diff --git a/src/arch/i386/prefix/pxeprefix.S b/src/arch/i386/prefix/pxeprefix.S
index 0e0b6c00..dea0b199 100644
--- a/src/arch/i386/prefix/pxeprefix.S
+++ b/src/arch/i386/prefix/pxeprefix.S
@@ -1,16 +1,9 @@
-/* Offsets of words containing ROM's CS and size (in 512 byte blocks)
- * from start of floppy boot block at 0x7c00
- * Offsets must match those in etherboot.h
- */
-#define FLOPPY_SEGMENT	0x7c0
-
 #define PXENV_UNDI_CLEANUP	0x02
 #define PXENV_UNDI_SHUTDOWN	0x05
 #define	PXENV_STOP_UNDI		0x15
 #define PXENV_UNLOAD_STACK	0x70
 #define PXENV_STOP_BASE		0x76
 
-#define PUSHA_SIZE		16
 #define PXE_STACK_MAGIC		0x57ac	/* 'STac' */
 
 	.text
@@ -22,21 +15,35 @@
  * Entry point:	set cs, ds, bp, print welcome message
  *****************************************************************************
  */	
-_prefix:
-	jmp	$FLOPPY_SEGMENT, $code_start-_prefix
+	jmp	$0x7c0, $code_start
 10:	.asciz	"PXE->EB "
 code_start:
-	pusha				/* Preserve all registers */
-	push	%ds
-	movw	%sp, %bp		/* %bp must be preserved, hence do
-					 * this after the pusha */
-	push	$PXE_STACK_MAGIC	/* PXE stack magic marker */
-
-	push	%cs			/* Set up data segment */
-	pop	%ds
-	mov	$0x40, %cx		/* Set up %fs for access to 40:13 */
-	mov	%cx, %fs
-	movw	$10b-_prefix, %si	/* Print welcome message */
+	/* Preserve registers for return to PXE stack */
+	pushfl
+	pushal
+	pushw	%gs
+	pushw	%fs
+	pushw	%es
+	pushw	%ds
+	pushw	%ss
+	pushw	%cs
+	pushw	$PXE_STACK_MAGIC	/* PXE stack magic marker */
+	/* Set up stack just below 0x7c00 */
+	pushw	%ss
+	popw	%es
+	movw	%sp, %di
+	xorw	%ax, %ax
+	movw	%ax, %ss
+	movw	$0x7c00, %sp
+	pushw	%es			/* Save old PXE stack pointer */
+	pushw	%di
+	/* Set up our other segment registers */
+	pushw	%cs
+	popw	%ds
+	movw	$0x40, %ax		/* BIOS data segment access */
+	movw	%ax, %fs
+	/* Print welcome message */
+	movw	$10b, %si
 	call	print_message
 
 /*****************************************************************************
@@ -44,12 +51,12 @@ code_start:
  *****************************************************************************
  */
 detect_pxe:
-	les	4+PUSHA_SIZE+2(%bp), %di /* !PXE structure */
+	les	%es:54(%di), %di /* !PXE structure */
 	cmpl	$0x45585021, %es:(%di)	/* '!PXE' signature */
 	je	detected_pxe
-	mov	$0x5650, %ax
+	movw	$0x5650, %ax
 	int	$0x1a
-	cmp	$0x564e, %ax
+	cmpw	$0x564e, %ax
 	jne	detected_nothing
 	cmpl	$0x4e455850, %es:(%bx)	/* 'PXEN' signature */
 	jne	detected_nothing
@@ -57,49 +64,45 @@ detect_pxe:
 	je	detected_pxenv
 
 detected_nothing:
-	movw	$10f-_prefix, %si
+	movw	$10f, %si
 	call	print_message
 	jmp	finished_with_error
 10:	.asciz	"No PXE "
 
