diff --git a/booter_fc/arm9/source/nds_loader_arm9.c b/booter_fc/arm9/source/nds_loader_arm9.c
index 6655a6fda9..cb0da2de9c 100644
--- a/booter_fc/arm9/source/nds_loader_arm9.c
+++ b/booter_fc/arm9/source/nds_loader_arm9.c
@@ -390,42 +390,7 @@ int runNdsFile (const char* filename, int argc, const char** argv)  {
 		runNds9i(filename);
 	}
 
-	bool loadFromRam = (runNds9(filename) || (isDSiMode() && access("sd:/", F_OK) != 0));
+	const bool loadFromRam = (runNds9(filename) || (isDSiMode() && access("sd:/", F_OK) != 0));
 	
-	bool havedsiSD = (argv[0][0]=='s' && argv[0][1]=='d');
-	installBootStub(havedsiSD);
-
 	return runNds (load_bin, load_bin_size, st.st_ino, true, (memcmp(io_dldi_data->friendlyName, "Default", 7) != 0), loadFromRam, argc, argv);
 }
-
-
-bool installBootStub(bool havedsiSD) {
-#ifndef _NO_BOOTSTUB_
-	extern char *fake_heap_end;
-	struct __bootstub *bootstub = (struct __bootstub *)fake_heap_end;
-	u32 *bootloader = (u32*)(fake_heap_end+bootstub_bin_size);
-
-	memcpy(bootstub,bootstub_bin,bootstub_bin_size);
-	memcpy(bootloader,load_bin,load_bin_size);
-	bool ret = false;
-
-	bootloader[8] = isDSiMode();
-	if ( havedsiSD) {
-		ret = true;
-		bootloader[3] = 0; // don't dldi patch
-		bootloader[7] = 1; // use internal dsi SD code
-	} else {
-		ret = dldiPatchLoader((data_t*)bootloader, load_bin_size,false);
-	}
-	bootstub->arm9reboot = (VoidFn)(((u32)bootstub->arm9reboot)+fake_heap_end);
-	bootstub->arm7reboot = (VoidFn)(((u32)bootstub->arm7reboot)+fake_heap_end);
-	bootstub->bootsize = load_bin_size;
-
-	DC_FlushAll();
-
-	return ret;
-#else
-	return true;
-#endif
-
-}
diff --git a/settings/bootloader/load.ld b/settings/bootloader/load.ld
index 53f37287ea..566c3e6dd9 100644
--- a/settings/bootloader/load.ld
+++ b/settings/bootloader/load.ld
@@ -4,7 +4,7 @@ ENTRY(_start)
 
 MEMORY {
 
-	vram	: ORIGIN = 0x06000000, LENGTH = 128K
+	vram	: ORIGIN = 0x06008000, LENGTH = 128K - 32K
 }
 
 __vram_start	=	ORIGIN(vram);
diff --git a/settings/bootloader/source/boot.c b/settings/bootloader/source/boot.c
index 165be33b27..1e2a21f037 100644
--- a/settings/bootloader/source/boot.c
+++ b/settings/bootloader/source/boot.c
@@ -143,7 +143,7 @@ void passArgs_ARM7 (void) {
 	argDst = (u32*)((ARM9_DST + ARM9_LEN + 3) & ~3);		// Word aligned
 
 	if (ARM9_LEN > 0x380000) {
-		argDst = (u32*)0x02FFA000;
+		argDst = (u32*)(0x02FFC000 - ((argSize/4)*4));
 	} else
 	if (dsiMode && (*(u8*)(NDS_HEAD + 0x012) & BIT(1))) {
 		u32 ARM9i_DST = *((u32*)(TWL_HEAD + 0x1C8));
@@ -448,6 +448,19 @@ int main (void) {
 #ifndef NO_SDMMC
 	sdRead = (dsiSD && dsiMode);
 #endif
+	if (wantToPatchDLDI) {
+		toncset((u32*)0x06000000, 0, 0x8000);
+		if (*(u32*)0x02FF8004 == 0x69684320) { // DLDI ' Chi' string
+			const u16 dldiFileSize = 1 << *(u8*)0x02FF800D;
+			tonccpy((u32*)0x06000000, (u32*)0x02FF8000, (dldiFileSize > 0x4000) ? 0x4000 : dldiFileSize);
+			dldiClearBss();
+		} else if (*(u32*)0x02FF8000 == 0x53535A4C) { // LZ77 flag
+			dldiDecompressBinary();
+		} else {
+			return -1;
+		}
+	}
+
 	if (*(u32*)(0x2FFFD0C) == 0x4E44544C) {
 		limitedModeMemoryPit();
 		*(u32*)(0x2FFFD0C) = 0;
diff --git a/settings/bootloader/source/dldi_patcher.c b/settings/bootloader/source/dldi_patcher.c
index 061b955728..9aa55f0024 100644
--- a/settings/bootloader/source/dldi_patcher.c
+++ b/settings/bootloader/source/dldi_patcher.c
@@ -22,6 +22,7 @@
 #ifndef NO_DLDI
 #include <string.h>
 #include <nds.h>
+#include "common/tonccpy.h"
 #include "dldi_patcher.h"
 
 #define FIX_ALL	0x01
@@ -76,7 +77,7 @@ static addr_t quickFind (const data_t* data, const data_t* search, size_t dataLe
 	addr_t i;
 	addr_t dataChunkEnd = (addr_t)(dataLen / sizeof(int));
 
-	for ( i = 0; i < dataChunkEnd; i++) {
+	for (i = 0; i < dataChunkEnd; i++) {
 		if (dataChunk[i] == searchChunk) {
 			if ((i*sizeof(int) + searchLen) > dataLen) {
 				return -1;
@@ -107,7 +108,7 @@ static void demangleMagicString(data_t *dest, const data_t *src) {
 
 #define DEVICE_TYPE_DLDI 0x49444C44
 
-extern data_t _dldi_start[];
+static data_t* _dldi_start = (data_t*)0x06000000;
 
 bool dldiPatchBinary (data_t *binData, u32 binSize) {
 
@@ -164,7 +165,7 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 	// Remember how much space is actually reserved
 	pDH[DO_allocatedSpace] = pAH[DO_allocatedSpace];
 	// Copy the DLDI patch into the application
-	memcpy (pAH, pDH, dldiFileSize);
+	tonccpy (pAH, pDH, dldiFileSize);
 
 	// Fix the section pointers in the header
 	writeAddr (pAH, DO_text_start, readAddr (pAH, DO_text_start) + relocationOffset);
@@ -184,7 +185,7 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 	writeAddr (pAH, DO_shutdown, readAddr (pAH, DO_shutdown) + relocationOffset);
 
 	// Put the correct DLDI magic string back into the DLDI header
-	memcpy (pAH, dldiMagicString, sizeof (dldiMagicString));
+	tonccpy (pAH, dldiMagicString, sizeof (dldiMagicString));
 
 	if (pDH[DO_fixSections] & FIX_ALL) { 
 		// Search through and fix pointers within the data section of the file
@@ -215,9 +216,96 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 
 	if (pDH[DO_fixSections] & FIX_BSS) { 
 		// Initialise the BSS to 0
-		memset (&pAH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+		toncset (&pAH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
 	}
 
