geenrate gep IR

Author: varun-r-mallya

pythonbpf/assign_pass.py

@@ -3,6 +3,8 @@
 from llvmlite import ir
 from pythonbpf.expr import eval_expr
 from pythonbpf.helper import emit_probe_read_kernel_str_call
+from pythonbpf.type_deducer import ctypes_to_ir
+from pythonbpf.vmlinux_parser.dependency_node import Field
 
 logger = logging.getLogger(__name__)
 
@@ -148,7 +150,16 @@ def handle_variable_assignment(
     val, val_type = val_result
     logger.info(f"Evaluated value for {var_name}: {val} of type {val_type}, {var_type}")
     if val_type != var_type:
-        if isinstance(val_type, ir.IntType) and isinstance(var_type, ir.IntType):
+        if isinstance(val_type, Field):
+            logger.info("Handling assignment to struct field")
+            #TODO: handling only ctype struct fields for now. Handle other stuff too later.
+            if var_type == ctypes_to_ir(val_type.type.__name__):
+                builder.store(val, var_ptr)
+                logger.info(f"Assigned ctype struct field to {var_name}")
+                return True
+            logger.error(f"Failed to assign ctype struct field to {var_name}: {val_type} != {var_type}")
+            return False
+        elif isinstance(val_type, ir.IntType) and isinstance(var_type, ir.IntType):
             # Allow implicit int widening
             if val_type.width < var_type.width:
                 val = builder.sext(val, var_type)

pythonbpf/codegen.py

@@ -157,7 +157,8 @@ def compile_to_ir(filename: str, output: str, loglevel=logging.INFO):
 
     module.add_named_metadata("llvm.ident", [f"PythonBPF {VERSION}"])
 
-    module_string = finalize_module(str(module))
+    module_string: str = finalize_module(str(module))
+    module_string += '\ndeclare ptr @llvm.preserve.struct.access.index.p0.p0(ptr, i32 immarg, i32 immarg) "nocallback" "nofree" "nosync" "nounwind" "willreturn" "memory(none)"'
 
     logger.info(f"IR written to {output}")
     with open(output, "w") as f:

pythonbpf/type_deducer.py

@@ -13,6 +13,8 @@
     "c_float": ir.FloatType(),
     "c_double": ir.DoubleType(),
     "c_void_p": ir.IntType(64),
+    "c_long": ir.IntType(64),
+    "c_longlong": ir.IntType(64),
     # Not so sure about this one
     "str": ir.PointerType(ir.IntType(8)),
 }

pythonbpf/vmlinux_parser/vmlinux_exports_handler.py

@@ -73,17 +73,6 @@ def get_vmlinux_enum_value(self, name):
             return value
         return None
 
-    def handle_vmlinux_struct(self, struct_name, module, builder):
-        """Handle vmlinux struct initializations"""
-        if self.is_vmlinux_struct(struct_name):
-            # TODO: Implement core-specific struct handling
-            # This will be more complex and depends on the BTF information
-            logger.info(f"Handling vmlinux struct {struct_name}")
-            # Return struct type and allocated pointer
-            # This is a stub, actual implementation will be more complex
-            return None
-        return None
-
     def handle_vmlinux_struct_field(
         self, struct_var_name, field_name, module, builder, local_sym_tab
     ):
@@ -97,26 +86,82 @@ def handle_vmlinux_struct_field(
             globvar_ir, field_data = self.get_field_type(
                 python_type.__name__, field_name
             )
+            builder.function.args[0].type = ir.PointerType(ir.IntType(8))
+            print(builder.function.args[0])
+            field_ptr = self.load_ctx_field(builder, builder.function.args[0], globvar_ir)
+            print(field_ptr)
+            # Return pointer to field and field type
+            return field_ptr, field_data
+        else:
+            raise RuntimeError("Variable accessed not found in symbol table")
 
-            # Load the offset value
-            offset = builder.load(globvar_ir)
+    @staticmethod
+    def load_ctx_field(builder, ctx_arg, offset_global):
+        """
+        Generate LLVM IR to load a field from BPF context using offset.
 
