diff --git a/asm/nonmatchings/code_dba20_len_350/get_area_byte.s b/asm/nonmatchings/code_dba20_len_350/get_area_byte.s
deleted file mode 100644
index e0546bd6eb..0000000000
--- a/asm/nonmatchings/code_dba20_len_350/get_area_byte.s
+++ /dev/null
@@ -1,16 +0,0 @@
-.set noat # allow manual use of $at
-.set noreorder # don't insert nops after branches
-
-
-glabel get_area_byte
-/* 0DBD50 80145650 3C02800E */  lui   $v0, 0x800e
-/* 0DBD54 80145654 00441021 */  addu  $v0, $v0, $a0
-/* 0DBD58 80145658 8042BF90 */  lb    $v0, -0x4070($v0)
-/* 0DBD5C 8014565C 03E00008 */  jr    $ra
-/* 0DBD60 80145660 00000000 */   nop   
-
-/* 0DBD64 80145664 00000000 */  nop   
-/* 0DBD68 80145668 00000000 */  nop   
-/* 0DBD6C 8014566C 00000000 */  nop   
-
-
diff --git a/asm/nonmatchings/code_dba20_len_350/get_area_flag.s b/asm/nonmatchings/code_dba20_len_350/get_area_flag.s
deleted file mode 100644
index cb334c5c93..0000000000
--- a/asm/nonmatchings/code_dba20_len_350/get_area_flag.s
+++ /dev/null
@@ -1,26 +0,0 @@
-.set noat # allow manual use of $at
-.set noreorder # don't insert nops after branches
-
-
-glabel get_area_flag
-/* 0DBCF0 801455F0 0080102D */  daddu $v0, $a0, $zero
-/* 0DBCF4 801455F4 04410002 */  bgez  $v0, .L80145600
-/* 0DBCF8 801455F8 0040182D */   daddu $v1, $v0, $zero
-/* 0DBCFC 801455FC 2443001F */  addiu $v1, $v0, 0x1f
-.L80145600:
-/* 0DBD00 80145600 00031943 */  sra   $v1, $v1, 5
-/* 0DBD04 80145604 00032140 */  sll   $a0, $v1, 5
-/* 0DBD08 80145608 00442023 */  subu  $a0, $v0, $a0
-/* 0DBD0C 8014560C 00031880 */  sll   $v1, $v1, 2
-/* 0DBD10 80145610 24020001 */  addiu $v0, $zero, 1
-/* 0DBD14 80145614 3C01800E */  lui   $at, 0x800e
-/* 0DBD18 80145618 00230821 */  addu  $at, $at, $v1
-/* 0DBD1C 8014561C 8C23BF70 */  lw    $v1, -0x4090($at)
-/* 0DBD20 80145620 00821004 */  sllv  $v0, $v0, $a0
-/* 0DBD24 80145624 00621024 */  and   $v0, $v1, $v0
-/* 0DBD28 80145628 54400001 */  bnel  $v0, $zero, .L80145630
-/* 0DBD2C 8014562C 24020001 */   addiu $v0, $zero, 1
-.L80145630:
-/* 0DBD30 80145630 03E00008 */  jr    $ra
-/* 0DBD34 80145634 00000000 */   nop   
-
diff --git a/asm/nonmatchings/code_dba20_len_350/get_global_byte.s b/asm/nonmatchings/code_dba20_len_350/get_global_byte.s
deleted file mode 100644
index 83900ec067..0000000000
--- a/asm/nonmatchings/code_dba20_len_350/get_global_byte.s
+++ /dev/null
@@ -1,11 +0,0 @@
-.set noat # allow manual use of $at
-.set noreorder # don't insert nops after branches
-
-
-glabel get_global_byte
-/* 0DBC38 80145538 3C02800E */  lui   $v0, 0x800e
-/* 0DBC3C 8014553C 00441021 */  addu  $v0, $v0, $a0
-/* 0DBC40 80145540 8042BD70 */  lb    $v0, -0x4290($v0)
-/* 0DBC44 80145544 03E00008 */  jr    $ra
-/* 0DBC48 80145548 00000000 */   nop   
-
diff --git a/asm/nonmatchings/code_dba20_len_350/set_area_byte.s b/asm/nonmatchings/code_dba20_len_350/set_area_byte.s
deleted file mode 100644
index 0d53dc2238..0000000000
--- a/asm/nonmatchings/code_dba20_len_350/set_area_byte.s
+++ /dev/null
@@ -1,12 +0,0 @@
-.set noat # allow manual use of $at
-.set noreorder # don't insert nops after branches
-
-
-glabel set_area_byte
-/* 0DBD38 80145638 3C02800E */  lui   $v0, 0x800e
-/* 0DBD3C 8014563C 2442ACC0 */  addiu $v0, $v0, -0x5340
-/* 0DBD40 80145640 00822021 */  addu  $a0, $a0, $v0
-/* 0DBD44 80145644 808212D0 */  lb    $v0, 0x12d0($a0)
-/* 0DBD48 80145648 03E00008 */  jr    $ra
-/* 0DBD4C 8014564C A08512D0 */   sb    $a1, 0x12d0($a0)
-
diff --git a/asm/nonmatchings/code_dba20_len_350/set_global_byte.s b/asm/nonmatchings/code_dba20_len_350/set_global_byte.s
deleted file mode 100644
index 5c8a96b50c..0000000000
--- a/asm/nonmatchings/code_dba20_len_350/set_global_byte.s
+++ /dev/null
@@ -1,12 +0,0 @@
-.set noat # allow manual use of $at
-.set noreorder # don't insert nops after branches
-
-
-glabel set_global_byte
-/* 0DBC20 80145520 3C02800E */  lui   $v0, 0x800e
-/* 0DBC24 80145524 2442ACC0 */  addiu $v0, $v0, -0x5340
-/* 0DBC28 80145528 00822021 */  addu  $a0, $a0, $v0
-/* 0DBC2C 8014552C 808210B0 */  lb    $v0, 0x10b0($a0)
-/* 0DBC30 80145530 03E00008 */  jr    $ra
-/* 0DBC34 80145534 A08510B0 */   sb    $a1, 0x10b0($a0)
-
diff --git a/asm/nonmatchings/code_e92d0_len_5da0/EnableModel.s b/asm/nonmatchings/code_e92d0_len_5da0/EnableModel.s
deleted file mode 100644
index 359a225eed..0000000000
--- a/asm/nonmatchings/code_e92d0_len_5da0/EnableModel.s
