-
Notifications
You must be signed in to change notification settings - Fork 14.9k
[MLIR][Python] remove liveOperations
#155114
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Conversation
8c71176
to
2a0e205
Compare
liveOpeartions
@llvm/pr-subscribers-mlir Author: Maksim Levental (makslevental) ChangesInspired by a renewed interest in #139721 (which has become a little stale...) <p align="center"> In the previous go-around (#92631) there were two issues which have been resolved
The other potentially breaking change is cc @0xMihir @hawkinsp @superbobry Patch is 22.39 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/155114.diff 10 Files Affected:
diff --git a/mlir/include/mlir-c/IR.h b/mlir/include/mlir-c/IR.h
index 71c7d4378677f..d20b46651383f 100644
--- a/mlir/include/mlir-c/IR.h
+++ b/mlir/include/mlir-c/IR.h
@@ -415,6 +415,8 @@ MLIR_CAPI_EXPORTED MlirOperation mlirModuleGetOperation(MlirModule module);
/// The returned module is null when the input operation was not a ModuleOp.
MLIR_CAPI_EXPORTED MlirModule mlirModuleFromOperation(MlirOperation op);
+MLIR_CAPI_EXPORTED bool mlirModuleEqual(MlirModule mod, MlirModule other);
+
//===----------------------------------------------------------------------===//
// Operation state.
//===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Bindings/Python/IRCore.cpp b/mlir/lib/Bindings/Python/IRCore.cpp
index 4b3a06cbce854..23c4ec7c7f944 100644
--- a/mlir/lib/Bindings/Python/IRCore.cpp
+++ b/mlir/lib/Bindings/Python/IRCore.cpp
@@ -702,84 +702,6 @@ size_t PyMlirContext::getLiveCount() {
return getLiveContexts().size();
}
-size_t PyMlirContext::getLiveOperationCount() {
- nb::ft_lock_guard lock(liveOperationsMutex);
- return liveOperations.size();
-}
-
-std::vector<PyOperation *> PyMlirContext::getLiveOperationObjects() {
- std::vector<PyOperation *> liveObjects;
- nb::ft_lock_guard lock(liveOperationsMutex);
- for (auto &entry : liveOperations)
- liveObjects.push_back(entry.second.second);
- return liveObjects;
-}
-
-size_t PyMlirContext::clearLiveOperations() {
-
- LiveOperationMap operations;
- {
- nb::ft_lock_guard lock(liveOperationsMutex);
- std::swap(operations, liveOperations);
- }
- for (auto &op : operations)
- op.second.second->setInvalid();
- size_t numInvalidated = operations.size();
- return numInvalidated;
-}
-
-void PyMlirContext::clearOperation(MlirOperation op) {
- PyOperation *py_op;
- {
- nb::ft_lock_guard lock(liveOperationsMutex);
- auto it = liveOperations.find(op.ptr);
- if (it == liveOperations.end()) {
- return;
- }
- py_op = it->second.second;
- liveOperations.erase(it);
- }
- py_op->setInvalid();
-}
-
-void PyMlirContext::clearOperationsInside(PyOperationBase &op) {
- typedef struct {
- PyOperation &rootOp;
- bool rootSeen;
- } callBackData;
- callBackData data{op.getOperation(), false};
- // Mark all ops below the op that the passmanager will be rooted
- // at (but not op itself - note the preorder) as invalid.
