Fix max_decode_depth to control decoder hierarchy and port generation (#18)

* Initial plan

* Fix max_decode_depth to properly control decoder hierarchy and port generation

Co-authored-by: arnavsacheti <36746504+arnavsacheti@users.noreply.github.com>

* Fix test that relied on old depth behavior

Co-authored-by: arnavsacheti <36746504+arnavsacheti@users.noreply.github.com>

* Update documentation for max_decode_depth parameter

Co-authored-by: arnavsacheti <36746504+arnavsacheti@users.noreply.github.com>

* fix format

* Add variable_depth RDL file and smoke tests for max_decode_depth parameter

Co-authored-by: arnavsacheti <36746504+arnavsacheti@users.noreply.github.com>

* Add variable depth tests for APB3 and AXI4-Lite CPUIFs

Co-authored-by: arnavsacheti <36746504+arnavsacheti@users.noreply.github.com>

* fix

* fix

* bump

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: arnavsacheti <36746504+arnavsacheti@users.noreply.github.com>

.github/workflows/release.yml

@@ -40,7 +40,6 @@ jobs:
       
       - name: Publish to PyPI
         if: startsWith(github.ref, 'refs/tags/')
-        env:
-          TWINE_USERNAME: __token__
-          TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }}
-        run: uvx twine upload dist/*
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          skip-existing: true

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "peakrdl-busdecoder"
-version = "0.2.0"
+version = "0.4.0"
 requires-python = ">=3.10"
 dependencies = ["jinja2>=3.1.6", "systemrdl-compiler~=1.30.1"]
 

src/peakrdl_busdecoder/__peakrdl__.py

@@ -116,7 +116,9 @@ class Exporter(ExporterSubcommandPlugin):
             type=int,
             default=1,
             help="""Maximum depth for address decoder to descend into nested
-            addressable components. Default is 1.
+            addressable components. Value of 0 decodes all levels (infinite depth).
+            Value of 1 decodes only top-level children. Value of 2 decodes top-level
+            and one level deeper, etc. Default is 1.
             """,
         )
 

src/peakrdl_busdecoder/cpuif/apb3/apb3_cpuif.py

@@ -52,8 +52,16 @@ class APB3Cpuif(BaseCpuif):
             fanin["cpuif_rd_ack"] = "'0"
             fanin["cpuif_rd_err"] = "'0"
         else:
-            fanin["cpuif_rd_ack"] = self.signal("PREADY", node, "i")
-            fanin["cpuif_rd_err"] = self.signal("PSLVERR", node, "i")
+            # Use intermediate signals for interface arrays to avoid
+            # non-constant indexing of interface arrays in procedural blocks
+            if self.is_interface and node.is_array and node.array_dimensions:
+                # Generate array index string [i0][i1]... for the intermediate signal
+                array_idx = "".join(f"[i{i}]" for i in range(len(node.array_dimensions)))
+                fanin["cpuif_rd_ack"] = f"{node.inst_name}_fanin_ready{array_idx}"
+                fanin["cpuif_rd_err"] = f"{node.inst_name}_fanin_err{array_idx}"
+            else:
+                fanin["cpuif_rd_ack"] = self.signal("PREADY", node, "i")
+                fanin["cpuif_rd_err"] = self.signal("PSLVERR", node, "i")
 
         return "\n".join(map(lambda kv: f"{kv[0]} = {kv[1]};", fanin.items()))
 
@@ -62,6 +70,23 @@ class APB3Cpuif(BaseCpuif):
         if node is None:
             fanin["cpuif_rd_data"] = "'0"
         else:
-            fanin["cpuif_rd_data"] = self.signal("PRDATA", node, "i")
+            # Use intermediate signals for interface arrays to avoid
+            # non-constant indexing of interface arrays in procedural blocks
+            if self.is_interface and node.is_array and node.array_dimensions:
+                # Generate array index string [i0][i1]... for the intermediate signal
+                array_idx = "".join(f"[i{i}]" for i in range(len(node.array_dimensions)))
+                fanin["cpuif_rd_data"] = f"{node.inst_name}_fanin_data{array_idx}"
+            else:
+                fanin["cpuif_rd_data"] = self.signal("PRDATA", node, "i")
 
         return "\n".join(map(lambda kv: f"{kv[0]} = {kv[1]};", fanin.items()))
+
+    def fanin_intermediate_assignments(
+        self, node: AddressableNode, inst_name: str, array_idx: str, master_prefix: str, indexed_path: str
+    ) -> list[str]:
+        """Generate intermediate signal assignments for APB3 interface arrays."""
+        return [
+            f"assign {inst_name}_fanin_ready{array_idx} = {master_prefix}{indexed_path}.PREADY;",
+            f"assign {inst_name}_fanin_err{array_idx} = {master_prefix}{indexed_path}.PSLVERR;",
+            f"assign {inst_name}_fanin_data{array_idx} = {master_prefix}{indexed_path}.PRDATA;",
+        ]

src/peakrdl_busdecoder/cpuif/apb3/apb3_tmpl.sv

@@ -26,6 +26,13 @@ assign {{cpuif.signal("PSLVERR")}} = cpuif_rd_err | cpuif_rd_sel.cpuif_err | cpu
 // Fanout CPU Bus interface signals
 //--------------------------------------------------------------------------
 {{fanout|walk(cpuif=cpuif)}}
+{%- if cpuif.is_interface %}
+
+//--------------------------------------------------------------------------
+// Intermediate signals for interface array fanin
+//--------------------------------------------------------------------------
+{{fanin_intermediate|walk(cpuif=cpuif)}}
+{%- endif %}
 
 //--------------------------------------------------------------------------
 // Fanin CPU Bus interface signals

src/peakrdl_busdecoder/cpuif/apb4/apb4_cpuif.py

@@ -55,8 +55,16 @@ class APB4Cpuif(BaseCpuif):
             fanin["cpuif_rd_ack"] = "'0"
             fanin["cpuif_rd_err"] = "'0"
         else:
-            fanin["cpuif_rd_ack"] = self.signal("PREADY", node, "i")
-            fanin["cpuif_rd_err"] = self.signal("PSLVERR", node, "i")
+            # Use intermediate signals for interface arrays to avoid
+            # non-constant indexing of interface arrays in procedural blocks
+            if self.is_interface and node.is_array and node.array_dimensions:
+                # Generate array index string [i0][i1]... for the intermediate signal
+                array_idx = "".join(f"[i{i}]" for i in range(len(node.array_dimensions)))
+                fanin["cpuif_rd_ack"] = f"{node.inst_name}_fanin_ready{array_idx}"
+                fanin["cpuif_rd_err"] = f"{node.inst_name}_fanin_err{array_idx}"
+            else:
+                fanin["cpuif_rd_ack"] = self.signal("PREADY", node, "i")
+                fanin["cpuif_rd_err"] = self.signal("PSLVERR", node, "i")
 
         return "\n".join(map(lambda kv: f"{kv[0]} = {kv[1]};", fanin.items()))
 
@@ -65,6 +73,23 @@ class APB4Cpuif(BaseCpuif):
         if node is None:
             fanin["cpuif_rd_data"] = "'0"
         else:
-            fanin["cpuif_rd_data"] = self.signal("PRDATA", node, "i")
+            # Use intermediate signals for interface arrays to avoid
+            # non-constant indexing of interface arrays in procedural blocks
+            if self.is_interface and node.is_array and node.array_dimensions:
+                # Generate array index string [i0][i1]... for the intermediate signal
+                array_idx = "".join(f"[i{i}]" for i in range(len(node.array_dimensions)))
+                fanin["cpuif_rd_data"] = f"{node.inst_name}_fanin_data{array_idx}"
+            else:
+                fanin["cpuif_rd_data"] = self.signal("PRDATA", node, "i")
 
         return "\n".join(map(lambda kv: f"{kv[0]} = {kv[1]};", fanin.items()))
+
+    def fanin_intermediate_assignments(
+        self, node: AddressableNode, inst_name: str, array_idx: str, master_prefix: str, indexed_path: str
+    ) -> list[str]:
+        """Generate intermediate signal assignments for APB4 interface arrays."""
+        return [
+            f"assign {inst_name}_fanin_ready{array_idx} = {master_prefix}{indexed_path}.PREADY;",
+            f"assign {inst_name}_fanin_err{array_idx} = {master_prefix}{indexed_path}.PSLVERR;",
+            f"assign {inst_name}_fanin_data{array_idx} = {master_prefix}{indexed_path}.PRDATA;",
+        ]

src/peakrdl_busdecoder/cpuif/apb4/apb4_tmpl.sv

@@ -29,6 +29,13 @@ assign {{cpuif.signal("PSLVERR")}} = cpuif_rd_err | cpuif_rd_sel.cpuif_err | cpu
 // Fanout CPU Bus interface signals
 //--------------------------------------------------------------------------
 {{fanout|walk(cpuif=cpuif)}}
+{%- if cpuif.is_interface %}
+
+//--------------------------------------------------------------------------
+// Intermediate signals for interface array fanin
+//--------------------------------------------------------------------------
+{{fanin_intermediate|walk(cpuif=cpuif)}}
+{%- endif %}
 
 //--------------------------------------------------------------------------
 // Fanin CPU Bus interface signals

src/peakrdl_busdecoder/cpuif/axi4lite/axi4_lite_cpuif.py

@@ -4,14 +4,14 @@ from systemrdl.node import AddressableNode
 
 from ...utils import get_indexed_path
 from ..base_cpuif import BaseCpuif
-from .axi4lite_interface import AXI4LiteSVInterface
+from .axi4_lite_interface import AXI4LiteSVInterface
 
 if TYPE_CHECKING:
     from ...exporter import BusDecoderExporter
 
 
 class AXI4LiteCpuif(BaseCpuif):
-    template_path = "axi4lite_tmpl.sv"
+    template_path = "axi4_lite_tmpl.sv"
 
     def __init__(self, exp: "BusDecoderExporter") -> None:
         super().__init__(exp)
@@ -68,9 +68,17 @@ class AXI4LiteCpuif(BaseCpuif):
             fanin["cpuif_rd_ack"] = "'0"
             fanin["cpuif_rd_err"] = "'0"
         else:
-            # Read side: ack comes from RVALID; err if RRESP[1] is set (SLVERR/DECERR)
-            fanin["cpuif_rd_ack"] = self.signal("RVALID", node, "i")
-            fanin["cpuif_rd_err"] = f"{self.signal('RRESP', node, 'i')}[1]"
+            # Use intermediate signals for interface arrays to avoid
+            # non-constant indexing of interface arrays in procedural blocks
+            if self.is_interface and node.is_array and node.array_dimensions:
+                # Generate array index string [i0][i1]... for the intermediate signal
+                array_idx = "".join(f"[i{i}]" for i in range(len(node.array_dimensions)))
+                fanin["cpuif_rd_ack"] = f"{node.inst_name}_fanin_ready{array_idx}"
+                fanin["cpuif_rd_err"] = f"{node.inst_name}_fanin_err{array_idx}"
+            else:
+                # Read side: ack comes from RVALID; err if RRESP[1] is set (SLVERR/DECERR)
+                fanin["cpuif_rd_ack"] = self.signal("RVALID", node, "i")
+                fanin["cpuif_rd_err"] = f"{self.signal('RRESP', node, 'i')}[1]"
 
         return "\n".join(f"{lhs} = {rhs};" for lhs, rhs in fanin.items())
 
@@ -79,6 +87,23 @@ class AXI4LiteCpuif(BaseCpuif):
         if node is None:
             fanin["cpuif_rd_data"] = "'0"
         else:
-            fanin["cpuif_rd_data"] = self.signal("RDATA", node, "i")
+            # Use intermediate signals for interface arrays to avoid
+            # non-constant indexing of interface arrays in procedural blocks
+            if self.is_interface and node.is_array and node.array_dimensions:
+                # Generate array index string [i0][i1]... for the intermediate signal
+                array_idx = "".join(f"[i{i}]" for i in range(len(node.array_dimensions)))
+                fanin["cpuif_rd_data"] = f"{node.inst_name}_fanin_data{array_idx}"
+            else:
+                fanin["cpuif_rd_data"] = self.signal("RDATA", node, "i")
 
         return "\n".join(f"{lhs} = {rhs};" for lhs, rhs in fanin.items())
+
+    def fanin_intermediate_assignments(
+        self, node: AddressableNode, inst_name: str, array_idx: str, master_prefix: str, indexed_path: str
+    ) -> list[str]:
+        """Generate intermediate signal assignments for AXI4-Lite interface arrays."""
+        return [
+            f"assign {inst_name}_fanin_ready{array_idx} = {master_prefix}{indexed_path}.RVALID;",
+            f"assign {inst_name}_fanin_err{array_idx} = {master_prefix}{indexed_path}.RRESP[1];",
+            f"assign {inst_name}_fanin_data{array_idx} = {master_prefix}{indexed_path}.RDATA;",
+        ]

src/peakrdl_busdecoder/cpuif/axi4lite/axi4_lite_cpuif_flat.py

@@ -4,7 +4,7 @@ from systemrdl.node import AddressableNode
 
 from ...utils import get_indexed_path
 from ..base_cpuif import BaseCpuif
-from .axi4lite_interface import AXI4LiteFlatInterface
+from .axi4_lite_interface import AXI4LiteFlatInterface
 
 if TYPE_CHECKING:
     from ...exporter import BusDecoderExporter
@@ -13,7 +13,7 @@ if TYPE_CHECKING:
 class AXI4LiteCpuifFlat(BaseCpuif):
     """Verilator-friendly variant that flattens the AXI4-Lite interface ports."""
 
-    template_path = "axi4lite_tmpl.sv"
+    template_path = "axi4_lite_tmpl.sv"
 
     def __init__(self, exp: "BusDecoderExporter") -> None:
         super().__init__(exp)

src/peakrdl_busdecoder/cpuif/axi4lite/axi4_lite_tmpl.sv

@@ -53,6 +53,13 @@ assign {{cpuif.signal("BRESP")}}  = (cpuif_wr_err | cpuif_wr_sel.cpuif_err | cpu
 // Fanout CPU Bus interface signals
 //--------------------------------------------------------------------------
 {{fanout|walk(cpuif=cpuif)}}
+{%- if cpuif.is_interface %}
+
+//--------------------------------------------------------------------------
+// Intermediate signals for interface array fanin
+//--------------------------------------------------------------------------
+{{fanin_intermediate|walk(cpuif=cpuif)}}
+{%- endif %}
 
 //--------------------------------------------------------------------------
 // Fanin CPU Bus interface signals

src/peakrdl_busdecoder/cpuif/base_cpuif.py

@@ -7,6 +7,7 @@ from systemrdl.node import AddressableNode
 
 from ..utils import clog2, get_indexed_path, is_pow2, roundup_pow2
 from .fanin_gen import FaninGenerator
+from .fanin_intermediate_gen import FaninIntermediateGenerator
 from .fanout_gen import FanoutGenerator
 
 if TYPE_CHECKING:
@@ -24,11 +25,7 @@ class BaseCpuif:
 
     @property
     def addressable_children(self) -> list[AddressableNode]:
-        return [
-            child
-            for child in self.exp.ds.top_node.children(unroll=self.unroll)
-            if isinstance(child, AddressableNode)
-        ]
+        return self.exp.ds.get_addressable_children_at_depth(unroll=self.unroll)
 
     @property
     def addr_width(self) -> int:
@@ -97,6 +94,7 @@ class BaseCpuif:
             "ds": self.exp.ds,
             "fanout": FanoutGenerator,
             "fanin": FaninGenerator,
+            "fanin_intermediate": FaninIntermediateGenerator,
         }
 
         template = jj_env.get_template(self.template_path)
@@ -116,3 +114,24 @@ class BaseCpuif:
 
     def readback(self, node: AddressableNode | None = None) -> str:
         raise NotImplementedError
+
+    def fanin_intermediate_assignments(
+        self, node: AddressableNode, inst_name: str, array_idx: str, master_prefix: str, indexed_path: str
+    ) -> list[str]:
+        """Generate intermediate signal assignments for interface array fanin.
+
+        This method should be implemented by cpuif classes that use interfaces.
+        It returns a list of assignment strings that copy signals from interface
+        arrays to intermediate unpacked arrays using constant (genvar) indexing.
+
+        Args:
+            node: The addressable node
+            inst_name: Instance name for the intermediate signals
+            array_idx: Array index string (e.g., "[gi0][gi1]")
+            master_prefix: Master interface prefix
+            indexed_path: Indexed path to the interface element
+
+        Returns:
+            List of assignment strings
+        """
+        return []  # Default: no intermediate assignments needed

src/peakrdl_busdecoder/cpuif/fanin_gen.py

@@ -27,6 +27,17 @@ class FaninGenerator(BusDecoderListener):
     def enter_AddressableComponent(self, node: AddressableNode) -> WalkerAction | None:
         action = super().enter_AddressableComponent(node)
 
+        should_generate = action == WalkerAction.SkipDescendants
+        if not should_generate and self._ds.max_decode_depth == 0:
+            for child in node.children():
+                if isinstance(child, AddressableNode):
+                    break
+            else:
+                should_generate = True
+
+        if not should_generate:
+            return action
+
         if node.array_dimensions:
             for i, dim in enumerate(node.array_dimensions):
                 fb = ForLoopBody(

src/peakrdl_busdecoder/cpuif/fanin_intermediate_gen.py

@@ -0,0 +1,142 @@
+"""Generator for intermediate signals needed for interface array fanin.
+
+When using SystemVerilog interface arrays, we cannot use variable indices
+in procedural blocks (like always_comb). This generator creates intermediate
+signals that copy from interface arrays using generate loops, which can then
+be safely accessed with variable indices in the fanin logic.
+"""
+
+from collections import deque
+from typing import TYPE_CHECKING
+
+from systemrdl.node import AddressableNode
+from systemrdl.walker import WalkerAction
+
+from ..body import Body, ForLoopBody
+from ..design_state import DesignState
+from ..listener import BusDecoderListener
+from ..utils import get_indexed_path
+
+if TYPE_CHECKING:
+    from .base_cpuif import BaseCpuif
+
+
+class FaninIntermediateGenerator(BusDecoderListener):
+    """Generates intermediate signals for interface array fanin."""
+
+    def __init__(self, ds: DesignState, cpuif: "BaseCpuif") -> None:
+        super().__init__(ds)
+        self._cpuif = cpuif
+        self._declarations: list[str] = []
+        self._stack: deque[Body] = deque()
+        self._stack.append(Body())
+
+    def enter_AddressableComponent(self, node: AddressableNode) -> WalkerAction | None:
+        action = super().enter_AddressableComponent(node)
+
+        # Only generate intermediates for interface arrays
+        # Check if cpuif has is_interface attribute (some implementations don't)
+        is_interface = getattr(self._cpuif, "is_interface", False)
+        if not is_interface or not node.array_dimensions:
+            return action
+
+        # Generate intermediate signal declarations
+        self._generate_intermediate_declarations(node)
+
+        # Generate assignment logic using generate loops
+        if node.array_dimensions:
+            for i, dim in enumerate(node.array_dimensions):
+                fb = ForLoopBody(
+                    "genvar",
+                    f"gi{i}",
+                    dim,
+                )
+                self._stack.append(fb)
+
+        # Generate assignments from interface array to intermediates
+        self._stack[-1] += self._generate_intermediate_assignments(node)
+
+        return action
+
+    def exit_AddressableComponent(self, node: AddressableNode) -> None:
+        is_interface = getattr(self._cpuif, "is_interface", False)
+        if is_interface and node.array_dimensions:
+            for _ in node.array_dimensions:
+                b = self._stack.pop()
+                if not b:
+                    continue
+                self._stack[-1] += b
+
+        super().exit_AddressableComponent(node)
+
+    def _generate_intermediate_declarations(self, node: AddressableNode) -> None:
+        """Generate intermediate signal declarations for a node."""
+        inst_name = node.inst_name
+
+        # Array dimensions should be checked before calling this function
+        if not node.array_dimensions:
+            return
+
+        # Calculate total array size
+        array_size = 1
+        for dim in node.array_dimensions:
+            array_size *= dim
+
+        # Create array dimension string
+        array_str = "".join(f"[{dim}]" for dim in node.array_dimensions)
+
+        # Generate declarations for each fanin signal
+        # For APB3/4: PREADY, PSLVERR, PRDATA
+        # These are the signals read in fanin
+        self._declarations.append(f"logic {inst_name}_fanin_ready{array_str};")
+        self._declarations.append(f"logic {inst_name}_fanin_err{array_str};")
+        self._declarations.append(
+            f"logic [{self._cpuif.data_width - 1}:0] {inst_name}_fanin_data{array_str};"
+        )
+
+    def _generate_intermediate_assignments(self, node: AddressableNode) -> str:
+        """Generate assignments from interface array to intermediate signals."""
+        inst_name = node.inst_name
+        indexed_path = get_indexed_path(node.parent, node, "gi", skip_kw_filter=True)
+
+        # Get master prefix - use getattr to avoid type errors
+        interface = getattr(self._cpuif, "_interface", None)
+        if interface is None:
+            return ""
+        master_prefix = interface.get_master_prefix()
+
+        # Array dimensions should be checked before calling this function
+        if not node.array_dimensions:
+            return ""
+
+        # Create indexed signal names for left-hand side
+        array_idx = "".join(f"[gi{i}]" for i in range(len(node.array_dimensions)))
+
+        # Delegate to cpuif to get the appropriate assignments for this interface type
+        assignments = self._cpuif.fanin_intermediate_assignments(
+            node, inst_name, array_idx, master_prefix, indexed_path
+        )
+
+        return "\n".join(assignments)
+
+    def get_declarations(self) -> str:
+        """Get all intermediate signal declarations."""
+        if not self._declarations:
+            return ""
+        return "\n".join(self._declarations)
+
+    def __str__(self) -> str:
+        """Get all intermediate signal declarations and assignments."""
+        if not self._declarations:
+            return ""
+
+        # Output declarations first
+        output = "\n".join(self._declarations)
+        output += "\n\n"
+
+        # Then output assignments
+        body_str = "\n".join(map(str, self._stack))
+        if body_str and body_str.strip():
+            output += body_str
+
+        return output

src/peakrdl_busdecoder/cpuif/fanout_gen.py

@@ -23,6 +23,17 @@ class FanoutGenerator(BusDecoderListener):
     def enter_AddressableComponent(self, node: AddressableNode) -> WalkerAction | None:
         action = super().enter_AddressableComponent(node)
 
+        should_generate = action == WalkerAction.SkipDescendants
+        if not should_generate and self._ds.max_decode_depth == 0:
+            for child in node.children():
+                if isinstance(child, AddressableNode):
+                    break
+            else:
+                should_generate = True
+
+        if not should_generate:
+            return action
+
         if node.array_dimensions:
             for i, dim in enumerate(node.array_dimensions):
                 fb = ForLoopBody(

src/peakrdl_busdecoder/decode_logic_gen.py

@@ -85,6 +85,20 @@ class DecodeLogicGenerator(BusDecoderListener):
     def enter_AddressableComponent(self, node: AddressableNode) -> WalkerAction | None:
         action = super().enter_AddressableComponent(node)
 
+        should_decode = action == WalkerAction.SkipDescendants
+
+        if not should_decode and self._ds.max_decode_depth == 0:
+            # When decoding all levels, treat leaf registers as decode boundary
+            for child in node.children():
+                if isinstance(child, AddressableNode):
+                    break
+            else:
+                should_decode = True
+
+        # Only generate select logic if we're at the decode boundary
+        if not should_decode:
+            return action
+
         conditions: list[str] = []
         conditions.extend(self.cpuif_addr_predicate(node))
         conditions.extend(self.cpuif_prot_predicate(node))
@@ -146,6 +160,8 @@ class DecodeLogicGenerator(BusDecoderListener):
     def __str__(self) -> str:
         body = self._decode_stack[-1]
         if isinstance(body, IfBody):
+            if len(body) == 0:
+                return f"{self._flavor.cpuif_select}.cpuif_err = 1'b1;"
             with body.cm(...) as b:
                 b += f"{self._flavor.cpuif_select}.cpuif_err = 1'b1;"
 

src/peakrdl_busdecoder/design_state.py

@@ -1,6 +1,6 @@
 from typing import TypedDict
 
-from systemrdl.node import AddrmapNode
+from systemrdl.node import AddressableNode, AddrmapNode
 from systemrdl.rdltypes.user_enum import UserEnum
 
 from .design_scanner import DesignScanner
@@ -72,3 +72,56 @@ class DesignState:
         if user_addr_width < self.addr_width:
             msg.fatal(f"User-specified address width shall be greater than or equal to {self.addr_width}.")
         self.addr_width = user_addr_width
+
+    def get_addressable_children_at_depth(self, unroll: bool = False) -> list[AddressableNode]:
+        """
+        Get addressable children at the decode boundary based on max_decode_depth.
+
+        max_decode_depth semantics:
+        - 0: decode all levels (return leaf registers)
+        - 1: decode only top level (return children at depth 1)
+        - 2: decode top + 1 level (return children at depth 2)
+        - N: decode down to depth N (return children at depth N)
+
+        Args:
+            unroll: Whether to unroll arrayed nodes
+
+        Returns:
+            List of addressable nodes at the decode boundary
+        """
+        from systemrdl.node import RegNode
+
+        def collect_nodes(node: AddressableNode, current_depth: int) -> list[AddressableNode]:
+            """Recursively collect nodes at the decode boundary."""
+            result: list[AddressableNode] = []
+
+            # For depth 0, collect all leaf registers
+            if self.max_decode_depth == 0:
+                # If this is a register, it's a leaf
+                if isinstance(node, RegNode):
+                    result.append(node)
+                else:
+                    # Recurse into children
+                    for child in node.children(unroll=unroll):
+                        if isinstance(child, AddressableNode):
+                            result.extend(collect_nodes(child, current_depth + 1))
+            else:
+                # For depth N, collect children at depth N
+                if current_depth == self.max_decode_depth:
+                    # We're at the decode boundary - return this node
+                    result.append(node)
+                elif current_depth < self.max_decode_depth:
+                    # We haven't reached the boundary yet - recurse
+                    for child in node.children(unroll=unroll):
+                        if isinstance(child, AddressableNode):
+                            result.extend(collect_nodes(child, current_depth + 1))
+
+            return result
+
+        # Start collecting from top node's children
+        nodes: list[AddressableNode] = []
+        for child in self.top_node.children(unroll=unroll):
+            if isinstance(child, AddressableNode):
+                nodes.extend(collect_nodes(child, 1))
+
+        return nodes

src/peakrdl_busdecoder/exporter.py

@@ -89,7 +89,9 @@ class BusDecoderExporter:
             interface. By default, arrayed nodes are kept as arrays.
         max_decode_depth: int
             Maximum depth for address decoder to descend into nested addressable
-            components. By default, the decoder descends 1 level deep.
+            components. A value of 0 decodes all levels (infinite depth). A value
+            of 1 decodes only top-level children. A value of 2 decodes top-level
+            and one level deeper, etc. By default, the decoder descends 1 level deep.
         """
         # If it is the root node, skip to top addrmap
         if isinstance(node, RootNode):

src/peakrdl_busdecoder/listener.py

@@ -15,7 +15,12 @@ class BusDecoderListener(RDLListener):
     def should_skip_node(self, node: AddressableNode) -> bool:
         """Check if this node should be skipped (not decoded)."""
         # Check if current depth exceeds max depth
-        if self._depth > self._ds.max_decode_depth:
+        # max_decode_depth semantics:
+        # - 0 means decode all levels (infinite)
+        # - 1 means decode only top level (depth 0)
+        # - 2 means decode top + 1 level (depth 0 and 1)
+        # - N means decode down to depth N-1
+        if self._ds.max_decode_depth > 0 and self._depth >= self._ds.max_decode_depth:
             return True
 
         # Check if this node only contains external addressable children

src/peakrdl_busdecoder/module_tmpl.sv

@@ -51,7 +51,7 @@ module {{ds.module_name}}
     //--------------------------------------------------------------------------
     always_comb begin
         // Default all write select signals to 0
-        cpuif_wr_sel = '0;
+        cpuif_wr_sel = '{default: '0};
 
         if (cpuif_req && cpuif_wr_en) begin
             // A write request is pending
@@ -66,7 +66,7 @@ module {{ds.module_name}}
     //--------------------------------------------------------------------------
     always_comb begin
         // Default all read select signals to 0
-        cpuif_rd_sel = '0;
+        cpuif_rd_sel = '{default: '0};
 
         if (cpuif_req && cpuif_rd_en) begin
             // A read request is pending

src/peakrdl_busdecoder/struct_gen.py

@@ -3,7 +3,7 @@ from collections import deque
 from systemrdl.node import AddressableNode
 from systemrdl.walker import WalkerAction
 
-from .body import Body, StructBody
+from .body import StructBody
 from .design_state import DesignState
 from .identifier_filter import kw_filter as kwf
 from .listener import BusDecoderListener
@@ -16,42 +16,53 @@ class StructGenerator(BusDecoderListener):
     ) -> None:
         super().__init__(ds)
 
-        self._stack: deque[Body] = deque()
-        self._stack.append(StructBody("cpuif_sel_t", True, True))
+        self._stack: list[StructBody] = [StructBody("cpuif_sel_t", True, False)]
+        self._struct_defs: list[StructBody] = []
+        self._created_struct_stack: deque[bool] = deque()  # Track if we created a struct for each node
 
     def enter_AddressableComponent(self, node: AddressableNode) -> WalkerAction | None:
         action = super().enter_AddressableComponent(node)
 
-        self._skip = False
-        if action == WalkerAction.SkipDescendants:
-            self._skip = True
+        skip = action == WalkerAction.SkipDescendants
 
-        if node.children():
+        # Only create nested struct if we're not skipping and node has addressable children
+        has_addressable_children = any(isinstance(child, AddressableNode) for child in node.children())
+        if has_addressable_children and not skip:
             # Push new body onto stack
-            body = StructBody(f"cpuif_sel_{node.inst_name}_t", True, True)
+            body = StructBody(f"cpuif_sel_{node.inst_name}_t", True, False)
             self._stack.append(body)
+            self._created_struct_stack.append(True)
+        else:
+            self._created_struct_stack.append(False)
 
         return action
 
     def exit_AddressableComponent(self, node: AddressableNode) -> None:
         type = "logic"
 
-        if node.children():
+        # Pop the created_struct flag
+        created_struct = self._created_struct_stack.pop()
+
+        # Only pop struct body if we created one
+        if created_struct:
             body = self._stack.pop()
-            if body and isinstance(body, StructBody) and not self._skip:
-                self._stack.appendleft(body)
+            if body:
+                self._struct_defs.append(body)
                 type = body.name
 
         name = kwf(node.inst_name)
 
         if node.array_dimensions:
             for dim in node.array_dimensions:
-                name = f"[{dim - 1}:0]{name}"
+                name = f"{name}[{dim}]"
 
         self._stack[-1] += f"{type} {name};"
 
         super().exit_AddressableComponent(node)
 
     def __str__(self) -> str:
-        self._stack[-1] += "logic cpuif_err;"
-        return "\n".join(map(str, self._stack))
+        if "logic cpuif_err;" not in self._stack[-1].lines:
+            self._stack[-1] += "logic cpuif_err;"
+        bodies = [str(body) for body in self._struct_defs]
+        bodies.append(str(self._stack[-1]))
+        return "\n".join(bodies)

