[Mlir-commits] [mlir] [mlir][vector] Migrate drop-lead-unit-dim to shape_cast (PR #196206)

Krzysztof Drewniak llvmlistbot at llvm.org
Fri May 8 13:52:42 PDT 2026


https://github.com/krzysz00 updated https://github.com/llvm/llvm-project/pull/196206

>From 5b2ae1b3f3e48f6629d4f6a3ea36b98e069ab99a Mon Sep 17 00:00:00 2001
From: Krzysztof Drewniak <Krzysztof.Drewniak at amd.com>
Date: Wed, 6 May 2026 19:53:50 +0000
Subject: [PATCH 1/8] [mlir][vector] Migrate drop-lead-unit-dim to shape_cast

Post-merge discussion on #195686 led to the conclusion that we should
change the behavior of drealeadunitdim to use shape_cast instead of
extracts and broadcasts since those are now the canonical form of such
unit-dimension striping. This commit implements that change.

The one exception is that vector contractions where the accumulator is
reduced to a scalar still use extract/broadcast.

AI: Codex 5.5 did most of the work on this one.
---
 .../Transforms/VectorDropLeadUnitDim.cpp      | 260 +++++++------
 .../vector-dropleadunitdim-transforms.mlir    | 341 ++++++++++--------
 .../Dialect/Vector/vector-transforms.mlir     |   4 +-
 3 files changed, 347 insertions(+), 258 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
index 26a702ef0f512..931a87ff83e9c 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
@@ -22,6 +22,11 @@
 using namespace mlir;
 using namespace mlir::vector;
 
+/// Return a smallVector of size `rank` containing all zeros.
+static SmallVector<int64_t> splatZero(int64_t rank) {
+  return SmallVector<int64_t>(rank, 0);
+}
+
 // Trims leading one dimensions from `oldType` and returns the result type.
 // Returns `vector<1xT>` if `oldType` only has one element.
 static VectorType trimLeadingOneDims(VectorType oldType) {
@@ -45,14 +50,82 @@ static VectorType trimLeadingOneDims(VectorType oldType) {
   return VectorType::get(newShape, oldType.getElementType(), newScalableDims);
 }
 
-/// Return a smallVector of size `rank` containing all zeros.
-static SmallVector<int64_t> splatZero(int64_t rank) {
-  return SmallVector<int64_t>(rank, 0);
+/// Returns `value` if it already has `newType`, otherwise inserts a
+/// vector.shape_cast to `newType`.
+static Value shapeCastVector(OpBuilder &b, Location loc, Value value,
+                             VectorType newType) {
+  if (value.getType() == newType)
+    return value;
+  return vector::ShapeCastOp::create(b, loc, newType, value);
+}
+
+static bool hasNonScalableUnitLeadingDims(VectorType type, int64_t dropCount) {
+  if (dropCount < 0 || dropCount > type.getRank())
+    return false;
+  ArrayRef<int64_t> leadingShape = type.getShape().take_front(dropCount);
+  ArrayRef<bool> leadingScalable = type.getScalableDims().take_front(dropCount);
+  return llvm::all_of(leadingShape, [](int64_t dim) { return dim == 1; }) &&
+         llvm::none_of(leadingScalable,
+                       [](bool scalable) { return scalable; });
+}
+
+static bool isNonScalableUnitDim(VectorType type, int64_t dim) {
+  return dim >= 0 && dim < type.getRank() && type.getShape()[dim] == 1 &&
+         !type.getScalableDims()[dim];
+}
+
+/// Shape-casts `operand` to the vector type obtained by dropping the first
+/// `dropCount` dimensions. Callers must ensure at least one vector dimension
+/// remains after the drop.
+static Value shapeCastDroppingLeadingDims(OpBuilder &b, Location loc,
+                                          Value operand, int64_t dropCount) {
+  auto oldType = cast<VectorType>(operand.getType());
+  assert(dropCount < oldType.getRank() &&
+         "shape_cast cannot drop all vector dimensions");
+  VectorType newType = VectorType::get(
+      oldType.getShape().drop_front(dropCount), oldType.getElementType(),
+      oldType.getScalableDims().drop_front(dropCount));
+  return shapeCastVector(b, loc, operand, newType);
+}
+
+static Value shapeCastDroppingDim(OpBuilder &b, Location loc, Value operand,
+                                  int64_t dim) {
+  auto oldType = cast<VectorType>(operand.getType());
+  assert(isNonScalableUnitDim(oldType, dim) &&
+         "expected a non-scalable unit dim to drop");
+
+  SmallVector<int64_t> newShape;
+  SmallVector<bool> newScalableDims;
+  for (int64_t i = 0, e = oldType.getRank(); i < e; ++i) {
+    if (i == dim)
+      continue;
+    newShape.push_back(oldType.getShape()[i]);
+    newScalableDims.push_back(oldType.getScalableDims()[i]);
+  }
+
+  return shapeCastVector(
+      b, loc, operand,
+      VectorType::get(newShape, oldType.getElementType(), newScalableDims));
+}
+
+static Value dropLeadingDimsForContraction(OpBuilder &b, Location loc,
+                                           Value operand, int64_t dropCount) {
+  auto oldType = cast<VectorType>(operand.getType());
+  assert(hasNonScalableUnitLeadingDims(oldType, dropCount) &&
+         "expected non-scalable leading unit dims to drop");
+
+  // vector.contract rejects 0-D vector accumulators/results. When every vector
+  // dimension is dropped, use the scalar path that vector.contract accepts.
+  if (dropCount == oldType.getRank())
+    return vector::ExtractOp::create(b, loc, operand, splatZero(dropCount));
+
+  return shapeCastDroppingLeadingDims(b, loc, operand, dropCount);
 }
+
 namespace {
 
 // Casts away leading one dimensions in vector.extract_strided_slice's vector
-// input by inserting vector.broadcast.
+// input by inserting vector.shape_cast.
 struct CastAwayExtractStridedSliceLeadingOneDim
     : public OpRewritePattern<vector::ExtractStridedSliceOp> {
   using Base::Base;
@@ -78,8 +151,8 @@ struct CastAwayExtractStridedSliceLeadingOneDim
 
     Location loc = extractOp.getLoc();
 
-    Value newSrcVector = vector::ExtractOp::create(
-        rewriter, loc, extractOp.getSource(), splatZero(dropCount));
+    Value newSrcVector =
+        shapeCastVector(rewriter, loc, extractOp.getSource(), newSrcType);
 
     // The offsets/sizes/strides attribute can have a less number of elements
     // than the input vector's rank: it is meant for the leading dimensions.
@@ -94,7 +167,7 @@ struct CastAwayExtractStridedSliceLeadingOneDim
         rewriter, loc, newDstType, newSrcVector, newOffsets, newSizes,
         newStrides);
 
-    rewriter.replaceOpWithNewOp<vector::BroadcastOp>(extractOp, oldDstType,
+    rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(extractOp, oldDstType,
                                                      newExtractOp);
 
     return success();
@@ -102,7 +175,7 @@ struct CastAwayExtractStridedSliceLeadingOneDim
 };
 
 // Casts away leading one dimensions in vector.insert_strided_slice's vector
-// inputs by inserting vector.broadcast.
+// inputs by inserting vector.shape_cast.
 struct CastAwayInsertStridedSliceLeadingOneDim
     : public OpRewritePattern<vector::InsertStridedSliceOp> {
   using Base::Base;
@@ -122,10 +195,10 @@ struct CastAwayInsertStridedSliceLeadingOneDim
     // Trim leading one dimensions from both operands.
     Location loc = insertOp.getLoc();
 
-    Value newSrcVector = vector::ExtractOp::create(
-        rewriter, loc, insertOp.getValueToStore(), splatZero(srcDropCount));
-    Value newDstVector = vector::ExtractOp::create(
-        rewriter, loc, insertOp.getDest(), splatZero(dstDropCount));
+    Value newSrcVector = shapeCastVector(rewriter, loc,
+                                         insertOp.getValueToStore(), newSrcType);
+    Value newDstVector =
+        shapeCastVector(rewriter, loc, insertOp.getDest(), newDstType);
 
     auto newOffsets = rewriter.getArrayAttr(
         insertOp.getOffsets().getValue().take_back(newDstType.getRank()));
@@ -136,7 +209,7 @@ struct CastAwayInsertStridedSliceLeadingOneDim
         rewriter, loc, newDstType, newSrcVector, newDstVector, newOffsets,
         newStrides);
 
-    rewriter.replaceOpWithNewOp<vector::BroadcastOp>(insertOp, oldDstType,
+    rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(insertOp, oldDstType,
                                                      newInsertOp);
 
     return success();
@@ -144,7 +217,7 @@ struct CastAwayInsertStridedSliceLeadingOneDim
 };
 
 // Casts away leading one dimensions in vector.insert's vector inputs by
-// inserting vector.broadcast.
+// inserting vector.shape_cast.
 struct CastAwayInsertLeadingOneDim : public OpRewritePattern<vector::InsertOp> {
   using Base::Base;
 
@@ -171,12 +244,11 @@ struct CastAwayInsertLeadingOneDim : public OpRewritePattern<vector::InsertOp> {
     Location loc = insertOp.getLoc();
 
     Value newSrcVector = insertOp.getValueToStore();
-    if (oldSrcRank != 0) {
-      newSrcVector = vector::ExtractOp::create(
-          rewriter, loc, insertOp.getValueToStore(), splatZero(srcDropCount));
-    }
-    Value newDstVector = vector::ExtractOp::create(
-        rewriter, loc, insertOp.getDest(), splatZero(dstDropCount));
+    if (oldSrcRank != 0)
+      newSrcVector = shapeCastVector(rewriter, loc, insertOp.getValueToStore(),
+                                     cast<VectorType>(newSrcType));
+    Value newDstVector =
+        shapeCastVector(rewriter, loc, insertOp.getDest(), newDstType);
 
     // New position rank needs to be computed in two steps: (1) if destination
     // type has leading unit dims, we also trim the position array accordingly,
@@ -193,7 +265,7 @@ struct CastAwayInsertLeadingOneDim : public OpRewritePattern<vector::InsertOp> {
     auto newInsertOp = vector::InsertOp::create(rewriter, loc, newSrcVector,
                                                 newDstVector, newPosition);
 
-    rewriter.replaceOpWithNewOp<vector::BroadcastOp>(insertOp, oldDstType,
+    rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(insertOp, oldDstType,
                                                      newInsertOp);
 
     return success();
@@ -201,20 +273,10 @@ struct CastAwayInsertLeadingOneDim : public OpRewritePattern<vector::InsertOp> {
 };
 
 static Value dropUnitDimsFromMask(OpBuilder &b, Location loc, Value mask,
-                                  VectorType newType, AffineMap newMap,
-                                  VectorType oldMaskType) {
+                                  VectorType newType, AffineMap newMap) {
   // Infer the type of the new mask from the new map.
   VectorType newMaskType = inferTransferOpMaskType(newType, newMap);
-
-  // If the new mask is broadcastable to the old result type, we can safely
-  // use a `vector.extract` to get the new mask. Otherwise the best we can
-  // do is shape cast.
-  if (vector::isBroadcastableTo(newMaskType, oldMaskType) ==
-      BroadcastableToResult::Success) {
-    int64_t dropDim = oldMaskType.getRank() - newMaskType.getRank();
-    return vector::ExtractOp::create(b, loc, mask, splatZero(dropDim));
-  }
-  return vector::ShapeCastOp::create(b, loc, newMaskType, mask);
+  return shapeCastVector(b, loc, mask, newMaskType);
 }
 
 // Turns vector.transfer_read on vector with leading 1 dimensions into
@@ -256,16 +318,14 @@ struct CastAwayTransferReadLeadingOneDim
           read.getInBoundsAttr().getValue().take_back(newType.getRank()));
 
     Value mask = Value();
-    if (read.getMask()) {
-      VectorType maskType = read.getMaskType();
+    if (read.getMask())
       mask = dropUnitDimsFromMask(rewriter, read.getLoc(), read.getMask(),
-                                  newType, newMap, maskType);
-    }
+                                  newType, newMap);
 
     auto newRead = vector::TransferReadOp::create(
         rewriter, read.getLoc(), newType, read.getBase(), read.getIndices(),
         AffineMapAttr::get(newMap), read.getPadding(), mask, inBoundsAttr);
-    rewriter.replaceOpWithNewOp<vector::BroadcastOp>(read, oldType, newRead);
+    rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(read, oldType, newRead);
 
     return success();
   }
@@ -295,8 +355,6 @@ struct CastAwayTransferWriteLeadingOneDim
     VectorType newType = trimLeadingOneDims(oldType);
     if (newType == oldType)
       return failure();
-    int64_t dropDim = oldType.getRank() - newType.getRank();
-
     AffineMap oldMap = write.getPermutationMap();
     ArrayRef<AffineExpr> newResults =
         oldMap.getResults().take_back(newType.getRank());
@@ -309,13 +367,12 @@ struct CastAwayTransferWriteLeadingOneDim
       inBoundsAttr = rewriter.getArrayAttr(
           write.getInBoundsAttr().getValue().take_back(newType.getRank()));
 
-    auto newVector = vector::ExtractOp::create(
-        rewriter, write.getLoc(), write.getVector(), splatZero(dropDim));
+    auto newVector =
+        shapeCastVector(rewriter, write.getLoc(), write.getVector(), newType);
 
     if (write.getMask()) {
-      VectorType maskType = write.getMaskType();
       Value newMask = dropUnitDimsFromMask(
-          rewriter, write.getLoc(), write.getMask(), newType, newMap, maskType);
+          rewriter, write.getLoc(), write.getMask(), newType, newMap);
       rewriter.replaceOpWithNewOp<vector::TransferWriteOp>(
           write, newVector, write.getBase(), write.getIndices(),
           AffineMapAttr::get(newMap), newMask, inBoundsAttr);
@@ -340,7 +397,7 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
     return failure();
   if (oldAccType.getRank() < 1)
     return failure();
-  if (oldAccType.getShape()[0] != 1)
+  if (!isNonScalableUnitDim(oldAccType, 0))
     return failure();
   // currently we support only dropping one dim but the pattern can be applied
   // greedily to drop more.
@@ -372,8 +429,8 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
 
   for (const auto &it : llvm::enumerate(oldIndexingMaps)) {
     // Check if the dim to be dropped exists as a leading dim in the operand
-    // if it does then we use vector.extract to drop it.
-    bool validExtract = false;
+    // if it does then we use vector.shape_cast to drop it.
+    bool needsShapeCast = false;
     SmallVector<AffineExpr> results;
     auto map = it.value();
     int64_t orginalZeroDim = it.value().getDimPosition(0);
@@ -415,8 +472,8 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
         }
       }
 
-      // Do the transpose now if needed so that we can drop the
-      // correct dim using extract later.
+      // Do the transpose now if needed so that we can drop the correct dim
+      // with shape_cast later.
       if (transposeNeeded) {
         map = AffineMap::get(map.getNumDims(), 0, transposeResults,
                              contractOp.getContext());
@@ -428,10 +485,9 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
     }
     // We have taken care to have the dim to be dropped be
     // the leading dim. If its still not leading that means it
-    // does not exist in this operand and hence we do not need
-    // an extract.
+    // does not exist in this operand and hence we do not need a shape_cast.
     if (map.getDimPosition(0) == dimToDrop)
-      validExtract = true;
+      needsShapeCast = true;
 
     for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
       int64_t currDim = map.getDimPosition(i);
@@ -444,13 +500,16 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
     }
     newIndexingMaps.push_back(AffineMap::get(map.getNumDims() - 1, 0, results,
                                              contractOp.getContext()));
-    // Extract if its a valid extraction, otherwise use the operand
-    // without extraction.
-    newOperands.push_back(validExtract
-                              ? vector::ExtractOp::create(rewriter, loc,
-                                                          operands[it.index()],
-                                                          splatZero(dropDim))
-                              : operands[it.index()]);
+    if (needsShapeCast) {
+      auto operandType = cast<VectorType>(operands[it.index()].getType());
+      if (operandType.getRank() < dropDim ||
+          !hasNonScalableUnitLeadingDims(operandType, dropDim))
+        return failure();
+      newOperands.push_back(dropLeadingDimsForContraction(
+          rewriter, loc, operands[it.index()], dropDim));
+    } else {
+      newOperands.push_back(operands[it.index()]);
+    }
   }
 
   // Depending on whether this vector.contract is masked, the replacing Op
@@ -461,12 +520,22 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
       rewriter.getArrayAttr(newIteratorTypes), contractOp.getKind());
 
   if (maskingOp) {
-    auto newMask = vector::ExtractOp::create(rewriter, loc, maskingOp.getMask(),
-                                             splatZero(dropDim));
+    auto oldMaskType = cast<VectorType>(maskingOp.getMask().getType());
+    if (oldMaskType.getRank() <= 1 ||
+        !isNonScalableUnitDim(oldMaskType, dimToDrop))
+      return failure();
+    Value newMask =
+        shapeCastDroppingDim(rewriter, loc, maskingOp.getMask(), dimToDrop);
 
     newOp = mlir::vector::maskOperation(rewriter, newOp, newMask);
   }
 
+  if (isa<VectorType>(newOp->getResults()[0].getType()))
+    return vector::ShapeCastOp::create(rewriter, loc,
+                                       contractOp->getResultTypes()[0],
+                                       newOp->getResults()[0])
+        .getResult();
+
   return vector::BroadcastOp::create(rewriter, loc,
                                      contractOp->getResultTypes()[0],
                                      newOp->getResults()[0])
@@ -476,9 +545,9 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
 namespace {
 
 /// Turns vector.contract on vector with leading 1 dimensions into
-/// vector.extract followed by vector.contract on vector without leading
+/// vector.shape_cast followed by vector.contract on vector without leading
 /// 1 dimensions. Also performs transpose of lhs and rhs operands if required
-/// prior to extract.
+/// prior to the shape cast.
 struct CastAwayContractionLeadingOneDim
     : public MaskableOpRewritePattern<vector::ContractionOp> {
   using MaskableOpRewritePattern::MaskableOpRewritePattern;
@@ -493,14 +562,15 @@ struct CastAwayContractionLeadingOneDim
 
 /// Looks at elementwise operations on vectors with at least one leading
 /// dimension equal 1, e.g. vector<1x[4]x1xf32> (but not vector<2x[4]x1xf32>),
-/// and cast aways the leading one dimensions (_plural_) and then broadcasts
-/// the results.
+/// and casts away the leading one dimensions (_plural_) with shape_cast.
 ///
 /// Example before:
 ///     %1 = arith.mulf %arg0, %arg1 : vector<1x4x1xf32>
 /// Example after:
-///    %2 = arith.mulf %0, %1 : vector<4x1xf32>
-///    %3 = vector.broadcast %2 : vector<4x1xf32> to vector<1x4x1xf32>
+///    %2 = vector.shape_cast %arg0 : vector<1x4x1xf32> to vector<4x1xf32>
+///    %3 = vector.shape_cast %arg1 : vector<1x4x1xf32> to vector<4x1xf32>
+///    %4 = arith.mulf %2, %3 : vector<4x1xf32>
+///    %5 = vector.shape_cast %4 : vector<4x1xf32> to vector<1x4x1xf32>
 ///
 /// Does support scalable vectors.
 class CastAwayElementwiseLeadingOneDim : public RewritePattern {
@@ -519,52 +589,29 @@ class CastAwayElementwiseLeadingOneDim : public RewritePattern {
     VectorType newVecType = trimLeadingOneDims(vecType);
     if (newVecType == vecType)
       return failure();
-    int64_t dropDim = vecType.getRank() - newVecType.getRank();
     SmallVector<Value, 4> newOperands;
     for (Value operand : op->getOperands()) {
-      if (auto opVecType = dyn_cast<VectorType>(operand.getType())) {
-        newOperands.push_back(vector::ExtractOp::create(
-            rewriter, op->getLoc(), operand, splatZero(dropDim)));
-      } else {
+      if (auto opVecType = dyn_cast<VectorType>(operand.getType()))
+        newOperands.push_back(shapeCastVector(
+            rewriter, op->getLoc(), operand, trimLeadingOneDims(opVecType)));
+      else
         newOperands.push_back(operand);
-      }
     }
     Operation *newOp =
         rewriter.create(op->getLoc(), op->getName().getIdentifier(),
                         newOperands, newVecType, op->getAttrs());
-    rewriter.replaceOpWithNewOp<vector::BroadcastOp>(op, vecType,
+    rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(op, vecType,
                                                      newOp->getResult(0));
     return success();
   }
 };
 } // namespace
 
-// Drops `dropDim` leading dimensions from `operand` using vector.extract when
-// those dims are all non-scalable units (the cheap, structural rewrite); falls
-// back to vector.shape_cast otherwise.
-static Value dropLeadingOneDimsFromOperand(OpBuilder &b, Location loc,
-                                           Value operand, int64_t nDropped) {
-  auto oldType = cast<VectorType>(operand.getType());
-  ArrayRef<int64_t> leadingShape = oldType.getShape().take_front(nDropped);
-  ArrayRef<bool> leadingScalable =
-      oldType.getScalableDims().take_front(nDropped);
-  bool extractable =
-      llvm::all_of(leadingShape, [](int64_t d) { return d == 1; }) &&
-      llvm::none_of(leadingScalable, [](bool s) { return s; });
-  if (extractable)
-    return vector::ExtractOp::create(b, loc, operand, splatZero(nDropped));
-  VectorType newType = VectorType::get(
-      oldType.getShape().drop_front(nDropped), oldType.getElementType(),
-      oldType.getScalableDims().drop_front(nDropped));
-  return vector::ShapeCastOp::create(b, loc, newType, operand);
-}
-
 namespace {
 
-// Drops leading 1 dimensions from load-like memory operaitons. REmoves leading
-// unit dimensions from the result types and then broadcasts back in those 1s,
-// while also extracting (or shape_cast-ing) any leading unit dimensions on
-// the input operands.
+// Drops leading unit dimensions from load-like memory operations by
+// shape_casting each vector operand and shape_casting the result back to the
+// original type.
 template <typename OpTy>
 struct CastAwayLoadLikeLeadingOneDim : public OpRewritePattern<OpTy> {
   using OpRewritePattern<OpTy>::OpRewritePattern;
@@ -583,7 +630,7 @@ struct CastAwayLoadLikeLeadingOneDim : public OpRewritePattern<OpTy> {
     for (Value operand : op->getOperands()) {
       if (isa<VectorType>(operand.getType())) {
         newOperands.push_back(
-            dropLeadingOneDimsFromOperand(rewriter, loc, operand, nDropped));
+            shapeCastDroppingLeadingDims(rewriter, loc, operand, nDropped));
       } else {
         newOperands.push_back(operand);
       }
@@ -592,15 +639,14 @@ struct CastAwayLoadLikeLeadingOneDim : public OpRewritePattern<OpTy> {
     Operation *newOp =
         rewriter.create(loc, op->getName().getIdentifier(), newOperands,
                         TypeRange{newResultType}, op->getAttrs());
-    rewriter.replaceOpWithNewOp<vector::BroadcastOp>(op, oldResultType,
+    rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(op, oldResultType,
                                                      newOp->getResult(0));
     return success();
   }
 };
 
-// Drops leading 1 dimensions from store-like memory ops. Extracts or
-// `shape_cast`s away those leading unit dimensions and leaves any scalar
-// operands alone.
+// Drops leading unit dimensions from store-like memory operations by
+// shape_casting each vector operand and leaving any scalar operands alone.
 template <typename OpTy>
 struct CastAwayStoreLikeLeadingOneDim : public OpRewritePattern<OpTy> {
   using OpRewritePattern<OpTy>::OpRewritePattern;
@@ -619,7 +665,7 @@ struct CastAwayStoreLikeLeadingOneDim : public OpRewritePattern<OpTy> {
     for (Value operand : op->getOperands()) {
       if (isa<VectorType>(operand.getType())) {
         newOperands.push_back(
-            dropLeadingOneDimsFromOperand(rewriter, loc, operand, nDropped));
+            shapeCastDroppingLeadingDims(rewriter, loc, operand, nDropped));
       } else {
         newOperands.push_back(operand);
       }
@@ -633,8 +679,8 @@ struct CastAwayStoreLikeLeadingOneDim : public OpRewritePattern<OpTy> {
   }
 };
 
-// Drops leading 1 dimensions from vector.constant_mask and inserts a
-// vector.broadcast back to the original shape.
+// Drops leading 1 dimensions from vector.constant_mask and shape_casts back to
+// the original shape.
 struct CastAwayConstantMaskLeadingOneDim
     : public OpRewritePattern<vector::ConstantMaskOp> {
   using Base::Base;
@@ -659,7 +705,7 @@ struct CastAwayConstantMaskLeadingOneDim
 
     auto newMask = vector::ConstantMaskOp::create(rewriter, mask.getLoc(),
                                                   newType, newDimSizes);
-    rewriter.replaceOpWithNewOp<vector::BroadcastOp>(mask, oldType, newMask);
+    rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(mask, oldType, newMask);
     return success();
   }
 };
diff --git a/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir b/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir
index bf01c8a8589d9..e5b1cc07319e3 100644
--- a/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir
+++ b/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir
@@ -5,13 +5,13 @@
 // CHECK-DAG: #[[$map2:.*]] = affine_map<(d0, d1, d2) -> (d0, d1)>
 
 // CHECK-LABEL: cast_away_contraction_leading_one_dims
-//  CHECK-NEXT:   %[[R0:.+]] =  vector.extract %{{.*}}[0] : vector<16x8xf32> from vector<1x16x8xf32>
-//  CHECK-NEXT:   %[[R1:.+]] =  vector.extract %{{.*}}[0] : vector<8x16xf32> from vector<1x8x16xf32>
-//  CHECK-NEXT:   %[[R2:.+]] =  vector.extract %{{.*}}[0] : vector<16x16xf32> from vector<1x16x16xf32>
+//  CHECK-NEXT:   %[[R0:.+]] = vector.shape_cast %{{.*}} : vector<1x16x8xf32> to vector<16x8xf32>
+//  CHECK-NEXT:   %[[R1:.+]] = vector.shape_cast %{{.*}} : vector<1x8x16xf32> to vector<8x16xf32>
+//  CHECK-NEXT:   %[[R2:.+]] = vector.shape_cast %{{.*}} : vector<1x16x16xf32> to vector<16x16xf32>
 //  CHECK-NEXT:   %[[R3:.+]] = vector.contract {indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]],
 //  CHECK-SAME:   iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>}
 //  CHECK-SAME:   %[[R0]], %[[R1]], %[[R2]] : vector<16x8xf32>, vector<8x16xf32> into vector<16x16xf32>
-//  CHECK-NEXT:   %[[R4:.+]] = vector.broadcast %[[R3]] : vector<16x16xf32> to vector<1x16x16xf32>
+//  CHECK-NEXT:   %[[R4:.+]] = vector.shape_cast %[[R3]] : vector<16x16xf32> to vector<1x16x16xf32>
 //  CHECK-NEXT:  return %[[R4]] : vector<1x16x16xf32>
 
 #contraction_accesses0 = [
@@ -36,14 +36,14 @@ func.func @cast_away_contraction_leading_one_dims(%arg0: vector<1x16x8xf32>, %ar
 
 // CHECK-LABEL:   func.func @cast_away_contraction_leading_one_dim_under_const_mask
 // CHECK:           %[[MASK:.*]] = vector.constant_mask [15, 15, 8] : vector<16x16x8xi1>
-// CHECK:           %[[R0:.*]] = vector.extract %{{.*}}[0] : vector<16x8xf32> from vector<1x16x8xf32>
-// CHECK:           %[[R1:.*]] = vector.extract %{{.*}}[0] : vector<8x16xf32> from vector<1x8x16xf32>
-// CHECK:           %[[R2:.*]] = vector.extract %{{.*}}[0] : vector<16x16xf32> from vector<1x16x16xf32>
+// CHECK:           %[[R0:.*]] = vector.shape_cast %{{.*}} : vector<1x16x8xf32> to vector<16x8xf32>
+// CHECK:           %[[R1:.*]] = vector.shape_cast %{{.*}} : vector<1x8x16xf32> to vector<8x16xf32>
+// CHECK:           %[[R2:.*]] = vector.shape_cast %{{.*}} : vector<1x16x16xf32> to vector<16x16xf32>
 // CHECK:           %[[CONTRACT:.*]] = vector.mask %[[MASK]] {
 // CHECK-SAME:        vector.contract {indexing_maps = [#[[$MAP_0]], #[[$MAP_1]], #[[$MAP_2]]], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>}
 // CHECK-SAME:          %[[R0]], %[[R1]], %[[R2]] : vector<16x8xf32>, vector<8x16xf32> into vector<16x16xf32>
 // CHECK-SAME:      } : vector<16x16x8xi1> -> vector<16x16xf32>
-// CHECK:           %[[RES:.*]] = vector.broadcast %[[CONTRACT]] : vector<16x16xf32> to vector<1x16x16xf32>
+// CHECK:           %[[RES:.*]] = vector.shape_cast %[[CONTRACT]] : vector<16x16xf32> to vector<1x16x16xf32>
 // CHECK:           return %[[RES]] : vector<1x16x16xf32>
 
 #contraction_accesses0 = [
@@ -70,15 +70,15 @@ func.func @cast_away_contraction_leading_one_dim_under_const_mask(%arg0: vector<
 // CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0, d1, d2) -> (d0, d1)>
 
 // CHECK-LABEL:   func.func @cast_away_contraction_leading_one_dim_under_mask
-// CHECK:           %[[R0:.*]] = vector.extract %{{.*}} : vector<16x8xf32> from vector<1x16x8xf32>
-// CHECK:           %[[R1:.*]] = vector.extract %{{.*}} : vector<8x16xf32> from vector<1x8x16xf32>
-// CHECK:           %[[R2:.*]] = vector.extract %{{.*}} : vector<16x16xf32> from vector<1x16x16xf32>
-// CHECK:           %[[M:.*]] = vector.extract %{{.*}} : vector<16x16x8xi1> from vector<1x16x16x8xi1>
+// CHECK:           %[[R0:.*]] = vector.shape_cast %{{.*}} : vector<1x16x8xf32> to vector<16x8xf32>
+// CHECK:           %[[R1:.*]] = vector.shape_cast %{{.*}} : vector<1x8x16xf32> to vector<8x16xf32>
+// CHECK:           %[[R2:.*]] = vector.shape_cast %{{.*}} : vector<1x16x16xf32> to vector<16x16xf32>
+// CHECK:           %[[M:.*]] = vector.shape_cast %{{.*}} : vector<1x16x16x8xi1> to vector<16x16x8xi1>
 // CHECK:           %[[CONTRACT:.*]] = vector.mask %[[M]] {
 // CHECK-SAME:      vector.contract {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP2]]], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>}
 // CHECK-SAME:          %[[R0]], %[[R1]], %[[R2]] : vector<16x8xf32>, vector<8x16xf32> into vector<16x16xf32>
 // CHECK-SAME:      } : vector<16x16x8xi1> -> vector<16x16xf32>
-// CHECK-NEXT:      %[[RES:.*]] = vector.broadcast %[[CONTRACT]] : vector<16x16xf32> to vector<1x16x16xf32>
+// CHECK-NEXT:      %[[RES:.*]] = vector.shape_cast %[[CONTRACT]] : vector<16x16xf32> to vector<1x16x16xf32>
 // CHECK-NEXT:      return %[[RES]] : vector<1x16x16xf32>
 
 #contraction_accesses0 = [
@@ -109,15 +109,14 @@ func.func @cast_away_contraction_leading_one_dim_under_mask(
 // CHECK-DAG: #[[$map2:.*]] = affine_map<(d0, d1) -> (d0)>
 
 // CHECK-LABEL: cast_away_contraction_leading_one_dims_transposeneeded
-//  CHECK-NEXT:   %[[R0:.+]] =  vector.extract %{{.*}}[0] : vector<8x16xf32> from vector<1x8x16xf32>
-//  CHECK-NEXT:   %[[R1:.+]] =  vector.extract %{{.*}}[0, 0] : vector<8xf32> from vector<1x1x8xf32>
-//  CHECK-NEXT:   %[[R2:.+]] =  vector.extract %{{.*}}[0, 0] : vector<16xf32> from vector<1x1x16xf32>
+//  CHECK-NEXT:   %[[R0:.+]] = vector.shape_cast %{{.*}} : vector<1x8x16xf32> to vector<8x16xf32>
+//  CHECK-NEXT:   %[[R1:.+]] = vector.shape_cast %{{.*}} : vector<1x1x8xf32> to vector<8xf32>
+//  CHECK-NEXT:   %[[R2:.+]] = vector.shape_cast %{{.*}} : vector<1x1x16xf32> to vector<16xf32>
 //  CHECK-NEXT:   %[[R3:.+]] = vector.contract {indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]],
 //  CHECK-SAME:   iterator_types = ["parallel", "reduction"], kind = #vector.kind<mul>}
 //  CHECK-SAME:   %[[R1]], %[[R0]], %[[R2]] : vector<8xf32>, vector<8x16xf32> into vector<16xf32>
-//  CHECK-NEXT:   %[[R4:.+]] = vector.broadcast %[[R3]] : vector<16xf32> to vector<1x16xf32>
-//  CHECK-NEXT:   %[[R5:.+]] = vector.broadcast %[[R4]] : vector<1x16xf32> to vector<1x1x16xf32>
-//  CHECK-NEXT:  return %[[R5]] : vector<1x1x16xf32>
+//  CHECK-NEXT:   %[[R4:.+]] = vector.shape_cast %[[R3]] : vector<16xf32> to vector<1x1x16xf32>
+//  CHECK-NEXT:  return %[[R4]] : vector<1x1x16xf32>
 
 #contraction_accesses1 = [
   affine_map<(l, i, j, k) -> (i, l, k)>,
@@ -141,15 +140,13 @@ func.func @cast_away_contraction_leading_one_dims_transposeneeded(%arg0: vector<
 // CHECK-DAG: #[[$map2:.*]] = affine_map<(d0, d1, d2) -> (d0, d1)>
 
 // CHECK-LABEL: cast_away_contraction_leading_one_dims_transposeneeded2
-//  CHECK-NEXT:   %[[R0:.+]] =  vector.transpose %{{.*}}[1, 0, 2] : vector<8x1x16xf32> to vector<1x8x16xf32>
-//  CHECK-NEXT:   %[[R1:.+]] =  vector.extract %[[R0]][0] : vector<8x16xf32> from vector<1x8x16xf32>
-//  CHECK-NEXT:   %[[R2:.+]] =  vector.transpose %{{.*}}[2, 0, 1] : vector<2x8x1xf32> to vector<1x2x8xf32>
-//  CHECK-NEXT:   %[[R3:.+]] =  vector.extract %[[R2]][0] : vector<2x8xf32> from vector<1x2x8xf32>
-//  CHECK-NEXT:   %[[R4:.+]] =  vector.extract %{{.*}}[0] : vector<2x16xf32> from vector<1x2x16xf32>
+//  CHECK-NEXT:   %[[R1:.+]] = vector.shape_cast %{{.*}} : vector<8x1x16xf32> to vector<8x16xf32>
+//  CHECK-NEXT:   %[[R3:.+]] = vector.shape_cast %{{.*}} : vector<2x8x1xf32> to vector<2x8xf32>
+//  CHECK-NEXT:   %[[R4:.+]] = vector.shape_cast %{{.*}} : vector<1x2x16xf32> to vector<2x16xf32>
 //  CHECK-NEXT:   %[[R5:.+]] = vector.contract {indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]],
 //  CHECK-SAME:   iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>}
 //  CHECK-SAME:   %[[R1]], %[[R3]], %[[R4]] : vector<8x16xf32>, vector<2x8xf32> into vector<2x16xf32>
-//  CHECK-NEXT:   %[[R6:.+]] = vector.broadcast %[[R5]] : vector<2x16xf32> to vector<1x2x16xf32>
+//  CHECK-NEXT:   %[[R6:.+]] = vector.shape_cast %[[R5]] : vector<2x16xf32> to vector<1x2x16xf32>
 //  CHECK-NEXT:  return %[[R6]] : vector<1x2x16xf32>
 
 #contraction_accesses2 = [
@@ -175,19 +172,14 @@ func.func @cast_away_contraction_leading_one_dims_transposeneeded2(%arg0: vector
 
 
 // CHECK-LABEL: cast_away_contraction_leading_one_dims_nonleadingunitdim_rank4
-//  CHECK-NEXT:   %[[R0:.+]] =  vector.extract %{{.*}}[0] : vector<8x1x16xf32> from vector<1x8x1x16xf32>
-//  CHECK-NEXT:   %[[R1:.+]] =  vector.extract %{{.*}}[0] : vector<2x8x1xf32> from vector<1x2x8x1xf32>
-//  CHECK-NEXT:   %[[R2:.+]] =  vector.transpose %[[R0]], [1, 0, 2] : vector<8x1x16xf32> to vector<1x8x16xf32>
-//  CHECK-NEXT:   %[[R3:.+]] =  vector.extract %[[R2]][0] : vector<8x16xf32> from vector<1x8x16xf32>
-//  CHECK-NEXT:   %[[R4:.+]] =  vector.transpose %[[R1]], [2, 0, 1] : vector<2x8x1xf32> to vector<1x2x8xf32>
-//  CHECK-NEXT:   %[[R5:.+]] =  vector.extract %[[R4]][0] : vector<2x8xf32> from vector<1x2x8xf32>
-//  CHECK-NEXT:   %[[R6:.+]] =  vector.extract %{{.*}}[0, 0] : vector<2x16xf32> from vector<1x1x2x16xf32>
+//  CHECK-NEXT:   %[[R3:.+]] =  vector.shape_cast %{{.*}} : vector<1x8x1x16xf32> to vector<8x16xf32>
+//  CHECK-NEXT:   %[[R5:.+]] =  vector.shape_cast %{{.*}} : vector<1x2x8x1xf32> to vector<2x8xf32>
+//  CHECK-NEXT:   %[[R6:.+]] =  vector.shape_cast %{{.*}} : vector<1x1x2x16xf32> to vector<2x16xf32>
 //  CHECK-NEXT:   %[[R7:.+]] =  vector.contract {indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]],
 //  CHECK-SAME:   iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>}
 //  CHECK-SAME:   %[[R3]], %[[R5]], %[[R6]] : vector<8x16xf32>, vector<2x8xf32> into vector<2x16xf32>
-//  CHECK-NEXT:   %[[R8:.+]] =  vector.broadcast %[[R7]] : vector<2x16xf32> to vector<1x2x16xf32>
-//  CHECK-NEXT:   %[[R9:.+]] =  vector.broadcast %[[R8]] : vector<1x2x16xf32> to vector<1x1x2x16xf32>
-//  CHECK-NEXT:  return %[[R9]] : vector<1x1x2x16xf32>
+//  CHECK-NEXT:   %[[R8:.+]] =  vector.shape_cast %[[R7]] : vector<2x16xf32> to vector<1x1x2x16xf32>
+//  CHECK-NEXT:  return %[[R8]] : vector<1x1x2x16xf32>
 
 #contraction_accesses2 = [
   affine_map<(m, l, i, j, k) -> (m, k, l, j)>,
@@ -211,17 +203,14 @@ func.func @cast_away_contraction_leading_one_dims_nonleadingunitdim_rank4(%arg0:
 // CHECK-DAG: #[[$map2:.*]] = affine_map<(d0, d1, d2) -> (d0, d1)>
 