 	return true;
 }
+
+void dldiDecompressBinary (void) {
+	extern void LZ77_Decompress(u8* source, u8* destination);
+	LZ77_Decompress((u8*)0x02FF8004, (u8*)_dldi_start);
+
+	addr_t memOffset;			// Offset of DLDI after the file is loaded into memory
+	addr_t relocationOffset;	// Value added to all offsets within the patch to fix it properly
+	addr_t ddmemOffset;			// Original offset used in the DLDI file
+	addr_t ddmemStart;			// Start of range that offsets can be in the DLDI file
+	addr_t ddmemEnd;			// End of range that offsets can be in the DLDI file
+	addr_t ddmemSize;			// Size of range that offsets can be in the DLDI file
+	addr_t addrIter;
+
+	data_t *pDH = _dldi_start;
+
+	memOffset = 0x06000000;
+	ddmemOffset = readAddr (pDH, DO_text_start);
+	relocationOffset = memOffset - ddmemOffset;
+
+	ddmemStart = readAddr (pDH, DO_text_start);
+	ddmemSize = (1 << pDH[DO_driverSize]);
+	ddmemEnd = ddmemStart + ddmemSize;
+
+	// Fix the section pointers in the header
+	writeAddr (pDH, DO_text_start, readAddr (pDH, DO_text_start) + relocationOffset);
+	writeAddr (pDH, DO_data_end, readAddr (pDH, DO_data_end) + relocationOffset);
+	writeAddr (pDH, DO_glue_start, readAddr (pDH, DO_glue_start) + relocationOffset);
+	writeAddr (pDH, DO_glue_end, readAddr (pDH, DO_glue_end) + relocationOffset);
+	writeAddr (pDH, DO_got_start, readAddr (pDH, DO_got_start) + relocationOffset);
+	writeAddr (pDH, DO_got_end, readAddr (pDH, DO_got_end) + relocationOffset);
+	writeAddr (pDH, DO_bss_start, readAddr (pDH, DO_bss_start) + relocationOffset);
+	writeAddr (pDH, DO_bss_end, readAddr (pDH, DO_bss_end) + relocationOffset);
+	// Fix the function pointers in the header
+	writeAddr (pDH, DO_startup, readAddr (pDH, DO_startup) + relocationOffset);
+	writeAddr (pDH, DO_isInserted, readAddr (pDH, DO_isInserted) + relocationOffset);
+	writeAddr (pDH, DO_readSectors, readAddr (pDH, DO_readSectors) + relocationOffset);
+	writeAddr (pDH, DO_writeSectors, readAddr (pDH, DO_writeSectors) + relocationOffset);
+	writeAddr (pDH, DO_clearStatus, readAddr (pDH, DO_clearStatus) + relocationOffset);
+	writeAddr (pDH, DO_shutdown, readAddr (pDH, DO_shutdown) + relocationOffset);
+
+	if (pDH[DO_fixSections] & FIX_ALL) { 
+		// Search through and fix pointers within the data section of the file
+		for (addrIter = (readAddr(pDH, DO_text_start) - ddmemStart); addrIter < (readAddr(pDH, DO_data_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_GLUE) { 
+		// Search through and fix pointers within the glue section of the file
+		for (addrIter = (readAddr(pDH, DO_glue_start) - ddmemStart); addrIter < (readAddr(pDH, DO_glue_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_GOT) { 
+		// Search through and fix pointers within the Global Offset Table section of the file
+		for (addrIter = (readAddr(pDH, DO_got_start) - ddmemStart); addrIter < (readAddr(pDH, DO_got_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_BSS) { 
+		// Initialise the BSS to 0
+		toncset (&pDH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+	}
+}
+
+void dldiClearBss (void) {
+	data_t *pDH = _dldi_start;
+
+	addr_t ddmemSize = (1 << pDH[DO_driverSize]);			// Size of range that offsets can be in the DLDI file
+
+	if (ddmemSize <= 0x4000) return; // For <= 16KB DLDI file, BSS has already been cleared
+
+	addr_t ddmemStart = readAddr (pDH, DO_text_start);		// Start of range that offsets can be in the DLDI file
+
+	if (pDH[DO_fixSections] & FIX_BSS) { 
+		// Initialise the BSS to 0
+		toncset (&pDH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+	}
+}
 #endif
\ No newline at end of file
diff --git a/settings/bootloader/source/dldi_patcher.h b/settings/bootloader/source/dldi_patcher.h
index f54c531ba1..16ea22a859 100644
--- a/settings/bootloader/source/dldi_patcher.h
+++ b/settings/bootloader/source/dldi_patcher.h
@@ -28,5 +28,7 @@
 typedef signed int addr_t;
 typedef unsigned char data_t;
 bool dldiPatchBinary (data_t *binData, u32 binSize);
+void dldiDecompressBinary (void);
+void dldiClearBss (void);
 
 #endif // DLDI_PATCHER_H
diff --git a/settings/bootloader/source/io_dldi.s b/settings/bootloader/source/io_dldi.s
index 08ac2cf002..82147a9140 100644
--- a/settings/bootloader/source/io_dldi.s
+++ b/settings/bootloader/source/io_dldi.s
@@ -32,73 +32,6 @@
 
 
 _dldi_start:
-#ifndef NO_DLDI
-
-@---------------------------------------------------------------------------------
-@ Driver patch file standard header -- 16 bytes
-#ifdef STANDARD_DLDI
-	.word	0xBF8DA5ED		@ Magic number to identify this region
-#else
-	.word	0xBF8DA5EE		@ Magic number to identify this region
-#endif
-	.asciz	" Chishm"		@ Identifying Magic string (8 bytes with null terminator)
-	.byte	0x01			@ Version number
-	.byte	0x0f		@ 32KiB	@ Log [base-2] of the size of this driver in bytes.
-	.byte	0x00			@ Sections to fix
-	.byte 	0x0f		@ 32KiB	@ Log [base-2] of the allocated space in bytes.
-	
-@---------------------------------------------------------------------------------
-@ Text identifier - can be anything up to 47 chars + terminating null -- 16 bytes
-	.align	4
-	.asciz "Loader (No interface)"
-
-@---------------------------------------------------------------------------------
-@ Offsets to important sections within the data	-- 32 bytes
-	.align	6
-	.word   _dldi_start		@ data start
-	.word   _dldi_end		@ data end
-	.word	0x00000000		@ Interworking glue start	-- Needs address fixing
-	.word	0x00000000		@ Interworking glue end
-	.word   0x00000000		@ GOT start			-- Needs address fixing
-	.word   0x00000000		@ GOT end
-	.word   0x00000000		@ bss start			-- Needs setting to zero
-	.word   0x00000000		@ bss end
-@---------------------------------------------------------------------------------
-@ IO_INTERFACE data -- 32 bytes
-_io_dldi:
-	.ascii	"DLDI"				@ ioType
-	.word	0x00000000			@ Features
-	.word	_DLDI_startup			@ 
-	.word	_DLDI_isInserted		@ 
-	.word	_DLDI_readSectors		@   Function pointers to standard device driver functions
-	.word	_DLDI_writeSectors		@ 
-	.word	_DLDI_clearStatus		@ 
-	.word	_DLDI_shutdown			@ 
-
-	
-@---------------------------------------------------------------------------------
-
-_DLDI_startup:
-_DLDI_isInserted:
-_DLDI_readSectors:
-_DLDI_writeSectors:
-_DLDI_clearStatus:
-_DLDI_shutdown:
-	mov		r0, #0x00		@ Return false for every function
-	bx		lr
-
-
-
-@---------------------------------------------------------------------------------
-	.align
-	.pool
-
-	.space (_dldi_start + 32768) - .	@ Fill to 32KiB
-
-_dldi_end:
-	.end
-@---------------------------------------------------------------------------------
-#else
 @---------------------------------------------------------------------------------
 @ IO_INTERFACE data -- 32 bytes
 _io_dldi:
@@ -111,14 +44,21 @@ _io_dldi:
 	.word	_DLDI_clearStatus		@
 	.word	_DLDI_shutdown			@
 
-	_DLDI_startup:
+_DLDI_startup:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x68]
 _DLDI_isInserted:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x6C]
 _DLDI_readSectors:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x70]
 _DLDI_writeSectors:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x74]
 _DLDI_clearStatus:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x78]
 _DLDI_shutdown:
-	mov		r0, #0x00		@ Return false for every function
-	bx		lr
-
-
-#endif
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x7C]
diff --git a/settings/bootloader/source/lzss.c b/settings/bootloader/source/lzss.c
new file mode 100644
index 0000000000..d4a113b3d9
--- /dev/null
+++ b/settings/bootloader/source/lzss.c
@@ -0,0 +1,27 @@
+#include <nds/ndstypes.h>
+#include <string.h>
+
+void LZ77_Decompress(u8* source, u8* destination) {
+	u32 leng = (source[1] | (source[2] << 8) | (source[3] << 16));
+	int Offs = 4;
+	int dstoffs = 0;
+	while (true) {
+		u8 header = source[Offs++];
+		for (int i = 0; i < 8; i++) {
+			if ((header & 0x80) == 0) destination[dstoffs++] = source[Offs++];
+			else
+			{
+				u8 a = source[Offs++];
+				u8 b = source[Offs++];
+				int offs = (((a & 0xF) << 8) | b) + 1;
+				int length = (a >> 4) + 3;
+				for (int j = 0; j < length; j++) {
+					destination[dstoffs] = destination[dstoffs - offs];
+					dstoffs++;
+				}
+			}
+			if (dstoffs >= (int)leng) return;
+			header <<= 1;
+		}
+	}
+}
diff --git a/title/arm9/source/main.cpp b/title/arm9/source/main.cpp
index f78ee52cfe..f86a7ebd59 100644
--- a/title/arm9/source/main.cpp
+++ b/title/arm9/source/main.cpp
@@ -1820,6 +1820,18 @@ int titleMode(void)
 		4: Run temporary DSiWare
 	*/
 
+	if ((strcmp(io_dldi_data->friendlyName, "NAND FLASH CARD LIBFATNRIO") == 0) && (*(u32*)0x02FF8000 != 0x53535A4C)) {
+		FILE* file = fopen("nitro:/dldi/nrio.lz77", "rb");
+		fread((void*)0x02FF8004, 1, 0x3FFC, file);
+		fclose(file);
+
+		*(u32*)0x02FF8000 = 0x53535A4C;
+	}
+
+	#ifndef _NO_BOOTSTUB_
+	installBootStub(sdFound());
+	#endif
+
 	if (isDSiMode() && sdFound()) {
 		if ((access("sd:/_nds/bios9i.bin", F_OK) != 0) && (access("sd:/_nds/bios9i_part1.bin", F_OK) != 0)) {
 			extern char copyBuf[0x8000];
@@ -1884,7 +1896,7 @@ int titleMode(void)
 		}
 	}
 
