--- uae/src/X86.S	2018/04/24 16:38:39	1.1.1.1
+++ uae/src/X86.S	2018/04/24 16:39:04	1.1.1.2
@@ -2,171 +2,168 @@
  * Feel free to send me Sparc/PPC/Alpha versions of this... :)
  */
 
-#ifdef __DOS__
-#define SYM(NAME) _##NAME
-#define _color_regs _bpl_info
+#ifdef X86_PPRO_OPT
+#define BYTE_REG(a,b) a
+#define BYTE_MOVE movzbl
 #else
-#define SYM(NAME) NAME
+#define BYTE_REG(a,b) b
+#define BYTE_MOVE movb
 #endif
 
+#ifndef USE_UNDERSCORE
+#define SYM(NAME) NAME
+#define FUNCTION_ALIGN .align 16
+/* Urgl... */
+#else
+#define SYM(NAME) _##NAME
 /* Urgl... */
-#define color_regs bpl_info
+#define FUNCTION_ALIGN .align 4
+#endif
 
-#ifndef __DOS__
+#ifndef USE_UNDERSCORE
 	.local	lastcolor
 	.comm	lastcolor,2
-	.align 16
 #else
 	.comm	lastcolor,2
-	.align 8
 #endif
 
-.globl pfield_linetoscr_full8
-pfield_linetoscr_full8:
+        .globl pfield_linetoscr_asm8
+	.type  pfield_linetoscr_asm8,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_asm8:
 	pushl %ebp
 	pushl %edi
 	pushl %esi
 	pushl %ebx
-	movl 20(%esp),%esi
-	movl 24(%esp),%edi
+	movl 24(%esp),%esi /* start */
+	movl 28(%esp),%edi /* lframe_end */
 	
-	xorl %edx,%edx
-	movl SYM(diwfirstword),%eax
-	cmpl $4,%eax
-	jc .Loops1
-	movl %edx,SYM(pixdata)-4(%eax)
-	movl %edx,SYM(spixstate)-4(%eax)
-.Loops1:	
-	movl SYM(diwlastword),%eax
-	movl %edx,SYM(pixdata)(%eax)
-	movl %edx,SYM(spixstate)(%eax)
-	
-	movl SYM(acolors),%edx
+	movl SYM(colors_for_drawing),%edx
 	movb %dl,%dh
 	movw %dx,%ax
-	bswapl %eax
+	sall $16,%eax
 	movw %dx,%ax
 	
+        movl 20(%esp),%ebx
 	movl SYM(xlinebuffer),%ecx
-	leal 4(%esi),%edx
-	cmpl %edx,SYM(diwfirstword)
-	jle .La_start_ok
+	jmp .La_do_start_line
 .La_start_line:
-	cmpl %edi,%esi
-	jge .La_start_ok
-	movl %eax,(%ecx)
-	addl $4,%ecx
-	movl %edx,%esi
-
-	leal 4(%esi),%edx
-	cmpl %edx,SYM(diwfirstword)
-	jg .La_start_line
-.La_start_ok:
-        movl %ecx,SYM(xlinebuffer)
-
-        cmpl $0,SYM(bplham)
-	je .La_noham
-	cmpl $6,SYM(bplplanecnt)
-	jne .La_noham
-	
-	/* HAM 6 */
-	movl SYM(xlinebuffer),%ebp
-	xorl %edx,%edx
-.La_hamloop:
-	cmpl %esi,SYM(diwlastword)
-	jle .La_hamend
-	cmpl %edi,%esi
-	jge .La_hamend
-		
+	movl %eax,(%ecx,%esi)
+	addl $4,%esi
+.La_do_start_line:
+	cmpl %esi,%edi
+	jnbe .La_start_line
+
+	call %ebx
+
+	/* Finish off the line */
+
+	movl SYM(xlinebuffer),%ecx
+        movl 36(%esp),%edi /* stoppos */
+	movl SYM(colors_for_drawing),%edx
+	movb %dl,%dh
+	movw %dx,%ax
+	sall $16,%eax
+	movw %dx,%ax
+	jmp .La_do_end_line
+.La_end_line:
+	movl %eax,(%ecx,%esi)
+	addl $4,%esi
+.La_do_end_line:
+	cmpl %esi,%edi
+	jnbe .La_end_line
+
+	popl %ebx
+	popl %esi
+	popl %edi
+	popl %ebp
+	ret
+
+        .globl pfield_linetoscr_ham6_asm8
+	.type  pfield_linetoscr_ham6_asm8,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_ham6_asm8:
+	movl 36(%esp),%edi /* diw_end */
+.Laa_hamloop:
+	cmpl %esi,%edi
+	jbe .Laa_hamend
+
 	movl SYM(ham_linebuf)+12(,%esi,4),%eax
-	movb SYM(xcolors)(,%eax,4),%bl
-	movl SYM(ham_linebuf)+8(,%esi,4),%eax
 	movb SYM(xcolors)(,%eax,4),%bh
-	bswapl %ebx
+	movl SYM(ham_linebuf)+8(,%esi,4),%eax
+	movb SYM(xcolors)(,%eax,4),%bl
+	sall $16,%ebx
 	movl SYM(ham_linebuf)+4(,%esi,4),%eax
 	movb SYM(xcolors)(,%eax,4),%bh
 	movl SYM(ham_linebuf)(,%esi,4),%eax
 	movb SYM(xcolors)(,%eax,4),%bl
-	movl %ebx,(%ebp)
-	addl $4,%ebp
+	movl %ebx,(%ecx,%esi)
 	addl $4,%esi
-	jmp .La_hamloop
+	jmp .Laa_hamloop
 
