cortex_a: fix endiannes issues on TI TMS570

The TI TMS470 and TMS570 series of processors are BE-32 processors,
despite BE-32 not being supported by ARM in the Cortex-R4 core. TI
hacked in BE-32 support, which requires odd swizzling in OpenOCD to
make memory reads and writes function correctly. In particular,
without this change, OpenOCD word reads and writes had the bytes
reversed, and halfword and byte packed reads were reading garbage.
In my testing, this change fixes these problems.

Change-Id: I21dd30f4b9003f20fcc85f674ab833407bb61f74
Signed-off-by: Seth LaForge <sethml@google.com>
Reviewed-on: http://openocd.zylin.com/2064
Tested-by: jenkins
Reviewed-by: Paul Fertser <fercerpav@gmail.com>
Reviewed-by: Andreas Fritiofson <andreas.fritiofson@gmail.com>

src/target/arm_adi_v5.c

@@ -295,18 +295,39 @@ int mem_ap_write(struct adiv5_dap *dap, const uint8_t *buffer, uint32_t size, ui
 	size_t nbytes = size * count;
 	const uint32_t csw_addrincr = addrinc ? CSW_ADDRINC_SINGLE : CSW_ADDRINC_OFF;
 	uint32_t csw_size;
+	uint32_t addr_xor;
 	int retval;
 
-	if (size == 4)
+	/* TI BE-32 Quirks mode:
+	 * Writes on big-endian TMS570 behave very strangely. Observed behavior:
+	 *   size   write address   bytes written in order
+	 *   4      TAR ^ 0         (val >> 24), (val >> 16), (val >> 8), (val)
+	 *   2      TAR ^ 2         (val >> 8), (val)
+	 *   1      TAR ^ 3         (val)
+	 * For example, if you attempt to write a single byte to address 0, the processor
+	 * will actually write a byte to address 3.
+	 *
+	 * To make writes of size < 4 work as expected, we xor a value with the address before
+	 * setting the TAP, and we set the TAP after every transfer rather then relying on
+	 * address increment. */
+
+	if (size == 4) {
 		csw_size = CSW_32BIT;
-	else if (size == 2)
+		addr_xor = 0;
+	} else if (size == 2) {
 		csw_size = CSW_16BIT;
-	else if (size == 1)
+		addr_xor = dap->ti_be_32_quirks ? 2 : 0;
+	} else if (size == 1) {
 		csw_size = CSW_8BIT;
-	else
+		addr_xor = dap->ti_be_32_quirks ? 3 : 0;
+	} else {
+		return ERROR_TARGET_UNALIGNED_ACCESS;
+	}
+
+	if (dap->unaligned_access_bad && (address % size != 0))
 		return ERROR_TARGET_UNALIGNED_ACCESS;
 
-	retval = dap_setup_accessport_tar(dap, address);
+	retval = dap_setup_accessport_tar(dap, address ^ addr_xor);
 	if (retval != ERROR_OK)
 		return retval;
 
@@ -328,6 +349,23 @@ int mem_ap_write(struct adiv5_dap *dap, const uint8_t *buffer, uint32_t size, ui
 		/* How many source bytes each transfer will consume, and their location in the DRW,
 		 * depends on the type of transfer and alignment. See ARM document IHI0031C. */
 		uint32_t outvalue = 0;
+		if (dap->ti_be_32_quirks) {
+			switch (this_size) {
+			case 4:
+				outvalue |= (uint32_t)*buffer++ << 8 * (3 ^ (address++ & 3) ^ addr_xor);
+				outvalue |= (uint32_t)*buffer++ << 8 * (3 ^ (address++ & 3) ^ addr_xor);
+				outvalue |= (uint32_t)*buffer++ << 8 * (3 ^ (address++ & 3) ^ addr_xor);
+				outvalue |= (uint32_t)*buffer++ << 8 * (3 ^ (address++ & 3) ^ addr_xor);
+				break;
+			case 2:
+				outvalue |= (uint32_t)*buffer++ << 8 * (1 ^ (address++ & 3) ^ addr_xor);
+				outvalue |= (uint32_t)*buffer++ << 8 * (1 ^ (address++ & 3) ^ addr_xor);
+				break;
+			case 1:
+				outvalue |= (uint32_t)*buffer++ << 8 * (0 ^ (address++ & 3) ^ addr_xor);
+				break;
+			}
+		} else {
 			switch (this_size) {
 			case 4:
 				outvalue |= (uint32_t)*buffer++ << 8 * (address++ & 3);
@@ -337,6 +375,7 @@ int mem_ap_write(struct adiv5_dap *dap, const uint8_t *buffer, uint32_t size, ui
 			case 1:
 				outvalue |= (uint32_t)*buffer++ << 8 * (address++ & 3);
 			}
+		}
 
 		nbytes -= this_size;
 
