Merge pull request #797 from riscv/Zve32

If XLEN=64 and vsew=64 fails, fall back to vsew=32.

3 files changed, 65 insertions, 39 deletions

diff --git a/src/target/riscv/riscv-013.c b/src/target/riscv/riscv-013.c
index 661197c..71cf287 100644
--- a/src/target/riscv/riscv-013.c
+++ b/src/target/riscv/riscv-013.c
@@ -116,12 +116,6 @@ struct trigger {
 	int unique_id;
 };
 
-typedef enum {
-	YNM_MAYBE,
-	YNM_YES,
-	YNM_NO
-} yes_no_maybe_t;
-
 #define HART_INDEX_MULTIPLE	-1
 #define HART_INDEX_UNKNOWN	-2
 
@@ -1738,8 +1732,8 @@ static int examine(struct target *target)
 			LOG_TARGET_WARNING(target, "Couldn't read vlenb; vector register access won't work.");
 		r->vlenb = 0;
 	} else {
-		LOG_TARGET_INFO(target, "Vector support with vlenb=%d", r->vlenb);
 		r->vlenb = vlenb;
+		LOG_TARGET_INFO(target, "Vector support with vlenb=%d", r->vlenb);
 	}
 
 	/* Now init registers based on what we discovered. */
@@ -1911,34 +1905,54 @@ COMMAND_HELPER(riscv013_print_info, struct target *target)
 	return 0;
 }
 
-static int prep_for_vector_access(struct target *target, uint64_t *vtype,
-		uint64_t *vl, unsigned *debug_vl)
+static int prep_for_vector_access(struct target *target, uint64_t *saved_vtype,
+		uint64_t *saved_vl, unsigned *debug_vl, unsigned *debug_vsew)
 {
 	RISCV_INFO(r);
 	/* TODO: this continuous save/restore is terrible for performance. */
 	/* Write vtype and vl. */
-	unsigned encoded_vsew;
-	switch (riscv_xlen(target)) {
-		case 32:
-			encoded_vsew = 2;
-			break;
-		case 64:
-			encoded_vsew = 3;
-			break;
-		default:
-			LOG_ERROR("Unsupported xlen: %d", riscv_xlen(target));
-			return ERROR_FAIL;
-	}
 
 	/* Save vtype and vl. */
-	if (register_read_direct(target, vtype, GDB_REGNO_VTYPE) != ERROR_OK)
+	if (register_read_direct(target, saved_vtype, GDB_REGNO_VTYPE) != ERROR_OK)
 		return ERROR_FAIL;
-	if (register_read_direct(target, vl, GDB_REGNO_VL) != ERROR_OK)
+	if (register_read_direct(target, saved_vl, GDB_REGNO_VL) != ERROR_OK)
 		return ERROR_FAIL;
 
-	if (register_write_direct(target, GDB_REGNO_VTYPE, encoded_vsew << 3) != ERROR_OK)
-		return ERROR_FAIL;
-	*debug_vl = DIV_ROUND_UP(r->vlenb * 8, riscv_xlen(target));
+	while (true) {
+		unsigned int encoded_vsew;
+		if (riscv_xlen(target) == 64 && r->vsew64_supported != YNM_NO) {
+			encoded_vsew = 3;
+			*debug_vsew = 64;
+		} else {
+			encoded_vsew = 2;
+			*debug_vsew = 32;
+		}
+
+		/* Set standard element width to match XLEN, for vmv instruction to move
+		 * the least significant bits into a GPR. */
+		if (register_write_direct(target, GDB_REGNO_VTYPE, encoded_vsew << 3) != ERROR_OK)
+			return ERROR_FAIL;
+
+		if (*debug_vsew == 64 && r->vsew64_supported == YNM_MAYBE) {
+			/* Check that it's supported. */
+			uint64_t vtype;
+			if (register_read_direct(target, &vtype, GDB_REGNO_VTYPE) != ERROR_OK)
+				return ERROR_FAIL;
+			if (vtype >> (riscv_xlen(target) - 1)) {
+				r->vsew64_supported = YNM_NO;
+				/* Try again. */
+				continue;
+			} else {
+				r->vsew64_supported = YNM_YES;
+			}
+		}
+		break;
+	}
+
+	/* Set the number of elements to be updated with results from a vector
+	 * instruction, for the vslide1down instruction.
+	 * Set it so the entire V register is updated. */
+	*debug_vl = DIV_ROUND_UP(r->vlenb * 8, *debug_vsew);
 	if (register_write_direct(target, GDB_REGNO_VL, *debug_vl) != ERROR_OK)
 		return ERROR_FAIL;
 
