aarch64: register access rewrite

All register access is now performed through common read/write
functions, which delegate the actual register access to the
armv8_common object. armv8_common contains function pointers
to direct read and write requests to the respective low-level
functions for each PE state.

The respective read/write functions are selected on debug state
entry.

At the same time, T32 opcodes are now formatted for ITR in
dpmv8_exec_opcode() and the T32_FMTITR macro is removed from global
visibility.

Change-Id: I9eaef017c7cc9e0c531e693c534901bfdbdb842c
Signed-off-by: Matthias Welwarsky <matthias.welwarsky@sysgo.com>

8 files changed, 580 insertions, 366 deletions

diff --git a/src/target/aarch64.c b/src/target/aarch64.c
index 41bea2e..02f17d8 100644
--- a/src/target/aarch64.c
+++ b/src/target/aarch64.c
@@ -637,13 +637,15 @@ static int aarch64_debug_entry(struct target *target)
 	int retval = ERROR_OK;
 	struct aarch64_common *aarch64 = target_to_aarch64(target);
 	struct armv8_common *armv8 = target_to_armv8(target);
+	struct arm_dpm *dpm = &armv8->dpm;
+	enum arm_state core_state;
 
-	LOG_DEBUG("dscr = 0x%08" PRIx32, aarch64->cpudbg_dscr);
+	LOG_DEBUG("%s dscr = 0x%08" PRIx32, target_name(target), aarch64->cpudbg_dscr);
 
-	/* REVISIT see A8 TRM 12.11.4 steps 2..3 -- make sure that any
-	 * imprecise data aborts get discarded by issuing a Data
-	 * Synchronization Barrier:  ARMV4_5_MCR(15, 0, 0, 7, 10, 4).
-	 */
+	dpm->dscr = aarch64->cpudbg_dscr;
+	core_state = armv8_dpm_get_core_state(dpm);
+	armv8_select_opcodes(armv8, core_state == ARM_STATE_AARCH64);
+	armv8_select_reg_access(armv8, core_state == ARM_STATE_AARCH64);
 
 	/* make sure to clear all sticky errors */
 	retval = mem_ap_write_atomic_u32(armv8->debug_ap,
@@ -677,11 +679,8 @@ static int aarch64_debug_entry(struct target *target)
 
 	retval = armv8_dpm_read_current_registers(&armv8->dpm);
 
-	if (armv8->post_debug_entry) {
+	if (retval == ERROR_OK && armv8->post_debug_entry)
 		retval = armv8->post_debug_entry(target);
-		if (retval != ERROR_OK)
-			return retval;
-	}
 
 	return retval;
 }
@@ -1457,7 +1456,7 @@ static int aarch64_write_apb_ap_memory(struct target *target,
 		/* Step 1.a+b - Write the address for read access into DBGDTRRX */
 		/* Step 1.c   - Copy value from DTR to R0 using instruction mrc DBGDTRTXint, r0 */
 		dpm->instr_write_data_dcc(dpm,
-				T32_FMTITR(ARMV4_5_MRC(14, 0, 0, 0, 5, 0)), address & ~0x3ULL);
+				ARMV4_5_MRC(14, 0, 0, 0, 5, 0), address & ~0x3ULL);
 
 	}
 	/* Step 1.d   - Change DCC to memory mode */
@@ -1486,6 +1485,8 @@ static int aarch64_write_apb_ap_memory(struct target *target,
 				armv8->debug_base + CPUV8_DBG_DSCR, &dscr);
 	if (retval != ERROR_OK)
 		goto error_free_buff_w;
+
+	dpm->dscr = dscr;
 	if (dscr & (DSCR_ERR | DSCR_SYS_ERROR_PEND)) {
 		/* Abort occurred - clear it and exit */
 		LOG_ERROR("abort occurred - dscr = 0x%08" PRIx32, dscr);
@@ -1586,9 +1587,9 @@ static int aarch64_read_apb_ap_memory(struct target *target,
 		/* Step 1.a+b - Write the address for read access into DBGDTRRXint */
 		/* Step 1.c   - Copy value from DTR to R0 using instruction mrc DBGDTRTXint, r0 */
 		retval += dpm->instr_write_data_dcc(dpm,
-				T32_FMTITR(ARMV4_5_MRC(14, 0, 0, 0, 5, 0)), address & ~0x3ULL);
+				ARMV4_5_MRC(14, 0, 0, 0, 5, 0), address & ~0x3ULL);
 		/* Step 1.d - Dummy operation to ensure EDSCR.Txfull == 1 */
-		retval += dpm->instr_execute(dpm, T32_FMTITR(ARMV4_5_MCR(14, 0, 0, 0, 5, 0)));
+		retval += dpm->instr_execute(dpm, ARMV4_5_MCR(14, 0, 0, 0, 5, 0));
 		/* Step 1.e - Change DCC to memory mode */
 		dscr = dscr | DSCR_MA;
 		retval +=  mem_ap_write_atomic_u32(armv8->debug_ap,
@@ -1648,6 +1649,9 @@ static int aarch64_read_apb_ap_memory(struct target *target,
 				armv8->debug_base + CPUV8_DBG_DSCR, &dscr);
 	if (retval != ERROR_OK)
 		goto error_free_buff_r;
+
+	dpm->dscr = dscr;
+
 	if (dscr & (DSCR_ERR | DSCR_SYS_ERROR_PEND)) {
 		/* Abort occurred - clear it and exit */
 		LOG_ERROR("abort occurred - dscr = 0x%08" PRIx32, dscr);
diff --git a/src/target/arm_dpm.h b/src/target/arm_dpm.h
index 63467d5..f8d1248 100644
--- a/src/target/arm_dpm.h
+++ b/src/target/arm_dpm.h
@@ -143,6 +143,9 @@ struct arm_dpm {
 	/** Recent value of DSCR. */
 	uint32_t dscr;
 
+	/** Recent exception level on armv8 */
+	unsigned int last_el;
+
 	/* FIXME -- read/write DCSR methods and symbols */
 };
 
@@ -150,10 +153,8 @@ int arm_dpm_setup(struct arm_dpm *dpm);
 int arm_dpm_initialize(struct arm_dpm *dpm);
 
 int arm_dpm_read_current_registers(struct arm_dpm *);
-int arm_dpm_read_current_registers_64(struct arm_dpm *);
 int dpm_modeswitch(struct arm_dpm *dpm, enum arm_mode mode);
 
-
 int arm_dpm_write_dirty_registers(struct arm_dpm *, bool bpwp);
 
 void arm_dpm_report_wfar(struct arm_dpm *, uint32_t wfar);
diff --git a/src/target/armv8.c b/src/target/armv8.c
index 9b160a9..32fe048 100644
--- a/src/target/armv8.c
+++ b/src/target/armv8.c
@@ -161,55 +161,333 @@ int armv8_mode_to_number(enum arm_mode mode)
 	}
 }
 
