Merge pull request #74 from pythonbpf/fix-vmlinux-ir-gen

Fix some issues with vmlinux struct usage in syntax
TODO: THIS DOES NOT FIX the XDP example. 
The issue with the XDP example here is that the range is not being tracked due to multiple loads from stack which is making the verifier lose track of the range on that value.

pythonbpf/allocation_pass.py

@@ -114,9 +114,22 @@ def _allocate_for_call(
         # Struct constructors
         elif call_type in structs_sym_tab:
             struct_info = structs_sym_tab[call_type]
-            var = builder.alloca(struct_info.ir_type, name=var_name)
-            local_sym_tab[var_name] = LocalSymbol(var, struct_info.ir_type, call_type)
-            logger.info(f"Pre-allocated {var_name} for struct {call_type}")
+            if len(rval.args) == 0:
+                # Zero-arg constructor: allocate the struct itself
+                var = builder.alloca(struct_info.ir_type, name=var_name)
+                local_sym_tab[var_name] = LocalSymbol(
+                    var, struct_info.ir_type, call_type
+                )
+                logger.info(f"Pre-allocated {var_name} for struct {call_type}")
+            else:
+                # Pointer cast: allocate as pointer to struct
+                ptr_type = ir.PointerType(struct_info.ir_type)
+                var = builder.alloca(ptr_type, name=var_name)
+                var.align = 8
+                local_sym_tab[var_name] = LocalSymbol(var, ptr_type, call_type)
+                logger.info(
+                    f"Pre-allocated {var_name} for struct pointer cast to {call_type}"
+                )
 
         elif VmlinuxHandlerRegistry.is_vmlinux_struct(call_type):
             # When calling struct_name(pointer), we're doing a cast, not construction
@@ -371,6 +384,7 @@ def _allocate_for_attribute(builder, var_name, rval, local_sym_tab, structs_sym_
                         f"Could not determine size for ctypes field {field_name}: {e}"
                     )
                     actual_ir_type = ir.IntType(64)
+                    field_size_bits = 64
 
             # Check if it's a nested vmlinux struct or complex type
             elif field.type.__module__ == "vmlinux":
@@ -379,23 +393,34 @@ def _allocate_for_attribute(builder, var_name, rval, local_sym_tab, structs_sym_
                     field.ctype_complex_type, ctypes._Pointer
                 ):
                     actual_ir_type = ir.IntType(64)  # Pointer is always 64-bit
+                    field_size_bits = 64
                 # For embedded structs, this is more complex - might need different handling
                 else:
                     logger.warning(
                         f"Field {field_name} is a nested vmlinux struct, using i64 for now"
                     )
                     actual_ir_type = ir.IntType(64)
+                    field_size_bits = 64
             else:
                 logger.warning(
                     f"Unknown field type module {field.type.__module__} for {field_name}"
                 )
                 actual_ir_type = ir.IntType(64)
+                field_size_bits = 64
 
-            # Allocate with the actual IR type
+            # Pre-allocate the tmp storage used by load_struct_field (so we don't alloca inside handler)
+            tmp_name = f"{struct_var}_{field_name}_tmp"
+            tmp_ir_type = ir.IntType(field_size_bits)
+            tmp_var = builder.alloca(tmp_ir_type, name=tmp_name)
+            tmp_var.align = tmp_ir_type.width // 8
+            local_sym_tab[tmp_name] = LocalSymbol(tmp_var, tmp_ir_type)
+            logger.info(
+                f"Pre-allocated temp {tmp_name} (i{field_size_bits}) for vmlinux field read {vmlinux_struct_name}.{field_name}"
+            )
+
+            # Allocate with the actual IR type for the destination var
             var = _allocate_with_type(builder, var_name, actual_ir_type)
-            local_sym_tab[var_name] = LocalSymbol(
-                var, actual_ir_type, field
-            )  # <-- Store Field metadata
+            local_sym_tab[var_name] = LocalSymbol(var, actual_ir_type, field)
 
             logger.info(
                 f"Pre-allocated {var_name} as {actual_ir_type} from vmlinux struct {vmlinux_struct_name}.{field_name}"

pythonbpf/assign_pass.py

@@ -174,6 +174,23 @@ def handle_variable_assignment(
                     f"Type mismatch: vmlinux struct pointer requires i64, got {var_type}"
                 )
                 return False
+        # Handle user-defined struct pointer casts
+        # val_type is a string (struct name), var_type is a pointer to the struct
+        if isinstance(val_type, str) and val_type in structs_sym_tab:
+            struct_info = structs_sym_tab[val_type]
+            expected_ptr_type = ir.PointerType(struct_info.ir_type)
+
+            # Check if var_type matches the expected pointer type
+            if isinstance(var_type, ir.PointerType) and var_type == expected_ptr_type:
+                # val is already the correct pointer type from inttoptr/bitcast
+                builder.store(val, var_ptr)
+                logger.info(f"Assigned user-defined struct pointer cast to {var_name}")
+                return True
+            else:
+                logger.error(
+                    f"Type mismatch: user-defined struct pointer cast requires pointer type, got {var_type}"
+                )
+                return False
         if isinstance(val_type, Field):
             logger.info("Handling assignment to struct field")
             # Special handling for struct_xdp_md i32 fields that are zero-extended to i64