+++ /dev/null
@@ -1,41 +0,0 @@
-.set noat # allow manual use of $at
-.set noreorder # don't insert nops after branches
-
-
-glabel EnableModel
-/* 0EDC38 802C9288 27BDFFE0 */  addiu $sp, $sp, -0x20
-/* 0EDC3C 802C928C AFB10014 */  sw    $s1, 0x14($sp)
-/* 0EDC40 802C9290 0080882D */  daddu $s1, $a0, $zero
-/* 0EDC44 802C9294 AFBF0018 */  sw    $ra, 0x18($sp)
-/* 0EDC48 802C9298 AFB00010 */  sw    $s0, 0x10($sp)
-/* 0EDC4C 802C929C 8E30000C */  lw    $s0, 0xc($s1)
-/* 0EDC50 802C92A0 8E050000 */  lw    $a1, ($s0)
-/* 0EDC54 802C92A4 0C0B1EAF */  jal   get_variable
-/* 0EDC58 802C92A8 26100004 */   addiu $s0, $s0, 4
-/* 0EDC5C 802C92AC 0C046C04 */  jal   get_model_list_index_from_tree_index
-/* 0EDC60 802C92B0 0040202D */   daddu $a0, $v0, $zero
-/* 0EDC64 802C92B4 0220202D */  daddu $a0, $s1, $zero
-/* 0EDC68 802C92B8 8E050000 */  lw    $a1, ($s0)
-/* 0EDC6C 802C92BC 0C0B1EAF */  jal   get_variable
-/* 0EDC70 802C92C0 0040802D */   daddu $s0, $v0, $zero
-/* 0EDC74 802C92C4 0200202D */  daddu $a0, $s0, $zero
-/* 0EDC78 802C92C8 0C046B4C */  jal   get_model_from_list_index
-/* 0EDC7C 802C92CC 0040802D */   daddu $s0, $v0, $zero
-/* 0EDC80 802C92D0 12000004 */  beqz  $s0, .L802C92E4
-/* 0EDC84 802C92D4 0040182D */   daddu $v1, $v0, $zero
-/* 0EDC88 802C92D8 94620000 */  lhu   $v0, ($v1)
-/* 0EDC8C 802C92DC 080B24BB */  j     .L802C92EC
-/* 0EDC90 802C92E0 3042FFFD */   andi  $v0, $v0, 0xfffd
-
-.L802C92E4:
-/* 0EDC94 802C92E4 94620000 */  lhu   $v0, ($v1)
-/* 0EDC98 802C92E8 34420002 */  ori   $v0, $v0, 2
-.L802C92EC:
-/* 0EDC9C 802C92EC A4620000 */  sh    $v0, ($v1)
-/* 0EDCA0 802C92F0 8FBF0018 */  lw    $ra, 0x18($sp)
-/* 0EDCA4 802C92F4 8FB10014 */  lw    $s1, 0x14($sp)
-/* 0EDCA8 802C92F8 8FB00010 */  lw    $s0, 0x10($sp)
-/* 0EDCAC 802C92FC 24020002 */  addiu $v0, $zero, 2
-/* 0EDCB0 802C9300 03E00008 */  jr    $ra
-/* 0EDCB4 802C9304 27BD0020 */   addiu $sp, $sp, 0x20
-
diff --git a/asm/nonmatchings/code_e92d0_len_5da0/GetModelCenter.s b/asm/nonmatchings/code_e92d0_len_5da0/GetModelCenter.s
deleted file mode 100644
index 942b0f1188..0000000000
--- a/asm/nonmatchings/code_e92d0_len_5da0/GetModelCenter.s
+++ /dev/null
@@ -1,38 +0,0 @@
-.set noat # allow manual use of $at
-.set noreorder # don't insert nops after branches
-
-
-glabel GetModelCenter
-/* 0ED930 802C8F80 27BDFFC0 */  addiu $sp, $sp, -0x40
-/* 0ED934 802C8F84 AFB00038 */  sw    $s0, 0x38($sp)
-/* 0ED938 802C8F88 0080802D */  daddu $s0, $a0, $zero
-/* 0ED93C 802C8F8C AFBF003C */  sw    $ra, 0x3c($sp)
-/* 0ED940 802C8F90 8E02000C */  lw    $v0, 0xc($s0)
-/* 0ED944 802C8F94 0C0B1EAF */  jal   get_variable
-/* 0ED948 802C8F98 8C450000 */   lw    $a1, ($v0)
-/* 0ED94C 802C8F9C 3044FFFF */  andi  $a0, $v0, 0xffff
-/* 0ED950 802C8FA0 27A50020 */  addiu $a1, $sp, 0x20
-/* 0ED954 802C8FA4 27A2002C */  addiu $v0, $sp, 0x2c
-/* 0ED958 802C8FA8 AFA20010 */  sw    $v0, 0x10($sp)
-/* 0ED95C 802C8FAC 27A20030 */  addiu $v0, $sp, 0x30
-/* 0ED960 802C8FB0 AFA20014 */  sw    $v0, 0x14($sp)
-/* 0ED964 802C8FB4 27A20034 */  addiu $v0, $sp, 0x34
-/* 0ED968 802C8FB8 27A60024 */  addiu $a2, $sp, 0x24
-/* 0ED96C 802C8FBC 27A70028 */  addiu $a3, $sp, 0x28
-/* 0ED970 802C8FC0 0C046C3B */  jal   get_model_center_and_size
-/* 0ED974 802C8FC4 AFA20018 */   sw    $v0, 0x18($sp)
-/* 0ED978 802C8FC8 C7A00020 */  lwc1  $f0, 0x20($sp)
-/* 0ED97C 802C8FCC 4600010D */  trunc.w.s $f4, $f0
-/* 0ED980 802C8FD0 E6040084 */  swc1  $f4, 0x84($s0)
-/* 0ED984 802C8FD4 C7A00024 */  lwc1  $f0, 0x24($sp)
-/* 0ED988 802C8FD8 C7A20028 */  lwc1  $f2, 0x28($sp)
-/* 0ED98C 802C8FDC 4600010D */  trunc.w.s $f4, $f0
-/* 0ED990 802C8FE0 E6040088 */  swc1  $f4, 0x88($s0)
-/* 0ED994 802C8FE4 4600110D */  trunc.w.s $f4, $f2
-/* 0ED998 802C8FE8 E604008C */  swc1  $f4, 0x8c($s0)
-/* 0ED99C 802C8FEC 8FBF003C */  lw    $ra, 0x3c($sp)
-/* 0ED9A0 802C8FF0 8FB00038 */  lw    $s0, 0x38($sp)
-/* 0ED9A4 802C8FF4 24020002 */  addiu $v0, $zero, 2
-/* 0ED9A8 802C8FF8 03E00008 */  jr    $ra
-/* 0ED9AC 802C8FFC 27BD0040 */   addiu $sp, $sp, 0x40
-
diff --git a/diff.py b/diff.py
index 013d935948..0e1da09deb 100755
--- a/diff.py
+++ b/diff.py
@@ -1,48 +1,71 @@
 #!/usr/bin/env python3
 import sys
-import re
-import os
-import ast
-import argparse
-import subprocess
-import collections
-import difflib
-import string
-import itertools
-import threading
-import queue
-import time
-
 
