Re: gdb -> OcdLibRemote -> Raven | Wiggler -> LittleEndian Arm

[Andrew Cagney <ac131313 at redhat dot com>]

This sounds very like the remote not realising that its got an endian 
swap and, consequently, failing to word-swap memory transfers.

I've attached a description of the problem as it relates to the PPC.

Andrew

Below is a worked through example of an little endian data
manipulation on the PowerPC.

Background (Programmer Model):

The PowerPC defines memory as being `byte' addressable.  Mostly
through out this example, I'll depict memory as only being one byte
size.

On a few occasions, I'll show the mapping between these bytes and the
64bit data bus of the PowerPC.

Hence memory can be viewed as:

		0:
		1:
		2:
		3:
		4:
		.
		.
		.

Some simple examples:

Big-endian store of ABCD at address 4.  Big endian is where the most
significant (A) byte is stored first (addr 4) in memory.

		3:	?
		4:	A
		5:	B
		6:	C
		7:	D
		8:	?

Little-endian store of ABCD at address 4.  Little endian is where the
least significant byte is stored first (addr 4) in memory.

		3:	?
		4:	D
		5:	C
		6:	B
		7:	A
		8:	?

Notice here that if a value is stored using one endian and read back
using the other, the order of bytes will be reversed vis:

	o	Write ABCD in Big-endian at address 4

	o	Read it as little endian at address 4 gives DCBA

What the PowerPC does (see the PowerPC manual for an introduction).
Below is a second way of looking at what it is doing.

Big endian case:

When in big endian mode, the PowerPC will behave as described above.
Nothing strange about this.

Little endian case (see table 2-29):

	Effective Address Modifications

	Data Width (bytes)	EA Mod

		8		No Change
		4		XOR with b'100
		2		XOR with b'110
		1		XOR with b'111

	[PowerPC 601 RISC Microprocessor User's Manual. MPC601UM/AD]

I'm going to look at the storing of the value ABCD at address 4, in
doing this I'm going to break this store into two separate cases a) as
individual bytes and b) as a 4 byte word.  In both cases, for the
machine to be be truly byte addressable and little endian, no
difference should be detectable.

a) Individual bytes

Using table 2-29, (byte address so XOR with binary 111), I translate
the `internal' address for each byte into its corresponding physical
address.  (Remember little endian, so D the list significant byte is
stored at address 4)

	Addr	Value	Baddr	Caddr
	 4	 D	4=100	011=3
	 5	 C	5=101	010=2
	 6	 B	6=110	001=1
	 7	 A	7=111	000=0

Hence each value is stored into memory as:

	0:	A
	1:	B
	2:	C
	3:	D
	4:	?

b) As a 4 byte quantity.

Again using table 2-29 and the rule `stored in big endian order' we
get that: 4 (= 100) xor b 100 (= 000) and hence again:

	0:	A
	1:	B
	2:	C
	3:	D
	4:	?

Notice that in each of these two examples, the data is:

	o	Stored in memory in big endian order

	o	Stored in locations in a different order

Mapping this model onto the 64 bit wide data path of the PowerPC:

The next step in trying to understand this is to look at the mapping
of this byte addressable programmable model onto the real 64 bit
PowerPC data path.  Again, to emphasize the mapping between the above
model and real hardware, I'll depict the data path vertically.  The
PowerPC has a 64 bit data path.  This data path can be broken into 8
data lanes, each 8 bits wide.

		Lane	Bits
		0:	0 - 7
		1:	8 - 15
		2:	16 -23
		3:	24 - 31
		4:	32 - 39
		5:	40 - 47
		6:	48 - 56
		7:	56 - 63

When the PowerPC performs a big endian store to address 4 with data
ABCD.  It put on to the address/data busses:

	Address:	4 (in binary 100)
	Data Lane:
		0:
		1:
		2:
		3:
		4:	A
		5:	B
		6:	C
		7:	D

For a little endian store (same address and data) it would output
(onto the address/data bus):

	Address:	0 (binary 000)
	Data Lane:
		0:	A
		1:	B
		2:	C
		3:	D
		4:
		5:
		6:
		7:

Or for a single byte consider a store of `*' at address 6.  6 (binary
110) XOR with 111 (as byte access) is 1 (001).  So for these two
cases, the data would appear on the byte lanes's as shown below:

	Address:	6(110)	1(001)
	Data Lane:
		0:
		1:		*
		2:
		3:
		4:
		5:
		6:	*
		7:

Impact on PowerPC <-> IO operations:

If the PowerPC is only making transfers between itself and memory, the
above translation doesn't matter.  If, however, the transfer is
between the PowerPC and a real IO device, then there is a serious
problem.

Consider, for instance, an IO controller with one register at `address
6' A programmer would assume a byte store at address 6 would go to the
IO controller.  Further the controller would expect that the data is
put onto data lane `6' of the data bus.

In big endian mode, this assumption is correct.  However, from the
above example, it can be seen that in little endian mode this is no
longer the case:

	o	The address output by the processor would be 1 not 6

	o	The data would appear on chanel 1 not 6


Following on from this, consider now a 2 byte register at address 6.
The registers could be accessed as either A@6 and B@7 or as a 2 byte
pair.  Clearly using our definition of little endian and big endian
stores, a write of AB should produce:

			Big	Little
	Data lane:
		5:	?	?
		6:	A	B
		7:	B	A
		8:	?	?

With the PowerPC in little endian mode, however, this doesn't happen.
Instead of putting on the bus:

	Address:	6
	Data lane
		6:	B
		7:	A

The PowerPC outputs onto it's address & data lines:

	Address:	0 (000 = 6(110) XOR 110)
	Data lane:
		0:	A
		1:	B

Oops!

Inpact on DMA operations:

Consider a far longer example of a dma device that is to store the
bytes ABCDEF into memory starting at address 6.  The endian ness of
the machine shouldn't matter as we're considering a byte stream.  So
in each case, the dma transfer should result in:

	5:	?
	6:	A
	7:	B
	8:	C
	9:	D
	a:	E
	b:	F
	c:	?

However applying the rules defined by table 2-29, the PowerPC in
little endian mode would look for the values at addresses:

	Value	Addr	PPC looks in:
	A	 6	 1
	B	 7	 0
	C	 8	 f
	D	 9	 e
	E	 a	 d
	F	 b	 c

Hence, for a little endian PowerPC to be able to correctly access the
bytes brought in via DMA they would need to appear in memory as:

	0:	B
	1:	A
	2:	?
	3:	?
	4:	?
	5:	?
	6:	?
	7:	?
	8:	?
	9:	?
	a:	?
	b:	?
	c:	F
	d:	E
	e:	D
	f:	C

Yuck!

Mapping these examples onto more traditional ways of drawing 64 bit
busses:

If you wish to re-draw these examples using more traditional diagrams
(eg 64but data buse) be certain to mark, in each case the mapping
between the physical data lanes and the logical PowerPC addresses.