-
-static int armv8_read_core_reg(struct target *target, struct reg *r,
-	int num, enum arm_mode mode)
+static int armv8_read_reg(struct armv8_common *armv8, int regnum, uint64_t *regval)
 {
-	uint64_t reg_value;
+	struct arm_dpm *dpm = &armv8->dpm;
 	int retval;
-	struct arm_reg *armv8_core_reg;
-	struct armv8_common *armv8 = target_to_armv8(target);
+	uint32_t value;
+	uint64_t value_64;
+
+	switch (regnum) {
+	case 0 ... 30:
+		retval = dpm->instr_read_data_dcc_64(dpm,
+				ARMV8_MSR_GP(SYSTEM_DBG_DBGDTR_EL0, regnum), &value_64);
+		break;
+	case ARMV8_SP:
+		retval = dpm->instr_read_data_r0_64(dpm,
+				ARMV8_MOVFSP_64(0), &value_64);
+		break;
+	case ARMV8_PC:
+		retval = dpm->instr_read_data_r0_64(dpm,
+				ARMV8_MRS_DLR(0), &value_64);
+		break;
+	case ARMV8_xPSR:
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS_DSPSR(0), &value);
+		value_64 = value;
+		break;
+	case ARMV8_ELR_EL1:
+		retval = dpm->instr_read_data_r0_64(dpm,
+				ARMV8_MRS(SYSTEM_ELR_EL1, 0), &value_64);
+		break;
+	case ARMV8_ELR_EL2:
+		retval = dpm->instr_read_data_r0_64(dpm,
+				ARMV8_MRS(SYSTEM_ELR_EL2, 0), &value_64);
+		break;
+	case ARMV8_ELR_EL3:
+		retval = dpm->instr_read_data_r0_64(dpm,
+				ARMV8_MRS(SYSTEM_ELR_EL3, 0), &value_64);
+		break;
+	case ARMV8_ESR_EL1:
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS(SYSTEM_ESR_EL1, 0), &value);
+		value_64 = value;
+		break;
+	case ARMV8_ESR_EL2:
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS(SYSTEM_ESR_EL2, 0), &value);
+		value_64 = value;
+		break;
+	case ARMV8_ESR_EL3:
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS(SYSTEM_ESR_EL3, 0), &value);
+		value_64 = value;
+		break;
+	case ARMV8_SPSR_EL1:
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS(SYSTEM_SPSR_EL1, 0), &value);
+		value_64 = value;
+		break;
+	case ARMV8_SPSR_EL2:
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS(SYSTEM_SPSR_EL2, 0), &value);
+		value_64 = value;
+		break;
+	case ARMV8_SPSR_EL3:
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS(SYSTEM_SPSR_EL3, 0), &value);
+		value_64 = value;
+		break;
+	default:
+		retval = ERROR_FAIL;
+		break;
+	}
 
-	assert(num < (int)armv8->arm.core_cache->num_regs);
+	if (retval == ERROR_OK && regval != NULL)
+		*regval = value_64;
 
-	armv8_core_reg = armv8->arm.core_cache->reg_list[num].arch_info;
-	retval = armv8->load_core_reg_u64(target,
-			armv8_core_reg->num, &reg_value);
+	return retval;
+}
 
-	buf_set_u64(armv8->arm.core_cache->reg_list[num].value, 0, 64, reg_value);
-	armv8->arm.core_cache->reg_list[num].valid = 1;
-	armv8->arm.core_cache->reg_list[num].dirty = 0;
+static int armv8_write_reg(struct armv8_common *armv8, int regnum, uint64_t value_64)
+{
+	struct arm_dpm *dpm = &armv8->dpm;
+	int retval;
+	uint32_t value;
+
+	switch (regnum) {
+	case 0 ... 30:
+		retval = dpm->instr_write_data_dcc_64(dpm,
+			ARMV8_MRS(SYSTEM_DBG_DBGDTR_EL0, regnum),
+			value_64);
+		break;
+	case ARMV8_SP:
+		retval = dpm->instr_write_data_r0_64(dpm,
+			ARMV8_MOVTSP_64(0),
+			value_64);
+		break;
+	case ARMV8_PC:
+		retval = dpm->instr_write_data_r0_64(dpm,
+			ARMV8_MSR_DLR(0),
+			value_64);
+		break;
+	case ARMV8_xPSR:
+		value = value_64;
+		retval = dpm->instr_write_data_r0(dpm,
+			ARMV8_MSR_DSPSR(0),
+			value);
+		break;
+	/* registers clobbered by taking exception in debug state */
+	case ARMV8_ELR_EL1:
+		retval = dpm->instr_write_data_r0_64(dpm,
+				ARMV8_MSR_GP(SYSTEM_ELR_EL1, 0), value_64);
+		break;
+	case ARMV8_ELR_EL2:
+		retval = dpm->instr_write_data_r0_64(dpm,
+				ARMV8_MSR_GP(SYSTEM_ELR_EL2, 0), value_64);
+		break;
+	case ARMV8_ELR_EL3:
+		retval = dpm->instr_write_data_r0_64(dpm,
+				ARMV8_MSR_GP(SYSTEM_ELR_EL3, 0), value_64);
+		break;
+	case ARMV8_ESR_EL1:
+		value = value_64;
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP(SYSTEM_ESR_EL1, 0), value);
+		break;
+	case ARMV8_ESR_EL2:
+		value = value_64;
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP(SYSTEM_ESR_EL2, 0), value);
+		break;
+	case ARMV8_ESR_EL3:
+		value = value_64;
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP(SYSTEM_ESR_EL3, 0), value);
+		break;
+	case ARMV8_SPSR_EL1:
+		value = value_64;
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP(SYSTEM_SPSR_EL1, 0), value);
+		break;
+	case ARMV8_SPSR_EL2:
+		value = value_64;
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP(SYSTEM_SPSR_EL2, 0), value);
+		break;
+	case ARMV8_SPSR_EL3:
+		value = value_64;
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP(SYSTEM_SPSR_EL3, 0), value);
+		break;
+	default:
+		retval = ERROR_FAIL;
+		break;
+	}
 
 	return retval;
 }
 
-#if 0
-static int armv8_write_core_reg(struct target *target, struct reg *r,
-	int num, enum arm_mode mode, target_addr_t value)
+static int armv8_read_reg32(struct armv8_common *armv8, int regnum, uint64_t *regval)
 {
+	struct arm_dpm *dpm = &armv8->dpm;
+	uint32_t value = 0;
 	int retval;
-	struct arm_reg *armv8_core_reg;
-	struct armv8_common *armv8 = target_to_armv8(target);
 
-	assert(num < (int)armv8->arm.core_cache->num_regs);
+	switch (regnum) {
+	case ARMV8_R0 ... ARMV8_R14:
+		/* return via DCC:  "MCR p14, 0, Rnum, c0, c5, 0" */
+		retval = dpm->instr_read_data_dcc(dpm,
+			ARMV4_5_MCR(14, 0, regnum, 0, 5, 0),
+			&value);
+		break;
+	case ARMV8_SP:
+		retval = dpm->instr_read_data_dcc(dpm,
+			ARMV4_5_MCR(14, 0, 13, 0, 5, 0),
+			&value);
+		break;
+	case ARMV8_PC:
+		retval = dpm->instr_read_data_r0(dpm,
+			ARMV8_MRC_DLR(0),
+			&value);
+		break;
+	case ARMV8_xPSR:
+		retval = dpm->instr_read_data_r0(dpm,
+			ARMV8_MRC_DSPSR(0),
+			&value);
+		break;
+	case ARMV8_ELR_EL1: /* mapped to LR_svc */
+		retval = dpm->instr_read_data_dcc(dpm,
+				ARMV4_5_MCR(14, 0, 14, 0, 5, 0),
+				&value);
+		break;
+	case ARMV8_ELR_EL2: /* mapped to ELR_hyp */
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS_T1(0, 14, 0, 1),
+				&value);
+		break;
+	case ARMV8_ELR_EL3: /* mapped to LR_mon */
+		retval = dpm->instr_read_data_dcc(dpm,
+				ARMV4_5_MCR(14, 0, 14, 0, 5, 0),
+				&value);
+		break;
+	case ARMV8_ESR_EL1: /* mapped to DFSR */
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV4_5_MRC(15, 0, 0, 5, 0, 0),
+				&value);
+		break;
+	case ARMV8_ESR_EL2: /* mapped to HSR */
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV4_5_MRC(15, 4, 0, 5, 2, 0),
+				&value);
+		break;
+	case ARMV8_ESR_EL3: /* FIXME: no equivalent in aarch32? */
+		retval = ERROR_FAIL;
+		break;
+	case ARMV8_SPSR_EL1: /* mapped to SPSR_svc */
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS_xPSR_T1(1, 0),
+				&value);
+		break;
+	case ARMV8_SPSR_EL2: /* mapped to SPSR_hyp */
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS_xPSR_T1(1, 0),
+				&value);
+		break;
+	case ARMV8_SPSR_EL3: /* mapped to SPSR_mon */
+		retval = dpm->instr_read_data_r0(dpm,
+				ARMV8_MRS_xPSR_T1(1, 0),
+				&value);
+		break;
+	default:
+		retval = ERROR_FAIL;
+		break;
+	}
+
+	if (retval == ERROR_OK && regval != NULL)
+		*regval = value;
 
