[mlir][sparse] use loop emitter to generate loop in sparsification

[wrengr]

Be sure to rebase over D136123, since that made a lot of changes to Sparsification.cpp
`

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.cpp

137

Why not use isZeroRankedTensorOrScalar here instead? If it's intentional, then you should add a comment explaining why rank-zero tensors don't continue.

219–221

Why not use MutableArrayRef instead of SmallVectorImpl?

228

Thanks for the clarifying rename :)

367

again, would probably be better to use MutableArrayRef unless you really need it to be a SmallVectorImpl

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.h

430–444

Should move all these fields to the bottom of the class together with the other fields

466

Should move this field to the bottom of the class, together with all the other fields.

[Peiming]

rebase

[Peiming]

address comments from Wren + rebase

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.cpp

219–221

Good suggestion! Simply because I do not know there is a MutableArrayRef.

It is probably a better way, as I can then have a default empty value for the parameter. I will try!

228

;-)

[Peiming]

rebase

[Peiming]

rebase..

[aartbik]

I got somewhere deeper into this revision! I still want to do a very careful 1;1 comparison with old sparsification code and new emitter but I am confident to get to that early next week!

mlir/include/mlir/Dialect/SparseTensor/Utils/Merger.h

235 ↗

(On Diff #468990)

we have the syntheticTensor field for that!

(btw you can send this single file change out by itself for faster review!
perhaps add a unit test to cover the new code in that revision)

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.h

282

It is still very confusing why L243 up to new emitter show up here as new?
You can you rebase to latest main?

305

... that (co-)iterate over sparse tensors.

(otherwise the () do not make much sense)

308

..generate the following..

326

this is very confusing given that we also generate for loops

just like it 3x

loopEmiter.exitCurrentLoop(); exit second k-loop
loopEmiter.exitCurrentLoop(); exit j-loop
loopEmiter.exitCurrentLoop();/ / exit i-loop

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.cpp

732

don't break the == 8
(also, unrelated change in own quick revision?)

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.h

308

Note that loops are really over lattices and tensor index expressions, so the dimensions may not always line up.

and please use tensor0_0, tensor0_1 etc. see below

333

when initializing
*ing form

334

space after comma

347

remove L308-310, does not add to much given total # methods

365

syntax p0- > end is confusing

use C flavored explanation

for (int i = ..., i < ...; i++)

372

Comment that this ends a loop
(sort of obvious but still nice to say something)

381

Note that in the original code and dump, we use tensor_d for this and i_tensor_idx for the loop index (see dumpBits()). It would be helpful for me as reviewer to see that slightly more familiar syntax for the constructs

393

co-iteration in comment and CoIteration in method name

402

Returns

408

Gets

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

162

I can't find my original question in all the history, but why do we need this change in this revision.
It should be NFC for now, right?

[Peiming]

rebase against main

[Peiming]

address comments.

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.cpp

732

I think it is automatically fomatted by emacs (clangd) ...

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.h

282

I rebased... and they are still here...

381

Yes, but here there is no concept of loop idx

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

162

This is a complicated story...

This does not matter whether we set it or not previously. In the previous implementation, you used sizes[i] (i is the loop index) for the loop bound. Note that, however, the sizes[i] are not initialized by the dimension of the dense tensor (because this is a complex affine, there is no corresponding loop index for it).

However, in loop emitter, it requires you to pass tid + dim, in that case, setting the level to dense will have the merger favors it over undefined dimension, and the loop bound will be computed from the dimension of the dense tensor (which is wrong!)

Apart from that, I think to avoid setting dimension level type makes more sense as well, because the complex affine does not actually corresponding to any loop index either, and the behavior is more consistent for affine expression on sparse tensor too.

[Peiming]

small fix

[Peiming]

rebase

[Peiming]

remove dup code.

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.h

282

You are right! Seems git failed to recognized this part is moved? (or I made some mistake merging two conflicts).

They should be removed.

[Peiming]

rebase

[Peiming]

rebase.

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.h

364

tensor_t_dim notation, also at L364, 366?

464

period at end, but see below for name/comment

465

Shall we make this just "hasOutput". By convention that is indeed always push to the end.
But the isLastOutput name suggests that this is a very special situation, where we just have output/no-output conditions...

466

Just for my own sanity, can you please indicate if I got the following right given the original names?

/// Universal dense indices and upper bounds (by index). The loops array
/// is updated with the value of the universal dense index in the current
/// loop. The sizes array is set once with the inferred dimension sizes.
std::vector<Value> loops;   // now in getLoopIdxValue?  what do you keep in tensors? [note that loops correspond to lattices more than tensors, hence my original name]
std::vector<Value> sizes;  // not here?
/// Buffers for storing dense and sparse numerical values (by tensor).
/// This array is set once during bufferization of all tensors.
std::vector<Value> buffers;    // now valBuffer?
/// Sparse storage schemes (1-D): pointers and indices (by tensor and index).
/// This array is set once during bufferization of all sparse tensors.
std::vector<std::vector<Value>> pointers;  // now ptrBuffer?
std::vector<std::vector<Value>> indices;    // now idxBuffer?
/// Sparse iteration information (by tensor and index). These arrays
/// are updated to remain current within the current loop.
std::vector<std::vector<Value>> highs;  // still highs?
std::vector<std::vector<Value>> pidxs;   // still pidxs?
std::vector<std::vector<Value>> idxs;    // now coords?

Note that I am not against a renaming per se, especially if the new name is better ;-)
But in changes like this, it would have been easier to first move everything over "as is"
and then rename later

475

This comment seems out of place now? The fields below are pointers/indices/values

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

86

Yeah, L86-125 feels really out of place now. But I am okay cleaning that up after this revision

452

can you try not to break the tensor index expressions in the comment, so it is a bit easier to read

461

Bit confusing comment. There is only one output, but the two references to it should match ;-)

1088

note that the original had a " if (!inits[b]) continue" here
I suspect this is still covered by the undef case now, although it seems a change in behavior?

1167

I don't get this. In general, I like how sparsification has become simpler due to loop emitter.
But this huge block of code in something that really just read like "startloop" .

Why don't we keep genLocals at least? That would also make the diff a lot easier to read?

1177

commented out? I would prefer to keep the method, since start/endLoopSeq and start/endLoop where originally intended to remain small
(sort of the very first loop emitter ;-)

mlir/test/Dialect/SparseTensor/sparse_2d.mlir

965

what happened here? missing > at end, but also different tensor?

1018

probably same reason?

1098

probably same reason?

mlir/test/Dialect/SparseTensor/sparse_3d.mlir

1129

missing > at end?

mlir/test/Dialect/SparseTensor/sparse_perm.mlir

68–70

did the order change here?

mlir/test/Dialect/SparseTensor/sparse_reshape.mlir

121

maybe DAG the load/add

269–270

maybe DAG the load/add

[Peiming]

address comments from Aart.

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.h

364

I delete L366 as we do not prepare for the next dimension now per offline discussion.

466

You are correct about all the variables names.

sizes are eliminated (or merged with highs for dense tensors)

loops are now managed with loopStack as loop emitter does not know the total number of loops it need to generated at beginning.

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

86

SG.

1088

I think it is because I change the loop from
for(int i = 0; i < bit.size(); i++) to iterator based loop.

The iterator based loop only visits set bit, so all the bits are guaranteed to be inited in the loop body.

1167

We still need to do the same translation even we keep the genLocals. It is because loop emitter and lattices uses different way to denote a loop.

The whole thing that the following blocks tries to do is translate between bit set -> vector<tid + dim>, but probably a until function for this will make it more readable (e.g., translateBitsToTidDimPair)?

WDYT?

1177

I actually prefer deleting it, so that the internal state of loop emitter are less likely to be broken.

But we can have some discussion on this.

mlir/test/Dialect/SparseTensor/sparse_2d.mlir

965

Yes, because we there might be multiple tensor_dim for the same loop index. Pick arbitrary one is fine.

mlir/test/Dialect/SparseTensor/sparse_3d.mlir

1129

No, becuase it is a sparse tensor... I was being lazy to include all the sparse encoding...

Let me know if you want me to include the entire string.

mlir/test/Dialect/SparseTensor/sparse_perm.mlir

68–70

No, the order does not change, for the same reason above (multiple tensors for the same loop idx and loop emitter happens to use a different one), I have to adjust the loop bound variable accordingly in CHECK tests.

[Peiming]

minor fixes.

[Peiming]

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

1167

Or maybe merger should be responsible for the translation? as the map should be managed by it as well.

[Peiming]

remove out dated comments.

[Peiming]

use util function to translate bits -> tid + dim.

