[MLIR] MLIREnzymeAnalysis fails to build: getSuccessorInputs removed in current MLIR

[agarret7]

MLIREnzymeAnalysis fails to build against current MLIR (getSuccessorInputs removed)

enzyme/Enzyme/MLIR/Analysis/ActivityAnalysis.cpp calls
RegionBranchOpInterface::getSuccessorInputs and
RegionBranchTerminatorOpInterface in a way that no longer compiles
against MLIR main (and LLVM/MLIR 22). Both APIs were changed as part of
the region branch interface refactor upstream.

Error

ActivityAnalysis.cpp:1169:28: error: 'class mlir::RegionBranchOpInterface'
    has no member named 'getSuccessorInputs'; did you mean 'getSuccessorRegions'?

ActivityAnalysis.cpp:1284:37: error: no match for call to
    '(...lambda...)(mlir::Operation*&, mlir::RegionBranchTerminatorOpInterface,
    std::deque<mlir::Block*>&)'
    note: no known conversion for argument 2 from
    'mlir::RegionBranchTerminatorOpInterface' to 'mlir::RegionBranchPoint'

Impact

MLIREnzymeAnalysis is a transitive dependency of MLIREnzymeTransforms
and enzymemlir-opt, so the entire MLIR test suite (test/MLIR/) is
currently broken on main.

Relevant upstream change

getSuccessorInputs was removed and RegionBranchTerminatorOpInterface
was replaced by RegionBranchPoint in the MLIR region branch interface
refactor. The fix likely involves updating the two call sites in
ActivityAnalysis.cpp to use the new API.

[wsmoses]

currently we only support a ~near llvm main commit for mlir because the api changes so much. This is in contrast to llvm where we support llvm major versions 15+, and main. Is there a particular reason you want a build against mlir 22?

[agarret7]

Yeah, I was thinking of exposing an Enzyme-enabled autodiff feature using Melior (prototype). Their bindings were upgraded to 22 a couple of months ago.

Perhaps I can add mlir 22 compatibility as a follow-up to #2859, or include it as part of the current PR?

[wsmoses]

yeah I'd do that in a separate PR, but also you should then add CI for that as well [in addition to the current mlir ci which builds from a specific near main commit]

[agarret7]

I have a PR in progress that fixes the getSuccessorInputs and related API changes (Op::create, PropertyRef, mlir::Complex, etc.). It's all mechanical renames, tested locally against llvmorg-22.1.0 with 130/130 MLIR tests passing.

However, while working on this I ran into a compatibility issue: enzyme's main currently tracks a near-main LLVM commit (87d42c13, which is LLVM 23-dev), and the LLVM 22 and 23-dev APIs are moving in opposite directions on some things (e.g. builder.create<Op>(...) is the correct form in 22 but deprecated in 23-dev in favor of Op::create(builder, ...)).

Would you accept a PR that adds #if LLVM_VERSION_MAJOR guards to the MLIR code (like the non-MLIR code already has) and adds llvmorg-22.1.0 to the CI matrix?