14 changed files with 166 additions and 138 deletions
--- a/10
+++ b/10
@ -9,8 +9,8 @@ endif
 include $(DEVKITARM)/3ds_rules

 export CITRO3D_MAJOR	:= 1
-export CITRO3D_MINOR	:= 7
-export CITRO3D_PATCH	:= 1
+export CITRO3D_MINOR	:= 5
+export CITRO3D_PATCH	:= 0

 VERSION	:=	$(CITRO3D_MAJOR).$(CITRO3D_MINOR).$(CITRO3D_PATCH)

@ -31,11 +31,11 @@ INCLUDES	:=	include
 #---------------------------------------------------------------------------------
 ARCH	:=	-march=armv6k -mtune=mpcore -mfloat-abi=hard -mtp=soft

-CFLAGS	:=	-g -Wall -Wno-sizeof-array-div -Werror -mword-relocations \
+CFLAGS	:=	-g -Wall -Werror -mword-relocations \
 			-ffunction-sections -fdata-sections \
 			$(ARCH) $(BUILD_CFLAGS)

-CFLAGS	+=	$(INCLUDE) -D__3DS__ -DCITRO3D_BUILD
+CFLAGS	+=	$(INCLUDE) -DARM11 -D_3DS -DCITRO3D_BUILD

 CXXFLAGS	:= $(CFLAGS) -fno-rtti -fno-exceptions -std=gnu++11

@ -114,7 +114,7 @@ debug:

 lib/libcitro3d.a : lib release $(SOURCES) $(INCLUDES)
 	@$(MAKE) BUILD=release OUTPUT=$(CURDIR)/$@ \
-	BUILD_CFLAGS="-DNDEBUG=1 -O2 -fomit-frame-pointer -fno-math-errno" \
+	BUILD_CFLAGS="-DNDEBUG=1 -O2 -fomit-frame-pointer" \
 	DEPSDIR=$(CURDIR)/release \
 	--no-print-directory -C release \
 	-f $(CURDIR)/Makefile
--- a/include/c3d/base.h
+++ b/include/c3d/base.h
@ -11,6 +11,7 @@ enum
 };

 bool C3D_Init(size_t cmdBufSize);
+void C3D_FlushAsync(void);
 void C3D_Fini(void);

 float C3D_GetCmdBufUsage(void);
@ -33,6 +34,23 @@ static inline void C3D_ImmDrawRestartPrim(void)
 	GPUCMD_AddWrite(GPUREG_RESTART_PRIMITIVE, 1);
 }

+static inline void C3D_FlushAwait(void)
+{
+	gspWaitForP3D();
+}
+
+static inline void C3D_Flush(void)
+{
+	C3D_FlushAsync();
+	C3D_FlushAwait();
+}
+
+static inline void C3D_VideoSync(void)
+{
+	gspWaitForEvent(GSPGPU_EVENT_VBlank0, false);
+	gfxSwapBuffersGpu();
+}
+
 // Fixed vertex attributes
 C3D_FVec* C3D_FixedAttribGetWritePtr(int id);

--- a/include/c3d/light.h
+++ b/include/c3d/light.h
@ -77,14 +77,6 @@ enum
 void C3D_LightEnvFresnel(C3D_LightEnv* env, GPU_FRESNELSEL selector);
 void C3D_LightEnvBumpMode(C3D_LightEnv* env, GPU_BUMPMODE mode);
 void C3D_LightEnvBumpSel(C3D_LightEnv* env, int texUnit);
-
-/**
- * @brief Configures whether to use the z component of the normal map.
- * @param[out] env    Pointer to light environment structure.
- * @param[in]  enable false if the z component is reconstructed from the xy components
- *                     of the normal map, true if the z component is taken from the normal map.
- */
-void C3D_LightEnvBumpNormalZ(C3D_LightEnv *env, bool enable);
 void C3D_LightEnvShadowMode(C3D_LightEnv* env, u32 mode);
 void C3D_LightEnvShadowSel(C3D_LightEnv* env, int texUnit);
 void C3D_LightEnvClampHighlights(C3D_LightEnv* env, bool clamp);
--- a/include/c3d/maths.h
+++ b/include/c3d/maths.h
@ -13,12 +13,7 @@
 * The one true circumference-to-radius ratio.
 * See http://tauday.com/tau-manifesto
 */
-#define M_TAU (6.28318530717958647692528676655900576)
-
-// Define the legacy circle constant as well
-#ifndef M_PI
-#define M_PI (M_TAU/2)
-#endif
+#define M_TAU (2*M_PI)