 def fail(msg):
     print(msg, file=sys.stderr)
     sys.exit(1)
 
-
-MISSING_PREREQUISITES = (
-    "Missing prerequisite python module {}. "
-    "Run `python3 -m pip install --user colorama ansiwrap attrs watchdog python-Levenshtein cxxfilt` to install prerequisites (cxxfilt only needed with --source)."
-)
-
-try:
-    import attr
-    from colorama import Fore, Style, Back
-    import ansiwrap
-    import watchdog
-except ModuleNotFoundError as e:
-    fail(MISSING_PREREQUISITES.format(e.name))
-
 # Prefer to use diff_settings.py from the current working directory
 sys.path.insert(0, ".")
 try:
     import diff_settings
 except ModuleNotFoundError:
     fail("Unable to find diff_settings.py in the same directory.")
+sys.path.pop(0)
 
-# ==== CONFIG ====
+# ==== COMMAND-LINE ====
+
+try:
+    import argcomplete  # type: ignore
+except ModuleNotFoundError:
+    argcomplete = None
+import argparse
 
 parser = argparse.ArgumentParser(description="Diff MIPS assembly.")
-parser.add_argument("start", help="Function name or address to start diffing from.")
+
+start_argument = parser.add_argument("start", help="Function name or address to start diffing from.")
+if argcomplete:
+    def complete_symbol(**kwargs):
+        prefix = kwargs["prefix"]
+        if prefix == "":
+            # skip reading the map file, which would
+            # result in a lot of useless completions
+            return []
+        parsed_args = kwargs["parsed_args"]
+        config = {}
+        diff_settings.apply(config, parsed_args)
+        mapfile = config.get("mapfile")
+        if not mapfile:
+            return []
+        completes = []
+        with open(mapfile) as f:
+            data = f.read()
+            # assume symbols are prefixed by a space character
+            search = f" {prefix}"
+            pos = data.find(search)
+            while pos != -1:
+                # skip the space character in the search string
+                pos += 1
+                # assume symbols are suffixed by either a space
+                # character or a (unix-style) line return
+                spacePos = data.find(" ", pos)
+                lineReturnPos = data.find("\n", pos)
+                if lineReturnPos == -1:
+                    endPos = spacePos
+                elif spacePos == -1:
+                    endPos = lineReturnPos
+                else:
+                    endPos = min(spacePos, lineReturnPos)
+                if endPos == -1:
+                    match = data[pos:]
+                    pos = -1
+                else:
+                    match = data[pos:endPos]
+                    pos = data.find(search, endPos)
+                completes.append(match)
+        return completes
+    start_argument.completer = complete_symbol
+
 parser.add_argument("end", nargs="?", help="Address to end diff at.")
 parser.add_argument(
     "-o",
@@ -52,6 +75,7 @@ parser.add_argument(
 )
 parser.add_argument(
     "-e",
+    "--elf",
     dest="diff_elf_symbol",
     help="Diff a given function in two ELFs, one being stripped and the other one non-stripped. Requires objdump from binutils 2.33+.",
 )
@@ -132,6 +156,14 @@ parser.add_argument(
     help="Automatically update when source/object files change. "
     "Recommended in combination with -m.",
 )
+parser.add_argument(
+    "-3",
+    "--threeway",
+    dest="threeway",
+    action="store_true",
+    help="Show a three-way diff between target asm, current asm, and asm "
+    "prior to -w rebuild. Requires -w.",
+)
 parser.add_argument(
     "--width",
     dest="column_width",
@@ -146,7 +178,6 @@ parser.add_argument(
     choices=["levenshtein", "difflib"],
     help="Diff algorithm to use.",
 )
-
 parser.add_argument(
     "--max-size",
     "--max-lines",
@@ -158,12 +189,44 @@ parser.add_argument(
 