 detected_pxenv: /* es:bx points to PXENV+ structure */
-	push	%es
-	push	%bx
-	push	%es:0x24(%bx)		/* UNDI code segment */
-	push	%es:0x26(%bx)		/* UNDI code size */
-	push	%es:0x20(%bx)		/* UNDI data segment */
-	push	%es:0x22(%bx)		/* UNDI data size */
+	pushw	%es
+	pushw	%bx
+	pushw	%es:0x24(%bx)		/* UNDI code segment */
+	pushw	%es:0x26(%bx)		/* UNDI code size */
+	pushw	%es:0x20(%bx)		/* UNDI data segment */
+	pushw	%es:0x22(%bx)		/* UNDI data size */
 	les	%es:0x0a(%bx), %di	/* Entry point to %es:%di */
-	movw	$10f-_prefix, %si
+	movw	$10f, %si
 	jmp	pxe_setup_done
 10:	.asciz	"PXENV+ "
 
 detected_pxe:	/* es:di points to !PXE structure */
-	push	%es
-	push	%di
-	push	%es:0x30(%di)		/* UNDI code segment */
-	push	%es:0x36(%di)		/* UNDI code size */
-	push	%es:0x28(%di)		/* UNDI data segment */
-	push	%es:0x2e(%di)		/* UNDI data size */
+	pushw	%es
+	pushw	%di
+	pushw	%es:0x30(%di)		/* UNDI code segment */
+	pushw	%es:0x36(%di)		/* UNDI code size */
+	pushw	%es:0x28(%di)		/* UNDI data segment */
+	pushw	%es:0x2e(%di)		/* UNDI data size */
 	les	%es:0x10(%di), %di	/* Entry point to %es:%di */
-	movw	$10f-_prefix, %si
+	movw	$10f, %si
 	jmp	pxe_setup_done
 10:	.asciz	"!PXE "
 
 pxe_setup_done:
-	mov	%es, pxe_entry_segment - _prefix
-	mov	%di, pxe_entry_offset - _prefix
-	pop	%ax
-	mov	%ax, undi_data_size - _prefix
-	pop	%ax
-	mov	%ax, undi_data_segment - _prefix
-	pop	%ax
-	mov	%ax, undi_code_size - _prefix
-	pop	%ax
-	mov	%ax, undi_code_segment - _prefix
+	movw	%es, pxe_entry_segment
+	movw	%di, pxe_entry_offset
+	popw	undi_data_size
+	popw	undi_data_segment
+	popw	undi_code_size
+	popw	undi_code_segment
 	call	print_message
-	pop	%di
-	pop	%es	/* Exit with %es:%di containing structure address */
+	popw	%di
+	popw	%es	/* Exit with %es:%di containing structure address */
 
 /*****************************************************************************
  * Print information about located structure
@@ -107,11 +110,11 @@ pxe_setup_done:
  */
 print_structure_information:
 	call	print_segoff	/* %es:%di contains address of structure */
-	les	%ds:(pxe_entry_segoff - _prefix), %di
+	les	pxe_entry_segoff, %di
 	call	print_segoff
-	les	%ds:(undi_code_segoff - _prefix), %di
+	les	undi_code_segoff, %di
 	call	print_segoff
-	les	%ds:(undi_data_segoff - _prefix), %di
+	les	undi_data_segoff, %di
 	call	print_segoff
 
 /*****************************************************************************
@@ -119,17 +122,17 @@ print_structure_information:
  *****************************************************************************
  */
 #ifdef PXELOADER_KEEP_ALL
-	xor	%ax, %ax		/* Force zero flag to show success */
+	xorw	%ax, %ax		/* Force zero flag to show success */
 	jmp	do_not_free_base_mem	/* Skip the unloading */
 #endif /* PXELOADER_KEEP_ALL */
 	
 unload_pxe:
-	mov	$PXENV_UNLOAD_STACK, %bx
+	movw	$PXENV_UNLOAD_STACK, %bx
 	call	pxe_call
-	mov	$PXENV_STOP_UNDI, %bx
+	movw	$PXENV_STOP_UNDI, %bx
 	call	pxe_call
 	pushfw				/* Ignore PXENV_UNDI_CLEANUP errors */
-	mov	$PXENV_UNDI_CLEANUP, %bx
+	movw	$PXENV_UNDI_CLEANUP, %bx
 	call	pxe_call
 	popfw
 	/* On exit, zero flag is set iff all calls were successful */
@@ -141,45 +144,45 @@ unload_pxe:
 free_base_mem:
 	jnz	do_not_free_base_mem	/* Using zero flag from unload_pxe */
 