 /**
 * @brief Convert an angle from revolutions to radians
--- a/include/c3d/renderqueue.h
+++ b/include/c3d/renderqueue.h
@ -64,11 +64,6 @@ C3D_RenderTarget* C3D_RenderTargetCreateFromTex(C3D_Tex* tex, GPU_TEXFACE face,
 void C3D_RenderTargetDelete(C3D_RenderTarget* target);
 void C3D_RenderTargetSetOutput(C3D_RenderTarget* target, gfxScreen_t screen, gfx3dSide_t side, u32 transferFlags);

-static inline void C3D_RenderTargetDetachOutput(C3D_RenderTarget* target)
-{
-	C3D_RenderTargetSetOutput(NULL, target->screen, target->side, 0);
-}
-
 static inline void C3D_RenderTargetClear(C3D_RenderTarget* target, C3D_ClearBits clearBits, u32 clearColor, u32 clearDepth)
 {
 	C3D_FrameBufClear(&target->frameBuf, clearBits, clearColor, clearDepth);
--- a/include/c3d/texture.h
+++ b/include/c3d/texture.h
@ -41,7 +41,7 @@ typedef struct
 	};
 } C3D_Tex;

-typedef struct CTR_ALIGN(8)
+typedef struct ALIGN(8)
 {
 	u16 width;
 	u16 height;
--- a/include/c3d/types.h
+++ b/include/c3d/types.h
@ -1,5 +1,5 @@
 #pragma once
-#if defined(__3DS__) || defined(_3DS)
+#ifdef _3DS
 #include <3ds.h>
 #else
 #include <stdbool.h>
--- a/include/tex3ds.h
+++ b/include/tex3ds.h
@ -38,7 +38,7 @@ extern "C" {
 #endif

 /** @brief Subtexture
- *  @note If top < bottom, the subtexture is rotated 1/4 revolution counter-clockwise
+ *  @note If top > bottom, the subtexture is rotated 1/4 revolution counter-clockwise
 */
 typedef struct Tex3DS_SubTexture
 {
--- a/source/base.c
+++ b/source/base.c
@ -8,6 +8,14 @@ C3D_Context __C3D_Context;

 static aptHookCookie hookCookie;

+__attribute__((weak)) void C3Di_RenderQueueWaitDone(void)
+{
+}
+
+__attribute__((weak)) void C3Di_RenderQueueExit(void)
+{
+}
+
 __attribute__((weak)) void C3Di_LightEnvUpdate(C3D_LightEnv* env)
 {
 	(void)env;
@ -42,12 +50,10 @@ static void C3Di_AptEventHook(APT_HookType hookType, C3D_UNUSED void* param)
 		case APTHOOK_ONSUSPEND:
 		{
 			C3Di_RenderQueueWaitDone();
-			C3Di_RenderQueueDisableVBlank();
 			break;
 		}
 		case APTHOOK_ONRESTORE:
 		{
-			C3Di_RenderQueueEnableVBlank();
 			ctx->flags |= C3DiF_AttrInfo | C3DiF_BufInfo | C3DiF_Effect | C3DiF_FrameBuf
 				| C3DiF_Viewport | C3DiF_Scissor | C3DiF_Program | C3DiF_VshCode | C3DiF_GshCode
 				| C3DiF_TexAll | C3DiF_TexEnvBuf | C3DiF_TexEnvAll | C3DiF_LightEnv | C3DiF_Gas;
@ -96,6 +102,10 @@ bool C3D_Init(size_t cmdBufSize)
 		return false;
 	}

