diff --git a/Makefile b/Makefile
index 16e2d07395..f1e7e75f22 100644
--- a/Makefile
+++ b/Makefile
@@ -53,9 +53,6 @@ endif
 # Version-specific settings
 ifeq ($(VERSION),gc-eu-mq)
   DEBUG := 0
-  CFLAGS += -DNON_MATCHING
-  CPPFLAGS += -DNON_MATCHING
-  COMPARE := 0
 else ifeq ($(VERSION),gc-eu-mq-dbg)
   DEBUG := 1
 else
@@ -312,9 +309,9 @@ $(BUILD_DIR)/src/code/jpegdecoder.o: CC := $(CC_OLD)
 # For using asm_processor on some files:
 #$(BUILD_DIR)/.../%.o: CC := $(PYTHON) tools/asm_processor/build.py $(CC) -- $(AS) $(ASFLAGS) --
 
-ifeq ($(PERMUTER),)  # permuter + reencode.sh misbehaves, permuter doesn't care about encoding ((ro)data diffs) so just don't use it in that case
-# Handle encoding (UTF-8 -> EUC-JP)
-$(BUILD_DIR)/%.o: CC := tools/reencode.sh $(CC)
+ifeq ($(PERMUTER),)  # permuter + preprocess.py misbehaves, permuter doesn't care about rodata diffs or bss ordering so just don't use it in that case
+# Handle encoding (UTF-8 -> EUC-JP) and custom pragmas
+$(BUILD_DIR)/src/%.o: CC := $(PYTHON) tools/preprocess.py $(CC)
 endif
 
 else
diff --git a/data/fault.bss.s b/data/fault.bss.s
deleted file mode 100644
index d8d76a0518..0000000000
--- a/data/fault.bss.s
+++ /dev/null
@@ -1,27 +0,0 @@
-.include "macro.inc"
-
-# assembler directives
-.set noat      # allow manual use of $at
-.set noreorder # don't insert nops after branches
-.set gp=64     # allow use of 64-bit general purpose registers
-
-.section .bss
-
-# Note: This file is only included in the MATCHING build, the data is imported for non-matching
-
-.balign 16
-
-glabel sFaultInstance
-    .space 4
-
-glabel sFaultAwaitingInput
-    .space 4
-
-glabel sFaultStack
-    .space 0x600
-
-glabel sFaultThreadInfo
-    .space 0x20
-
-glabel gFaultMgr
-    .space 0x850
diff --git a/data/fault_drawer.bss.s b/data/fault_drawer.bss.s
deleted file mode 100644
index 89a29d8bb4..0000000000
--- a/data/fault_drawer.bss.s
+++ /dev/null
@@ -1,19 +0,0 @@
-.include "macro.inc"
-
-# assembler directives
-.set noat      # allow manual use of $at
-.set noreorder # don't insert nops after branches
-.set gp=64     # allow use of 64-bit general purpose registers
-
-.section .bss
-
-# Note: This file is only included in the MATCHING build, the data is imported for non-matching
-
-.balign 16
-
-glabel sFaultDrawer
-    .space 0x3C
-    .space 0x04 # padding
-
-glabel D_8016B6C0
-    .space 0x20
diff --git a/include/padmgr.h b/include/padmgr.h
index 65705949ea..ba86315c6f 100644
--- a/include/padmgr.h
+++ b/include/padmgr.h
@@ -44,8 +44,6 @@ typedef struct PadMgr {
     /* 0x0464 */ void* retraceCallbackArg;
 } PadMgr; // size = 0x468
 
-extern PadMgr gPadMgr;
-
 // Initialization
 
 void PadMgr_Init(PadMgr* padMgr, OSMesgQueue* serialEventQueue, IrqMgr* irqMgr, OSId id, OSPri priority, void* stack);
diff --git a/include/variables.h b/include/variables.h
index 1c7e4ebc85..a67d99246b 100644
--- a/include/variables.h
+++ b/include/variables.h
@@ -174,25 +174,38 @@ extern u8 gSampleBankTable[];
 extern SaveContext gSaveContext;
 extern RegEditor* gRegEditor;
 
+extern u8 gUseCutsceneCam;
+extern u16 D_8015FCCC;
+extern char D_8015FCD0[20];
+extern u8 D_8015FCE4;
 extern u16 gCamAtSplinePointsAppliedFrame;
 extern u16 gCamEyePointAppliedFrame;
 extern u16 gCamAtPointAppliedFrame;
-extern u8 gUseCutsceneCam;
 
+extern LightningStrike gLightningStrike;
+// TODO: These variables are here for BSS ordering but ideally they should not
+// be extern. This could be fixed by putting more stuff (e.g. struct definitions)
+// between gLightningStrike and gCustomLensFlareOn.
+extern s16 sLightningFlashAlpha;
+extern s16 sSunDepthTestX;
+extern s16 sSunDepthTestY;
 extern u8 gCustomLensFlareOn;
 extern Vec3f gCustomLensFlarePos;
 extern s16 gLensFlareScale;
 extern f32 gLensFlareColorIntensity;
 extern s16 gLensFlareGlareStrength;
-extern LightningStrike gLightningStrike;
 extern MapData* gMapData;
 extern f32 gBossMarkScale;
+extern u32 D_8016139C;
 extern PauseMapMarksData* gLoadedPauseMarkDataTable;
+extern TransitionTile gTransitionTile;
 extern s32 gTransitionTileState;
+extern VisMono gPlayVisMono;
 extern Color_RGBA8_u32 gVisMonoColor;
 extern PreNmiBuff* gAppNmiBufferPtr;
-extern Scheduler gScheduler;
 extern uintptr_t gSegments[NUM_SEGMENTS];
+extern Scheduler gScheduler;
+extern PadMgr gPadMgr;
 extern volatile OSTime gAudioThreadUpdateTimeTotalPerGfxTask;
 extern volatile OSTime gGfxTaskSentToNextReadyMinusAudioThreadUpdateTime;
 extern volatile OSTime gRSPAudioTimeTotal;
@@ -204,6 +217,7 @@ extern volatile OSTime gAudioThreadUpdateTimeAcc;
 extern volatile OSTime gRSPAudioTimeAcc;
 extern volatile OSTime gRSPGfxTimeAcc;
 extern volatile OSTime gRSPOtherTimeAcc;
+extern volatile OSTime D_8016A578;
 extern volatile OSTime gRDPTimeAcc;
 
