[REFACTOR][NODE] Migrate ReprPrinter to tvm-ffi __ffi_repr__ mechanism by tqchen · Pull Request #19461 · apache/tvm

tqchen · 2026-04-27T16:07:58Z

This PR phases out the legacy ReprPrinter machinery (include/tvm/node/repr_printer.h, src/node/repr_printer.cc, src/node/container_printing.cc) by migrating ~169 dispatch sites to the tvm-ffi __ffi_repr__ mechanism.

Summary of changes:

All ~169 TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable).set_dispatch<T> sites have been migrated: most non-user-facing types now use the dataclass auto-default; user-facing types (script/printer/, Trace, DPL pattern syntax, Span/SequentialSpan/SourceName, data_layout Layout/BijectiveLayout, target_kind/virtual_device) retain custom hooks via __ffi_repr__.
Adds include/tvm/node/repr.h with operator<<(ostream, ObjectRef/Any/Variant) delegating to ffi::ReprPrint. Adds src/node/repr.cc with Dump() debug helpers and the node.AsRepr FFI for Python compatibility.
Note: null ObjectRef now reprs as None (Python-compatible) rather than (nullptr); one test was updated for this (test_tir_transform_vectorize.py::test_illegal_extent). The auto-default uses the full type-key (e.g. arith.ModularSet(...)) per design direction.

Test results:

all-platform-minimal: 75p/77s
tirx-base: 273p/2s
arith: 982p/1s/8xf (preexisting xfails)
ir+target: 236p/7s
tvmscript: 771p/1xf
tirx-transform+analysis: 374p/9xf/1xp
s_tir (excl. 2 broken test files): 1279p/36s/2xf + 15 preexisting CUDA sketch failures
relax: covered (no repr-pinned tests)
Pre-commit clean

Replace all ~169 TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable).set_dispatch<T> dispatch sites with the tvm-ffi reflection-based __ffi_repr__ mechanism. Key changes: - Add include/tvm/node/repr.h: ostream operator<< for ObjectRef/Any/Variant delegating to ffi::ReprPrint; AccessStep/AccessPath ostream operators - Add src/node/repr.cc: Dump() helpers, node.AsRepr FFI, AccessPath kRepr - Strip repr_printer.h to a compatibility shim (forward-declares only) - Pattern A (auto-default): remove dispatch, let ffi::ReprPrint generate dataclass-style TypeKey(field=val, ...) repr for ~140 internal types - Pattern B (custom kRepr): register TypeAttrDef<T>.def(kRepr, lambda) for user-facing types: Span, Layout, VirtualDevice, TargetKind, TensorInfo, s_tir Trace - Pattern C (TVM_SCRIPT_REPR): use existing macro for IR/TIR/Relax nodes that delegate to TVMScriptPrinter - Fix double-registration bugs introduced during migration (trace.cc, expr.cc, tirx/ir/expr.cc) - Add explicit refl::ObjectDef<T>() calls for CostModelNode/PyCostModelNode and PyMutatorNode to preserve Python @register_object type index allocation that was implicitly provided by the removed set_dispatch side effect - Update test_tir_transform_vectorize.py: null Target now repr as "None" instead of "(nullptr)" per the new Python-compatible repr convention

gemini-code-assist

Code Review

This pull request migrates the TVM node representation system from the legacy ReprPrinter to a new FFI-based mechanism using ffi::ReprPrint. The changes involve removing numerous static functor dispatches across the codebase and replacing them with kRepr attribute definitions within FFI initialization blocks. Key additions include the tvm/node/repr.h header and src/node/repr.cc implementation, while the legacy repr_printer files are deprecated. Feedback focuses on improving the robustness and consistency of the new implementation, specifically addressing a potential null pointer dereference in access path formatting, removing an unused header, and ensuring the consistent use of the fn_repr callback for recursive printing to maintain architectural integrity.

gemini-code-assist · 2026-04-27T16:13:22Z

+ffi::String FormatAccessPath(const ffi::reflection::AccessPath& path) {
+  std::vector<ffi::reflection::AccessStep> steps;
+  const ffi::reflection::AccessPathObj* cur = path.get();
+  while (cur->step.defined()) {


Potential null pointer dereference. If cur->parent is not defined (which happens at the root of the access path), cur will become null in the next iteration, and cur->step.defined() will crash. You should check if cur is valid before accessing its members.

