Caching potentially causes intermittent segfault

[joelberkeley]

I was seeing intermittent segfaults in this line

Enzyme/enzyme/Enzyme/MLIR/Passes/EnzymeMLIRPass.cpp

Line 342 in ad6c5f6

dyn_cast<AutoDiffFunctionInterface>(newFunc.getOperation())) {

The segfaults stopped if I deleted this cache fetch

Enzyme/enzyme/Enzyme/MLIR/Interfaces/EnzymeLogicReverse.cpp

Lines 219 to 223 in 9b78a1c

	{
	auto cachedFn = ReverseCachedFunctions.find(tup);
	if (cachedFn != ReverseCachedFunctions.end())
	return cachedFn->second;
	}

This was observed when differentiating several MLIR graphs, such as below. Each graph would be run anywhere between 2 to 400 times, each run with different constants. I ran it using the C++ API, not command line.

module @root {
  func.func @main() -> tensor<f64> {
    %cst = stablehlo.constant dense<1.000000e+00> : tensor<f64>
    %cst_0 = stablehlo.constant dense<0.000000e+00> : tensor<f64>
    %0 = enzyme.autodiff_region(%cst_0, %cst) {
    ^bb0(%arg0: tensor<f64>):
      enzyme.yield %arg0 : tensor<f64>
    } attributes {activity = [#enzyme<activity enzyme_active>], ret_activity = [#enzyme<activity enzyme_activenoneed>]} : (tensor<f64>, tensor<f64>) -> tensor<f64>
    return %0 : tensor<f64>
  }
}

These are my passes

{-#
  external_resources: {
    mlir_reproducer: {
      pipeline: "builtin.module(outline-enzyme-regions, enzyme{postpasses=canonicalize,remove-unnecessary-enzyme-ops verifyPostPasses=true}, arith-raise{stablehlo=true})",
      disable_threading: false,
      verify_each: true
    }
  }
#-}