-.La_hamend:
-	movl %ebp,SYM(xlinebuffer)
-	jmp .La_finish_line
-	
-.La_noham:
-	cmpl $0,SYM(bpldualpf)
-	je .La_normal
-	
-	/* Dual Playfield */
-	
-	xorl %edx,%edx
-
-	movl $SYM(dblpf_ind1),%ebp
-	cmpl $0,SYM(bpldualpfpri)
-	je .La_dp_loop
-	movl $SYM(dblpf_ind2),%ebp
-.La_dp_loop:
-	cmpl %esi,SYM(diwlastword)
-	jle .La_finish_line
-	
-	cmpl %edi,%esi
-	jge .La_finish_line
+.Laa_hamend:
+        ret
 
+        .globl pfield_linetoscr_dualpf_asm8
+	.type  pfield_linetoscr_dualpf_asm8,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_dualpf_asm8:
+	xorl %edx,%edx
+	movl 44(%esp),%ebp /* lookup_no */
+.Lab_dp_loop:
+        movl 36(%esp),%eax /* diw_end */
 	xorl %edx,%edx
+	cmpl %esi,%eax
+	jbe .La_dp_done
+
 	movl SYM(pixdata)(%esi),%eax
 	movl SYM(spixstate)(%esi),%ebx
 	testl %ebx,%ebx
-	jne .La_dp_sprites
+	jne .Lab_dp_sprites
 	
 	/* no sprites */
-
-	movb %al,%dl
-	movb (%ebp,%edx,4),%dl
-	movb SYM(acolors)(,%edx,4),%cl
-	movb %ah,%dl
-	movb (%ebp,%edx,4),%dl
-	movb SYM(acolors)(,%edx,4),%ch
+	
+	BYTE_MOVE %al,BYTE_REG(%edx,%dl)
+	BYTE_MOVE (%ebp,%edx,4),BYTE_REG(%edx,%dl)
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
+	BYTE_MOVE %ah,BYTE_REG(%edx,%dl)
+	BYTE_MOVE (%ebp,%edx,4),BYTE_REG(%edx,%dl)
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
 	bswapl %eax
 	bswapl %ecx
-	movb %al,%dl
-	movb (%ebp,%edx,4),%dl
-	movb SYM(acolors)(,%edx,4),%cl
-	movb %ah,%dl
-	movb (%ebp,%edx,4),%dl
-	movb SYM(acolors)(,%edx,4),%ch
+	BYTE_MOVE %al,BYTE_REG(%edx,%dl)
+	BYTE_MOVE (%ebp,%edx,4),BYTE_REG(%edx,%dl)
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
+	BYTE_MOVE %ah,BYTE_REG(%edx,%dl)
+	BYTE_MOVE (%ebp,%edx,4),BYTE_REG(%edx,%dl)
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
 	bswapl %ecx
 	
 	movl SYM(xlinebuffer),%eax
-	movl %ecx,(%eax)
-	addl $4,SYM(xlinebuffer)
+	movl %ecx,(%eax,%esi)
 	addl $4,%esi
-	jmp .La_dp_loop
+	jmp .Lab_dp_loop
 	
-.La_dp_sprites:
-
+.Lab_dp_sprites:
         /* dual playfield with sprites */
 
 	movb %al,%dl
 	testb %bl,%bl
 	je .La3
-	movb SYM(acolors)(,%edx,4),%cl
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
 	jmp .La1
 .La3:
 	movb (%ebp,%edx,4),%dl
-	movb SYM(acolors)(,%edx,4),%cl
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
 .La1:
-
 	movb %ah,%dl
 	testb %bh,%bh
 	je .La6
-	movb SYM(acolors)(,%edx,4),%ch
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
 	jmp .La4
 .La6:
 	movb (%ebp,%edx,4),%dl
-	movb SYM(acolors)(,%edx,4),%ch
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
 .La4:
 	bswapl %ecx	
 	bswapl %eax
@@ -175,358 +172,518 @@ pfield_linetoscr_full8:
 	movb %al,%dl
 	testb %bl,%bl
 	je .La9
-	movb SYM(acolors)(,%edx,4),%cl
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
 	jmp .La7
 .La9:
 	movb (%ebp,%edx,4),%dl
-	movb SYM(acolors)(,%edx,4),%cl
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
 .La7:
 