 extern SfxBankEntry D_8016BAD0[9];
diff --git a/spec b/spec
index 287659d1b9..1630ba28cc 100644
--- a/spec
+++ b/spec
@@ -448,10 +448,6 @@ beginseg
 #endif
     include "$(BUILD_DIR)/src/code/fault.o"
     include "$(BUILD_DIR)/src/code/fault_drawer.o"
-#ifndef NON_MATCHING
-    include "$(BUILD_DIR)/data/fault.bss.o"
-    include "$(BUILD_DIR)/data/fault_drawer.bss.o"
-#endif
     include "$(BUILD_DIR)/src/code/kanread.o"
 #if OOT_DEBUG
     include "$(BUILD_DIR)/src/code/ucode_disas.o"
diff --git a/src/boot/idle.c b/src/boot/idle.c
index dec3006b01..3d62994c47 100644
--- a/src/boot/idle.c
+++ b/src/boot/idle.c
@@ -1,7 +1,12 @@
 #include "global.h"
 #include "terminal.h"
 
-OSThread gMainThread;
+// For retail BSS ordering, the block number of sMainThread must be 0 or
+// just above (the exact upper bound depends on the block numbers assigned to
+// extern variables declared in headers).
+#pragma increment_block_number 60
+
+OSThread sMainThread;
 STACK(sMainStack, 0x900);
 StackEntry sMainStackInfo;
 OSMesg sPiMgrCmdBuff[50];
@@ -86,8 +91,8 @@ void Idle_ThreadEntry(void* arg) {
     osViSwapBuffer((void*)0x803DA80); //! @bug Invalid vram address (probably intended to be 0x803DA800)
     osCreatePiManager(OS_PRIORITY_PIMGR, &gPiMgrCmdQueue, sPiMgrCmdBuff, ARRAY_COUNT(sPiMgrCmdBuff));
     StackCheck_Init(&sMainStackInfo, sMainStack, STACK_TOP(sMainStack), 0, 0x400, "main");
-    osCreateThread(&gMainThread, THREAD_ID_MAIN, Main_ThreadEntry, arg, STACK_TOP(sMainStack), THREAD_PRI_MAIN_INIT);
-    osStartThread(&gMainThread);
+    osCreateThread(&sMainThread, THREAD_ID_MAIN, Main_ThreadEntry, arg, STACK_TOP(sMainStack), THREAD_PRI_MAIN_INIT);
+    osStartThread(&sMainThread);
     osSetThreadPri(NULL, OS_PRIORITY_IDLE);
 
     for (;;) {}
diff --git a/src/code/fault.c b/src/code/fault.c
index 8b05fb8208..662c1a9f1a 100644
--- a/src/code/fault.c
+++ b/src/code/fault.c
@@ -44,6 +44,15 @@
 #include "terminal.h"
 #include "alloca.h"
 
+// For retail BSS ordering, the block number of sFaultInstance must be 0 or
+// just above (the exact upper bound depends on the block numbers assigned to
+// extern variables declared in headers).
+#if OOT_DEBUG
+#pragma increment_block_number 20
+#else
+#pragma increment_block_number 25
+#endif
+
 void FaultDrawer_Init(void);
 void FaultDrawer_SetOsSyncPrintfEnabled(u32 enabled);
 void FaultDrawer_DrawRecImpl(s32 xStart, s32 yStart, s32 xEnd, s32 yEnd, u16 color);
@@ -76,21 +85,11 @@ const char* sFpExceptionNames[] = {
     "Unimplemented operation", "Invalid operation", "Division by zero", "Overflow", "Underflow", "Inexact operation",
 };
 
-#ifndef NON_MATCHING
-// TODO: match .bss (has reordering issues)
-extern FaultMgr* sFaultInstance;
-extern u8 sFaultAwaitingInput;
-extern STACK(sFaultStack, 0x600);
-extern StackEntry sFaultThreadInfo;
-extern FaultMgr gFaultMgr;
-#else
-// Non-matching version for struct shiftability
 FaultMgr* sFaultInstance;
 u8 sFaultAwaitingInput;
 STACK(sFaultStack, 0x600);
 StackEntry sFaultThreadInfo;
 FaultMgr gFaultMgr;
-#endif
 
 typedef struct {
     /* 0x00 */ s32 (*callback)(void*, void*);
diff --git a/src/code/fault_drawer.c b/src/code/fault_drawer.c
index 8f5fbcbe49..5c82b7a0bb 100644
--- a/src/code/fault_drawer.c
+++ b/src/code/fault_drawer.c
@@ -99,15 +99,8 @@ FaultDrawer sFaultDrawerDefault = {
     NULL,
 };
 