+	GPUCMD_SetBuffer(ctx->cmdBuf, ctx->cmdBufSize, 0);
+	GX_BindQueue(&ctx->gxQueue);
+	gxCmdQueueRun(&ctx->gxQueue);
+
 	ctx->flags = C3DiF_Active | C3DiF_TexEnvBuf | C3DiF_TexEnvAll | C3DiF_Effect | C3DiF_TexStatus | C3DiF_TexAll;

 	// TODO: replace with direct struct access
@ -127,7 +137,6 @@ bool C3D_Init(size_t cmdBufSize)
 	ctx->fixedAttribDirty = 0;
 	ctx->fixedAttribEverDirty = 0;

-	C3Di_RenderQueueInit();
 	aptHook(&hookCookie, C3Di_AptEventHook, NULL);

 	return true;
@ -327,6 +336,25 @@ bool C3Di_SplitFrame(u32** pBuf, u32* pSize)
 	return true;
 }

+void C3D_FlushAsync(void)
+{
+	C3D_Context* ctx = C3Di_GetContext();
+
+	if (!(ctx->flags & C3DiF_Active))
+		return;
+
+	u32* cmdBuf;
+	u32 cmdBufSize;
+	C3Di_SplitFrame(&cmdBuf, &cmdBufSize);
+	GPUCMD_SetBuffer(ctx->cmdBuf, ctx->cmdBufSize, 0);
+
+	//take advantage of GX_FlushCacheRegions to flush gsp heap
+	extern u32 __ctru_linear_heap;
+	extern u32 __ctru_linear_heap_size;
+	GX_FlushCacheRegions(cmdBuf, cmdBufSize*4, (u32 *) __ctru_linear_heap, __ctru_linear_heap_size, NULL, 0);
+	GX_ProcessCommandList(cmdBuf, cmdBufSize*4, 0x0);
+}
+
 float C3D_GetCmdBufUsage(void)
 {
 	return C3Di_GetContext()->cmdBufUsage;
@ -339,8 +367,11 @@ void C3D_Fini(void)
 	if (!(ctx->flags & C3DiF_Active))
 		return;

-	aptUnhook(&hookCookie);
 	C3Di_RenderQueueExit();
+	aptUnhook(&hookCookie);
+	gxCmdQueueStop(&ctx->gxQueue);
+	gxCmdQueueWait(&ctx->gxQueue, -1);
+	GX_BindQueue(NULL);
 	free(ctx->gxQueue.entries);
 	linearFree(ctx->cmdBuf);
 	ctx->flags = 0;
--- a/source/internal.h
+++ b/source/internal.h
@ -124,18 +124,6 @@ static inline bool C3Di_TexIs2D(C3D_Tex* tex)
 	return !typeIsCube(C3D_TexGetType(tex));
 }

-static inline bool addrIsVRAM(const void* addr)
-{
-	u32 vaddr = (u32)addr;
-	return vaddr >= OS_VRAM_VADDR && vaddr < OS_VRAM_VADDR + OS_VRAM_SIZE;
-}
-
-static inline vramAllocPos addrGetVRAMBank(const void* addr)
-{
-	u32 vaddr = (u32)addr;
-	return vaddr < OS_VRAM_VADDR + OS_VRAM_SIZE/2 ? VRAM_ALLOC_A : VRAM_ALLOC_B;
-}
-
 void C3Di_UpdateContext(void);
 void C3Di_AttrInfoBind(C3D_AttrInfo* info);
 void C3Di_BufInfoBind(C3D_BufInfo* info);
@ -152,9 +140,3 @@ void C3Di_LoadShaderUniforms(shaderInstance_s* si);
 void C3Di_ClearShaderUniforms(GPU_SHADER_TYPE type);

 bool C3Di_SplitFrame(u32** pBuf, u32* pSize);
-
-void C3Di_RenderQueueInit(void);
-void C3Di_RenderQueueExit(void);
-void C3Di_RenderQueueWaitDone(void);
-void C3Di_RenderQueueEnableVBlank(void);
-void C3Di_RenderQueueDisableVBlank(void);
--- a/source/lightenv.c
+++ b/source/lightenv.c
@ -250,14 +250,6 @@ void C3D_LightEnvBumpSel(C3D_LightEnv* env, int texUnit)
 	env->flags |= C3DF_LightEnv_Dirty;
 }