-	armv8_core_reg = armv8->arm.core_cache->reg_list[num].arch_info;
-	retval = armv8->store_core_reg_u64(target,
-					    armv8_core_reg->num,
-					    value);
-	if (retval != ERROR_OK) {
-		LOG_ERROR("JTAG failure");
-		armv8->arm.core_cache->reg_list[num].dirty = armv8->arm.core_cache->reg_list[num].valid;
-		return ERROR_JTAG_DEVICE_ERROR;
+	return retval;
+}
+
+static int armv8_write_reg32(struct armv8_common *armv8, int regnum, uint64_t value)
+{
+	struct arm_dpm *dpm = &armv8->dpm;
+	int retval;
+
+	switch (regnum) {
+	case ARMV8_R0 ... ARMV8_R14:
+		/* load register from DCC:  "MRC p14, 0, Rnum, c0, c5, 0" */
+		retval = dpm->instr_write_data_dcc(dpm,
+				ARMV4_5_MRC(14, 0, regnum, 0, 5, 0), value);
+		break;
+	case ARMV8_SP:
+		retval = dpm->instr_write_data_dcc(dpm,
+			ARMV4_5_MRC(14, 0, 13, 0, 5, 0),
+			value);
+			break;
+	case ARMV8_PC:/* PC
+		 * read r0 from DCC; then "MOV pc, r0" */
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MCR_DLR(0), value);
+		break;
+	case ARMV8_xPSR: /* CPSR */
+		/* read r0 from DCC, then "MCR r0, DSPSR" */
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MCR_DSPSR(0), value);
+		break;
+	case ARMV8_ELR_EL1: /* mapped to LR_svc */
+		retval = dpm->instr_write_data_dcc(dpm,
+				ARMV4_5_MRC(14, 0, 14, 0, 5, 0),
+				value);
+		break;
+	case ARMV8_ELR_EL2: /* mapped to ELR_hyp */
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP_T1(0, 14, 0, 1),
+				value);
+		break;
+	case ARMV8_ELR_EL3: /* mapped to LR_mon */
+		retval = dpm->instr_write_data_dcc(dpm,
+				ARMV4_5_MRC(14, 0, 14, 0, 5, 0),
+				value);
+		break;
+	case ARMV8_ESR_EL1: /* mapped to DFSR */
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV4_5_MCR(15, 0, 0, 5, 0, 0),
+				value);
+		break;
+	case ARMV8_ESR_EL2: /* mapped to HSR */
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV4_5_MCR(15, 4, 0, 5, 2, 0),
+				value);
+		break;
+	case ARMV8_ESR_EL3: /* FIXME: no equivalent in aarch32? */
+		retval = ERROR_FAIL;
+		break;
+	case ARMV8_SPSR_EL1: /* mapped to SPSR_svc */
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP_xPSR_T1(1, 0, 15),
+				value);
+		break;
+	case ARMV8_SPSR_EL2: /* mapped to SPSR_hyp */
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP_xPSR_T1(1, 0, 15),
+				value);
+		break;
+	case ARMV8_SPSR_EL3: /* mapped to SPSR_mon */
+		retval = dpm->instr_write_data_r0(dpm,
+				ARMV8_MSR_GP_xPSR_T1(1, 0, 15),
+				value);
+		break;
+	default:
+		retval = ERROR_FAIL;
+		break;
 	}
 
-	LOG_DEBUG("write core reg %i value 0x%" PRIx64 "", num, value);
-	armv8->arm.core_cache->reg_list[num].valid = 1;
-	armv8->arm.core_cache->reg_list[num].dirty = 0;
+	return retval;
 
-	return ERROR_OK;
 }
-#endif
+
+void armv8_select_reg_access(struct armv8_common *armv8, bool is_aarch64)
+{
+	if (is_aarch64) {
+		armv8->read_reg_u64 = armv8_read_reg;
+		armv8->write_reg_u64 = armv8_write_reg;
+	} else {
+		armv8->read_reg_u64 = armv8_read_reg32;
+		armv8->write_reg_u64 = armv8_write_reg32;
+	}
+}
 
 /*  retrieve core id cluster id  */
 int armv8_read_mpidr(struct armv8_common *armv8)
@@ -306,26 +584,26 @@ static void armv8_show_fault_registers32(struct armv8_common *armv8)
 
 	/* c5/c0 - {data, instruction} fault status registers */
 	retval = dpm->instr_read_data_r0(dpm,
-			T32_FMTITR(ARMV4_5_MRC(15, 0, 0, 5, 0, 0)),
+			ARMV4_5_MRC(15, 0, 0, 5, 0, 0),
 			&dfsr);
 	if (retval != ERROR_OK)
 		goto done;
 
 	retval = dpm->instr_read_data_r0(dpm,
-			T32_FMTITR(ARMV4_5_MRC(15, 0, 0, 5, 0, 1)),
+			ARMV4_5_MRC(15, 0, 0, 5, 0, 1),
 			&ifsr);
 	if (retval != ERROR_OK)
 		goto done;
 
 	/* c6/c0 - {data, instruction} fault address registers */
 	retval = dpm->instr_read_data_r0(dpm,
-			T32_FMTITR(ARMV4_5_MRC(15, 0, 0, 6, 0, 0)),
+			ARMV4_5_MRC(15, 0, 0, 6, 0, 0),
 			&dfar);
 	if (retval != ERROR_OK)
 		goto done;
 
 	retval = dpm->instr_read_data_r0(dpm,
-			T32_FMTITR(ARMV4_5_MRC(15, 0, 0, 6, 0, 2)),
+			ARMV4_5_MRC(15, 0, 0, 6, 0, 2),
 			&ifar);
 	if (retval != ERROR_OK)
 		goto done;
@@ -386,7 +664,7 @@ static __unused int armv8_read_ttbcr32(struct target *target)
 		goto done;
 	/*  MRC p15,0,<Rt>,c2,c0,2 ; Read CP15 Translation Table Base Control Register*/
 	retval = dpm->instr_read_data_r0(dpm,
-			T32_FMTITR(ARMV4_5_MRC(15, 0, 0, 2, 0, 2)),
+			ARMV4_5_MRC(15, 0, 0, 2, 0, 2),
 			&ttbcr);
 	if (retval != ERROR_OK)
 		goto done;
@@ -431,7 +709,7 @@ static __unused int armv8_read_ttbcr(struct target *target)
 	memset(&armv8->armv8_mmu.ttbr1_used, 0, sizeof(armv8->armv8_mmu.ttbr1_used));
 	memset(&armv8->armv8_mmu.ttbr0_mask, 0, sizeof(armv8->armv8_mmu.ttbr0_mask));
 
-	switch (armv8_curel_from_core_mode(arm)) {
+	switch (armv8_curel_from_core_mode(arm->core_mode)) {
 	case SYSTEM_CUREL_EL3:
 		retval = dpm->instr_read_data_r0(dpm,
 				ARMV8_MRS(SYSTEM_TCR_EL3, 0),
@@ -519,7 +797,7 @@ int armv8_mmu_translate_va_pa(struct target *target, target_addr_t va,
 	if (retval != ERROR_OK)
 		return retval;
 
-	switch (armv8_curel_from_core_mode(arm)) {
+	switch (armv8_curel_from_core_mode(arm->core_mode)) {
 	case SYSTEM_CUREL_EL0:
 		instr = ARMV8_SYS(SYSTEM_ATS12E0R, 0);
 		/* can only execute instruction at EL2 */
@@ -602,11 +880,6 @@ int armv8_init_arch_info(struct target *target, struct armv8_common *armv8)
 	armv8->arm.common_magic = ARM_COMMON_MAGIC;
 	armv8->common_magic = ARMV8_COMMON_MAGIC;
 
