[amd-rmotiram]

mlir-imp: fix import of If bodies whose outputs reuse the If's output names

Summary

The 12 whisper "merged" decoders (whisper-base and whisper-large-v3-turbo decoder_model_merged{,_bnb4,_int8,_q4,_quantized,_uint8}) fail during the ONNX -> MLIR import:

register_captured_alias: null Value for 'present.0.decoder.key'

Their If branches return values named the same as the If's own outputs. ONNX allows this: inside a branch, the branch's value shadows the outer name. The importer mishandled it in two places in morphizen/mlir-imp/src/mlir-graph.cpp.

Problem

flowchart TD
    A["Import If: output names declared<br/>(placeholders, no Value yet)"] --> B["Import branch bodies"]
    B --> C["Branch looks up present.0.decoder.key"]
    C --> D["Not local, found in parent<br/>as an empty placeholder"]
    D --> E["Treated as a capture<br/>register_captured_alias(name, null)"]
    E --> F["CHECK fails: null Value"]
    style F fill:#f8d7da,stroke:#c00

    G["Build main-graph name table<br/>(post-order: children first)"] --> H["Branch value claims the name"]
    H --> I["If output dropped (emplace keeps first)"]
    I --> J["Partitioning fails<br/>morphizen-ep.cpp:894 / :461"]
    style J fill:#f8d7da,stroke:#c00

Loading

The second failure only shows up once the first is fixed.

Solution

flowchart TD
    C["Branch looks up a name"] --> L{"Local?"}
    L -- Yes --> U["Use local value"]
    L -- No --> P{"In parent AND<br/>has a Value?"}
    P -- Yes --> R["Real capture"]
    P -- No --> S["Not found: branch creates its own<br/>value (shadows outer name)"]
    style R fill:#d4edda,stroke:#080
    style S fill:#d4edda,stroke:#080

    G["Build main-graph name table<br/>(pre-order: parent first)"] --> H["If claims its output names first"]
    H --> I["Same-named branch values don't override"]
    style I fill:#d4edda,stroke:#080

Loading

• get_node_arg_index: an outer name only counts as a capture if it already has a Value. Genuine captures always do, because outer nodes are imported in topological order.
• initialize_node_args_map: walk pre-order, matching populate_node_arg_indexes.
• add_node: log a warning when a named input has no producer and is replaced by none. Before this change, that case asserted.

Testing (Strix Halo, gfx1151)

• All 12 merged decoders now pass import. They then stop at a separate pipeline error: 'onnx.If' op failed to build output inits (fully-static output shapes required).
• New test/numeric/tests/test_if_subgraph_names.py (then and else branches) and the existing test_dangling_outputs.py pass: 6/6. The new test fails without this fix.
• Smoke: resnet50_fp16, mobilenetv3_fp32, mosaic_fp16 and whisper-base encoder are unchanged. whisper-base decoder_model.onnx crashes (0xC0000005) both with and without this change.

Known gaps

• No end-to-end numeric check on the 12 real models yet, because the onnx.If static-shape error above blocks them.
• Tests cover If only. Loop/Scan/nested-If shadowing uses the same code path but isn't directly tested, and there is no lit test.
• A malformed graph referencing a never-produced outer value no longer asserts. It now gets the add_node warning and a none input. ORT's load-time validation normally rejects such models first.

[github-actions]

Thanks for opening a PR!

This project follows LLVM's incremental-development and AI-tool-use
guidance. See CONTRIBUTING.md
for the project workflow.

Before requesting review, please check that:

1. The change is focused. Substantial work links the relevant issue
   or design discussion.
2. The PR documents relevant test results and updates affected
   documentation.
3. If AI tools provided substantial assistance, the description
   explains what was assisted and how it was validated, and commit
   trailers identify the tool. The contributor has reviewed and
   understands the result.

Reviewers are assigned through
CODEOWNERS where ownership
is configured.