Suggested change

while (cur->step.defined()) {

while (cur && cur->step.defined()) {

This concern is now moot. The refactor in e9a5055 deletes FormatAccessPath entirely — there is no more manual linked-list walk and no cur = cur->parent.get() cast. AccessPathObj's kRepr hook is now recursive: it returns "" at the root and otherwise delegates parent repr and step repr through fn_repr. Static analysis has nothing to flag.

gemini-code-assist · 2026-04-27T16:13:23Z

+#include <sstream>
+#include <vector>
+
+#include "../support/str_escape.h"


The include ../support/str_escape.h appears to be unused in this file. It was likely carried over from the legacy repr_printer.cc but is no longer needed as ffi::ReprPrint handles string escaping internally.

Confirmed and fixed in commit 1aba295. The orphan include has been removed.

gemini-code-assist · 2026-04-27T16:13:23Z

+// REPR_LAMBDA bodies compile without modification.
+struct PatternReprPrinterHelper {
+  std::ostringstream stream;
+  void Print(const ObjectRef& x) { stream << ffi::ReprPrint(ffi::Any(x)); }


For recursive printing within the __ffi_repr__ mechanism, it is recommended to use the provided fn_repr function rather than calling ffi::ReprPrint directly. This ensures that any context or state maintained by the printer (such as cycle detection or indentation, if implemented) is correctly propagated through the recursion.

void Print(const ObjectRef& x) { stream << fn_repr(ffi::AnyView(x)).cast<ffi::String>(); } ffi::Function fn_repr; PatternReprPrinterHelper(ffi::Function fn_repr) : fn_repr(fn_repr) {}

Good catch — we had previously dismissed this as equivalent, but the user has confirmed fn_repr should be used. Fixed in ff016a8: PatternReprPrinterHelper now stores the fn_repr callback and its Print method routes through it (fn_repr(AnyView(x)).cast()) rather than calling ffi::ReprPrint directly. The RELAX_PATTERN_PRINTER_DEF macro now names fn_repr and passes it to the helper constructor.

gemini-code-assist · 2026-04-27T16:13:23Z

+                                            os << "TensorInfo(\"" << ti->dtype << "\", [";
+                                            bool first = true;
+                                            for (int64_t v : ti->shape) {
+                                              if (!first) os << ", ";
+                                              os << v;
+                                              first = false;
+                                            }
+                                            os << "])";
+                                            return os.str();


Instead of manually iterating over the shape array, you can leverage the provided fn_repr function. Since ti->shape is an ffi::Array, the repr system already knows how to format it consistently.

Suggested change

os << "TensorInfo(\"" << ti->dtype << "\", [";

bool first = true;

for (int64_t v : ti->shape) {

if (!first) os << ", ";

os << v;

first = false;

}

os << "])";

return os.str();

refl::TypeAttrDef<TensorInfoNode>().def(refl::type_attr::kRepr,

[](TensorInfo ti, ffi::Function fn_repr) -> ffi::String {

std::ostringstream os;

os << "TensorInfo(\"" << ti->dtype << "\", "

<< fn_repr(ffi::AnyView(ti->shape)).cast<ffi::String>() << ")";

return os.str();

});

Prior dismissal reversed per user direction — fn_repr should be threaded. Fixed in ff016a8: the lambda now names its fn_repr parameter. The shape elements are raw int64_t from ffi::Shape iteration, so there is no ObjectRef sub-element dispatch to route through fn_repr; the manual loop format [v1, v2, ...] is preserved as intended (fn_repr(ti->shape) would produce Shape(...) format, which the user explicitly rejected).

gemini-code-assist · 2026-04-27T16:13:23Z

+TVM_FFI_STATIC_INIT_BLOCK() {
+  namespace refl = tvm::ffi::reflection;
+  refl::TypeAttrDef<VirtualDeviceNode>().def(
+      refl::type_attr::kRepr, [](VirtualDevice vd, ffi::Function) -> ffi::String {


The fn_repr argument is ignored here. While node->target->str() works for Target, it's better to use the provided fn_repr for consistency with the new printing architecture.