-	arm->read_core_reg = armv8_read_core_reg;
-#if 0
-	arm->write_core_reg = armv8_write_core_reg;
-#endif
-
 	armv8->armv8_mmu.armv8_cache.l2_cache = NULL;
 	armv8->armv8_mmu.armv8_cache.info = -1;
 	armv8->armv8_mmu.armv8_cache.flush_all_data_cache = NULL;
@@ -673,46 +946,59 @@ static const struct {
 	unsigned id;
 	const char *name;
 	unsigned bits;
+	enum arm_mode mode;
 	enum reg_type type;
 	const char *group;
 	const char *feature;
 } armv8_regs[] = {
-	{ ARMV8_R0,  "x0",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R1,  "x1",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R2,  "x2",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R3,  "x3",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R4,  "x4",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R5,  "x5",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R6,  "x6",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R7,  "x7",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R8,  "x8",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R9,  "x9",  64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R10, "x10", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R11, "x11", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R12, "x12", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R13, "x13", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R14, "x14", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R15, "x15", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R16, "x16", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R17, "x17", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R18, "x18", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R19, "x19", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R20, "x20", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R21, "x21", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R22, "x22", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R23, "x23", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R24, "x24", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R25, "x25", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R26, "x26", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R27, "x27", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R28, "x28", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R29, "x29", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_R30, "x30", 64, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
-
-	{ ARMV8_R31, "sp", 64, REG_TYPE_DATA_PTR, "general", "org.gnu.gdb.aarch64.core" },
-	{ ARMV8_PC,  "pc", 64, REG_TYPE_CODE_PTR, "general", "org.gnu.gdb.aarch64.core" },
-
-	{ ARMV8_xPSR, "CPSR", 32, REG_TYPE_UINT32, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R0,  "x0",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R1,  "x1",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R2,  "x2",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R3,  "x3",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R4,  "x4",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R5,  "x5",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R6,  "x6",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R7,  "x7",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R8,  "x8",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R9,  "x9",  64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R10, "x10", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R11, "x11", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R12, "x12", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R13, "x13", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R14, "x14", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R15, "x15", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R16, "x16", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R17, "x17", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R18, "x18", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R19, "x19", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R20, "x20", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R21, "x21", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R22, "x22", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R23, "x23", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R24, "x24", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R25, "x25", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R26, "x26", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R27, "x27", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R28, "x28", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R29, "x29", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_R30, "x30", 64, ARM_MODE_ANY, REG_TYPE_UINT64, "general", "org.gnu.gdb.aarch64.core" },
+
+	{ ARMV8_SP, "sp", 64, ARM_MODE_ANY, REG_TYPE_DATA_PTR, "general", "org.gnu.gdb.aarch64.core" },
+	{ ARMV8_PC,  "pc", 64, ARM_MODE_ANY, REG_TYPE_CODE_PTR, "general", "org.gnu.gdb.aarch64.core" },
+
+	{ ARMV8_xPSR, "CPSR", 32, ARM_MODE_ANY, REG_TYPE_UINT32, "general", "org.gnu.gdb.aarch64.core" },
+
+	{ ARMV8_ELR_EL1, "ELR_EL1", 64, ARMV8_64_EL1H, REG_TYPE_CODE_PTR, "banked", "net.sourceforge.openocd.banked" },
+	{ ARMV8_ESR_EL1, "ESR_EL1", 32, ARMV8_64_EL1H, REG_TYPE_UINT32, "banked", "net.sourceforge.openocd.banked" },
+	{ ARMV8_SPSR_EL1, "SPSR_EL1", 32, ARMV8_64_EL1H, REG_TYPE_UINT32, "banked", "net.sourceforge.openocd.banked" },
+
+	{ ARMV8_ELR_EL2, "ELR_EL2", 64, ARMV8_64_EL2H, REG_TYPE_CODE_PTR, "banked", "net.sourceforge.openocd.banked" },
+	{ ARMV8_ESR_EL2, "ESR_EL2", 32, ARMV8_64_EL2H, REG_TYPE_UINT32, "banked", "net.sourceforge.openocd.banked" },
+	{ ARMV8_SPSR_EL2, "SPSR_EL2", 32, ARMV8_64_EL2H, REG_TYPE_UINT32, "banked", "net.sourceforge.openocd.banked" },
+
+	{ ARMV8_ELR_EL3, "ELR_EL3", 64, ARMV8_64_EL3H, REG_TYPE_CODE_PTR, "banked", "net.sourceforge.openocd.banked" },
+	{ ARMV8_ESR_EL3, "ESR_EL3", 32, ARMV8_64_EL3H, REG_TYPE_UINT32, "banked", "net.sourceforge.openocd.banked" },
+	{ ARMV8_SPSR_EL3, "SPSR_EL3", 32, ARMV8_64_EL3H, REG_TYPE_UINT32, "banked", "net.sourceforge.openocd.banked" },
 };
 
 #define ARMV8_NUM_REGS ARRAY_SIZE(armv8_regs)
@@ -782,6 +1068,7 @@ struct reg_cache *armv8_build_reg_cache(struct target *target)
 
 	for (i = 0; i < num_regs; i++) {
 		arch_info[i].num = armv8_regs[i].id;
+		arch_info[i].mode = armv8_regs[i].mode;
 		arch_info[i].target = target;
 		arch_info[i].arm = arm;
 
@@ -847,11 +1134,18 @@ int armv8_get_gdb_reg_list(struct target *target,
 
 	switch (reg_class) {
 	case REG_CLASS_GENERAL:
+		*reg_list_size = ARMV8_ELR_EL1;
+		*reg_list = malloc(sizeof(struct reg *) * (*reg_list_size));
+
+		for (i = 0; i < *reg_list_size; i++)
+				(*reg_list)[i] = armv8_reg_current(arm, i);
+
+		return ERROR_OK;
 	case REG_CLASS_ALL:
 		*reg_list_size = ARMV8_LAST_REG;
 		*reg_list = malloc(sizeof(struct reg *) * (*reg_list_size));
 
-		for (i = 0; i < ARMV8_LAST_REG; i++)
+		for (i = 0; i < *reg_list_size; i++)
 				(*reg_list)[i] = armv8_reg_current(arm, i);
 
 		return ERROR_OK;
diff --git a/src/target/armv8.h b/src/target/armv8.h
index 497e482..85845e6 100644
--- a/src/target/armv8.h
+++ b/src/target/armv8.h
@@ -26,7 +26,7 @@
 #include "armv8_dpm.h"
 
 enum {
-	ARMV8_R0,
+	ARMV8_R0 = 0,
 	ARMV8_R1,
 	ARMV8_R2,
 	ARMV8_R3,
@@ -57,11 +57,23 @@ enum {
 	ARMV8_R28,
 	ARMV8_R29,
 	ARMV8_R30,
-	ARMV8_R31,
 
+	ARMV8_SP = 31,
 	ARMV8_PC = 32,
 	ARMV8_xPSR = 33,
 
+	ARMV8_ELR_EL1 = 34,
+	ARMV8_ESR_EL1 = 35,
+	ARMV8_SPSR_EL1 = 36,
+
+	ARMV8_ELR_EL2 = 37,
+	ARMV8_ESR_EL2 = 38,
+	ARMV8_SPSR_EL2 = 39,
+
+	ARMV8_ELR_EL3 = 40,
+	ARMV8_ESR_EL3 = 41,
+	ARMV8_SPSR_EL3 = 42,
+
 	ARMV8_LAST_REG,
 };
 
@@ -162,8 +174,8 @@ struct armv8_common {
 	struct armv8_mmu_common armv8_mmu;
 