-#ifndef NON_MATCHING
-// TODO: match .bss (has reordering issues)
-extern FaultDrawer sFaultDrawer;
-extern char D_8016B6C0[0x20];
-#else
-// Non-matching version for struct shiftability
 FaultDrawer sFaultDrawer;
 char D_8016B6C0[0x20];
-#endif
 
 void FaultDrawer_SetOsSyncPrintfEnabled(u32 enabled) {
     sFaultDrawer.osSyncPrintfEnabled = enabled;
diff --git a/src/code/sys_math3d.c b/src/code/sys_math3d.c
index 3e50249b12..a3b8ca2d61 100644
--- a/src/code/sys_math3d.c
+++ b/src/code/sys_math3d.c
@@ -1,6 +1,10 @@
 #include "global.h"
 #include "terminal.h"
 
+// For retail BSS ordering, the block number of cbf in Math3D_CylVsCylOverlapCenterDist
+// must be 0.
+#pragma increment_block_number 190
+
 s32 Math3D_LineVsLineClosestTwoPoints(Vec3f* lineAPointA, Vec3f* lineAPointB, Vec3f* lineBPointA, Vec3f* lineBPointB,
                                       Vec3f* lineAClosestToB, Vec3f* lineBClosestToA);
 s32 Math3D_TriLineIntersect(Vec3f* v0, Vec3f* v1, Vec3f* v2, f32 nx, f32 ny, f32 nz, f32 originDist, Vec3f* linePointA,
diff --git a/src/code/z_actor.c b/src/code/z_actor.c
index dfa9d88dd8..5e8b86aaa6 100644
--- a/src/code/z_actor.c
+++ b/src/code/z_actor.c
@@ -8,6 +8,9 @@
 #include "assets/objects/gameplay_dangeon_keep/gameplay_dangeon_keep.h"
 #include "assets/objects/object_bdoor/object_bdoor.h"
 
+// For retail BSS ordering, the block number of sCurCeilingPoly
+// must be between 2 and 243 inclusive.
+
 static CollisionPoly* sCurCeilingPoly;
 static s32 sCurCeilingBgId;
 
@@ -1894,7 +1897,9 @@ s32 func_8002F9EC(PlayState* play, Actor* actor, CollisionPoly* poly, s32 bgId,
     return false;
 }
 
-// Local data used for Farore's Wind light (stored in BSS, possibly a struct?)
+#pragma increment_block_number 22
+
+// Local data used for Farore's Wind light (stored in BSS)
 LightInfo D_8015BC00;
 LightNode* D_8015BC10;
 s32 D_8015BC14;
diff --git a/src/code/z_camera.c b/src/code/z_camera.c
index a54d640700..4caafe18c9 100644
--- a/src/code/z_camera.c
+++ b/src/code/z_camera.c
@@ -4,6 +4,10 @@
 #include "terminal.h"
 #include "overlays/actors/ovl_En_Horse/z_en_horse.h"
 
+// For retail BSS ordering, the block number of D_8015BD7C
+// must be between 88 and 123 inclusive.
+#pragma increment_block_number 30
+
 s16 Camera_RequestSettingImpl(Camera* camera, s16 requestedSetting, s16 flags);
 s32 Camera_RequestModeImpl(Camera* camera, s16 requestedMode, u8 forceModeChange);
 s32 Camera_UpdateWater(Camera* camera);
@@ -3628,6 +3632,8 @@ s32 Camera_KeepOn3(Camera* camera) {
     return 1;
 }
 
+#pragma increment_block_number 100
+
 s32 Camera_KeepOn4(Camera* camera) {
     static Vec3f D_8015BD50;
     static Vec3f D_8015BD60;
diff --git a/src/code/z_collision_check.c b/src/code/z_collision_check.c
index 14822cf1cd..118a445d7b 100644
--- a/src/code/z_collision_check.c
+++ b/src/code/z_collision_check.c
@@ -7,6 +7,10 @@ typedef void (*ColChkApplyFunc)(PlayState*, CollisionCheckContext*, Collider*);
 typedef void (*ColChkVsFunc)(PlayState*, CollisionCheckContext*, Collider*, Collider*);
 typedef s32 (*ColChkLineFunc)(PlayState*, CollisionCheckContext*, Collider*, Vec3f*, Vec3f*);
 
+// For retail BSS ordering, the block number of sparkInit in CollisionCheck_BlueBlood
+// must be between 183 and 255 inclusive.
+#pragma increment_block_number 50
+
 #if OOT_DEBUG
 /**
  * Draws a red triangle with vertices vA, vB, and vC.
@@ -2238,8 +2242,10 @@ void CollisionCheck_ATCylVsACQuad(PlayState* play, CollisionCheckContext* colChk
     }
 }
 
+#if OOT_DEBUG
 static s8 sBssDummy0;
 static s8 sBssDummy1;
+#endif
 
 void CollisionCheck_ATQuadVsACCyl(PlayState* play, CollisionCheckContext* colChkCtx, Collider* atCol, Collider* acCol) {
     static TriNorm tri1;
@@ -2300,10 +2306,12 @@ void CollisionCheck_ATQuadVsACCyl(PlayState* play, CollisionCheckContext* colChk
     }
 }
 
+#if OOT_DEBUG
 static s8 sBssDummy3;
 static s8 sBssDummy4;
 static s8 sBssDummy5;
 static s8 sBssDummy6;
+#endif
 
 void CollisionCheck_ATTrisVsACTris(PlayState* play, CollisionCheckContext* colChkCtx, Collider* atCol,
                                    Collider* acCol) {
@@ -2350,10 +2358,12 @@ void CollisionCheck_ATTrisVsACTris(PlayState* play, CollisionCheckContext* colCh
     }
 }
 