pythonbpf/expr/expr_pass.py

@@ -168,7 +168,11 @@ def get_operand_value(
             var_type = var.type
             base_type, depth = get_base_type_and_depth(var_type)
             logger.info(f"var is {var}, base_type is {base_type}, depth is {depth}")
-            val = deref_to_depth(func, builder, var, depth)
+            if depth == 1:
+                val = builder.load(var)
+                return val
+            else:
+                val = deref_to_depth(func, builder, var, depth)
             return val
         else:
             # Check if it's a vmlinux enum/constant
@@ -545,7 +549,7 @@ def _handle_boolean_op(
 
 
 # ============================================================================
-# VMLinux casting
+# Struct casting (including vmlinux struct casting)
 # ============================================================================
 
 
@@ -593,17 +597,85 @@ def _handle_vmlinux_cast(
     # Cast the integer/value to a pointer to the struct
     # If arg_val is an integer type, we need to inttoptr it
     ptr_type = ir.PointerType()
-    # TODO: add a integer check here later
-    if ctypes_to_ir(arg_type.type.__name__):
-        # Cast integer to pointer
-        casted_ptr = builder.inttoptr(arg_val, ptr_type)
+    # TODO: add a field value type check here
+    # print(arg_type)
+    if isinstance(arg_type, Field):
+        if ctypes_to_ir(arg_type.type.__name__):
+            # Cast integer to pointer
+            casted_ptr = builder.inttoptr(arg_val, ptr_type)
+        else:
+            logger.error(f"Unsupported type for vmlinux cast: {arg_type}")
+            return None
     else:
-        logger.error(f"Unsupported type for vmlinux cast: {arg_type}")
-        return None
+        casted_ptr = builder.inttoptr(arg_val, ptr_type)
 
     return casted_ptr, vmlinux_struct_type
 
 
+def _handle_user_defined_struct_cast(
+    func,
+    module,
+    builder,
+    expr,
+    local_sym_tab,
+    map_sym_tab,
+    structs_sym_tab,
+):
+    """Handle user-defined struct cast expressions like iphdr(nh).
+
+    This casts a pointer/integer value to a pointer to the user-defined struct,
+    similar to how vmlinux struct casts work but for user-defined @struct types.
+    """
+    if len(expr.args) != 1:
+        logger.info("User-defined struct cast takes exactly one argument")
+        return None
+
+    # Get the struct name
+    struct_name = expr.func.id
+
+    if struct_name not in structs_sym_tab:
+        logger.error(f"Struct {struct_name} not found in structs_sym_tab")
+        return None
+
+    struct_info = structs_sym_tab[struct_name]
+
+    # Evaluate the argument (e.g.,
+    # an address/pointer value)
+    arg_result = eval_expr(
+        func,
+        module,
+        builder,
+        expr.args[0],
+        local_sym_tab,
+        map_sym_tab,
+        structs_sym_tab,
+    )
+
+    if arg_result is None:
+        logger.info("Failed to evaluate argument to user-defined struct cast")
+        return None
+
+    arg_val, arg_type = arg_result
+
+    # Cast the integer/pointer value to a pointer to the struct type
+    # The struct pointer type is a pointer to the struct's IR type
+    struct_ptr_type = ir.PointerType(struct_info.ir_type)
+
+    # If arg_val is an integer type (like i64), convert to pointer using inttoptr
+    if isinstance(arg_val.type, ir.IntType):
+        casted_ptr = builder.inttoptr(arg_val, struct_ptr_type)
+        logger.info(f"Cast integer to pointer for struct {struct_name}")
+    elif isinstance(arg_val.type, ir.PointerType):
+        # If already a pointer, bitcast to the struct pointer type
+        casted_ptr = builder.bitcast(arg_val, struct_ptr_type)
+        logger.info(f"Bitcast pointer to struct pointer for {struct_name}")
+    else:
+        logger.error(f"Unsupported type for user-defined struct cast: {arg_val.type}")
+        return None
+
+    return casted_ptr, struct_name
+
+
 # ============================================================================
 # Expression Dispatcher
 # ============================================================================
@@ -649,6 +721,16 @@ def eval_expr(
                 map_sym_tab,
                 structs_sym_tab,
             )
+        if isinstance(expr.func, ast.Name) and (expr.func.id in structs_sym_tab):
+            return _handle_user_defined_struct_cast(
+                func,
+                module,
+                builder,
+                expr,
+                local_sym_tab,
+                map_sym_tab,
+                structs_sym_tab,
+            )
 