 	/* Direct processor core register read and writes */
-	int (*load_core_reg_u64)(struct target *target, uint32_t num, uint64_t *value);
-	int (*store_core_reg_u64)(struct target *target, uint32_t num, uint64_t value);
+	int (*read_reg_u64)(struct armv8_common *armv8, int num, uint64_t *value);
+	int (*write_reg_u64)(struct armv8_common *armv8, int num, uint64_t value);
 
 	int (*examine_debug_reason)(struct target *target);
 	int (*post_debug_entry)(struct target *target);
@@ -270,10 +282,32 @@ int armv8_handle_cache_info_command(struct command_context *cmd_ctx,
 
 void armv8_set_cpsr(struct arm *arm, uint32_t cpsr);
 
-static inline int armv8_curel_from_core_mode(struct arm *arm)
+static inline unsigned int armv8_curel_from_core_mode(enum arm_mode core_mode)
 {
-	return (arm->core_mode >> 6) & 3;
+	switch (core_mode) {
+	/* Aarch32 modes */
+	case ARM_MODE_USR:
+		return 0;
+	case ARM_MODE_SVC:
+	case ARM_MODE_ABT: /* FIXME: EL3? */
+	case ARM_MODE_IRQ: /* FIXME: EL3? */
+	case ARM_MODE_FIQ: /* FIXME: EL3? */
+	case ARM_MODE_UND: /* FIXME: EL3? */
+	case ARM_MODE_SYS: /* FIXME: EL3? */
+		return 1;
+	/* case ARM_MODE_HYP:
+	 *     return 2;
+	 */
+	case ARM_MODE_MON:
+		return 3;
+	/* all Aarch64 modes */
+	default:
+		return (core_mode >> 6) & 3;
+	}
 }
+
+void armv8_select_reg_access(struct armv8_common *armv8, bool is_aarch64);
+
 extern const struct command_registration armv8_command_handlers[];
 
 #endif
diff --git a/src/target/armv8_dpm.c b/src/target/armv8_dpm.c
index 4bb2075..a92f92c 100644
--- a/src/target/armv8_dpm.c
+++ b/src/target/armv8_dpm.c
@@ -28,6 +28,8 @@
 
 #include "helper/time_support.h"
 
+/* T32 ITR format */
+#define T32_FMTITR(instr) (((instr & 0x0000FFFF) << 16) | ((instr & 0xFFFF0000) >> 16))
 
 /**
  * @file
@@ -42,6 +44,29 @@
  * to minimize needless differences in debug behavior between cores.
  */
 
+/**
+ * Get core state from EDSCR, without necessity to retrieve CPSR
+ */
+enum arm_state armv8_dpm_get_core_state(struct arm_dpm *dpm)
+{
+	int el = (dpm->dscr >> 8) & 0x3;
+	int rw = (dpm->dscr >> 10) & 0xF;
+	int pos;
+
+	dpm->last_el = el;
+
+	/* find the first '0' in DSCR.RW */
+	for (pos = 3; pos >= 0; pos--) {
+		if ((rw & (1 << pos)) == 0)
+			break;
+	}
+
+	if (el > pos)
+		return ARM_STATE_AARCH64;
+
+	return ARM_STATE_ARM;
+}
+
 /*----------------------------------------------------------------------*/
 
 static int dpmv8_write_dcc(struct armv8_common *armv8, uint32_t data)
@@ -166,6 +191,9 @@ static int dpmv8_dpm_prepare(struct arm_dpm *dpm)
 		}
 	}
 
+	/* update the stored copy of dscr */
+	dpm->dscr = dscr;
+
 	/* this "should never happen" ... */
 	if (dscr & DSCR_DTR_RX_FULL) {
 		LOG_ERROR("DSCR_DTR_RX_FULL, dscr 0x%08" PRIx32, dscr);
@@ -218,6 +246,9 @@ static int dpmv8_exec_opcode(struct arm_dpm *dpm,
 		}
 	}
 
+	if (armv8_dpm_get_core_state(dpm) != ARM_STATE_AARCH64)
+		opcode = T32_FMTITR(opcode);
+
 	retval = mem_ap_write_u32(armv8->debug_ap,
 			armv8->debug_base + CPUV8_DBG_ITR, opcode);
 	if (retval != ERROR_OK)
@@ -237,6 +268,20 @@ static int dpmv8_exec_opcode(struct arm_dpm *dpm,
 		}
 	} while ((dscr & DSCR_ITE) == 0);	/* Wait for InstrCompl bit to be set */
 
+	/* update dscr and el after each command execution */
+	dpm->dscr = dscr;
+	if (dpm->last_el != ((dscr >> 8) & 3))
+		LOG_DEBUG("EL %i -> %i", dpm->last_el, (dscr >> 8) & 3);
+	dpm->last_el = (dscr >> 8) & 3;
+
+	if (dscr & DSCR_ERR) {
+		LOG_ERROR("Opcode 0x%08"PRIx32", DSCR.ERR=1, DSCR.EL=%i", opcode, dpm->last_el);
+		/* clear the sticky error condition */
+		mem_ap_write_atomic_u32(armv8->debug_ap,
+				armv8->debug_base + CPUV8_DBG_DRCR, DRCR_CSE);
+		retval = ERROR_FAIL;
+	}
+
 	if (p_dscr)
 		*p_dscr = dscr;
 
@@ -471,7 +516,7 @@ static int dpmv8_mrc(struct target *target, int cpnum,
 
 	/* read coprocessor register into R0; return via DCC */
 	retval = dpm->instr_read_data_r0(dpm,
-			T32_FMTITR(ARMV4_5_MRC(cpnum, op1, 0, CRn, CRm, op2)),
+			ARMV4_5_MRC(cpnum, op1, 0, CRn, CRm, op2),
 			value);
 
 	/* (void) */ dpm->finish(dpm);
@@ -496,7 +541,7 @@ static int dpmv8_mcr(struct target *target, int cpnum,
 
 	/* read DCC into r0; then write coprocessor register from R0 */
 	retval = dpm->instr_write_data_r0(dpm,
-			T32_FMTITR(ARMV4_5_MCR(cpnum, op1, 0, CRn, CRm, op2)),
+			ARMV4_5_MCR(cpnum, op1, 0, CRn, CRm, op2),
 			value);
 
 	/* (void) */ dpm->finish(dpm);
@@ -616,210 +661,50 @@ int dpmv8_modeswitch(struct arm_dpm *dpm, enum arm_mode mode)
 	return retval;
 }
 