tests/cocotb/apb3/smoke/test_variable_depth.py

@@ -0,0 +1,211 @@
+"""APB3 smoke tests for variable depth design testing max_decode_depth parameter."""
+
+import cocotb
+from cocotb.triggers import Timer
+
+
+class _Apb3SlaveShim:
+    def __init__(self, dut):
+        prefix = "s_apb"
+        self.PSEL = getattr(dut, f"{prefix}_PSEL")
+        self.PENABLE = getattr(dut, f"{prefix}_PENABLE")
+        self.PWRITE = getattr(dut, f"{prefix}_PWRITE")
+        self.PADDR = getattr(dut, f"{prefix}_PADDR")
+        self.PWDATA = getattr(dut, f"{prefix}_PWDATA")
+        self.PRDATA = getattr(dut, f"{prefix}_PRDATA")
+        self.PREADY = getattr(dut, f"{prefix}_PREADY")
+        self.PSLVERR = getattr(dut, f"{prefix}_PSLVERR")
+
+
+class _Apb3MasterShim:
+    def __init__(self, dut, base: str):
+        self.PSEL = getattr(dut, f"{base}_PSEL")
+        self.PENABLE = getattr(dut, f"{base}_PENABLE")
+        self.PWRITE = getattr(dut, f"{base}_PWRITE")
+        self.PADDR = getattr(dut, f"{base}_PADDR")
+        self.PWDATA = getattr(dut, f"{base}_PWDATA")
+        self.PRDATA = getattr(dut, f"{base}_PRDATA")
+        self.PREADY = getattr(dut, f"{base}_PREADY")
+        self.PSLVERR = getattr(dut, f"{base}_PSLVERR")
+
+
+def _apb3_slave(dut):
+    return getattr(dut, "s_apb", None) or _Apb3SlaveShim(dut)
+
+
+def _apb3_master(dut, base: str):
+    return getattr(dut, base, None) or _Apb3MasterShim(dut, base)
+
+
+@cocotb.test()
+async def test_depth_1(dut):
+    """Test max_decode_depth=1 - should have interface for inner1 only."""
+    s_apb = _apb3_slave(dut)
+    
+    # At depth 1, we should have m_apb_inner1 but not deeper interfaces
+    inner1 = _apb3_master(dut, "m_apb_inner1")
+    
+    # Default slave side inputs
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    s_apb.PWRITE.value = 0
+    s_apb.PADDR.value = 0
+    s_apb.PWDATA.value = 0
+    
+    inner1.PRDATA.value = 0
+    inner1.PREADY.value = 0
+    inner1.PSLVERR.value = 0
+    
+    await Timer(1, units="ns")
+    
+    # Write to address 0x0 (should select inner1)
+    inner1.PREADY.value = 1
+    s_apb.PADDR.value = 0x0
+    s_apb.PWDATA.value = 0x12345678
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(inner1.PSEL.value) == 1, "inner1 must be selected"
+    assert int(inner1.PWRITE.value) == 1, "Write should propagate"
+    assert int(s_apb.PREADY.value) == 1, "Ready should mirror master"
+
+
+@cocotb.test()
+async def test_depth_2(dut):
+    """Test max_decode_depth=2 - should have interfaces for reg1 and inner2."""
+    s_apb = _apb3_slave(dut)
+    
+    # At depth 2, we should have m_apb_reg1 and m_apb_inner2
+    reg1 = _apb3_master(dut, "m_apb_reg1")
+    inner2 = _apb3_master(dut, "m_apb_inner2")
+    
+    # Default slave side inputs
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    s_apb.PWRITE.value = 0
+    s_apb.PADDR.value = 0
+    s_apb.PWDATA.value = 0
+    
+    reg1.PRDATA.value = 0
+    reg1.PREADY.value = 0
+    reg1.PSLVERR.value = 0
+    
+    inner2.PRDATA.value = 0
+    inner2.PREADY.value = 0
+    inner2.PSLVERR.value = 0
+    
+    await Timer(1, units="ns")
+    
+    # Write to address 0x0 (should select reg1)
+    reg1.PREADY.value = 1
+    s_apb.PADDR.value = 0x0
+    s_apb.PWDATA.value = 0xABCDEF01
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg1.PSEL.value) == 1, "reg1 must be selected for address 0x0"
+    assert int(inner2.PSEL.value) == 0, "inner2 should not be selected"
+    
+    # Reset
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    reg1.PREADY.value = 0
+    await Timer(1, units="ns")
+    
+    # Write to address 0x10 (should select inner2)
+    inner2.PREADY.value = 1
+    s_apb.PADDR.value = 0x10
+    s_apb.PWDATA.value = 0x23456789
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(inner2.PSEL.value) == 1, "inner2 must be selected for address 0x10"
+    assert int(reg1.PSEL.value) == 0, "reg1 should not be selected"
+
+
+@cocotb.test()
+async def test_depth_0(dut):
+    """Test max_decode_depth=0 - should have interfaces for all leaf registers."""
+    s_apb = _apb3_slave(dut)
+    
+    # At depth 0, we should have all leaf registers: reg1, reg2, reg2b
+    reg1 = _apb3_master(dut, "m_apb_reg1")
+    reg2 = _apb3_master(dut, "m_apb_reg2")
+    reg2b = _apb3_master(dut, "m_apb_reg2b")
+    
+    # Default slave side inputs
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    s_apb.PWRITE.value = 0
+    s_apb.PADDR.value = 0
+    s_apb.PWDATA.value = 0
+    
+    for master in [reg1, reg2, reg2b]:
+        master.PRDATA.value = 0
+        master.PREADY.value = 0
+        master.PSLVERR.value = 0
+    
+    await Timer(1, units="ns")
+    
+    # Write to address 0x0 (should select reg1)
+    reg1.PREADY.value = 1
+    s_apb.PADDR.value = 0x0
+    s_apb.PWDATA.value = 0x11111111
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg1.PSEL.value) == 1, "reg1 must be selected for address 0x0"
+    assert int(reg2.PSEL.value) == 0, "reg2 should not be selected"
+    assert int(reg2b.PSEL.value) == 0, "reg2b should not be selected"
+    
+    # Reset
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    reg1.PREADY.value = 0
+    await Timer(1, units="ns")
+    
+    # Write to address 0x10 (should select reg2)
+    reg2.PREADY.value = 1
+    s_apb.PADDR.value = 0x10
+    s_apb.PWDATA.value = 0x22222222
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg2.PSEL.value) == 1, "reg2 must be selected for address 0x10"
+    assert int(reg1.PSEL.value) == 0, "reg1 should not be selected"
+    assert int(reg2b.PSEL.value) == 0, "reg2b should not be selected"
+    
+    # Reset
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    reg2.PREADY.value = 0
+    await Timer(1, units="ns")
+    
+    # Write to address 0x14 (should select reg2b)
+    reg2b.PREADY.value = 1
+    s_apb.PADDR.value = 0x14
+    s_apb.PWDATA.value = 0x33333333
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg2b.PSEL.value) == 1, "reg2b must be selected for address 0x14"
+    assert int(reg1.PSEL.value) == 0, "reg1 should not be selected"
+    assert int(reg2.PSEL.value) == 0, "reg2 should not be selected"

tests/cocotb/apb3/smoke/test_variable_depth_runner.py

@@ -0,0 +1,128 @@
+"""Pytest wrapper launching the APB3 cocotb smoke test for variable depth."""
+
+from pathlib import Path
+
+import pytest
+
+from peakrdl_busdecoder.cpuif.apb3.apb3_cpuif_flat import APB3CpuifFlat
+
+try:  # pragma: no cover - optional dependency shim
+    from cocotb.runner import get_runner
+except ImportError:  # pragma: no cover
+    from cocotb_tools.runner import get_runner
+
+from tests.cocotb_lib.utils import compile_rdl_and_export, get_verilog_sources
+
+
+@pytest.mark.simulation
+@pytest.mark.verilator
+def test_apb3_variable_depth_1(tmp_path: Path) -> None:
+    """Test APB3 design with max_decode_depth=1."""
+    repo_root = Path(__file__).resolve().parents[4]
+
+    module_path, package_path = compile_rdl_and_export(
+        str(repo_root / "tests" / "cocotb_lib" / "variable_depth.rdl"),
+        "variable_depth",
+        tmp_path,
+        cpuif_cls=APB3CpuifFlat,
+        max_decode_depth=1,
+    )
+
+    sources = get_verilog_sources(
+        module_path,
+        package_path,
+        [repo_root / "hdl-src" / "apb3_intf.sv"],
+    )
+
+    runner = get_runner("verilator")
+    build_dir = tmp_path / "sim_build"
+
+    runner.build(
+        sources=sources,
+        hdl_toplevel=module_path.stem,
+        build_dir=build_dir,
+    )
+
+    runner.test(
+        hdl_toplevel=module_path.stem,
+        test_module="tests.cocotb.apb3.smoke.test_variable_depth",
+        build_dir=build_dir,
+        log_file=str(tmp_path / "sim_depth1.log"),
+        testcase="test_depth_1",
+    )
+
+
+@pytest.mark.simulation
+@pytest.mark.verilator
+def test_apb3_variable_depth_2(tmp_path: Path) -> None:
+    """Test APB3 design with max_decode_depth=2."""
+    repo_root = Path(__file__).resolve().parents[4]
+
+    module_path, package_path = compile_rdl_and_export(
+        str(repo_root / "tests" / "cocotb_lib" / "variable_depth.rdl"),
+        "variable_depth",
+        tmp_path,
+        cpuif_cls=APB3CpuifFlat,
+        max_decode_depth=2,
+    )
+
+    sources = get_verilog_sources(
+        module_path,
+        package_path,
+        [repo_root / "hdl-src" / "apb3_intf.sv"],
+    )
+
+    runner = get_runner("verilator")
+    build_dir = tmp_path / "sim_build"
+
+    runner.build(
+        sources=sources,
+        hdl_toplevel=module_path.stem,
+        build_dir=build_dir,
+    )
+
+    runner.test(
+        hdl_toplevel=module_path.stem,
+        test_module="tests.cocotb.apb3.smoke.test_variable_depth",
+        build_dir=build_dir,
+        log_file=str(tmp_path / "sim_depth2.log"),
+        testcase="test_depth_2",
+    )
+
+
+@pytest.mark.simulation
+@pytest.mark.verilator
+def test_apb3_variable_depth_0(tmp_path: Path) -> None:
+    """Test APB3 design with max_decode_depth=0 (all levels)."""
+    repo_root = Path(__file__).resolve().parents[4]
+
+    module_path, package_path = compile_rdl_and_export(
+        str(repo_root / "tests" / "cocotb_lib" / "variable_depth.rdl"),
+        "variable_depth",
+        tmp_path,
+        cpuif_cls=APB3CpuifFlat,
+        max_decode_depth=0,
+    )
+
+    sources = get_verilog_sources(
+        module_path,
+        package_path,
+        [repo_root / "hdl-src" / "apb3_intf.sv"],
+    )
+
+    runner = get_runner("verilator")
+    build_dir = tmp_path / "sim_build"
+
+    runner.build(
+        sources=sources,
+        hdl_toplevel=module_path.stem,
+        build_dir=build_dir,
+    )
+
+    runner.test(
+        hdl_toplevel=module_path.stem,
+        test_module="tests.cocotb.apb3.smoke.test_variable_depth",
+        build_dir=build_dir,
+        log_file=str(tmp_path / "sim_depth0.log"),
+        testcase="test_depth_0",
+    )