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

1088

Ah neat, of course!

mlir/test/Dialect/SparseTensor/sparse_3d.mlir

1129

Ah, well, you can use

#sparse_tensor.encoding<{{.*}}>>

to avoid clutter, but the missing > at the end looks like it was a typo ;-)

[Peiming]

fix CHECK tests.

[Peiming]

fix typo

[aartbik]

Thanks for your patience during the review, Peiming!
And now, SHIP IT!

From now on, we prefer our loop emitting to be done in its own class ;-)

[Peiming]

fix CHECK test

[Peiming]

rebase

[Peiming]

rebase

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.h

 
 namespace sparse_tensor {
 
 /// Shorthand aliases for the `emitCInterface` argument to `getFunc()`,
 /// `createFuncCall()`, and `replaceOpWithFuncCall()`.
 enum class EmitCInterface : bool { Off = false, On = true };
 
 //===----------------------------------------------------------------------===//
-// SparseTensorLoopEmiter class, manages sparse tensors and helps to generate
-// loop structure to (co-iterate) sparse tensors.
-//
-// An example usage:
-// To generate following loops over T1<?x?> and T2<?x?>
-//
-// for i in T1[0] {
-//   for j : T2[0] {
-//     for k : T1[1] {}
-//     for k : T2[1] {}
-//   }
-// }
-//
-// One can use
-//
-// SparseTensorLoopEmiter loopEmiter({T1, T1});
-// loopEmiter.initializeLoopEmit();
-// loopEmiter.enterLoopOverTensorAtDim(T1, 0);
-// loopEmiter.enterLoopOverTensorAtDim(T2, 0);
-// loopEmiter.enterLoopOverTensorAtDim(T1, 1);
-// loopEmiter.exitCurrentLoop();
-// loopEmiter.enterLoopOverTensorAtDim(T2, 1);
-// for 0 -> 3:
-//    loopEmiter.exitCurrentLoop();
-//===----------------------------------------------------------------------===//
-
-// TODO: Sparsification should also rely on this class to generate loops.
-class SparseTensorLoopEmitter {
-public:
-  /// Constructor: take an array of tensors inputs, on which the generated loops
-  /// will iterate on. The index of the tensor in the array is also the
-  /// tensor id (tid) used in related functions.
-  explicit SparseTensorLoopEmitter(ValueRange tensors,
-                                   bool isLastOutput = false);
-
-  ///
-  /// Core functions.
-  ///
-
-  /// Starts a loop emitting session:
-  /// 1. Generates all the buffers needed to iterate tensors.
-  /// 2. Generates the lo/hi bounds to iterate tensors[0].
-  void initializeLoopEmit(OpBuilder &builder, Location loc);
-
-  // TODO: Gets rid of `dim` in the argument list? Track the dimension we
-  // are currently at internally. Then it would be enterNextDimForTensor.
-
-  /// Emits loop over tensor[dim], it assumes that loops between
-  /// tensor[0...dim - 1] have already been generated.
-  /// It also prepares to enter tensor[dim + 1].
-  Operation *enterLoopOverTensorAtDim(OpBuilder &builder, Location loc,
-                                      size_t tid, size_t dim,
-                                      ArrayRef<Value> reduc = {});
-
-  /// Emits a coiteration loop over a set of tensors.
-  // TODO: not yet implemented
-  void enterCoiterationOverTensorsAtDims(OpBuilder &builder, Location loc,
-                                         ArrayRef<size_t> ts,
-                                         ArrayRef<size_t> ds);
-
-  /// Emits extra locals, since the locals might not be in simplified lattices
-  /// point used to generate the loops, but are still required to generates
-  /// expressions.
-  Value emitExtraLocalsForTensorsAtDims(OpBuilder &builder, Location loc,
-                                        size_t tid, size_t dim);
-
-  void exitCurrentLoop();
-
-  /// Return the array of coordinate for all the loop generated till now.
-  void getCoordinateArray(SmallVectorImpl<Value> &coords) {
-    for (auto &l : loopStack)
-      coords.push_back(l.idx);
-  }
-
-  ///
-  /// Getters.
-  ///
-
-  Value getTensorValueBuffer(size_t tid) { return valBuffer[tid]; }
-  Value getLastLevelTensorPointerIndex(size_t tid) {
-    return pidxs[tid].back();
-  };
-
-private:
-  struct LoopLevelInfo {
-    LoopLevelInfo(ArrayRef<size_t> ts, ArrayRef<size_t> ds, Value idx)
-        : tensors(ts), dims(ds), idx(idx) {}
-    llvm::SmallVector<size_t, 4> tensors;
-    llvm::SmallVector<size_t, 4> dims;
-    Value idx;
-  };
-
-  /// Return false if tid[dim] is a dense dimension that does not need to be
-  /// prepared (to be used by sparsification for needUniv).
-  bool prepareLoopOverTensorAtDim(OpBuilder &builder, Location loc, size_t tid,
-                                  size_t dim);
-
-  /// Input (TODO: and output) tensors.
-  std::vector<Value> tensors;
-  /// The dim type array for each tensor.
-  std::vector<std::vector<DimLevelType>> dims;
-  /// Sparse iteration information (by tensor and dim). These arrays
-  /// are updated to remain current within the current loop.
-  std::vector<std::vector<Value>> pidxs;
-  std::vector<std::vector<Value>> coord;
-  std::vector<std::vector<Value>> highs;
-  /// Universal dense indices and upper bounds (by index). The sizes array is
-  /// set once with the inferred dimension sizes.
-  std::vector<std::vector<Value>> sizes;
-  std::vector<std::vector<Value>> ptrBuffer; // to_pointers
-  std::vector<std::vector<Value>> idxBuffer; // to_indices
-  std::vector<Value> valBuffer;              // to_value
-
-  bool isLastOutput; // Is the last tensor output tensor
-  std::vector<LoopLevelInfo> loopStack;
-  // TODO: not yet used, it should track the current level for each tensor
-  // to help eliminate `dim` paramters from above APIs.
-  std::vector<size_t> curLv;
-};
-
-//===----------------------------------------------------------------------===//
 // ExecutionEngine/SparseTensorUtils helper functions.
 //===----------------------------------------------------------------------===//
 
 /// Converts an overhead storage bitwidth to its internal type-encoding.
 OverheadType overheadTypeEncoding(unsigned width);
 
 /// Converts an overhead storage type to its internal type-encoding.
 OverheadType overheadTypeEncoding(Type tp);

 }
 
 /// Generates a constant of the internal dimension level type encoding.
 inline Value constantDimLevelTypeEncoding(OpBuilder &builder, Location loc,
                                           DimLevelType dlt) {
   return constantI8(builder, loc, static_cast<uint8_t>(dlt));
 }
 
+inline bool isZeroRankedTensorOrScalar(Type type) {
+  auto rtp = type.dyn_cast<RankedTensorType>();
+  return !rtp || rtp.getRank() == 0;
+}
+
+//===----------------------------------------------------------------------===//
+// SparseTensorLoopEmiter class, manages sparse tensors and helps to generate
+// loop structure to (co)-iterate sparse tensors.
+//
+// An example usage:
+// To generate the following loops over T1<?x?> and T2<?x?>
+//
+// for i in TENSOR_1_0 {
+//   for j : TENSOR_2_0 {
+//     for k : TENSOR_1_1 {}
+//     for k : TENSOR_2_1 {}
+//   }
+// }
+//
+// One can use
+//
+// SparseTensorLoopEmiter loopEmiter({T1, T1});
+// loopEmiter.initializeLoopEmit();
+// loopEmiter.enterLoopOverTensorAtDim(T1, 0);
+// loopEmiter.enterLoopOverTensorAtDim(T2, 0);
+// loopEmiter.enterLoopOverTensorAtDim(T1, 1);
+// loopEmiter.exitCurrentLoop();
+// loopEmiter.enterLoopOverTensorAtDim(T2, 1);
+// loopEmiter.exitCurrentLoop(); // exit k
+// loopEmiter.exitCurrentLoop(); // exit j
+// loopEmiter.exitCurrentLoop(); // exit i
+//===----------------------------------------------------------------------===//
+
+// TODO: Sparsification should also rely on this class to generate loops.
+class SparseTensorLoopEmitter {
+public:
+  /// Optional callback function to setup dense output tensors when
+  /// initializing the loop emitter (e.g., to fill a dense output with zeros).
+  using OutputUpdater = function_ref<Value(OpBuilder &builder, Location loc,
+                                           Value memref, Value tensor)>;
+
+  /// Constructor: take an array of tensors inputs, on which the generated loops
+  /// will iterate on. The index of the tensor in the array is also the
+  /// tensor id (tid) used in related functions.
+  /// If isSparseOut is set, loop emitter assume that the sparse output tensor
+  /// is empty, and will always generate loops on it based on the dim sizes.
+  explicit SparseTensorLoopEmitter(ValueRange tensors, bool hasOutput = false,
+                                   bool isSparseOut = false);
+
+  /// Starts a loop emitting session by generating all the buffers needed to
+  /// iterate tensors.
+  void initializeLoopEmit(OpBuilder &builder, Location loc,
+                          OutputUpdater updater = nullptr);
+
+  /// Enters a new loop sequence, the loops within the same sequence starts from
+  /// the break points of previous loop instead of starting over from 0.
+  /// e.g.,
+  /// {
+  ///   // loop sequence start.
+  ///   p0 = while(xxx)
+  ///     ...
+  ///     break p0
+  ///
+  ///   // Starts loop from p0
+  ///   for (i = p0; i < end; i++)
+  ///     ...
+  ///   // loop sequence end.
+  /// }
+  void enterNewLoopSeq(OpBuilder &builder, Location loc, ArrayRef<size_t> tids,
+                       ArrayRef<size_t> dims);
+
+  // exit the current loop sequence, this will reset universal index to 0.
+  void exitCurrentLoopSeq() {
+    assert(loopSeqStack.size() == loopStack.size() + 1);
+    loopSeqStack.pop_back();
+  }
+
+  // TODO: Gets rid of `dim` in the argument list? Track the dimension we
+  // are currently at internally. Then it would be enterNextDimForTensor.
+  // Still need a way to specify the dim for non annoated dense tensor though,
+  // as it can be accessed out of order.
+  /// Emits loop over tensor_tid_dim, it assumes that loops between
+  /// tensor_tid_[0, dim - 1] have already been generated.
+  /// The function will also perform in-place update on the `reduc` vector to
+  /// return the reduction variable used inside the generated loop.
+  Operation *enterLoopOverTensorAtDim(OpBuilder &builder, Location loc,
+                                      size_t tid, size_t dim,
+                                      MutableArrayRef<Value> reduc = {},
+                                      bool isParallel = false,
+                                      ArrayRef<size_t> extraTids = {},
+                                      ArrayRef<size_t> extraDims = {});
+
+  /// Emits a co-iteration loop over a set of tensors.
+  Operation *enterCoIterationOverTensorsAtDims(
+      OpBuilder &builder, Location loc, ArrayRef<size_t> tids,