@@ -1972,20 +1986,21 @@ static int riscv013_get_register_buf(struct target *target,
 		return ERROR_FAIL;
 
 	uint64_t vtype, vl;
-	unsigned debug_vl;
-	if (prep_for_vector_access(target, &vtype, &vl, &debug_vl) != ERROR_OK)
+	unsigned int debug_vl, debug_vsew;
+	if (prep_for_vector_access(target, &vtype, &vl, &debug_vl, &debug_vsew) != ERROR_OK)
 		return ERROR_FAIL;
 
 	unsigned vnum = regno - GDB_REGNO_V0;
-	unsigned xlen = riscv_xlen(target);
-
-	struct riscv_program program;
-	riscv_program_init(&program, target);
-	riscv_program_insert(&program, vmv_x_s(S0, vnum));
-	riscv_program_insert(&program, vslide1down_vx(vnum, vnum, S0, true));
 
 	int result = ERROR_OK;
 	for (unsigned i = 0; i < debug_vl; i++) {
+		/* Can't reuse the same program because riscv_program_exec() adds
+		 * ebreak to the end every time. */
+		struct riscv_program program;
+		riscv_program_init(&program, target);
+		riscv_program_insert(&program, vmv_x_s(S0, vnum));
+		riscv_program_insert(&program, vslide1down_vx(vnum, vnum, S0, true));
+
 		/* Executing the program might result in an exception if there is some
 		 * issue with the vector implementation/instructions we're using. If that
 		 * happens, attempt to restore as usual. We may have clobbered the
@@ -1997,8 +2012,10 @@ static int riscv013_get_register_buf(struct target *target,
 			uint64_t v;
 			if (register_read_direct(target, &v, GDB_REGNO_S0) != ERROR_OK)
 				return ERROR_FAIL;
-			buf_set_u64(value, xlen * i, xlen, v);
+			buf_set_u64(value, debug_vsew * i, debug_vsew, v);
 		} else {
+			LOG_TARGET_ERROR(target, "Failed to execute vmv/vslide1down while reading %s",
+					 gdb_regno_name(regno));
 			break;
 		}
 	}
@@ -2028,12 +2045,11 @@ static int riscv013_set_register_buf(struct target *target,
 		return ERROR_FAIL;
 
 	uint64_t vtype, vl;
-	unsigned debug_vl;
-	if (prep_for_vector_access(target, &vtype, &vl, &debug_vl) != ERROR_OK)
+	unsigned int debug_vl, debug_vsew;
+	if (prep_for_vector_access(target, &vtype, &vl, &debug_vl, &debug_vsew) != ERROR_OK)
 		return ERROR_FAIL;
 
 	unsigned vnum = regno - GDB_REGNO_V0;
-	unsigned xlen = riscv_xlen(target);
 
 	struct riscv_program program;
 	riscv_program_init(&program, target);
@@ -2041,7 +2057,7 @@ static int riscv013_set_register_buf(struct target *target,
 	int result = ERROR_OK;
 	for (unsigned i = 0; i < debug_vl; i++) {
 		if (register_write_direct(target, GDB_REGNO_S0,
-					buf_get_u64(value, xlen * i, xlen)) != ERROR_OK)
+					buf_get_u64(value, debug_vsew * i, debug_vsew)) != ERROR_OK)
 			return ERROR_FAIL;
 		result = riscv_program_exec(&program, target);
 		if (result != ERROR_OK)
diff --git a/src/target/riscv/riscv.c b/src/target/riscv/riscv.c
index d198010..058a810 100644
--- a/src/target/riscv/riscv.c
+++ b/src/target/riscv/riscv.c
@@ -3926,6 +3926,8 @@ void riscv_info_init(struct target *target, riscv_info_t *r)
 
 	INIT_LIST_HEAD(&r->expose_csr);
 	INIT_LIST_HEAD(&r->expose_custom);
+
+	r->vsew64_supported = YNM_MAYBE;
 }
 
 static int riscv_resume_go_all_harts(struct target *target)
diff --git a/src/target/riscv/riscv.h b/src/target/riscv/riscv.h
index 87ec13f..6a1d7a4 100644
--- a/src/target/riscv/riscv.h
+++ b/src/target/riscv/riscv.h
@@ -41,6 +41,12 @@ typedef uint64_t riscv_reg_t;
 typedef uint32_t riscv_insn_t;
 typedef uint64_t riscv_addr_t;
 
+typedef enum {
+	YNM_MAYBE,
+	YNM_YES,
+	YNM_NO
+} yes_no_maybe_t;
+
 enum riscv_mem_access_method {
 	RISCV_MEM_ACCESS_UNSPECIFIED,
 	RISCV_MEM_ACCESS_PROGBUF,
@@ -253,6 +259,8 @@ typedef struct {
 
 	/* Track when we were last asked to do something substantial. */
 	int64_t last_activity;
+
+	yes_no_maybe_t vsew64_supported;
 } riscv_info_t;
 
 COMMAND_HELPER(riscv_print_info_line, const char *section, const char *key,