@@ -344,9 +383,9 @@ int mem_ap_write(struct adiv5_dap *dap, const uint8_t *buffer, uint32_t size, ui
 		if (retval != ERROR_OK)
 			break;
 
-		/* Rewrite TAR if it wrapped */
+		/* Rewrite TAR if it wrapped or we're xoring addresses */
-		if (addrinc && address % dap->tar_autoincr_block < size && nbytes > 0) {
+		if (addrinc && (addr_xor || (address % dap->tar_autoincr_block < size && nbytes > 0))) {
-			retval = dap_setup_accessport_tar(dap, address);
+			retval = dap_setup_accessport_tar(dap, address ^ addr_xor);
 			if (retval != ERROR_OK)
 				break;
 		}
@@ -389,6 +428,13 @@ int mem_ap_read(struct adiv5_dap *dap, uint8_t *buffer, uint32_t size, uint32_t
 	uint32_t address = adr;
 	int retval;
 
+	/* TI BE-32 Quirks mode:
+	 * Reads on big-endian TMS570 behave strangely differently than writes.
+	 * They read from the physical address requested, but with DRW byte-reversed.
+	 * For example, a byte read from address 0 will place the result in the high bytes of DRW.
+	 * Also, packed 8-bit and 16-bit transfers seem to sometimes return garbage in some bytes,
+	 * so avoid them. */
+
 	if (size == 4)
 		csw_size = CSW_32BIT;
 	else if (size == 2)
@@ -398,6 +444,9 @@ int mem_ap_read(struct adiv5_dap *dap, uint8_t *buffer, uint32_t size, uint32_t
 	else
 		return ERROR_TARGET_UNALIGNED_ACCESS;
 
+	if (dap->unaligned_access_bad && (adr % size != 0))
+		return ERROR_TARGET_UNALIGNED_ACCESS;
+
 	/* Allocate buffer to hold the sequence of DRW reads that will be made. This is a significant
 	 * over-allocation if packed transfers are going to be used, but determining the real need at
 	 * this point would be messy. */
@@ -478,6 +527,17 @@ int mem_ap_read(struct adiv5_dap *dap, uint8_t *buffer, uint32_t size, uint32_t
 			this_size = 4;
 		}
 
+		if (dap->ti_be_32_quirks) {
+			switch (this_size) {
+			case 4:
+				*buffer++ = *read_ptr >> 8 * (3 - (address++ & 3));
+				*buffer++ = *read_ptr >> 8 * (3 - (address++ & 3));
+			case 2:
+				*buffer++ = *read_ptr >> 8 * (3 - (address++ & 3));
+			case 1:
+				*buffer++ = *read_ptr >> 8 * (3 - (address++ & 3));
+			}
+		} else {
 			switch (this_size) {
 			case 4:
 				*buffer++ = *read_ptr >> 8 * (address++ & 3);
@@ -487,6 +547,7 @@ int mem_ap_read(struct adiv5_dap *dap, uint8_t *buffer, uint32_t size, uint32_t
 			case 1:
 				*buffer++ = *read_ptr >> 8 * (address++ & 3);
 			}
+		}
 
 		read_ptr++;
 		nbytes -= this_size;
@@ -845,7 +906,7 @@ int ahbap_debugport_init(struct adiv5_dap *dap)
 	dap_syssec(dap);
 
 	/* check that we support packed transfers */
-	uint32_t csw;
+	uint32_t csw, cfg;
 
 	retval = dap_setup_accessport(dap, CSW_8BIT | CSW_ADDRINC_PACKED, 0);
 	if (retval != ERROR_OK)
@@ -855,6 +916,10 @@ int ahbap_debugport_init(struct adiv5_dap *dap)
 	if (retval != ERROR_OK)
 		return retval;
 
+	retval = dap_queue_ap_read(dap, AP_REG_CFG, &cfg);
+	if (retval != ERROR_OK)
+		return retval;
+
 	retval = dap_run(dap);
 	if (retval != ERROR_OK)
 		return retval;
@@ -864,9 +929,25 @@ int ahbap_debugport_init(struct adiv5_dap *dap)
 	else
 		dap->packed_transfers = false;
 
+	/* Packed transfers on TI BE-32 processors do not work correctly in
+	 * many cases. */
+	if (dap->ti_be_32_quirks)
+		dap->packed_transfers = false;
+
 	LOG_DEBUG("MEM_AP Packed Transfers: %s",
 			dap->packed_transfers ? "enabled" : "disabled");
 
+	/* The ARM ADI spec leaves implementation-defined whether unaligned
+	 * memory accesses work, only work partially, or cause a sticky error.
+	 * On TI BE-32 processors, reads seem to return garbage in some bytes
+	 * and unaligned writes seem to cause a sticky error.
+	 * TODO: it would be nice to have a way to detect whether unaligned
+	 * operations are supported on other processors. */
+	dap->unaligned_access_bad = dap->ti_be_32_quirks;
+
+	LOG_DEBUG("MEM_AP CFG: large data %d, long address %d, big-endian %d",
+			!!(cfg & 0x04), !!(cfg & 0x02), !!(cfg & 0x01));
+
 	return ERROR_OK;
 }
 
@@ -1664,6 +1745,32 @@ COMMAND_HANDLER(dap_apid_command)
 	return retval;
 }
 
+COMMAND_HANDLER(dap_ti_be_32_quirks_command)
+{
+	struct target *target = get_current_target(CMD_CTX);
+	struct arm *arm = target_to_arm(target);
+	struct adiv5_dap *dap = arm->dap;
+
+	uint32_t enable = dap->ti_be_32_quirks;
+
+	switch (CMD_ARGC) {
+	case 0:
+		break;
+	case 1:
+		COMMAND_PARSE_NUMBER(u32, CMD_ARGV[0], enable);
+		if (enable > 1)
+			return ERROR_COMMAND_SYNTAX_ERROR;
+		break;
+	default:
+		return ERROR_COMMAND_SYNTAX_ERROR;
+	}
+	dap->ti_be_32_quirks = enable;
+	command_print(CMD_CTX, "TI BE-32 quirks mode %s",
+		enable ? "enabled" : "disabled");
+
+	return 0;
+}
+
 static const struct command_registration dap_commands[] = {
 	{
 		.name = "info",
@@ -1713,6 +1820,13 @@ static const struct command_registration dap_commands[] = {
 			"bus access [0-255]",
 		.usage = "[cycles]",
 	},
+	{
+		.name = "ti_be_32_quirks",
+		.handler = dap_ti_be_32_quirks_command,
+		.mode = COMMAND_CONFIG,
+		.help = "set/get quirks mode for TI TMS450/TMS570 processors",
+		.usage = "[enable]",
+	},
 	COMMAND_REGISTRATION_DONE
 };
 

src/target/arm_adi_v5.h

@@ -189,6 +189,15 @@ struct adiv5_dap {
 
 	/* true if packed transfers are supported by the MEM-AP */
 	bool packed_transfers;
+
+	/* true if unaligned memory access is not supported by the MEM-AP */
+	bool unaligned_access_bad;
+
+	/* The TI TMS470 and TMS570 series processors use a BE-32 memory ordering
+	 * despite lack of support in the ARMv7 architecture. Memory access through
+	 * the AHB-AP has strange byte ordering these processors, and we need to
+	 * swizzle appropriately. */
+	bool ti_be_32_quirks;
 };
 
 /**

tcl/target/ti_tms570.cfg

@@ -40,6 +40,9 @@ set _TARGETNAME $_CHIPNAME.cpu
 target create $_TARGETNAME cortex_r4 -endian $_ENDIAN \
 	-chain-position $_CHIPNAME.dap -coreid 0 -dbgbase 0x00001003
 
+# TMS570 uses quirky BE-32 mode
+$_TARGETNAME dap ti_be_32_quirks 1
+
 $_TARGETNAME configure -event gdb-attach {
 	cortex_r4 dbginit
 	halt