Coverity static analysis fixes - May 2026

[jantonguirao]

Summary

Three changes:

• complex_pipeline_test.cc — Coverity-flagged use-after-free in RunCheckpointingTest. Split checkpoint_h into distinct checkpoint1_h / checkpoint2_h locals so there is no apparent reuse of a handle after a RAII transfer. The test keeps using the raw C API for output handles (it is intentionally a C API test).
• pipeline_test_utils.h — initialize daliPipelineOutputs_h raw_out_h = nullptr in the PopOutputs and PopOutputsAsync test helpers. CHECK_DALI here is EXPECT_EQ (non-fatal), so on a failing call the helpers were wrapping an uninitialized handle in PipelineOutputsHandle and the destructor would call daliPipelineOutputsDestroy on garbage. With this init the destructor sees nullptr on the failure path and the original EXPECT_EQ failure stays visible. The C API itself is unchanged: per the convention shared with the CUDA runtime, output parameters are not modified on non-success returns (only DALI_NO_DATA writes nullptr, which these functions do not return), so initialization is the caller's responsibility.
• operator_factory.h — Coverity-flagged copy-when-could-move in Registerer. Take Registerer::name by value and std::move both name and creator into OperatorRegistry::Register, eliminating two copies on each operator registration. Also std::move(name) in the Register<Backend> template overload so the name reaches the inner Register without an extra copy.

The remaining Coverity findings from the same snapshot (an unchecked-return-value in FileLabelLoaderBase and an uncaught-exception in vendored third_party/pybind11) are intentionally not addressed here.

Test plan

• DALI CI green (build + tests) on this branch
• No new regressions in CAPI2_SerializedPipelineTest.Checkpointing / CAPI2_PipelineBuilderTest.Checkpointing
• Next Coverity scan no longer flags the use-after-free in RunCheckpointingTest or the copy-when-could-move in Registerer

[jantonguirao]

!build

[greptile-apps]

Greptile Summary

This PR addresses three Coverity static-analysis findings across test and registration infrastructure. All changes are narrow, mechanical, and leave public API and runtime behaviour unchanged.

• complex_pipeline_test.cc: Replaces the single reused checkpoint_h raw-handle variable with two distinct locals (checkpoint1_h / checkpoint2_h) so the RAII wrapper is clearly constructed from a fresh handle each time, eliminating the apparent use-after-transfer.
• pipeline_test_utils.h: Zero-initialises raw_out_h in PopOutputs and PopOutputsAsync so that when the non-fatal EXPECT_EQ inside CHECK_DALI fires, the PipelineOutputsHandle RAII wrapper receives nullptr rather than an indeterminate value and the destructor does not call daliPipelineOutputsDestroy on garbage.
• operator_factory.h: Changes Registerer's constructor to accept name by value and std::move both name and creator into OperatorRegistry::Register, and also moves name in the Register<Backend> template overload, removing two per-registration copies.

Confidence Score: 5/5

All three changes are narrow Coverity fixes with no functional impact — safe to merge.

Each change is a targeted, well-scoped fix: renaming a reused raw-handle local, adding a defensive null initialisation in test helpers, and converting constructor arguments to move semantics. None of them alter observable behaviour or touch production data paths beyond the registration fast-path at static-init time.

No files require special attention.

Important Files Changed

Filename	Overview
dali/c_api_2/op_test/complex_pipeline_test.cc	Splits reused checkpoint_h into distinct checkpoint1_h/checkpoint2_h locals to eliminate Coverity use-after-RAII-transfer warning; logic is unchanged.
dali/c_api_2/pipeline_test_utils.h	Initialises raw_out_h = nullptr in PopOutputs and PopOutputsAsync; prevents RAII wrapper from calling destroy on an uninitialised handle when the non-fatal CHECK_DALI assertion fires.
dali/pipeline/operator/operator_factory.h	Takes Registerer::name by value and moves both name and creator into Register; also moves name in the Register<Backend> template overload, eliminating redundant copies on every operator registration.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    subgraph complex_pipeline_test.cc
        A[daliPipelineGetCheckpoint] -->|writes| B[checkpoint1_h]
        B --> C[CheckpointHandle checkpoint1]
        D[daliPipelineDeserializeCheckpoint] -->|writes| E[checkpoint2_h]
        E --> F[CheckpointHandle checkpoint2]
    end

    subgraph pipeline_test_utils.h
        G["raw_out_h = nullptr"] --> H[daliPipelinePopOutputs]
        H -->|success| I[PipelineOutputsHandle wraps valid ptr]
        H -->|EXPECT_EQ fails| J[PipelineOutputsHandle wraps nullptr - safe destructor]
    end

    subgraph operator_factory.h
        K["Registerer(std::string name, Creator creator)"] -->|std::move both| L["OperatorRegistry::Register(std::string, Creator, ...)"]
        M["Register-Backend-(std::string name, Creator creator)"] -->|std::move both| L
    end

Loading

_{Reviews (5): Last reviewed commit: "Address review: drop pipeline.cc change,..." | Re-trigger Greptile}

[Copilot]

Pull request overview

This PR addresses two Coverity findings in DALI: it updates operator registration to avoid extra copies during static registration, and it refactors a checkpointing test to use RAII for output handles and distinct checkpoint variables.

Changes:

• Changed Registerer to take operator names by value and move both the name and creator into OperatorRegistry::Register.
• Replaced manual output-pop/destroy code in RunCheckpointingTest with the PopOutputs RAII helper.
• Split a reused checkpoint raw handle into separate checkpoint1_h and checkpoint2_h variables in the checkpointing test.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
dali/pipeline/operator/operator_factory.h	Updates registration helper argument passing to reduce copies during operator registration.
dali/c_api_2/op_test/complex_pipeline_test.cc	Refactors checkpointing test resource management and raw checkpoint handle naming.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[mzient]

1. The test was meant to be run on raw C API. It's a C API test, after all.
2. The comment about daliPipelinePopOutputs is correct - the function is broken and needs fixing.

[jantonguirao]

Fixed

[dali-automaton]

CI MESSAGE: [50302622]: BUILD STARTED

[mzient]

This is against the convention. If the function fails, it should not modify the output variable.
FWIW, CUDA runtime APIs behave the same way - if the arguments are invalid, no output is returned:

#include <cuda_runtime.h>
#include <stdio.h>

int main() {
  cudaStream_t s = (cudaStream_t)0x12345678;
  cudaError_t err = cudaStreamCreateWithFlags(&s, 0x1234568);
  printf("%i %p\n", err, s);
  return 0;
}

Output:

1 0x12345678

[mzient]

Likewise.

[dali-automaton]

CI MESSAGE: [50302622]: BUILD PASSED

[jantonguirao]

!build

[dali-automaton]

CI MESSAGE: [50328210]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [50328210]: BUILD PASSED