Fix build failure from direct ONNX schema include in layernorm optimizer tests

[JonathanC-ARM]

Description

This fixes a macOS AppleClang build failure in onnxruntime_test_all caused by directly including ONNX’s onnx/defs/schema.h from graph_transform_test_layernorm.cc.

With the macOS arm64 CI build, ORT enables -Wshorten-64-to-32 and treats warnings as errors. ONNX’s schema header pulls in protobuf v21.12 headers, and protobuf’s google/protobuf/parse_context.h emits several -Wshorten-64-to-32 diagnostics under AppleClang. Because the test file included onnx/defs/schema.h directly, it bypassed ORT’s existing core/graph/onnx_protobuf.h wrapper, which already suppresses this known protobuf warning around ONNX/protobuf includes.

The fix is to include/use the existing ORT ONNX protobuf wrapper instead of directly including onnx/defs/schema.h.

Repro

On macOS arm64 with AppleClang/Xcode, run:

./build.sh \
  --config "${ORT_BUILD_CONFIG:-Release}" \
  --use_xcode \
  --apple_sysroot macosx \
  --build_shared_lib \
  --parallel \
  --skip_tests \
  --osx_arch arm64 \
  --apple_deploy_target 14 \
  --cmake_generator Ninja

A faster targeted repro is:

python3 tools/ci_build/build.py \
  --config Release \
  --use_xcode \
  --apple_sysroot macosx \
  --build_shared_lib \
  --parallel \
  --skip_tests \
  --osx_arch arm64 \
  --apple_deploy_target 14 \
  --cmake_generator Ninja \
  --build_dir build/repro_ci \
  --update --build \
  --target onnxruntime_test_all

Before this fix, the build fails compiling:

onnxruntime/test/optimizer/graph_transform_test_layernorm.cc

with errors like:

google/protobuf/parse_context.h:328:47: error: implicit conversion loses integer precision: 'long' to 'int' [-Werror,-Wshorten-64-to-32]
  int chunk_size = buffer_end_ + kSlopBytes - ptr;

Root cause

graph_transform_test_layernorm.cc directly included:

#include "onnx/defs/schema.h"

This exposed protobuf v21.12 headers to ORT’s normal AppleClang warning policy:

• -Wshorten-64-to-32 is enabled
• warnings are treated as errors
• protobuf v21.12 emits known -Wshorten-64-to-32 warnings in parse_context.h

ORT already has core/graph/onnx_protobuf.h, which wraps ONNX/protobuf includes with the appropriate diagnostic suppression. The direct ONNX include bypassed that wrapper.

Summary of fix

Replace the direct ONNX schema include in graph_transform_test_layernorm.cc with ORT’s existing ONNX protobuf wrapper include, so protobuf headers are included under the existing warning suppression.

Validation

Validated locally on macOS arm64 / Xcode 16.4 / AppleClang 17 with:

python3 tools/ci_build/build.py \
  --config Release \
  --use_xcode \
  --apple_sysroot macosx \
  --build_shared_lib \
  --parallel \
  --skip_tests \
  --osx_arch arm64 \
  --apple_deploy_target 14 \
  --cmake_generator Ninja \
  --build_dir build/repro_ci \
  --build \
  --target onnxruntime_test_all

and then the broader build step:

python3 tools/ci_build/build.py \
  --config Release \
  --use_xcode \
  --apple_sysroot macosx \
  --build_shared_lib \
  --parallel \
  --skip_tests \
  --osx_arch arm64 \
  --apple_deploy_target 14 \
  --cmake_generator Ninja \
  --build_dir build/repro_ci \
  --build

Both completed successfully after the fix.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Fixes a macOS AppleClang build failure in onnxruntime_test_all by avoiding direct inclusion of ONNX schema headers in a LayerNorm optimizer test, instead relying on ORT’s existing ONNX/protobuf wrapper that suppresses known Protobuf warnings under strict warning settings.

Changes:

• Replace direct include of onnx/defs/schema.h with core/graph/onnx_protobuf.h in the LayerNorm graph transform test.
• Ensure protobuf headers are included under ORT’s diagnostic suppression on AppleClang (arm64).