-            # Convert pointer to integer
-            i64_type = ir.IntType(64)
-            ptr_as_int = builder.ptrtoint(var_info.var, i64_type)
+        Args:
+            builder: llvmlite IRBuilder instance
+            ctx_arg: The context pointer argument (ptr/i8*)
+            offset_global: Global variable containing the field offset (i64)
 
-            # Add the offset
-            field_addr = builder.add(ptr_as_int, offset)
+        Returns:
+            The loaded value (i64 register)
+        """
 
-            # Convert back to pointer
-            i8_ptr_type = ir.IntType(8).as_pointer()
-            field_ptr_i8 = builder.inttoptr(field_addr, i8_ptr_type)
-            logger.info(f"field_ptr_i8: {field_ptr_i8}")
+        # Load the offset value
+        offset = builder.load(offset_global)
 
-            # Return pointer to field and field type
-            return field_ptr_i8, field_data
+        # Ensure ctx_arg is treated as i8* (byte pointer)
+        i8_type = ir.IntType(8)
+        i8_ptr_type = ir.PointerType(i8_type)
+
+        # Cast ctx_arg to i8* if it isn't already
+        if str(ctx_arg.type) != str(i8_ptr_type):
+            ctx_i8_ptr = builder.bitcast(ctx_arg, i8_ptr_type)
         else:
-            raise RuntimeError("Variable accessed not found in symbol table")
+            ctx_i8_ptr = ctx_arg
+
+        # GEP with explicit type - this is the key fix
+        field_ptr = builder.gep(
+            ctx_i8_ptr,
+            [offset],
+            inbounds=False,
+        )
+
+        # Get or declare the BPF passthrough intrinsic
+        module = builder.function.module
+
+        try:
+            passthrough_fn = module.globals.get('llvm.bpf.passthrough.p0.p0')
+            if passthrough_fn is None:
+                raise KeyError
+        except (KeyError, AttributeError):
+            passthrough_type = ir.FunctionType(
+                i8_ptr_type,
+                [ir.IntType(32), i8_ptr_type]
+            )
+            passthrough_fn = ir.Function(
+                module,
+                passthrough_type,
+                name='llvm.bpf.passthrough.p0.p0'
+            )
+
+        # Call passthrough to satisfy BPF verifier
+        verified_ptr = builder.call(
+            passthrough_fn,
+            [ir.Constant(ir.IntType(32), 0), field_ptr],
+        )
+
+        # Bitcast to i64* (assuming field is 64-bit, adjust if needed)
+        i64_ptr_type = ir.PointerType(ir.IntType(64))
+        typed_ptr = builder.bitcast(verified_ptr, i64_ptr_type)
+
+        # Load and return the value
+        value = builder.load(typed_ptr)
+
+        return value
+
 
     def has_field(self, struct_name, field_name):
         """Check if a vmlinux struct has a specific field"""

tests/c-form/struct_field_tests.bpf.c

@@ -19,8 +19,8 @@ There is no point of
 SEC("tp/syscalls/sys_enter_execve")
 int handle_setuid_entry(struct trace_event_raw_sys_enter *ctx) {
   // Access each argument separately with clear variable assignments
-  long int arg0 = ctx->id;
-  bpf_printk("args[0]: %d", arg0);
+  long int id = ctx->id;
+  bpf_printk("This is context field %d", id);
   /*
    * the IR to aim for is
    * %2 = alloca ptr, align 8

tests/failing_tests/vmlinux/struct_field_access.py

@@ -14,11 +14,9 @@
 @bpf
 @section("tracepoint/syscalls/sys_enter_execve")
 def hello_world(ctx: struct_trace_event_raw_sys_enter) -> c_int64:
-    a = 2 + TASK_COMM_LEN + TASK_COMM_LEN
     b = ctx.id
-    print(f"Hello, World{TASK_COMM_LEN} and {a}")
     print(f"This is context field {b}")
-    return c_int64(TASK_COMM_LEN + 2)
+    return c_int64(0)
 
 
 @bpf
@@ -28,4 +26,4 @@ def LICENSE() -> str:
 
 
 compile_to_ir("struct_field_access.py", "struct_field_access.ll", loglevel=logging.INFO)
-# compile()
+compile()