 	movb %ah,%dl
 	testb %bh,%bh
 	je .La12
-	movb SYM(acolors)(,%edx,4),%ch
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
 	jmp .La10
 .La12:
 	movb (%ebp,%edx,4),%dl
-	movb SYM(acolors)(,%edx,4),%ch
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
 .La10:
-	addl $4,%esi
-	
 	bswapl %ecx
 	movl SYM(xlinebuffer),%eax
-	movl %ecx,(%eax)
-	addl $4,SYM(xlinebuffer)
-	
-	jmp .La_dp_loop
-	
-	/* normal case */	
-.La_normal:
+	movl %ecx,(%eax,%esi)
+	addl $4,%esi
+	jmp .Lab_dp_loop
+
+.La_dp_done:
+	ret
+
+        .globl pfield_linetoscr_ehb_asm8
+	.type  pfield_linetoscr_ehb_asm8,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_ehb_asm8:
 	xorl %edx,%edx
 	movl SYM(xlinebuffer),%ebx
-.La_normal_loop:
-	cmpl %esi,SYM(diwlastword)
-	jle .La_normal_end
-	cmpl %edi,%esi
-	jge .La_normal_end
+.La_ehb_loop:
+	movl 36(%esp),%edi /* diw_end */
+	cmpl %esi,%edi
+	jbe .La_ehb_end
 	
 	movl SYM(pixdata)(%esi),%eax
-	
-	movb %al,%dl
-	movb SYM(acolors)(,%edx,4),%cl
-	movb %ah,%dl
-	movb SYM(acolors)(,%edx,4),%ch
-	bswapl %eax
-	bswapl %ecx
-	movb %al,%dl
-	movb SYM(acolors)(,%edx,4),%cl
-	movb %ah,%dl
-	movb SYM(acolors)(,%edx,4),%ch
-	bswapl %ecx
-	
-	movl %ecx,(%ebx)
-	addl $4,%ebx
+	BYTE_MOVE %al,BYTE_REG(%edx,%dl)
+	cmpb $32,%al
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
+	jc   .La_ehb1
+	movw SYM(colors_for_drawing)+32*4-32*2(,%edx,2),%di
+	shrl $1,%edi
+	andl $0x777,%edi
+	movb SYM(xcolors)(,%edi,4),%cl
+.La_ehb1:
+	BYTE_MOVE %ah,BYTE_REG(%edx,%dl)
+	cmpb $32,%ah
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
+	jc   .La_ehb2
+	movw SYM(colors_for_drawing)+32*4-32*2(,%edx,2),%di
+	shrl $1,%edi
+	andl $0x777,%edi
+	movb SYM(xcolors)(,%edi,4),%ch
+.La_ehb2:
+	sarl $16,%eax
+	sall $16,%ecx
+	BYTE_MOVE %al,BYTE_REG(%edx,%dl)
+	cmpb $32,%al
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
+	jc   .La_ehb3
+	movw SYM(colors_for_drawing)+32*4-32*2(,%edx,2),%di
+	shrl $1,%edi
+	andl $0x777,%edi
+	movb SYM(xcolors)(,%edi,4),%cl
+.La_ehb3:
+        BYTE_MOVE %ah,BYTE_REG(%edx,%dl)
+	cmpb $32,%ah
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
+	jc   .La_ehb4
+	movw SYM(colors_for_drawing)+32*4-32*2(,%edx,2),%di
+	shrl $1,%edi
+	andl $0x777,%edi
+	movb SYM(xcolors)(,%edi,4),%ch
+.La_ehb4:
+	roll $16,%ecx
+
+	movl %ecx,(%ebx,%esi)
 	addl $4,%esi
-	jmp .La_normal_loop
+	jmp .La_ehb_loop
 
-.La_normal_end:
-	movl %ebx,SYM(xlinebuffer)
+.La_ehb_end:
+        ret
 
-.La_finish_line:
-	/* Finish off the line */
-	
-	movl SYM(acolors),%edx
-	movb %dl,%dh
-	movw %dx,%ax
-	bswapl %eax
-	movw %dx,%ax
-	
-	movl SYM(xlinebuffer),%ecx
-.La_fin_loop:
-	cmpl %edi,%esi
-	jge .La_fin_end
-	
-	movl %eax,(%ecx)
-	addl $4,%ecx
+        .globl pfield_linetoscr_normal_asm8
+	.type  pfield_linetoscr_normal_asm8,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_normal_asm8:
+	xorl %edx,%edx
+	xorl %ebx,%ebx
+	movl 36(%esp),%edi /* diw_end */
+	movl SYM(xlinebuffer),%ebp
+	jmp .La_normal_doloop
+	FUNCTION_ALIGN
+.La_normal_loop:
+        BYTE_MOVE SYM(pixdata)(%esi),BYTE_REG(%edx,%dl)
+	BYTE_MOVE SYM(pixdata)+4(%esi),BYTE_REG(%ebx,%bl)
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
+	movb SYM(colors_for_drawing)(,%ebx,4),%al
+	BYTE_MOVE SYM(pixdata)+1(%esi),BYTE_REG(%edx,%dl)
+	BYTE_MOVE SYM(pixdata)+5(%esi),BYTE_REG(%ebx,%bl)
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
+	movb SYM(colors_for_drawing)(,%ebx,4),%ah
+
+	sall $16,%ecx
+	BYTE_MOVE SYM(pixdata)+2(%esi),BYTE_REG(%edx,%dl)
+	sall $16,%eax
+
+	BYTE_MOVE SYM(pixdata)+6(%esi),BYTE_REG(%ebx,%bl)
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
+	movb SYM(colors_for_drawing)(,%ebx,4),%al
+	BYTE_MOVE SYM(pixdata)+3(%esi),BYTE_REG(%edx,%dl)
+	BYTE_MOVE SYM(pixdata)+7(%esi),BYTE_REG(%ebx,%bl)
+	movb SYM(colors_for_drawing)(,%edx,4),%ch
 	addl $4,%esi
-	jmp  .La_fin_loop
+	roll $16,%ecx
+	movb SYM(colors_for_drawing)(,%ebx,4),%ah
+	roll $16,%eax
+
+	movl %ecx,-4(%ebp,%esi)
+	cmpl %esi,%edi
+	jbe .La_normal_end
+	movl %eax,(%ebp,%esi)
+	addl $4,%esi
+.La_normal_doloop:
+	cmpl %esi,%edi
+	jnbe .La_normal_loop
 