+#if OOT_DEBUG
 static s8 sBssDummy7;
 static s8 sBssDummy8;
 static s8 sBssDummy9;
 static s8 sBssDummy10;
+#endif
 
 void CollisionCheck_ATTrisVsACQuad(PlayState* play, CollisionCheckContext* colChkCtx, Collider* atCol,
                                    Collider* acCol) {
@@ -2586,9 +2596,10 @@ static ColChkApplyFunc sColChkApplyFuncs[] = {
  */
 void CollisionCheck_SetHitEffects(PlayState* play, CollisionCheckContext* colChkCtx) {
     Collider** acColP;
+    Collider* acCol;
 
     for (acColP = colChkCtx->colAC; acColP < colChkCtx->colAC + colChkCtx->colACCount; acColP++) {
-        Collider* acCol = *acColP;
+        acCol = *acColP;
 
         if (acCol != NULL && acCol->acFlags & AC_ON) {
             if (acCol->actor != NULL && acCol->actor->update == NULL) {
@@ -2635,9 +2646,10 @@ static ColChkVsFunc sACVsFuncs[COLSHAPE_MAX][COLSHAPE_MAX] = {
  */
 void CollisionCheck_AC(PlayState* play, CollisionCheckContext* colChkCtx, Collider* atCol) {
     Collider** acColP;
+    Collider* acCol;
 
     for (acColP = colChkCtx->colAC; acColP < colChkCtx->colAC + colChkCtx->colACCount; acColP++) {
-        Collider* acCol = *acColP;
+        acCol = *acColP;
 
         if (acCol != NULL && acCol->acFlags & AC_ON) {
             if (acCol->actor != NULL && acCol->actor->update == NULL) {
@@ -2661,12 +2673,13 @@ void CollisionCheck_AC(PlayState* play, CollisionCheckContext* colChkCtx, Collid
  */
 void CollisionCheck_AT(PlayState* play, CollisionCheckContext* colChkCtx) {
     Collider** atColP;
+    Collider* atCol;
 
     if (colChkCtx->colATCount == 0 || colChkCtx->colACCount == 0) {
         return;
     }
     for (atColP = colChkCtx->colAT; atColP < colChkCtx->colAT + colChkCtx->colATCount; atColP++) {
-        Collider* atCol = *atColP;
+        atCol = *atColP;
 
         if (atCol != NULL && atCol->atFlags & AT_ON) {
             if (atCol->actor != NULL && atCol->actor->update == NULL) {
@@ -2684,6 +2697,8 @@ typedef enum {
     /* 2 */ MASSTYPE_NORMAL
 } ColChkMassType;
 
+#pragma increment_block_number 253
+
 /**
  * Get mass type. Immovable colliders cannot be pushed, while heavy colliders can only be pushed by heavy and immovable
  * colliders.
@@ -3134,9 +3149,10 @@ static ColChkApplyFunc sApplyDamageFuncs[COLSHAPE_MAX] = {
  */
 void CollisionCheck_Damage(PlayState* play, CollisionCheckContext* colChkCtx) {
     s32 i;
+    Collider* col;
 
     for (i = 0; i < colChkCtx->colACCount; i++) {
-        Collider* col = colChkCtx->colAC[i];
+        col = colChkCtx->colAC[i];
 
         if (col == NULL) {
             continue;
@@ -3312,12 +3328,14 @@ void Collider_SetTrisDim(PlayState* play, ColliderTris* tris, s32 elemIndex, Col
     Collider_SetTrisElementDim(play, &trisElem->dim, src);
 }
 
+#if OOT_DEBUG
 // Due to an unknown reason, bss ordering changed between the 2 static Vec3f variables in the function below.
 // In order to reproduce this behavior, we need a specific number of bss variables in the file before that point.
 // For this, we introduce a certain amount of dummy variables throughout the file, which we fit inside padding added
 // by the compiler between structs like TriNorm and/or Vec3f, so they don't take space in bss.
 static s8 sBssDummy11;
 static s8 sBssDummy12;
+#endif
 
 /**
  * Updates the world spheres for all of the collider's JntSph elements attached to the specified limb
diff --git a/src/code/z_common_data.c b/src/code/z_common_data.c
index 2d7c36fc81..47aecd9e6a 100644
--- a/src/code/z_common_data.c
+++ b/src/code/z_common_data.c
@@ -1,5 +1,10 @@
 #include "global.h"
 
+// For retail BSS ordering, the block number of D_8015FA88 must be 0 or
+// just above (the exact upper bound depends on the block numbers assigned to
+// extern variables declared in headers).
+#pragma increment_block_number 60
+
 ALIGNED(16) SaveContext gSaveContext;
 u32 D_8015FA88;
 u32 D_8015FA8C;
diff --git a/src/code/z_kaleido_scope_call.c b/src/code/z_kaleido_scope_call.c
index be1d9aa59e..7b53b8521a 100644
--- a/src/code/z_kaleido_scope_call.c
+++ b/src/code/z_kaleido_scope_call.c
@@ -1,6 +1,11 @@
 #include "global.h"
 #include "terminal.h"
 
+// For retail BSS ordering, the block number of sKaleidoScopeUpdateFunc must be 0 or
+// just above (the exact upper bound depends on the block numbers assigned to
+// extern variables declared in headers).
+#pragma increment_block_number 60
+
 void (*sKaleidoScopeUpdateFunc)(PlayState* play);
 void (*sKaleidoScopeDrawFunc)(PlayState* play);
 f32 gBossMarkScale;
diff --git a/src/code/z_kankyo.c b/src/code/z_kankyo.c
index db80807355..ce79a7df93 100644
--- a/src/code/z_kankyo.c
+++ b/src/code/z_kankyo.c
@@ -4,6 +4,11 @@
 #include "assets/objects/gameplay_keep/gameplay_keep.h"
 #include "assets/objects/gameplay_field_keep/gameplay_field_keep.h"
 