- MlirOperationWalkCallback invalidatingCallback = [](MlirOperation op,
- void *userData) {
- callBackData *data = static_cast<callBackData *>(userData);
- if (LLVM_LIKELY(data->rootSeen))
- data->rootOp.getOperation().getContext()->clearOperation(op);
- else
- data->rootSeen = true;
- return MlirWalkResult::MlirWalkResultAdvance;
- };
- mlirOperationWalk(op.getOperation(), invalidatingCallback,
- static_cast<void *>(&data), MlirWalkPreOrder);
-}
-void PyMlirContext::clearOperationsInside(MlirOperation op) {
- PyOperationRef opRef = PyOperation::forOperation(getRef(), op);
- clearOperationsInside(opRef->getOperation());
-}
-
-void PyMlirContext::clearOperationAndInside(PyOperationBase &op) {
- MlirOperationWalkCallback invalidatingCallback = [](MlirOperation op,
- void *userData) {
- PyMlirContextRef &contextRef = *static_cast<PyMlirContextRef *>(userData);
- contextRef->clearOperation(op);
- return MlirWalkResult::MlirWalkResultAdvance;
- };
- mlirOperationWalk(op.getOperation(), invalidatingCallback,
- &op.getOperation().getContext(), MlirWalkPreOrder);
-}
-
-size_t PyMlirContext::getLiveModuleCount() { return liveModules.size(); }
-
nb::object PyMlirContext::contextEnter(nb::object context) {
return PyThreadContextEntry::pushContext(context);
}
@@ -1151,38 +1073,20 @@ PyLocation &DefaultingPyLocation::resolve() {
PyModule::PyModule(PyMlirContextRef contextRef, MlirModule module)
: BaseContextObject(std::move(contextRef)), module(module) {}
-PyModule::~PyModule() {
- nb::gil_scoped_acquire acquire;
- auto &liveModules = getContext()->liveModules;
- assert(liveModules.count(module.ptr) == 1 &&
- "destroying module not in live map");
- liveModules.erase(module.ptr);
- mlirModuleDestroy(module);
-}
+PyModule::~PyModule() { mlirModuleDestroy(module); }
PyModuleRef PyModule::forModule(MlirModule module) {
MlirContext context = mlirModuleGetContext(module);
PyMlirContextRef contextRef = PyMlirContext::forContext(context);
- nb::gil_scoped_acquire acquire;
- auto &liveModules = contextRef->liveModules;
- auto it = liveModules.find(module.ptr);
- if (it == liveModules.end()) {
- // Create.
- PyModule *unownedModule = new PyModule(std::move(contextRef), module);
- // Note that the default return value policy on cast is automatic_reference,
- // which does not take ownership (delete will not be called).
- // Just be explicit.
- nb::object pyRef = nb::cast(unownedModule, nb::rv_policy::take_ownership);
- unownedModule->handle = pyRef;
- liveModules[module.ptr] =
- std::make_pair(unownedModule->handle, unownedModule);
- return PyModuleRef(unownedModule, std::move(pyRef));
- }
- // Use existing.
- PyModule *existing = it->second.second;
- nb::object pyRef = nb::borrow<nb::object>(it->second.first);
- return PyModuleRef(existing, std::move(pyRef));
+ // Create.
+ PyModule *unownedModule = new PyModule(std::move(contextRef), module);
+ // Note that the default return value policy on cast is automatic_reference,
+ // which does not take ownership (delete will not be called).
+ // Just be explicit.
+ nb::object pyRef = nb::cast(unownedModule, nb::rv_policy::take_ownership);
+ unownedModule->handle = pyRef;
+ return PyModuleRef(unownedModule, std::move(pyRef));
}
nb::object PyModule::createFromCapsule(nb::object capsule) {
@@ -1207,16 +1111,8 @@ PyOperation::~PyOperation() {
// If the operation has already been invalidated there is nothing to do.
if (!valid)
return;
-
- // Otherwise, invalidate the operation and remove it from live map when it is
- // attached.
- if (isAttached()) {
- getContext()->clearOperation(*this);
- } else {
- // And destroy it when it is detached, i.e. owned by Python, in which case
- // all nested operations must be invalidated at removed from the live map as
- // well.
- erase();
+ if (!isAttached()) {
+ mlirOperationDestroy(operation);
}
}
@@ -1252,35 +1148,16 @@ PyOperationRef PyOperation::createInstance(PyMlirContextRef contextRef,
PyOperationRef PyOperation::forOperation(PyMlirContextRef contextRef,
MlirOperation operation,
nb::object parentKeepAlive) {
- nb::ft_lock_guard lock(contextRef->liveOperationsMutex);
- auto &liveOperations = contextRef->liveOperations;
- auto it = liveOperations.find(operation.ptr);
- if (it == liveOperations.end()) {
- // Create.
- PyOperationRef result = createInstance(std::move(contextRef), operation,
- std::move(parentKeepAlive));
- liveOperations[operation.ptr] =
- std::make_pair(result.getObject(), result.get());
- return result;
- }
- // Use existing.
- PyOperation *existing = it->second.second;
- nb::object pyRef = nb::borrow<nb::object>(it->second.first);
- return PyOperationRef(existing, std::move(pyRef));
+ // Create.
