diff --git a/CMakeLists.txt b/CMakeLists.txt
index f0fc47f..c8c8243 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -127,6 +127,7 @@ set(SOURCES_COMMON
   src/usb/classes/ff-bulk.c
   src/usb/configure.c
   src/usb/pipes.c
+  src/usb/read4.S
   src/usb/setup.c
   src/usb/string.c
   src/usb/usb.c
diff --git a/src/usb/classes/ff-bulk.c b/src/usb/classes/ff-bulk.c
index 745ce5f..c90af33 100644
--- a/src/usb/classes/ff-bulk.c
+++ b/src/usb/classes/ff-bulk.c
@@ -12,7 +12,7 @@ static void notify_read(int endpoint);
 static usb_dc_interface_t dc_interface = {
 	.bLength              = sizeof(usb_dc_interface_t),
 	.bDescriptorType      = USB_DC_INTERFACE,
-	.bInterfaceNumber     = -1 /* Set by driver */,
+	.bInterfaceNumber     = -1, /* Set by driver */
 	.bAlternateSetting    = 0,
 	.bNumEndpoints        = 2,
 	.bInterfaceClass      = 0xff, /* Vendor-Specific */
@@ -283,10 +283,11 @@ void usb_fxlink_set_notifier(void (*notifier_function)(void))
 void usb_fxlink_drop_transaction(void)
 {
 	int block = USB_READ_BLOCK;
+	char buf[512];
 
 	while(1) {
 		timeout_t tm = timeout_make_ms(1000);
-		int rc = usb_read_async(usb_ff_bulk_input(), NULL, 512,
+		int rc = usb_read_async(usb_ff_bulk_input(), buf, 512,
 			USB_READ_WAIT | block, NULL, &tm, GINT_CALL_NULL);
 
 		/* Break on error or short read (end of transaction) */
diff --git a/src/usb/pipes.c b/src/usb/pipes.c
index e988f44..316f664 100644
--- a/src/usb/pipes.c
+++ b/src/usb/pipes.c
@@ -36,7 +36,7 @@ void usb_pipe_configure(int address, endpoint_t const *ep)
 	USB.PIPECFG.TYPE    = type;
 	USB.PIPECFG.BFRE    = 0;
 	/* Enable continuous mode on all bulk transfer pipes
-	   TODO: Also make it double mode*/
+	   TODO: Also make it double mode */
 	USB.PIPECFG.DBLB    = 0;
 	USB.PIPECFG.CNTMD   = (type == TYPE_BULK);
 	USB.PIPECFG.SHTNAK  = 1;
@@ -532,11 +532,17 @@ static void USB_LOG_TR(char const *p, asyncio_op_t *t, char const *fmt, ...)
 	int E = USB.INTENB0.BRDYE;
 	USB.INTENB0.BRDYE = 0;
 
+	char shbuf[16];
+	if(t->shbuf_size >= 4)
+		sprintf(shbuf, "!!%d", t->shbuf_size);
+	else
+		snprintf(shbuf, t->shbuf_size * 2 + 1, "%08x", t->shbuf);
+
 	char str[128];
-	snprintf(str, sizeof str - 1, "%s: %s buf=%d%s%s req=%d/%d%s ",
+	snprintf(str, sizeof str - 1, "%s: %s buf=%d%s%s req=%d/%d%s |%s| ",
 		p, t->type == ASYNCIO_READ ? "READ" : "NONE",
 		t->buffer_used, t->cont_r ? "+":"", t->interrupt_r ? "!":"",
-		t->round_size, t->size, t->autoclose_r ? "#" : "");
+		t->round_size, t->size, t->autoclose_r ? "#" : "", shbuf);
 
 	va_list args;
 	va_start(args, fmt);
@@ -623,31 +629,9 @@ static bool read_round(asyncio_op_t *t, int pipe)
 	if(t->controller == D0F) FIFO = &USB.D0FIFO;
 	if(t->controller == D1F) FIFO = &USB.D1FIFO;
 