+// For retail BSS ordering, the block number of sLensFlareUnused must be lower
+// than the extern variables declared in the header (e.g. gLightningStrike)
+// while the block number of sNGameOverLightNode must be higher.
+#pragma increment_block_number 80
+
 typedef enum {
     /* 0x00 */ LIGHTNING_BOLT_START,
     /* 0x01 */ LIGHTNING_BOLT_WAIT,
@@ -183,7 +188,7 @@ f32 sSandstormLerpScale = 0.0f;
 
 u8 gCustomLensFlareOn;
 Vec3f gCustomLensFlarePos;
-s16 gLensFlareUnused;
+s16 sLensFlareUnused;
 s16 gLensFlareScale;
 f32 gLensFlareColorIntensity;
 s16 gLensFlareGlareStrength;
@@ -207,6 +212,11 @@ s16 sLightningFlashAlpha;
 s16 sSunDepthTestX;
 s16 sSunDepthTestY;
 
+// These variables could be moved farther down in the file to reduce the amount
+// of block number padding here, but currently this causes BSS ordering issues
+// for debug.
+#pragma increment_block_number 230
+
 LightNode* sNGameOverLightNode;
 LightInfo sNGameOverLightInfo;
 LightNode* sSGameOverLightNode;
@@ -2208,7 +2218,7 @@ void Environment_DrawCustomLensFlare(PlayState* play) {
         pos.y = gCustomLensFlarePos.y;
         pos.z = gCustomLensFlarePos.z;
 
-        Environment_DrawLensFlare(play, &play->envCtx, &play->view, play->state.gfxCtx, pos, gLensFlareUnused,
+        Environment_DrawLensFlare(play, &play->envCtx, &play->view, play->state.gfxCtx, pos, sLensFlareUnused,
                                   gLensFlareScale, gLensFlareColorIntensity, gLensFlareGlareStrength, false);
     }
 }
diff --git a/src/code/z_play.c b/src/code/z_play.c
index c418c40899..f86986839b 100644
--- a/src/code/z_play.c
+++ b/src/code/z_play.c
@@ -9,9 +9,9 @@ UNK_TYPE D_8012D1F4 = 0; // unused
 
 Input* D_8012D1F8 = NULL;
 
-TransitionTile sTransitionTile;
+TransitionTile gTransitionTile;
 s32 gTransitionTileState;
-VisMono sPlayVisMono;
+VisMono gPlayVisMono;
 Color_RGBA8_u32 gVisMonoColor;
 
 #if OOT_DEBUG
@@ -199,7 +199,7 @@ void Play_Destroy(GameState* thisx) {
     CollisionCheck_DestroyContext(this, &this->colChkCtx);
 
     if (gTransitionTileState == TRANS_TILE_READY) {
-        TransitionTile_Destroy(&sTransitionTile);
+        TransitionTile_Destroy(&gTransitionTile);
         gTransitionTileState = TRANS_TILE_OFF;
     }
 
@@ -211,7 +211,7 @@ void Play_Destroy(GameState* thisx) {
 
     Letterbox_Destroy();
     TransitionFade_Destroy(&this->transitionFadeFlash);
-    VisMono_Destroy(&sPlayVisMono);
+    VisMono_Destroy(&gPlayVisMono);
 
     if (gSaveContext.save.linkAge != this->linkAgeOnLoad) {
         Inventory_SwapAgeEquipment();
@@ -409,7 +409,7 @@ void Play_Init(GameState* thisx) {
     TransitionFade_SetType(&this->transitionFadeFlash, TRANS_INSTANCE_TYPE_FADE_FLASH);
     TransitionFade_SetColor(&this->transitionFadeFlash, RGBA8(160, 160, 160, 255));
     TransitionFade_Start(&this->transitionFadeFlash);
-    VisMono_Init(&sPlayVisMono);
+    VisMono_Init(&gPlayVisMono);
     gVisMonoColor.a = 0;
     CutsceneFlags_UnsetAll(this);
 
@@ -521,18 +521,18 @@ void Play_Update(PlayState* this) {
         if (gTransitionTileState != TRANS_TILE_OFF) {
             switch (gTransitionTileState) {
                 case TRANS_TILE_PROCESS:
-                    if (TransitionTile_Init(&sTransitionTile, 10, 7) == NULL) {
+                    if (TransitionTile_Init(&gTransitionTile, 10, 7) == NULL) {
                         PRINTF("fbdemo_init呼出し失敗！\n"); // "fbdemo_init call failed!"
                         gTransitionTileState = TRANS_TILE_OFF;
                     } else {
-                        sTransitionTile.zBuffer = (u16*)gZBuffer;
+                        gTransitionTile.zBuffer = (u16*)gZBuffer;
                         gTransitionTileState = TRANS_TILE_READY;
                         R_UPDATE_RATE = 1;
                     }
                     break;
 
                 case TRANS_TILE_READY:
-                    TransitionTile_Update(&sTransitionTile);
+                    TransitionTile_Update(&gTransitionTile);
                     break;
 
                 default:
@@ -687,7 +687,7 @@ void Play_Update(PlayState* this) {
                             this->transitionMode = TRANS_MODE_OFF;
 
                             if (gTransitionTileState == TRANS_TILE_READY) {
-                                TransitionTile_Destroy(&sTransitionTile);
+                                TransitionTile_Destroy(&gTransitionTile);
                                 gTransitionTileState = TRANS_TILE_OFF;
                                 R_UPDATE_RATE = 3;
                             }
@@ -1135,8 +1135,8 @@ void Play_Draw(PlayState* this) {
             TransitionFade_Draw(&this->transitionFadeFlash, &gfxP);
 
             if (gVisMonoColor.a > 0) {
-                sPlayVisMono.vis.primColor.rgba = gVisMonoColor.rgba;
-                VisMono_Draw(&sPlayVisMono, &gfxP);
+                gPlayVisMono.vis.primColor.rgba = gVisMonoColor.rgba;
+                VisMono_Draw(&gPlayVisMono, &gfxP);
             }
 
             gSPEndDisplayList(gfxP++);
@@ -1147,7 +1147,7 @@ void Play_Draw(PlayState* this) {
         if (gTransitionTileState == TRANS_TILE_READY) {
             Gfx* sp88 = POLY_OPA_DISP;
 