tests/cocotb/apb4/smoke/test_variable_depth.py

@@ -0,0 +1,227 @@
+"""APB4 smoke tests for variable depth design testing max_decode_depth parameter."""
+
+import cocotb
+from cocotb.triggers import Timer
+
+
+class _Apb4SlaveShim:
+    def __init__(self, dut):
+        prefix = "s_apb"
+        self.PSEL = getattr(dut, f"{prefix}_PSEL")
+        self.PENABLE = getattr(dut, f"{prefix}_PENABLE")
+        self.PWRITE = getattr(dut, f"{prefix}_PWRITE")
+        self.PADDR = getattr(dut, f"{prefix}_PADDR")
+        self.PPROT = getattr(dut, f"{prefix}_PPROT")
+        self.PWDATA = getattr(dut, f"{prefix}_PWDATA")
+        self.PSTRB = getattr(dut, f"{prefix}_PSTRB")
+        self.PRDATA = getattr(dut, f"{prefix}_PRDATA")
+        self.PREADY = getattr(dut, f"{prefix}_PREADY")
+        self.PSLVERR = getattr(dut, f"{prefix}_PSLVERR")
+
+
+class _Apb4MasterShim:
+    def __init__(self, dut, base: str):
+        self.PSEL = getattr(dut, f"{base}_PSEL")
+        self.PENABLE = getattr(dut, f"{base}_PENABLE")
+        self.PWRITE = getattr(dut, f"{base}_PWRITE")
+        self.PADDR = getattr(dut, f"{base}_PADDR")
+        self.PPROT = getattr(dut, f"{base}_PPROT")
+        self.PWDATA = getattr(dut, f"{base}_PWDATA")
+        self.PSTRB = getattr(dut, f"{base}_PSTRB")
+        self.PRDATA = getattr(dut, f"{base}_PRDATA")
+        self.PREADY = getattr(dut, f"{base}_PREADY")
+        self.PSLVERR = getattr(dut, f"{base}_PSLVERR")
+
+
+def _apb4_slave(dut):
+    return getattr(dut, "s_apb", None) or _Apb4SlaveShim(dut)
+
+
+def _apb4_master(dut, base: str):
+    return getattr(dut, base, None) or _Apb4MasterShim(dut, base)
+
+
+@cocotb.test()
+async def test_depth_1(dut):
+    """Test max_decode_depth=1 - should have interface for inner1 only."""
+    s_apb = _apb4_slave(dut)
+    
+    # At depth 1, we should have m_apb_inner1 but not deeper interfaces
+    inner1 = _apb4_master(dut, "m_apb_inner1")
+    
+    # Default slave side inputs
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    s_apb.PWRITE.value = 0
+    s_apb.PADDR.value = 0
+    s_apb.PWDATA.value = 0
+    s_apb.PPROT.value = 0
+    s_apb.PSTRB.value = 0
+    
+    inner1.PRDATA.value = 0
+    inner1.PREADY.value = 0
+    inner1.PSLVERR.value = 0
+    
+    await Timer(1, units="ns")
+    
+    # Write to address 0x0 (should select inner1)
+    inner1.PREADY.value = 1
+    s_apb.PADDR.value = 0x0
+    s_apb.PWDATA.value = 0x12345678
+    s_apb.PSTRB.value = 0xF
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(inner1.PSEL.value) == 1, "inner1 must be selected"
+    assert int(inner1.PWRITE.value) == 1, "Write should propagate"
+    assert int(s_apb.PREADY.value) == 1, "Ready should mirror master"
+
+
+@cocotb.test()
+async def test_depth_2(dut):
+    """Test max_decode_depth=2 - should have interfaces for reg1 and inner2."""
+    s_apb = _apb4_slave(dut)
+    
+    # At depth 2, we should have m_apb_reg1 and m_apb_inner2
+    reg1 = _apb4_master(dut, "m_apb_reg1")
+    inner2 = _apb4_master(dut, "m_apb_inner2")
+    
+    # Default slave side inputs
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    s_apb.PWRITE.value = 0
+    s_apb.PADDR.value = 0
+    s_apb.PWDATA.value = 0
+    s_apb.PPROT.value = 0
+    s_apb.PSTRB.value = 0
+    
+    reg1.PRDATA.value = 0
+    reg1.PREADY.value = 0
+    reg1.PSLVERR.value = 0
+    
+    inner2.PRDATA.value = 0
+    inner2.PREADY.value = 0
+    inner2.PSLVERR.value = 0
+    
+    await Timer(1, units="ns")
+    
+    # Write to address 0x0 (should select reg1)
+    reg1.PREADY.value = 1
+    s_apb.PADDR.value = 0x0
+    s_apb.PWDATA.value = 0xABCDEF01
+    s_apb.PSTRB.value = 0xF
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg1.PSEL.value) == 1, "reg1 must be selected for address 0x0"
+    assert int(inner2.PSEL.value) == 0, "inner2 should not be selected"
+    
+    # Reset
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    reg1.PREADY.value = 0
+    await Timer(1, units="ns")
+    
+    # Write to address 0x10 (should select inner2)
+    inner2.PREADY.value = 1
+    s_apb.PADDR.value = 0x10
+    s_apb.PWDATA.value = 0x23456789
+    s_apb.PSTRB.value = 0xF
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(inner2.PSEL.value) == 1, "inner2 must be selected for address 0x10"
+    assert int(reg1.PSEL.value) == 0, "reg1 should not be selected"
+
+
+@cocotb.test()
+async def test_depth_0(dut):
+    """Test max_decode_depth=0 - should have interfaces for all leaf registers."""
+    s_apb = _apb4_slave(dut)
+    
+    # At depth 0, we should have all leaf registers: reg1, reg2, reg2b
+    reg1 = _apb4_master(dut, "m_apb_reg1")
+    reg2 = _apb4_master(dut, "m_apb_reg2")
+    reg2b = _apb4_master(dut, "m_apb_reg2b")
+    
+    # Default slave side inputs
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    s_apb.PWRITE.value = 0
+    s_apb.PADDR.value = 0
+    s_apb.PWDATA.value = 0
+    s_apb.PPROT.value = 0
+    s_apb.PSTRB.value = 0
+    
+    for master in [reg1, reg2, reg2b]:
+        master.PRDATA.value = 0
+        master.PREADY.value = 0
+        master.PSLVERR.value = 0
+    
+    await Timer(1, units="ns")
+    
+    # Write to address 0x0 (should select reg1)
+    reg1.PREADY.value = 1
+    s_apb.PADDR.value = 0x0
+    s_apb.PWDATA.value = 0x11111111
+    s_apb.PSTRB.value = 0xF
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg1.PSEL.value) == 1, "reg1 must be selected for address 0x0"
+    assert int(reg2.PSEL.value) == 0, "reg2 should not be selected"
+    assert int(reg2b.PSEL.value) == 0, "reg2b should not be selected"
+    
+    # Reset
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    reg1.PREADY.value = 0
+    await Timer(1, units="ns")
+    
+    # Write to address 0x10 (should select reg2)
+    reg2.PREADY.value = 1
+    s_apb.PADDR.value = 0x10
+    s_apb.PWDATA.value = 0x22222222
+    s_apb.PSTRB.value = 0xF
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg2.PSEL.value) == 1, "reg2 must be selected for address 0x10"
+    assert int(reg1.PSEL.value) == 0, "reg1 should not be selected"
+    assert int(reg2b.PSEL.value) == 0, "reg2b should not be selected"
+    
+    # Reset
+    s_apb.PSEL.value = 0
+    s_apb.PENABLE.value = 0
+    reg2.PREADY.value = 0
+    await Timer(1, units="ns")
+    
+    # Write to address 0x14 (should select reg2b)
+    reg2b.PREADY.value = 1
+    s_apb.PADDR.value = 0x14
+    s_apb.PWDATA.value = 0x33333333
+    s_apb.PSTRB.value = 0xF
+    s_apb.PWRITE.value = 1
+    s_apb.PSEL.value = 1
+    s_apb.PENABLE.value = 1
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg2b.PSEL.value) == 1, "reg2b must be selected for address 0x14"
+    assert int(reg1.PSEL.value) == 0, "reg1 should not be selected"
+    assert int(reg2.PSEL.value) == 0, "reg2 should not be selected"

tests/cocotb/apb4/smoke/test_variable_depth_runner.py

@@ -0,0 +1,131 @@
+"""Pytest wrapper launching the APB4 cocotb smoke test for variable depth."""
+
+from pathlib import Path
+
+import pytest
+
+from peakrdl_busdecoder.cpuif.apb4.apb4_cpuif_flat import APB4CpuifFlat
+
+try:  # pragma: no cover - optional dependency shim
+    from cocotb.runner import get_runner
+except ImportError:  # pragma: no cover
+    from cocotb_tools.runner import get_runner
+
+from tests.cocotb_lib.utils import compile_rdl_and_export, get_verilog_sources
+
+
+@pytest.mark.simulation
+@pytest.mark.verilator
+def test_apb4_variable_depth_1(tmp_path: Path) -> None:
+    """Test APB4 design with max_decode_depth=1."""
+    repo_root = Path(__file__).resolve().parents[4]
+
+    module_path, package_path = compile_rdl_and_export(
+        str(repo_root / "tests" / "cocotb_lib" / "variable_depth.rdl"),
+        "variable_depth",
+        tmp_path,
+        cpuif_cls=APB4CpuifFlat,
+        max_decode_depth=1,
+    )
+
+    sources = get_verilog_sources(
+        module_path,
+        package_path,
+        [repo_root / "hdl-src" / "apb4_intf.sv"],
+    )
+
+    runner = get_runner("verilator")
+    build_dir = tmp_path / "sim_build"
+
+    runner.build(
+        sources=sources,
+        hdl_toplevel=module_path.stem,
+        build_dir=build_dir,
+        log_file=str(tmp_path / "build_depth_1.log"),
+    )
+
+    runner.test(
+        hdl_toplevel=module_path.stem,
+        test_module="tests.cocotb.apb4.smoke.test_variable_depth",
+        build_dir=build_dir,
+        log_file=str(tmp_path / "sim_depth1.log"),
+        testcase="test_depth_1",
+    )
+
+
+@pytest.mark.simulation
+@pytest.mark.verilator
+def test_apb4_variable_depth_2(tmp_path: Path) -> None:
+    """Test APB4 design with max_decode_depth=2."""
+    repo_root = Path(__file__).resolve().parents[4]
+
+    module_path, package_path = compile_rdl_and_export(
+        str(repo_root / "tests" / "cocotb_lib" / "variable_depth.rdl"),
+        "variable_depth",
+        tmp_path,
+        cpuif_cls=APB4CpuifFlat,
+        max_decode_depth=2,
+    )
+
+    sources = get_verilog_sources(
+        module_path,
+        package_path,
+        [repo_root / "hdl-src" / "apb4_intf.sv"],
+    )
+
+    runner = get_runner("verilator")
+    build_dir = tmp_path / "sim_build"
+
+    runner.build(
+        sources=sources,
+        hdl_toplevel=module_path.stem,
+        build_dir=build_dir,
+        log_file=str(tmp_path / "build_depth_2.log"),
+    )
+
+    runner.test(
+        hdl_toplevel=module_path.stem,
+        test_module="tests.cocotb.apb4.smoke.test_variable_depth",
+        build_dir=build_dir,
+        log_file=str(tmp_path / "sim_depth_2.log"),
+        testcase="test_depth_2",
+    )
+
+
+@pytest.mark.simulation
+@pytest.mark.verilator
+def test_apb4_variable_depth_0(tmp_path: Path) -> None:
+    """Test APB4 design with max_decode_depth=0 (all levels)."""
+    repo_root = Path(__file__).resolve().parents[4]
+
+    module_path, package_path = compile_rdl_and_export(
+        str(repo_root / "tests" / "cocotb_lib" / "variable_depth.rdl"),
+        "variable_depth",
+        tmp_path,
+        cpuif_cls=APB4CpuifFlat,
+        max_decode_depth=0,
+    )
+
+    sources = get_verilog_sources(
+        module_path,
+        package_path,
+        [repo_root / "hdl-src" / "apb4_intf.sv"],
+    )
+
+    runner = get_runner("verilator")
+    build_dir = tmp_path / "sim_build"
+
+    runner.build(
+        sources=sources,
+        hdl_toplevel=module_path.stem,
+        build_dir=build_dir,
+        log_file=str(tmp_path / "build_depth_0.log"),
+    )
+
+    runner.test(
+        hdl_toplevel=module_path.stem,
+        test_module="tests.cocotb.apb4.smoke.test_variable_depth",
+        build_dir=build_dir,
+        log_file=str(tmp_path / "sim_depth_0.log"),
+        testcase="test_depth_0",
+    )