Suggested change

refl::type_attr::kRepr, [](VirtualDevice vd, ffi::Function) -> ffi::String {

refl::type_attr::kRepr, [](VirtualDevice vd, ffi::Function fn_repr) -> ffi::String {

Prior dismissal reversed per user direction — fn_repr should be used for consistency with the kRepr design. Fixed in ff016a8: the lambda now names fn_repr and uses fn_repr(AnyView(node->target)).castffi::String() instead of node->target->str(). This produces the Target(...) dataclass repr rather than the raw target-string format, which is consistent with how other ObjectRef sub-fields are rendered via fn_repr.

The include was carried over from the legacy repr_printer.cc but StrEscape is not referenced in the new repr.cc.

Manual linked-list traversal in FormatAccessPath duplicated work that ffi::ReprPrint's __ffi_repr__ dispatch already does for free; the manual walk also invited a null-deref concern from static analysis. Recurse through parent and step via the kRepr hook; delete the manual walker. AccessStep's formatting is now inline in its own kRepr lambda, using the fn_repr argument for map-key sub-repr as intended.

## Summary `include/tvm/node/` is a leftover separation from when "node" was a distinct concept from "ir". Today everything in `include/tvm/node/` is just lower-level IR plumbing routinely included from `ir/`. This PR collapses the two by moving surviving headers into `ir/`, deleting dead ones, and redirecting the rest to tvm-ffi where the machinery already lives. Main changes: - Migrate `include/tvm/node/{functor,cast,script_printer,attr_registry_map,repr}.h` → `include/tvm/ir/` (functor renamed to `node_functor.h` to preserve type-name connection) - Delete `repr_printer.h`, `structural_equal.h`, `structural_hash.h` (post-#19461 shims and forwarding stubs; redirect 3 cpptest consumers to `tvm/ffi/extra/structural_equal.h`) - Remove inline `AccessStep`/`AccessPath` `operator<<` definitions (existing `__ffi_repr__` registrations already cover these; migrate to generic ObjectRef streaming via kRepr) - Move `src/node/{repr,script_printer}.cc` → `src/ir/` - Delete `src/node/` and empty `include/tvm/node/` directories entirely (no shims) - Update 35 includers across IR, relax, script, target, tirx, and tests

… favor of default representation (#19709) Python-side meta-schedule classes (`PyCostModel`, `PyFeatureExtractor`, `PyMeasureCallback`, `PyScheduleRule`, `PyMutator`, `PyPostproc`) carried an `f_as_string` callback whose only purpose was to produce a repr-style string (`s_tir.meta_schedule.<SubclassName>(0x...)`) for `str(...)`. This mechanism stopped working after #19461 migrated `ReprPrinter` to the tvm-ffi `__ffi_repr__` mechanism and intentionally removed the per-type `set_dispatch<Py*Node>` hooks that called back into `f_as_string`, which broke three `*_as_string` tests: - `test_meta_schedule_cost_model.py::test_meta_schedule_cost_model_as_string` - `test_meta_schedule_feature_extractor.py::test_meta_schedule_feature_extractor_as_string` - `test_meta_schedule_measure_callback.py::test_meta_schedule_measure_callback_as_string` Rather than restoring the old behavior, this PR removes the mechanism entirely: the string it produced is just a repr, and tvm-ffi reflection already provides an auto-generated default repr for every object. Keeping a dedicated Python → FFI → C++ callback chain alive only to reproduce that is not worth the complexity.