-	void *dataptr = t->data_r;
-	if(dataptr) {
-		for(int i = 0; i < round_size / 4; i++) {
-			*(uint32_t *)dataptr = *FIFO;
-			dataptr += 4;
-		}
-		if(round_size & 2) {
-			*(uint16_t *)dataptr = *(uint16_t volatile *)FIFO;
-			dataptr += 2;
-		}
-		if(round_size & 1) {
-			*(uint8_t *)dataptr = *(uint8_t volatile *)FIFO;
-			dataptr += 1;
-		}
-	}
-	else {
-		volatile int x;
-		for(int i = 0; i < round_size / 4; i++)
-			x = *FIFO;
-		if(round_size & 2)
-			x = *(uint16_t volatile *)FIFO;
-		if(round_size & 1)
-			x = *(uint8_t volatile *)FIFO;
-		(void)x;
-	}
+	int fifosize = t->buffer_used - t->shbuf_size;
+	usb_pipe_read4(t->data_r, round_size, FIFO, fifosize, &t->shbuf,
+		&t->shbuf_size);
 
 	finish_read_round(t, pipe);
 	return false;
diff --git a/src/usb/read4.S b/src/usb/read4.S
new file mode 100644
index 0000000..cbf837b
--- /dev/null
+++ b/src/usb/read4.S
@@ -0,0 +1,181 @@
+.global _usb_pipe_read4
+
+/* User buffer and round size */
+#define _data r4
+#define _datasize r5
+/* FIFO address and amount of data available in there */
+#define _fifo r6
+#define _fifosize r7
+/* Short buffer address and *pointer to* its size */
+#define _buf r8
+#define _bufsize r9
+
+/* Copy _datasize bytes from _fifo to _data, using USB FIFO access rules and
+   storing excess data in the short buffer.
+   Requires: 1 ≤ _datasize ≤ _fifosize + *_bufsize */
+_usb_pipe_read4:
+	mov.l	@(4, r15), r1
+	mov.l	r8, @-r15
+	mov.l	r9, @-r15
+
+	mov	r1, _bufsize
+	mov.b	@_bufsize, r1
+
+/* Step #1: If _datasize ≤ *_bufsize < 4, then we fill user data from the short
+   buffer, and return immediately. */
+
+	/* 1 cycle lost here due to early use of r1 */
+
+	mov.l	@(8, r15), _buf
+	cmp/ge	_datasize, r1
+
+	bt	.short_buffer_only
+	tst	r1, r1
+
+/* Step #2: If *_bufsize > 0, copy *_bufsize bytes from the short buffer to
+  _data, emptying the short buffer. */
+
+	bt	2f
+	mov	_buf, r3
+
+1:	mov.b	@r3+, r2
+	dt	_datasize
+
+	dt	r1
+ 	mov.b	r2, @_data
+
+	bf.s	1b
+	add	#1, _data
+
+	mov.b	r1, @_bufsize
+	nop
+
+/* Step #3: Copy longwords from the FIFO to the user buffer as long as there is
+   at least 4 bytes in both. Since at this stage _datasize ≤ _fifosize we can
+   simply check _datasize. */
+
+2:	/* Update _datasize and _fifosize in advance */
+	mov	#-4, r0
+	and	_datasize, r0
+
+	/* If _datasize ≤ 3, skip this step */
+	mov	#3, r3
+	cmp/ge	_datasize, r3
+
+	bt.s	3f
+	sub	r0, _fifosize
+
+	tst	r3, _data
+	nop
+
+	/* r2 is the number of 4-byte reads; since _datasize > 3, r2 ≥ 1 */
+	mov	_datasize, r2
+	shlr2	r2
+
+	bf.s	.unaligned
+	and	r3, _datasize
+
+#define VERSION 0
+
+.aligned:
+
+#if VERSION == 0
+	mov.l	@_fifo, r3
+	dt	r2
+
+	mov.l	r3, @_data
+	add	#4, _data
+
+	bf.s	.aligned
+	nop
+#elif VERSION == 1
+	mov	_fifo, r3
+	nop
+
+9:	movs.l	@r3, x0
+	dt	r2
+
+	bf.s	9b
+	movs.l	x0, @_data+
+#elif VERSION == 2
+	ldrs	.al_b
+	ldre	.al_e
+	ldrc	r2
+	mov	_fifo, r3
+
+.al_b:	movs.l	@r3, x0
+.al_e:	movs.l	x0, @_data+
+#endif
+
+	bra	3f
+	nop
+
+.unaligned:
+	mov.l	@_fifo, r0
+	dt	r2
+
+	mov.b	r0, @(3, _data)
+	nop
+
+	shlr8	r0
+	mov.b	r0, @(2, _data)
+
+	shlr8	r0
+	mov.b	r0, @(1, _data)
+
+	shlr8	r0
+	mov.b	r0, @_data
+
+	bf.s	.unaligned
+	add	#4, _data
+
+/* Step #4: Load the final bytes of the round from the FIFO into the short
+   buffer (can be either nothing or a standard 4-byte read), then finish with a
+   copy to user data. */
+
+3:	/* If we finished the read, don't load anything */
+	tst	_datasize, _datasize
+	mov	#3, r3
+
+	bt	.epilogue
+	cmp/hi	r3, _fifosize
+
+	bf	4f
+	mov	#4, _fifosize
+
+	/* Make a 4-byte read from the FIFO. If there are less than 4 bytes
+	   left this will pad with zeros. */
+4:	mov.l	@_fifo, r3
+	mov	_fifosize, r1
+
+	mov.l	r3, @_buf
+	nop
+
+/* Step #1/#5: Copy 0 < _datasize ≤ *_bufsize bytes from the short buffer to
+   _data, then return. r1 must be *_bufsize. */
+.short_buffer_only:
+	mov.l	@_buf, r3
+	mov	#0, r0
+
+	sub	_datasize, r1
+	nop
+
+	/* The loop accesses memory with @(r0, _) but also shifts the contents
+	   of _buf in r3 (it's more convenient to use the available EX slot
+	   than shift by 8*_datasize outside the loop) */
+
+5:	mov.b	@(r0, _buf), r2
+	dt	_datasize
+
+	mov.b	r2, @(r0, _data)
+	add	#1, r0
+
+	bf.s	5b
+	shll8	r3
+
+.epilogue:
+	mov.l	r3, @_buf
+	mov.b	r1, @_bufsize
+	mov.l	@r15+, r9
+	rts
+	mov.l	@r15+, r8
diff --git a/src/usb/usb_private.h b/src/usb/usb_private.h
index 58be9c7..68f1a20 100644
--- a/src/usb/usb_private.h
+++ b/src/usb/usb_private.h
@@ -175,6 +175,23 @@ void usb_pipe_write4(void const *data, int size, uint32_t volatile *buffer,
 void usb_pipe_flush4(uint32_t buffer, int buffer_size,
    uint32_t volatile *FIFO);
 
+/* usb_pipe_read4(): Copy arbitrary ranges of memory from a 4-byte USB FIFO
+
+   This function performs arbitrarily-aligned reads of any size from a 4-byte
+   USB FIFO register to regular memory. It performs only 4-byte reads on the
+   FIFO (except when reading the last few bytes in the buffer) and copies read
+   data to the supplied buffer. Any excess bytes read from the FIFO are stored
+   in a short buffer to be used on the next call.
+
+   @data         Buffer to read into
+   @data_size    Read size (1 ≤ data_size ≤ fifo_size + *buffer_size)
+   @fifo         USB FIFO register to read from
+   @fifo_size    Amount of data left in the FIFO (excluding short buffer!)
+   @buffer       Address of short buffer
+   @buffer_size  Address of short buffer's size tracker */
+void usb_pipe_read4(void *data, int size, uint32_t volatile *FIFO,
+   int fifo_size, uint32_t volatile *buffer, uint8_t volatile *buffer_size);
+
 //---
 // Timeout waits
 //---