[Mlir-commits] [mlir] [mlir][dataflow] Delete visitRegionSuccessors not run code (NFC) (PR #174790)

[Matthias Springer]

================
@@ -307,29 +307,20 @@ void AbstractSparseForwardDataFlowAnalysis::visitRegionSuccessors(
            "expected the same number of successor inputs as operands");
 
     unsigned firstIndex = 0;
-    if (inputs.size() != lattices.size()) {
-      if (!point->isBlockStart()) {
-        if (!inputs.empty())
-          firstIndex = cast<OpResult>(inputs.front()).getResultNumber();
-        visitNonControlFlowArgumentsImpl(
----------------
matthias-springer wrote:

Actually, I now see that there is an implicit assumption built into the data flow analysis framework:
```
  // The results of a region branch operation are determined by control-flow.
  if (auto branch = dyn_cast<RegionBranchOpInterface>(op)) {
    visitRegionSuccessors(getProgramPointAfter(branch), branch,
                          /*successor=*/{branch, branch->getResults()},
                          resultLattices);
    return success();
  }
```

I'm not sure if this comment is correct. We are treating block arguments and op results in a non-uniform way.

Before merging this PR, we should first clarify if op results of a `RegionBranchOpInterface` op must be successor inputs. (While I have no real-world counter-example where this would be a problem, it still seems a bit strange to me.) You found something interesting here!


https://github.com/llvm/llvm-project/pull/174790