-	mov	undi_code_segment - _prefix, %bx
-	mov	undi_data_segment - _prefix, %cx
-	mov	undi_code_size - _prefix, %ax
-	cmp	%bx, %cx
+	movw	undi_code_segment, %bx
+	movw	undi_data_segment, %cx
+	movw	undi_code_size, %ax
+	cmpw	%bx, %cx
 	jb	1f
-	mov	%cx, %bx
-	mov	undi_data_size - _prefix, %ax
-1:	add	$0x0f, %ax		/* Round up to next segment */
-	shr	$4, %ax
-	add	%bx, %ax		/* Highest segment address into %ax */
-	add	$(1024 / 16 - 1), %ax	/* Round up to next kb */
-	shr	$6, %ax			/* New free basemem size in %ax */
-	mov	%fs:(0x13), %bx		/* Old free base memory in %bx */
-	mov	%ax, %fs:(0x13)		/* Store new free base memory size */
+	movw	%cx, %bx
+	movw	undi_data_size, %ax
+1:	addw	$0x0f, %ax		/* Round up to next segment */
+	shrw	$4, %ax
+	addw	%bx, %ax		/* Highest segment address into %ax */
+	addw	$(1024 / 16 - 1), %ax	/* Round up to next kb */
+	shrw	$6, %ax			/* New free basemem size in %ax */
+	movw	%fs:(0x13), %bx		/* Old free base memory in %bx */
+	movw	%ax, %fs:(0x13)		/* Store new free base memory size */
 
 	/* Note that zero_mem_loop will also zero out our stack, so make
 	 * sure the stack is empty at this point.
 	 */
-	mov	%ax, %dx
-	sub	%bx, %dx		/* numberof kb to zero in %dx */
-	shl	$6, %bx			/* Segment address into %bx */
+	movw	%ax, %dx
+	subw	%bx, %dx		/* numberof kb to zero in %dx */
+	shlw	$6, %bx			/* Segment address into %bx */
 zero_mem_loop:
-	mov	%bx, %es		/* kB boundary into %es:00 */
-	xor	%ax, %ax
-	xor	%di, %di
-	mov	$0x400, %cx
+	movw	%bx, %es		/* kB boundary into %es:00 */
+	xorw	%ax, %ax
+	xorw	%di, %di
+	movw	$0x400, %cx
 	rep	stosb			/* fill kB with zeroes */
-	add	$(1024 / 16), %bx
-	dec	%dx
+	addw	$(1024 / 16), %bx
+	decw	%dx
 	jnz	zero_mem_loop
 	/* Will exit here with zero flag set, so no need to set it explicitly
 	 * in order to indicate success.
 	 */
 	
 do_not_free_base_mem:
-	pushf				/* Save success (zero) flag status */
-	mov	%fs:(0x13), %ax		/* Free base memory in %ax */
+	pushfw				/* Save success (zero) flag status */
+	movw	%fs:(0x13), %ax		/* Free base memory in %ax */
 	call	print_hex_word		/* Print free base memory */
-	popf				/* Restore success (zero) flag */
+	popfw				/* Restore success (zero) flag */
 
 /*****************************************************************************
  * Exit point
@@ -188,92 +191,16 @@ do_not_free_base_mem:
  *****************************************************************************
  */	
 finished:
-	movw	$10f-_prefix, %si
+	movw	$10f, %si
 	jz	1f
 finished_with_error:
-	movw	$20f-_prefix, %si
+	movw	$20f, %si
 1:
 	call	print_message
-	jmp	99f
+	jmp	run_etherboot
 10:	.asciz " ok\n"
 20:	.asciz " err\n"
 
-	
-	/* We place a stack here.  It doesn't get used until after all
-	 * the above code is finished, so we can happily write all
-	 * over it.  Putting the stack here ensures that it doesn't
-	 * accidentally go over the 512 byte threshold, which would
-	 * cause problems when returning via start32's prefix
-	 * relocation mechanism.
-	 */
-_estack:	
-99:
-
-/*****************************************************************************
- * Run Etherboot main code
- *****************************************************************************
- */	
-run_etherboot:
-	/* Very temporarily switch stacks to one internal to the
-	 * prefix.  Do this because the stack provided by the PXE ROM
-	 * could be absolutely anywhere, including in an area of
-	 * memory that the call to prelocate will vapourise...
-	 */
-	pushw	%ss			/* PXE stack pointer to ES:DI */
-	popw	%es
-	movw	%sp, %di
-	pushw	%ds			/* Set up stack in "safe" area */
-	popw	%ss
-	movw	$_estack-_prefix, %sp
-	pushw	%es			/* Record PXE stack pointer */
-	pushw	%di
-	/* Relocate payload and stack to claimed base memory */
-	pushw	$4			/* Preserve old PXE stack pointer */
-	call	prelocate
-	popw	%ax			/* Remove parameter */
-	pushl	$4			/* Preserve old PXE stack pointer */
-	pushw	$0			/* Indicate prefix exit mechanism */
-	jmp	_start			/* Run Etherboot */
-
-	.section ".text16", "ax", @progbits
-prefix_exit:
-	pushw	%cs			/* Set %ds, %bp for access to text */
-	popw	%ds
-	call	1f
-1:	popw	%bp
-	popw	%di			/* Old PXE stack to %es:di */
-	popw	%es
-	cmpw	$PXE_STACK_MAGIC, %es:0(%di)  /* See if PXE stack intact */
-	jne	exit_via_int18
-exit_via_pxe:				/* Stack OK, return to PXE */
-	push	%es			/* Restore PXE stack pointer */
-	pop	%ss
-	mov	%di, %sp
-	pop	%ax			/* Discard PXE_STACK_MAGIC marker */
-	leaw	(10f-1b)(%bp), %si
-	call	print_exit_message
-	pop	%ds			/* Restore PXE's DS */
-	popa				/* Restore PXE's other registers */
-	movw	$0, %ax			/* Return PXENV_STATUS_SUCCESS */
-	lret				/* Return control to PXE ROM */
-10:	.asciz	"EB->PXE\r\n"
-exit_via_int18:				/* Stack damaged, do int 18 */
-	leaw	(10f-1b)(%bp), %si
-	call	print_exit_message
-	int	$0x18
-10:	.asciz	"EB->BIOS\r\n"
-print_exit_message:	
-	movw	$0x0007, %bx		/* page 0, attribute 7 (normal) */
-	movb	$0x0e, %ah		/* write char, tty mode */
-1: 	lodsb
-	testb	%al, %al
-	je	2f
-	int	$0x10
-	jmp	1b
-2:	ret
-prefix_exit_end:
-	.previous
-
 /*****************************************************************************
  * Subroutine: print character in %al (with LF -> LF,CR translation)
  *****************************************************************************
@@ -305,17 +232,17 @@ print_message:
  *****************************************************************************
  */
 print_hex_word:
-	mov	$4, %cx
+	movw	$4, %cx
 1:
-	push	%ax
-	shr	$12, %ax
+	pushw	%ax
+	shrw	$12, %ax
 	/* Courtesy of Norbert Juffa <norbert.juffa@amd.com> */
-	cmp	$10, %al	
-	sbb	$0x69, %al
+	cmpb	$10, %al	
+	sbbb	$0x69, %al
 	das
 	call	print_character
-	pop	%ax
-	shl	$4, %ax
+	popw	%ax
+	shlw	$4, %ax
 	loop	1b
 	ret
 	