-.La_fin_end:
-        movl %ecx,SYM(xlinebuffer)
-	
-	popl %ebx
-	popl %esi
-	popl %edi
-	popl %ebp
-	ret
+.La_normal_end:
+        ret
 
-#ifndef __DOS__
-	.align 16
-#else
-	.align 8
-#endif
-.globl pfield_linetoscr_full16
-pfield_linetoscr_full16:
+        .globl pfield_linetoscr_hdouble_asm8
+	.type  pfield_linetoscr_hdouble_asm8,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_hdouble_asm8:
+        shrl $1,%esi
+	xorl %edx,%edx
+	xorl %ebx,%ebx
+	movl 36(%esp),%edi /* diw_end */
+	shrl $1,%edi
+	movl SYM(xlinebuffer),%ebp
+	jmp .La_hdouble_doloop
+	FUNCTION_ALIGN
+.La_hdouble_loop:
+        BYTE_MOVE SYM(pixdata)+1(%esi),BYTE_REG(%edx,%dl)
+	BYTE_MOVE SYM(pixdata)+3(%esi),BYTE_REG(%ebx,%bl)
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
+	movb SYM(colors_for_drawing)(,%ebx,4),%al
+	BYTE_MOVE SYM(pixdata)+0(%esi),BYTE_REG(%edx,%dl)
+	movb %cl,%ch
+	BYTE_MOVE SYM(pixdata)+2(%esi),BYTE_REG(%ebx,%bl)
+	movb %al,%ah
+	sall $16,%ecx
+	movb SYM(colors_for_drawing)(,%edx,4),%cl
+	sall $16,%eax
+	movb SYM(colors_for_drawing)(,%ebx,4),%al
+	addl $2,%esi
+	movb %cl,%ch
+	movb %al,%ah
+
+	movl %ecx,-4(%ebp,%esi,2)
+	cmpl %esi,%edi
+	jbe .La_hdouble_end
+ 	movl %eax,(%ebp,%esi,2)
+	addl $2,%esi
+.La_hdouble_doloop:
+	cmpl %esi,%edi
+	jnbe .La_hdouble_loop
+
+.La_hdouble_end:
+        addl %esi,%esi
+        ret
+
+/* 16 bit functions */
+
+        .globl pfield_linetoscr_asm16
+	.type  pfield_linetoscr_asm16,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_asm16:
 	pushl %ebp
 	pushl %edi
 	pushl %esi
 	pushl %ebx
+	movl 24(%esp),%esi /* start */
+	movl 28(%esp),%edi /* lframe_end */
 	
-	movl 20(%esp),%esi
-	movl 24(%esp),%edi
-	
-	xorl %edx,%edx
-	movl SYM(diwfirstword),%eax
-	cmpl $4,%eax
-	jc .Lboops1
-	movl %edx,SYM(pixdata)-4(%eax)
-	movl %edx,SYM(spixstate)-4(%eax)
-.Lboops1:	
-	movl SYM(diwlastword),%eax
-	movl %edx,SYM(pixdata)(%eax)
-	movl %edx,SYM(spixstate)(%eax)
-	
-	movl SYM(acolors),%edx
-	movw %dx,%ax
+	movl SYM(colors_for_drawing),%eax
+	movl %eax,%edx
 	sall $16,%eax
 	movw %dx,%ax
 	