 // CHECK-LABEL: cast_away_contraction_leading_one_dims_nonleadingunitdim_rank4_acctranspose
-//  CHECK-NEXT:   %[[R0:.+]] =  vector.transpose %{{.*}}, [2, 0, 1, 3] : vector<1x8x1x16xf32> to vector<1x1x8x16xf32>
-//  CHECK-NEXT:   %[[R1:.+]] =  vector.transpose %{{.*}}, [3, 0, 1, 2] : vector<1x2x8x1xf32> to vector<1x1x2x8xf32>
-//  CHECK-NEXT:   %[[R2:.+]] =  vector.extract %[[R0]][0, 0] : vector<8x16xf32> from vector<1x1x8x16xf32>
-//  CHECK-NEXT:   %[[R3:.+]] =  vector.extract %[[R1]][0, 0] : vector<2x8xf32> from vector<1x1x2x8xf32>
-//  CHECK-NEXT:   %[[R4:.+]] =  vector.extract %{{.*}}[0, 0] : vector<2x16xf32> from vector<1x1x2x16xf32>
+//  CHECK-NEXT:   %[[R2:.+]] =  vector.shape_cast %{{.*}} : vector<1x8x1x16xf32> to vector<8x16xf32>
+//  CHECK-NEXT:   %[[R3:.+]] =  vector.shape_cast %{{.*}} : vector<1x2x8x1xf32> to vector<2x8xf32>
+//  CHECK-NEXT:   %[[R4:.+]] =  vector.shape_cast %{{.*}} : vector<1x1x2x16xf32> to vector<2x16xf32>
 //  CHECK-NEXT:   %[[R5:.+]] =  vector.contract {indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]],
 //  CHECK-SAME:   iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>}
 //  CHECK-SAME:   %[[R2]], %[[R3]], %[[R4]] : vector<8x16xf32>, vector<2x8xf32> into vector<2x16xf32>
-//  CHECK-NEXT:   %[[R6:.+]] =  vector.broadcast %[[R5]] : vector<2x16xf32> to vector<1x2x16xf32>
-//  CHECK-NEXT:   %[[R7:.+]] =  vector.broadcast %[[R6]] : vector<1x2x16xf32> to vector<1x1x2x16xf32>
-//  CHECK-NEXT:  return %[[R7]] : vector<1x1x2x16xf32>
+//  CHECK-NEXT:   %[[R6:.+]] =  vector.shape_cast %[[R5]] : vector<2x16xf32> to vector<1x1x2x16xf32>
+//  CHECK-NEXT:  return %[[R6]] : vector<1x1x2x16xf32>
 
 #contraction_accesses3 = [
   affine_map<(m, l, i, j, k) -> (m, k, l, j)>,
@@ -256,7 +245,7 @@ func.func @cast_away_contraction_does_not_transpose_leading_unit_dims(%lhs: vect
 // CHECK-DAG: #[[$map_dp1:.*]] = affine_map<(d0) -> ()>
 
 // CHECK-LABEL: cast_away_contraction_leading_one_dims_to_dot_product
-//  CHECK-NEXT:   %[[R0:.+]] = vector.extract %{{.*}}[0] : vector<64xf32> from vector<1x64xf32>
+//  CHECK-NEXT:   %[[R0:.+]] = vector.shape_cast %{{.*}} : vector<1x64xf32> to vector<64xf32>
 //  CHECK-NEXT:   %[[R1:.+]] = vector.extract %{{.*}}[0] : f32 from vector<1xf32>
 //  CHECK-NEXT:   %[[R2:.+]] = vector.contract {indexing_maps = [#[[$map_dp0]], #[[$map_dp0]], #[[$map_dp1]]],
 //  CHECK-SAME:   iterator_types = ["reduction"], kind = #vector.kind<add>}
@@ -269,45 +258,82 @@ func.func @cast_away_contraction_leading_one_dims_to_dot_product(%arg0: vector<6
   return %0 : vector<1xf32>
 }
 
+// -----
+
+// CHECK-DAG: #[[$DOT_MAP:.*]] = affine_map<(d0) -> (d0)>
+// CHECK-DAG: #[[$SCALAR_MAP:.*]] = affine_map<(d0) -> ()>
+
+// CHECK-LABEL: cast_away_masked_contraction_with_rank1_acc
+//  CHECK-NEXT:   %[[RHS:.+]] = vector.shape_cast %{{.*}} : vector<1x64xf32> to vector<64xf32>
+//  CHECK-NEXT:   %[[ACC:.+]] = vector.extract %{{.*}}[0] : f32 from vector<1xf32>
+//  CHECK-NEXT:   %[[MASK:.+]] = vector.shape_cast %{{.*}} : vector<64x1xi1> to vector<64xi1>
+//  CHECK-NEXT:   %[[DOT:.+]] = vector.mask %[[MASK]] {
+//  CHECK-SAME:     vector.contract {indexing_maps = [#[[$DOT_MAP]], #[[$DOT_MAP]], #[[$SCALAR_MAP]]], iterator_types = ["reduction"], kind = #vector.kind<add>}
+//  CHECK-SAME:     %{{.*}}, %[[RHS]], %[[ACC]] : vector<64xf32>, vector<64xf32> into f32
+//  CHECK-SAME:   } : vector<64xi1> -> f32
+//  CHECK-NEXT:   %[[RES:.+]] = vector.broadcast %[[DOT]] : f32 to vector<1xf32>
+//  CHECK-NEXT:   return %[[RES]] : vector<1xf32>
+
+func.func @cast_away_masked_contraction_with_rank1_acc(%arg0: vector<64xf32>, %arg1: vector<1x64xf32>, %arg2: vector<1xf32>, %mask: vector<64x1xi1>) -> vector<1xf32> {
+  %0 = vector.mask %mask {
+    vector.contract {indexing_maps = [affine_map<(d0, d1) -> (d0)>, affine_map<(d0, d1) -> (d1, d0)>, affine_map<(d0, d1) -> (d1)>], iterator_types = ["reduction", "parallel"], kind = #vector.kind<add>} %arg0, %arg1, %arg2 : vector<64xf32>, vector<1x64xf32> into vector<1xf32>
+  } : vector<64x1xi1> -> vector<1xf32>
+  return %0 : vector<1xf32>
+}
+
+// -----
+
+// CHECK-LABEL: do_not_cast_away_contraction_with_scalable_rank1_acc
+//  CHECK-NOT: vector.shape_cast
+//  CHECK-NOT: vector.extract
+//  CHECK-NOT: vector.broadcast
+//  CHECK-NEXT: vector.contract
+//  CHECK-NEXT: return
+
+func.func @do_not_cast_away_contraction_with_scalable_rank1_acc(%arg0: vector<64xf32>, %arg1: vector<[1]x64xf32>, %arg2: vector<[1]xf32>) -> vector<[1]xf32> {
+  %0 = vector.contract {indexing_maps = [affine_map<(d0, d1) -> (d0)>, affine_map<(d0, d1) -> (d1, d0)>, affine_map<(d0, d1) -> (d1)>], iterator_types = ["reduction", "parallel"], kind = #vector.kind<add>} %arg0, %arg1, %arg2 : vector<64xf32>, vector<[1]x64xf32> into vector<[1]xf32>
+  return %0 : vector<[1]xf32>
+}
+
 // -----
 // CHECK-LABEL: func @cast_away_extract_strided_slice_leading_one_dims
 func.func @cast_away_extract_strided_slice_leading_one_dims(%arg0: vector<1x8x8xf16>) -> vector<1x1x8xf16> {
-  // CHECK:     %[[SRC:.+]] = vector.extract %{{.*}}[0] : vector<8x8xf16> from vector<1x8x8xf16>
+  // CHECK:     %[[SRC:.+]] = vector.shape_cast %{{.*}} : vector<1x8x8xf16> to vector<8x8xf16>
   // CHECK: %[[EXTRACT:.+]] = vector.extract_strided_slice %[[SRC]] {offsets = [4], sizes = [1], strides = [1]} : vector<8x8xf16> to vector<1x8xf16>
   %0 = vector.extract_strided_slice %arg0 {offsets = [0, 4], sizes = [1, 1], strides = [1, 1]} : vector<1x8x8xf16> to vector<1x1x8xf16>
-  // CHECK:     %[[RET:.+]] = vector.broadcast %[[EXTRACT]] : vector<1x8xf16> to vector<1x1x8xf16>
+  // CHECK:     %[[RET:.+]] = vector.shape_cast %[[EXTRACT]] : vector<1x8xf16> to vector<1x1x8xf16>
   // CHECK: return %[[RET]]
   return %0: vector<1x1x8xf16>
 }
 
 // CHECK-LABEL: func @cast_away_extract_strided_slice_leading_one_dims_scalable
 func.func @cast_away_extract_strided_slice_leading_one_dims_scalable(%arg0: vector<1x8x[8]xf16>) -> vector<1x1x[8]xf16> {
-  // CHECK:     %[[SRC:.+]] = vector.extract %{{.*}}[0] : vector<8x[8]xf16> from vector<1x8x[8]xf16>
+  // CHECK:     %[[SRC:.+]] = vector.shape_cast %{{.*}} : vector<1x8x[8]xf16> to vector<8x[8]xf16>
   // CHECK: %[[EXTRACT:.+]] = vector.extract_strided_slice %[[SRC]] {offsets = [4], sizes = [1], strides = [1]} : vector<8x[8]xf16> to vector<1x[8]xf16>
   %0 = vector.extract_strided_slice %arg0 {offsets = [0, 4], sizes = [1, 1], strides = [1, 1]} : vector<1x8x[8]xf16> to vector<1x1x[8]xf16>
-  // CHECK:     %[[RET:.+]] = vector.broadcast %[[EXTRACT]] : vector<1x[8]xf16> to vector<1x1x[8]xf16>
+  // CHECK:     %[[RET:.+]] = vector.shape_cast %[[EXTRACT]] : vector<1x[8]xf16> to vector<1x1x[8]xf16>
   // CHECK: return %[[RET]]
   return %0: vector<1x1x[8]xf16>
 }
 
 // CHECK-LABEL: func @cast_away_insert_strided_slice_leading_one_dims
 func.func @cast_away_insert_strided_slice_leading_one_dims(%arg0: vector<1x8xf16>, %arg1: vector<1x8x8xf16>) -> vector<1x8x8xf16> {
-  // CHECK:    %[[SRC:.+]] = vector.extract %{{.*}}[0] : vector<8xf16> from vector<1x8xf16>
-  // CHECK:    %[[DST:.+]] = vector.extract %{{.*}}[0] : vector<8x8xf16> from vector<1x8x8xf16>
+  // CHECK:    %[[SRC:.+]] = vector.shape_cast %{{.*}} : vector<1x8xf16> to vector<8xf16>
+  // CHECK:    %[[DST:.+]] = vector.shape_cast %{{.*}} : vector<1x8x8xf16> to vector<8x8xf16>
   // CHECK: %[[INSERT:.+]] = vector.insert_strided_slice %[[SRC]], %[[DST]] {offsets = [0, 0], strides = [1]} : vector<8xf16> into vector<8x8xf16>
   %0 = vector.insert_strided_slice %arg0, %arg1 {offsets = [0, 0, 0], strides = [1, 1]} : vector<1x8xf16> into vector<1x8x8xf16>
-  // CHECK:    %[[RET:.+]] = vector.broadcast %[[INSERT]] : vector<8x8xf16> to vector<1x8x8xf16>
+  // CHECK:    %[[RET:.+]] = vector.shape_cast %[[INSERT]] : vector<8x8xf16> to vector<1x8x8xf16>
   // CHECK: return %[[RET]]
   return %0: vector<1x8x8xf16>
 }
 
 // CHECK-LABEL: func @cast_away_insert_strided_slice_leading_one_dims_scalable
 func.func @cast_away_insert_strided_slice_leading_one_dims_scalable(%arg0: vector<1x[8]xf16>, %arg1: vector<1x8x[8]xf16>) -> vector<1x8x[8]xf16> {
-  // CHECK:    %[[SRC:.+]] = vector.extract %{{.*}}[0] : vector<[8]xf16> from vector<1x[8]xf16>
-  // CHECK:    %[[DST:.+]] = vector.extract %{{.*}}[0] : vector<8x[8]xf16> from vector<1x8x[8]xf16>
+  // CHECK:    %[[SRC:.+]] = vector.shape_cast %{{.*}} : vector<1x[8]xf16> to vector<[8]xf16>
+  // CHECK:    %[[DST:.+]] = vector.shape_cast %{{.*}} : vector<1x8x[8]xf16> to vector<8x[8]xf16>
   // CHECK: %[[INSERT:.+]] = vector.insert_strided_slice %[[SRC]], %[[DST]] {offsets = [0, 0], strides = [1]} : vector<[8]xf16> into vector<8x[8]xf16>
   %0 = vector.insert_strided_slice %arg0, %arg1 {offsets = [0, 0, 0], strides = [1, 1]} : vector<1x[8]xf16> into vector<1x8x[8]xf16>
-  // CHECK:    %[[RET:.+]] = vector.broadcast %[[INSERT]] : vector<8x[8]xf16> to vector<1x8x[8]xf16>
+  // CHECK:    %[[RET:.+]] = vector.shape_cast %[[INSERT]] : vector<8x[8]xf16> to vector<1x8x[8]xf16>
   // CHECK: return %[[RET]]
   return %0: vector<1x8x[8]xf16>
 }
@@ -315,8 +341,7 @@ func.func @cast_away_insert_strided_slice_leading_one_dims_scalable(%arg0: vecto
 // CHECK-LABEL: func @cast_away_insert_strided_slice_leading_one_dims_one_element
 //  CHECK-SAME: %[[ARG0:.+]]: vector<1x1xf16>, %{{.+}}: vector<1x1x1xf16>
 func.func @cast_away_insert_strided_slice_leading_one_dims_one_element(%arg0: vector<1x1xf16>, %arg1: vector<1x1x1xf16>) -> vector<1x1x1xf16> {
-  // CHECK: %[[EXT:.+]] = vector.extract %{{.*}}[0] : vector<1xf16> from vector<1x1xf16>
-  // CHECK: %[[B:.+]] = vector.broadcast %[[EXT]] : vector<1xf16> to vector<1x1x1xf16>
+  // CHECK: %[[B:.+]] = vector.shape_cast %{{.*}} : vector<1x1xf16> to vector<1x1x1xf16>
   %0 = vector.insert_strided_slice %arg0, %arg1 {offsets = [0, 0, 0], strides = [1, 1]} : vector<1x1xf16> into vector<1x1x1xf16>
   // CHECK: return %[[B]]
   return %0: vector<1x1x1xf16>
@@ -325,8 +350,7 @@ func.func @cast_away_insert_strided_slice_leading_one_dims_one_element(%arg0: ve
 // CHECK-LABEL: func @cast_away_insert_strided_slice_leading_one_dims_one_element_scalable
 //  CHECK-SAME: %[[ARG0:.+]]: vector<1x[1]xf16>, %{{.+}}: vector<1x1x[1]xf16>
 func.func @cast_away_insert_strided_slice_leading_one_dims_one_element_scalable(%arg0: vector<1x[1]xf16>, %arg1: vector<1x1x[1]xf16>) -> vector<1x1x[1]xf16> {
-  // CHECK: %[[EXT:.+]] = vector.extract %{{.*}}[0] : vector<[1]xf16> from vector<1x[1]xf16>
-  // CHECK: %[[B:.+]] = vector.broadcast %[[EXT]] : vector<[1]xf16> to vector<1x1x[1]xf16>
+  // CHECK: %[[B:.+]] = vector.shape_cast %{{.*}} : vector<1x[1]xf16> to vector<1x1x[1]xf16>
   %0 = vector.insert_strided_slice %arg0, %arg1 {offsets = [0, 0, 0], strides = [1, 1]} : vector<1x[1]xf16> into vector<1x1x[1]xf16>
   // CHECK: return %[[B]]
   return %0: vector<1x1x[1]xf16>
@@ -339,7 +363,7 @@ func.func @cast_away_transfer_read_leading_one_dims(%arg0: memref<1x4x8x16xf16>)
   // CHECK: %[[F0:.+]] = arith.constant 0.000000e+00 : f16
   %f0 = arith.constant 0. : f16
   // CHECK: %[[READ:.+]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]], %[[C0]], %[[C0]]], %[[F0]] {in_bounds = [true]} : memref<1x4x8x16xf16>, vector<4xf16>
-  // CHECK: %[[CAST:.+]] = vector.broadcast %[[READ]] : vector<4xf16> to vector<1x4xf16>
+  // CHECK: %[[CAST:.+]] = vector.shape_cast %[[READ]] : vector<4xf16> to vector<1x4xf16>
   %0 = vector.transfer_read %arg0[%c0, %c0, %c0, %c0], %f0 {in_bounds = [true, true]} : memref<1x4x8x16xf16>, vector<1x4xf16>
   // CHECK: return %[[CAST]]
   return %0: vector<1x4xf16>
@@ -351,9 +375,9 @@ func.func @cast_away_masked_transfer_read_leading_one_dims(%arg0: memref<1x4x8x1
   %c0 = arith.constant 0 : index
   // CHECK: %[[F0:.+]] = arith.constant 0.000000e+00 : f16
   %f0 = arith.constant 0. : f16
-  // CHECK: %[[MASK_CAST:.+]] = vector.extract %{{.*}}[0] : vector<4xi1> from vector<1x4xi1>
+  // CHECK: %[[MASK_CAST:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi1> to vector<4xi1>
   // CHECK: %[[READ:.+]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]], %[[C0]], %[[C0]]], %[[F0]], %[[MASK_CAST]] {in_bounds = [true]} : memref<1x4x8x16xf16>, vector<4xf16>
-  // CHECK: %[[CAST:.+]] = vector.broadcast %[[READ]] : vector<4xf16> to vector<1x4xf16>
+  // CHECK: %[[CAST:.+]] = vector.shape_cast %[[READ]] : vector<4xf16> to vector<1x4xf16>
   %0 = vector.transfer_read %arg0[%c0, %c0, %c0, %c0], %f0, %arg1 {in_bounds = [true, true]} : memref<1x4x8x16xf16>, vector<1x4xf16>
   // CHECK: return %[[CAST]]
   return %0: vector<1x4xf16>
@@ -363,7 +387,7 @@ func.func @cast_away_masked_transfer_read_leading_one_dims(%arg0: memref<1x4x8x1
 func.func @cast_away_transfer_read_leading_one_dims_one_element(%arg0: memref<1x1x1x1xf16>) -> vector<1x1xf16> {
   %c0 = arith.constant 0 : index
   %f0 = arith.constant 0. : f16
-  // CHECK: vector.broadcast %{{.+}} : vector<1xf16> to vector<1x1xf16>
+  // CHECK: vector.shape_cast %{{.+}} : vector<1xf16> to vector<1x1xf16>
   %0 = vector.transfer_read %arg0[%c0, %c0, %c0, %c0], %f0 {in_bounds = [true, true]} : memref<1x1x1x1xf16>, vector<1x1xf16>
   return %0: vector<1x1xf16>
 }
@@ -380,7 +404,7 @@ func.func @cast_away_nontrivial_map_masked_transfer_read(%arg0: memref<1x4x8xf16
   // CHECK: %[[MASK_CAST:.+]] = vector.shape_cast %{{.*}} : vector<1x4x1xi1> to vector<4xi1>
   // CHECK: %[[READ:.+]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]], %[[C0]]], %[[F0]], %[[MASK_CAST]] {in_bounds = [true]
   // CHECK-SAME: permutation_map = #[[$MAP]]} : memref<1x4x8xf16>, vector<4xf16>
-  // CHECK: %[[CAST:.+]] = vector.broadcast %[[READ]] : vector<4xf16> to vector<1x1x4xf16>
+  // CHECK: %[[CAST:.+]] = vector.shape_cast %[[READ]] : vector<4xf16> to vector<1x1x4xf16>
   %0 = vector.transfer_read %arg0[%c0, %c0, %c0], %f0, %arg1 {in_bounds = [true, true, true],
                             permutation_map = affine_map<(d0, d1, d2) -> (d0, d2, d1)>} : memref<1x4x8xf16>, vector<1x1x4xf16>
   // CHECK: return %[[CAST]]
@@ -391,7 +415,7 @@ func.func @cast_away_nontrivial_map_masked_transfer_read(%arg0: memref<1x4x8xf16
 
 // CHECK-LABEL: func @not_insert_cast_fo4_transfer_read_under_mask
 // CHECK:      %[[MASK:.+]] = vector.constant_mask
-// CHECK:      %[[CASTED_MASK:.+]] = vector.broadcast %[[MASK]]
+// CHECK:      %[[CASTED_MASK:.+]] = vector.shape_cast %[[MASK]]
 // CHECK:      %[[RET:.+]] = vector.mask %[[CASTED_MASK]] {
 // CHECK-SAME:   vector.transfer_read {{.*}} : memref<1x1x4xf16>, vector<1x4xf16> }
 // CHECK:      return %[[RET]] : vector<1x4xf16>
@@ -411,7 +435,7 @@ func.func @not_insert_cast_fo4_transfer_read_under_mask(%arg0: memref<1x1x4xf16>
 func.func @cast_away_transfer_write_leading_one_dims(%arg0: memref<1x4x8x16xf16>, %arg1: vector<1x4xf16>) {
   // CHECK: %[[C0:.+]] = arith.constant 0 : index
   %c0 = arith.constant 0 : index
-  // CHECK: %[[CAST:.+]] = vector.extract %{{.*}}[0] : vector<4xf16> from vector<1x4xf16>
+  // CHECK: %[[CAST:.+]] = vector.shape_cast %{{.*}} : vector<1x4xf16> to vector<4xf16>
   // CHECK: vector.transfer_write %[[CAST]], %{{.*}}[%[[C0]], %[[C0]], %[[C0]], %[[C0]]] {in_bounds = [true]} : vector<4xf16>, memref<1x4x8x16xf16>
 
   vector.transfer_write %arg1, %arg0[%c0, %c0, %c0, %c0] {in_bounds = [true, true]} : vector<1x4xf16>, memref<1x4x8x16xf16>
@@ -422,8 +446,8 @@ func.func @cast_away_transfer_write_leading_one_dims(%arg0: memref<1x4x8x16xf16>
 func.func @cast_away_masked_transfer_write_leading_one_dims(%arg0: memref<1x4x8x16xf16>, %arg1: vector<1x4xf16>, %arg2: vector<1x4xi1>) {
   // CHECK: %[[C0:.+]] = arith.constant 0 : index
   %c0 = arith.constant 0 : index
-  // CHECK: %[[CAST:.+]] = vector.extract %{{.*}}[0] : vector<4xf16> from vector<1x4xf16>
-  // CHECK: %[[MASK_CAST:.+]] = vector.extract %{{.*}}[0] : vector<4xi1> from vector<1x4xi1>
+  // CHECK: %[[CAST:.+]] = vector.shape_cast %{{.*}} : vector<1x4xf16> to vector<4xf16>
+  // CHECK: %[[MASK_CAST:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi1> to vector<4xi1>
   // CHECK: vector.transfer_write %[[CAST]], %{{.*}}[%[[C0]], %[[C0]], %[[C0]], %[[C0]]], %[[MASK_CAST]] {in_bounds = [true]} : vector<4xf16>, memref<1x4x8x16xf16>
 
   vector.transfer_write %arg1, %arg0[%c0, %c0, %c0, %c0], %arg2 {in_bounds = [true, true]} : vector<1x4xf16>, memref<1x4x8x16xf16>
@@ -433,7 +457,7 @@ func.func @cast_away_masked_transfer_write_leading_one_dims(%arg0: memref<1x4x8x
 // CHECK-LABEL: func @cast_away_transfer_write_leading_one_dims_one_element
 func.func @cast_away_transfer_write_leading_one_dims_one_element(%arg0: memref<1x1x1x1xf16>, %arg1: vector<1x1xf16>) {
   %c0 = arith.constant 0 : index
-  // CHECK: vector.extract %{{.+}}[0] : vector<1xf16> from vector<1x1xf16>
+  // CHECK: vector.shape_cast %{{.+}} : vector<1x1xf16> to vector<1xf16>
   vector.transfer_write %arg1, %arg0[%c0, %c0, %c0, %c0] {in_bounds = [true, true]} : vector<1x1xf16>, memref<1x1x1x1xf16>
   return
 }
@@ -442,7 +466,7 @@ func.func @cast_away_transfer_write_leading_one_dims_one_element(%arg0: memref<1
 
 // CHECK-LABEL: func @not_insert_cast_for_transfer_write_under_mask
 // CHECK:      %[[MASK:.+]] = vector.constant_mask
-// CHECK:      %[[CASTED_MASK:.+]] = vector.broadcast %[[MASK]]
+// CHECK:      %[[CASTED_MASK:.+]] = vector.shape_cast %[[MASK]]
 // CHECK:      vector.mask %[[CASTED_MASK]] {
 // CHECK-SAME:   vector.transfer_write {{.*}} : vector<1x4xf16>, memref<1x1x4xf16> }
 // CHECK:      return
@@ -462,7 +486,7 @@ func.func @not_insert_cast_for_transfer_write_under_mask(%arg0: memref<1x1x4xf16
 func.func @cast_away_nontrivial_map_masked_transfer_write(%arg0: memref<1x4x8xf16>, %arg1: vector<1x1x4xf16>, %arg2: vector<1x4x1xi1>) {
   // CHECK: %[[C0:.+]] = arith.constant 0 : index
   %c0 = arith.constant 0 : index
-  // CHECK: %[[CAST:.+]] = vector.extract %{{.*}}[0, 0] : vector<4xf16> from vector<1x1x4xf16>
+  // CHECK: %[[CAST:.+]] = vector.shape_cast %{{.*}} : vector<1x1x4xf16> to vector<4xf16>
   // CHECK: %[[MASK_CAST:.+]] = vector.shape_cast %{{.*}} : vector<1x4x1xi1> to vector<4xi1>
   // CHECK: vector.transfer_write %[[CAST]], %{{.*}}[%[[C0]], %[[C0]], %[[C0]]], %[[MASK_CAST]] {in_bounds = [true]
   // CHECK-SAME: permutation_map = #[[$MAP]]} : vector<4xf16>, memref<1x4x8xf16>
@@ -479,25 +503,25 @@ func.func @cast_away_elementwise_leading_one_dims(
   %arg0: vector<1x1x8xf32>, %arg1: f32, %arg2: vector<1x4xf32>,
   %arg3: vector<1x4xf32>, %arg4: i1) ->
   (vector<1x1x8xf32>, vector<1x4xi1>, vector<1x4xf32>, vector<1x4xf32>) {
-  // CHECK:  vector.extract %{{.*}}[0, 0] : vector<8xf32> from vector<1x1x8xf32>
-  // CHECK:  vector.extract %{{.*}}[0, 0] : vector<8xf32> from vector<1x1x8xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<1x1x8xf32> to vector<8xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<1x1x8xf32> to vector<8xf32>
   // CHECK:  arith.addf %{{.*}}, %{{.*}} : vector<8xf32>
-  // CHECK:  vector.broadcast %{{.*}} : vector<8xf32> to vector<1x1x8xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<8xf32> to vector<1x1x8xf32>
   %0 = arith.addf %arg0, %arg0 : vector<1x1x8xf32>
-  // CHECK:  vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
-  // CHECK:  vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
   // CHECK:  arith.cmpf ogt, %{{.*}}, %{{.*}} : vector<4xf32>
-  // CHECK:  vector.broadcast %{{.*}} : vector<4xi1> to vector<1x4xi1>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<4xi1> to vector<1x4xi1>
   %1 = arith.cmpf ogt, %arg2, %arg3 : vector<1x4xf32>
-  // CHECK:  vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
-  // CHECK:  vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
   // CHECK:  select %{{.*}}, %{{.*}}, %{{.*}} : vector<4xi1>, vector<4xf32>
-  // CHECK:  vector.broadcast %{{.*}} : vector<4xf32> to vector<1x4xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<4xf32> to vector<1x4xf32>
   %2 = arith.select %1, %arg3, %arg2 : vector<1x4xi1>, vector<1x4xf32>
-  // CHECK:  vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
-  // CHECK:  vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
   // CHECK:  select %arg4, %12, %{{.*}} : vector<4xf32>
-  // CHECK:  vector.broadcast %{{.*}} : vector<4xf32> to vector<1x4xf32>
+  // CHECK:  vector.shape_cast %{{.*}} : vector<4xf32> to vector<1x4xf32>
   %3 = arith.select %arg4, %arg3, %arg2 : vector<1x4xf32>
   return %0, %1, %2, %3: vector<1x1x8xf32>, vector<1x4xi1>, vector<1x4xf32>, vector<1x4xf32>
 }
@@ -506,10 +530,10 @@ func.func @cast_away_elementwise_leading_one_dims(
 
 // CHECK-LABEL: func @cast_away_insert_leading_one_dims_scalar
 //  CHECK-SAME: (%[[S:.+]]: f32, %[[V:.+]]: vector<1x1x4xf32>)
-//       CHECK:   %[[EXTRACT:.+]] = vector.extract %[[V]][0, 0] : vector<4xf32> from vector<1x1x4xf32>
-//       CHECK:   %[[INSERT:.+]] = vector.insert %[[S]], %[[EXTRACT]] [0] : f32 into vector<4xf32>
-//       CHECK:   %[[BCAST:.+]] = vector.broadcast %[[INSERT]] : vector<4xf32> to vector<1x1x4xf32>
-//       CHECK:   return %[[BCAST]]
+//       CHECK:   %[[DST_CAST:.+]] = vector.shape_cast %[[V]] : vector<1x1x4xf32> to vector<4xf32>
+//       CHECK:   %[[INSERT:.+]] = vector.insert %[[S]], %[[DST_CAST]] [0] : f32 into vector<4xf32>
+//       CHECK:   %[[RESULT_CAST:.+]] = vector.shape_cast %[[INSERT]] : vector<4xf32> to vector<1x1x4xf32>
+//       CHECK:   return %[[RESULT_CAST]]
 func.func @cast_away_insert_leading_one_dims_scalar(%s: f32, %v: vector<1x1x4xf32>) -> vector<1x1x4xf32> {
   %0 = vector.insert %s, %v [0, 0, 0] : f32 into vector<1x1x4xf32>
   return %0: vector<1x1x4xf32>
@@ -521,10 +545,10 @@ func.func @cast_away_insert_leading_one_dims_scalar(%s: f32, %v: vector<1x1x4xf3
 // CHECK-SAME:    %[[S:.*]]: f32,
 // CHECK-SAME:    %[[V:.*]]: vector<1x1x[4]xf32>) -> vector<1x1x[4]xf32> {
 func.func @cast_away_insert_leading_one_dims_scalar_scalable(%s: f32, %v: vector<1x1x[4]xf32>) -> vector<1x1x[4]xf32> {
-// CHECK:           %[[EXTRACT:.*]] = vector.extract %[[V]][0, 0] : vector<[4]xf32> from vector<1x1x[4]xf32>
-// CHECK:           %[[INSERT:.*]] = vector.insert %[[S]], %[[EXTRACT]] [0] : f32 into vector<[4]xf32>
-// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[INSERT]] : vector<[4]xf32> to vector<1x1x[4]xf32>
-// CHECK:           return %[[BCAST]] : vector<1x1x[4]xf32>
+// CHECK:           %[[DST_CAST:.*]] = vector.shape_cast %[[V]] : vector<1x1x[4]xf32> to vector<[4]xf32>
+// CHECK:           %[[INSERT:.*]] = vector.insert %[[S]], %[[DST_CAST]] [0] : f32 into vector<[4]xf32>
+// CHECK:           %[[RESULT_CAST:.*]] = vector.shape_cast %[[INSERT]] : vector<[4]xf32> to vector<1x1x[4]xf32>
+// CHECK:           return %[[RESULT_CAST]] : vector<1x1x[4]xf32>
   %0 = vector.insert %s, %v [0, 0, 0] : f32 into vector<1x1x[4]xf32>
   return %0: vector<1x1x[4]xf32>
 }
@@ -535,10 +559,10 @@ func.func @cast_away_insert_leading_one_dims_scalar_scalable(%s: f32, %v: vector
 // CHECK-SAME:    %[[S:.*]]: f32,
 // CHECK-SAME:    %[[V:.*]]: vector<1x[1]x4xf32>) -> vector<1x[1]x4xf32> {
 func.func @cast_away_insert_leading_one_dims_scalar_skip_scalable_dim(%s: f32, %v: vector<1x[1]x4xf32>) -> vector<1x[1]x4xf32> {
-// CHECK:           %[[EXTRACT:.*]] = vector.extract %[[V]][0] : vector<[1]x4xf32> from vector<1x[1]x4xf32>
-// CHECK:           %[[INSERT:.*]] = vector.insert %[[S]], %[[EXTRACT]] [0, 0] : f32 into vector<[1]x4xf32>
-// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[INSERT]] : vector<[1]x4xf32> to vector<1x[1]x4xf32>
-// CHECK:           return %[[BCAST]] : vector<1x[1]x4xf32>
+// CHECK:           %[[DST_CAST:.*]] = vector.shape_cast %[[V]] : vector<1x[1]x4xf32> to vector<[1]x4xf32>
+// CHECK:           %[[INSERT:.*]] = vector.insert %[[S]], %[[DST_CAST]] [0, 0] : f32 into vector<[1]x4xf32>
+// CHECK:           %[[RESULT_CAST:.*]] = vector.shape_cast %[[INSERT]] : vector<[1]x4xf32> to vector<1x[1]x4xf32>
+// CHECK:           return %[[RESULT_CAST]] : vector<1x[1]x4xf32>
   %0 = vector.insert %s, %v [0, 0, 0] : f32 into vector<1x[1]x4xf32>
   return %0: vector<1x[1]x4xf32>
 }
@@ -547,8 +571,8 @@ func.func @cast_away_insert_leading_one_dims_scalar_skip_scalable_dim(%s: f32, %
 
 // CHECK-LABEL: func @cast_away_insert_leading_one_dims_rank1
 //  CHECK-SAME: (%[[S:.+]]: vector<4xf32>, %[[V:.+]]: vector<1x1x4xf32>)
-//       CHECK:   %[[BCAST:.+]] = vector.broadcast %[[S]] : vector<4xf32> to vector<1x1x4xf32>
-//       CHECK:   return %[[BCAST]]
+//       CHECK:   %[[RESULT_CAST:.+]] = vector.shape_cast %[[S]] : vector<4xf32> to vector<1x1x4xf32>
+//       CHECK:   return %[[RESULT_CAST]]
 func.func @cast_away_insert_leading_one_dims_rank1(%s: vector<4xf32>, %v: vector<1x1x4xf32>) -> vector<1x1x4xf32> {
   %0 = vector.insert %s, %v [0, 0] : vector<4xf32> into vector<1x1x4xf32>
   return %0: vector<1x1x4xf32>
@@ -559,8 +583,8 @@ func.func @cast_away_insert_leading_one_dims_rank1(%s: vector<4xf32>, %v: vector
 // CHECK-LABEL:   func.func @cast_away_insert_leading_one_dims_rank1_scalable(
 // CHECK-SAME:    %[[S:.*]]: vector<[4]xf32>,
 // CHECK-SAME:    %[[V:.*]]: vector<1x1x[4]xf32>) -> vector<1x1x[4]xf32> {
-// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[S]] : vector<[4]xf32> to vector<1x1x[4]xf32>
-// CHECK:           return %[[BCAST]] : vector<1x1x[4]xf32>
+// CHECK:           %[[RESULT_CAST:.*]] = vector.shape_cast %[[S]] : vector<[4]xf32> to vector<1x1x[4]xf32>
+// CHECK:           return %[[RESULT_CAST]] : vector<1x1x[4]xf32>
 func.func @cast_away_insert_leading_one_dims_rank1_scalable(%s: vector<[4]xf32>, %v: vector<1x1x[4]xf32>) -> vector<1x1x[4]xf32> {
   %0 = vector.insert %s, %v [0, 0] : vector<[4]xf32> into vector<1x1x[4]xf32>
   return %0: vector<1x1x[4]xf32>
@@ -570,9 +594,8 @@ func.func @cast_away_insert_leading_one_dims_rank1_scalable(%s: vector<[4]xf32>,
 
 // CHECK-LABEL: func @cast_away_insert_leading_one_dims_rank2
 //  CHECK-SAME: (%[[S:.+]]: vector<1x4xf32>, %[[V:.+]]: vector<1x1x4xf32>)
-//       CHECK:   %[[EXTRACT:.+]] = vector.extract %[[S]][0] : vector<4xf32> from vector<1x4xf32>
-//       CHECK:   %[[BCAST:.+]] = vector.broadcast %[[EXTRACT]] : vector<4xf32> to vector<1x1x4xf32>
-//       CHECK:   return %[[BCAST]]
+//       CHECK:   %[[SRC_CAST:.+]] = vector.shape_cast %[[S]] : vector<1x4xf32> to vector<1x1x4xf32>
+//       CHECK:   return %[[SRC_CAST]]
 func.func @cast_away_insert_leading_one_dims_rank2(%s: vector<1x4xf32>, %v: vector<1x1x4xf32>) -> vector<1x1x4xf32> {
   %0 = vector.insert %s, %v [0] : vector<1x4xf32> into vector<1x1x4xf32>
   return %0: vector<1x1x4xf32>
@@ -583,9 +606,8 @@ func.func @cast_away_insert_leading_one_dims_rank2(%s: vector<1x4xf32>, %v: vect
 // CHECK-LABEL:   func.func @cast_away_insert_leading_one_dims_rank2_scalable(
 // CHECK-SAME:    %[[S:.*]]: vector<1x[4]xf32>,
 // CHECK-SAME:    %[[V:.*]]: vector<1x1x[4]xf32>) -> vector<1x1x[4]xf32> {
-// CHECK:           %[[EXTRACT:.*]] = vector.extract %[[S]][0] : vector<[4]xf32> from vector<1x[4]xf32>
-// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[EXTRACT]] : vector<[4]xf32> to vector<1x1x[4]xf32>
-// CHECK:           return %[[BCAST]] : vector<1x1x[4]xf32>
+// CHECK:           %[[SRC_CAST:.*]] = vector.shape_cast %[[S]] : vector<1x[4]xf32> to vector<1x1x[4]xf32>
+// CHECK:           return %[[SRC_CAST]] : vector<1x1x[4]xf32>
 func.func @cast_away_insert_leading_one_dims_rank2_scalable(%s: vector<1x[4]xf32>, %v: vector<1x1x[4]xf32>) -> vector<1x1x[4]xf32> {
   %0 = vector.insert %s, %v [0] : vector<1x[4]xf32> into vector<1x1x[4]xf32>
   return %0: vector<1x1x[4]xf32>
@@ -595,11 +617,11 @@ func.func @cast_away_insert_leading_one_dims_rank2_scalable(%s: vector<1x[4]xf32
 
