Merge pull request #1 from blepabyte/06-10-fix_descend_logic_and_lru_…

…cache_support

fix: descend logic and lru_cache support

maxray/__init__.py

@@ -2,8 +2,9 @@
 
 import inspect
 from contextvars import ContextVar
+from contextlib import contextmanager
 from dataclasses import dataclass
-from functools import wraps
+from functools import wraps, _lru_cache_wrapper
 from typing import Any, Callable
 from result import Result, Ok, Err
 
@@ -13,6 +14,22 @@
 logger.disable("maxray")
 
 
+def _set_logging(enabled: bool):
+    if enabled:
+        logger.enable("maxray")
+    else:
+        logger.disable("maxray")
+
+
+@contextmanager
+def _with_logging():
+    try:
+        _set_logging(True)
+        yield
+    finally:
+        _set_logging(False)
+
+
 def transform(writer):
     """
     Decorator that rewrites the source code of a function to wrap every single expression by passing it through the `writer(expr, ctx)` callable.
@@ -37,14 +54,19 @@ def xray(walker, **kwargs):
 
 
 _GLOBAL_SKIP_MODULES = {
+    "builtins",  # duh... object.__init__ is a wrapper_descriptor
     "abc",  # excessive inheritance and super calls in scikit-learn
     "inspect",  # don't want to screw this module up
     "pathlib",  # internally used in transform for checking source file exists
     "re",  # internals of regexp have a lot of uninteresting step methods
     "copy",  # pytorch spends so much time in here
+    "functools",  # partialmethod causes mis-bound `self` in TQDM
     "typing",
     "importlib",
+    "ctypes",
     "loguru",  # used internally in transform - accidental patching will cause inf recursion
+    "uuid",  # used for generating _MAXRAY_TRANSFORM_ID
+    "maxray",
 }
 # TODO: probably just skip the entire Python standard library...
 
@@ -65,6 +87,17 @@ class W_erHook:
 
 _MAXRAY_REGISTERED_HOOKS: list[W_erHook] = []
 
+
+def descend_allowed(x, ctx: NodeContext):
+    num_active_hooks = 0
+    for hook in _MAXRAY_REGISTERED_HOOKS:
+        if hook.active_call_state.get():
+            num_active_hooks += 1
+            if not hook.descend_predicate(x, ctx):
+                return False
+    return num_active_hooks > 0
+
+
 _GLOBAL_WRITER_ACTIVE_FLAG = ContextVar("writer_active (global)", default=False)
 
 # We don't want to recompile the same function over and over - so our cache needs to be global
@@ -84,7 +117,11 @@ def callable_allowed_for_transform(x, ctx: NodeContext):
         and callable(x)
         and callable(getattr(x, "__hash__", None))
         and getattr(type(x), "__module__", None) not in {"ctypes"}
-        and (inspect.isfunction(x) or inspect.ismethod(x))
+        and (
+            inspect.isfunction(x)
+            or inspect.ismethod(x)
+            or isinstance(x, _lru_cache_wrapper)
+        )
     )
 
 
@@ -94,6 +131,7 @@ def instance_init_allowed_for_transform(x, ctx: NodeContext):
     """
     return (
         type(x) is type
+        and getattr(x, "__module__", None) not in {"ctypes"}
         and hasattr(x, "__init__")
         and not hasattr(x, "_MAXRAY_TRANSFORMED")
     )
@@ -103,47 +141,47 @@ def instance_call_allowed_for_transform(x, ctx: NodeContext):
     """
     Decides whether the __call__ method can be transformed.
     """
+    if getattr(x, "__module__", None) in _GLOBAL_SKIP_MODULES:
+        return False
+
     return (
         type(x) is type
+        and getattr(x, "__module__", None) not in {"ctypes"}
         and hasattr(x, "__call__")
         and not hasattr(x, "_MAXRAY_TRANSFORMED")
     )
 
 
 def _maxray_walker_handler(x, ctx: NodeContext):
-    # We ignore writer calls triggered by code execution in other writers to prevent easily getting stuck in recursive hell
-    if _GLOBAL_WRITER_ACTIVE_FLAG.get():
-        return x
-
     # 1.  logic to recursively patch callables
     # 1a. special-case callables: __init__ and __call__
     if instance_init_allowed_for_transform(x, ctx):
-        # TODO: should we somehow delay doing this until before an actual call?
         match recompile_fn_with_transform(
             x.__init__, _maxray_walker_handler, special_use_instance_type=x
         ):
             case Ok(init_patch):
-                logger.info(f"Patching __init__ for class {x}")
+                logger.debug(f"Patching __init__ for class {x}")
                 setattr(x, "__init__", init_patch)
+            # TODO: consolidate error handling and reporting
+            # case Err(bad):
+            #     logger.error(bad)
 
     elif instance_call_allowed_for_transform(x, ctx):
-        # TODO: should we somehow delay doing this until before an actual call?
         match recompile_fn_with_transform(
             x.__call__, _maxray_walker_handler, special_use_instance_type=x
         ):
             case Ok(call_patch):
-                logger.info(f"Patching __call__ for class {x}")
+                logger.debug(f"Patching __call__ for class {x}")
                 setattr(x, "__call__", call_patch)
+            # case Err(bad):
+            #     logger.error(bad)
 
     # 1b. normal functions or bound methods or method descriptors like @classmethod and @staticmethod
     elif callable_allowed_for_transform(x, ctx):
         # We can only cache functions - as caching invokes __hash__, which may fail badly on incompletely-initialised class instances w/ __call__ methods, like torch._ops.OpOverload
         if x in _MAXRAY_FN_CACHE:
             x = _MAXRAY_FN_CACHE[x]
-        elif not any(
-            hook.active_call_state.get() and hook.descend_predicate(x, ctx)
-            for hook in _MAXRAY_REGISTERED_HOOKS
-        ):
+        elif not descend_allowed(x, ctx):
             # user-defined filters for which nodes (not) to descend into
             pass
         else:
@@ -185,6 +223,11 @@ def _maxray_walker_handler(x, ctx: NodeContext):
                     # Errors in functions that have been recursively compiled are less important
                     logger.warning(f"Failed to transform in walker handler: {e}")
 
+    # We ignore writer calls triggered by code execution in other writers to prevent easily getting stuck in recursive hell
+    # This happens *after* checking and patching callables to still allow for explicitly patching a callable/method by calling this handler
+    if _GLOBAL_WRITER_ACTIVE_FLAG.get():
+        return x
+
     # 2. run the active hooks
     global_write_active_token = _GLOBAL_WRITER_ACTIVE_FLAG.set(True)
     try:
@@ -288,6 +331,8 @@ def fn_with_context_update(*args, **kwargs):
                     ACTIVE_FLAG.reset(prev_token)
 
         fn_with_context_update._MAXRAY_TRANSFORMED = True
+        # TODO: set correctly everywhere
+        # fn_with_context_update._MAXRAY_TRANSFORM_ID = ...
         return fn_with_context_update
 
     return recursive_transform