+        movl 20(%esp),%ebx
 	movl SYM(xlinebuffer),%ecx
-	leal 4(%esi),%edx
-	cmpl %edx,SYM(diwfirstword)
-	jle .Lb_start_ok
+	jmp .Lb_do_start_line
 .Lb_start_line:
-	cmpl %edi,%esi
-	jge .Lb_start_ok
-	movl %eax,(%ecx)
-	movl %eax,4(%ecx)
-	addl $8,%ecx
-	movl %edx,%esi
-
-	leal 4(%esi),%edx
-	cmpl %edx,SYM(diwfirstword)
-	jg .Lb_start_line
-.Lb_start_ok:
-        movl %ecx,SYM(xlinebuffer)
-
-        cmpl $0,SYM(bplham)
-	je .Lb_noham
-	cmpl $6,SYM(bplplanecnt)
-	jne .Lb_noham
-	
-	/* HAM 6 */
-	movl SYM(xlinebuffer),%ebp
-	xorl %edx,%edx
+	movl %eax,(%ecx,%esi,2)
+	movl %eax,4(%ecx,%esi,2)
+	addl $4,%esi
+.Lb_do_start_line:
+	cmpl %esi,%edi
+	jnbe .Lb_start_line
+
+	call %ebx
+
+	/* Finish off the line */
+
+	movl SYM(xlinebuffer),%ecx
+        movl 36(%esp),%edi /* stoppos */
+	movl SYM(colors_for_drawing),%eax
+	movl %eax,%edx
+	sall $16,%eax
+	movw %dx,%ax
+	jmp .Lb_do_end_line
+.Lb_end_line:
+	movl %eax,(%ecx,%esi,2)
+	movl %eax,4(%ecx,%esi,2)
+	addl $4,%esi
+.Lb_do_end_line:
+	cmpl %esi,%edi
+	jnbe .Lb_end_line
+
+	popl %ebx
+	popl %esi
+	popl %edi
+	popl %ebp
+	ret
+
+        .globl pfield_linetoscr_ham6_asm16
+	.type  pfield_linetoscr_ham6_asm16,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_ham6_asm16:
+	movl 36(%esp),%edi /* diw_end */
 .Lb_hamloop:
-	cmpl %esi,SYM(diwlastword)
-	jle .Lb_hamend
-	cmpl %edi,%esi
-	jge .Lb_hamend
+	cmpl %esi,%edi
+	jbe .Lb_hamend
 
 	movl SYM(ham_linebuf)+4(,%esi,4),%eax
-	movw SYM(xcolors)(,%eax,4),%bx
+	movl SYM(xcolors)(,%eax,4),%ebx
 	sall $16,%ebx
 	movl SYM(ham_linebuf)(,%esi,4),%eax
 	movw SYM(xcolors)(,%eax,4),%bx
-	movl %ebx,(%ebp)
+	movl %ebx,(%ecx,%esi,2)
 
 	movl SYM(ham_linebuf)+12(,%esi,4),%eax
-	movw SYM(xcolors)(,%eax,4),%bx
+	movl SYM(xcolors)(,%eax,4),%ebx
 	sall $16,%ebx
 	movl SYM(ham_linebuf)+8(,%esi,4),%eax
 	movw SYM(xcolors)(,%eax,4),%bx
-	movl %ebx,4(%ebp)
-	addl $8,%ebp
+	movl %ebx,4(%ecx,%esi,2)
+
 	addl $4,%esi
 	jmp .Lb_hamloop
-
 .Lb_hamend:
-	movl %ebp,SYM(xlinebuffer)
-	jmp .Lb_finish_line
-	
-.Lb_noham:
-	cmpl $0,SYM(bpldualpf)
-	je .Lb_normal
-	
-	/* Dual Playfield */
-	
-	xorl %edx,%edx
-	movl $SYM(dblpf_ind1),%ebp
-	cmpl $0,SYM(bpldualpfpri)
-	je .Lb_dp_loop
-	movl $SYM(dblpf_ind2),%ebp
-.Lb_dp_loop:
-	cmpl %esi,SYM(diwlastword)
-	jle .Lb_finish_line
-	
-	cmpl %edi,%esi
-	jge .Lb_finish_line
+        ret
 
+        .globl pfield_linetoscr_dualpf_asm16
+	.type  pfield_linetoscr_dualpf_asm16,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_dualpf_asm16:
+	movl 44(%esp),%ebp /* lookup_no */
+.Lb_dp_loop:
+        movl 36(%esp),%eax /* diw_end */
 	xorl %edx,%edx
+	cmpl %esi,%eax
+	jbe .Lb_dp_done
+
 	movl SYM(pixdata)(%esi),%eax
 	movl SYM(spixstate)(%esi),%ebx
 	testl %ebx,%ebx
 	jne .Lb_dp_sprites
 	
 	/* no sprites */
-
-        movl SYM(xlinebuffer),%ebx
+	
+	movl SYM(xlinebuffer),%ebx
 
 	movb %ah,%dl
 	movb (%ebp,%edx,4),%dl
-	movw SYM(acolors)(,%edx,4),%cx
+	movl SYM(colors_for_drawing)(,%edx,4),%ecx
 	sall $16,%ecx
 	movb %al,%dl
 	movb (%ebp,%edx,4),%dl
-	movw SYM(acolors)(,%edx,4),%cx
-	
-	movl %ecx,(%ebx)
-
-	bswapl %eax
-	
-	movb %al,%dl
+	movw SYM(colors_for_drawing)(,%edx,4),%cx
+	sarl $16,%eax
+	movl %ecx,(%ebx,%esi,2)
+	movb %ah,%dl
 	movb (%ebp,%edx,4),%dl
-	movw SYM(acolors)(,%edx,4),%cx
+	movl SYM(colors_for_drawing)(,%edx,4),%ecx
+	movb %al,%dl
 	sall $16,%ecx
-	movb %ah,%dl
 	movb (%ebp,%edx,4),%dl
-	movw SYM(acolors)(,%edx,4),%cx
-	
-	movl %ecx,4(%ebx)
-	addl $8,SYM(xlinebuffer)
+	movw SYM(colors_for_drawing)(,%edx,4),%cx
+	movl %ecx,4(%ebx,%esi,2)
+
 	addl $4,%esi
 	jmp .Lb_dp_loop
 	
