[Mlir-commits] [mlir] [MLIR][Affine] Fix private memref creation bug in affine fusion (PR #126028)

[Arnab Dutta]

================
@@ -2297,3 +2298,40 @@ FailureOr<AffineValueMap> mlir::affine::simplifyConstrainedMinMaxOp(
   affine::canonicalizeMapAndOperands(&newMap, &newOperands);
   return AffineValueMap(newMap, newOperands);
 }
+
+Block *mlir::affine::findInnermostCommonBlockInScope(Operation *a,
+                                                     Operation *b) {
+  Region *aScope = mlir::affine::getAffineScope(a);
+  Region *bScope = mlir::affine::getAffineScope(b);
+  if (aScope != bScope)
+    return nullptr;
+
+  // Get the block ancestry of `a` while stopping at the affine scope.
+  auto getBlockAncestry = [&](Operation *op,
+                              SmallVectorImpl<Block *> &ancestry) {
+    Operation *curOp = op;
+    do {
+      ancestry.push_back(curOp->getBlock());
+      if (curOp->getParentRegion() == aScope)
+        break;
+      curOp = curOp->getParentOp();
+    } while (curOp);
+    assert(curOp && "can't reach root op without passing through affine scope");
+    std::reverse(ancestry.begin(), ancestry.end());
----------------
arnab-polymage wrote:

llvm::reverse?

https://github.com/llvm/llvm-project/pull/126028