-	bool is3DS = fifoGetValue32(FIFO_USER_05) != 0xD2;
+	const bool is3DS = fifoGetValue32(FIFO_USER_05) != 0xD2;
 
 	useTwlCfg = (REG_SCFG_EXT!=0 && (*(u8*)0x02000400 != 0) && (*(u8*)0x02000401 == 0) && (*(u8*)0x02000402 == 0) && (*(u8*)0x02000404 == 0) && (*(u8*)0x02000448 != 0));
 	bool nandMounted = false;
diff --git a/title/bootloader/load.ld b/title/bootloader/load.ld
index 53f37287ea..566c3e6dd9 100644
--- a/title/bootloader/load.ld
+++ b/title/bootloader/load.ld
@@ -4,7 +4,7 @@ ENTRY(_start)
 
 MEMORY {
 
-	vram	: ORIGIN = 0x06000000, LENGTH = 128K
+	vram	: ORIGIN = 0x06008000, LENGTH = 128K - 32K
 }
 
 __vram_start	=	ORIGIN(vram);
diff --git a/title/bootloader/source/boot.c b/title/bootloader/source/boot.c
index 5c33316c2d..2aa0b38400 100644
--- a/title/bootloader/source/boot.c
+++ b/title/bootloader/source/boot.c
@@ -148,7 +148,7 @@ void passArgs_ARM7 (void) {
 	argDst = (u32*)((ARM9_DST + ARM9_LEN + 3) & ~3);		// Word aligned
 
 	if (ARM9_LEN > 0x380000) {
-		argDst = (u32*)0x02FFA000;
+		argDst = (u32*)(0x02FFC000 - ((argSize/4)*4));
 	} else if (dsiMode && (*(u8*)(NDS_HEAD + 0x012) & BIT(1))) {
 		u32 ARM9i_DST = *((u32*)(TWL_HEAD + 0x1C8));
 		u32 ARM9i_LEN = *((u32*)(TWL_HEAD + 0x1CC));
@@ -635,6 +635,19 @@ int main (void) {
 #ifndef NO_SDMMC
 	sdRead = (dsiSD && dsiMode);
 #endif
+	if (wantToPatchDLDI) {
+		toncset((u32*)0x06000000, 0, 0x8000);
+		if (*(u32*)0x02FF8004 == 0x69684320) { // DLDI ' Chi' string
+			const u16 dldiFileSize = 1 << *(u8*)0x02FF800D;
+			tonccpy((u32*)0x06000000, (u32*)0x02FF8000, (dldiFileSize > 0x4000) ? 0x4000 : dldiFileSize);
+			dldiClearBss();
+		} else if (*(u32*)0x02FF8000 == 0x53535A4C) { // LZ77 flag
+			dldiDecompressBinary();
+		} else {
+			return -1;
+		}
+	}
+
 	if (*(u32*)(0x2FFFD0C) == 0x4E44544C) {
 		limitedModeMemoryPit();
 		*(u32*)(0x2FFFD0C) = 0;
diff --git a/title/bootloader/source/dldi_patcher.c b/title/bootloader/source/dldi_patcher.c
index 69f292e436..9aa55f0024 100644
--- a/title/bootloader/source/dldi_patcher.c
+++ b/title/bootloader/source/dldi_patcher.c
@@ -22,6 +22,7 @@
 #ifndef NO_DLDI
 #include <string.h>
 #include <nds.h>
+#include "common/tonccpy.h"
 #include "dldi_patcher.h"
 
 #define FIX_ALL	0x01
@@ -107,7 +108,7 @@ static void demangleMagicString(data_t *dest, const data_t *src) {
 
 #define DEVICE_TYPE_DLDI 0x49444C44
 
-extern data_t _dldi_start[];
+static data_t* _dldi_start = (data_t*)0x06000000;
 
 bool dldiPatchBinary (data_t *binData, u32 binSize) {
 
@@ -164,7 +165,7 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 	// Remember how much space is actually reserved
 	pDH[DO_allocatedSpace] = pAH[DO_allocatedSpace];
 	// Copy the DLDI patch into the application
-	memcpy (pAH, pDH, dldiFileSize);
+	tonccpy (pAH, pDH, dldiFileSize);
 
 	// Fix the section pointers in the header
 	writeAddr (pAH, DO_text_start, readAddr (pAH, DO_text_start) + relocationOffset);
@@ -184,7 +185,7 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 	writeAddr (pAH, DO_shutdown, readAddr (pAH, DO_shutdown) + relocationOffset);
 
 	// Put the correct DLDI magic string back into the DLDI header
-	memcpy (pAH, dldiMagicString, sizeof (dldiMagicString));
+	tonccpy (pAH, dldiMagicString, sizeof (dldiMagicString));
 
 	if (pDH[DO_fixSections] & FIX_ALL) { 
 		// Search through and fix pointers within the data section of the file
@@ -215,9 +216,96 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 
 	if (pDH[DO_fixSections] & FIX_BSS) { 
 		// Initialise the BSS to 0
-		memset (&pAH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+		toncset (&pAH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
 	}
 
 	return true;
 }
+
+void dldiDecompressBinary (void) {
+	extern void LZ77_Decompress(u8* source, u8* destination);
+	LZ77_Decompress((u8*)0x02FF8004, (u8*)_dldi_start);
+
+	addr_t memOffset;			// Offset of DLDI after the file is loaded into memory
+	addr_t relocationOffset;	// Value added to all offsets within the patch to fix it properly
+	addr_t ddmemOffset;			// Original offset used in the DLDI file
+	addr_t ddmemStart;			// Start of range that offsets can be in the DLDI file
+	addr_t ddmemEnd;			// End of range that offsets can be in the DLDI file
+	addr_t ddmemSize;			// Size of range that offsets can be in the DLDI file
+	addr_t addrIter;
+
+	data_t *pDH = _dldi_start;
+
+	memOffset = 0x06000000;
+	ddmemOffset = readAddr (pDH, DO_text_start);
+	relocationOffset = memOffset - ddmemOffset;
+
+	ddmemStart = readAddr (pDH, DO_text_start);
+	ddmemSize = (1 << pDH[DO_driverSize]);
+	ddmemEnd = ddmemStart + ddmemSize;
+
+	// Fix the section pointers in the header
+	writeAddr (pDH, DO_text_start, readAddr (pDH, DO_text_start) + relocationOffset);
+	writeAddr (pDH, DO_data_end, readAddr (pDH, DO_data_end) + relocationOffset);
+	writeAddr (pDH, DO_glue_start, readAddr (pDH, DO_glue_start) + relocationOffset);
+	writeAddr (pDH, DO_glue_end, readAddr (pDH, DO_glue_end) + relocationOffset);
+	writeAddr (pDH, DO_got_start, readAddr (pDH, DO_got_start) + relocationOffset);
+	writeAddr (pDH, DO_got_end, readAddr (pDH, DO_got_end) + relocationOffset);
+	writeAddr (pDH, DO_bss_start, readAddr (pDH, DO_bss_start) + relocationOffset);
+	writeAddr (pDH, DO_bss_end, readAddr (pDH, DO_bss_end) + relocationOffset);
+	// Fix the function pointers in the header
+	writeAddr (pDH, DO_startup, readAddr (pDH, DO_startup) + relocationOffset);
+	writeAddr (pDH, DO_isInserted, readAddr (pDH, DO_isInserted) + relocationOffset);
+	writeAddr (pDH, DO_readSectors, readAddr (pDH, DO_readSectors) + relocationOffset);
+	writeAddr (pDH, DO_writeSectors, readAddr (pDH, DO_writeSectors) + relocationOffset);
+	writeAddr (pDH, DO_clearStatus, readAddr (pDH, DO_clearStatus) + relocationOffset);
+	writeAddr (pDH, DO_shutdown, readAddr (pDH, DO_shutdown) + relocationOffset);
+
+	if (pDH[DO_fixSections] & FIX_ALL) { 
+		// Search through and fix pointers within the data section of the file
+		for (addrIter = (readAddr(pDH, DO_text_start) - ddmemStart); addrIter < (readAddr(pDH, DO_data_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_GLUE) { 
+		// Search through and fix pointers within the glue section of the file
+		for (addrIter = (readAddr(pDH, DO_glue_start) - ddmemStart); addrIter < (readAddr(pDH, DO_glue_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_GOT) { 
+		// Search through and fix pointers within the Global Offset Table section of the file
+		for (addrIter = (readAddr(pDH, DO_got_start) - ddmemStart); addrIter < (readAddr(pDH, DO_got_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_BSS) { 
+		// Initialise the BSS to 0
+		toncset (&pDH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+	}
+}
+
+void dldiClearBss (void) {
+	data_t *pDH = _dldi_start;
+
+	addr_t ddmemSize = (1 << pDH[DO_driverSize]);			// Size of range that offsets can be in the DLDI file
+
+	if (ddmemSize <= 0x4000) return; // For <= 16KB DLDI file, BSS has already been cleared
+
+	addr_t ddmemStart = readAddr (pDH, DO_text_start);		// Start of range that offsets can be in the DLDI file
+
+	if (pDH[DO_fixSections] & FIX_BSS) { 
+		// Initialise the BSS to 0
+		toncset (&pDH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+	}
+}
 #endif
\ No newline at end of file
diff --git a/title/bootloader/source/dldi_patcher.h b/title/bootloader/source/dldi_patcher.h
index f54c531ba1..16ea22a859 100644
--- a/title/bootloader/source/dldi_patcher.h
+++ b/title/bootloader/source/dldi_patcher.h
@@ -28,5 +28,7 @@
 typedef signed int addr_t;
 typedef unsigned char data_t;
 bool dldiPatchBinary (data_t *binData, u32 binSize);
+void dldiDecompressBinary (void);
+void dldiClearBss (void);
 