… favor of default representation (apache#19709) Python-side meta-schedule classes (`PyCostModel`, `PyFeatureExtractor`, `PyMeasureCallback`, `PyScheduleRule`, `PyMutator`, `PyPostproc`) carried an `f_as_string` callback whose only purpose was to produce a repr-style string (`s_tir.meta_schedule.<SubclassName>(0x...)`) for `str(...)`. This mechanism stopped working after apache#19461 migrated `ReprPrinter` to the tvm-ffi `__ffi_repr__` mechanism and intentionally removed the per-type `set_dispatch<Py*Node>` hooks that called back into `f_as_string`, which broke three `*_as_string` tests: - `test_meta_schedule_cost_model.py::test_meta_schedule_cost_model_as_string` - `test_meta_schedule_feature_extractor.py::test_meta_schedule_feature_extractor_as_string` - `test_meta_schedule_measure_callback.py::test_meta_schedule_measure_callback_as_string` Rather than restoring the old behavior, this PR removes the mechanism entirely: the string it produced is just a repr, and tvm-ffi reflection already provides an auto-generated default repr for every object. Keeping a dedicated Python → FFI → C++ callback chain alive only to reproduce that is not worth the complexity. (cherry picked from commit 219f1d8)

tqchen added 3 commits April 27, 2026 16:03

[REFACTOR] Apply clang-format to repr.cc and script_printer.cc

032908b

[REFACTOR] Apply clang-format to remaining changed files

c14dbf4

gemini-code-assist Bot reviewed Apr 27, 2026

View reviewed changes

tqchen added 2 commits April 27, 2026 16:54

[REFACTOR][NODE] Remove orphan str_escape.h include from repr.cc

1aba295

The include was carried over from the legacy repr_printer.cc but StrEscape is not referenced in the new repr.cc.

tlopex approved these changes Apr 27, 2026

View reviewed changes

tlopex merged commit 200f4fd into apache:main Apr 27, 2026
10 checks passed

This was referenced Apr 27, 2026

[REFACTOR][NODE] Use fn_repr inside kRepr lambdas, not ffi::ReprPrint #19462

Merged

[REFACTOR][IR] Migrate include/tvm/node into include/tvm/ir #19463

Merged

petersktang mentioned this pull request May 1, 2026

[Bug] undefined symbol _ZN3tvm3ffi9ReprPrintERKNS0_3AnyE in libtvm_compiler.so #19486

Closed

ysh329 mentioned this pull request May 6, 2026

[Release] v0.24.0 Release Candidate Notes #19513

Closed

tlopex mentioned this pull request Jun 9, 2026

[Refactor][Meta-schedule] Remove meta-schedule as_string mechanism in favor of default representation #19709

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[REFACTOR][NODE] Migrate ReprPrinter to tvm-ffi __ffi_repr__ mechanism#19461

[REFACTOR][NODE] Migrate ReprPrinter to tvm-ffi __ffi_repr__ mechanism#19461
tlopex merged 5 commits into
apache:mainfrom
tqchen:tvm-migrate-to-tvm-ffi-repr-mechanism

tqchen commented Apr 27, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Uh oh!

tqchen Apr 27, 2026

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Uh oh!

tqchen Apr 27, 2026

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Uh oh!

tqchen Apr 27, 2026

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Uh oh!

tqchen Apr 27, 2026

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Uh oh!

tqchen Apr 27, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

	while (cur->step.defined()) {
	while (cur && cur->step.defined()) {

	refl::type_attr::kRepr, [](VirtualDevice vd, ffi::Function) -> ffi::String {
	refl::type_attr::kRepr, [](VirtualDevice vd, ffi::Function fn_repr) -> ffi::String {

Uh oh!

Conversation

tqchen commented Apr 27, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

tqchen Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

tqchen Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

tqchen Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

tqchen Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

tqchen Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants