stage0: fix some bugs

src/boot/stage0.s

@@ -159,6 +159,8 @@ GLOBL _start
 	rep stosb
 	mov	%di, %sp	/* %sp = 0x7800:8000 = 0x80000 */
 
+	sti			/* we're safe again, i think */
+
 	/* say hello */
 	mov	$msg_loader_info, %si
 	call	print
@@ -336,52 +338,67 @@ LOCAL start2
 	and	%ebx, %edi	/* %edi = 0 (crc32 returns the inverted value) */
 	jnz	err_bad_gpt_csum
 
-	/* search the partition table for our GUID */
+	/*
+	 * Step 5: Search the GPT for stage1
+	 */
+
 	pop	%bx			/* restore entry size */				/* '1 ^ */
-	sub	$(bussy_guid_end - bussy_guid), %bx /* entry size -= sizeof(bussy_guid) */
-	/* this loop uses %ax for temporarily storing `entry count` and
-	 * `sizeof(bussy_guid)` because xchg with %ax is only one byte */
+	xor	%si, %si
 1:	push	%cx			/* save remaining entry count */			/* '1 v */
-	mov	$bussy_guid, %si
+	push	%si			/* save current position */				/* '2 v */
+	mov	$bussy_guid, %di
 	mov	$(bussy_guid_end - bussy_guid), %cx
 	repe cmpsb %es:(%di), %fs:(%si)
+	pop	%si			/* restore current position */				/* '2 ^ */
 	pop	%cx			/* restore remaining entry count */			/* '1 ^ */
 	je	2f			/* found it */
-	add	%bx, %di		/* %di += entry size */
+	add	%bx, %si		/* %si += entry size */
 	loopne	1b			/* only loop if %di didn't wrap around */
 	jmp	err_no_stage1
 
-	/* gotcha! now load the first 64 K from the partition */
+	/*
+	 * Step 6: Load stage1 to the beginning of low memory (0x0500)
+	 */
+
 2:	push	%dx			/* save drive number */					/* '1 v */
-	xor	%ax, %ax
-	mov	$1, %dl			/* %dh is known to be 0x00 */
-	divw	(drive_params + 0x18)	/* %ax = 0x10000 / sector size = sector count */
+	mov	$0x8000, %ax
+	xor	%dx, %dx
+	divw	(drive_params + 0x18)	/* %ax = 0x8000 / sector size = sector count */
 	xchg	%ax, %cx		/* %cx = sector count (xchg saves 1 byte) */
-	mov	%fs:0x20(%di), %eax	/* %eax = LBA[31:0] */
-	mov	%fs:0x24(%di), %bx	/* %bx = LBA[47:32] */
+	mov	%fs:0x20(%si), %eax	/* %eax = LBA[31:0] */
+	mov	%fs:0x24(%si), %bx	/* %bx = LBA[47:32] */
 	xor	%di, %di
 	pop	%dx			/* restore drive number */				/* '1 ^ */
 	call	read_lba
 
-	/* check the stage1 header */
-	lods	%fs:(%di), %ax		/* offset 0x00: magic number */
-	cmp	$0xacab, %ax
+	/* these values are for stage1 */
+	push	%eax			/* boot partition LBA[31:0] */				/* '1 v */
+	push	%bx			/* boot partition LBA[47:32] */				/* '2 v */
+	push	%dx			/* boot drive number */					/* '3 v */
+
+	/*
+	 * Step 7: Check stage1's magic number and CRC
+	 */
+
+	xor	%si, %si
+	cmp	$0xacab, %fs:(%si)	/* offset 0x00: magic number */
 	jne	err_bad_stage1_magic
-	lods	%fs:(%di), %ax		/* offset 0x02: byte count for CRC */
-	xchg	%cx, %ax
+	mov	%fs:0x02(%si), %cx	/* offset 0x02: byte count for CRC */
 	xor	%eax, %eax
-	xchg	%fs:(%di), %eax		/* offset 0x04: CRC (replace with 0) */
-	push	%ax			/* save CRC[15:0] (crc32 clobbers %al) */		/* '1 v */
+	xchg	%fs:0x04(%si), %eax	/* offset 0x04: CRC (replace with 0) */
+	push	%ax			/* save CRC[15:0] (crc32 clobbers %al) */		/* '4 v */
 	call	crc32
-	pop	%ax			/* restore CRC[15:0] */					/* '1 ^ */
-	and	%ebx, %eax		/* check CRC and clear %eax */
+	pop	%ax			/* restore CRC[15:0] */					/* '4 ^ */
+	xor	%ebx, %eax		/* check CRC and clear %eax */
 	jnz	err_bad_stage1_csum
 
 	/*
 	 * Step 8: Jump to stage1 (finally)
+	 *
+	 * The three remaining items on the stack are for stage1
 	 */
 
-	ljmp	$0x0000, $0x0508	/* entry point is right after the 8-byte header */
+	ljmp	$0x0000, $0x0510	/* entry point is right after the 16-byte header */
 END start2
 
 /*