-static int dpmv8_read_reg32(struct arm_dpm *dpm, struct reg *r, unsigned regnum)
-{
-	uint32_t value;
-	int retval = ERROR_FAIL;
-	bool valid = true;
-
-	switch (regnum) {
-	case 0 ... 14:
-		/* return via DCC:  "MCR p14, 0, Rnum, c0, c5, 0" */
-		retval = dpm->instr_read_data_dcc(dpm,
-			T32_FMTITR(ARMV4_5_MCR(14, 0, regnum, 0, 5, 0)),
-			&value);
-		break;
-		case ARMV8_R31:
-			retval = dpm->instr_read_data_dcc(dpm,
-				T32_FMTITR(ARMV4_5_MCR(14, 0, 13, 0, 5, 0)),
-				&value);
-				break;
-		case ARMV8_PC:
-			retval = dpm->instr_read_data_r0(dpm,
-				T32_FMTITR(ARMV8_MRC_DLR(0)),
-				&value);
-			break;
-		case ARMV8_xPSR:
-			retval = dpm->instr_read_data_r0(dpm,
-				T32_FMTITR(ARMV8_MRC_DSPSR(0)),
-				&value);
-			break;
-		default:
-			LOG_DEBUG("READ: %s ignored", r->name);
-			retval = ERROR_OK;
-			value = 0xFFFFFFFF;
-			valid = false;
-			break;
-	}
-
-	if (retval == ERROR_OK) {
-		r->valid = valid;
-		r->dirty = false;
-		buf_set_u64(r->value, 0, 32, value);
-		LOG_DEBUG("READ: %s, %8.8x", r->name, (unsigned) value);
-	}
-	return retval;
-}
-
-static int dpmv8_write_reg32(struct arm_dpm *dpm, struct reg *r, unsigned regnum)
-{
-	int retval;
-	uint64_t value = buf_get_u64(r->value, 0, 32);
-
-	switch (regnum) {
-		case 0 ... 14:
-			/* load register from DCC:  "MRC p14, 0, Rnum, c0, c5, 0" */
-			retval = dpm->instr_write_data_dcc(dpm,
-					T32_FMTITR(ARMV4_5_MRC(14, 0, regnum, 0, 5, 0)), value);
-			break;
-		case ARMV8_PC:/* PC
-			 * read r0 from DCC; then "MOV pc, r0" */
-			retval = dpm->instr_write_data_r0(dpm,
-					T32_FMTITR(ARMV8_MCR_DLR(0)), value);
-			break;
-		case ARMV8_xPSR: /* CPSR */
-			/* read r0 from DCC, then "MCR r0, DSPSR" */
-			retval = dpm->instr_write_data_r0(dpm,
-					T32_FMTITR(ARMV8_MCR_DSPSR(0)), value);
-			break;
-		default:
-			retval = ERROR_OK;
-			LOG_DEBUG("WRITE: %s ignored", r->name);
-			break;
-	}
-
-	if (retval == ERROR_OK) {
-		r->dirty = false;
-		LOG_DEBUG("WRITE: %s, %8.8x", r->name, (unsigned) value);
-	}
-
-	return retval;
-}
-
-/* just read the register -- rely on the core mode being right */
+/*
+ * Common register read, relies on armv8_select_reg_access() having been called.
+ */
 static int dpmv8_read_reg(struct arm_dpm *dpm, struct reg *r, unsigned regnum)
 {
-	uint32_t value;
+	struct armv8_common *armv8 = dpm->arm->arch_info;
 	uint64_t value_64;
-	int retval = ERROR_FAIL;
+	int retval;
 
-	switch (regnum) {
-		case 0 ... 30:
-			retval = dpm->instr_read_data_dcc_64(dpm,
-				ARMV8_MSR_GP(SYSTEM_DBG_DBGDTR_EL0, regnum),
-				&value_64);
-			break;
-		case ARMV8_R31:
-			retval = dpm->instr_read_data_r0_64(dpm,
-				ARMV8_MOVFSP_64(0),
-				&value_64);
-			break;
-		case ARMV8_PC:
-			retval = dpm->instr_read_data_r0_64(dpm,
-				ARMV8_MRS_DLR(0),
-				&value_64);
-			break;
-		case ARMV8_xPSR:
-			retval = dpm->instr_read_data_r0(dpm,
-				ARMV8_MRS_DSPSR(0),
-				&value);
-			break;
-		default:
-			LOG_DEBUG("READ: %s fail", r->name);
-			break;
-	}
+	retval = armv8->read_reg_u64(armv8, regnum, &value_64);
 
 	if (retval == ERROR_OK) {
 		r->valid = true;
 		r->dirty = false;
-		if (r->size == 64) {
-			buf_set_u64(r->value, 0, 64, value_64);
+		buf_set_u64(r->value, 0, r->size, value_64);
+		if (r->size == 64)
 			LOG_DEBUG("READ: %s, %16.8llx", r->name, (unsigned long long) value_64);
-		} else {
-			buf_set_u32(r->value, 0, 32, value);
-			LOG_DEBUG("READ: %s, %8.8x", r->name, (unsigned) value);
-		}
+		else
+			LOG_DEBUG("READ: %s, %8.8x", r->name, (unsigned int) value_64);
 	}
-	return retval;
+	return ERROR_OK;
 }
 
-/* just write the register -- rely on the core mode being right */
+/*
+ * Common register write, relies on armv8_select_reg_access() having been called.
+ */
 static int dpmv8_write_reg(struct arm_dpm *dpm, struct reg *r, unsigned regnum)
 {
+	struct armv8_common *armv8 = dpm->arm->arch_info;
 	int retval = ERROR_FAIL;
-	uint32_t value = 0xFFFFFFFF;
-	uint64_t value_64 = 0xFFFFFFFFFFFFFFFF;
-
-	switch (regnum) {
-		case 0 ... 30:
-			value_64 = buf_get_u64(r->value, 0, 64);
-			retval = dpm->instr_write_data_dcc_64(dpm,
-				ARMV8_MRS(SYSTEM_DBG_DBGDTR_EL0, regnum),
-				value_64);
-			break;
-		case ARMV8_R31:
-			value_64 = buf_get_u64(r->value, 0, 64);
-			retval = dpm->instr_write_data_r0_64(dpm,
-				ARMV8_MOVTSP_64(0),
-				value_64);
-			break;
-		case ARMV8_PC:
-			value_64 = buf_get_u64(r->value, 0, 64);
-			retval = dpm->instr_write_data_r0_64(dpm,
-				ARMV8_MSR_DLR(0),
-				value_64);
-			break;
-		case ARMV8_xPSR:
-			value = buf_get_u32(r->value, 0, 32);
-			retval = dpm->instr_write_data_r0(dpm,
-				ARMV8_MSR_DSPSR(0),
-				value);
-			break;
-		default:
-			LOG_DEBUG("write: %s fail", r->name);
-			break;
-	}
+	uint64_t value_64;
 
+	value_64 = buf_get_u64(r->value, 0, r->size);
 
+	retval = armv8->write_reg_u64(armv8, regnum, value_64);
 	if (retval == ERROR_OK) {
 		r->dirty = false;
 		if (r->size == 64)
-			LOG_DEBUG("WRITE: %s, %16.8llx", r->name, (unsigned long long) value_64);
+			LOG_DEBUG("WRITE: %s, %16.8llx", r->name, (unsigned long long)value_64);
 		else
-			LOG_DEBUG("WRITE: %s, %8.8x", r->name, (unsigned) value);
+			LOG_DEBUG("WRITE: %s, %8.8x", r->name, (unsigned int)value_64);
 	}
 
-	return retval;
-}
-
-static inline enum arm_state dpm_get_core_state(uint32_t dscr)
-{
-	int el = (dscr >> 8) & 0x3;
-	int rw = (dscr >> 10) & 0xF;
-
-	LOG_DEBUG("EL:%i, RW:0x%x", el, rw);
-
-	/* DSCR.RW = 0b1111 - all EL are using AArch64 state */
-	if (rw == 0xF)
-		return ARM_STATE_AARCH64;
-
-	/* DSCR.RW = 0b1110 - all EL > 0 are using AArch64 state */
-	if (rw == 0xE && el > 0)
-		return ARM_STATE_AARCH64;
-
-	/* DSCR.RW = 0b110x - all EL > 1 are using Aarch64 state */
-	if ((rw & 0xE) == 0xC && el > 1)
-		return ARM_STATE_AARCH64;
-
-	/* DSCR.RW = 0b10xx - all EL > 2 are using Aarch64 state */
-	if ((rw & 0xC) == 0x8 && el > 2)
-		return ARM_STATE_AARCH64;
-
-	/* DSCR.RW = 0b0xxx - all EL are using AArch32 state */
-	if ((rw & 0x8) == 0)
-		return ARM_STATE_ARM;
-
-	return ARM_STATE_ARM;
+	return ERROR_OK;
 }
 
 /**
@@ -833,60 +718,59 @@ int armv8_dpm_read_current_registers(struct arm_dpm *dpm)
 {
 	struct arm *arm = dpm->arm;
 	struct armv8_common *armv8 = (struct armv8_common *)arm->arch_info;
-	enum arm_state core_state;
+	struct reg_cache *cache;
+	struct reg *r;
 	uint32_t cpsr;
-
 	int retval;
-	struct reg *r;
 
 	retval = dpm->prepare(dpm);
 	if (retval != ERROR_OK)
 		return retval;
 
