			; Hue (H of HSV)
+proc	img._.composite_rgb_11			; Hue (H of HSV)
 	push		eax ebx ecx edx
 
 	movd		eax, xmm3
@@ -749,7 +798,7 @@ proc	xcf._.composite_rgb_11			; Hue (H of HSV)
 endp
 
 
-proc	xcf._.composite_rgb_12			; Saturation (S of HSV)
+proc	img._.composite_rgb_12			; Saturation (S of HSV)
 	push		eax ebx ecx edx
 
 	movd		eax, xmm3
@@ -807,7 +856,7 @@ proc	xcf._.composite_rgb_12			; Saturation (S of HSV)
 endp
 
 
-proc	xcf._.composite_rgb_13			; Color (H and S of HSL)
+proc	img._.composite_rgb_13			; Color (H and S of HSL)
 	push		eax ebx ecx edx
 
 	movd		eax, xmm3
@@ -854,7 +903,7 @@ proc	xcf._.composite_rgb_13			; Color (H and S of HSL)
 endp
 
 
-proc	xcf._.composite_rgb_14			; Value (V of HSV)
+proc	img._.composite_rgb_14			; Value (V of HSV)
 	push		eax ebx ecx edx
 
 	movd		eax, xmm3
@@ -909,7 +958,7 @@ proc	xcf._.composite_rgb_14			; Value (V of HSV)
 endp
 
 
-proc	xcf._.composite_rgb_15			; Divide
+proc	img._.composite_rgb_15			; Divide
 	push		eax ebx ecx
 
 	movd		eax, xmm3
@@ -1006,7 +1055,7 @@ proc	xcf._.composite_rgb_15			; Divide
 endp
 
 
-proc	xcf._.composite_rgb_16			; Dodge
+proc	img._.composite_rgb_16			; Dodge
 	push		eax ebx ecx
 
 	movd		eax, xmm3
@@ -1107,7 +1156,7 @@ proc	xcf._.composite_rgb_16			; Dodge
 endp
 
 
-proc	xcf._.composite_rgb_17			; Burn
+proc	img._.composite_rgb_17			; Burn
 	push		eax ebx ecx
 
 	movd		eax, xmm3
@@ -1212,7 +1261,7 @@ proc	xcf._.composite_rgb_17			; Burn
 endp
 
 
-proc	xcf._.composite_rgb_18			; Hard Light
+proc	img._.composite_rgb_18			; Hard Light
 	push		eax ebx ecx
 
 	movd		eax, xmm3
@@ -1297,7 +1346,7 @@ proc	xcf._.composite_rgb_18			; Hard Light
 endp
 
 
-proc	xcf._.composite_rgb_20			; Grain Extract
+proc	img._.composite_rgb_20			; Grain Extract
 
 	punpcklbw	xmm2, xmm0
 	punpcklbw	xmm3, xmm0
@@ -1311,7 +1360,7 @@ proc	xcf._.composite_rgb_20			; Grain Extract
 endp
 
 
-proc	xcf._.composite_rgb_21			; Grain Merge
+proc	img._.composite_rgb_21			; Grain Merge
 
 	punpcklbw	xmm2, xmm0
 	punpcklbw	xmm3, xmm0