+      ArrayRef<size_t> dims, bool needsUniv, MutableArrayRef<Value> reduc = {},
+      ArrayRef<size_t> extraTids = {}, ArrayRef<size_t> extraDims = {});
+
+  SmallVector<Value, 2> exitCurrentLoop(OpBuilder &builder, Location loc,
+                                        ArrayRef<Value> reduc = {});
+
+  /// Returns the array of coordinate for all the loop generated till now.
+  void getCoordinateArray(SmallVectorImpl<Value> &coords) const {
+    for (auto &l : loopStack)
+      coords.push_back(l.iv);
+  }
+
+  /// Gets loop induction variable at the given level.
+  Value getLoopIV(size_t level) const {
+    if (level < loopStack.size())
+      return loopStack[level].iv;
+    return nullptr;
+  }
+
+  ///
+  /// Getters.
+  ///
+  const std::vector<std::vector<Value>> &getPidxs() const { return pidxs; };
+  const std::vector<std::vector<Value>> &getCoord() const { return coord; };
+  const std::vector<std::vector<Value>> &getHighs() const { return highs; };
+  const std::vector<std::vector<Value>> &getPtrBuffer() const {
+    return ptrBuffer;
+  };
+  const std::vector<std::vector<Value>> &getIdxBuffer() const {
+    return idxBuffer;
+  };
+  const std::vector<Value> &getValBuffer() const { return valBuffer; };
+
+private:
+  struct LoopLevelInfo {
+    LoopLevelInfo(ArrayRef<size_t> tids, ArrayRef<size_t> dims, Operation *loop,
+                  Value iv)
+        : tids(tids), dims(dims), loop(loop), iv(iv) {}
+    // TODO: maybe use a vector<pair> for tid and dim?
+    // The set of tensors that the loop is operating on
+    const llvm::SmallVector<size_t, 4> tids;
+    // The corresponding dims for the tensors
+    const llvm::SmallVector<size_t, 4> dims;
+    const Operation *loop; // the loop operation
+    const Value iv;        // the induction variable for the loop
+  };
+
+  /// Linearizes address for dense dimension (i.e., p = (i * d0) + j).
+  Value genAddress(OpBuilder &builder, Location loc, size_t tid, size_t dim,
+                   Value iv) {
+    Value p = dim == 0 ? constantIndex(builder, loc, 0) : pidxs[tid][dim - 1];
+    Value mul = builder.create<arith::MulIOp>(loc, highs[tid][dim], p);
+    Value add = builder.create<arith::AddIOp>(loc, mul, iv);
+    return add;
+  }
+
+  bool isOutputTensor(size_t tid) {
+    return hasOutput && tid == tensors.size() - 1;
+  }
+
+  /// Setups [lo, hi] for iterating tensor[dim], it assumes that tensor[0
+  /// ...dims-1] has already been setup.
+  void prepareLoopOverTensorAtDim(OpBuilder &builder, Location loc, size_t tid,
+                                  size_t dim);
+
+  /// Emits extra locals, since the locals might not be in simplified lattices
+  /// point used to generate the loops, but are still required to generates
+  /// expressions.
+  void emitExtraLocalsForTensorsAtDenseDims(OpBuilder &builder, Location loc,
+                                            ArrayRef<size_t> tids,
+                                            ArrayRef<size_t> dims);
+
+  /// Exits a for loop, returns the reduction results, e.g.,
+  /// %ret = for () {
+  ///   ...
+  ///   yield %val
+  /// }
+  /// Return %ret to user, while %val is provided by users (`reduc`)
+  SmallVector<Value, 2> exitForLoop(OpBuilder &builder, Location loc,
+                                    ArrayRef<Value> reduc);
+
+  /// Exits a while loop, returns the reduction results.
+  SmallVector<Value, 2> exitCoiterationLoop(OpBuilder &builder, Location loc,
+                                            ArrayRef<Value> reduc);
+
+  // Whether the loop emitter needs to treat the last tensor as the output
+  // tensor.
+  bool hasOutput;
+  /// Input and (optional) output tensors.
+  std::vector<Value> tensors;
+  /// The dim type array for each tensor.
+  std::vector<std::vector<DimLevelType>> dimTypes;
+  /// Sparse iteration information (by tensor and dim). These arrays
+  /// are updated to remain current within the current loop.
+  std::vector<std::vector<Value>> pidxs;
+  std::vector<std::vector<Value>> coord;
+  std::vector<std::vector<Value>> highs;
+  std::vector<std::vector<Value>> ptrBuffer; // to_pointers
+  std::vector<std::vector<Value>> idxBuffer; // to_indices
+  std::vector<Value> valBuffer;              // to_value
+
+  // Loop Stack, stores the information of all the nested loops that are alive.
+  std::vector<LoopLevelInfo> loopStack;
+
+  // Loop Sequence Stack, stores the unversial index for the current loop
+  // sequence.
+  std::vector<Value> loopSeqStack;
+
+  // TODO: not yet used, it should track the current level for each tensor
+  // to help eliminate `dim` paramters from above APIs.
+  // std::vector<size_t> curLv;
+};
+
 } // namespace sparse_tensor
 } // namespace mlir
 