 # Project-specific flags, e.g. different versions/make arguments.
 if hasattr(diff_settings, "add_custom_arguments"):
-    diff_settings.add_custom_arguments(parser)
+    diff_settings.add_custom_arguments(parser)  # type: ignore
+
+if argcomplete:
+    argcomplete.autocomplete(parser)
+
+# ==== IMPORTS ====
+
+import re
+import os
+import ast
+import subprocess
+import difflib
+import string
+import itertools
+import threading
+import queue
+import time
+from typing import Any, Dict, List, NamedTuple, Optional, Set, Tuple, Union
+
+
+MISSING_PREREQUISITES = (
+    "Missing prerequisite python module {}. "
+    "Run `python3 -m pip install --user colorama ansiwrap watchdog python-Levenshtein cxxfilt` to install prerequisites (cxxfilt only needed with --source)."
+)
+
+try:
+    from colorama import Fore, Style, Back  # type: ignore
+    import ansiwrap  # type: ignore
+    import watchdog  # type: ignore
+except ModuleNotFoundError as e:
+    fail(MISSING_PREREQUISITES.format(e.name))
+
+# ==== CONFIG ====
 
 args = parser.parse_args()
 
 # Set imgs, map file and make flags in a project-specific manner.
-config = {}
+config: Dict[str, Any] = {}
 diff_settings.apply(config, args)
 
 arch = config.get("arch", "mips")
@@ -199,13 +262,13 @@ FS_WATCH_EXTENSIONS = [".c", ".h"]
 
 if args.algorithm == "levenshtein":
     try:
-        import Levenshtein
+        import Levenshtein  # type: ignore
     except ModuleNotFoundError as e:
         fail(MISSING_PREREQUISITES.format(e.name))
 
 if args.source:
     try:
-        import cxxfilt
+        import cxxfilt  # type: ignore
     except ModuleNotFoundError as e:
         fail(MISSING_PREREQUISITES.format(e.name))
 
@@ -538,22 +601,21 @@ def process_mips_reloc(row, prev):
     return before + repl + after
 
 
-def cleanup_whitespace(line):
-    return "".join(f"{o:<8s}" for o in line.strip().split("\t"))
+def pad_mnemonic(line):
+    if "\t" not in line:
+        return line
+    mn, args = line.split("\t", 1)
+    return f"{mn:<7s} {args}"
 
 
-Line = collections.namedtuple(
-    "Line",
-    [
-        "mnemonic",
-        "diff_row",
-        "original",
-        "line_num",
-        "branch_target",
-        "source_lines",
-        "comment",
-    ],
-)
+class Line(NamedTuple):
+    mnemonic: str
+    diff_row: str
+    original: str
+    line_num: str
+    branch_target: Optional[str]
+    source_lines: List[str]
+    comment: Optional[str]
 
 
 def process(lines):
@@ -565,6 +627,7 @@ def process(lines):
             lines.pop()
 
     output = []
+    stop_after_delay_slot = False
     for row in lines:
         if args.diff_obj and (">:" in row or not row):
             continue
@@ -634,18 +697,15 @@ def process(lines):
         source_lines = []
 
         if args.stop_jrra and mnemonic == "jr" and row_parts[1].strip() == "ra":
+            stop_after_delay_slot = True
+        elif stop_after_delay_slot:
             break
 
-    # Cleanup whitespace, after relocation fixups have happened
-    output = [
-        line._replace(original=cleanup_whitespace(line.original)) for line in output
-    ]
-
     return output
 
 
 def format_single_line_diff(line1, line2, column_width):
-    return f"{ansi_ljust(line1,column_width)}{ansi_ljust(line2,column_width)}"
+    return f"{ansi_ljust(line1,column_width)}{line2}"
 
 
 class SymbolColorer:
@@ -743,11 +803,27 @@ def diff_sequences(seq1, seq2):
     return Levenshtein.opcodes(seq1, seq2)
 
 
-def do_diff(basedump, mydump):
-    output = []
+class OutputLine:
+    base: Optional[str]
+    fmt2: str
+    key2: str
 
-    # TODO: status line?
-    # output.append(sha1sum(mydump))
+    def __init__(self, base: Optional[str], fmt2: str, key2: str) -> None:
+        self.base = base
+        self.fmt2 = fmt2
+        self.key2 = key2
+
+    def __eq__(self, other: object) -> bool:
+        if not isinstance(other, OutputLine):
+            return NotImplemented
+        return self.key2 == other.key2
+
+    def __hash__(self) -> int:
+        return hash(self.key2)
+
+
+def do_diff(basedump: str, mydump: str) -> List[OutputLine]:
+    output: List[OutputLine] = []
 
     lines1 = process(basedump.split("\n"))
     lines2 = process(mydump.split("\n"))
@@ -758,8 +834,8 @@ def do_diff(basedump, mydump):
     sc4 = SymbolColorer(4)
     sc5 = SymbolColorer(0)
     sc6 = SymbolColorer(0)
-    bts1 = set()
-    bts2 = set()
+    bts1: Set[str] = set()
+    bts2: Set[str] = set()
 
     if args.show_branches:
         for (lines, btset, sc) in [
@@ -856,40 +932,36 @@ def do_diff(basedump, mydump):
                 out2 = line2.original
             elif line1:
                 line_prefix = "<"
-                line_color1 = line_color2 = sym_color = Fore.RED
+                line_color1 = sym_color = Fore.RED
                 out1 = line1.original
                 out2 = ""
             elif line2:
                 line_prefix = ">"
-                line_color1 = line_color2 = sym_color = Fore.GREEN
+                line_color2 = sym_color = Fore.GREEN
                 out1 = ""
                 out2 = line2.original
 
-            in_arrow1 = "  "
-            in_arrow2 = "  "
-            out_arrow1 = ""
-            out_arrow2 = ""
-
-            if args.show_branches and line1:
-                if line1.line_num in bts1:
-                    in_arrow1 = sc5.color_symbol(line1.line_num, "~>") + line_color1
-                if line1.branch_target is not None:
-                    out_arrow1 = " " + sc5.color_symbol(line1.branch_target + ":", "~>")
-            if args.show_branches and line2:
-                if line2.line_num in bts2:
-                    in_arrow2 = sc6.color_symbol(line2.line_num, "~>") + line_color2
-                if line2.branch_target is not None:
-                    out_arrow2 = " " + sc6.color_symbol(line2.branch_target + ":", "~>")
-
             if args.source and line2 and line2.comment:
                 out2 += f" {line2.comment}"
 
-            line_num1 = line1.line_num if line1 else ""
-            line_num2 = line2.line_num if line2 else ""
+            def format_part(out: str, line: Optional[Line], line_color: str, btset: Set[str], sc: SymbolColorer) -> Optional[str]:
+                if line is None:
+                    return None
+                in_arrow = "  "
+                out_arrow = ""
+                if args.show_branches:
+                    if line.line_num in btset:
+                        in_arrow = sc.color_symbol(line.line_num, "~>") + line_color
+                    if line.branch_target is not None:
+                        out_arrow = " " + sc.color_symbol(line.branch_target + ":", "~>")
+                out = pad_mnemonic(out)
+                return f"{line_color}{line.line_num} {in_arrow} {out}{Style.RESET_ALL}{out_arrow}"
 
-            out1 = f"{line_color1}{line_num1} {in_arrow1} {out1}{Style.RESET_ALL}{out_arrow1}"
-            out2 = f"{line_color2}{line_num2} {in_arrow2} {out2}{Style.RESET_ALL}{out_arrow2}"
-            mid = f"{sym_color}{line_prefix} "
+            part1 = format_part(out1, line1, line_color1, bts1, sc5)
+            part2 = format_part(out2, line2, line_color2, bts2, sc6)
+            key2 = line2.original if line2 else ""
+
+            mid = f"{sym_color}{line_prefix}"
 
             if line2:
                 for source_line in line2.source_lines:
@@ -907,22 +979,82 @@ def do_diff(basedump, mydump):
                                 )
                             except:
                                 pass
-                    output.append(
-                        format_single_line_diff(
-                            "",
-                            f"  {color}{source_line}{Style.RESET_ALL}",
-                            args.column_width,
-                        )
-                    )
+                    output.append(OutputLine(None, f"  {color}{source_line}{Style.RESET_ALL}", source_line))
 
-            output.append(format_single_line_diff(out1, mid + out2, args.column_width))
+            fmt2 = mid + " " + (part2 or "")
+            output.append(OutputLine(part1, fmt2, key2))
 
-    return output[args.skip_lines :]
+    return output
+
+
+def chunk_diff(diff: List[OutputLine]) -> List[Union[List[OutputLine], OutputLine]]:
+    cur_right: List[OutputLine] = []
+    chunks: List[Union[List[OutputLine], OutputLine]] = []
+    for output_line in diff:
+        if output_line.base is not None:
+            chunks.append(cur_right)
+            chunks.append(output_line)
+            cur_right = []
+        else:
+            cur_right.append(output_line)
+    chunks.append(cur_right)
+    return chunks
+
+
+def format_diff(old_diff: List[OutputLine], new_diff: List[OutputLine]) -> Tuple[str, List[str]]:
+    old_chunks = chunk_diff(old_diff)
+    new_chunks = chunk_diff(new_diff)
+    output: List[Tuple[str, OutputLine, OutputLine]] = []
+    assert len(old_chunks) == len(new_chunks), "same target"
+    empty = OutputLine("", "", "")
+    for old_chunk, new_chunk in zip(old_chunks, new_chunks):
+        if isinstance(old_chunk, list):
+            assert isinstance(new_chunk, list)
+            if not old_chunk and not new_chunk:
+                # Most of the time lines sync up without insertions/deletions,
+                # and there's no interdiffing to be done.
+                continue
+            differ = difflib.SequenceMatcher(a=old_chunk, b=new_chunk, autojunk=False)
+            for (tag, i1, i2, j1, j2) in differ.get_opcodes():
+                if tag in ["equal", "replace"]:
+                    for i, j in zip(range(i1, i2), range(j1, j2)):
+                        output.append(("", old_chunk[i], new_chunk[j]))
+                elif tag == "insert":
+                    for j in range(j1, j2):
+                        output.append(("", empty, new_chunk[j]))
+                else:
+                    for i in range(i1, i2):
+                        output.append(("", old_chunk[i], empty))
+        else:
+            assert isinstance(new_chunk, OutputLine)
+            # old_chunk.base and new_chunk.base have the same text since
+            # both diffs are based on the same target, but they might
+            # differ in color. Use the new version.
+            output.append((new_chunk.base or "", old_chunk, new_chunk))
+
+    # TODO: status line, with e.g. approximate permuter score?
+    width = args.column_width
+    if args.threeway:
+        header_line = "TARGET".ljust(width) + "  CURRENT".ljust(width) + "  PREVIOUS"
+        diff_lines = [
+            ansi_ljust(base, width)
+            + ansi_ljust(new.fmt2, width)
+            + (old.fmt2 or "-" if old != new else "")
+            for (base, old, new) in output
+        ]
+    else:
+        header_line = ""
+        diff_lines = [
+            ansi_ljust(base, width) + new.fmt2
+            for (base, old, new) in output
+            if base or new.key2
+        ]
+    return header_line, diff_lines
 
 
 def debounced_fs_watch(targets, outq, debounce_delay):
-    import watchdog.events
-    import watchdog.observers
+    import watchdog.events  # type: ignore
+    import watchdog.observers  # type: ignore
 
     class WatchEventHandler(watchdog.events.FileSystemEventHandler):
         def __init__(self, queue, file_targets):
@@ -993,12 +1125,18 @@ class Display:
         self.basedump = basedump
         self.mydump = mydump
         self.emsg = None
+        self.last_diff_output = None
 
     def run_less(self):
         if self.emsg is not None:
             output = self.emsg
         else:
-            output = "\n".join(do_diff(self.basedump, self.mydump))
+            diff_output = do_diff(self.basedump, self.mydump)
+            last_diff_output = self.last_diff_output or diff_output
+            self.last_diff_output = diff_output
+            header, diff_lines = format_diff(last_diff_output, diff_output)
+            header_lines = [header] if header else []
+            output = "\n".join(header_lines + diff_lines[args.skip_lines :])
 
         # Pipe the output through 'tail' and only then to less, to ensure the
         # write call doesn't block. ('tail' has to buffer all its input before
diff --git a/include/common_structs.h b/include/common_structs.h
index d21831fa92..0efa114222 100644
--- a/include/common_structs.h
+++ b/include/common_structs.h
@@ -1429,4 +1429,33 @@ typedef struct EncounterStatus {
     /* 0x98 */ s32 unk_98;
 } EncounterStatus; // size = 0x9C
 
+typedef struct SaveData {
+    /* 0x0000 */ char magicString[16]; /* "Mario Story 006" string */
+    /* 0x0010 */ s8 pad[32]; /* always zero */
+    /* 0x0030 */ s32 crc1;
+    /* 0x0034 */ s32 crc2;
+    /* 0x0038 */ s32 saveSlot;
+    /* 0x003C */ s32 saveCount;
+    /* 0x0040 */ struct PlayerData player;
+    /* 0x0380 */ char unk_380[0xE0];
+    /* 0x0460 */ s32 starPoints;
+    /* 0x0464 */ char unk_464[4];
+    /* 0x0468 */ s16 areaID;
+    /* 0x046A */ s16 mapID;
+    /* 0x046C */ s16 entryID;
+    /* 0x046E */ char unk_46E[2];
+    /* 0x0470 */ s32 enemyDefeatFlags[720];
+    /* 0x0FB0 */ s8 globalFlags[256];
+    /* 0x10B0 */ s8 globalBytes[512];
+    /* 0x12B0 */ s32 areaFlags[8];
+    /* 0x12D0 */ s8 areaBytes[16];
+    /* 0x12E0 */ char unk_12E0[6];
+    /* 0x12E6 */ s16 savePos[3];
+    /* 0x12EC */ s32 unk_12EC;
+    /* 0x12F0 */ s8 unk_12F0[12]; /* player name starts at 4th char */
+    /* 0x12FC */ s32 unk_12FC;
+    /* 0x1300 */ s32 unk_1300;
+    /* 0x1304 */ char unk_1304[0x7C];
+} SaveData; // size = 0x1380
+
 #endif
diff --git a/include/functions.h b/include/functions.h
index 6754c373ec..f62d9d992a 100644
--- a/include/functions.h
+++ b/include/functions.h
@@ -10,8 +10,14 @@ void osCleanupThread(void);
 s32 heap_malloc(s32 size);
 HeapNode* _heap_create(void* addr, s32 size);
 
+s8 get_global_byte(s32 index);
+s32 get_global_flag(s32 index);
+s8 get_area_byte(s32 index);
+s32 get_area_flag(s32 index);
+
 void clone_model(u16 srcModelID, u16 newModelID);
 Model* get_model_from_list_index(s32 listIndex);
+void get_model_center_and_size(s32 modelID, f32* centerX, f32* centerY, f32* centerZ, f32* sizeX, f32* sizeY, f32* sizeZ);
 
 void update_collider_transform(s16 colliderID);
 void get_collider_center(s32 colliderID, f32* x, f32* y, f32* z);
diff --git a/include/variables.h b/include/variables.h
index f2078ce060..79dbce9b8c 100644
--- a/include/variables.h
+++ b/include/variables.h
@@ -92,4 +92,9 @@ extern UNK_TYPE D_80108558;
 extern UNK_TYPE D_8010F250; // play_sound state struct?
 extern s8 D_8010EBB0;
 
+extern SaveData gCurrentSaveFile;
+
+extern u32* gMapFlags;
+extern u32* gMapVars;
+
 #endif
diff --git a/src/code_a5dd0_len_114e0.c b/src/code_a5dd0_len_114e0.c
index c154b19a60..a9577cdb92 100644
--- a/src/code_a5dd0_len_114e0.c
+++ b/src/code_a5dd0_len_114e0.c
@@ -191,7 +191,7 @@ INCLUDE_ASM("code_a5dd0_len_114e0", get_model_list_index_from_tree_index);
 
 INCLUDE_ASM("code_a5dd0_len_114e0", func_8011B090);
 
-INCLUDE_ASM("code_a5dd0_len_114e0", get_model_center_and_size);
+void INCLUDE_ASM("code_a5dd0_len_114e0", get_model_center_and_size, s32 modelID, f32* centerX, f32* centerY, f32* centerZ, f32* sizeX, f32* sizeY, f32* sizeZ);
 
 INCLUDE_ASM("code_a5dd0_len_114e0", func_8011B1C0);
 
diff --git a/src/code_dba20_len_350.c b/src/code_dba20_len_350.c
index 068ec82156..69b2579604 100644
--- a/src/code_dba20_len_350.c
+++ b/src/code_dba20_len_350.c
@@ -7,19 +7,104 @@ INCLUDE_ASM("code_dba20_len_350", clear_area_flags);
 INCLUDE_ASM("code_dba20_len_350", clear_global_flag);
 
 INCLUDE_ASM("code_dba20_len_350", set_global_flag);
+/*s32 set_global_flag(s32 index) {
+    //SaveData* saveFile = &gCurrentSaveFile;
+    s32 flag;
 
-INCLUDE_ASM("code_dba20_len_350", get_global_flag);
+    if (index <= -120000000) {
+        index += 130000000;
+    }
 
-INCLUDE_ASM("code_dba20_len_350", set_global_byte);
+    flag = gCurrentSaveFile->globalFlags[index / 32] & (1 << (index % 32));
 
-INCLUDE_ASM("code_dba20_len_350", get_global_byte);
+    if (flag) {
+        flag = 1;
+    }
+
+    gCurrentSaveFile->globalFlags[index / 32] |= (1 << (index % 32));
+    return flag;
+}*/
+
+s32 INCLUDE_ASM("code_dba20_len_350", get_global_flag, s32 index);
+/*s32 get_global_flag(s32 index) {
+    s32 bitIdx;
+    s32 wordIdx;
+    s32 bit;
+    s32 phi_return;
+
+    if (index <= -120000000) {
+        index += 130000000;
+    } 
+    wordIdx = index / 32;
+    bitIdx = index % 32;
+    bit = gCurrentSaveFile->globalFlags[wordIdx] & (1 << bitIdx);
+
+    if (bit != 0) {
+        bit = 1;
+    }
+    return bit; 
+    //return (bit != 0) ? 1 : bit; // ??? surely this is `bit != 0`
+}*/
+
+s8 set_global_byte(s32 index, s8 value) {
+    SaveData* saveFile = &gCurrentSaveFile;
+    s32 ret = saveFile->globalBytes[index];
+
+    saveFile->globalBytes[index] = value;
+    return ret;
+}
+
+s8 get_global_byte(s32 index) {
+    return gCurrentSaveFile.globalBytes[index];
+}
 
 INCLUDE_ASM("code_dba20_len_350", clear_area_flag);
 
 INCLUDE_ASM("code_dba20_len_350", set_area_flag);
+/*s32 set_area_flag(s32 index) {
+    SaveData* saveFile = &gCurrentSaveFile;
+    s32 flag;
+    s32 flagIdx;
+    s32 flagShift;
 
-INCLUDE_ASM("code_dba20_len_350", get_area_flag);
+    flagIdx = index / 32;
+    flagShift = index % 32;
 
-INCLUDE_ASM("code_dba20_len_350", set_area_byte);
+    flag = saveFile->areaFlags[flagIdx] & (1 << flagShift);
 
-INCLUDE_ASM("code_dba20_len_350", get_area_byte);
+    if (flag != 0) {
+        flag = 1;
+    }
+
+    saveFile->areaFlags[flagIdx] |= (1 << flagShift);
+
+    return flag;
+}*/
+
+s32 get_area_flag(s32 index) {
+    s32 flag;
+    s32 flagIdx;
+    s32 flagShift;
+
+    flagIdx = index / 32;
+    flagShift = index % 32;
+
+    flag = gCurrentSaveFile.areaFlags[flagIdx] & (1 << flagShift);
+
+    if (flag != 0) {
+        flag = 1;
+    }
+    return flag;
+}
+
+s8 set_area_byte(s32 index, s8 value) {
+    SaveData* saveFile = &gCurrentSaveFile;
+    s32 ret = saveFile->areaBytes[index];
+
+    saveFile->areaBytes[index] = value;
+    return ret;
+}
+
+s8 get_area_byte(s32 index) {
+    return gCurrentSaveFile.areaBytes[index];
+}
diff --git a/src/code_e92d0_len_5da0.c b/src/code_e92d0_len_5da0.c
index 064305180b..e21bad3142 100644
--- a/src/code_e92d0_len_5da0.c
+++ b/src/code_e92d0_len_5da0.c
@@ -719,6 +719,57 @@ s32 func_802C73B8(ScriptInstance* script) {
 INCLUDE_ASM("code_e92d0_len_5da0", si_execute_next_command);
 
 s32 INCLUDE_ASM("code_e92d0_len_5da0", get_variable, ScriptInstance* script, Bytecode var);
+// TODO: consider renaming to si_get_variable
+/*s32 get_variable(ScriptInstance* script, Bytecode var) {
+    s32 abs_value;
+    s32 word_index;
+    s32 bit_index;
+
+    if (var <= -270000000) {
+        return var;
+    } else if (var <= -250000000) {
+        return var;
+    } else if (var <= -220000000) {
+        return (s32) fixed_var_to_float(var);
+    } else if (var <= -200000000) {
+        var += 210000000;
+        word_index = var / 32;
+        bit_index = var % 32;
+        return (script->flagArray[word_index] & (1 << bit_index)) != 0;
+    } else if (var <= -180000000) {
+        var += 190000000;
+        var = script->array[var];
+        return (var > -270000000 && var < -220000000) ? (s32) fixed_var_to_float(var) : var;
+    } else if (var <= -160000000) {
+        return get_global_byte(var + 170000000);
+    } else if (var <= -140000000) {
+        return get_area_byte(var + 150000000);
+    } else if (var <= -120000000) {
+        return get_global_flag(var + 130000000);
+    } else if (var <= -100000000) {
+        return get_area_flag(var + 110000000);
+    } else if (var <= -80000000) {
+        s32 avar = var + 90000000;
+        word_index = avar / 32;
+        bit_index = avar % 32;
+        return (gMapFlags[word_index] & (1 << bit_index)) != 0;
+    } else if (var <= -60000000) {
+        var += 70000000;
+        word_index = var / 32;
+        bit_index = var % 32;
+        return (script->varFlags[word_index] & (1 << bit_index)) != 0;
+    } else if (var <= -40000000) {
+        var += 50000000;
+        var = gMapVars[var];
+        return (var > -270000000 && var < -220000000) ? (s32) fixed_var_to_float(var) : var;
+    } else if (var <= -20000000) {
+        var += 30000000;
+        var = script->varTable[var];
+        return (var > -270000000 && var < -220000000) ? (s32) fixed_var_to_float(var) : var;
+    } else {
+        return var;
+    }
+}*/
 
 INCLUDE_ASM("code_e92d0_len_5da0", get_variable_index);
 
@@ -848,7 +899,21 @@ ApiStatus CloneModel(ScriptInstance* script, s32 isInitialCall) {
     return ApiStatus_DONE2;
 }
 
-INCLUDE_API_ASM("code_e92d0_len_5da0", GetModelCenter);
+ApiStatus GetModelCenter(ScriptInstance* script, s32 isInitialCall) {
+    Bytecode* thisPos = script->ptrReadPos;
+    f32 centerX;
+    f32 centerY;
+    f32 centerZ;
+    f32 sizeX;
+    f32 sizeY;
+    f32 sizeZ;
+
+    get_model_center_and_size(get_variable(script, *thisPos++) & 0xFFFF, &centerX, &centerY, &centerZ, &sizeX, &sizeY, &sizeZ);
+    script->varTable[0] = centerX;
+    script->varTable[1] = centerY;
+    script->varTable[2] = centerZ;
+    return ApiStatus_DONE2;
+}
 
 INCLUDE_API_ASM("code_e92d0_len_5da0", SetTexPanner);
 
@@ -856,7 +921,20 @@ INCLUDE_API_ASM("code_e92d0_len_5da0", SetModelFlag10);
 
 INCLUDE_API_ASM("code_e92d0_len_5da0", EnableTexPanning);
 
-INCLUDE_API_ASM("code_e92d0_len_5da0", EnableModel);
+ApiStatus EnableModel(ScriptInstance* script, s32 isInitialCall) {
+    Bytecode* thisPos = script->ptrReadPos;
+    Bytecode treeIndex = get_variable(script, *thisPos++);
+    s32 listIndex = get_model_list_index_from_tree_index(treeIndex);
+    Bytecode flag = get_variable(script, *thisPos++);
+    Model* model = get_model_from_list_index(listIndex);
+
+    if (flag != 0) {
+        model->flags &= ~0x2;
+    } else {
+        model->flags |= 0x2;
+    }
+    return ApiStatus_DONE2;
+}
 
 ApiStatus SetGroupEnabled(ScriptInstance* script, s32 isInitialCall) {
     Bytecode* thisPos = script->ptrReadPos;
diff --git a/undefined_syms.txt b/undefined_syms.txt
index c5a390091c..1956a2254b 100644
--- a/undefined_syms.txt
+++ b/undefined_syms.txt
@@ -112,6 +112,11 @@ gScriptIdList = 0x802D1898;
 gScriptIndexList = 0x802DAA98;
 gScriptListCount = 0x802DAC98;
 
+gCurrentSaveFile = 0x800DACC0;
+
+gMapFlags = 0x802DA480;
+gMapVars = 0x802DA484;
+
 func_802A91F8 = 0x802A91F8;
 func_802A9208 = 0x802A9208;
 func_802A9228 = 0x802A9228;