Recompilation Pipeline Feature for the exploration of interesting data properties

[StoeckOverflow]

This Draft PR introduces a basic recompilation pipeline designed to support recompilation during the execution of loops. The implemented functionality includes:

1. Recompile Kernel: A C++ kernel that accepts MLIR code as a string, along with variadic inputs and outputs, and recompiles it.
2. Recompile Pass: A new pass that extracts the loop body of ForOps, feeding it as input to the Recompile Kernel. This allows the loop to be recompiled dynamically at each iteration.

---

Key Features

1. Dynamic Loop Recompilation:

   • The Recompile Pass wraps around ForOp loop bodies, triggering recompilation of the loop during execution.
   • Recompilation happens on a per-iteration basis as a proof-of-concept.

2. Extensible Recompilation Control:

   • Future iterations will allow for configurable recompilation conditions (e.g., recompiling after n iterations or under specific runtime conditions).
   • Potential to customize the PassManager for the RecompileOp to utilize either the full pass pipeline or a subset of it.

3. Testing:

   • A RecompileOp Kernel Test validates the core functionality of the C++ Recompile Kernel.
   • A script-level test ensures the overall pipeline works end-to-end, testing integration between MLIR and the dynamic recompilation mechanism.

---

Known Issues

There is currently an issue with the invocation of the recompilation kernel that needs to be resolved. Specifically:

• Error:

loc("-":5:10): error: 'daphne.cast' op using value defined outside the region
Failed to parse source string.
loc("-":5:10): error: 'daphne.cast' op using value defined outside the region
Failed to parse source string.
[error]: Execution error: The kernel-function _decRef__Structure failed during runtime with the following message [ DataObjectFactory::destroy() must not be called with nullptr ]
   | Source file -> "/home/dominic/daphne/test/api/cli/recompilation/recompilation_for_test_1.daphne":5:14
   | 
 5 | P = (X + Y)*2 + 1e-6;
   |               ^~~

• Investigation Path:
  This issue appears to be related to how arguments are forwarded to the kernel. A deeper analysis of the LLVM code generated by the program may provide insights into the root cause and potential solutions.

---

Areas for Further Development

This PR serves as a foundation for exploring the following enhancements:

1. Recompilation Invocation Logic: Implementing heuristics to decide when recompilation is necessary.
2. Partial vs. Full Pass Pipeline Execution: Determining whether the PassManager for the RecompileOp should execute the entire pass pipeline or only specific passes.

Additionally:

• Improve Script-Level Tests: The current script-level tests, while sufficient to verify basic functionality, produce NaN results. Developing more robust and precise tests can better validate the pipeline and help identify edge cases or hidden issues.

---

Tests and Validation

1. C++ Recompile Kernel Test:

   • Verifies the Recompile Kernel processes MLIR strings correctly with the given inputs and outputs.

2. Pipeline Integration Test:

   • Validates the end-to-end functionality of the recompilation pipeline through script-level tests.