 #endif // DLDI_PATCHER_H
diff --git a/title/bootloader/source/io_dldi.s b/title/bootloader/source/io_dldi.s
index 08ac2cf002..82147a9140 100644
--- a/title/bootloader/source/io_dldi.s
+++ b/title/bootloader/source/io_dldi.s
@@ -32,73 +32,6 @@
 
 
 _dldi_start:
-#ifndef NO_DLDI
-
-@---------------------------------------------------------------------------------
-@ Driver patch file standard header -- 16 bytes
-#ifdef STANDARD_DLDI
-	.word	0xBF8DA5ED		@ Magic number to identify this region
-#else
-	.word	0xBF8DA5EE		@ Magic number to identify this region
-#endif
-	.asciz	" Chishm"		@ Identifying Magic string (8 bytes with null terminator)
-	.byte	0x01			@ Version number
-	.byte	0x0f		@ 32KiB	@ Log [base-2] of the size of this driver in bytes.
-	.byte	0x00			@ Sections to fix
-	.byte 	0x0f		@ 32KiB	@ Log [base-2] of the allocated space in bytes.
-	
-@---------------------------------------------------------------------------------
-@ Text identifier - can be anything up to 47 chars + terminating null -- 16 bytes
-	.align	4
-	.asciz "Loader (No interface)"
-
-@---------------------------------------------------------------------------------
-@ Offsets to important sections within the data	-- 32 bytes
-	.align	6
-	.word   _dldi_start		@ data start
-	.word   _dldi_end		@ data end
-	.word	0x00000000		@ Interworking glue start	-- Needs address fixing
-	.word	0x00000000		@ Interworking glue end
-	.word   0x00000000		@ GOT start			-- Needs address fixing
-	.word   0x00000000		@ GOT end
-	.word   0x00000000		@ bss start			-- Needs setting to zero
-	.word   0x00000000		@ bss end
-@---------------------------------------------------------------------------------
-@ IO_INTERFACE data -- 32 bytes
-_io_dldi:
-	.ascii	"DLDI"				@ ioType
-	.word	0x00000000			@ Features
-	.word	_DLDI_startup			@ 
-	.word	_DLDI_isInserted		@ 
-	.word	_DLDI_readSectors		@   Function pointers to standard device driver functions
-	.word	_DLDI_writeSectors		@ 
-	.word	_DLDI_clearStatus		@ 
-	.word	_DLDI_shutdown			@ 
-
-	
-@---------------------------------------------------------------------------------
-
-_DLDI_startup:
-_DLDI_isInserted:
-_DLDI_readSectors:
-_DLDI_writeSectors:
-_DLDI_clearStatus:
-_DLDI_shutdown:
-	mov		r0, #0x00		@ Return false for every function
-	bx		lr
-
-
-
-@---------------------------------------------------------------------------------
-	.align
-	.pool
-
-	.space (_dldi_start + 32768) - .	@ Fill to 32KiB
-
-_dldi_end:
-	.end
-@---------------------------------------------------------------------------------
-#else
 @---------------------------------------------------------------------------------
 @ IO_INTERFACE data -- 32 bytes
 _io_dldi:
@@ -111,14 +44,21 @@ _io_dldi:
 	.word	_DLDI_clearStatus		@
 	.word	_DLDI_shutdown			@
 
-	_DLDI_startup:
+_DLDI_startup:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x68]
 _DLDI_isInserted:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x6C]
 _DLDI_readSectors:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x70]
 _DLDI_writeSectors:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x74]
 _DLDI_clearStatus:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x78]
 _DLDI_shutdown:
-	mov		r0, #0x00		@ Return false for every function
-	bx		lr
-
-
-#endif
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x7C]
diff --git a/title/bootloader/source/load_crt0.s b/title/bootloader/source/load_crt0.s
index bb841834c1..aa1b015ae3 100644
--- a/title/bootloader/source/load_crt0.s
+++ b/title/bootloader/source/load_crt0.s
@@ -55,7 +55,7 @@ argStart:
 argSize:
 	.word	0x00000000
 dldiOffset:
-	.word	_dldi_start - _start
+	.word	0
 dsiSD:
 	.word	0
 dsiMode:
diff --git a/title/bootloader/source/lzss.c b/title/bootloader/source/lzss.c
new file mode 100644
index 0000000000..d4a113b3d9
--- /dev/null
+++ b/title/bootloader/source/lzss.c
@@ -0,0 +1,27 @@
+#include <nds/ndstypes.h>
+#include <string.h>
+
+void LZ77_Decompress(u8* source, u8* destination) {
+	u32 leng = (source[1] | (source[2] << 8) | (source[3] << 16));
+	int Offs = 4;
+	int dstoffs = 0;
+	while (true) {
+		u8 header = source[Offs++];
+		for (int i = 0; i < 8; i++) {
+			if ((header & 0x80) == 0) destination[dstoffs++] = source[Offs++];
+			else
+			{
+				u8 a = source[Offs++];
+				u8 b = source[Offs++];
+				int offs = (((a & 0xF) << 8) | b) + 1;
+				int length = (a >> 4) + 3;
+				for (int j = 0; j < length; j++) {
+					destination[dstoffs] = destination[dstoffs - offs];
+					dstoffs++;
+				}
+			}
+			if (dstoffs >= (int)leng) return;
+			header <<= 1;
+		}
+	}
+}
diff --git a/title/nitrofiles/dldi/nrio.lz77 b/title/nitrofiles/dldi/nrio.lz77
new file mode 100644
index 0000000000..fc556a403e
Binary files /dev/null and b/title/nitrofiles/dldi/nrio.lz77 differ
diff --git a/universal/bootloader_app/load.ld b/universal/bootloader_app/load.ld
index 53f37287ea..566c3e6dd9 100644
--- a/universal/bootloader_app/load.ld
+++ b/universal/bootloader_app/load.ld
@@ -4,7 +4,7 @@ ENTRY(_start)
 
 MEMORY {
 
-	vram	: ORIGIN = 0x06000000, LENGTH = 128K
+	vram	: ORIGIN = 0x06008000, LENGTH = 128K - 32K
 }
 
 __vram_start	=	ORIGIN(vram);
diff --git a/universal/bootloader_app/source/boot.c b/universal/bootloader_app/source/boot.c
index 314ac9bb49..17b5218703 100644
--- a/universal/bootloader_app/source/boot.c
+++ b/universal/bootloader_app/source/boot.c
@@ -388,6 +388,19 @@ int main (void) {
 #ifndef NO_SDMMC
 	sdRead = (dsiSD && dsiMode);
 #endif
+	if (wantToPatchDLDI) {
+		toncset((u32*)0x06000000, 0, 0x8000);
+		if (*(u32*)0x02FF8004 == 0x69684320) { // DLDI ' Chi' string
+			const u16 dldiFileSize = 1 << *(u8*)0x02FF800D;
+			tonccpy((u32*)0x06000000, (u32*)0x02FF8000, (dldiFileSize > 0x4000) ? 0x4000 : dldiFileSize);
+			dldiClearBss();
+		} else if (*(u32*)0x02FF8000 == 0x53535A4C) { // LZ77 flag
+			dldiDecompressBinary();
+		} else {
+			return -1;
+		}
+	}
+
 	u32 fileCluster = storedFileCluster;
 	if (!loadFromRam) {
 		// Init card
diff --git a/universal/bootloader_app/source/dldi_patcher.c b/universal/bootloader_app/source/dldi_patcher.c
index 69f292e436..9aa55f0024 100644
--- a/universal/bootloader_app/source/dldi_patcher.c
+++ b/universal/bootloader_app/source/dldi_patcher.c
@@ -22,6 +22,7 @@
 #ifndef NO_DLDI
 #include <string.h>
 #include <nds.h>
+#include "common/tonccpy.h"
 #include "dldi_patcher.h"
 
 #define FIX_ALL	0x01
@@ -107,7 +108,7 @@ static void demangleMagicString(data_t *dest, const data_t *src) {
 
 #define DEVICE_TYPE_DLDI 0x49444C44
 
-extern data_t _dldi_start[];
+static data_t* _dldi_start = (data_t*)0x06000000;
 
 bool dldiPatchBinary (data_t *binData, u32 binSize) {
 
@@ -164,7 +165,7 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 	// Remember how much space is actually reserved
 	pDH[DO_allocatedSpace] = pAH[DO_allocatedSpace];
 	// Copy the DLDI patch into the application
-	memcpy (pAH, pDH, dldiFileSize);
+	tonccpy (pAH, pDH, dldiFileSize);
 