+ return createInstance(std::move(contextRef), operation,
+ std::move(parentKeepAlive));
}
PyOperationRef PyOperation::createDetached(PyMlirContextRef contextRef,
MlirOperation operation,
nb::object parentKeepAlive) {
- nb::ft_lock_guard lock(contextRef->liveOperationsMutex);
- auto &liveOperations = contextRef->liveOperations;
- assert(liveOperations.count(operation.ptr) == 0 &&
- "cannot create detached operation that already exists");
- (void)liveOperations;
PyOperationRef created = createInstance(std::move(contextRef), operation,
std::move(parentKeepAlive));
- liveOperations[operation.ptr] =
- std::make_pair(created.getObject(), created.get());
created->attached = false;
return created;
}
@@ -1652,7 +1529,6 @@ nb::object PyOperation::createOpView() {
void PyOperation::erase() {
checkValid();
- getContext()->clearOperationAndInside(*this);
mlirOperationDestroy(operation);
}
@@ -3023,14 +2899,6 @@ void mlir::python::populateIRCore(nb::module_ &m) {
PyMlirContextRef ref = PyMlirContext::forContext(self.get());
return ref.releaseObject();
})
- .def("_get_live_operation_count", &PyMlirContext::getLiveOperationCount)
- .def("_get_live_operation_objects",
- &PyMlirContext::getLiveOperationObjects)
- .def("_clear_live_operations", &PyMlirContext::clearLiveOperations)
- .def("_clear_live_operations_inside",
- nb::overload_cast<MlirOperation>(
- &PyMlirContext::clearOperationsInside))
- .def("_get_live_module_count", &PyMlirContext::getLiveModuleCount)
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyMlirContext::getCapsule)
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyMlirContext::createFromCapsule)
.def("__enter__", &PyMlirContext::contextEnter)
@@ -3349,6 +3217,7 @@ void mlir::python::populateIRCore(nb::module_ &m) {
nb::class_<PyModule>(m, "Module", nb::is_weak_referenceable())
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyModule::getCapsule)
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyModule::createFromCapsule)
+ .def("_clear_referrent", &PyModule::clearReferrent)
.def_static(
"parse",
[](const std::string &moduleAsm, DefaultingPyMlirContext context) {
@@ -3428,7 +3297,13 @@ void mlir::python::populateIRCore(nb::module_ &m) {
// Defer to the operation's __str__.
return self.attr("operation").attr("__str__")();
},
- kOperationStrDunderDocstring);
+ kOperationStrDunderDocstring)
+ .def(
+ "__eq__",
+ [](PyModule &self, PyModule &other) {
+ return mlirModuleEqual(self.get(), other.get());
+ },
+ "other"_a);
//----------------------------------------------------------------------------
// Mapping of Operation.
@@ -3440,7 +3315,8 @@ void mlir::python::populateIRCore(nb::module_ &m) {
})
.def("__eq__",
[](PyOperationBase &self, PyOperationBase &other) {
- return &self.getOperation() == &other.getOperation();
+ return mlirOperationEqual(self.getOperation().get(),
+ other.getOperation().get());
})
.def("__eq__",
[](PyOperationBase &self, nb::object other) { return false; })
diff --git a/mlir/lib/Bindings/Python/IRModule.h b/mlir/lib/Bindings/Python/IRModule.h
index fa16ae3ce3294..e1a71c627a902 100644
--- a/mlir/lib/Bindings/Python/IRModule.h
+++ b/mlir/lib/Bindings/Python/IRModule.h
@@ -218,40 +218,6 @@ class PyMlirContext {
/// Gets the count of live context objects. Used for testing.
static size_t getLiveCount();
- /// Get a list of Python objects which are still in the live context map.
- std::vector<PyOperation *> getLiveOperationObjects();
-
- /// Gets the count of live operations associated with this context.
- /// Used for testing.
- size_t getLiveOperationCount();
-
- /// Clears the live operations map, returning the number of entries which were
- /// invalidated. To be used as a safety mechanism so that API end-users can't
- /// corrupt by holding references they shouldn't have accessed in the first
- /// place.
- size_t clearLiveOperations();
-
- /// Removes an operation from the live operations map and sets it invalid.
- /// This is useful for when some non-bindings code destroys the operation and
- /// the bindings need to made aware. For example, in the case when pass
- /// manager is run.
