[Mlir-commits] [mlir] [XeGPU] Handle scf.while result types from scf.condition (PR #202931)

[Mehdi Amini]

================
@@ -537,9 +526,28 @@ void xegpu::doSCFStructuralTypeConversionWithTensorType(
       return WalkResult::advance();
     });
 
+    // For scf.while, the parent op results correspond to scf.condition
+    // operands, not to the after-region scf.yield operands.
+    op->walk([](scf::ConditionOp conditionOp) {
+      auto whileOp = dyn_cast<scf::WhileOp>(conditionOp->getParentOp());
+      if (!whileOp)
+        return;
----------------
joker-eph wrote:

Is this possible? Otherwise seems like we should just cast

https://github.com/llvm/llvm-project/pull/202931