-	core_state = dpm_get_core_state(dpm->dscr);
-
-	armv8_select_opcodes(armv8, core_state);
+	cache = arm->core_cache;
 
 	/* read R0 first (it's used for scratch), then CPSR */
-	r = arm->core_cache->reg_list + 0;
+	r = cache->reg_list + 0;
 	if (!r->valid) {
-		retval = core_state == ARM_STATE_AARCH64 ?
-					dpmv8_read_reg(dpm, r, 0) : dpmv8_read_reg32(dpm, r, 0);
+		retval = dpmv8_read_reg(dpm, r, 0);
 		if (retval != ERROR_OK)
 			goto fail;
 	}
 	r->dirty = true;
 
 	/* read cpsr to r0 and get it back */
-	retval = dpm->instr_read_data_r0(dpm, armv8_opcode(armv8, READ_REG_DSPSR), &cpsr);
+	retval = dpm->instr_read_data_r0(dpm,
+			armv8_opcode(armv8, READ_REG_DSPSR), &cpsr);
 	if (retval != ERROR_OK)
 		goto fail;
 
-	/* update core mode and state, plus shadow mapping for R8..R14 */
+	/* update core mode and state */
 	armv8_set_cpsr(arm, cpsr);
 
-	/* REVISIT we can probably avoid reading R1..R14, saving time... */
-	for (unsigned i = 1; i < arm->core_cache->num_regs ; i++) {
+	for (unsigned int i = 1; i < cache->num_regs ; i++) {
+		struct arm_reg *arm_reg;
+
 		r = armv8_reg_current(arm, i);
 		if (r->valid)
 			continue;
 
-		retval = core_state == ARM_STATE_AARCH64 ?
-					dpmv8_read_reg(dpm, r, i) : dpmv8_read_reg32(dpm, r, i);
+		/*
+		 * Only read registers that are available from the
+		 * current EL (or core mode).
+		 */
+		arm_reg = r->arch_info;
+		if (arm_reg->mode != ARM_MODE_ANY &&
+				dpm->last_el != armv8_curel_from_core_mode(arm_reg->mode))
+			continue;
 
+		retval = dpmv8_read_reg(dpm, r, i);
 		if (retval != ERROR_OK)
 			goto fail;
-	}
-
-	/* NOTE: SPSR ignored (if it's even relevant). */
 
-	/* REVISIT the debugger can trigger various exceptions.  See the
-	 * ARMv7A architecture spec, section C5.7, for more info about
-	 * what defenses are needed; v6 debug has the most issues.
-	 */
+	}
 
 fail:
-	/* (void) */ dpm->finish(dpm);
+	dpm->finish(dpm);
 	return retval;
 }
 
@@ -951,7 +835,6 @@ int armv8_dpm_write_dirty_registers(struct arm_dpm *dpm, bool bpwp)
 	struct arm *arm = dpm->arm;
 	struct reg_cache *cache = arm->core_cache;
 	int retval;
-	bool is_aarch64 = arm->core_state == ARM_STATE_AARCH64;
 
 	retval = dpm->prepare(dpm);
 	if (retval != ERROR_OK)
@@ -991,59 +874,48 @@ int armv8_dpm_write_dirty_registers(struct arm_dpm *dpm, bool bpwp)
 	 * be queued, and need (efficient/batched) flushing later.
 	 */
 
-	/* Scan the registers until we find one that's both dirty and
-	 * eligible for flushing.  Flush that and everything else that
-	 * shares the same core mode setting.  Typically this won't
-	 * actually find anything to do...
-	 */
+	/* Restore original core mode and state */
+	retval = dpmv8_modeswitch(dpm, ARM_MODE_ANY);
+	if (retval != ERROR_OK)
+		goto done;
 
 	/* check everything except our scratch register R0 */
 	for (unsigned i = 1; i < cache->num_regs; i++) {
 		struct arm_reg *r;
-		unsigned regnum;
 
-		/* also skip PC, CPSR, and non-dirty */
-		if (i == (arm->core_cache->num_regs - 2))
+		/* skip PC and CPSR */
+		if (i == ARMV8_PC || i == ARMV8_xPSR)
 			continue;
-		if (arm->cpsr == cache->reg_list + i)
+		/* skip invalid */
+		if (!cache->reg_list[i].valid)
 			continue;
+		/* skip non-dirty */
 		if (!cache->reg_list[i].dirty)
 			continue;
 
+		/* skip all registers not on the current EL */
 		r = cache->reg_list[i].arch_info;
-		regnum = r->num;
+		if (r->mode != ARM_MODE_ANY &&
+				dpm->last_el != armv8_curel_from_core_mode(r->mode))
+			continue;
 
-		retval = is_aarch64 ? dpmv8_write_reg(dpm, &cache->reg_list[i], regnum)
-				: dpmv8_write_reg32(dpm, &cache->reg_list[i], regnum);
+		retval = dpmv8_write_reg(dpm, &cache->reg_list[i], i);
 		if (retval != ERROR_OK)
-			goto done;
+			break;
 	}
 
-
-	/* Restore original CPSR ... assuming either that we changed it,
-	 * or it's dirty.  Must write PC to ensure the return address is
-	 * defined, and must not write it before CPSR.
-	 */
-	retval = dpmv8_modeswitch(dpm, ARM_MODE_ANY);
-	if (retval != ERROR_OK)
-		goto done;
-	arm->cpsr->dirty = false;
-
-	retval = is_aarch64 ? dpmv8_write_reg(dpm, arm->pc, (arm->core_cache->num_regs - 2))
-			: dpmv8_write_reg32(dpm, arm->pc, (arm->core_cache->num_regs - 2));
-	if (retval != ERROR_OK)
-		goto done;
-	arm->pc->dirty = false;
-
+	/* flush CPSR and PC */
+	if (retval == ERROR_OK)
+		retval = dpmv8_write_reg(dpm, &cache->reg_list[ARMV8_xPSR], ARMV8_xPSR);
+	if (retval == ERROR_OK)
+		retval = dpmv8_write_reg(dpm, &cache->reg_list[ARMV8_PC], ARMV8_PC);
 	/* flush R0 -- it's *very* dirty by now */
-	retval = is_aarch64 ? dpmv8_write_reg(dpm, &cache->reg_list[0], 0)
-			: dpmv8_write_reg32(dpm, &cache->reg_list[0], 0);
-	if (retval != ERROR_OK)
-		goto done;
-	cache->reg_list[0].dirty = false;
-
-	/* (void) */ dpm->finish(dpm);
+	if (retval == ERROR_OK)
+		retval = dpmv8_write_reg(dpm, &cache->reg_list[0], 0);
+	if (retval == ERROR_OK)
+		dpm->instr_cpsr_sync(dpm);
 done:
+	dpm->finish(dpm);
 	return retval;
 }
 
@@ -1061,19 +933,18 @@ static int armv8_dpm_read_core_reg(struct target *target, struct reg *r,
 	int retval;
 	int max = arm->core_cache->num_regs;
 
-	if (regnum < 0 || regnum > max)
+	if (regnum < 0 || regnum >= max)
 		return ERROR_COMMAND_SYNTAX_ERROR;
 
-	/* REVISIT what happens if we try to read SPSR in a core mode
+	/*
+	 * REVISIT what happens if we try to read SPSR in a core mode
 	 * which has no such register?
 	 */
-
 	retval = dpm->prepare(dpm);
 	if (retval != ERROR_OK)
 		return retval;
 
-	retval = arm->core_state == ARM_STATE_AARCH64 ?
-			dpmv8_read_reg(dpm, r, regnum) : dpmv8_read_reg32(dpm, r, regnum);
+	retval = dpmv8_read_reg(dpm, r, regnum);
 	if (retval != ERROR_OK)
 		goto fail;
 
