[WIP] ORT aborts with the `linspace` implementation when input is empty

[Copilot]

Fix ORT abort with linspace implementation when input is empty

Analysis

• [x] Explore repository structure and understand the issue
• [x] Reproduce the issue: confirmed it's a floating point exception (abort) during division by zero
• [x] Identify the root cause: The issue occurs in the Div operation when dividing by zero with integer types
• [x] Locate relevant source files: onnxruntime/core/providers/cpu/math/element_wise_ops.cc
• [x] Understand the current implementation uses Eigen arithmetic which doesn't check for division by zero

Implementation

• [x] Added #include <type_traits> for compile-time type checking
• [x] Created CheckDivisionByZero<T>() helper function that:
  • Uses if constexpr (std::is_integral_v<T>) to only check integer types
  • Iterates through denominator tensor to detect any zero values
  • Returns proper error status instead of allowing SIGFPE
• [x] Modified Div<T>::Compute() to call validation before performing division
• [x] Added comprehensive test cases for both integer and floating point division by zero
• [x] Verified code compiles successfully
• [x] Validated logic with standalone tests

Changes Summary

• 25 lines added to onnxruntime/core/providers/cpu/math/element_wise_ops.cc
• 44 lines added to onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc
• Total: 69 lines added, 0 deleted (minimal surgical changes)

Root Cause

The linspace implementation creates a model that performs division by zero:

1. Range(0, 1, 1) produces [0]
2. Sub(1, 1) produces 0
3. Div(1, 0) attempts to divide by zero, causing floating point exception instead of proper error message

Solution

For integer types (int32_t, int64_t, etc.), division by zero now returns: INVALID_ARGUMENT: "Division by zero error in Div operator"

For floating point types (float, double), behavior remains unchanged (produces infinity per IEEE 754).

Testing & Validation

• [x] Standalone validation confirms logic works correctly
• [x] Integer division by zero detected and returns proper error
• [x] Float division by zero allowed (produces infinity as expected)
• [x] Code compiles without errors
• [x] Original reproduction case still fails as expected (fix needs full build to test)

Next steps: Full build and integration testing will validate that the fix resolves the original crashing issue.

Fixes #16998.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.