@@ -324,15 +251,15 @@ print_hex_word:
  *****************************************************************************
  */
 print_segoff:
-	push	%di
-	push	%es
-	pop	%ax
+	pushw	%di
+	pushw	%es
+	popw	%ax
 	call	print_hex_word
 	movb	$0x3a,%al			/* ':' */
 	call	print_character
-	pop	%ax
+	popw	%ax
 	call	print_hex_word
-	mov	$0x20, %al			/* ' ' */
+	movb	$0x20, %al			/* ' ' */
 	call	print_character
 	ret
 	
@@ -346,24 +273,24 @@ print_segoff:
  */
 pxe_call:
 	/* Set up registers for PXENV+ API.  %bx already set up */
-	push	%ds
-	pop	%es
-	mov	$pxe_parameter_structure - _prefix, %di
+	pushw	%ds
+	popw	%es
+	movw	$pxe_parameter_structure, %di
 	/* Set up stack for !PXE API */
 	pushw   %cs
 	pushw	%di
 	pushw	%bx
 	/* Make the API call */
-	lcall	*(pxe_entry_segoff - _prefix)
+	lcall	*pxe_entry_segoff
 	/* Reset the stack */
-	add	$6, %sp
-	mov	pxe_parameter_structure - _prefix, %ax
-	push	%ax
+	addw	$6, %sp
+	movw	pxe_parameter_structure, %ax
+	pushw	%ax
 	call	print_hex_word
-	mov	$0x20, %ax		/* ' ' */
+	movw	$0x20, %ax		/* ' ' */
 	call	print_character
-	pop	%bx
-	or	%bx, pxe_overall_status - _prefix
+	popw	%bx
+	orw	%bx, pxe_overall_status
 	ret
 
 /*****************************************************************************
@@ -389,7 +316,73 @@ pxe_parameter_structure:
 	.word	0
 	.word	0,0,0,0,0
 
-end_of_pxeloader:
+/*****************************************************************************
+ * Run Etherboot main code
+ *****************************************************************************
+ */	
+run_etherboot:
+	/* Install Etherboot */
+	call	install
 
-	.balign 16, 0
-payload:
+	/* Jump to .text16 segment with %ds pointing to .data16*/
+	movw	%bx, %ds
+	pushw	%ax
+	pushw	$1f
+	lret
+	.section ".text16", "ax", @progbits
+1:
+	/* Original PXE stack pointer to es:di.  We must hold it in
+	 * registers, because our current stack may be vapourised by
+	 * the time main() returns.  (main() will still be able to
+	 * return, because prot_call() transfers the return address to
+	 * the internal stack and back again).
+	 */
+	popw	%di
+	popw	%es
+
+	/* Run main program */
+	pushl	$main
+	pushw	%cs
+	call	prot_call
+	popl	%eax /* discard */
+
+	/* If original PXE stack is intact, return via PXE, else via INT 18 */
+	cmpw	$PXE_STACK_MAGIC, %es:0(%di)
+	jne	exit_via_int18
+exit_via_pxe:				/* Stack OK, return to PXE */
+	movw	$exit_via_pxe_message, %si
+	call	print_exit_message
+	pushw	%es			/* Restore original PXE stack */
+	popw	%ss
+	movw	%di, %sp
+	popw	%ax /* discard PXE_STACK_MAGIC */
+	popw	%ax /* discard %cs */
+	popw	%ax /* discard %ss */
+	popw	%ds
+	popw	%es
+	popw	%fs
+	popw	%gs
+	popal
+	popfl
+	xorw	%ax, %ax		/* Return PXENV_STATUS_SUCCESS */
+	lret
+exit_via_int18:				/* Stack damaged, do int 18 */
+	movw	$exit_via_int18_message, %si
+	call	print_exit_message
+	int	$0x18
+
+print_exit_message:	
+	movw	$0x0007, %bx		/* page 0, attribute 7 (normal) */
+	movb	$0x0e, %ah		/* write char, tty mode */
+1: 	lodsb
+	testb	%al, %al
+	je	2f
+	int	$0x10
+	jmp	1b
+2:	ret
+
+	.section ".data16", "aw", @progbits
+exit_via_pxe_message:
+	.asciz	"EB->PXE\r\n"
+exit_via_int18_message:
+	.asciz	"EB->BIOS\r\n"