         result = CallHandlerRegistry.handle_call(
             expr, module, builder, func, local_sym_tab, map_sym_tab, structs_sym_tab

pythonbpf/type_deducer.py

@@ -18,6 +18,10 @@ mapping = {
     "c_longlong": ir.IntType(64),
     "c_uint": ir.IntType(32),
     "c_int": ir.IntType(32),
+    "c_ushort": ir.IntType(16),
+    "c_short": ir.IntType(16),
+    "c_ubyte": ir.IntType(8),
+    "c_byte": ir.IntType(8),
     # Not so sure about this one
     "str": ir.PointerType(ir.IntType(8)),
 }

pythonbpf/vmlinux_parser/vmlinux_exports_handler.py

@@ -77,7 +77,7 @@ class VmlinuxHandler:
         return None
 
     def get_vmlinux_enum_value(self, name):
-        """Handle vmlinux enum constants by returning LLVM IR constants"""
+        """Handle vmlinux.enum constants by returning LLVM IR constants"""
         if self.is_vmlinux_enum(name):
             value = self.vmlinux_symtab[name].value
             logger.info(f"The value of vmlinux enum {name} = {value}")
@@ -119,9 +119,18 @@ class VmlinuxHandler:
                 # Load the struct pointer from the local variable
                 struct_ptr = builder.load(var_info.var)
 
+                # Determine the preallocated tmp name that assignment pass should have created
+                tmp_name = f"{struct_var_name}_{field_name}_tmp"
+
                 # Use bpf_probe_read_kernel for non-context struct field access
                 field_value = self.load_struct_field(
-                    builder, struct_ptr, globvar_ir, field_data, struct_name
+                    builder,
+                    struct_ptr,
+                    globvar_ir,
+                    field_data,
+                    struct_name,
+                    local_sym_tab,
+                    tmp_name,
                 )
                 # Return field value and field type
                 return field_value, field_data
@@ -130,7 +139,13 @@ class VmlinuxHandler:
 
     @staticmethod
     def load_struct_field(
-        builder, struct_ptr_int, offset_global, field_data, struct_name=None
+        builder,
+        struct_ptr_int,
+        offset_global,
+        field_data,
+        struct_name=None,
+        local_sym_tab=None,
+        tmp_name: str | None = None,
     ):
         """
         Generate LLVM IR to load a field from a regular (non-context) struct using bpf_probe_read_kernel.
@@ -141,6 +156,8 @@ class VmlinuxHandler:
             offset_global: Global variable containing the field offset (i64)
             field_data: contains data about the field
             struct_name: Name of the struct being accessed (optional)
+            local_sym_tab: symbol table (optional) - used to locate preallocated tmp storage
+            tmp_name: name of the preallocated temporary storage to use (preferred)
         Returns:
             The loaded value
         """
@@ -203,9 +220,18 @@ class VmlinuxHandler:
                 else:
                     logger.warning("Complex vmlinux field type, using default 64 bits")
 
-        # Allocate local storage for the field value
-        local_storage = builder.alloca(ir.IntType(int_width))
-        local_storage_i8_ptr = builder.bitcast(local_storage, i8_ptr_type)
+        # Use preallocated temporary storage if provided by allocation pass
+
+        local_storage_i8_ptr = None
+        if tmp_name and local_sym_tab and tmp_name in local_sym_tab:
+            # Expect the tmp to be an alloca created during allocation pass
+            tmp_alloca = local_sym_tab[tmp_name].var
+            local_storage_i8_ptr = builder.bitcast(tmp_alloca, i8_ptr_type)
+        else:
+            # Fallback: allocate inline (not ideal, but preserves behavior)
+            local_storage = builder.alloca(ir.IntType(int_width))
+            local_storage_i8_ptr = builder.bitcast(local_storage, i8_ptr_type)
+            logger.warning(f"Temp storage '{tmp_name}' not found. Allocating inline")
 
         # Use bpf_probe_read_kernel to safely read the field
         # This generates:
@@ -219,7 +245,9 @@ class VmlinuxHandler:
         )
 
         # Load the value from local storage
-        value = builder.load(local_storage)
+        value = builder.load(
+            builder.bitcast(local_storage_i8_ptr, ir.PointerType(ir.IntType(int_width)))
+        )
 
         # Zero-extend i32 to i64 if needed
         if needs_zext: