[flang-commits] [flang] [clang-tools-extra] [clang] [mlir] [llvm] [compiler-rt] [mlir][vector] Fix invalid `LoadOp` indices being created (PR #76292)

[Mehdi Amini via flang-commits]

================
@@ -866,16 +866,41 @@ struct TransferOpConversion : public VectorToSCFPattern<OpTy> {
     this->setHasBoundedRewriteRecursion();
   }
 
+  static void getMaskBufferLoadIndices(OpTy xferOp, Value castedMaskBuffer,
+                                       SmallVector<Value, 8> &loadIndices,
+                                       Value iv) {
+    assert(xferOp.getMask() && "Expected transfer op to have mask");
+
+    // Add load indices from the previous iteration.
+    // The mask buffer depends on the permutation map, which makes determining
+    // the indices quite complex, so this is why we need to "look back" to the
+    // previous iteration to find the right indices.
+    Value maskBuffer = getMaskBuffer(xferOp);
+    for (OpOperand &use : maskBuffer.getUses()) {
+      // If there is no previous load op, then the indices are empty.
+      if (auto loadOp = dyn_cast<memref::LoadOp>(use.getOwner())) {
----------------
joker-eph wrote:

```suggestion
    for (Operation *user : maskBuffer.getUsers()) {
      // If there is no previous load op, then the indices are empty.
      if (auto loadOp = dyn_cast<memref::LoadOp>(user)) {
```

https://github.com/llvm/llvm-project/pull/76292