 // CHECK-LABEL: func @cast_away_insert_leading_one_dims_rank2_one_dest
 //  CHECK-SAME: (%[[S:.+]]: vector<1x4xf32>, %[[V:.+]]: vector<1x2x1x4xf32>)
-//       CHECK:   %[[EXTRACTS:.+]] = vector.extract %[[S]][0] : vector<4xf32> from vector<1x4xf32>
-//       CHECK:   %[[EXTRACTV:.+]] = vector.extract %[[V]][0] : vector<2x1x4xf32> from vector<1x2x1x4xf32>
-//       CHECK:   %[[INSERT:.+]] = vector.insert %[[EXTRACTS]], %[[EXTRACTV]] [1, 0] : vector<4xf32> into vector<2x1x4xf32>
-//       CHECK:   %[[BCAST:.+]] = vector.broadcast %[[INSERT]] : vector<2x1x4xf32> to vector<1x2x1x4xf32>
-//       CHECK:   return %[[BCAST]]
+//       CHECK:   %[[SRC_CAST:.+]] = vector.shape_cast %[[S]] : vector<1x4xf32> to vector<4xf32>
+//       CHECK:   %[[DST_CAST:.+]] = vector.shape_cast %[[V]] : vector<1x2x1x4xf32> to vector<2x1x4xf32>
+//       CHECK:   %[[INSERT:.+]] = vector.insert %[[SRC_CAST]], %[[DST_CAST]] [1, 0] : vector<4xf32> into vector<2x1x4xf32>
+//       CHECK:   %[[RESULT_CAST:.+]] = vector.shape_cast %[[INSERT]] : vector<2x1x4xf32> to vector<1x2x1x4xf32>
+//       CHECK:   return %[[RESULT_CAST]]
 func.func @cast_away_insert_leading_one_dims_rank2_one_dest(%s: vector<1x4xf32>, %v: vector<1x2x1x4xf32>) -> vector<1x2x1x4xf32> {
   %0 = vector.insert %s, %v [0, 1] : vector<1x4xf32> into vector<1x2x1x4xf32>
   return %0: vector<1x2x1x4xf32>
@@ -610,11 +632,11 @@ func.func @cast_away_insert_leading_one_dims_rank2_one_dest(%s: vector<1x4xf32>,
 // CHECK-LABEL:   func.func @cast_away_insert_leading_one_dims_rank2_one_dest_scalable(
 // CHECK-SAME:      %[[S:.*]]: vector<1x[4]xf32>,
 // CHECK-SAME:      %[[V:.*]]: vector<1x2x1x[4]xf32>) -> vector<1x2x1x[4]xf32> {
-// CHECK:           %[[EXTRACTS:.*]] = vector.extract %[[S]][0] : vector<[4]xf32> from vector<1x[4]xf32>
-// CHECK:           %[[EXTRACTV:.*]] = vector.extract %[[V]][0] : vector<2x1x[4]xf32> from vector<1x2x1x[4]xf32>
-// CHECK:           %[[INSERT:.*]] = vector.insert %[[EXTRACTS]], %[[EXTRACTV]] [1, 0] : vector<[4]xf32> into vector<2x1x[4]xf32>
-// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[INSERT]] : vector<2x1x[4]xf32> to vector<1x2x1x[4]xf32>
-// CHECK:           return %[[BCAST]] : vector<1x2x1x[4]xf32>
+// CHECK:           %[[SRC_CAST:.*]] = vector.shape_cast %[[S]] : vector<1x[4]xf32> to vector<[4]xf32>
+// CHECK:           %[[DST_CAST:.*]] = vector.shape_cast %[[V]] : vector<1x2x1x[4]xf32> to vector<2x1x[4]xf32>
+// CHECK:           %[[INSERT:.*]] = vector.insert %[[SRC_CAST]], %[[DST_CAST]] [1, 0] : vector<[4]xf32> into vector<2x1x[4]xf32>
+// CHECK:           %[[RESULT_CAST:.*]] = vector.shape_cast %[[INSERT]] : vector<2x1x[4]xf32> to vector<1x2x1x[4]xf32>
+// CHECK:           return %[[RESULT_CAST]] : vector<1x2x1x[4]xf32>
 func.func @cast_away_insert_leading_one_dims_rank2_one_dest_scalable(%s: vector<1x[4]xf32>, %v: vector<1x2x1x[4]xf32>) -> vector<1x2x1x[4]xf32> {
   %0 = vector.insert %s, %v [0, 1] : vector<1x[4]xf32> into vector<1x2x1x[4]xf32>
   return %0: vector<1x2x1x[4]xf32>
@@ -624,8 +646,8 @@ func.func @cast_away_insert_leading_one_dims_rank2_one_dest_scalable(%s: vector<
 
 // CHECK-LABEL: func @cast_away_insert_leading_one_dims_non_one_dest
 //  CHECK-SAME: (%[[S:.+]]: vector<1x4xf32>, %[[V:.+]]: vector<8x1x4xf32>)
-//       CHECK:   %[[EXTRACT:.+]] = vector.extract %[[S]][0] : vector<4xf32> from vector<1x4xf32>
-//       CHECK:   %[[INSERT:.+]] = vector.insert %[[EXTRACT]], %[[V]] [5, 0] : vector<4xf32> into vector<8x1x4xf32>
+//       CHECK:   %[[SRC_CAST:.+]] = vector.shape_cast %[[S]] : vector<1x4xf32> to vector<4xf32>
+//       CHECK:   %[[INSERT:.+]] = vector.insert %[[SRC_CAST]], %[[V]] [5, 0] : vector<4xf32> into vector<8x1x4xf32>
 //       CHECK:   return %[[INSERT]]
 func.func @cast_away_insert_leading_one_dims_non_one_dest(%s: vector<1x4xf32>, %v: vector<8x1x4xf32>) -> vector<8x1x4xf32> {
   %0 = vector.insert %s, %v [5] : vector<1x4xf32> into vector<8x1x4xf32>
@@ -637,8 +659,8 @@ func.func @cast_away_insert_leading_one_dims_non_one_dest(%s: vector<1x4xf32>, %
 // CHECK-LABEL:   func.func @cast_away_insert_leading_one_dims_non_one_dest_scalable(
 // CHECK-SAME:      %[[S:.*]]: vector<1x[4]xf32>,
 // CHECK-SAME:      %[[V:.*]]: vector<8x1x[4]xf32>) -> vector<8x1x[4]xf32> {
-// CHECK:           %[[EXTRACT:.*]] = vector.extract %[[S]][0] : vector<[4]xf32> from vector<1x[4]xf32>
-// CHECK:           %[[INSERT:.*]] = vector.insert %[[EXTRACT]], %[[V]] [5, 0] : vector<[4]xf32> into vector<8x1x[4]xf32>
+// CHECK:           %[[SRC_CAST:.*]] = vector.shape_cast %[[S]] : vector<1x[4]xf32> to vector<[4]xf32>
+// CHECK:           %[[INSERT:.*]] = vector.insert %[[SRC_CAST]], %[[V]] [5, 0] : vector<[4]xf32> into vector<8x1x[4]xf32>
 // CHECK:           return %[[INSERT]] : vector<8x1x[4]xf32>
 func.func @cast_away_insert_leading_one_dims_non_one_dest_scalable(%s: vector<1x[4]xf32>, %v: vector<8x1x[4]xf32>) -> vector<8x1x[4]xf32> {
   %0 = vector.insert %s, %v [5] : vector<1x[4]xf32> into vector<8x1x[4]xf32>
@@ -649,11 +671,11 @@ func.func @cast_away_insert_leading_one_dims_non_one_dest_scalable(%s: vector<1x
 
 // CHECK-LABEL: func @cast_away_insert_leading_one_dims_one_two_dest
 //  CHECK-SAME: (%[[S:.+]]: vector<1x8xi1>, %[[V:.+]]: vector<1x1x8x1x8xi1>)
-//       CHECK:   %[[EXTRACTS:.+]] = vector.extract %[[S]][0] : vector<8xi1> from vector<1x8xi1>
-//       CHECK:   %[[EXTRACTV:.+]] = vector.extract %[[V]][0, 0] : vector<8x1x8xi1> from vector<1x1x8x1x8xi1>
-//       CHECK:   %[[INSERT:.+]] = vector.insert %[[EXTRACTS]], %[[EXTRACTV]] [7, 0] : vector<8xi1> into vector<8x1x8xi1>
-//       CHECK:   %[[BCAST:.+]] = vector.broadcast %[[INSERT]] : vector<8x1x8xi1> to vector<1x1x8x1x8xi1>
-//       CHECK:   return %[[BCAST]]
+//       CHECK:   %[[SRC_CAST:.+]] = vector.shape_cast %[[S]] : vector<1x8xi1> to vector<8xi1>
+//       CHECK:   %[[DST_CAST:.+]] = vector.shape_cast %[[V]] : vector<1x1x8x1x8xi1> to vector<8x1x8xi1>
+//       CHECK:   %[[INSERT:.+]] = vector.insert %[[SRC_CAST]], %[[DST_CAST]] [7, 0] : vector<8xi1> into vector<8x1x8xi1>
+//       CHECK:   %[[RESULT_CAST:.+]] = vector.shape_cast %[[INSERT]] : vector<8x1x8xi1> to vector<1x1x8x1x8xi1>
+//       CHECK:   return %[[RESULT_CAST]]
 func.func @cast_away_insert_leading_one_dims_one_two_dest(%s: vector<1x8xi1>, %v: vector<1x1x8x1x8xi1>) -> vector<1x1x8x1x8xi1> {
   %0 = vector.insert %s, %v [0, 0, 7] : vector<1x8xi1> into vector<1x1x8x1x8xi1>
   return %0: vector<1x1x8x1x8xi1>
@@ -664,11 +686,11 @@ func.func @cast_away_insert_leading_one_dims_one_two_dest(%s: vector<1x8xi1>, %v
 // CHECK-LABEL:   func.func @cast_away_insert_leading_one_dims_one_two_dest_scalable(
 // CHECK-SAME:      %[[S:.*]]: vector<1x[8]xi1>,
 // CHECK-SAME:      %[[V:.*]]: vector<1x1x8x1x[8]xi1>) -> vector<1x1x8x1x[8]xi1> {
-// CHECK:           %[[EXTRACTS:.*]] = vector.extract %[[S]][0] : vector<[8]xi1> from vector<1x[8]xi1>
-// CHECK:           %[[EXTRACTV:.*]] = vector.extract %[[V]][0, 0] : vector<8x1x[8]xi1> from vector<1x1x8x1x[8]xi1>
-// CHECK:           %[[INSERT:.*]] = vector.insert %[[EXTRACTS]], %[[EXTRACTV]] [7, 0] : vector<[8]xi1> into vector<8x1x[8]xi1>
-// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[INSERT]] : vector<8x1x[8]xi1> to vector<1x1x8x1x[8]xi1>
-// CHECK:           return %[[BCAST]] : vector<1x1x8x1x[8]xi1>
+// CHECK:           %[[SRC_CAST:.*]] = vector.shape_cast %[[S]] : vector<1x[8]xi1> to vector<[8]xi1>
+// CHECK:           %[[DST_CAST:.*]] = vector.shape_cast %[[V]] : vector<1x1x8x1x[8]xi1> to vector<8x1x[8]xi1>
+// CHECK:           %[[INSERT:.*]] = vector.insert %[[SRC_CAST]], %[[DST_CAST]] [7, 0] : vector<[8]xi1> into vector<8x1x[8]xi1>
+// CHECK:           %[[RESULT_CAST:.*]] = vector.shape_cast %[[INSERT]] : vector<8x1x[8]xi1> to vector<1x1x8x1x[8]xi1>
+// CHECK:           return %[[RESULT_CAST]] : vector<1x1x8x1x[8]xi1>
 func.func @cast_away_insert_leading_one_dims_one_two_dest_scalable(%s: vector<1x[8]xi1>, %v: vector<1x1x8x1x[8]xi1>) -> vector<1x1x8x1x[8]xi1> {
   %0 = vector.insert %s, %v [0, 0, 7] : vector<1x[8]xi1> into vector<1x1x8x1x[8]xi1>
   return %0: vector<1x1x8x1x[8]xi1>
@@ -678,8 +700,8 @@ func.func @cast_away_insert_leading_one_dims_one_two_dest_scalable(%s: vector<1x
 
 // CHECK-LABEL:   func.func @cast_away_constant_mask() -> vector<1x1x8x2x1xi1> {
 // CHECK:           %[[MASK:.*]] = vector.constant_mask [6, 1, 1] : vector<8x2x1xi1>
-// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[MASK]] : vector<8x2x1xi1> to vector<1x1x8x2x1xi1>
-// CHECK:           return %[[BCAST]] : vector<1x1x8x2x1xi1>
+// CHECK:           %[[MASK_CAST:.*]] = vector.shape_cast %[[MASK]] : vector<8x2x1xi1> to vector<1x1x8x2x1xi1>
+// CHECK:           return %[[MASK_CAST]] : vector<1x1x8x2x1xi1>
 func.func @cast_away_constant_mask() -> vector<1x1x8x2x1xi1> {
   %0 = vector.constant_mask [1, 1, 6, 1, 1] : vector<1x1x8x2x1xi1>
   return %0: vector<1x1x8x2x1xi1>
@@ -698,7 +720,7 @@ func.func @drop_unit_dims_scalar_cond_select(%cond: i1, %arg0: vector<1x16xi1>,
 
 // CHECK-LABEL: func.func @cast_away_load_leading_one_dims
 // CHECK:         %[[L:.+]] = vector.load %{{.*}}[%{{.*}}, %{{.*}}] : memref<8x16xf32>, vector<4xf32>
-// CHECK:         %[[B:.+]] = vector.broadcast %[[L]] : vector<4xf32> to vector<1x4xf32>
+// CHECK:         %[[B:.+]] = vector.shape_cast %[[L]] : vector<4xf32> to vector<1x4xf32>
 // CHECK:         return %[[B]] : vector<1x4xf32>
 func.func @cast_away_load_leading_one_dims(%base: memref<8x16xf32>, %i: index, %j: index) -> vector<1x4xf32> {
   %0 = vector.load %base[%i, %j] : memref<8x16xf32>, vector<1x4xf32>
@@ -707,11 +729,22 @@ func.func @cast_away_load_leading_one_dims(%base: memref<8x16xf32>, %i: index, %
 
 // -----
 
+// CHECK-LABEL: func.func @cast_away_load_leading_one_dims_scalable
+// CHECK:         %[[L:.+]] = vector.load %{{.*}}[%{{.*}}, %{{.*}}] : memref<?x?xf32>, vector<[4]xf32>
+// CHECK:         %[[B:.+]] = vector.shape_cast %[[L]] : vector<[4]xf32> to vector<1x[4]xf32>
+// CHECK:         return %[[B]] : vector<1x[4]xf32>
+func.func @cast_away_load_leading_one_dims_scalable(%base: memref<?x?xf32>, %i: index, %j: index) -> vector<1x[4]xf32> {
+  %0 = vector.load %base[%i, %j] : memref<?x?xf32>, vector<1x[4]xf32>
+  return %0 : vector<1x[4]xf32>
+}
+
+// -----
+
 // CHECK-LABEL: func.func @cast_away_maskedload_leading_one_dims
-// CHECK:         %[[M:.+]] = vector.extract %{{.*}}[0] : vector<4xi1> from vector<1x4xi1>
-// CHECK:         %[[P:.+]] = vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+// CHECK:         %[[M:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi1> to vector<4xi1>
+// CHECK:         %[[P:.+]] = vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
 // CHECK:         %[[L:.+]] = vector.maskedload %{{.*}}[%{{.*}}], %[[M]], %[[P]] : memref<16xf32>, vector<4xi1>, vector<4xf32> into vector<4xf32>
-// CHECK:         %[[B:.+]] = vector.broadcast %[[L]] : vector<4xf32> to vector<1x4xf32>
+// CHECK:         %[[B:.+]] = vector.shape_cast %[[L]] : vector<4xf32> to vector<1x4xf32>
 // CHECK:         return %[[B]] : vector<1x4xf32>
 func.func @cast_away_maskedload_leading_one_dims(%base: memref<16xf32>, %i: index, %mask: vector<1x4xi1>, %pass: vector<1x4xf32>) -> vector<1x4xf32> {
   %0 = vector.maskedload %base[%i], %mask, %pass : memref<16xf32>, vector<1x4xi1>, vector<1x4xf32> into vector<1x4xf32>
@@ -721,10 +754,10 @@ func.func @cast_away_maskedload_leading_one_dims(%base: memref<16xf32>, %i: inde
 // -----
 
 // CHECK-LABEL: func.func @cast_away_expandload_leading_one_dims
-// CHECK:         %[[M:.+]] = vector.extract %{{.*}}[0] : vector<4xi1> from vector<1x4xi1>
-// CHECK:         %[[P:.+]] = vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+// CHECK:         %[[M:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi1> to vector<4xi1>
+// CHECK:         %[[P:.+]] = vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
 // CHECK:         %[[L:.+]] = vector.expandload %{{.*}}[%{{.*}}], %[[M]], %[[P]] : memref<16xf32>, vector<4xi1>, vector<4xf32> into vector<4xf32>
-// CHECK:         %[[B:.+]] = vector.broadcast %[[L]] : vector<4xf32> to vector<1x4xf32>
+// CHECK:         %[[B:.+]] = vector.shape_cast %[[L]] : vector<4xf32> to vector<1x4xf32>
 // CHECK:         return %[[B]] : vector<1x4xf32>
 func.func @cast_away_expandload_leading_one_dims(%base: memref<16xf32>, %i: index, %mask: vector<1x4xi1>, %pass: vector<1x4xf32>) -> vector<1x4xf32> {
   %0 = vector.expandload %base[%i], %mask, %pass : memref<16xf32>, vector<1x4xi1>, vector<1x4xf32> into vector<1x4xf32>
@@ -734,11 +767,11 @@ func.func @cast_away_expandload_leading_one_dims(%base: memref<16xf32>, %i: inde
 // -----
 
 // CHECK-LABEL: func.func @cast_away_gather_leading_one_dims
-// CHECK:         %[[I:.+]] = vector.extract %{{.*}}[0] : vector<4xi32> from vector<1x4xi32>
-// CHECK:         %[[M:.+]] = vector.extract %{{.*}}[0] : vector<4xi1> from vector<1x4xi1>
-// CHECK:         %[[P:.+]] = vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+// CHECK:         %[[I:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi32> to vector<4xi32>
+// CHECK:         %[[M:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi1> to vector<4xi1>
+// CHECK:         %[[P:.+]] = vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
 // CHECK:         %[[G:.+]] = vector.gather %{{.*}}[%{{.*}}] [%[[I]]], %[[M]], %[[P]] : memref<16xf32>, vector<4xi32>, vector<4xi1>, vector<4xf32> into vector<4xf32>
-// CHECK:         %[[B:.+]] = vector.broadcast %[[G]] : vector<4xf32> to vector<1x4xf32>
+// CHECK:         %[[B:.+]] = vector.shape_cast %[[G]] : vector<4xf32> to vector<1x4xf32>
 // CHECK:         return %[[B]] : vector<1x4xf32>
 func.func @cast_away_gather_leading_one_dims(%base: memref<16xf32>, %i: index, %idx: vector<1x4xi32>, %mask: vector<1x4xi1>, %pass: vector<1x4xf32>) -> vector<1x4xf32> {
   %0 = vector.gather %base[%i] [%idx], %mask, %pass : memref<16xf32>, vector<1x4xi32>, vector<1x4xi1>, vector<1x4xf32> into vector<1x4xf32>
@@ -748,7 +781,7 @@ func.func @cast_away_gather_leading_one_dims(%base: memref<16xf32>, %i: index, %
 // -----
 
 // CHECK-LABEL: func.func @cast_away_store_leading_one_dims
-// CHECK:         %[[V:.+]] = vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+// CHECK:         %[[V:.+]] = vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
 // CHECK:         vector.store %[[V]], %{{.*}}[%{{.*}}, %{{.*}}] : memref<8x16xf32>, vector<4xf32>
 func.func @cast_away_store_leading_one_dims(%val: vector<1x4xf32>, %base: memref<8x16xf32>, %i: index, %j: index) {
   vector.store %val, %base[%i, %j] : memref<8x16xf32>, vector<1x4xf32>
@@ -757,9 +790,19 @@ func.func @cast_away_store_leading_one_dims(%val: vector<1x4xf32>, %base: memref
 
 // -----
 
+// CHECK-LABEL: func.func @cast_away_store_leading_one_dims_scalable
+// CHECK:         %[[V:.+]] = vector.shape_cast %{{.*}} : vector<1x[4]xf32> to vector<[4]xf32>
+// CHECK:         vector.store %[[V]], %{{.*}}[%{{.*}}, %{{.*}}] : memref<?x?xf32>, vector<[4]xf32>
+func.func @cast_away_store_leading_one_dims_scalable(%val: vector<1x[4]xf32>, %base: memref<?x?xf32>, %i: index, %j: index) {
+  vector.store %val, %base[%i, %j] : memref<?x?xf32>, vector<1x[4]xf32>
+  return
+}
+
+// -----
+
 // CHECK-LABEL: func.func @cast_away_maskedstore_leading_one_dims
-// CHECK:         %[[M:.+]] = vector.extract %{{.*}}[0] : vector<4xi1> from vector<1x4xi1>
-// CHECK:         %[[V:.+]] = vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+// CHECK:         %[[M:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi1> to vector<4xi1>
+// CHECK:         %[[V:.+]] = vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
 // CHECK:         vector.maskedstore %{{.*}}[%{{.*}}], %[[M]], %[[V]] : memref<16xf32>, vector<4xi1>, vector<4xf32>
 func.func @cast_away_maskedstore_leading_one_dims(%base: memref<16xf32>, %i: index, %mask: vector<1x4xi1>, %val: vector<1x4xf32>) {
   vector.maskedstore %base[%i], %mask, %val : memref<16xf32>, vector<1x4xi1>, vector<1x4xf32>
@@ -769,8 +812,8 @@ func.func @cast_away_maskedstore_leading_one_dims(%base: memref<16xf32>, %i: ind
 // -----
 