- ///
- /// Note that this does *NOT* clear the nested operations.
- void clearOperation(MlirOperation op);
-
- /// Clears all operations nested inside the given op using
- /// `clearOperation(MlirOperation)`.
- void clearOperationsInside(PyOperationBase &op);
- void clearOperationsInside(MlirOperation op);
-
- /// Clears the operaiton _and_ all operations inside using
- /// `clearOperation(MlirOperation)`.
- void clearOperationAndInside(PyOperationBase &op);
-
- /// Gets the count of live modules associated with this context.
- /// Used for testing.
- size_t getLiveModuleCount();
-
/// Enter and exit the context manager.
static nanobind::object contextEnter(nanobind::object context);
void contextExit(const nanobind::object &excType,
@@ -278,25 +244,6 @@ class PyMlirContext {
static nanobind::ft_mutex live_contexts_mutex;
static LiveContextMap &getLiveContexts();
- // Interns all live modules associated with this context. Modules tracked
- // in this map are valid. When a module is invalidated, it is removed
- // from this map, and while it still exists as an instance, any
- // attempt to access it will raise an error.
- using LiveModuleMap =
- llvm::DenseMap<const void *, std::pair<nanobind::handle, PyModule *>>;
- LiveModuleMap liveModules;
-
- // Interns all live operations associated with this context. Operations
- // tracked in this map are valid. When an operation is invalidated, it is
- // removed from this map, and while it still exists as an instance, any
- // attempt to access it will raise an error.
- using LiveOperationMap =
- llvm::DenseMap<void *, std::pair<nanobind::handle, PyOperation *>>;
- nanobind::ft_mutex liveOperationsMutex;
-
- // Guarded by liveOperationsMutex in free-threading mode.
- LiveOperationMap liveOperations;
-
bool emitErrorDiagnostics = false;
MlirContext context;
@@ -575,6 +522,8 @@ class PyModule : public BaseContextObject {
/// is taken by calling this function.
static nanobind::object createFromCapsule(nanobind::object capsule);
+ void clearReferrent() { module = {nullptr}; }
+
private:
PyModule(PyMlirContextRef contextRef, MlirModule module);
MlirModule module;
diff --git a/mlir/lib/Bindings/Python/Pass.cpp b/mlir/lib/Bindings/Python/Pass.cpp
index 20017e25b69bb..817479ee2421b 100644
--- a/mlir/lib/Bindings/Python/Pass.cpp
+++ b/mlir/lib/Bindings/Python/Pass.cpp
@@ -159,11 +159,7 @@ void mlir::python::populatePassManagerSubmodule(nb::module_ &m) {
"ValueError if the pipeline can't be parsed.")
.def(
"run",
- [](PyPassManager &passManager, PyOperationBase &op,
- bool invalidateOps) {
- if (invalidateOps) {
- op.getOperation().getContext()->clearOperationsInside(op);
- }
+ [](PyPassManager &passManager, PyOperationBase &op) {
// Actually run the pass manager.
PyMlirContext::ErrorCapture errors(op.getOperation().getContext());
MlirLogicalResult status = mlirPassManagerRunOnOp(
@@ -172,7 +168,7 @@ void mlir::python::populatePassManagerSubmodule(nb::module_ &m) {
throw MLIRError("Failure while executing pass pipeline",
errors.take());
},
- "operation"_a, "invalidate_ops"_a = true,
+ "operation"_a,
"Run the pass manager on the provided operation, raising an "
"MLIRError on failure.")
.def(
diff --git a/mlir/lib/Bindings/Python/TransformInterpreter.cpp b/mlir/lib/Bindings/Python/TransformInterpreter.cpp
index f9b0fed62778f..920bca886f617 100644
--- a/mlir/lib/Bindings/Python/TransformInterpreter.cpp
+++ b/mlir/lib/Bindings/Python/TransformInterpreter.cpp
@@ -67,7 +67,6 @@ static void populateTransformInterpreterSubmodule(nb::module_ &m) {
// root. This is awkward, but we don't have access to PyMlirContext
// object here otherwise.