@@ -1101,12 +972,11 @@ static int armv8_dpm_write_core_reg(struct target *target, struct reg *r,
 	if (retval != ERROR_OK)
 		return retval;
 
-	retval = arm->core_state == ARM_STATE_AARCH64 ?
-			dpmv8_write_reg(dpm, r, regnum) : dpmv8_write_reg32(dpm, r, regnum);
+	retval = dpmv8_write_reg(dpm, r, regnum);
 
 	/* always clean up, regardless of error */
+	dpm->finish(dpm);
 
-	/* (void) */ dpm->finish(dpm);
 	return retval;
 }
 
@@ -1402,6 +1272,7 @@ void armv8_dpm_report_dscr(struct arm_dpm *dpm, uint32_t dscr)
 	struct target *target = dpm->arm->target;
 
 	dpm->dscr = dscr;
+	dpm->last_el = (dscr >> 8) & 3;
 
 	/* Examine debug reason */
 	switch (DSCR_ENTRY(dscr)) {
diff --git a/src/target/armv8_dpm.h b/src/target/armv8_dpm.h
index 416f4df..6470461 100644
--- a/src/target/armv8_dpm.h
+++ b/src/target/armv8_dpm.h
@@ -111,5 +111,6 @@ void armv8_dpm_report_wfar(struct arm_dpm *, uint64_t wfar);
 #define PRSR_SDR				(1 << 11)
 
 void armv8_dpm_report_dscr(struct arm_dpm *dpm, uint32_t dcsr);
+enum arm_state armv8_dpm_get_core_state(struct arm_dpm *dpm);
 
 #endif /* OPENOCD_TARGET_ARM_DPM_H */
diff --git a/src/target/armv8_opcodes.c b/src/target/armv8_opcodes.c
index 78b60e0..3e18a02 100644
--- a/src/target/armv8_opcodes.c
+++ b/src/target/armv8_opcodes.c
@@ -37,16 +37,16 @@ static const uint32_t a64_opcodes[ARMV8_OPC_NUM] = {
 };
 
 static const uint32_t t32_opcodes[ARMV8_OPC_NUM] = {
-		[READ_REG_CLIDR]	= T32_FMTITR(ARMV4_5_MRC(15, 1, 0, 0, 0, 1)),
-		[READ_REG_CSSELR]	= T32_FMTITR(ARMV4_5_MRC(15, 2, 0, 0, 0, 0)),
-		[READ_REG_CCSIDR]	= T32_FMTITR(ARMV4_5_MRC(15, 1, 0, 0, 0, 0)),
-		[WRITE_REG_CSSELR]	= T32_FMTITR(ARMV4_5_MCR(15, 2, 0, 0, 0, 0)),
-		[READ_REG_MPIDR]	= T32_FMTITR(ARMV4_5_MRC(15, 0, 0, 0, 0, 5)),
-		[READ_REG_DTRRX]	= T32_FMTITR(ARMV4_5_MRC(14, 0, 0, 0, 5, 0)),
-		[WRITE_REG_DTRTX]	= T32_FMTITR(ARMV4_5_MCR(14, 0, 0, 0, 5, 0)),
-		[WRITE_REG_DSPSR]	= T32_FMTITR(ARMV8_MCR_DSPSR(0)),
-		[READ_REG_DSPSR]	= T32_FMTITR(ARMV8_MRC_DSPSR(0)),
-		[ARMV8_OPC_DSB_SY]	= T32_FMTITR(ARMV8_DSB_SY_T1),
+		[READ_REG_CLIDR]	= ARMV4_5_MRC(15, 1, 0, 0, 0, 1),
+		[READ_REG_CSSELR]	= ARMV4_5_MRC(15, 2, 0, 0, 0, 0),
+		[READ_REG_CCSIDR]	= ARMV4_5_MRC(15, 1, 0, 0, 0, 0),
+		[WRITE_REG_CSSELR]	= ARMV4_5_MCR(15, 2, 0, 0, 0, 0),
+		[READ_REG_MPIDR]	= ARMV4_5_MRC(15, 0, 0, 0, 0, 5),
+		[READ_REG_DTRRX]	= ARMV4_5_MRC(14, 0, 0, 0, 5, 0),
+		[WRITE_REG_DTRTX]	= ARMV4_5_MCR(14, 0, 0, 0, 5, 0),
+		[WRITE_REG_DSPSR]	= ARMV8_MCR_DSPSR(0),
+		[READ_REG_DSPSR]	= ARMV8_MRC_DSPSR(0),
+		[ARMV8_OPC_DSB_SY]	= ARMV8_DSB_SY_T1,
 };
 
 void armv8_select_opcodes(struct armv8_common *armv8, bool state_is_aarch64)
diff --git a/src/target/armv8_opcodes.h b/src/target/armv8_opcodes.h
index 41abe04..cb20c84 100644
--- a/src/target/armv8_opcodes.h
+++ b/src/target/armv8_opcodes.h
@@ -66,7 +66,6 @@
 
 #define SYSTEM_ISR_EL1			0b1100011000001000
 
-
 #define SYSTEM_DBG_DSPSR_EL0    0b1101101000101000
 #define SYSTEM_DBG_DLR_EL0		0b1101101000101001
 #define SYSTEM_DBG_DTRRX_EL0	0b1001100000101000
@@ -103,14 +102,19 @@
 #define SYSTEM_ATS1E2R			0b0110001111000000
 #define SYSTEM_ATS1E3R			0b0111001111000000
 
+/* fault status and fault address */
+#define SYSTEM_FAR_EL1			0b1100001100000000
+#define SYSTEM_FAR_EL2			0b1110001100000000
+#define SYSTEM_FAR_EL3			0b1111001100000000
+#define SYSTEM_ESR_EL1			0b1100001010010000
+#define SYSTEM_ESR_EL2			0b1110001010010000
+#define SYSTEM_ESR_EL3			0b1111001010010000
+
 #define ARMV8_MRS_DSPSR(Rt)	(0xd53b4500 | (Rt))
 #define ARMV8_MSR_DSPSR(Rt)	(0xd51b4500 | (Rt))
 #define ARMV8_MRS_DLR(Rt)	(0xd53b4520 | (Rt))
 #define ARMV8_MSR_DLR(Rt)	(0xd51b4520 | (Rt))
 
-/* T32 ITR format */
-#define T32_FMTITR(instr) (((instr & 0x0000FFFF) << 16) | ((instr & 0xFFFF0000) >> 16))
-
 /* T32 instruction to access coprocessor registers */
 #define ARMV8_MCR_T1(cp, CRn, opc1, CRm, opc2, Rt) ARMV4_5_MCR(cp, opc1, Rt, CRn, CRm, opc2)
 #define ARMV8_MRC_T1(cp, CRn, opc1, CRm, opc2, Rt) ARMV4_5_MRC(cp, opc1, Rt, CRn, CRm, opc2)
@@ -136,6 +140,11 @@
 #define ARMV8_MSR_IM(Op1, CRm, Op2) \
 	(0xd500401f | ((Op1) << 16)  | ((CRm) << 8) | ((Op2) << 5))
 
+#define ARMV8_MRS_T1(R, M1, Rd, M) (0xF3E08020 | (R << 20) | (M1 << 16) | (Rd << 8) | (M << 4))
+#define ARMV8_MRS_xPSR_T1(R, Rd) (0xF3EF8000 | (R << 20) | (Rd << 8))
+#define ARMV8_MSR_GP_T1(R, M1, Rd, M) (0xF3808020 | (R << 20) | (M1 << 8) | (Rd << 16) | (M << 4))
+#define ARMV8_MSR_GP_xPSR_T1(R, Rn, mask) (0xF3808000 | (R << 20) | (Rn << 16) | (mask << 8))
+
 #define ARMV8_BKPT(Im) (0xD4200000 | ((Im & 0xffff) << 5))
 #define ARMV8_HLT(Im) (0x0D4400000 | ((Im & 0xffff) << 5))