-            TransitionTile_Draw(&sTransitionTile, &sp88);
+            TransitionTile_Draw(&gTransitionTile, &sp88);
             POLY_OPA_DISP = sp88;
             goto Play_Draw_DrawOverlayElements;
         }
diff --git a/src/overlays/actors/ovl_Boss_Ganon/z_boss_ganon.c b/src/overlays/actors/ovl_Boss_Ganon/z_boss_ganon.c
index 0498724369..0d5b973618 100644
--- a/src/overlays/actors/ovl_Boss_Ganon/z_boss_ganon.c
+++ b/src/overlays/actors/ovl_Boss_Ganon/z_boss_ganon.c
@@ -100,12 +100,22 @@ static ColliderCylinderInit sLightBallCylinderInit = {
 static u8 D_808E4C58[] = { 0, 12, 10, 12, 14, 16, 12, 14, 16, 12, 14, 16, 12, 14, 16, 10, 16, 14 };
 static Vec3f sZeroVec = { 0.0f, 0.0f, 0.0f };
 
+// For retail BSS ordering, the block number of sGanondorf must be 0 or just above.
+
+// TODO: There's probably a way to do this with less padding by spreading the variables out and moving
+// data around. It would be easier if we had more options for controlling BSS ordering in debug.
+#pragma increment_block_number 50
+
 static EnGanonMant* sCape;
 
+#pragma increment_block_number 200
+
 static s32 sSeed1;
 static s32 sSeed2;
 static s32 sSeed3;
 
+#pragma increment_block_number 200
+
 static BossGanon* sGanondorf;
 
 static EnZl3* sZelda;
diff --git a/src/overlays/actors/ovl_En_Xc/z_en_xc.c b/src/overlays/actors/ovl_En_Xc/z_en_xc.c
index 3809daae4d..17eff561ad 100644
--- a/src/overlays/actors/ovl_En_Xc/z_en_xc.c
+++ b/src/overlays/actors/ovl_En_Xc/z_en_xc.c
@@ -14,6 +14,9 @@
 #include "assets/scenes/dungeons/ice_doukutu/ice_doukutu_scene.h"
 #include "terminal.h"
 
+// For retail BSS ordering, the block number of sSfxPos
+// must be between 0 and 213 inclusive.
+
 #define FLAGS ACTOR_FLAG_4
 
 void EnXc_Init(Actor* thisx, PlayState* play);
@@ -387,6 +390,8 @@ s32 EnXc_SerenadeCS(EnXc* this, PlayState* play) {
 void EnXc_DoNothing(EnXc* this, PlayState* play) {
 }
 
+static Vec3f sSfxPos;
+
 void EnXc_SetWalkingSFX(EnXc* this, PlayState* play) {
     s32 pad[2];
     u32 sfxId;
@@ -432,7 +437,6 @@ void EnXc_SetLandingSFX(EnXc* this, PlayState* play) {
 }
 
 void EnXc_SetColossusAppearSFX(EnXc* this, PlayState* play) {
-    static Vec3f sXyzDist;
     s16 sceneId;
 
     if (gSaveContext.sceneLayer == 4) {
@@ -445,14 +449,14 @@ void EnXc_SetColossusAppearSFX(EnXc* this, PlayState* play) {
             if (csCurFrame == 119) {
                 Vec3f pos = { -611.0f, 728.0f, -2.0f };
 
-                SkinMatrix_Vec3fMtxFMultXYZW(&play->viewProjectionMtxF, &pos, &sXyzDist, wDest);
-                Sfx_PlaySfxAtPos(&sXyzDist, NA_SE_EV_JUMP_CONC);
+                SkinMatrix_Vec3fMtxFMultXYZW(&play->viewProjectionMtxF, &pos, &sSfxPos, wDest);
+                Sfx_PlaySfxAtPos(&sSfxPos, NA_SE_EV_JUMP_CONC);
             } else if (csCurFrame == 164) {
                 Vec3f pos = { -1069.0f, 38.0f, 0.0f };
                 s32 pad;
 
-                SkinMatrix_Vec3fMtxFMultXYZW(&play->viewProjectionMtxF, &pos, &sXyzDist, wDest);
-                Sfx_PlaySfxAtPos(&sXyzDist, NA_SE_PL_WALK_GROUND + SURFACE_SFX_OFFSET_STONE);
+                SkinMatrix_Vec3fMtxFMultXYZW(&play->viewProjectionMtxF, &pos, &sSfxPos, wDest);
+                Sfx_PlaySfxAtPos(&sSfxPos, NA_SE_PL_WALK_GROUND + SURFACE_SFX_OFFSET_STONE);
             }
         }
     }
@@ -466,8 +470,6 @@ void func_80B3D118(PlayState* play) {
     }
 }
 
-static Vec3f D_80B42DA0;
-
 void EnXc_SetColossusWindSFX(PlayState* play) {
     if (gSaveContext.sceneLayer == 4) {
         static s32 D_80B41D90 = 0;
@@ -1394,7 +1396,10 @@ void func_80B3F3D8(void) {
     Sfx_PlaySfxCentered2(NA_SE_PL_SKIP);
 }
 
+#pragma increment_block_number 20
+
 void EnXc_PlayDiveSFX(Vec3f* src, PlayState* play) {
+    static Vec3f D_80B42DA0;
     f32 wDest[2];
 