nb::object obj = nb::cast(payloadRoot);
- obj.attr("context").attr("_clear_live_operations_inside")(payloadRoot);
MlirLogicalResult result = mlirTransformApplyNamedSequence(
payloadRoot, transformRoot, transformModule, options.options);
diff --git a/mlir/lib/CAPI/IR/IR.cpp b/mlir/lib/CAPI/IR/IR.cpp
index 8491553dab76f..7ea799c836889 100644
--- a/mlir/lib/CAPI/IR/IR.cpp
+++ b/mlir/lib/CAPI/IR/IR.cpp
@@ -465,6 +465,10 @@ MlirModule mlirModuleFromOperation(MlirOperation op) {
return wrap(dyn_cast<ModuleOp>(unwrap(op)));
}
+bool mlirModuleEqual(MlirModule mod, MlirModule other) {
+ return unwrap(mod) == unwrap(other);
+}
+
//===----------------------------------------------------------------------===//
// Operation state API.
//===----------------------------------------------------------------------===//
diff --git a/mlir/test/python/ir/module.py b/mlir/test/python/ir/module.py
index 6065e59fd6ed9..4a9cf9d44a8aa 100644
--- a/mlir/test/python/ir/module.py
+++ b/mlir/test/python/ir/module.py
@@ -121,27 +121,16 @@ def testRoundtripBinary():
def testModuleOperation():
ctx = Context()
module = Module.parse(r"""module @successfulParse {}""", ctx)
- assert ctx._get_live_module_count() == 1
op1 = module.operation
- assert ctx._get_live_operation_count() == 1
- live_ops = ctx._get_live_operation_objects()
- assert len(live_ops) == 1
- assert live_ops[0] is op1
- live_ops = None
# CHECK: module @successfulParse
print(op1)
# Ensure that operations are the same on multiple calls.
op2 = module.operation
- assert ctx._get_live_operation_count() == 1
- assert op1 is op2
+ assert op1 == op2
# Test live operation clearing.
op1 = module.operation
- assert ctx._get_live_operation_count() == 1
- num_invalidated = ctx._clear_live_operations()
- assert num_invalidated == 1
- assert ctx._get_live_operation_count() == 0
op1 = None
gc.collect()
op1 = module.operation
@@ -155,9 +144,6 @@ def testModuleOperation():
op1 = None
op2 = None
gc.collect()
- print("LIVE OPERATIONS:", ctx._get_live_operation_count())
- assert ctx._get_live_operation_count() == 0
- assert ctx._get_live_module_count() == 0
# CHECK-LABEL: TEST: testModuleCapsule
@@ -165,16 +151,14 @@ def testModuleOperation():
def testModuleCapsule():
ctx = Context()
module = Module.parse(r"""module @successfulParse {}""", ctx)
- assert ctx._get_live_module_count() == 1
# CHECK: "mlir.ir.Module._CAPIPtr"
module_capsule = module._CAPIPtr
print(module_capsule)
module_dup = Module._CAPICreate(module_capsule)
- assert module is module_dup
+ module._clear_referrent()
assert module_dup.context is ctx
# Gc and verify destructed.
module = None
module_capsule = None
module_dup = None
gc.collect()
- assert ctx._get_live_module_count() == 0
diff --git a/mlir/test/python/ir/operation.py b/mlir/test/python/ir/operation.py
index bf16e3f75d60d..c95a97b9731a2 100644
--- a/mlir/test/python/ir/operation.py
+++ b/mlir/test/python/ir/operation.py
@@ -907,7 +907,7 @@ def testCapsuleConversions():
m_capsule = m._CAPIPtr
assert '"mlir.ir.Operation._CAPIPtr"' in repr(m_capsule)
m2 = Operation._CAPICreate(m_capsule)
- assert m2 is m
+ assert m2 == m
# CHECK-LABEL: TEST: testOperationErase
diff --git a/mlir/test/python/ir/symbol_table.py b/mlir/test/python/ir/symbol_table.py
index 8b6d7ea5a197d..7afd539271d21 100644
--- a/mlir/test/python/ir/symbol_table.py
+++ b/mlir/test/python/ir/symbol_table.py
@@ -56,14 +56,6 @@ def testSymbolTableInsert():
print(m1)
assert "bar" not in symbol_table
- try:
- print(bar)
- except RuntimeError as e:
- if "the operation has been invalidated" not in str(e):
- raise
- else:
- assert False, "expected RuntimeError due to invalidated operation"
-
qux = m2.body.operations[0]
m1.body.append(qux)
symbol_table.insert(qux)
diff --git a/mlir/test/python/pass_manager.py b/mlir/test/python/pass_manager.py
index e26d42bb32913..0896cd9784641 100644
--- a/mlir/test/python/pass_manager.py
...