tests/cocotb/axi4lite/smoke/test_variable_depth.py

@@ -0,0 +1,271 @@
+"""AXI4-Lite smoke tests for variable depth design testing max_decode_depth parameter."""
+
+import cocotb
+from cocotb.triggers import Timer
+
+
+class _AxilSlaveShim:
+    def __init__(self, dut):
+        prefix = "s_axil"
+        self.AWREADY = getattr(dut, f"{prefix}_AWREADY")
+        self.AWVALID = getattr(dut, f"{prefix}_AWVALID")
+        self.AWADDR = getattr(dut, f"{prefix}_AWADDR")
+        self.AWPROT = getattr(dut, f"{prefix}_AWPROT")
+        self.WREADY = getattr(dut, f"{prefix}_WREADY")
+        self.WVALID = getattr(dut, f"{prefix}_WVALID")
+        self.WDATA = getattr(dut, f"{prefix}_WDATA")
+        self.WSTRB = getattr(dut, f"{prefix}_WSTRB")
+        self.BREADY = getattr(dut, f"{prefix}_BREADY")
+        self.BVALID = getattr(dut, f"{prefix}_BVALID")
+        self.BRESP = getattr(dut, f"{prefix}_BRESP")
+        self.ARREADY = getattr(dut, f"{prefix}_ARREADY")
+        self.ARVALID = getattr(dut, f"{prefix}_ARVALID")
+        self.ARADDR = getattr(dut, f"{prefix}_ARADDR")
+        self.ARPROT = getattr(dut, f"{prefix}_ARPROT")
+        self.RREADY = getattr(dut, f"{prefix}_RREADY")
+        self.RVALID = getattr(dut, f"{prefix}_RVALID")
+        self.RDATA = getattr(dut, f"{prefix}_RDATA")
+        self.RRESP = getattr(dut, f"{prefix}_RRESP")
+
+
+class _AxilMasterShim:
+    def __init__(self, dut, base: str):
+        self.AWREADY = getattr(dut, f"{base}_AWREADY")
+        self.AWVALID = getattr(dut, f"{base}_AWVALID")
+        self.AWADDR = getattr(dut, f"{base}_AWADDR")
+        self.AWPROT = getattr(dut, f"{base}_AWPROT")
+        self.WREADY = getattr(dut, f"{base}_WREADY")
+        self.WVALID = getattr(dut, f"{base}_WVALID")
+        self.WDATA = getattr(dut, f"{base}_WDATA")
+        self.WSTRB = getattr(dut, f"{base}_WSTRB")
+        self.BREADY = getattr(dut, f"{base}_BREADY")
+        self.BVALID = getattr(dut, f"{base}_BVALID")
+        self.BRESP = getattr(dut, f"{base}_BRESP")
+        self.ARREADY = getattr(dut, f"{base}_ARREADY")
+        self.ARVALID = getattr(dut, f"{base}_ARVALID")
+        self.ARADDR = getattr(dut, f"{base}_ARADDR")
+        self.ARPROT = getattr(dut, f"{base}_ARPROT")
+        self.RREADY = getattr(dut, f"{base}_RREADY")
+        self.RVALID = getattr(dut, f"{base}_RVALID")
+        self.RDATA = getattr(dut, f"{base}_RDATA")
+        self.RRESP = getattr(dut, f"{base}_RRESP")
+
+
+def _axil_slave(dut):
+    return getattr(dut, "s_axil", None) or _AxilSlaveShim(dut)
+
+
+def _axil_master(dut, base: str):
+    return getattr(dut, base, None) or _AxilMasterShim(dut, base)
+
+
+@cocotb.test()
+async def test_depth_1(dut):
+    """Test max_decode_depth=1 - should have interface for inner1 only."""
+    s_axil = _axil_slave(dut)
+    
+    # At depth 1, we should have m_axil_inner1 but not deeper interfaces
+    inner1 = _axil_master(dut, "m_axil_inner1")
+    
+    # Default slave side inputs
+    s_axil.AWVALID.value = 0
+    s_axil.AWADDR.value = 0
+    s_axil.AWPROT.value = 0
+    s_axil.WVALID.value = 0
+    s_axil.WDATA.value = 0
+    s_axil.WSTRB.value = 0
+    s_axil.BREADY.value = 0
+    s_axil.ARVALID.value = 0
+    s_axil.ARADDR.value = 0
+    s_axil.ARPROT.value = 0
+    s_axil.RREADY.value = 0
+    
+    inner1.AWREADY.value = 0
+    inner1.WREADY.value = 0
+    inner1.BVALID.value = 0
+    inner1.BRESP.value = 0
+    inner1.ARREADY.value = 0
+    inner1.RVALID.value = 0
+    inner1.RDATA.value = 0
+    inner1.RRESP.value = 0
+    
+    await Timer(1, units="ns")
+    
+    # Write to address 0x0 (should select inner1)
+    inner1.AWREADY.value = 1
+    inner1.WREADY.value = 1
+    s_axil.AWVALID.value = 1
+    s_axil.AWADDR.value = 0x0
+    s_axil.WVALID.value = 1
+    s_axil.WDATA.value = 0x12345678
+    s_axil.WSTRB.value = 0xF
+    
+    await Timer(1, units="ns")
+    
+    assert int(inner1.AWVALID.value) == 1, "inner1 write address valid must be set"
+    assert int(inner1.WVALID.value) == 1, "inner1 write data valid must be set"
+
+
+@cocotb.test()
+async def test_depth_2(dut):
+    """Test max_decode_depth=2 - should have interfaces for reg1 and inner2."""
+    s_axil = _axil_slave(dut)
+    
+    # At depth 2, we should have m_axil_reg1 and m_axil_inner2
+    reg1 = _axil_master(dut, "m_axil_reg1")
+    inner2 = _axil_master(dut, "m_axil_inner2")
+    
+    # Default slave side inputs
+    s_axil.AWVALID.value = 0
+    s_axil.AWADDR.value = 0
+    s_axil.AWPROT.value = 0
+    s_axil.WVALID.value = 0
+    s_axil.WDATA.value = 0
+    s_axil.WSTRB.value = 0
+    s_axil.BREADY.value = 0
+    s_axil.ARVALID.value = 0
+    s_axil.ARADDR.value = 0
+    s_axil.ARPROT.value = 0
+    s_axil.RREADY.value = 0
+    
+    for master in [reg1, inner2]:
+        master.AWREADY.value = 0
+        master.WREADY.value = 0
+        master.BVALID.value = 0
+        master.BRESP.value = 0
+        master.ARREADY.value = 0
+        master.RVALID.value = 0
+        master.RDATA.value = 0
+        master.RRESP.value = 0
+    
+    await Timer(1, units="ns")
+    
+    # Write to address 0x0 (should select reg1)
+    reg1.AWREADY.value = 1
+    reg1.WREADY.value = 1
+    s_axil.AWVALID.value = 1
+    s_axil.AWADDR.value = 0x0
+    s_axil.WVALID.value = 1
+    s_axil.WDATA.value = 0xABCDEF01
+    s_axil.WSTRB.value = 0xF
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg1.AWVALID.value) == 1, "reg1 must be selected for address 0x0"
+    assert int(inner2.AWVALID.value) == 0, "inner2 should not be selected"
+    
+    # Reset
+    s_axil.AWVALID.value = 0
+    s_axil.WVALID.value = 0
+    reg1.AWREADY.value = 0
+    reg1.WREADY.value = 0
+    await Timer(1, units="ns")
+    
+    # Write to address 0x10 (should select inner2)
+    inner2.AWREADY.value = 1
+    inner2.WREADY.value = 1
+    s_axil.AWVALID.value = 1
+    s_axil.AWADDR.value = 0x10
+    s_axil.WVALID.value = 1
+    s_axil.WDATA.value = 0x23456789
+    s_axil.WSTRB.value = 0xF
+    
+    await Timer(1, units="ns")
+    
+    assert int(inner2.AWVALID.value) == 1, "inner2 must be selected for address 0x10"
+    assert int(reg1.AWVALID.value) == 0, "reg1 should not be selected"
+
+
+@cocotb.test()
+async def test_depth_0(dut):
+    """Test max_decode_depth=0 - should have interfaces for all leaf registers."""
+    s_axil = _axil_slave(dut)
+    
+    # At depth 0, we should have all leaf registers: reg1, reg2, reg2b
+    reg1 = _axil_master(dut, "m_axil_reg1")
+    reg2 = _axil_master(dut, "m_axil_reg2")
+    reg2b = _axil_master(dut, "m_axil_reg2b")
+    
+    # Default slave side inputs
+    s_axil.AWVALID.value = 0
+    s_axil.AWADDR.value = 0
+    s_axil.AWPROT.value = 0
+    s_axil.WVALID.value = 0
+    s_axil.WDATA.value = 0
+    s_axil.WSTRB.value = 0
+    s_axil.BREADY.value = 0
+    s_axil.ARVALID.value = 0
+    s_axil.ARADDR.value = 0
+    s_axil.ARPROT.value = 0
+    s_axil.RREADY.value = 0
+    
+    for master in [reg1, reg2, reg2b]:
+        master.AWREADY.value = 0
+        master.WREADY.value = 0
+        master.BVALID.value = 0
+        master.BRESP.value = 0
+        master.ARREADY.value = 0
+        master.RVALID.value = 0
+        master.RDATA.value = 0
+        master.RRESP.value = 0
+    
+    await Timer(1, units="ns")
+    
+    # Write to address 0x0 (should select reg1)
+    reg1.AWREADY.value = 1
+    reg1.WREADY.value = 1
+    s_axil.AWVALID.value = 1
+    s_axil.AWADDR.value = 0x0
+    s_axil.WVALID.value = 1
+    s_axil.WDATA.value = 0x11111111
+    s_axil.WSTRB.value = 0xF
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg1.AWVALID.value) == 1, "reg1 must be selected for address 0x0"
+    assert int(reg2.AWVALID.value) == 0, "reg2 should not be selected"
+    assert int(reg2b.AWVALID.value) == 0, "reg2b should not be selected"
+    
+    # Reset
+    s_axil.AWVALID.value = 0
+    s_axil.WVALID.value = 0
+    reg1.AWREADY.value = 0
+    reg1.WREADY.value = 0
+    await Timer(1, units="ns")
+    
+    # Write to address 0x10 (should select reg2)
+    reg2.AWREADY.value = 1
+    reg2.WREADY.value = 1
+    s_axil.AWVALID.value = 1
+    s_axil.AWADDR.value = 0x10
+    s_axil.WVALID.value = 1
+    s_axil.WDATA.value = 0x22222222
+    s_axil.WSTRB.value = 0xF
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg2.AWVALID.value) == 1, "reg2 must be selected for address 0x10"
+    assert int(reg1.AWVALID.value) == 0, "reg1 should not be selected"
+    assert int(reg2b.AWVALID.value) == 0, "reg2b should not be selected"
+    
+    # Reset
+    s_axil.AWVALID.value = 0
+    s_axil.WVALID.value = 0
+    reg2.AWREADY.value = 0
+    reg2.WREADY.value = 0
+    await Timer(1, units="ns")
+    
+    # Write to address 0x14 (should select reg2b)
+    reg2b.AWREADY.value = 1
+    reg2b.WREADY.value = 1
+    s_axil.AWVALID.value = 1
+    s_axil.AWADDR.value = 0x14
+    s_axil.WVALID.value = 1
+    s_axil.WDATA.value = 0x33333333
+    s_axil.WSTRB.value = 0xF
+    
+    await Timer(1, units="ns")
+    
+    assert int(reg2b.AWVALID.value) == 1, "reg2b must be selected for address 0x14"
+    assert int(reg1.AWVALID.value) == 0, "reg1 should not be selected"
+    assert int(reg2.AWVALID.value) == 0, "reg2 should not be selected"

tests/cocotb/axi4lite/smoke/test_variable_depth_runner.py

@@ -0,0 +1,128 @@
+"""Pytest wrapper launching the AXI4-Lite cocotb smoke test for variable depth."""
+
+from pathlib import Path
+
+import pytest
+
+from peakrdl_busdecoder.cpuif.axi4lite.axi4_lite_cpuif_flat import AXI4LiteCpuifFlat
+
+try:  # pragma: no cover - optional dependency shim
+    from cocotb.runner import get_runner
+except ImportError:  # pragma: no cover
+    from cocotb_tools.runner import get_runner
+
+from tests.cocotb_lib.utils import compile_rdl_and_export, get_verilog_sources
+
+
+@pytest.mark.simulation
+@pytest.mark.verilator
+def test_axi4lite_variable_depth_1(tmp_path: Path) -> None:
+    """Test AXI4-Lite design with max_decode_depth=1."""
+    repo_root = Path(__file__).resolve().parents[4]
+
+    module_path, package_path = compile_rdl_and_export(
+        str(repo_root / "tests" / "cocotb_lib" / "variable_depth.rdl"),
+        "variable_depth",
+        tmp_path,
+        cpuif_cls=AXI4LiteCpuifFlat,
+        max_decode_depth=1,
+    )
+
+    sources = get_verilog_sources(
+        module_path,
+        package_path,
+        [repo_root / "hdl-src" / "axi4lite_intf.sv"],
+    )
+
+    runner = get_runner("verilator")
+    build_dir = tmp_path / "sim_build"
+
+    runner.build(
+        sources=sources,
+        hdl_toplevel=module_path.stem,
+        build_dir=build_dir,
+    )
+
+    runner.test(
+        hdl_toplevel=module_path.stem,
+        test_module="tests.cocotb.axi4lite.smoke.test_variable_depth",
+        build_dir=build_dir,
+        log_file=str(tmp_path / "sim_depth1.log"),
+        testcase="test_depth_1",
+    )
+
+
+@pytest.mark.simulation
+@pytest.mark.verilator
+def test_axi4lite_variable_depth_2(tmp_path: Path) -> None:
+    """Test AXI4-Lite design with max_decode_depth=2."""
+    repo_root = Path(__file__).resolve().parents[4]
+
+    module_path, package_path = compile_rdl_and_export(
+        str(repo_root / "tests" / "cocotb_lib" / "variable_depth.rdl"),
+        "variable_depth",
+        tmp_path,
+        cpuif_cls=AXI4LiteCpuifFlat,
+        max_decode_depth=2,
+    )
+
+    sources = get_verilog_sources(
+        module_path,
+        package_path,
+        [repo_root / "hdl-src" / "axi4lite_intf.sv"],
+    )
+
+    runner = get_runner("verilator")
+    build_dir = tmp_path / "sim_build"
+
+    runner.build(
+        sources=sources,
+        hdl_toplevel=module_path.stem,
+        build_dir=build_dir,
+    )
+
+    runner.test(
+        hdl_toplevel=module_path.stem,
+        test_module="tests.cocotb.axi4lite.smoke.test_variable_depth",
+        build_dir=build_dir,
+        log_file=str(tmp_path / "sim_depth2.log"),
+        testcase="test_depth_2",
+    )
+
+
+@pytest.mark.simulation
+@pytest.mark.verilator
+def test_axi4lite_variable_depth_0(tmp_path: Path) -> None:
+    """Test AXI4-Lite design with max_decode_depth=0 (all levels)."""
+    repo_root = Path(__file__).resolve().parents[4]
+
+    module_path, package_path = compile_rdl_and_export(
+        str(repo_root / "tests" / "cocotb_lib" / "variable_depth.rdl"),
+        "variable_depth",
+        tmp_path,
+        cpuif_cls=AXI4LiteCpuifFlat,
+        max_decode_depth=0,
+    )
+
+    sources = get_verilog_sources(
+        module_path,
+        package_path,
+        [repo_root / "hdl-src" / "axi4lite_intf.sv"],
+    )
+
+    runner = get_runner("verilator")
+    build_dir = tmp_path / "sim_build"
+
+    runner.build(
+        sources=sources,
+        hdl_toplevel=module_path.stem,
+        build_dir=build_dir,
+    )
+
+    runner.test(
+        hdl_toplevel=module_path.stem,
+        test_module="tests.cocotb.axi4lite.smoke.test_variable_depth",
+        build_dir=build_dir,
+        log_file=str(tmp_path / "sim_depth0.log"),
+        testcase="test_depth_0",
+    )

tests/cocotb_lib/variable_depth.rdl

@@ -0,0 +1,31 @@
+// Variable depth register hierarchy for testing max_decode_depth parameter
+addrmap level2 {
+    reg {
+        field {
+            sw=rw;
+            hw=r;
+        } data2[31:0];
+    } reg2 @ 0x0;
+    
+    reg {
+        field {
+            sw=rw;
+            hw=r;
+        } data2b[31:0];
+    } reg2b @ 0x4;
+};
+
+addrmap level1 {
+    reg {
+        field {
+            sw=rw;
+            hw=r;
+        } data1[31:0];
+    } reg1 @ 0x0;
+    
+    level2 inner2 @ 0x10;
+};
+
+addrmap variable_depth {
+    level1 inner1 @ 0x0;
+};

tests/conftest.py

@@ -2,7 +2,7 @@
 
 from __future__ import annotations
 
-collect_ignore_glob = ["cocotb/*/smoke/test_register_access.py"]
+collect_ignore_glob = ["cocotb/*/smoke/test_register_access.py", "cocotb/*/smoke/test_variable_depth.py"]
 
 import os
 from collections.abc import Callable

tests/exporter/test_bus_decoder_exporter.py

@@ -89,7 +89,8 @@ class TestBusDecoderExporter:
 
         exporter = BusDecoderExporter()
         output_dir = str(tmp_path)
-        exporter.export(top, output_dir, cpuif_cls=APB4Cpuif)
+        # Use depth=0 to descend all the way to registers
+        exporter.export(top, output_dir, cpuif_cls=APB4Cpuif, max_decode_depth=0)
 
         # Check that output files are created
         module_file = tmp_path / "outer_block.sv"

tests/generator/test_questa_compatibility.py

@@ -0,0 +1,131 @@
+"""Test Questa simulator compatibility for instance arrays."""
+
+from collections.abc import Callable
+from pathlib import Path
+from tempfile import TemporaryDirectory
+
+from systemrdl.node import AddrmapNode
+
+from peakrdl_busdecoder import BusDecoderExporter
+from peakrdl_busdecoder.cpuif.apb4 import APB4Cpuif
+
+
+def test_instance_array_questa_compatibility(compile_rdl: Callable[..., AddrmapNode]) -> None:
+    """Test that instance arrays generate Questa-compatible code.
+    
+    This test ensures that:
+    - Struct members for arrays use unpacked array syntax (name[dim])
+    - NOT packed bit-vector syntax ([dim-1:0]name)
+    - Struct is unpacked (not packed)
+    - Array indexing with loop variables works correctly
+    
+    This fixes the error: "Nonconstant index into instance array"
+    """
+    rdl_source = """
+    addrmap test_map {
+        reg {
+            field {
+                sw=rw;
+                hw=r;
+            } data[31:0];
+        } my_reg[4] @ 0x0 += 0x10;
+    };
+    """
+    top = compile_rdl(rdl_source, top="test_map")
+    
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif)
+        
+        # Read the generated module
+        module_file = Path(tmpdir) / "test_map.sv"
+        content = module_file.read_text()
+        
+        # Should use unpacked struct
+        assert "typedef struct {" in content
+        assert "typedef struct packed" not in content
+        
+        # Should use unpacked array syntax for array members
+        assert "logic my_reg[4];" in content
+        
+        # Should NOT use packed bit-vector syntax
+        assert "[3:0]my_reg" not in content
+        
+        # Should have proper array indexing in decode logic
+        assert "cpuif_wr_sel.my_reg[i0] = 1'b1;" in content
+        assert "cpuif_rd_sel.my_reg[i0] = 1'b1;" in content
+        
+        # Should have proper array indexing in fanout/fanin logic
+        assert "cpuif_wr_sel.my_reg[gi0]" in content or "cpuif_rd_sel.my_reg[gi0]" in content
+        assert "cpuif_wr_sel.my_reg[i0]" in content or "cpuif_rd_sel.my_reg[i0]" in content
+
+
+def test_multidimensional_array_questa_compatibility(compile_rdl: Callable[..., AddrmapNode]) -> None:
+    """Test that multidimensional instance arrays generate Questa-compatible code."""
+    rdl_source = """
+    addrmap test_map {
+        reg {
+            field {
+                sw=rw;
+                hw=r;
+            } data[31:0];
+        } my_reg[2][3] @ 0x0 += 0x10;
+    };
+    """
+    top = compile_rdl(rdl_source, top="test_map")
+    
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif)
+        
+        # Read the generated module
+        module_file = Path(tmpdir) / "test_map.sv"
+        content = module_file.read_text()
+        
+        # Should use unpacked struct with multidimensional array
+        assert "typedef struct {" in content
+        
+        # Should use unpacked array syntax for multidimensional arrays
+        assert "logic my_reg[2][3];" in content
+        
+        # Should NOT use packed bit-vector syntax
+        assert "[1:0][2:0]my_reg" not in content
+        assert "[5:0]my_reg" not in content
+
+
+def test_nested_instance_array_questa_compatibility(compile_rdl: Callable[..., AddrmapNode]) -> None:
+    """Test that nested instance arrays generate Questa-compatible code."""
+    rdl_source = """
+    addrmap inner_map {
+        reg {
+            field {
+                sw=rw;
+                hw=r;
+            } data[31:0];
+        } inner_reg[2] @ 0x0 += 0x10;
+    };
+    
+    addrmap outer_map {
+        inner_map inner[3] @ 0x0 += 0x100;
+    };
+    """
+    top = compile_rdl(rdl_source, top="outer_map")
+    
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif)
+        
+        # Read the generated module
+        module_file = Path(tmpdir) / "outer_map.sv"
+        content = module_file.read_text()
+        
+        # Should use unpacked struct
+        assert "typedef struct {" in content
+        
+        # Inner should be an array
+        # The exact syntax may vary, but it should be unpacked
+        # Look for the pattern of unpacked arrays, not packed bit-vectors
+        assert "inner[3]" in content or "logic inner" in content
+        
+        # Should NOT use packed bit-vector syntax like [2:0]inner
+        assert "[2:0]inner" not in content