 	// Fix the section pointers in the header
 	writeAddr (pAH, DO_text_start, readAddr (pAH, DO_text_start) + relocationOffset);
@@ -184,7 +185,7 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 	writeAddr (pAH, DO_shutdown, readAddr (pAH, DO_shutdown) + relocationOffset);
 
 	// Put the correct DLDI magic string back into the DLDI header
-	memcpy (pAH, dldiMagicString, sizeof (dldiMagicString));
+	tonccpy (pAH, dldiMagicString, sizeof (dldiMagicString));
 
 	if (pDH[DO_fixSections] & FIX_ALL) { 
 		// Search through and fix pointers within the data section of the file
@@ -215,9 +216,96 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 
 	if (pDH[DO_fixSections] & FIX_BSS) { 
 		// Initialise the BSS to 0
-		memset (&pAH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+		toncset (&pAH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
 	}
 
 	return true;
 }
+
+void dldiDecompressBinary (void) {
+	extern void LZ77_Decompress(u8* source, u8* destination);
+	LZ77_Decompress((u8*)0x02FF8004, (u8*)_dldi_start);
+
+	addr_t memOffset;			// Offset of DLDI after the file is loaded into memory
+	addr_t relocationOffset;	// Value added to all offsets within the patch to fix it properly
+	addr_t ddmemOffset;			// Original offset used in the DLDI file
+	addr_t ddmemStart;			// Start of range that offsets can be in the DLDI file
+	addr_t ddmemEnd;			// End of range that offsets can be in the DLDI file
+	addr_t ddmemSize;			// Size of range that offsets can be in the DLDI file
+	addr_t addrIter;
+
+	data_t *pDH = _dldi_start;
+
+	memOffset = 0x06000000;
+	ddmemOffset = readAddr (pDH, DO_text_start);
+	relocationOffset = memOffset - ddmemOffset;
+
+	ddmemStart = readAddr (pDH, DO_text_start);
+	ddmemSize = (1 << pDH[DO_driverSize]);
+	ddmemEnd = ddmemStart + ddmemSize;
+
+	// Fix the section pointers in the header
+	writeAddr (pDH, DO_text_start, readAddr (pDH, DO_text_start) + relocationOffset);
+	writeAddr (pDH, DO_data_end, readAddr (pDH, DO_data_end) + relocationOffset);
+	writeAddr (pDH, DO_glue_start, readAddr (pDH, DO_glue_start) + relocationOffset);
+	writeAddr (pDH, DO_glue_end, readAddr (pDH, DO_glue_end) + relocationOffset);
+	writeAddr (pDH, DO_got_start, readAddr (pDH, DO_got_start) + relocationOffset);
+	writeAddr (pDH, DO_got_end, readAddr (pDH, DO_got_end) + relocationOffset);
+	writeAddr (pDH, DO_bss_start, readAddr (pDH, DO_bss_start) + relocationOffset);
+	writeAddr (pDH, DO_bss_end, readAddr (pDH, DO_bss_end) + relocationOffset);
+	// Fix the function pointers in the header
+	writeAddr (pDH, DO_startup, readAddr (pDH, DO_startup) + relocationOffset);
+	writeAddr (pDH, DO_isInserted, readAddr (pDH, DO_isInserted) + relocationOffset);
+	writeAddr (pDH, DO_readSectors, readAddr (pDH, DO_readSectors) + relocationOffset);
+	writeAddr (pDH, DO_writeSectors, readAddr (pDH, DO_writeSectors) + relocationOffset);
+	writeAddr (pDH, DO_clearStatus, readAddr (pDH, DO_clearStatus) + relocationOffset);
+	writeAddr (pDH, DO_shutdown, readAddr (pDH, DO_shutdown) + relocationOffset);
+
+	if (pDH[DO_fixSections] & FIX_ALL) { 
+		// Search through and fix pointers within the data section of the file
+		for (addrIter = (readAddr(pDH, DO_text_start) - ddmemStart); addrIter < (readAddr(pDH, DO_data_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_GLUE) { 
+		// Search through and fix pointers within the glue section of the file
+		for (addrIter = (readAddr(pDH, DO_glue_start) - ddmemStart); addrIter < (readAddr(pDH, DO_glue_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_GOT) { 
+		// Search through and fix pointers within the Global Offset Table section of the file
+		for (addrIter = (readAddr(pDH, DO_got_start) - ddmemStart); addrIter < (readAddr(pDH, DO_got_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_BSS) { 
+		// Initialise the BSS to 0
+		toncset (&pDH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+	}
+}
+
+void dldiClearBss (void) {
+	data_t *pDH = _dldi_start;
+
+	addr_t ddmemSize = (1 << pDH[DO_driverSize]);			// Size of range that offsets can be in the DLDI file
+
+	if (ddmemSize <= 0x4000) return; // For <= 16KB DLDI file, BSS has already been cleared
+
+	addr_t ddmemStart = readAddr (pDH, DO_text_start);		// Start of range that offsets can be in the DLDI file
+
+	if (pDH[DO_fixSections] & FIX_BSS) { 
+		// Initialise the BSS to 0
+		toncset (&pDH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+	}
+}
 #endif
\ No newline at end of file
diff --git a/universal/bootloader_app/source/dldi_patcher.h b/universal/bootloader_app/source/dldi_patcher.h
index f54c531ba1..16ea22a859 100644
--- a/universal/bootloader_app/source/dldi_patcher.h
+++ b/universal/bootloader_app/source/dldi_patcher.h
@@ -28,5 +28,7 @@
 typedef signed int addr_t;
 typedef unsigned char data_t;
 bool dldiPatchBinary (data_t *binData, u32 binSize);
+void dldiDecompressBinary (void);
+void dldiClearBss (void);
 
 #endif // DLDI_PATCHER_H
diff --git a/universal/bootloader_app/source/io_dldi.s b/universal/bootloader_app/source/io_dldi.s
index 08ac2cf002..82147a9140 100644
--- a/universal/bootloader_app/source/io_dldi.s
+++ b/universal/bootloader_app/source/io_dldi.s
@@ -32,73 +32,6 @@
 
 
 _dldi_start:
-#ifndef NO_DLDI
-
-@---------------------------------------------------------------------------------
-@ Driver patch file standard header -- 16 bytes
-#ifdef STANDARD_DLDI
-	.word	0xBF8DA5ED		@ Magic number to identify this region
-#else
-	.word	0xBF8DA5EE		@ Magic number to identify this region
-#endif
-	.asciz	" Chishm"		@ Identifying Magic string (8 bytes with null terminator)
-	.byte	0x01			@ Version number
-	.byte	0x0f		@ 32KiB	@ Log [base-2] of the size of this driver in bytes.
-	.byte	0x00			@ Sections to fix
-	.byte 	0x0f		@ 32KiB	@ Log [base-2] of the allocated space in bytes.
-	
-@---------------------------------------------------------------------------------
-@ Text identifier - can be anything up to 47 chars + terminating null -- 16 bytes
-	.align	4
-	.asciz "Loader (No interface)"
-
-@---------------------------------------------------------------------------------
-@ Offsets to important sections within the data	-- 32 bytes
-	.align	6
-	.word   _dldi_start		@ data start
-	.word   _dldi_end		@ data end
-	.word	0x00000000		@ Interworking glue start	-- Needs address fixing
-	.word	0x00000000		@ Interworking glue end
-	.word   0x00000000		@ GOT start			-- Needs address fixing
-	.word   0x00000000		@ GOT end
-	.word   0x00000000		@ bss start			-- Needs setting to zero
-	.word   0x00000000		@ bss end
-@---------------------------------------------------------------------------------
-@ IO_INTERFACE data -- 32 bytes
-_io_dldi:
-	.ascii	"DLDI"				@ ioType
-	.word	0x00000000			@ Features
-	.word	_DLDI_startup			@ 
-	.word	_DLDI_isInserted		@ 
-	.word	_DLDI_readSectors		@   Function pointers to standard device driver functions
-	.word	_DLDI_writeSectors		@ 
-	.word	_DLDI_clearStatus		@ 
-	.word	_DLDI_shutdown			@ 
-
-	
-@---------------------------------------------------------------------------------
-
-_DLDI_startup:
-_DLDI_isInserted:
-_DLDI_readSectors:
-_DLDI_writeSectors:
-_DLDI_clearStatus:
-_DLDI_shutdown:
-	mov		r0, #0x00		@ Return false for every function
-	bx		lr
-
-
-
-@---------------------------------------------------------------------------------
-	.align
-	.pool
-
-	.space (_dldi_start + 32768) - .	@ Fill to 32KiB
-
-_dldi_end:
-	.end
-@---------------------------------------------------------------------------------
-#else
 @---------------------------------------------------------------------------------
 @ IO_INTERFACE data -- 32 bytes
 _io_dldi:
@@ -111,14 +44,21 @@ _io_dldi:
 	.word	_DLDI_clearStatus		@
 	.word	_DLDI_shutdown			@
 