[truncated]
|
liveOpeartions
liveOperations
3de0c41
to
b32e0f3
Compare
liveOperations
liveOperations
6041b0c
to
266540a
Compare
I kicked off a manual build on sanitizer-x86_64-linux-bootstrap-asan |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It's been too long ago that I looked at that code, so I can't give meaningful input on the design. Just a small comment on the documentation which might help readers (see inline).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Could we perhaps pull in more of the why of removal into PR description to make it easier for folks who read the change in git log?
I'd not be opposed to removing Module._CAPICreate (Module's special casing in Python here is a bit historical). But if this already addresses the race then SGTM.
✅ With the latest revision this PR passed the C/C++ code formatter. |
2056f23
to
6419117
Compare
6419117
to
9343c1f
Compare
Module, when used purely from the Python API, can't alias anyway, so we can use | ||
it as a top-level ref type without a live-list for interning. If the API ever | ||
changes such that this cannot be guaranteed (i.e. by letting you marshal a | ||
native-defined Module in), then there would need to be a live table for it too. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This isn't even valid/true anymore - we do have (before this PR) a liveModules
live-list and Module._CAPICreate
is exactly for "letting you marshal a native-defined Module in".
9343c1f
to
0f79609
Compare
fdcc2f9
to
465a4b4
Compare
465a4b4
to
685fbb5
Compare
purposes of the discussion here. Metaphorically, one can think of this similarly to how STL container iterators are invalidated once the container itself is changed. The "best practices" recommendation is to structure your code such that | ||
|
||
1. First, query/manipulate various Python wrapper objects `py_op1`, `py_op2`, `py_op3`, etc.; | ||
2. Second, Transform the AST/erase operations/etc. via a single root object; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
s/Transform/transform/
purposes of the discussion here. Metaphorically, one can think of this similarly to how STL container iterators are invalidated once the container itself is changed. The "best practices" recommendation is to structure your code such that | ||
|
||
1. First, query/manipulate various Python wrapper objects `py_op1`, `py_op2`, `py_op3`, etc.; | ||
2. Second, Transform the AST/erase operations/etc. via a single root object; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This is effectively saying, scope your change based on nesting (e.g., change leafs first before going up in hierarchy, and/or only in very rare cases consider nested op post modifying parent).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
sure yea that's a very sophisticated way to put it :) i'll add that
|
||
1. First, query/manipulate various Python wrapper objects `py_op1`, `py_op2`, `py_op3`, etc.; | ||
2. Second, Transform the AST/erase operations/etc. via a single root object; | ||
3. End. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm not sure what End means here. Invalidate/forget all queried nodes?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
yea effectively - colloquially i meant "don't do anything else with the stuff you touched in steps 1 and 2" - but i'll change it to what you said
Historical context:
PyMlirContext::liveOperations
was an optimization meant to cut down on the number of Python object allocations and (partially) a mechanism for updating validity of ops after transformation. E.g. during walking/transforming the AST. See original patch here.Inspired by a renewed interest in #139721 (which has become a little stale...)
In the previous go-around (#92631) there were two issues which have been resolved
ops that were "fetched" under a root op which has been transformed are no longer reported as invalid. We simply "formally forbid" this;
Module._CAPICreate(module_capsule)
must now be followed by amodule._clear_mlir_module()
to prevent double-freeing of the actualModuleOp
object (i.e. calling the dtor on theOwningOpRef<ModuleOp>
):Module._CAPICreate
API altogether and replace it with something likeModule._move(module)
which will do bothModule._CAPICreate
andmodule._clear_mlir_module
.Note, the other approach I explored last year was a weakref system for
mlir::Operation
which would effectively hoist thisliveOperations
thing into MLIR core. Possibly doable but I now believe it's a bad idea.The other potentially breaking change is
is
, which checks object equality rather than value equality, will now reportFalse
because we are always allocatingnew
Python objects (ie that's the whole point of this change). Users wanting to check equality forOperation
andModule
should use==
.cc @0xMihir @hawkinsp @superbobry @christopherbate @chhzh123
EDIT: Sanity check: no one is using Module._CAPICreate