tests/generator/test_struct_generator.py

@@ -96,3 +96,10 @@ class TestStructGenerator:
         assert "[" in result and "]" in result
         # Should reference the register
         assert "my_regs" in result
+        # Should use unpacked array syntax (name[size]), not packed bit-vector ([size:0]name)
+        assert "my_regs[4]" in result
+        # Should NOT use packed bit-vector syntax
+        assert "[3:0]my_regs" not in result
+        # Should be unpacked struct, not packed
+        assert "typedef struct {" in result
+        assert "typedef struct packed" not in result

tests/unit/test_external_nested.py

@@ -41,9 +41,9 @@ def test_external_nested_components_generate_correct_decoder(external_nested_rdl
         assert "cpuif_rd_sel.multicast.response" not in content
         
         # Verify struct is flat (no nested structs for external children)
-        assert "typedef struct packed" in content
+        assert "typedef struct" in content
         assert "logic multicast;" in content
-        assert "logic [15:0]port;" in content
+        assert "logic port[16];" in content
 
 
 def test_external_nested_components_generate_correct_interfaces(external_nested_rdl: AddrmapNode) -> None:
@@ -101,7 +101,8 @@ def test_non_external_nested_components_are_descended(compile_rdl: Callable[...,
     
     with TemporaryDirectory() as tmpdir:
         exporter = BusDecoderExporter()
-        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif)
+        # Use depth=0 to descend all the way down to registers
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif, max_decode_depth=0)
         
         # Read the generated module
         module_file = Path(tmpdir) / "outer_block.sv"

tests/unit/test_max_decode_depth.py

@@ -0,0 +1,256 @@
+"""Test max_decode_depth parameter behavior."""
+
+from collections.abc import Callable
+from pathlib import Path
+from tempfile import TemporaryDirectory
+
+from systemrdl.node import AddrmapNode
+
+from peakrdl_busdecoder import BusDecoderExporter
+from peakrdl_busdecoder.cpuif.apb4 import APB4Cpuif
+
+
+def test_depth_1_generates_top_level_interface_only(compile_rdl: Callable[..., AddrmapNode]) -> None:
+    """Test that depth=1 generates interface only for top-level children."""
+    rdl_source = """
+    addrmap level1 {
+        reg {
+            field { sw=rw; hw=r; } data1[31:0];
+        } reg1 @ 0x0;
+    };
+
+    addrmap level0 {
+        level1 inner1 @ 0x0;
+    };
+    """
+    top = compile_rdl(rdl_source, top="level0")
+
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif, max_decode_depth=1)
+
+        module_file = Path(tmpdir) / "level0.sv"
+        content = module_file.read_text()
+
+        # Should have interface for inner1 only
+        assert "m_apb_inner1" in content
+        # Should NOT have interface for reg1
+        assert "m_apb_reg1" not in content
+        
+        # Struct should have inner1 but not nested structure
+        assert "logic inner1;" in content
+
+
+def test_depth_2_generates_second_level_interfaces(compile_rdl: Callable[..., AddrmapNode]) -> None:
+    """Test that depth=2 generates interfaces for second-level children."""
+    rdl_source = """
+    addrmap level2 {
+        reg {
+            field { sw=rw; hw=r; } data2[31:0];
+        } reg2 @ 0x0;
+    };
+
+    addrmap level1 {
+        reg {
+            field { sw=rw; hw=r; } data1[31:0];
+        } reg1 @ 0x0;
+        
+        level2 inner2 @ 0x10;
+    };
+
+    addrmap level0 {
+        level1 inner1 @ 0x0;
+    };
+    """
+    top = compile_rdl(rdl_source, top="level0")
+
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif, max_decode_depth=2)
+
+        module_file = Path(tmpdir) / "level0.sv"
+        content = module_file.read_text()
+
+        # Should have interfaces for reg1 and inner2
+        assert "m_apb_reg1" in content
+        assert "m_apb_inner2" in content
+        # Should NOT have interface for inner1 or reg2
+        assert "m_apb_inner1" not in content
+        assert "m_apb_reg2" not in content
+        
+        # Struct should be hierarchical with inner1.reg1 and inner1.inner2
+        assert "cpuif_sel_inner1_t" in content
+        assert "logic reg1;" in content
+        assert "logic inner2;" in content
+
+
+def test_depth_0_decodes_all_levels(compile_rdl: Callable[..., AddrmapNode]) -> None:
+    """Test that depth=0 decodes all the way down to registers."""
+    rdl_source = """
+    addrmap level2 {
+        reg {
+            field { sw=rw; hw=r; } data2[31:0];
+        } reg2 @ 0x0;
+        
+        reg {
+            field { sw=rw; hw=r; } data2b[31:0];
+        } reg2b @ 0x4;
+    };
+
+    addrmap level1 {
+        reg {
+            field { sw=rw; hw=r; } data1[31:0];
+        } reg1 @ 0x0;
+        
+        level2 inner2 @ 0x10;
+    };
+
+    addrmap level0 {
+        level1 inner1 @ 0x0;
+    };
+    """
+    top = compile_rdl(rdl_source, top="level0")
+
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif, max_decode_depth=0)
+
+        module_file = Path(tmpdir) / "level0.sv"
+        content = module_file.read_text()
+
+        # Should have interfaces for all leaf registers
+        assert "m_apb_reg1" in content
+        assert "m_apb_reg2" in content
+        assert "m_apb_reg2b" in content
+        # Should NOT have interfaces for addrmaps
+        assert "m_apb_inner1" not in content
+        assert "m_apb_inner2" not in content
+        
+        # Struct should be fully hierarchical
+        assert "cpuif_sel_inner1_t" in content
+        assert "cpuif_sel_inner2_t" in content
+
+
+def test_depth_affects_decode_logic(compile_rdl: Callable[..., AddrmapNode]) -> None:
+    """Test that decode logic changes based on max_decode_depth."""
+    rdl_source = """
+    addrmap level1 {
+        reg {
+            field { sw=rw; hw=r; } data1[31:0];
+        } reg1 @ 0x0;
+    };
+
+    addrmap level0 {
+        level1 inner1 @ 0x0;
+    };
+    """
+    top = compile_rdl(rdl_source, top="level0")
+
+    # Test depth=1: should set cpuif_wr_sel.inner1
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif, max_decode_depth=1)
+
+        module_file = Path(tmpdir) / "level0.sv"
+        content = module_file.read_text()
+
+        assert "cpuif_wr_sel.inner1 = 1'b1;" in content
+        assert "cpuif_wr_sel.inner1.reg1" not in content
+
+    # Test depth=2: should set cpuif_wr_sel.inner1.reg1
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif, max_decode_depth=2)
+
+        module_file = Path(tmpdir) / "level0.sv"
+        content = module_file.read_text()
+
+        assert "cpuif_wr_sel.inner1.reg1 = 1'b1;" in content
+        assert "cpuif_wr_sel.inner1 = 1'b1;" not in content
+
+
+def test_depth_affects_fanout_fanin(compile_rdl: Callable[..., AddrmapNode]) -> None:
+    """Test that fanout/fanin logic changes based on max_decode_depth."""
+    rdl_source = """
+    addrmap level1 {
+        reg {
+            field { sw=rw; hw=r; } data1[31:0];
+        } reg1 @ 0x0;
+    };
+
+    addrmap level0 {
+        level1 inner1 @ 0x0;
+    };
+    """
+    top = compile_rdl(rdl_source, top="level0")
+
+    # Test depth=1: should have fanout for inner1
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif, max_decode_depth=1)
+
+        module_file = Path(tmpdir) / "level0.sv"
+        content = module_file.read_text()
+
+        assert "m_apb_inner1.PSEL" in content
+        assert "m_apb_reg1.PSEL" not in content
+
+    # Test depth=2: should have fanout for reg1
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif, max_decode_depth=2)
+
+        module_file = Path(tmpdir) / "level0.sv"
+        content = module_file.read_text()
+
+        assert "m_apb_reg1.PSEL" in content
+        assert "m_apb_inner1.PSEL" not in content
+
+
+def test_depth_3_with_deep_hierarchy(compile_rdl: Callable[..., AddrmapNode]) -> None:
+    """Test depth=3 with a 4-level deep hierarchy."""
+    rdl_source = """
+    addrmap level3 {
+        reg {
+            field { sw=rw; hw=r; } data3[31:0];
+        } reg3 @ 0x0;
+    };
+
+    addrmap level2 {
+        reg {
+            field { sw=rw; hw=r; } data2[31:0];
+        } reg2 @ 0x0;
+        
+        level3 inner3 @ 0x10;
+    };
+
+    addrmap level1 {
+        reg {
+            field { sw=rw; hw=r; } data1[31:0];
+        } reg1 @ 0x0;
+        
+        level2 inner2 @ 0x10;
+    };
+
+    addrmap level0 {
+        level1 inner1 @ 0x0;
+    };
+    """
+    top = compile_rdl(rdl_source, top="level0")
+
+    with TemporaryDirectory() as tmpdir:
+        exporter = BusDecoderExporter()
+        exporter.export(top, tmpdir, cpuif_cls=APB4Cpuif, max_decode_depth=3)
+
+        module_file = Path(tmpdir) / "level0.sv"
+        content = module_file.read_text()
+
+        # Should have interfaces at depth 3: reg2, inner3
+        # (reg1 is at depth 2, not 3)
+        assert "m_apb_reg2" in content
+        assert "m_apb_inner3" in content
+        # Should NOT have interfaces at other depths
+        assert "m_apb_inner1" not in content
+        assert "m_apb_inner2" not in content
+        assert "m_apb_reg1" not in content
+        assert "m_apb_reg3" not in content

tools/shims/xargs

@@ -9,7 +9,6 @@ invocation constructed from stdin tokens.
 
 from __future__ import annotations
 
-import os
 import subprocess
 import sys
 

uv.lock

@@ -608,7 +608,7 @@ wheels = [
 
 [[package]]
 name = "peakrdl-busdecoder"
-version = "0.2.0"
+version = "0.4.0"
 source = { editable = "." }
 dependencies = [
     { name = "jinja2" },