-	_DLDI_startup:
+_DLDI_startup:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x68]
 _DLDI_isInserted:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x6C]
 _DLDI_readSectors:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x70]
 _DLDI_writeSectors:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x74]
 _DLDI_clearStatus:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x78]
 _DLDI_shutdown:
-	mov		r0, #0x00		@ Return false for every function
-	bx		lr
-
-
-#endif
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x7C]
diff --git a/universal/bootloader_app/source/lzss.c b/universal/bootloader_app/source/lzss.c
new file mode 100644
index 0000000000..d4a113b3d9
--- /dev/null
+++ b/universal/bootloader_app/source/lzss.c
@@ -0,0 +1,27 @@
+#include <nds/ndstypes.h>
+#include <string.h>
+
+void LZ77_Decompress(u8* source, u8* destination) {
+	u32 leng = (source[1] | (source[2] << 8) | (source[3] << 16));
+	int Offs = 4;
+	int dstoffs = 0;
+	while (true) {
+		u8 header = source[Offs++];
+		for (int i = 0; i < 8; i++) {
+			if ((header & 0x80) == 0) destination[dstoffs++] = source[Offs++];
+			else
+			{
+				u8 a = source[Offs++];
+				u8 b = source[Offs++];
+				int offs = (((a & 0xF) << 8) | b) + 1;
+				int length = (a >> 4) + 3;
+				for (int j = 0; j < length; j++) {
+					destination[dstoffs] = destination[dstoffs - offs];
+					dstoffs++;
+				}
+			}
+			if (dstoffs >= (int)leng) return;
+			header <<= 1;
+		}
+	}
+}
diff --git a/universal/bootloader_menu/load.ld b/universal/bootloader_menu/load.ld
index 53f37287ea..566c3e6dd9 100644
--- a/universal/bootloader_menu/load.ld
+++ b/universal/bootloader_menu/load.ld
@@ -4,7 +4,7 @@ ENTRY(_start)
 
 MEMORY {
 
-	vram	: ORIGIN = 0x06000000, LENGTH = 128K
+	vram	: ORIGIN = 0x06008000, LENGTH = 128K - 32K
 }
 
 __vram_start	=	ORIGIN(vram);
diff --git a/universal/bootloader_menu/source/boot.c b/universal/bootloader_menu/source/boot.c
index 0849a578e7..5fb22323c8 100644
--- a/universal/bootloader_menu/source/boot.c
+++ b/universal/bootloader_menu/source/boot.c
@@ -148,7 +148,7 @@ void passArgs_ARM7 (void) {
 	argDst = (u32*)((ARM9_DST + ARM9_LEN + 3) & ~3);		// Word aligned
 
 	if (ARM9_LEN > 0x380000) {
-		argDst = (u32*)0x02FFA000;
+		argDst = (u32*)(0x02FFC000 - ((argSize/4)*4));
 	} else
 	if (dsiMode && (*(u8*)(NDS_HEAD + 0x012) & BIT(1))) {
 		u32 ARM9i_DST = *((u32*)(TWL_HEAD + 0x1C8));
@@ -579,6 +579,19 @@ int main (void) {
 #ifndef NO_SDMMC
 	sdRead = (dsiSD && dsiMode);
 #endif
+	if (wantToPatchDLDI) {
+		toncset((u32*)0x06000000, 0, 0x8000);
+		if (*(u32*)0x02FF8004 == 0x69684320) { // DLDI ' Chi' string
+			const u16 dldiFileSize = 1 << *(u8*)0x02FF800D;
+			tonccpy((u32*)0x06000000, (u32*)0x02FF8000, (dldiFileSize > 0x4000) ? 0x4000 : dldiFileSize);
+			dldiClearBss();
+		} else if (*(u32*)0x02FF8000 == 0x53535A4C) { // LZ77 flag
+			dldiDecompressBinary();
+		} else {
+			return -1;
+		}
+	}
+
 	u32 fileCluster = storedFileCluster;
 	if (!loadFromRam) {
 		// Init card
diff --git a/universal/bootloader_menu/source/dldi_patcher.c b/universal/bootloader_menu/source/dldi_patcher.c
index 69f292e436..9aa55f0024 100644
--- a/universal/bootloader_menu/source/dldi_patcher.c
+++ b/universal/bootloader_menu/source/dldi_patcher.c
@@ -22,6 +22,7 @@
 #ifndef NO_DLDI
 #include <string.h>
 #include <nds.h>
+#include "common/tonccpy.h"
 #include "dldi_patcher.h"
 
 #define FIX_ALL	0x01
@@ -107,7 +108,7 @@ static void demangleMagicString(data_t *dest, const data_t *src) {
 
 #define DEVICE_TYPE_DLDI 0x49444C44
 
-extern data_t _dldi_start[];
+static data_t* _dldi_start = (data_t*)0x06000000;
 
 bool dldiPatchBinary (data_t *binData, u32 binSize) {
 
@@ -164,7 +165,7 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 	// Remember how much space is actually reserved
 	pDH[DO_allocatedSpace] = pAH[DO_allocatedSpace];
 	// Copy the DLDI patch into the application
-	memcpy (pAH, pDH, dldiFileSize);
+	tonccpy (pAH, pDH, dldiFileSize);
 
 	// Fix the section pointers in the header
 	writeAddr (pAH, DO_text_start, readAddr (pAH, DO_text_start) + relocationOffset);
@@ -184,7 +185,7 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 	writeAddr (pAH, DO_shutdown, readAddr (pAH, DO_shutdown) + relocationOffset);
 
 	// Put the correct DLDI magic string back into the DLDI header
-	memcpy (pAH, dldiMagicString, sizeof (dldiMagicString));
+	tonccpy (pAH, dldiMagicString, sizeof (dldiMagicString));
 
 	if (pDH[DO_fixSections] & FIX_ALL) { 
 		// Search through and fix pointers within the data section of the file
@@ -215,9 +216,96 @@ bool dldiPatchBinary (data_t *binData, u32 binSize) {
 
 	if (pDH[DO_fixSections] & FIX_BSS) { 
 		// Initialise the BSS to 0
-		memset (&pAH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+		toncset (&pAH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
 	}
 
 	return true;
 }
+
+void dldiDecompressBinary (void) {
+	extern void LZ77_Decompress(u8* source, u8* destination);
+	LZ77_Decompress((u8*)0x02FF8004, (u8*)_dldi_start);
+
+	addr_t memOffset;			// Offset of DLDI after the file is loaded into memory
+	addr_t relocationOffset;	// Value added to all offsets within the patch to fix it properly
+	addr_t ddmemOffset;			// Original offset used in the DLDI file
+	addr_t ddmemStart;			// Start of range that offsets can be in the DLDI file
+	addr_t ddmemEnd;			// End of range that offsets can be in the DLDI file
+	addr_t ddmemSize;			// Size of range that offsets can be in the DLDI file
+	addr_t addrIter;
+
+	data_t *pDH = _dldi_start;
+
+	memOffset = 0x06000000;
+	ddmemOffset = readAddr (pDH, DO_text_start);
+	relocationOffset = memOffset - ddmemOffset;
+
+	ddmemStart = readAddr (pDH, DO_text_start);
+	ddmemSize = (1 << pDH[DO_driverSize]);
+	ddmemEnd = ddmemStart + ddmemSize;
+
+	// Fix the section pointers in the header
+	writeAddr (pDH, DO_text_start, readAddr (pDH, DO_text_start) + relocationOffset);
+	writeAddr (pDH, DO_data_end, readAddr (pDH, DO_data_end) + relocationOffset);
+	writeAddr (pDH, DO_glue_start, readAddr (pDH, DO_glue_start) + relocationOffset);
+	writeAddr (pDH, DO_glue_end, readAddr (pDH, DO_glue_end) + relocationOffset);
+	writeAddr (pDH, DO_got_start, readAddr (pDH, DO_got_start) + relocationOffset);
+	writeAddr (pDH, DO_got_end, readAddr (pDH, DO_got_end) + relocationOffset);
+	writeAddr (pDH, DO_bss_start, readAddr (pDH, DO_bss_start) + relocationOffset);
+	writeAddr (pDH, DO_bss_end, readAddr (pDH, DO_bss_end) + relocationOffset);
+	// Fix the function pointers in the header
+	writeAddr (pDH, DO_startup, readAddr (pDH, DO_startup) + relocationOffset);
+	writeAddr (pDH, DO_isInserted, readAddr (pDH, DO_isInserted) + relocationOffset);
+	writeAddr (pDH, DO_readSectors, readAddr (pDH, DO_readSectors) + relocationOffset);
+	writeAddr (pDH, DO_writeSectors, readAddr (pDH, DO_writeSectors) + relocationOffset);
+	writeAddr (pDH, DO_clearStatus, readAddr (pDH, DO_clearStatus) + relocationOffset);
+	writeAddr (pDH, DO_shutdown, readAddr (pDH, DO_shutdown) + relocationOffset);
+
+	if (pDH[DO_fixSections] & FIX_ALL) { 
+		// Search through and fix pointers within the data section of the file
+		for (addrIter = (readAddr(pDH, DO_text_start) - ddmemStart); addrIter < (readAddr(pDH, DO_data_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_GLUE) { 
+		// Search through and fix pointers within the glue section of the file
+		for (addrIter = (readAddr(pDH, DO_glue_start) - ddmemStart); addrIter < (readAddr(pDH, DO_glue_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_GOT) { 
+		// Search through and fix pointers within the Global Offset Table section of the file
+		for (addrIter = (readAddr(pDH, DO_got_start) - ddmemStart); addrIter < (readAddr(pDH, DO_got_end) - ddmemStart); addrIter++) {
+			if ((ddmemStart <= readAddr(pDH, addrIter)) && (readAddr(pDH, addrIter) < ddmemEnd)) {
+				writeAddr (pDH, addrIter, readAddr(pDH, addrIter) + relocationOffset);
+			}
+		}
+	}
+
+	if (pDH[DO_fixSections] & FIX_BSS) { 
+		// Initialise the BSS to 0
+		toncset (&pDH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+	}
+}
+
+void dldiClearBss (void) {
+	data_t *pDH = _dldi_start;
+
+	addr_t ddmemSize = (1 << pDH[DO_driverSize]);			// Size of range that offsets can be in the DLDI file
+
+	if (ddmemSize <= 0x4000) return; // For <= 16KB DLDI file, BSS has already been cleared
+
+	addr_t ddmemStart = readAddr (pDH, DO_text_start);		// Start of range that offsets can be in the DLDI file
+
+	if (pDH[DO_fixSections] & FIX_BSS) { 
+		// Initialise the BSS to 0
+		toncset (&pDH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
+	}
+}
 #endif
\ No newline at end of file
diff --git a/universal/bootloader_menu/source/dldi_patcher.h b/universal/bootloader_menu/source/dldi_patcher.h
index f54c531ba1..16ea22a859 100644
--- a/universal/bootloader_menu/source/dldi_patcher.h
+++ b/universal/bootloader_menu/source/dldi_patcher.h
@@ -28,5 +28,7 @@
 typedef signed int addr_t;
 typedef unsigned char data_t;
 bool dldiPatchBinary (data_t *binData, u32 binSize);
+void dldiDecompressBinary (void);
+void dldiClearBss (void);
 
 #endif // DLDI_PATCHER_H
diff --git a/universal/bootloader_menu/source/io_dldi.s b/universal/bootloader_menu/source/io_dldi.s
index 08ac2cf002..82147a9140 100644
--- a/universal/bootloader_menu/source/io_dldi.s
+++ b/universal/bootloader_menu/source/io_dldi.s
@@ -32,73 +32,6 @@
 
 
 _dldi_start:
-#ifndef NO_DLDI
-
-@---------------------------------------------------------------------------------
-@ Driver patch file standard header -- 16 bytes
-#ifdef STANDARD_DLDI
-	.word	0xBF8DA5ED		@ Magic number to identify this region
-#else
-	.word	0xBF8DA5EE		@ Magic number to identify this region
-#endif
-	.asciz	" Chishm"		@ Identifying Magic string (8 bytes with null terminator)
-	.byte	0x01			@ Version number
-	.byte	0x0f		@ 32KiB	@ Log [base-2] of the size of this driver in bytes.
-	.byte	0x00			@ Sections to fix
-	.byte 	0x0f		@ 32KiB	@ Log [base-2] of the allocated space in bytes.
-	
-@---------------------------------------------------------------------------------
-@ Text identifier - can be anything up to 47 chars + terminating null -- 16 bytes
-	.align	4
-	.asciz "Loader (No interface)"
-
-@---------------------------------------------------------------------------------
-@ Offsets to important sections within the data	-- 32 bytes
-	.align	6
-	.word   _dldi_start		@ data start
-	.word   _dldi_end		@ data end
-	.word	0x00000000		@ Interworking glue start	-- Needs address fixing
-	.word	0x00000000		@ Interworking glue end
-	.word   0x00000000		@ GOT start			-- Needs address fixing
-	.word   0x00000000		@ GOT end
-	.word   0x00000000		@ bss start			-- Needs setting to zero
-	.word   0x00000000		@ bss end
-@---------------------------------------------------------------------------------
-@ IO_INTERFACE data -- 32 bytes
-_io_dldi:
-	.ascii	"DLDI"				@ ioType
-	.word	0x00000000			@ Features
-	.word	_DLDI_startup			@ 
-	.word	_DLDI_isInserted		@ 
-	.word	_DLDI_readSectors		@   Function pointers to standard device driver functions
-	.word	_DLDI_writeSectors		@ 
-	.word	_DLDI_clearStatus		@ 
-	.word	_DLDI_shutdown			@ 
-
-	
-@---------------------------------------------------------------------------------
-
-_DLDI_startup:
-_DLDI_isInserted:
-_DLDI_readSectors:
-_DLDI_writeSectors:
-_DLDI_clearStatus:
-_DLDI_shutdown:
-	mov		r0, #0x00		@ Return false for every function
-	bx		lr
-
-
-
-@---------------------------------------------------------------------------------
-	.align
-	.pool
-
-	.space (_dldi_start + 32768) - .	@ Fill to 32KiB
-
-_dldi_end:
-	.end
-@---------------------------------------------------------------------------------
-#else
 @---------------------------------------------------------------------------------
 @ IO_INTERFACE data -- 32 bytes
 _io_dldi:
@@ -111,14 +44,21 @@ _io_dldi:
 	.word	_DLDI_clearStatus		@
 	.word	_DLDI_shutdown			@
 
-	_DLDI_startup:
+_DLDI_startup:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x68]
 _DLDI_isInserted:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x6C]
 _DLDI_readSectors:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x70]
 _DLDI_writeSectors:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x74]
 _DLDI_clearStatus:
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x78]
 _DLDI_shutdown:
-	mov		r0, #0x00		@ Return false for every function
-	bx		lr
-
-
-#endif
+	mov	r12, #0x06000000
+	ldr pc, [r12, #0x7C]
diff --git a/universal/bootloader_menu/source/load_crt0.s b/universal/bootloader_menu/source/load_crt0.s
index bb841834c1..aa1b015ae3 100644
--- a/universal/bootloader_menu/source/load_crt0.s
+++ b/universal/bootloader_menu/source/load_crt0.s
@@ -55,7 +55,7 @@ argStart:
 argSize:
 	.word	0x00000000
 dldiOffset:
-	.word	_dldi_start - _start
+	.word	0
 dsiSD:
 	.word	0
 dsiMode:
diff --git a/universal/bootloader_menu/source/lzss.c b/universal/bootloader_menu/source/lzss.c
new file mode 100644
index 0000000000..d4a113b3d9
--- /dev/null
+++ b/universal/bootloader_menu/source/lzss.c
@@ -0,0 +1,27 @@
+#include <nds/ndstypes.h>
+#include <string.h>
+
+void LZ77_Decompress(u8* source, u8* destination) {
+	u32 leng = (source[1] | (source[2] << 8) | (source[3] << 16));
+	int Offs = 4;
+	int dstoffs = 0;
+	while (true) {
+		u8 header = source[Offs++];
+		for (int i = 0; i < 8; i++) {
+			if ((header & 0x80) == 0) destination[dstoffs++] = source[Offs++];
+			else
+			{
+				u8 a = source[Offs++];
+				u8 b = source[Offs++];
+				int offs = (((a & 0xF) << 8) | b) + 1;
+				int length = (a >> 4) + 3;
+				for (int j = 0; j < length; j++) {
+					destination[dstoffs] = destination[dstoffs - offs];
+					dstoffs++;
+				}
+			}
+			if (dstoffs >= (int)leng) return;
+			header <<= 1;
+		}
+	}
+}
diff --git a/universal/bootstub/bootstub.s b/universal/bootstub/bootstub.s
index 579fb9904d..a1ec52b674 100644
--- a/universal/bootstub/bootstub.s
+++ b/universal/bootstub/bootstub.s
@@ -122,7 +122,7 @@ hook7from9:
 	adr	r0, _loader
 	ldr	r2, _loader_size
 	mov	r1, #0x06800000
-	add	r1, r1, #0x40000
+	add	r1, r1, #0x48000
 	add	r2, r0, r2
 _copyloader:
 	ldr	r4, [r0], #4
@@ -140,6 +140,7 @@ _copyloader:
 
 	ldr	r0, arm7bootaddr
 	mov	r1, #0x06000000
+	add	r1, r1, #0x8000
 	str	r1, [r0]
 	
 	ldr	r0, resetcode
diff --git a/universal/include/common/nds_loader_arm9.h b/universal/include/common/nds_loader_arm9.h
index fd56741ead..05f7ea1b19 100644
--- a/universal/include/common/nds_loader_arm9.h
+++ b/universal/include/common/nds_loader_arm9.h
@@ -34,7 +34,7 @@ int runNds (const void* loader, u32 loaderSize, u32 cluster, bool initDisc, bool
 
 int runNdsFile (const char* filename, int argc, const char** argv, bool dldiPatchNds, bool clearMasterBright, bool dsModeSwitch, bool boostCpu, bool boostVram, bool tscTgds, int language);
 
-bool installBootStub(bool havedsiSD);
+void installBootStub(bool havedsiSD);
 
 #ifdef __cplusplus
 }
diff --git a/universal/source/nds_loader/nds_loader_arm9.c b/universal/source/nds_loader/nds_loader_arm9.c
index 52e822db50..dab7d8bb73 100644
--- a/universal/source/nds_loader/nds_loader_arm9.c
+++ b/universal/source/nds_loader/nds_loader_arm9.c
@@ -36,7 +36,7 @@
 #endif
 
 #include "common/nds_loader_arm9.h"
-#define LCDC_BANK_C (u16*)0x06840000
+#define LCDC_BANK_C (u16*)0x06848000
 #define STORED_FILE_CLUSTER (*(((u32*)LCDC_BANK_C) + 1))
 #define INIT_DISC (*(((u32*)LCDC_BANK_C) + 2))
 #define WANT_TO_PATCH_DLDI (*(((u32*)LCDC_BANK_C) + 3))
@@ -126,7 +126,7 @@ static void writeAddr (data_t *mem, addr_t offset, addr_t value) {
 	((addr_t*)mem)[offset/sizeof(addr_t)] = value;
 }
 
-static addr_t quickFind (const data_t* data, const data_t* search, size_t dataLen, size_t searchLen) {
+/* static addr_t quickFind (const data_t* data, const data_t* search, size_t dataLen, size_t searchLen) {
 	const int* dataChunk = (const int*) data;
 	int searchChunk = ((const int*)search)[0];
 	addr_t i;
@@ -144,18 +144,17 @@ static addr_t quickFind (const data_t* data, const data_t* search, size_t dataLe
 	}
 
 	return -1;
-}
+} */
 
 // Normal DLDI uses "\xED\xA5\x8D\xBF Chishm"
 // Bootloader string is different to avoid being patched
-static const data_t dldiMagicLoaderString[] = "\xEE\xA5\x8D\xBF Chishm";	// Different to a normal DLDI file
+// static const data_t dldiMagicLoaderString[] = "\xEE\xA5\x8D\xBF Chishm";	// Different to a normal DLDI file
 
 #define DEVICE_TYPE_DLDI 0x49444C44
 
-static bool dldiPatchLoader (data_t *binData, u32 binSize, bool clearBSS)
+static bool dldiPatchLoader (void)
 {
 	addr_t memOffset;			// Offset of DLDI after the file is loaded into memory
-	addr_t patchOffset;			// Position of patch destination in the file
 	addr_t relocationOffset;	// Value added to all offsets within the patch to fix it properly
 	addr_t ddmemOffset;			// Original offset used in the DLDI file
 	addr_t ddmemStart;			// Start of range that offsets can be in the DLDI file
@@ -164,39 +163,28 @@ static bool dldiPatchLoader (data_t *binData, u32 binSize, bool clearBSS)
 
 	addr_t addrIter;
 
-	data_t *pDH;
-	data_t *pAH;
-
 	size_t dldiFileSize = 0;
 
-	// Find the DLDI reserved space in the file
-	patchOffset = quickFind (binData, dldiMagicLoaderString, binSize, sizeof(dldiMagicLoaderString));
+	data_t *pDH = (data_t*)(io_dldi_data);
 
-	if (patchOffset < 0) {
-		// does not have a DLDI section
-		return false;
-	}
-
-	pDH = (data_t*)(io_dldi_data);
-
-	pAH = &(binData[patchOffset]);
+	data_t *pAH = (data_t*)0x02FF8000;
 
 	if (*((u32*)(pDH + DO_ioType)) == DEVICE_TYPE_DLDI) {
 		// No DLDI patch
 		return false;
 	}
 
-	if (pDH[DO_driverSize] > pAH[DO_allocatedSpace]) {
-		// Not enough space for patch
-		return false;
+	if (*(u32*)0x02FF8004 == 0x69684320 || *(u32*)0x02FF8000 == 0x53535A4C) {
+		return true; // Skip patching for existing or compressed DLDI file
 	}
 
 	dldiFileSize = 1 << pDH[DO_driverSize];
 
-	memOffset = readAddr (pAH, DO_text_start);
+	/* memOffset = readAddr (pAH, DO_text_start);
 	if (memOffset == 0) {
 			memOffset = readAddr (pAH, DO_startup) - DO_code;
-	}
+	} */
+	memOffset = 0x06000000;
 	ddmemOffset = readAddr (pDH, DO_text_start);
 	relocationOffset = memOffset - ddmemOffset;
 
@@ -207,7 +195,7 @@ static bool dldiPatchLoader (data_t *binData, u32 binSize, bool clearBSS)
 	// Remember how much space is actually reserved
 	pDH[DO_allocatedSpace] = pAH[DO_allocatedSpace];
 	// Copy the DLDI patch into the application
-	tonccpy (pAH, pDH, dldiFileSize);
+	tonccpy (pAH, pDH, (dldiFileSize > 0x4000) ? 0x4000 : dldiFileSize);
 
 	// Fix the section pointers in the header
 	writeAddr (pAH, DO_text_start, readAddr (pAH, DO_text_start) + relocationOffset);
@@ -253,7 +241,7 @@ static bool dldiPatchLoader (data_t *binData, u32 binSize, bool clearBSS)
 		}
 	}
 
-	if (clearBSS && (pDH[DO_fixSections] & FIX_BSS)) {
+	if (/* clearBSS && */ (pDH[DO_fixSections] & FIX_BSS) && dldiFileSize <= 0x4000) {
 		// Initialise the BSS to 0, only if the disc is being re-inited
 		toncset (&pAH[readAddr(pDH, DO_bss_start) - ddmemStart] , 0, readAddr(pDH, DO_bss_end) - readAddr(pDH, DO_bss_start));
 	}
@@ -328,7 +316,7 @@ int runNds (const void* loader, u32 loaderSize, u32 cluster, bool initDisc, bool
 
 	if (dldiPatchNds) {
 		// Patch the loader with a DLDI for the card
-		if (!dldiPatchLoader ((data_t*)LCDC_BANK_C, loaderSize, initDisc)) {
+		if (!dldiPatchLoader ()) {
 			return 3;
 		}
 	}
@@ -354,7 +342,7 @@ int runNds (const void* loader, u32 loaderSize, u32 cluster, bool initDisc, bool
 	*((vu32*)0x02FFFE04) = (u32)0xE59FF018;
 	*((vu32*)0x02FFFE24) = (u32)0x02FFFE04;
 
-	resetARM7(0x06000000);
+	resetARM7(0x06008000);
 
 	swiSoftReset();
 	return true;
@@ -507,7 +495,7 @@ int runNdsFile (const char* filename, int argc, const char** argv, bool dldiPatc
 		filename = argv[0];
 	}
 
-	bool havedsiSD = (access("sd:/", F_OK) == 0);
+	const bool havedsiSD = (access("sd:/", F_OK) == 0);
 
 	if (REG_SCFG_EXT != 0 && havedsiSD) {
 		// Check for Unlaunch
@@ -526,7 +514,9 @@ int runNdsFile (const char* filename, int argc, const char** argv, bool dldiPatc
 
 	bool loadFromRam = runNds9(filename, dsModeSwitch);
 
+	#ifndef _NO_BOOTSTUB_
 	installBootStub(havedsiSD);
+	#endif
 
 	return runNds (load_bin, load_bin_size, st.st_ino, true, (dldiPatchNds && memcmp(io_dldi_data->friendlyName, "Default", 7) != 0), loadFromRam, filename, argc, argv, clearMasterBright, dsModeSwitch, lockScfg, boostCpu, boostVram, tscTgds, language);
 }
@@ -551,34 +541,31 @@ int runNdsFile (const char* filename, int argc, const char** argv, bool dldiPatc
 	.word	0
 */
 
-bool installBootStub(bool havedsiSD) {
 #ifndef _NO_BOOTSTUB_
+void installBootStub(bool havedsiSD) {
+	static bool installed = false;
+	if (installed) return;
+
 	extern char *fake_heap_end;
 	struct __bootstub *bootstub = (struct __bootstub *)fake_heap_end;
 	u32 *bootloader = (u32*)(fake_heap_end+bootstub_bin_size);
 
-	memcpy(bootstub,bootstub_bin,bootstub_bin_size);
-	memcpy(bootloader,load_bin,load_bin_size);
-	bool ret = false;
+	tonccpy(bootstub,bootstub_bin,bootstub_bin_size);
+	tonccpy(bootloader,load_bin,load_bin_size);
 
 	bootloader[8] = isDSiMode();
 	if (havedsiSD) {
-		ret = true;
 		bootloader[3] = 0; // don't dldi patch
 		bootloader[7] = 1; // use internal dsi SD code
 	} else {
-		ret = dldiPatchLoader((data_t*)bootloader, load_bin_size,false);
+		dldiPatchLoader ();
 	}
 	bootstub->arm9reboot = (VoidFn)(((u32)bootstub->arm9reboot)+fake_heap_end);
 	bootstub->arm7reboot = (VoidFn)(((u32)bootstub->arm7reboot)+fake_heap_end);
 	bootstub->bootsize = load_bin_size;
 
 	DC_FlushAll();
-
-	return ret;
-#else
-	return true;
-#endif
-
+	installed = true;
 }
+#endif