-void C3D_LightEnvBumpNormalZ(C3D_LightEnv *env, bool usez) {
-	if (usez)
-		env->conf.config[0] |= BIT(30);
-	else
-		env->conf.config[0] &= ~BIT(30);
-	env->flags |= C3DF_LightEnv_Dirty;
-}
-
 void C3D_LightEnvShadowMode(C3D_LightEnv* env, u32 mode)
 {
 	mode &= 0xF<<16;
--- a/source/renderqueue.c
+++ b/source/renderqueue.c
@ -8,16 +8,22 @@ static C3D_RenderTarget *linkedTarget[3];

 static TickCounter gpuTime, cpuTime;

+#define STAGE_HAS_TRANSFER(n)   BIT(0+(n))
+#define STAGE_HAS_ANY_TRANSFER  (7<<0)
+#define STAGE_NEED_TRANSFER(n)  BIT(3+(n))
+#define STAGE_NEED_TOP_TRANSFER (STAGE_NEED_TRANSFER(0)|STAGE_NEED_TRANSFER(1))
+#define STAGE_NEED_BOT_TRANSFER STAGE_NEED_TRANSFER(2)
+#define STAGE_WAIT_TRANSFER     BIT(6)
+
+static bool initialized;
 static bool inFrame, inSafeTransfer, measureGpuTime;
-static bool needSwapTop, needSwapBot, isTopStereo;
+static u8 frameStage;
 static float framerate = 60.0f;
 static float framerateCounter[2] = { 60.0f, 60.0f };
 static u32 frameCounter[2];
 static void (* frameEndCb)(void*);
 static void* frameEndCbData;

-static void C3Di_RenderTargetDestroy(C3D_RenderTarget* target);
-
 static bool framerateLimit(int id)
 {
 	framerateCounter[id] -= framerate;
@ -31,12 +37,44 @@ static bool framerateLimit(int id)

 static void onVBlank0(C3D_UNUSED void* unused)
 {
+	if (frameStage & STAGE_NEED_TOP_TRANSFER)
+	{
+		C3D_RenderTarget *left = linkedTarget[0], *right = linkedTarget[1];
+		if (left && !(frameStage&STAGE_NEED_TRANSFER(0)))
+			left = NULL;
+		if (right && !(frameStage&STAGE_NEED_TRANSFER(1)))
+			right = NULL;
+		if (gfxIs3D() && !right)
+			right = left;
+
+		frameStage &= ~STAGE_NEED_TOP_TRANSFER;
+		if (left || right)
+		{
+			frameStage |= STAGE_WAIT_TRANSFER;
+			if (left)
+				C3D_FrameBufTransfer(&left->frameBuf, GFX_TOP, GFX_LEFT, left->transferFlags);
+			if (right)
+				C3D_FrameBufTransfer(&right->frameBuf, GFX_TOP, GFX_RIGHT, right->transferFlags);
+			gfxConfigScreen(GFX_TOP, false);
+		}
+	}
 	if (framerateLimit(0))
 		frameCounter[0]++;
 }

 static void onVBlank1(C3D_UNUSED void* unused)
 {
+	if (frameStage & STAGE_NEED_BOT_TRANSFER)
+	{
+		frameStage &= ~STAGE_NEED_BOT_TRANSFER;
+		C3D_RenderTarget* target = linkedTarget[2];
+		if (target)
+		{
+			frameStage |= STAGE_WAIT_TRANSFER;
+			C3D_FrameBufTransfer(&target->frameBuf, GFX_BOTTOM, GFX_LEFT, target->transferFlags);
+			gfxConfigScreen(GFX_BOTTOM, false);
+		}
+	}
 	if (framerateLimit(1))
 		frameCounter[1]++;
 }
@ -57,18 +95,12 @@ static void onQueueFinish(gxCmdQueue_s* queue)
 			gxCmdQueueClear(queue);
 		}
 	}
+	else if (frameStage & STAGE_WAIT_TRANSFER)
+		frameStage &= ~STAGE_WAIT_TRANSFER;
 	else
 	{
-		if (needSwapTop)
-		{
-			gfxScreenSwapBuffers(GFX_TOP, isTopStereo);
-			needSwapTop = false;
-		}
-		if (needSwapBot)
-		{
-			gfxScreenSwapBuffers(GFX_BOTTOM, false);
-			needSwapBot = false;
-		}
+		u8 needs = frameStage & STAGE_HAS_ANY_TRANSFER;
+		frameStage = (frameStage&~STAGE_HAS_ANY_TRANSFER) | (needs<<3);
 	}
 }

@ -94,60 +126,72 @@ static bool C3Di_WaitAndClearQueue(s64 timeout)
 	gxCmdQueue_s* queue = &C3Di_GetContext()->gxQueue;
 	if (!gxCmdQueueWait(queue, timeout))
 		return false;
+	if (timeout==0 && frameStage)
+		return false;
+	while (frameStage)
+		gspWaitForAnyEvent();
 	gxCmdQueueStop(queue);
 	gxCmdQueueClear(queue);
 	return true;
 }

-void C3Di_RenderQueueEnableVBlank(void)
+static void C3Di_RenderQueueInit(void)
 {
 	gspSetEventCallback(GSPGPU_EVENT_VBlank0, onVBlank0, NULL, false);
 	gspSetEventCallback(GSPGPU_EVENT_VBlank1, onVBlank1, NULL, false);
+	gxCmdQueueSetCallback(&C3Di_GetContext()->gxQueue, onQueueFinish, NULL);
 }

-void C3Di_RenderQueueDisableVBlank(void)
-{
-	gspSetEventCallback(GSPGPU_EVENT_VBlank0, NULL, NULL, false);
-	gspSetEventCallback(GSPGPU_EVENT_VBlank1, NULL, NULL, false);
-}
-
-void C3Di_RenderQueueInit(void)
-{
-	C3D_Context* ctx = C3Di_GetContext();
-
-	C3Di_RenderQueueEnableVBlank();
-
-	GX_BindQueue(&ctx->gxQueue);
-	gxCmdQueueSetCallback(&ctx->gxQueue, onQueueFinish, NULL);
-	gxCmdQueueRun(&ctx->gxQueue);
-}
+static void C3Di_RenderTargetDestroy(C3D_RenderTarget* target);

 void C3Di_RenderQueueExit(void)
 {
 	int i;
 	C3D_RenderTarget *a, *next;

+	if (!initialized)
+		return;
+
 	C3Di_WaitAndClearQueue(-1);
-	gxCmdQueueSetCallback(&C3Di_GetContext()->gxQueue, NULL, NULL);
-	GX_BindQueue(NULL);
-
-	C3Di_RenderQueueDisableVBlank();
-
-	for (i = 0; i < 3; i ++)
-		linkedTarget[i] = NULL;
-
 	for (a = firstTarget; a; a = next)
 	{
 		next = a->next;
 		C3Di_RenderTargetDestroy(a);
 	}
+
+	gspSetEventCallback(GSPGPU_EVENT_VBlank0, NULL, NULL, false);
+	gspSetEventCallback(GSPGPU_EVENT_VBlank1, NULL, NULL, false);
+	gxCmdQueueSetCallback(&C3Di_GetContext()->gxQueue, NULL, NULL);
+
+	for (i = 0; i < 3; i ++)
+		linkedTarget[i] = NULL;
+
+	initialized = false;
 }

 void C3Di_RenderQueueWaitDone(void)
 {
+	if (!initialized)
+		return;
 	C3Di_WaitAndClearQueue(-1);
 }

+static bool checkRenderQueueInit(void)
+{
+	C3D_Context* ctx = C3Di_GetContext();
+
+	if (!(ctx->flags & C3DiF_Active))
+		return false;
+
+	if (!initialized)
+	{
+		C3Di_RenderQueueInit();
+		initialized = true;
+	}
+
+	return true;
+}
+
 float C3D_FrameRate(float fps)
 {
 	float old = framerate;
@ -162,17 +206,13 @@ float C3D_FrameRate(float fps)

 bool C3D_FrameBegin(u8 flags)
 {
-	C3D_Context* ctx = C3Di_GetContext();
 	if (inFrame) return false;
-
 	if (flags & C3D_FRAME_SYNCDRAW)
 		C3D_FrameSync();
 	if (!C3Di_WaitAndClearQueue((flags & C3D_FRAME_NONBLOCK) ? 0 : -1))
 		return false;
-
 	inFrame = true;
 	osTickCounterStart(&cpuTime);
-	GPUCMD_SetBuffer(ctx->cmdBuf, ctx->cmdBufSize, 0);
 	return true;
 }

@ -197,15 +237,13 @@ void C3D_FrameSplit(u8 flags)
 void C3D_FrameEnd(u8 flags)
 {
 	C3D_Context* ctx = C3Di_GetContext();
-	if (!inFrame) return;

 	if (frameEndCb)
 		frameEndCb(frameEndCbData);

 	C3D_FrameSplit(flags);
-	GPUCMD_SetBuffer(NULL, 0, 0);
-	osTickCounterUpdate(&cpuTime);
 	inFrame = false;
+	osTickCounterUpdate(&cpuTime);

 	// Flush the entire linear memory if the user did not explicitly mandate to flush the command list
 	if (!(flags & GX_CMDLIST_FLUSH))
@ -217,24 +255,16 @@ void C3D_FrameEnd(u8 flags)

 	int i;
 	C3D_RenderTarget* target;
-	isTopStereo = false;
 	for (i = 2; i >= 0; i --)
 	{
 		target = linkedTarget[i];
 		if (!target || !target->used)
 			continue;
 		target->used = false;
-		C3D_FrameBufTransfer(&target->frameBuf, target->screen, target->side, target->transferFlags);
-		if (target->screen == GFX_TOP)
-		{
-			needSwapTop = true;
-			if (target->side == GFX_RIGHT)
-				isTopStereo = true;
-		}
-		else if (target->screen == GFX_BOTTOM)
-			needSwapBot = true;
+		frameStage |= STAGE_HAS_TRANSFER(i);
 	}

+	GPUCMD_SetBuffer(ctx->cmdBuf, ctx->cmdBufSize, 0);
 	measureGpuTime = true;
 	osTickCounterStart(&gpuTime);
 	gxCmdQueueRun(&ctx->gxQueue);
@ -277,6 +307,8 @@ static void C3Di_RenderTargetFinishInit(C3D_RenderTarget* target)

 C3D_RenderTarget* C3D_RenderTargetCreate(int width, int height, GPU_COLORBUF colorFmt, C3D_DEPTHTYPE depthFmt)
 {
+	if (!checkRenderQueueInit()) goto _fail0;
+
 	GPU_DEPTHBUF depthFmtReal = GPU_RB_DEPTH16;
 	void* depthBuf = NULL;
 	void* colorBuf = vramAlloc(C3D_CalcColorBufSize(width,height,colorFmt));
@ -284,10 +316,7 @@ C3D_RenderTarget* C3D_RenderTargetCreate(int width, int height, GPU_COLORBUF col
 	if (C3D_DEPTHTYPE_OK(depthFmt))
 	{
 		depthFmtReal = C3D_DEPTHTYPE_VAL(depthFmt);
-		size_t depthSize = C3D_CalcDepthBufSize(width,height,depthFmtReal);
-		vramAllocPos vramBank = addrGetVRAMBank(colorBuf);
-		depthBuf = vramAllocAt(depthSize, vramBank ^ VRAM_ALLOC_ANY); // Attempt opposite bank first...
-		if (!depthBuf) depthBuf = vramAllocAt(depthSize, vramBank); // ... if that fails, attempt same bank
+		depthBuf = vramAlloc(C3D_CalcDepthBufSize(width,height,depthFmtReal));
 		if (!depthBuf) goto _fail1;
 	}