 .Lb_dp_sprites:
-
         /* dual playfield with sprites */
 
 	movb %ah,%dl
 	testb %bh,%bh
 	je .Lb3
-	movw SYM(acolors)(,%edx,4),%cx
+	movl SYM(colors_for_drawing)(,%edx,4),%ecx
 	jmp .Lb1
 .Lb3:
 	movb (%ebp,%edx,4),%dl
-	movw SYM(acolors)(,%edx,4),%cx
+	movl SYM(colors_for_drawing)(,%edx,4),%ecx
 .Lb1:
         sall $16,%ecx
-
 	movb %al,%dl
+	sarl $16,%eax
 	testb %bl,%bl
 	je .Lb6
-	movw SYM(acolors)(,%edx,4),%cx
+	movw SYM(colors_for_drawing)(,%edx,4),%cx
 	jmp .Lb4
 .Lb6:
 	movb (%ebp,%edx,4),%dl
-	mov SYM(acolors)(,%edx,4),%cx
+	movw SYM(colors_for_drawing)(,%edx,4),%cx
 .Lb4:
-	movl SYM(xlinebuffer),%edx
-	movl %ecx,(%edx)
-	xorl %edx,%edx
-
-	bswapl %eax
-	bswapl %ebx
+	movl SYM(xlinebuffer),%edi
+	sarl $16,%ebx
+	movl %ecx,(%edi,%esi,2)
 	
-	movb %al,%dl
-	testb %bl,%bl
+	movb %ah,%dl
+	testb %bh,%bh
 	je .Lb9
-	movw SYM(acolors)(,%edx,4),%cx
+	movl SYM(colors_for_drawing)(,%edx,4),%ecx
 	jmp .Lb7
 .Lb9:
 	movb (%ebp,%edx,4),%dl
-	movw SYM(acolors)(,%edx,4),%cx
+	movl SYM(colors_for_drawing)(,%edx,4),%ecx
 .Lb7:
         sall $16,%ecx
-
-	movb %ah,%dl
-	testb %bh,%bh
+	movb %al,%dl
+	testb %bl,%bl
 	je .Lb12
-	movw SYM(acolors)(,%edx,4),%cx
+	movw SYM(colors_for_drawing)(,%edx,4),%cx
 	jmp .Lb10
 .Lb12:
 	movb (%ebp,%edx,4),%dl
-	movw SYM(acolors)(,%edx,4),%cx
+	movw SYM(colors_for_drawing)(,%edx,4),%cx
 .Lb10:
-	addl $4,%esi
-	
 	movl SYM(xlinebuffer),%eax
-	movl %ecx,4(%eax)
-	addl $8,SYM(xlinebuffer)
-	
+	movl %ecx,4(%eax,%esi,2)
+	addl $4,%esi
 	jmp .Lb_dp_loop
-	
-	/* normal case */	
-.Lb_normal:
+
+.Lb_dp_done:
+	ret
+
+        .globl pfield_linetoscr_ehb_asm16
+	.type  pfield_linetoscr_ehb_asm16,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_ehb_asm16:
 	xorl %edx,%edx
-	movl SYM(xlinebuffer),%ebx
-.Lb_normal_loop:
-	cmpl %esi,SYM(diwlastword)
-	jle .Lb_normal_end
-	cmpl %edi,%esi
-	jge .Lb_normal_end
+	xorl %ebx,%ebx
+.Lb_ehb_loop:
+	movl 36(%esp),%edi /* diw_end */
+	cmpl %esi,%edi
+	jbe .Lb_ehb_end
 	
 	movl SYM(pixdata)(%esi),%eax
-	
-	movb %ah,%dl
-	movw SYM(acolors)(,%edx,4),%cx
+	movb %ah,%bl
+	cmpb $32,%ah
+	movl SYM(colors_for_drawing)(,%ebx,4),%ecx
+	jc   .Lb_ehb2
+	movw SYM(colors_for_drawing)+32*4-32*2(,%ebx,2),%di
+	shrl $1,%edi
+	andl $0x777,%edi
+	movl SYM(xcolors)(,%edi,4),%ecx
+.Lb_ehb2:
 	sall $16,%ecx
 	movb %al,%dl
-	movw SYM(acolors)(,%edx,4),%cx
-	movl %ecx,(%ebx)
-	bswapl %eax
-	movb %al,%dl
-	movw SYM(acolors)(,%edx,4),%cx
+	movw SYM(colors_for_drawing)(,%edx,4),%cx
+	cmpb $32,%al
+	jc   .Lb_ehb1
+
+	movw SYM(colors_for_drawing)+32*4-32*2(,%edx,2),%di
+	shrl $1,%edi
+	andl $0x777,%edi
+	movw SYM(xcolors)(,%edi,4),%cx
+.Lb_ehb1:
+	movl SYM(xlinebuffer),%edi
+	sarl $16,%eax
+	movl %ecx,(%edi,%esi,2)
+	movb %ah,%bl
+	movl SYM(colors_for_drawing)(,%ebx,4),%ecx
+	cmpb $32,%ah
+	jc   .Lb_ehb4
+
+	movw SYM(colors_for_drawing)+32*4-32*2(,%ebx,2),%di
+	shrl $1,%edi
+	andl $0x777,%edi
+	movl SYM(xcolors)(,%edi,4),%ecx
+.Lb_ehb4:
 	sall $16,%ecx
-	movb %ah,%dl
-	movw SYM(acolors)(,%edx,4),%cx
-	movl %ecx,4(%ebx)
-	
-	addl $8,%ebx
+	movb %al,%dl
+	movw SYM(colors_for_drawing)(,%edx,4),%cx
+	cmpb $32,%al
+	jc   .Lb_ehb3
+
+	movw SYM(colors_for_drawing)+32*4-32*2(,%edx,2),%di
+	shrl $1,%edi
+	andl $0x777,%edi
+	movw SYM(xcolors)(,%edi,4),%cx
+.Lb_ehb3:
+
+	movl SYM(xlinebuffer),%edi
+	movl %ecx,4(%edi,%esi,2)
 	addl $4,%esi
-	jmp .Lb_normal_loop
+	jmp .Lb_ehb_loop
 