     SkinMatrix_Vec3fMtxFMultXYZW(&play->viewProjectionMtxF, src, &D_80B42DA0, wDest);
diff --git a/src/overlays/actors/ovl_Fishing/z_fishing.c b/src/overlays/actors/ovl_Fishing/z_fishing.c
index 7181accf39..bf1e59de7f 100644
--- a/src/overlays/actors/ovl_Fishing/z_fishing.c
+++ b/src/overlays/actors/ovl_Fishing/z_fishing.c
@@ -10,6 +10,9 @@
 #include "assets/objects/object_fish/object_fish.h"
 #include "terminal.h"
 
+// For retail BSS ordering, the block number of sStreamSfxProjectedPos must be 0.
+#pragma increment_block_number 201
+
 #define FLAGS ACTOR_FLAG_4
 
 #define WATER_SURFACE_Y(play) play->colCtx.colHeader->waterBoxes->ySurface
diff --git a/src/overlays/actors/ovl_player_actor/z_player.c b/src/overlays/actors/ovl_player_actor/z_player.c
index 6f4cc33c3f..fd2cc36c48 100644
--- a/src/overlays/actors/ovl_player_actor/z_player.c
+++ b/src/overlays/actors/ovl_player_actor/z_player.c
@@ -353,11 +353,24 @@ void Player_Action_80850E84(Player* this, PlayState* play);
 void Player_Action_CsAction(Player* this, PlayState* play);
 
 // .bss part 1
+
+// For retail BSS ordering, the block number of sDogSpawnPos in Player_Update
+// must be between 0 and 53 inclusive.
+
+// TODO: There's probably a way to do this with less padding by spreading the variables out and moving
+// data around. It would be easier if we had more options for controlling BSS ordering in debug.
+#pragma increment_block_number 30
+
 static s32 D_80858AA0;
+
+#pragma increment_block_number 250
+
 static s32 D_80858AA4;
 static Vec3f sInteractWallCheckResult;
 static Input* sControlInput;
 
+#pragma increment_block_number 50
+
 // .data
 
 static u8 sUpperBodyLimbCopyMap[PLAYER_LIMB_MAX] = {
@@ -11376,8 +11389,6 @@ typedef struct {
 
 static BunnyEarKinematics sBunnyEarKinematics;
 
-static Vec3s D_80858AD8[25];
-
 static Gfx* sMaskDlists[PLAYER_MASK_MAX - 1] = {
     gLinkChildKeatonMaskDL, gLinkChildSkullMaskDL, gLinkChildSpookyMaskDL, gLinkChildBunnyHoodDL,
     gLinkChildGoronMaskDL,  gLinkChildZoraMaskDL,  gLinkChildGerudoMaskDL, gLinkChildMaskOfTruthDL,
@@ -14080,6 +14091,7 @@ void Player_Action_80850AEC(Player* this, PlayState* play) {
 
 void Player_Action_80850C68(Player* this, PlayState* play) {
     if ((this->av2.actionVar2 != 0) && ((this->unk_858 != 0.0f) || (this->unk_85C != 0.0f))) {
+        static Vec3s D_80858AD8[25];
         f32 updateScale = R_UPDATE_RATE * 0.5f;
 
         this->skelAnime.curFrame += this->skelAnime.playSpeed * updateScale;
diff --git a/tools/disasm/gc-eu-mq/variables.txt b/tools/disasm/gc-eu-mq/variables.txt
index 7be392076d..7bcb196227 100644
--- a/tools/disasm/gc-eu-mq/variables.txt
+++ b/tools/disasm/gc-eu-mq/variables.txt
@@ -41,7 +41,7 @@ sIdleThread = 0x80006590; // size:0x1B0
 sIdleThreadStack = 0x80006740; // size:0x400
 sIdleThreadInfo = 0x80006B40; // size:0x1C
 sBootThreadStack = 0x80006B60; // size:0x400
-gMainThread = 0x80006F60; // size:0x1B0
+sMainThread = 0x80006F60; // size:0x1B0
 sMainStack = 0x80007110; // size:0x900
 sMainStackInfo = 0x80007A10; // size:0x1C
 sPiMgrCmdBuff = 0x80007A30; // size:0xC8
@@ -570,9 +570,9 @@ sKaleidoScopeUpdateFunc = 0x8011B3C0; // size:0x4
 sKaleidoScopeDrawFunc = 0x8011B3C4; // size:0x4
 gBossMarkScale = 0x8011B3C8; // size:0x4
 gLoadedPauseMarkDataTable = 0x8011B3D0; // size:0x4
-sTransitionTile = 0x8011B3E0; // size:0xE0
+gTransitionTile = 0x8011B3E0; // size:0xE0
 gTransitionTileState = 0x8011B4C0; // size:0x4
-sPlayVisMono = 0x8011B4C8; // size:0x18
+gPlayVisMono = 0x8011B4C8; // size:0x18
 gVisMonoColor = 0x8011B4E0; // size:0x4
 D_801664D0 = 0x8011B4F0; // size:0x20
 sVisCvg = 0x8011B510; // size:0x10
diff --git a/tools/ido_block_numbers.py b/tools/ido_block_numbers.py
index 8858481dc2..2deb64c79c 100755
--- a/tools/ido_block_numbers.py
+++ b/tools/ido_block_numbers.py
@@ -222,7 +222,7 @@ def find_compiler_command_line(filename, oot_version):
     for line in make_output:
         parts = line.split()
         if "-o" in parts and str(filename) in parts:
-            makefile_command_line = parts
+            compiler_command_line = parts
             found += 1
 
     if found != 1:
@@ -231,49 +231,29 @@ def find_compiler_command_line(filename, oot_version):
         )
         sys.exit(1)
 