@ -316,7 +345,8 @@ _fail0:

 C3D_RenderTarget* C3D_RenderTargetCreateFromTex(C3D_Tex* tex, GPU_TEXFACE face, int level, C3D_DEPTHTYPE depthFmt)
 {
-	if (!addrIsVRAM(tex->data)) return NULL; // Render targets must be in VRAM
+	if (!checkRenderQueueInit()) return NULL;
+
 	C3D_RenderTarget* target = C3Di_RenderTargetNew();
 	if (!target) return NULL;

@ -326,10 +356,7 @@ C3D_RenderTarget* C3D_RenderTargetCreateFromTex(C3D_Tex* tex, GPU_TEXFACE face,
 	if (C3D_DEPTHTYPE_OK(depthFmt))
 	{
 		GPU_DEPTHBUF depthFmtReal = C3D_DEPTHTYPE_VAL(depthFmt);
-		size_t depthSize = C3D_CalcDepthBufSize(fb->width,fb->height,depthFmtReal);
-		vramAllocPos vramBank = addrGetVRAMBank(tex->data);
-		void* depthBuf = vramAllocAt(depthSize, vramBank ^ VRAM_ALLOC_ANY); // Attempt opposite bank first...
-		if (!depthBuf) depthBuf = vramAllocAt(depthSize, vramBank); // ... if that fails, attempt same bank
+		void* depthBuf = vramAlloc(C3D_CalcDepthBufSize(fb->width,fb->height,depthFmtReal));
 		if (!depthBuf)
 		{
 			free(target);
@ -362,10 +389,7 @@ void C3D_RenderTargetDelete(C3D_RenderTarget* target)
 {
 	if (inFrame)
 		svcBreak(USERBREAK_PANIC); // Shouldn't happen.
-	if (target->linked)
-		C3D_RenderTargetDetachOutput(target);
-	else
-		C3Di_WaitAndClearQueue(-1);
+	C3Di_WaitAndClearQueue(-1);
 	C3Di_RenderTargetDestroy(target);
 }

@ -375,19 +399,12 @@ void C3D_RenderTargetSetOutput(C3D_RenderTarget* target, gfxScreen_t screen, gfx
 	if (screen==GFX_BOTTOM) id = 2;
 	else if (side==GFX_RIGHT) id = 1;
 	if (linkedTarget[id])
-	{
 		linkedTarget[id]->linked = false;
-		if (!inFrame)
-			C3Di_WaitAndClearQueue(-1);
-	}
 	linkedTarget[id] = target;
-	if (target)
-	{
-		target->linked = true;
-		target->transferFlags = transferFlags;
-		target->screen = screen;
-		target->side = side;
-	}
+	target->linked = true;
+	target->transferFlags = transferFlags;
+	target->screen = screen;
+	target->side = side;
 }

 static void C3Di_SafeDisplayTransfer(u32* inadr, u32 indim, u32* outadr, u32 outdim, u32 flags)
--- a/source/texture.c
+++ b/source/texture.c
@ -30,6 +30,12 @@ static inline size_t fmtSize(GPU_TEXCOLOR fmt)
 	}
 }

+static inline bool addrIsVRAM(const void* addr)
+{
+	u32 vaddr = (u32)addr;
+	return vaddr >= 0x1F000000 && vaddr < 0x1F600000;
+}
+
 static inline bool checkTexSize(u32 size)
 {
 	if (size < 8 || size > 1024)
--- a/test/3ds/Makefile
+++ b/test/3ds/Makefile
@ -47,10 +47,10 @@ ICON            :=
 ARCH     := -march=armv6k -mtune=mpcore -mfloat-abi=hard -mtp=soft

 CFLAGS   := -g -Wall -O3 -mword-relocations \
-            -ffunction-sections \
+            -fomit-frame-pointer -ffunction-sections \
            $(ARCH)

-CFLAGS   +=  $(INCLUDE) -D__3DS__
+CFLAGS   +=  $(INCLUDE) -DARM11 -D_3DS

 CXXFLAGS := $(CFLAGS) -fno-rtti -std=gnu++11