-.Lb_normal_end:
-	movl %ebx,SYM(xlinebuffer)
+.Lb_ehb_end:
+        ret
 
-.Lb_finish_line:
-	/* Finish off the line */
-	
-	movw SYM(acolors),%edx
-	movw %dx,%ax
+	.globl pfield_linetoscr_normal_asm16
+	.type  pfield_linetoscr_normal_asm16,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_normal_asm16:
+	movl 36(%esp),%edi /* diw_end */
+	jmp .Lb_normal_doloop
+
+	FUNCTION_ALIGN
+.Lb_normal_loop:
+	xorl %edx,%edx	
+	BYTE_MOVE SYM(pixdata)+1(%esi),BYTE_REG(%edx,%dl)
+	xorl %ebx,%ebx
+	BYTE_MOVE SYM(pixdata)+5(%esi),BYTE_REG(%ebx,%bl)
+
+	movl SYM(colors_for_drawing)(,%edx,4),%edi
+	movl SYM(colors_for_drawing)(,%ebx,4),%ecx
+	sall $16,%edi
+	BYTE_MOVE SYM(pixdata)+0(%esi),BYTE_REG(%edx,%dl)
+	sall $16,%ecx
+	BYTE_MOVE SYM(pixdata)+4(%esi),BYTE_REG(%ebx,%bl)
+	movw SYM(colors_for_drawing)(,%edx,4),%di
+	movw SYM(colors_for_drawing)(,%ebx,4),%cx
+	BYTE_MOVE SYM(pixdata)+3(%esi),BYTE_REG(%edx,%dl)
+	BYTE_MOVE SYM(pixdata)+7(%esi),BYTE_REG(%ebx,%bl)
+	movl SYM(colors_for_drawing)(,%edx,4),%eax
+	movl SYM(colors_for_drawing)(,%ebx,4),%ebp
 	sall $16,%eax
-	movw %dx,%ax
-	
-	movl SYM(xlinebuffer),%ecx
-.Lb_fin_loop:
-	cmpl %edi,%esi
-	jge .Lb_fin_end
-	
-	movl %eax,(%ecx)
-	movl %eax,4(%ecx)
-	addl $8,%ecx
+	BYTE_MOVE SYM(pixdata)+2(%esi),BYTE_REG(%edx,%dl)
+	sall $16,%ebp
+	BYTE_MOVE SYM(pixdata)+6(%esi),BYTE_REG(%ebx,%bl)
+	movw SYM(colors_for_drawing)(,%edx,4),%ax
 	addl $4,%esi
-	jmp  .Lb_fin_loop
+	movw SYM(colors_for_drawing)(,%ebx,4),%bp
 
-.Lb_fin_end:
-        movl %ecx,SYM(xlinebuffer)
-	
-	popl %ebx
-	popl %esi
-	popl %edi
-	popl %ebp
-	ret
+	movl SYM(xlinebuffer),%ebx
+	movl %edi,-8(%ebx,%esi,2)
+	movl 36(%esp),%edi
+	movl %eax,-4(%ebx,%esi,2)
+
+	cmpl %esi,%edi
+	jbe .Lb_normal_done
+
+	movl %ecx,(%ebx,%esi,2)
+	movl %ebp,4(%ebx,%esi,2)
+
+	addl $4,%esi
+.Lb_normal_doloop:
+	cmpl %esi,%edi
+	jnbe .Lb_normal_loop
+
+.Lb_normal_done:
+        ret
+
+	.globl pfield_linetoscr_hdouble_asm16
+	.type  pfield_linetoscr_hdouble_asm16,@function
+	FUNCTION_ALIGN
+pfield_linetoscr_hdouble_asm16:
+	shrl $1,%esi
+	movl 36(%esp),%eax /* diw_end */
+	shrl $1,%eax
+	movl SYM(xlinebuffer),%ebp
+	jmp .Lb_hdouble_doloop
+
+	FUNCTION_ALIGN
+.Lb_hdouble_loop:
+	xorl %edx,%edx	
+	BYTE_MOVE SYM(pixdata)(%esi),BYTE_REG(%edx,%dl)
+	xorl %ebx,%ebx
+	BYTE_MOVE SYM(pixdata)+1(%esi),BYTE_REG(%ebx,%bl)
 
-        .ident	"Crux CC 3.14.15"
+	movl SYM(colors_for_drawing)(,%edx,4),%edi
+	movl SYM(colors_for_drawing)(,%ebx,4),%ecx
+	movl %edi,%edx
+	movl %ecx,%ebx
+	sall $16,%edi
+	andl $0xFFFF,%edx
+	sall $16,%ecx
+	andl $0xFFFF,%ebx
+	addl $2,%esi
+	orl %edx,%edi
+	orl %ebx,%ecx
+	movl %edi,-8(%ebp,%esi,4)
+	movl %ecx,-4(%ebp,%esi,4)
+
+.Lb_hdouble_doloop:
+	cmpl %esi,%eax
+	jnbe .Lb_hdouble_loop
+
+.Lb_hdouble_done:
+        addl %esi,%esi
+        ret
 
-#ifndef __DOS__
-	.align 16
-.globl DitherLine
+        .globl DitherLine
+#ifndef USE_UNDERSCORE
 	.type	 DitherLine,@function
-#else
-	.align 8
-.globl DitherLine
 #endif
+	FUNCTION_ALIGN
 DitherLine:
 	pushl %ebp
 	pushl %edi
@@ -608,25 +765,22 @@ DitherLine:
 .Li_fast:
 	movl 24(%esp),%esi
 	xorl %edx,%edx
+	xorl %ecx,%ecx
+	FUNCTION_ALIGN
 .Li_fast_loop:
-	movl (%esi),%ecx
-	movw %cx,%dx
-	movb (%edx,%ebp),%al
-	
-	sarl $16,%ecx
-	movw %cx,%dx
-	movb 4096(%ebp,%edx),%ah
-	bswapl %eax
+	movw (%esi),%edx
+	movw 2(%esi),%cx
+	movb (%edx,%ebp),%al	
+	movw 4(%esi),%dx
+	movb 4096(%ebp,%ecx),%ah
+	
+	movw 6(%esi),%cx
+	sall $16,%eax
+	movb 8192(%ebp,%edx),%al
 	
-	movl 4(%esi),%ecx
-	movw %cx,%dx
-	movb 8192(%ebp,%edx),%ah
-	
-	sarl $16,%ecx
-	movw %cx,%dx
-	movb 12288(%ebp,%edx),%al
+	movb 12288(%ebp,%ecx),%ah
 	
-	bswapl %eax
+	roll $16,%eax
 	movl %eax,(%edi)
 	addl $4,%edi
 	addl $8,%esi
@@ -640,3 +794,98 @@ DitherLine:
 	popl %edi
 	popl %ebp
 	ret
+#if 0	
+.globl compiler_do_rts
+        /* Entry: EDX == regs.regs + 15 */
+compiler_do_rts:
+        movl (%edx),%esi
+        addl address_space,%esi
+	movl jsr_num,%ecx
+	movl (%esi),%esi
+	orl %ecx,%ecx
+	bswapl %esi
+	jz cdrts_noway
+	decl %ecx
+	cmpl jsr_rets(,%ecx,4),%esi
+	jne cdrts_noway
+	movl jsr_hash(,%ecx,4),%ebx
+	movl (%ebx),%ebx
+	orl %ebx,%ebx
+	jz cdrts_noway
+	addl $4,(%edx)
+	movl %ecx,jsr_num
+	jmp %ebx
+cdrts_noway:
+        ret
+#endif
+#if 0
+.globl longget_stub
+longget_stub:
+        pushl %ecx
+        pushl %eax
+	pushl %edx
+	call SYM(longget)
+	popl %ecx
+	movl %eax,%ecx
+	popl %eax
+        ret
+
+.globl wordget_stub
+wordget_stub:
+        pushl %ecx
+        pushl %eax
+	pushl %edx
+	call SYM(wordget)
+	popl %ecx
+	movl %eax,%ecx
+	popl %eax
+        ret
+
+.globl byteget_stub
+byteget_stub:
+        pushl %ecx
+        pushl %eax
+	pushl %edx
+	call SYM(byteget)
+	popl %ecx
+	movl %eax,%ecx
+	popl %eax
+        ret
+
+.globl longput_stub
+longput_stub:
+        pushl %ecx
+        pushl %eax
+	pushl %ebx
+	pushl %edx
+	call SYM(longput)
+	movl %eax,%ecx
+	addl $8,%esp
+	popl %eax
+        ret
+
+.globl wordput_stub
+wordput_stub:
+        pushl %ecx
+        pushl %eax
+	pushl %ebx
+	pushl %edx
+	call SYM(wordput)
+	movl %eax,%ecx
+	addl $8,%esp
+	popl %eax
+        ret
+
+.globl byteput_stub
+byteput_stub:
+        pushl %ecx
+        pushl %eax
+	pushl %ebx
+	pushl %edx
+	call SYM(byteput)
+	movl %eax,%ecx
+	addl $8,%esp
+	popl %eax
+        ret
+#endif
+	.ident	"Crux CC 3.14.15"