 // CHECK-LABEL: func.func @cast_away_compressstore_leading_one_dims
-// CHECK:         %[[M:.+]] = vector.extract %{{.*}}[0] : vector<4xi1> from vector<1x4xi1>
-// CHECK:         %[[V:.+]] = vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+// CHECK:         %[[M:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi1> to vector<4xi1>
+// CHECK:         %[[V:.+]] = vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
 // CHECK:         vector.compressstore %{{.*}}[%{{.*}}], %[[M]], %[[V]] : memref<16xf32>, vector<4xi1>, vector<4xf32>
 func.func @cast_away_compressstore_leading_one_dims(%base: memref<16xf32>, %i: index, %mask: vector<1x4xi1>, %val: vector<1x4xf32>) {
   vector.compressstore %base[%i], %mask, %val : memref<16xf32>, vector<1x4xi1>, vector<1x4xf32>
@@ -780,9 +823,9 @@ func.func @cast_away_compressstore_leading_one_dims(%base: memref<16xf32>, %i: i
 // -----
 
 // CHECK-LABEL: func.func @cast_away_scatter_leading_one_dims
-// CHECK:         %[[I:.+]] = vector.extract %{{.*}}[0] : vector<4xi32> from vector<1x4xi32>
-// CHECK:         %[[M:.+]] = vector.extract %{{.*}}[0] : vector<4xi1> from vector<1x4xi1>
-// CHECK:         %[[V:.+]] = vector.extract %{{.*}}[0] : vector<4xf32> from vector<1x4xf32>
+// CHECK:         %[[I:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi32> to vector<4xi32>
+// CHECK:         %[[M:.+]] = vector.shape_cast %{{.*}} : vector<1x4xi1> to vector<4xi1>
+// CHECK:         %[[V:.+]] = vector.shape_cast %{{.*}} : vector<1x4xf32> to vector<4xf32>
 // CHECK:         vector.scatter %{{.*}}[%{{.*}}] [%[[I]]], %[[M]], %[[V]] : memref<16xf32>, vector<4xi32>, vector<4xi1>, vector<4xf32>
 func.func @cast_away_scatter_leading_one_dims(%base: memref<16xf32>, %i: index, %idx: vector<1x4xi32>, %mask: vector<1x4xi1>, %val: vector<1x4xf32>) {
   vector.scatter %base[%i] [%idx], %mask, %val : memref<16xf32>, vector<1x4xi32>, vector<1x4xi1>, vector<1x4xf32>
diff --git a/mlir/test/Dialect/Vector/vector-transforms.mlir b/mlir/test/Dialect/Vector/vector-transforms.mlir
index de12a87253a67..37ca4358d97bd 100644
--- a/mlir/test/Dialect/Vector/vector-transforms.mlir
+++ b/mlir/test/Dialect/Vector/vector-transforms.mlir
@@ -36,7 +36,7 @@ func.func @no_change(%arg0: vector<2x[4]x1xf32>, %arg1: vector<2x[4]x1xf32>) ->
 
 // CHECK-LABEL:   func.func @cast_away_leading_one_dim(
 // CHECK:           %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} : vector<4x1xf32>
-// CHECK:           vector.broadcast %[[MUL]] : vector<4x1xf32> to vector<1x4x1xf32>
+// CHECK:           vector.shape_cast %[[MUL]] : vector<4x1xf32> to vector<1x4x1xf32>
 func.func @cast_away_leading_one_dim(%arg0: vector<1x4x1xf32>, %arg1: vector<1x4x1xf32>) -> vector<1x4x1xf32> {
   %1 = arith.mulf %arg0, %arg1 : vector<1x4x1xf32>
   return %1: vector<1x4x1xf32>
@@ -44,7 +44,7 @@ func.func @cast_away_leading_one_dim(%arg0: vector<1x4x1xf32>, %arg1: vector<1x4
 
 // CHECK-LABEL:   func.func @cast_away_leading_one_dim_scalable(
 // CHECK:           %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} : vector<[4]x1xf32>
-// CHECK:           vector.broadcast %[[MUL]] : vector<[4]x1xf32> to vector<1x[4]x1xf32>
+// CHECK:           vector.shape_cast %[[MUL]] : vector<[4]x1xf32> to vector<1x[4]x1xf32>
 func.func @cast_away_leading_one_dim_scalable(%arg0: vector<1x[4]x1xf32>, %arg1: vector<1x[4]x1xf32>) -> vector<1x[4]x1xf32> {
   %1 = arith.mulf %arg0, %arg1 : vector<1x[4]x1xf32>
   return %1: vector<1x[4]x1xf32>

>From c3e4ac9247ffc5ec9fa86e05108bc0ecd3aec4a0 Mon Sep 17 00:00:00 2001
From: Krzysztof Drewniak <Krzysztof.Drewniak at amd.com>
Date: Wed, 6 May 2026 23:18:18 +0000
Subject: [PATCH 2/8] Clang-format

---
 .../Vector/Transforms/VectorDropLeadUnitDim.cpp   | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
index 931a87ff83e9c..c6cdca156c9f4 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
@@ -65,8 +65,7 @@ static bool hasNonScalableUnitLeadingDims(VectorType type, int64_t dropCount) {
   ArrayRef<int64_t> leadingShape = type.getShape().take_front(dropCount);
   ArrayRef<bool> leadingScalable = type.getScalableDims().take_front(dropCount);
   return llvm::all_of(leadingShape, [](int64_t dim) { return dim == 1; }) &&
-         llvm::none_of(leadingScalable,
-                       [](bool scalable) { return scalable; });
+         llvm::none_of(leadingScalable, [](bool scalable) { return scalable; });
 }
 
 static bool isNonScalableUnitDim(VectorType type, int64_t dim) {
@@ -195,8 +194,8 @@ struct CastAwayInsertStridedSliceLeadingOneDim
     // Trim leading one dimensions from both operands.
     Location loc = insertOp.getLoc();
 
-    Value newSrcVector = shapeCastVector(rewriter, loc,
-                                         insertOp.getValueToStore(), newSrcType);
+    Value newSrcVector =
+        shapeCastVector(rewriter, loc, insertOp.getValueToStore(), newSrcType);
     Value newDstVector =
         shapeCastVector(rewriter, loc, insertOp.getDest(), newDstType);
 
@@ -371,8 +370,8 @@ struct CastAwayTransferWriteLeadingOneDim
         shapeCastVector(rewriter, write.getLoc(), write.getVector(), newType);
 
     if (write.getMask()) {
-      Value newMask = dropUnitDimsFromMask(
-          rewriter, write.getLoc(), write.getMask(), newType, newMap);
+      Value newMask = dropUnitDimsFromMask(rewriter, write.getLoc(),
+                                           write.getMask(), newType, newMap);
       rewriter.replaceOpWithNewOp<vector::TransferWriteOp>(
           write, newVector, write.getBase(), write.getIndices(),
           AffineMapAttr::get(newMap), newMask, inBoundsAttr);
@@ -592,8 +591,8 @@ class CastAwayElementwiseLeadingOneDim : public RewritePattern {
     SmallVector<Value, 4> newOperands;
     for (Value operand : op->getOperands()) {
       if (auto opVecType = dyn_cast<VectorType>(operand.getType()))
-        newOperands.push_back(shapeCastVector(
-            rewriter, op->getLoc(), operand, trimLeadingOneDims(opVecType)));
+        newOperands.push_back(shapeCastVector(rewriter, op->getLoc(), operand,
+                                              trimLeadingOneDims(opVecType)));
       else
         newOperands.push_back(operand);
     }

>From f8ce0504ab2b6311c427f67bf36284057b822525 Mon Sep 17 00:00:00 2001
From: Krzysztof Drewniak <Krzysztof.Drewniak at amd.com>
Date: Fri, 8 May 2026 00:27:11 +0000
Subject: [PATCH 3/8] Review comments (dropped unneeded bounds checks, fixed
 modify-before-failing)

---
 .../Transforms/VectorDropLeadUnitDim.cpp      | 126 +++++++++++-------
 .../vector-dropleadunitdim-transforms.mlir    |  15 +++
 2 files changed, 90 insertions(+), 51 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
index c6cdca156c9f4..33efc5a1e0546 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
@@ -7,7 +7,9 @@
 //===----------------------------------------------------------------------===//
 
 #include <numeric>
+#include <utility>
 
+#include "mlir/Dialect/Utils/IndexingUtils.h"
 #include "mlir/Dialect/Utils/StructuredOpsUtils.h"
 #include "mlir/Dialect/Vector/IR/VectorOps.h"
 #include "mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h"
@@ -15,6 +17,7 @@
 #include "mlir/Dialect/Vector/Utils/VectorUtils.h"
 #include "mlir/IR/Builders.h"
 #include "mlir/IR/TypeUtilities.h"
+#include "llvm/ADT/Repeated.h"
 #include "llvm/ADT/STLExtras.h"
 
 #define DEBUG_TYPE "vector-drop-unit-dim"
@@ -22,11 +25,6 @@
 using namespace mlir;
 using namespace mlir::vector;
 
-/// Return a smallVector of size `rank` containing all zeros.
-static SmallVector<int64_t> splatZero(int64_t rank) {
-  return SmallVector<int64_t>(rank, 0);
-}
-
 // Trims leading one dimensions from `oldType` and returns the result type.
 // Returns `vector<1xT>` if `oldType` only has one element.
 static VectorType trimLeadingOneDims(VectorType oldType) {
@@ -60,8 +58,8 @@ static Value shapeCastVector(OpBuilder &b, Location loc, Value value,
 }
 
 static bool hasNonScalableUnitLeadingDims(VectorType type, int64_t dropCount) {
-  if (dropCount < 0 || dropCount > type.getRank())
-    return false;
+  assert(dropCount >= 0 && dropCount <= type.getRank() &&
+         "expected a valid leading dimension count");
   ArrayRef<int64_t> leadingShape = type.getShape().take_front(dropCount);
   ArrayRef<bool> leadingScalable = type.getScalableDims().take_front(dropCount);
   return llvm::all_of(leadingShape, [](int64_t dim) { return dim == 1; }) &&
@@ -69,8 +67,14 @@ static bool hasNonScalableUnitLeadingDims(VectorType type, int64_t dropCount) {
 }
 
 static bool isNonScalableUnitDim(VectorType type, int64_t dim) {
-  return dim >= 0 && dim < type.getRank() && type.getShape()[dim] == 1 &&
-         !type.getScalableDims()[dim];
+  return type.getShape()[dim] == 1 && !type.getScalableDims()[dim];
+}
+
+static VectorType transposeVectorType(VectorType type,
+                                      ArrayRef<int64_t> permutation) {
+  return VectorType::get(applyPermutation(type.getShape(), permutation),
+                         type.getElementType(),
+                         applyPermutation(type.getScalableDims(), permutation));
 }
 
 /// Shape-casts `operand` to the vector type obtained by dropping the first
@@ -115,8 +119,10 @@ static Value dropLeadingDimsForContraction(OpBuilder &b, Location loc,
 
   // vector.contract rejects 0-D vector accumulators/results. When every vector
   // dimension is dropped, use the scalar path that vector.contract accepts.
-  if (dropCount == oldType.getRank())
-    return vector::ExtractOp::create(b, loc, operand, splatZero(dropCount));
+  if (dropCount == oldType.getRank()) {
+    llvm::Repeated<int64_t> zeros(static_cast<size_t>(dropCount), 0);
+    return vector::ExtractOp::create(b, loc, operand, llvm::to_vector(zeros));
+  }
 
   return shapeCastDroppingLeadingDims(b, loc, operand, dropCount);
 }
@@ -387,6 +393,16 @@ struct CastAwayTransferWriteLeadingOneDim
 
 } // namespace
 
+namespace {
+struct ContractionOperandDropInfo {
+  AffineMap map;
+  bool needsShapeCast = false;
+  bool transposeNeeded = false;
+  bool transposeNonOuterUnitDims = false;
+  SmallVector<int64_t> permutation;
+};
+} // namespace
+
 FailureOr<Value>
 mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
                                                MaskingOpInterface maskingOp,
@@ -423,35 +439,34 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
 
   SmallVector<Value> operands = {contractOp.getLhs(), contractOp.getRhs(),
                                  contractOp.getAcc()};
+  SmallVector<ContractionOperandDropInfo> operandDropInfo;
   SmallVector<Value> newOperands;
   auto loc = contractOp.getLoc();
 
   for (const auto &it : llvm::enumerate(oldIndexingMaps)) {
     // Check if the dim to be dropped exists as a leading dim in the operand
     // if it does then we use vector.shape_cast to drop it.
-    bool needsShapeCast = false;
+    ContractionOperandDropInfo dropInfo;
     SmallVector<AffineExpr> results;
-    auto map = it.value();
-    int64_t orginalZeroDim = it.value().getDimPosition(0);
-    if (orginalZeroDim != dimToDrop) {
+    dropInfo.map = it.value();
+    int64_t originalZeroDim = it.value().getDimPosition(0);
+    if (originalZeroDim != dimToDrop) {
       // There are two reasons to be in this path, 1. We need to
       // transpose the operand to make the dim to be dropped
       // leading. 2. The dim to be dropped does not exist and in
       // that case we dont want to add a unit transpose but we must
       // check all the indices to make sure this is the case.
-      bool transposeNeeded = false;
-      SmallVector<int64_t> perm;
       SmallVector<AffineExpr> transposeResults;
 
-      for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
-        int64_t currDim = map.getDimPosition(i);
+      for (int64_t i = 0, e = dropInfo.map.getNumResults(); i < e; ++i) {
+        int64_t currDim = dropInfo.map.getDimPosition(i);
         if (currDim == dimToDrop) {
-          transposeNeeded = true;
-          perm.insert(perm.begin(), i);
+          dropInfo.transposeNeeded = true;
+          dropInfo.permutation.insert(dropInfo.permutation.begin(), i);
           auto targetExpr = rewriter.getAffineDimExpr(currDim);
           transposeResults.insert(transposeResults.begin(), targetExpr);
         } else {
-          perm.push_back(i);
+          dropInfo.permutation.push_back(i);
           auto targetExpr = rewriter.getAffineDimExpr(currDim);
           transposeResults.push_back(targetExpr);
         }
@@ -460,36 +475,31 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
       // Checks if only the outer, unit dimensions (of size 1) are permuted.
       // Such transposes do not materially effect the underlying vector and can
       // be omitted. EG: perm [1, 0, 2] applied to vector<1x1x8xi32>
-      bool transposeNonOuterUnitDims = false;
       auto operandShape = cast<ShapedType>(operands[it.index()].getType());
-      for (auto [index, dim] :
-           llvm::enumerate(ArrayRef<int64_t>(perm).drop_back(1))) {
+      for (auto [index, dim] : llvm::enumerate(
+               ArrayRef<int64_t>(dropInfo.permutation).drop_back(1))) {
         if (dim != static_cast<int64_t>(index) &&
             operandShape.getDimSize(index) != 1) {
-          transposeNonOuterUnitDims = true;
+          dropInfo.transposeNonOuterUnitDims = true;
           break;
         }
       }
 
       // Do the transpose now if needed so that we can drop the correct dim
       // with shape_cast later.
-      if (transposeNeeded) {
-        map = AffineMap::get(map.getNumDims(), 0, transposeResults,
-                             contractOp.getContext());
-        if (transposeNonOuterUnitDims) {
-          operands[it.index()] = rewriter.createOrFold<vector::TransposeOp>(
-              loc, operands[it.index()], perm);
-        }
-      }
+      if (dropInfo.transposeNeeded)
+        dropInfo.map =
+            AffineMap::get(dropInfo.map.getNumDims(), 0, transposeResults,
+                           contractOp.getContext());
     }
     // We have taken care to have the dim to be dropped be
     // the leading dim. If its still not leading that means it
     // does not exist in this operand and hence we do not need a shape_cast.
-    if (map.getDimPosition(0) == dimToDrop)
-      needsShapeCast = true;
+    if (dropInfo.map.getDimPosition(0) == dimToDrop)
+      dropInfo.needsShapeCast = true;
 
-    for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
-      int64_t currDim = map.getDimPosition(i);
+    for (int64_t i = 0, e = dropInfo.map.getNumResults(); i < e; ++i) {
+      int64_t currDim = dropInfo.map.getDimPosition(i);
       if (currDim == dimToDrop)
         // This is the dim we are dropping.
         continue;
@@ -497,18 +507,36 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
           currDim < dimToDrop ? currDim : currDim - 1);
       results.push_back(targetExpr);
     }
-    newIndexingMaps.push_back(AffineMap::get(map.getNumDims() - 1, 0, results,
-                                             contractOp.getContext()));
-    if (needsShapeCast) {
+    newIndexingMaps.push_back(AffineMap::get(dropInfo.map.getNumDims() - 1, 0,
+                                             results, contractOp.getContext()));
+    if (dropInfo.needsShapeCast) {
       auto operandType = cast<VectorType>(operands[it.index()].getType());
-      if (operandType.getRank() < dropDim ||
-          !hasNonScalableUnitLeadingDims(operandType, dropDim))
+      VectorType typeToDrop =
+          dropInfo.transposeNeeded && dropInfo.transposeNonOuterUnitDims
+              ? transposeVectorType(operandType, dropInfo.permutation)
+              : operandType;
+      if (!hasNonScalableUnitLeadingDims(typeToDrop, dropDim))
         return failure();
-      newOperands.push_back(dropLeadingDimsForContraction(
-          rewriter, loc, operands[it.index()], dropDim));
-    } else {
-      newOperands.push_back(operands[it.index()]);
     }
+    operandDropInfo.push_back(std::move(dropInfo));
+  }
+
+  if (maskingOp) {
+    auto oldMaskType = cast<VectorType>(maskingOp.getMask().getType());
+    if (oldMaskType.getRank() <= 1 ||
+        !isNonScalableUnitDim(oldMaskType, dimToDrop))
+      return failure();
+  }
+
+  for (const auto &it : llvm::enumerate(operandDropInfo)) {
+    Value operand = operands[it.index()];
+    const ContractionOperandDropInfo &dropInfo = it.value();
+    if (dropInfo.transposeNeeded && dropInfo.transposeNonOuterUnitDims)
+      operand = rewriter.createOrFold<vector::TransposeOp>(
+          loc, operand, dropInfo.permutation);
+    if (dropInfo.needsShapeCast)
+      operand = dropLeadingDimsForContraction(rewriter, loc, operand, dropDim);
+    newOperands.push_back(operand);
   }
 
   // Depending on whether this vector.contract is masked, the replacing Op
@@ -519,10 +547,6 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
       rewriter.getArrayAttr(newIteratorTypes), contractOp.getKind());
 
   if (maskingOp) {
-    auto oldMaskType = cast<VectorType>(maskingOp.getMask().getType());
-    if (oldMaskType.getRank() <= 1 ||
-        !isNonScalableUnitDim(oldMaskType, dimToDrop))
-      return failure();
     Value newMask =
         shapeCastDroppingDim(rewriter, loc, maskingOp.getMask(), dimToDrop);
 
diff --git a/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir b/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir
index e5b1cc07319e3..ab2f566ff0cb0 100644
--- a/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir
+++ b/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir
@@ -296,6 +296,21 @@ func.func @do_not_cast_away_contraction_with_scalable_rank1_acc(%arg0: vector<64
 }
 