-    # Assume command line is of the form:
-    # tools/reencode.sh [COMPILER] [COMPILER_ARGS]
-    compiler_command_line = makefile_command_line[1:]
-
     print(f'Command line: {" ".join(compiler_command_line)}', file=sys.stderr)
     return compiler_command_line
 
 
 def generate_symbol_table(command_line):
-    # Find source file in compiler arguments
-    source_file = None
-    args = []
-    for arg in command_line:
-        if arg.endswith(".c"):
-            source_file = Path(arg)
-        else:
-            args.append(arg)
-
-    if source_file is None:
-        raise Exception("No source file found")
-
-    source_contents = source_file.read_text()
+    # Assume command line is of the form:
+    # python3 tools/preprocess.py [COMPILER] [COMPILER_ARGS] [INPUT_FILE]
+    input_file = Path(command_line[-1])
+    rest = command_line[:-1]
 
-    stem = "reencode_tmp"
-    input_file = Path(f"{stem}.c")
+    stem = input_file.stem
     symbol_table_file = Path(f"{stem}.T")
     ucode_file = Path(f"{stem}.B")
 
     try:
-        # Write temporary file with #line directive to simulate asm-processor
-        with open(input_file, "w") as f:
-            f.write('#line 1 "{}"\n'.format(source_file))
-            f.write(source_contents)
-
         # Invoke compiler
         # -Hf stops compilation after cfe so we can inspect the symbol table
-        subprocess.run(args + ["-Hf", input_file], check=True)
+        subprocess.run(rest + ["-Hf", input_file], check=True)
 
         # Read symbol table
         return symbol_table_file.read_bytes()
     finally:
         # Cleanup
-        input_file.unlink(missing_ok=True)
         symbol_table_file.unlink(missing_ok=True)
         ucode_file.unlink(missing_ok=True)
 
diff --git a/tools/preprocess.py b/tools/preprocess.py
new file mode 100755
index 0000000000..d84c3eef21
--- /dev/null
+++ b/tools/preprocess.py
@@ -0,0 +1,66 @@
+#!/usr/bin/env python3
+
+# SPDX-FileCopyrightText: © 2024 ZeldaRET
+# SPDX-License-Identifier: CC0-1.0
+
+# Usage: preprocess.py [compile command minus input file...] [single input file]
+# Preprocess a C file to:
+# * Re-encode from UTF-8 to EUC-JP (the repo uses UTF-8 for text encoding, but
+#   the strings in the ROM are encoded in EUC-JP)
+# * Replace `#pragma increment_block_number N` with `N` fake structs for
+#   controlling BSS ordering
+
+from pathlib import Path
+import os
+import tempfile
+import subprocess
+import sys
+
+
+def fail(message):
+    print(message, file=sys.stderr)
+    sys.exit(1)
+
+
+def process_file(filename, input, output):
+    output.write(f'#line 1 "{filename}"\n')
+    for i, line in enumerate(input, start=1):
+        if line.startswith("#pragma increment_block_number"):
+            parts = line.split()
+            if len(parts) != 3:
+                fail(
+                    f"{filename}:{i}: increment_block_number must be followed by an integer"
+                )
+            try:
+                amount = int(parts[2])
+            except ValueError:
+                fail(
+                    f"{filename}:{i}: increment_block_number must be followed by an integer"
+                )
+            # Write fake structs for BSS ordering
+            for j in range(amount):
+                output.write(f"struct DummyStruct_{i:05}_{j:03};\n")
+            output.write(f'#line {i + 1} "{filename}"\n')
+        else:
+            output.write(line)
+
+
+def main():
+    filename = Path(sys.argv[-1])
+    with tempfile.TemporaryDirectory(prefix="oot_") as tmpdir:
+        tmpfile = Path(tmpdir) / filename.name
+
+        with open(filename, mode="r", encoding="utf-8") as input:
+            with open(tmpfile, mode="w", encoding="euc-jp") as output:
+                process_file(filename, input, output)
+
+        compile_command = sys.argv[1:-1] + ["-I", filename.parent, tmpfile]
+        process = subprocess.run(compile_command)
+        return process.returncode
+
+
+if __name__ == "__main__":
+    try:
+        sys.exit(main())
+    except KeyboardInterrupt:
+        sys.exit(1)
diff --git a/tools/reencode.sh b/tools/reencode.sh
deleted file mode 100755
index d027faf5b9..0000000000
--- a/tools/reencode.sh
+++ /dev/null
@@ -1,26 +0,0 @@
-#!/bin/bash
-# The repo uses UTF-8 for text encoding, but the strings in the ROM are encoded in EUC-JP.
-# This means for matching the source must be re-encoded to EUC-JP before IDO compiles it.
-# This is conceptually equivalent to `gcc --finput-charset=UTF-8 --fexec-charset=EUC-JP`,
-# except IDO has no equivalent arguments.
-# Usage: reencode.sh [compile command minus input file...] [single input file]
-
-set -e
-
-# The last argument, the input source file to be compiled
-srcfile="${@: -1}"
-
-# Create a temporary file, and remove it on script exit
-tempfile=`mktemp`_oot.c
-trap "rm -f $tempfile" EXIT
-
-# Re-encode from UTF-8 to EUC-JP
-{
-    printf '#line 1 "%s"\n' "$srcfile"  # linemarker
-    cat "$srcfile"
-} | iconv -f UTF-8 -t EUC-JP > "$tempfile"
-
-# All arguments but the last, forming the compile command
-# Also include the source file's directory to have the include path as if we compiled the original source.
-# Pass the EUC-JP encoded temporary file for compilation.
-"${@:1:$# - 1}" -I `dirname $srcfile` $tempfile