 // -----
+
+// CHECK-LABEL: do_not_cast_away_contraction_with_scalable_operand_dim
+//  CHECK-NOT: vector.shape_cast
+//  CHECK-NOT: vector.extract
+//  CHECK-NOT: vector.broadcast
+//  CHECK-NEXT: vector.contract
+//  CHECK-NEXT: return
+
+func.func @do_not_cast_away_contraction_with_scalable_operand_dim(%arg0: vector<64xf32>, %arg1: vector<[1]x64xf32>, %arg2: vector<1xf32>) -> vector<1xf32> {
+  %0 = vector.contract {indexing_maps = [affine_map<(d0, d1) -> (d0)>, affine_map<(d0, d1) -> (d1, d0)>, affine_map<(d0, d1) -> (d1)>], iterator_types = ["reduction", "parallel"], kind = #vector.kind<add>} %arg0, %arg1, %arg2 : vector<64xf32>, vector<[1]x64xf32> into vector<1xf32>
+  return %0 : vector<1xf32>
+}
+
+// -----
+
 // CHECK-LABEL: func @cast_away_extract_strided_slice_leading_one_dims
 func.func @cast_away_extract_strided_slice_leading_one_dims(%arg0: vector<1x8x8xf16>) -> vector<1x1x8xf16> {
   // CHECK:     %[[SRC:.+]] = vector.shape_cast %{{.*}} : vector<1x8x8xf16> to vector<8x8xf16>

>From e2d5d377ec04929d647551c8c164437493b729e5 Mon Sep 17 00:00:00 2001
From: Krzysztof Drewniak <Krzysztof.Drewniak at amd.com>
Date: Fri, 8 May 2026 18:29:15 +0000
Subject: [PATCH 4/8] Human changes

---
 .../Transforms/VectorDropLeadUnitDim.cpp      | 262 +++++++++---------
 .../vector-dropleadunitdim-transforms.mlir    |   8 +-
 2 files changed, 132 insertions(+), 138 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
index 33efc5a1e0546..9ba53c978d36d 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
@@ -7,9 +7,7 @@
 //===----------------------------------------------------------------------===//
 
 #include <numeric>
-#include <utility>
 
-#include "mlir/Dialect/Utils/IndexingUtils.h"
 #include "mlir/Dialect/Utils/StructuredOpsUtils.h"
 #include "mlir/Dialect/Vector/IR/VectorOps.h"
 #include "mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h"
@@ -48,83 +46,89 @@ static VectorType trimLeadingOneDims(VectorType oldType) {
   return VectorType::get(newShape, oldType.getElementType(), newScalableDims);
 }
 
-/// Returns `value` if it already has `newType`, otherwise inserts a
-/// vector.shape_cast to `newType`.
-static Value shapeCastVector(OpBuilder &b, Location loc, Value value,
-                             VectorType newType) {
-  if (value.getType() == newType)
-    return value;
-  return vector::ShapeCastOp::create(b, loc, newType, value);
-}
-
-static bool hasNonScalableUnitLeadingDims(VectorType type, int64_t dropCount) {
-  assert(dropCount >= 0 && dropCount <= type.getRank() &&
-         "expected a valid leading dimension count");
-  ArrayRef<int64_t> leadingShape = type.getShape().take_front(dropCount);
-  ArrayRef<bool> leadingScalable = type.getScalableDims().take_front(dropCount);
-  return llvm::all_of(leadingShape, [](int64_t dim) { return dim == 1; }) &&
-         llvm::none_of(leadingScalable, [](bool scalable) { return scalable; });
-}
-
 static bool isNonScalableUnitDim(VectorType type, int64_t dim) {
+  assert(dim >= 0 && dim < type.getRank() &&
+         "expected a valid vector dimension");
   return type.getShape()[dim] == 1 && !type.getScalableDims()[dim];
 }
 
-static VectorType transposeVectorType(VectorType type,
-                                      ArrayRef<int64_t> permutation) {
-  return VectorType::get(applyPermutation(type.getShape(), permutation),
-                         type.getElementType(),
-                         applyPermutation(type.getScalableDims(), permutation));
+/// Returns true if the first `k` dimensions of `type` are non-scalable unit
+/// dimensions.
+static bool leadingDimsAreUnit(VectorType type, int64_t k) {
+  assert(k >= 0 && k <= type.getRank() &&
+         "expected a valid leading dimension count");
+  return llvm::all_of(llvm::seq<int64_t>(0, k), [&](int64_t dim) {
+    return isNonScalableUnitDim(type, dim);
+  });
 }
 
-/// Shape-casts `operand` to the vector type obtained by dropping the first
-/// `dropCount` dimensions. Callers must ensure at least one vector dimension
-/// remains after the drop.
-static Value shapeCastDroppingLeadingDims(OpBuilder &b, Location loc,
-                                          Value operand, int64_t dropCount) {
-  auto oldType = cast<VectorType>(operand.getType());
-  assert(dropCount < oldType.getRank() &&
-         "shape_cast cannot drop all vector dimensions");
-  VectorType newType = VectorType::get(
-      oldType.getShape().drop_front(dropCount), oldType.getElementType(),
-      oldType.getScalableDims().drop_front(dropCount));
-  return shapeCastVector(b, loc, operand, newType);
+static bool leadingDimsAreUnitAfterPermutation(VectorType type,
+                                               ArrayRef<int64_t> permutation,
+                                               int64_t k) {
+  assert(k >= 0 && k <= static_cast<int64_t>(permutation.size()) &&
+         "expected a valid leading dimension count");
+  return llvm::all_of(permutation.take_front(k), [&](int64_t dim) {
+    return isNonScalableUnitDim(type, dim);
+  });
 }
 
-static Value shapeCastDroppingDim(OpBuilder &b, Location loc, Value operand,
-                                  int64_t dim) {
+/// Shape-casts `operand` to the vector type obtained by dropping dimension
+/// `dim`, which must be non-scalable and unit-sized.
+static Value dropUnitDim(OpBuilder &b, Location loc, Value operand,
+                         int64_t dim) {
   auto oldType = cast<VectorType>(operand.getType());
   assert(isNonScalableUnitDim(oldType, dim) &&
          "expected a non-scalable unit dim to drop");
+  int64_t rank = oldType.getRank();
+  assert(rank > 1 && "cannot shape_cast to a 0-D vector");
 
   SmallVector<int64_t> newShape;
   SmallVector<bool> newScalableDims;
-  for (int64_t i = 0, e = oldType.getRank(); i < e; ++i) {
-    if (i == dim)
+  newShape.reserve(rank - 1);
+  newScalableDims.reserve(rank - 1);
+  for (auto [i, size, scalable] :
+       llvm::enumerate(oldType.getShape(), oldType.getScalableDims())) {
+    if (static_cast<int64_t>(i) == dim)
       continue;
-    newShape.push_back(oldType.getShape()[i]);
-    newScalableDims.push_back(oldType.getScalableDims()[i]);
+    newShape.push_back(size);
+    newScalableDims.push_back(scalable);
   }
 
-  return shapeCastVector(
-      b, loc, operand,
-      VectorType::get(newShape, oldType.getElementType(), newScalableDims));
+  return b.createOrFold<vector::ShapeCastOp>(
+      loc, VectorType::get(newShape, oldType.getElementType(), newScalableDims),
+      operand);
 }
 
-static Value dropLeadingDimsForContraction(OpBuilder &b, Location loc,
-                                           Value operand, int64_t dropCount) {
+/// Shape-casts `operand` to the vector type obtained by dropping the first
+/// `k` non-scalable unit dimensions. Callers must ensure at least one vector
+/// dimension remains after the drop.
+static Value dropLeadingUnitDims(OpBuilder &b, Location loc, Value operand,
+                                 int64_t k) {
+  auto oldType = cast<VectorType>(operand.getType());
+  assert(leadingDimsAreUnit(oldType, k) &&
+         "expected non-scalable leading unit dims to drop");
+  assert(k < oldType.getRank() &&
+         "shape_cast cannot drop all vector dimensions");
+  VectorType newType = VectorType::get(oldType.getShape().drop_front(k),
+                                       oldType.getElementType(),
+                                       oldType.getScalableDims().drop_front(k));
+  return b.createOrFold<vector::ShapeCastOp>(loc, newType, operand);
+}
+
+/// Like `dropLeadingUnitDims` except that if all dimensions would be dropped,
+/// the single element inside that vector is extracted and returned.
+static Value dropLeadingUnitDims0DIsScalar(OpBuilder &b, Location loc,
+                                           Value operand, int64_t k) {
   auto oldType = cast<VectorType>(operand.getType());
-  assert(hasNonScalableUnitLeadingDims(oldType, dropCount) &&
+  assert(leadingDimsAreUnit(oldType, k) &&
          "expected non-scalable leading unit dims to drop");
 
-  // vector.contract rejects 0-D vector accumulators/results. When every vector
-  // dimension is dropped, use the scalar path that vector.contract accepts.
-  if (dropCount == oldType.getRank()) {
-    llvm::Repeated<int64_t> zeros(static_cast<size_t>(dropCount), 0);
+  if (k == oldType.getRank()) {
+    llvm::Repeated<int64_t> zeros(static_cast<size_t>(k), 0);
     return vector::ExtractOp::create(b, loc, operand, llvm::to_vector(zeros));
   }
 
-  return shapeCastDroppingLeadingDims(b, loc, operand, dropCount);
+  return dropLeadingUnitDims(b, loc, operand, k);
 }
 
 namespace {
@@ -156,8 +160,8 @@ struct CastAwayExtractStridedSliceLeadingOneDim
 
     Location loc = extractOp.getLoc();
 
-    Value newSrcVector =
-        shapeCastVector(rewriter, loc, extractOp.getSource(), newSrcType);
+    Value newSrcVector = rewriter.createOrFold<vector::ShapeCastOp>(
+        loc, newSrcType, extractOp.getSource());
 
     // The offsets/sizes/strides attribute can have a less number of elements
     // than the input vector's rank: it is meant for the leading dimensions.
@@ -200,10 +204,10 @@ struct CastAwayInsertStridedSliceLeadingOneDim
     // Trim leading one dimensions from both operands.
     Location loc = insertOp.getLoc();
 
-    Value newSrcVector =
-        shapeCastVector(rewriter, loc, insertOp.getValueToStore(), newSrcType);
-    Value newDstVector =
-        shapeCastVector(rewriter, loc, insertOp.getDest(), newDstType);
+    Value newSrcVector = rewriter.createOrFold<vector::ShapeCastOp>(
+        loc, newSrcType, insertOp.getValueToStore());
+    Value newDstVector = rewriter.createOrFold<vector::ShapeCastOp>(
+        loc, newDstType, insertOp.getDest());
 
     auto newOffsets = rewriter.getArrayAttr(
         insertOp.getOffsets().getValue().take_back(newDstType.getRank()));
@@ -250,10 +254,10 @@ struct CastAwayInsertLeadingOneDim : public OpRewritePattern<vector::InsertOp> {
 
     Value newSrcVector = insertOp.getValueToStore();
     if (oldSrcRank != 0)
-      newSrcVector = shapeCastVector(rewriter, loc, insertOp.getValueToStore(),
-                                     cast<VectorType>(newSrcType));
-    Value newDstVector =
-        shapeCastVector(rewriter, loc, insertOp.getDest(), newDstType);
+      newSrcVector = rewriter.createOrFold<vector::ShapeCastOp>(
+          loc, cast<VectorType>(newSrcType), insertOp.getValueToStore());
+    Value newDstVector = rewriter.createOrFold<vector::ShapeCastOp>(
+        loc, newDstType, insertOp.getDest());
 
     // New position rank needs to be computed in two steps: (1) if destination
     // type has leading unit dims, we also trim the position array accordingly,
@@ -281,7 +285,7 @@ static Value dropUnitDimsFromMask(OpBuilder &b, Location loc, Value mask,
                                   VectorType newType, AffineMap newMap) {
   // Infer the type of the new mask from the new map.
   VectorType newMaskType = inferTransferOpMaskType(newType, newMap);
-  return shapeCastVector(b, loc, mask, newMaskType);
+  return b.createOrFold<vector::ShapeCastOp>(loc, newMaskType, mask);
 }
 
 // Turns vector.transfer_read on vector with leading 1 dimensions into
@@ -372,8 +376,8 @@ struct CastAwayTransferWriteLeadingOneDim
       inBoundsAttr = rewriter.getArrayAttr(
           write.getInBoundsAttr().getValue().take_back(newType.getRank()));
 
-    auto newVector =
-        shapeCastVector(rewriter, write.getLoc(), write.getVector(), newType);
+    auto newVector = rewriter.createOrFold<vector::ShapeCastOp>(
+        write.getLoc(), newType, write.getVector());
 
     if (write.getMask()) {
       Value newMask = dropUnitDimsFromMask(rewriter, write.getLoc(),
@@ -393,16 +397,6 @@ struct CastAwayTransferWriteLeadingOneDim
 
 } // namespace
 
-namespace {
-struct ContractionOperandDropInfo {
-  AffineMap map;
-  bool needsShapeCast = false;
-  bool transposeNeeded = false;
-  bool transposeNonOuterUnitDims = false;
-  SmallVector<int64_t> permutation;
-};
-} // namespace
-
 FailureOr<Value>
 mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
                                                MaskingOpInterface maskingOp,
@@ -439,16 +433,22 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
 
   SmallVector<Value> operands = {contractOp.getLhs(), contractOp.getRhs(),
                                  contractOp.getAcc()};
-  SmallVector<ContractionOperandDropInfo> operandDropInfo;
   SmallVector<Value> newOperands;
   auto loc = contractOp.getLoc();
 
+  if (maskingOp) {
+    auto oldMaskType = cast<VectorType>(maskingOp.getMask().getType());
+    if (oldMaskType.getRank() <= 1 || dimToDrop >= oldMaskType.getRank() ||
+        !isNonScalableUnitDim(oldMaskType, dimToDrop))
+      return failure();
+  }
+
   for (const auto &it : llvm::enumerate(oldIndexingMaps)) {
     // Check if the dim to be dropped exists as a leading dim in the operand
     // if it does then we use vector.shape_cast to drop it.
-    ContractionOperandDropInfo dropInfo;
+    bool needsDrop = false;
     SmallVector<AffineExpr> results;
-    dropInfo.map = it.value();
+    auto map = it.value();
     int64_t originalZeroDim = it.value().getDimPosition(0);
     if (originalZeroDim != dimToDrop) {
       // There are two reasons to be in this path, 1. We need to
@@ -456,17 +456,19 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
       // leading. 2. The dim to be dropped does not exist and in
       // that case we dont want to add a unit transpose but we must
       // check all the indices to make sure this is the case.
+      bool transposeNeeded = false;
+      SmallVector<int64_t> perm;
       SmallVector<AffineExpr> transposeResults;
 
-      for (int64_t i = 0, e = dropInfo.map.getNumResults(); i < e; ++i) {
-        int64_t currDim = dropInfo.map.getDimPosition(i);
+      for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
+        int64_t currDim = map.getDimPosition(i);
         if (currDim == dimToDrop) {
-          dropInfo.transposeNeeded = true;
-          dropInfo.permutation.insert(dropInfo.permutation.begin(), i);
+          transposeNeeded = true;
+          perm.insert(perm.begin(), i);
           auto targetExpr = rewriter.getAffineDimExpr(currDim);
           transposeResults.insert(transposeResults.begin(), targetExpr);
         } else {
-          dropInfo.permutation.push_back(i);
+          perm.push_back(i);
           auto targetExpr = rewriter.getAffineDimExpr(currDim);
           transposeResults.push_back(targetExpr);
         }
@@ -475,31 +477,48 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
       // Checks if only the outer, unit dimensions (of size 1) are permuted.
       // Such transposes do not materially effect the underlying vector and can
       // be omitted. EG: perm [1, 0, 2] applied to vector<1x1x8xi32>
-      auto operandShape = cast<ShapedType>(operands[it.index()].getType());
-      for (auto [index, dim] : llvm::enumerate(
-               ArrayRef<int64_t>(dropInfo.permutation).drop_back(1))) {
+      bool transposeNonOuterUnitDims = false;
+      auto operandType = cast<VectorType>(operands[it.index()].getType());
+      for (auto [index, dim] :
+           llvm::enumerate(ArrayRef<int64_t>(perm).drop_back(1))) {
         if (dim != static_cast<int64_t>(index) &&
-            operandShape.getDimSize(index) != 1) {
-          dropInfo.transposeNonOuterUnitDims = true;
+            !isNonScalableUnitDim(operandType, index)) {
+          transposeNonOuterUnitDims = true;
           break;
         }
       }
 
       // Do the transpose now if needed so that we can drop the correct dim
       // with shape_cast later.
-      if (dropInfo.transposeNeeded)
-        dropInfo.map =
-            AffineMap::get(dropInfo.map.getNumDims(), 0, transposeResults,
-                           contractOp.getContext());
+      if (transposeNeeded) {
+        map = AffineMap::get(map.getNumDims(), 0, transposeResults,
+                             contractOp.getContext());
+        if (map.getDimPosition(0) == dimToDrop) {
+          bool leadingDimsCanBeDropped =
+              transposeNonOuterUnitDims
+                  ? leadingDimsAreUnitAfterPermutation(operandType, perm,
+                                                       dropDim)
+                  : leadingDimsAreUnit(operandType, dropDim);
+          if (!leadingDimsCanBeDropped)
+            return failure();
+        }
+        if (transposeNonOuterUnitDims)
+          operands[it.index()] = rewriter.createOrFold<vector::TransposeOp>(
+              loc, operands[it.index()], perm);
+      }
     }
     // We have taken care to have the dim to be dropped be
     // the leading dim. If its still not leading that means it
     // does not exist in this operand and hence we do not need a shape_cast.
-    if (dropInfo.map.getDimPosition(0) == dimToDrop)
-      dropInfo.needsShapeCast = true;
+    if (map.getDimPosition(0) == dimToDrop)
+      needsDrop = true;
+    if (needsDrop && originalZeroDim == dimToDrop &&
+        !leadingDimsAreUnit(cast<VectorType>(operands[it.index()].getType()),
+                            dropDim))
+      return failure();
 
-    for (int64_t i = 0, e = dropInfo.map.getNumResults(); i < e; ++i) {
-      int64_t currDim = dropInfo.map.getDimPosition(i);
+    for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
+      int64_t currDim = map.getDimPosition(i);
       if (currDim == dimToDrop)
         // This is the dim we are dropping.
         continue;
@@ -507,36 +526,12 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
           currDim < dimToDrop ? currDim : currDim - 1);
       results.push_back(targetExpr);
     }
-    newIndexingMaps.push_back(AffineMap::get(dropInfo.map.getNumDims() - 1, 0,
-                                             results, contractOp.getContext()));
-    if (dropInfo.needsShapeCast) {
-      auto operandType = cast<VectorType>(operands[it.index()].getType());
-      VectorType typeToDrop =
-          dropInfo.transposeNeeded && dropInfo.transposeNonOuterUnitDims
-              ? transposeVectorType(operandType, dropInfo.permutation)
-              : operandType;
-      if (!hasNonScalableUnitLeadingDims(typeToDrop, dropDim))
-        return failure();
-    }
-    operandDropInfo.push_back(std::move(dropInfo));
-  }
-
-  if (maskingOp) {
-    auto oldMaskType = cast<VectorType>(maskingOp.getMask().getType());
-    if (oldMaskType.getRank() <= 1 ||
-        !isNonScalableUnitDim(oldMaskType, dimToDrop))
-      return failure();
-  }
-
-  for (const auto &it : llvm::enumerate(operandDropInfo)) {
-    Value operand = operands[it.index()];
-    const ContractionOperandDropInfo &dropInfo = it.value();
-    if (dropInfo.transposeNeeded && dropInfo.transposeNonOuterUnitDims)
-      operand = rewriter.createOrFold<vector::TransposeOp>(
-          loc, operand, dropInfo.permutation);
-    if (dropInfo.needsShapeCast)
-      operand = dropLeadingDimsForContraction(rewriter, loc, operand, dropDim);
-    newOperands.push_back(operand);
+    newIndexingMaps.push_back(AffineMap::get(map.getNumDims() - 1, 0, results,
+                                             contractOp.getContext()));
+    newOperands.push_back(
+        needsDrop ? dropLeadingUnitDims0DIsScalar(rewriter, loc,
+                                                  operands[it.index()], dropDim)
+                  : operands[it.index()]);
   }
 
   // Depending on whether this vector.contract is masked, the replacing Op
@@ -547,8 +542,7 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
       rewriter.getArrayAttr(newIteratorTypes), contractOp.getKind());
 
   if (maskingOp) {
-    Value newMask =
-        shapeCastDroppingDim(rewriter, loc, maskingOp.getMask(), dimToDrop);
+    Value newMask = dropUnitDim(rewriter, loc, maskingOp.getMask(), dimToDrop);
 
     newOp = mlir::vector::maskOperation(rewriter, newOp, newMask);
   }
@@ -615,8 +609,8 @@ class CastAwayElementwiseLeadingOneDim : public RewritePattern {
     SmallVector<Value, 4> newOperands;
     for (Value operand : op->getOperands()) {
       if (auto opVecType = dyn_cast<VectorType>(operand.getType()))
-        newOperands.push_back(shapeCastVector(rewriter, op->getLoc(), operand,
-                                              trimLeadingOneDims(opVecType)));
+        newOperands.push_back(rewriter.createOrFold<vector::ShapeCastOp>(
+            op->getLoc(), trimLeadingOneDims(opVecType), operand));
       else
         newOperands.push_back(operand);
     }
@@ -653,7 +647,7 @@ struct CastAwayLoadLikeLeadingOneDim : public OpRewritePattern<OpTy> {
     for (Value operand : op->getOperands()) {
       if (isa<VectorType>(operand.getType())) {
         newOperands.push_back(
-            shapeCastDroppingLeadingDims(rewriter, loc, operand, nDropped));
+            dropLeadingUnitDims(rewriter, loc, operand, nDropped));
       } else {
         newOperands.push_back(operand);
       }
@@ -688,7 +682,7 @@ struct CastAwayStoreLikeLeadingOneDim : public OpRewritePattern<OpTy> {
     for (Value operand : op->getOperands()) {
       if (isa<VectorType>(operand.getType())) {
         newOperands.push_back(
-            shapeCastDroppingLeadingDims(rewriter, loc, operand, nDropped));
+            dropLeadingUnitDims(rewriter, loc, operand, nDropped));
       } else {
         newOperands.push_back(operand);
       }
diff --git a/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir b/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir
index ab2f566ff0cb0..81b68baadca55 100644
--- a/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir
+++ b/mlir/test/Dialect/Vector/vector-dropleadunitdim-transforms.mlir
@@ -283,28 +283,28 @@ func.func @cast_away_masked_contraction_with_rank1_acc(%arg0: vector<64xf32>, %a
 
 // -----
 
-// CHECK-LABEL: do_not_cast_away_contraction_with_scalable_rank1_acc
+// CHECK-LABEL: negative_cast_away_contraction_with_scalable_rank1_acc
 //  CHECK-NOT: vector.shape_cast
 //  CHECK-NOT: vector.extract
 //  CHECK-NOT: vector.broadcast
 //  CHECK-NEXT: vector.contract
 //  CHECK-NEXT: return
 
-func.func @do_not_cast_away_contraction_with_scalable_rank1_acc(%arg0: vector<64xf32>, %arg1: vector<[1]x64xf32>, %arg2: vector<[1]xf32>) -> vector<[1]xf32> {
+func.func @negative_cast_away_contraction_with_scalable_rank1_acc(%arg0: vector<64xf32>, %arg1: vector<[1]x64xf32>, %arg2: vector<[1]xf32>) -> vector<[1]xf32> {
   %0 = vector.contract {indexing_maps = [affine_map<(d0, d1) -> (d0)>, affine_map<(d0, d1) -> (d1, d0)>, affine_map<(d0, d1) -> (d1)>], iterator_types = ["reduction", "parallel"], kind = #vector.kind<add>} %arg0, %arg1, %arg2 : vector<64xf32>, vector<[1]x64xf32> into vector<[1]xf32>
   return %0 : vector<[1]xf32>
 }
 
 // -----
 
-// CHECK-LABEL: do_not_cast_away_contraction_with_scalable_operand_dim
+// CHECK-LABEL: negative_cast_away_contraction_with_scalable_operand_dim
 //  CHECK-NOT: vector.shape_cast
 //  CHECK-NOT: vector.extract
 //  CHECK-NOT: vector.broadcast
 //  CHECK-NEXT: vector.contract
 //  CHECK-NEXT: return
 
-func.func @do_not_cast_away_contraction_with_scalable_operand_dim(%arg0: vector<64xf32>, %arg1: vector<[1]x64xf32>, %arg2: vector<1xf32>) -> vector<1xf32> {
+func.func @negative_cast_away_contraction_with_scalable_operand_dim(%arg0: vector<64xf32>, %arg1: vector<[1]x64xf32>, %arg2: vector<1xf32>) -> vector<1xf32> {
   %0 = vector.contract {indexing_maps = [affine_map<(d0, d1) -> (d0)>, affine_map<(d0, d1) -> (d1, d0)>, affine_map<(d0, d1) -> (d1)>], iterator_types = ["reduction", "parallel"], kind = #vector.kind<add>} %arg0, %arg1, %arg2 : vector<64xf32>, vector<[1]x64xf32> into vector<1xf32>
   return %0 : vector<1xf32>
 }

>From bf0cbc5272fff7d5f158b409cea3f0b064570405 Mon Sep 17 00:00:00 2001
From: Krzysztof Drewniak <Krzysztof.Drewniak at amd.com>
Date: Fri, 8 May 2026 18:38:31 +0000
Subject: [PATCH 5/8] Name change

---
 .../Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp   | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
index 9ba53c978d36d..11561012aaff9 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
@@ -446,7 +446,7 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
   for (const auto &it : llvm::enumerate(oldIndexingMaps)) {
     // Check if the dim to be dropped exists as a leading dim in the operand
     // if it does then we use vector.shape_cast to drop it.
-    bool needsDrop = false;
+    bool needsCast = false;
     SmallVector<AffineExpr> results;
     auto map = it.value();
     int64_t originalZeroDim = it.value().getDimPosition(0);
@@ -511,8 +511,8 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
     // the leading dim. If its still not leading that means it
     // does not exist in this operand and hence we do not need a shape_cast.
     if (map.getDimPosition(0) == dimToDrop)
-      needsDrop = true;
-    if (needsDrop && originalZeroDim == dimToDrop &&
+      needsCast = true;
+    if (needsCast && originalZeroDim == dimToDrop &&
         !leadingDimsAreUnit(cast<VectorType>(operands[it.index()].getType()),
                             dropDim))
       return failure();
@@ -529,7 +529,7 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
     newIndexingMaps.push_back(AffineMap::get(map.getNumDims() - 1, 0, results,
                                              contractOp.getContext()));
     newOperands.push_back(
-        needsDrop ? dropLeadingUnitDims0DIsScalar(rewriter, loc,
+        needsCast ? dropLeadingUnitDims0DIsScalar(rewriter, loc,
                                                   operands[it.index()], dropDim)
                   : operands[it.index()]);
   }

>From 7348f5b6a270acd29e8916866f7ff985c9f84ead Mon Sep 17 00:00:00 2001
From: Krzysztof Drewniak <Krzysztof.Drewniak at amd.com>
Date: Fri, 8 May 2026 20:21:26 +0000
Subject: [PATCH 6/8] Simplify things since we don't need transposes

---
 .../Transforms/VectorDropLeadUnitDim.cpp      | 138 ++++++++++--------
 1 file changed, 77 insertions(+), 61 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
index 11561012aaff9..c408cc3e2a745 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
@@ -7,7 +7,9 @@
 //===----------------------------------------------------------------------===//
 
 #include <numeric>
+#include <utility>
 
+#include "mlir/Dialect/Utils/IndexingUtils.h"
 #include "mlir/Dialect/Utils/StructuredOpsUtils.h"
 #include "mlir/Dialect/Vector/IR/VectorOps.h"
 #include "mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h"
@@ -115,6 +117,19 @@ static Value dropLeadingUnitDims(OpBuilder &b, Location loc, Value operand,
   return b.createOrFold<vector::ShapeCastOp>(loc, newType, operand);
 }
 
+/// Returns the vector type obtained by applying `permutation` to `type`.
+static VectorType permuteVectorType(VectorType type,
+                                    ArrayRef<int64_t> permutation) {
+  assert(static_cast<int64_t>(permutation.size()) == type.getRank() &&
+         "expected a permutation matching the operand rank");
+  SmallVector<int64_t> permutedShape =
+      applyPermutation(type.getShape(), permutation);
+  SmallVector<bool> permutedScalableDims =
+      applyPermutation(type.getScalableDims(), permutation);
+  return VectorType::get(permutedShape, type.getElementType(),
+                         permutedScalableDims);
+}
+
 /// Like `dropLeadingUnitDims` except that if all dimensions would be dropped,
 /// the single element inside that vector is extracted and returned.
 static Value dropLeadingUnitDims0DIsScalar(OpBuilder &b, Location loc,
@@ -128,7 +143,10 @@ static Value dropLeadingUnitDims0DIsScalar(OpBuilder &b, Location loc,
     return vector::ExtractOp::create(b, loc, operand, llvm::to_vector(zeros));
   }
 
-  return dropLeadingUnitDims(b, loc, operand, k);
+  VectorType newType = VectorType::get(oldType.getShape().drop_front(k),
+                                       oldType.getElementType(),
+                                       oldType.getScalableDims().drop_front(k));
+  return vector::ShapeCastOp::create(b, loc, newType, operand);
 }
 
 namespace {
@@ -397,6 +415,15 @@ struct CastAwayTransferWriteLeadingOneDim
 
 } // namespace
 
+namespace {
+struct VectorContractOperandCastPlan {
+  AffineMap map;
+  SmallVector<int64_t> permutation;
+  bool dropLeadingUnitDim = false;
+  bool permuteOperand = false;
+};
+} // namespace
+
 FailureOr<Value>
 mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
                                                MaskingOpInterface maskingOp,
@@ -433,6 +460,7 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
 
   SmallVector<Value> operands = {contractOp.getLhs(), contractOp.getRhs(),
                                  contractOp.getAcc()};
+  SmallVector<VectorContractOperandCastPlan> operandCastPlans;
   SmallVector<Value> newOperands;
   auto loc = contractOp.getLoc();
 
@@ -446,79 +474,56 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
   for (const auto &it : llvm::enumerate(oldIndexingMaps)) {
     // Check if the dim to be dropped exists as a leading dim in the operand
     // if it does then we use vector.shape_cast to drop it.
-    bool needsCast = false;
+    VectorContractOperandCastPlan plan;
     SmallVector<AffineExpr> results;
-    auto map = it.value();
-    int64_t originalZeroDim = it.value().getDimPosition(0);
+    plan.map = it.value();
+    int64_t originalZeroDim = plan.map.getDimPosition(0);
     if (originalZeroDim != dimToDrop) {
       // There are two reasons to be in this path, 1. We need to
-      // transpose the operand to make the dim to be dropped
+      // permute the operand type to make the dim to be dropped
       // leading. 2. The dim to be dropped does not exist and in
-      // that case we dont want to add a unit transpose but we must
+      // that case we dont want to add a unit permutation but we must
       // check all the indices to make sure this is the case.
-      bool transposeNeeded = false;
-      SmallVector<int64_t> perm;
-      SmallVector<AffineExpr> transposeResults;
+      SmallVector<AffineExpr> permutedResults;
 
-      for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
-        int64_t currDim = map.getDimPosition(i);
+      for (int64_t i = 0, e = plan.map.getNumResults(); i < e; ++i) {
+        int64_t currDim = plan.map.getDimPosition(i);
         if (currDim == dimToDrop) {
-          transposeNeeded = true;
-          perm.insert(perm.begin(), i);
+          plan.permuteOperand = true;
+          plan.permutation.insert(plan.permutation.begin(), i);
           auto targetExpr = rewriter.getAffineDimExpr(currDim);
-          transposeResults.insert(transposeResults.begin(), targetExpr);
+          permutedResults.insert(permutedResults.begin(), targetExpr);
         } else {
-          perm.push_back(i);
+          plan.permutation.push_back(i);
           auto targetExpr = rewriter.getAffineDimExpr(currDim);
-          transposeResults.push_back(targetExpr);
+          permutedResults.push_back(targetExpr);
         }
       }
 
-      // Checks if only the outer, unit dimensions (of size 1) are permuted.
-      // Such transposes do not materially effect the underlying vector and can
-      // be omitted. EG: perm [1, 0, 2] applied to vector<1x1x8xi32>
-      bool transposeNonOuterUnitDims = false;
-      auto operandType = cast<VectorType>(operands[it.index()].getType());
-      for (auto [index, dim] :
-           llvm::enumerate(ArrayRef<int64_t>(perm).drop_back(1))) {
-        if (dim != static_cast<int64_t>(index) &&
-            !isNonScalableUnitDim(operandType, index)) {
-          transposeNonOuterUnitDims = true;
-          break;
-        }
-      }
-
-      // Do the transpose now if needed so that we can drop the correct dim
-      // with shape_cast later.
-      if (transposeNeeded) {
-        map = AffineMap::get(map.getNumDims(), 0, transposeResults,
-                             contractOp.getContext());
-        if (map.getDimPosition(0) == dimToDrop) {
-          bool leadingDimsCanBeDropped =
-              transposeNonOuterUnitDims
-                  ? leadingDimsAreUnitAfterPermutation(operandType, perm,
-                                                       dropDim)
-                  : leadingDimsAreUnit(operandType, dropDim);
-          if (!leadingDimsCanBeDropped)
+      // Update the map now so that the later shape_cast drops the correct dim.
+      if (plan.permuteOperand) {
+        plan.map = AffineMap::get(plan.map.getNumDims(), 0, permutedResults,
+                                  contractOp.getContext());
+        if (plan.map.getDimPosition(0) == dimToDrop) {
+          auto operandType = cast<VectorType>(operands[it.index()].getType());
+          if (!leadingDimsAreUnitAfterPermutation(operandType, plan.permutation,
+                                                  dropDim))
             return failure();
         }
-        if (transposeNonOuterUnitDims)
-          operands[it.index()] = rewriter.createOrFold<vector::TransposeOp>(
-              loc, operands[it.index()], perm);
       }
     }
     // We have taken care to have the dim to be dropped be
     // the leading dim. If its still not leading that means it
     // does not exist in this operand and hence we do not need a shape_cast.
-    if (map.getDimPosition(0) == dimToDrop)
-      needsCast = true;
-    if (needsCast && originalZeroDim == dimToDrop &&
+    if (plan.map.getDimPosition(0) == dimToDrop)
+      plan.dropLeadingUnitDim = true;
+    if (plan.dropLeadingUnitDim && originalZeroDim == dimToDrop &&
         !leadingDimsAreUnit(cast<VectorType>(operands[it.index()].getType()),
                             dropDim))
       return failure();
 
-    for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
-      int64_t currDim = map.getDimPosition(i);
+    for (int64_t i = 0, e = plan.map.getNumResults(); i < e; ++i) {
+      int64_t currDim = plan.map.getDimPosition(i);
       if (currDim == dimToDrop)
         // This is the dim we are dropping.
         continue;
@@ -526,12 +531,23 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
           currDim < dimToDrop ? currDim : currDim - 1);
       results.push_back(targetExpr);
     }
-    newIndexingMaps.push_back(AffineMap::get(map.getNumDims() - 1, 0, results,
-                                             contractOp.getContext()));
-    newOperands.push_back(
-        needsCast ? dropLeadingUnitDims0DIsScalar(rewriter, loc,
-                                                  operands[it.index()], dropDim)
-                  : operands[it.index()]);
+    newIndexingMaps.push_back(AffineMap::get(plan.map.getNumDims() - 1, 0,
+                                             results, contractOp.getContext()));
+    operandCastPlans.push_back(std::move(plan));
+  }
+
+  for (auto [plan, operand] : llvm::zip_equal(operandCastPlans, operands)) {
+    Value newOperand = operand;
+    if (plan.permuteOperand)
+      newOperand = rewriter.createOrFold<vector::ShapeCastOp>(
+          loc,
+          permuteVectorType(cast<VectorType>(newOperand.getType()),
+                            plan.permutation),
+          newOperand);
+    if (plan.dropLeadingUnitDim)
+      newOperand =
+          dropLeadingUnitDims0DIsScalar(rewriter, loc, newOperand, dropDim);
+    newOperands.push_back(newOperand);
   }
 
   // Depending on whether this vector.contract is masked, the replacing Op
@@ -547,13 +563,13 @@ mlir::vector::castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
     newOp = mlir::vector::maskOperation(rewriter, newOp, newMask);
   }
 
-  if (isa<VectorType>(newOp->getResults()[0].getType()))
-    return vector::ShapeCastOp::create(rewriter, loc,
+  if (!isa<VectorType>(newOp->getResults()[0].getType()))
+    return vector::BroadcastOp::create(rewriter, loc,
                                        contractOp->getResultTypes()[0],
                                        newOp->getResults()[0])
         .getResult();
 
-  return vector::BroadcastOp::create(rewriter, loc,
+  return vector::ShapeCastOp::create(rewriter, loc,
                                      contractOp->getResultTypes()[0],
                                      newOp->getResults()[0])
       .getResult();
@@ -563,8 +579,8 @@ namespace {
 
 /// Turns vector.contract on vector with leading 1 dimensions into
 /// vector.shape_cast followed by vector.contract on vector without leading
-/// 1 dimensions. Also performs transpose of lhs and rhs operands if required
-/// prior to the shape cast.
+/// 1 dimensions. Non-leading unit dimensions are dropped via direct
+/// shape_casts.
 struct CastAwayContractionLeadingOneDim
     : public MaskableOpRewritePattern<vector::ContractionOp> {
   using MaskableOpRewritePattern::MaskableOpRewritePattern;

>From ab5f5a7d9b28f6940e37bf8ecc90c890d97ed927 Mon Sep 17 00:00:00 2001
From: Krzysztof Drewniak <Krzysztof.Drewniak at amd.com>
Date: Fri, 8 May 2026 20:38:25 +0000
Subject: [PATCH 7/8] Stray to_vector

---
 mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
index c408cc3e2a745..a1ecd1086fdbe 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
@@ -140,7 +140,7 @@ static Value dropLeadingUnitDims0DIsScalar(OpBuilder &b, Location loc,
 
   if (k == oldType.getRank()) {
     llvm::Repeated<int64_t> zeros(static_cast<size_t>(k), 0);
-    return vector::ExtractOp::create(b, loc, operand, llvm::to_vector(zeros));
+    return vector::ExtractOp::create(b, loc, operand, zeros);
   }
 
   VectorType newType = VectorType::get(oldType.getShape().drop_front(k),

>From 7717cf468fb424ce2f425c2042a94e1109e3b563 Mon Sep 17 00:00:00 2001
From: Krzysztof Drewniak <Krzysztof.Drewniak at amd.com>
Date: Fri, 8 May 2026 20:52:17 +0000
Subject: [PATCH 8/8] Ok, actually, we need OpFoldResult range

---
 mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
index a1ecd1086fdbe..8be2fa071a9d7 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorDropLeadUnitDim.cpp
@@ -139,7 +139,7 @@ static Value dropLeadingUnitDims0DIsScalar(OpBuilder &b, Location loc,
          "expected non-scalable leading unit dims to drop");
 
   if (k == oldType.getRank()) {
-    llvm::Repeated<int64_t> zeros(static_cast<size_t>(k), 0);
+    SmallVector<int64_t> zeros(k, static_cast<int64_t>(0));
     return vector::ExtractOp::create(b, loc, operand, zeros);
   }
 



More information about the Mlir-commits mailing list