 #endif // MLIR_DIALECT_SPARSETENSOR_TRANSFORMS_CODEGENUTILS_H_

mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.cpp

static Value genIndexAndValueForDense(OpBuilder &builder, Location loc,
   return val;
 }
 
 //===----------------------------------------------------------------------===//
 // Sparse tensor loop emitter class implementations
 //===----------------------------------------------------------------------===//
 
 SparseTensorLoopEmitter::SparseTensorLoopEmitter(ValueRange tensors,
-                                                 bool isLastOutput)
-    : tensors(tensors.begin(), tensors.end()), dims(tensors.size()),
-      pidxs(tensors.size()), coord(tensors.size()), highs(tensors.size()),
-      sizes(tensors.size()), ptrBuffer(tensors.size()),
-      idxBuffer(tensors.size()), valBuffer(tensors.size()),
-      isLastOutput(isLastOutput), loopStack(), curLv(tensors.size(), 0) {
-  for (size_t i = 0, e = tensors.size(); i < e; i++) {
-    auto t = tensors[i];
-    auto rtp = t.getType().dyn_cast<RankedTensorType>();
-    if (!rtp) // a scalar (0-dimension tensors)
+                                                 bool hasOutput,
+                                                 bool isSparseOut)
+    : hasOutput(hasOutput), tensors(tensors.begin(), tensors.end()),
+      dimTypes(tensors.size()), pidxs(tensors.size()), coord(tensors.size()),
+      highs(tensors.size()), ptrBuffer(tensors.size()),
+      idxBuffer(tensors.size()), valBuffer(tensors.size()), loopStack() {
+  for (size_t tid = 0, e = tensors.size(); tid < e; tid++) {
+    auto t = tensors[tid];
+    // a scalar or 0-dimension tensors
+    if (isZeroRankedTensorOrScalar(t.getType()))
       continue;
-
+    auto rtp = t.getType().cast<RankedTensorType>();
     auto rank = static_cast<size_t>(rtp.getRank());
     auto enc = getSparseTensorEncoding(rtp);
-    if (enc)
+    // We always treat sparse output tensor as dense so that we always iterate
+    // it based on dim size.
+    if (enc && !(isOutputTensor(tid) && isSparseOut))
       for (auto dimTp : enc.getDimLevelType())
-        dims[i].push_back(dimTp);
+        dimTypes[tid].push_back(dimTp);
     else
-      dims[i].assign(rank, DimLevelType::Dense);
+      dimTypes[tid].assign(rank, DimLevelType::Dense);
 
     // Initialize using empty value.
-    pidxs[i].assign(rank, Value());
-    coord[i].assign(rank, Value());
-    highs[i].assign(rank, Value());
-    sizes[i].assign(rank, Value());
-    ptrBuffer[i].assign(rank, Value());
-    idxBuffer[i].assign(rank, Value());
+    pidxs[tid].assign(rank, Value());
+    coord[tid].assign(rank, Value());
+    highs[tid].assign(rank, Value());
+    ptrBuffer[tid].assign(rank, Value());
+    idxBuffer[tid].assign(rank, Value());
   }
 }
 
-void SparseTensorLoopEmitter::initializeLoopEmit(OpBuilder &builder,
-                                                 Location loc) {
+void SparseTensorLoopEmitter::initializeLoopEmit(
+    OpBuilder &builder, Location loc,
+    SparseTensorLoopEmitter::OutputUpdater updater) {
   // For every tensor, find lower and upper bound on dimensions, set the
   // same bounds on loop indices, and obtain dense or sparse buffer(s).
-  // TODO: Provides ability to generate loop on output buffer (with undef
-  // dim level in Merger in GenericOp Sparsification).
   for (size_t t = 0, e = tensors.size(); t < e; t++) {
     auto tensor = tensors[t];
-    auto rtp = tensor.getType().cast<RankedTensorType>();
+    auto rtp = tensor.getType().dyn_cast<RankedTensorType>();
+    if (!rtp)
+      // Skips only scalar, zero ranked tensor still need to be bufferized and
+      // (probably) filled with zeros by users.
+      continue;
     auto rank = rtp.getRank();
     auto shape = rtp.getShape();
     auto enc = getSparseTensorEncoding(rtp);
     auto dynShape = {ShapedType::kDynamicSize};
     // Scan all dimensions of current tensor.
     for (int64_t d = 0; d < rank; d++) {
       // This should be called only once at beginning.
-      assert(!ptrBuffer[t][d] && !idxBuffer[t][d] && !sizes[t][d] &&
-             !highs[t][d]);
+      assert(!ptrBuffer[t][d] && !idxBuffer[t][d] && !highs[t][d]);
       // Handle sparse storage schemes.
-      if (isCompressedDLT(dims[t][d])) {
+      if (isCompressedDLT(dimTypes[t][d])) {
         auto ptrTp =
             MemRefType::get(dynShape, getPointerOverheadType(builder, enc));
         auto indTp =
             MemRefType::get(dynShape, getIndexOverheadType(builder, enc));
         auto dim = builder.getIndexAttr(d);
         // Generate sparse primitives to obtains pointer and indices.
         ptrBuffer[t][d] = builder.create<ToPointersOp>(loc, ptrTp, tensor, dim);
         idxBuffer[t][d] = builder.create<ToIndicesOp>(loc, indTp, tensor, dim);
-      } else if (isSingletonDLT(dims[t][d])) {
+      } else if (isSingletonDLT(dimTypes[t][d])) {
         // Singleton dimension, fetch indices.
         auto indTp =
             MemRefType::get(dynShape, getIndexOverheadType(builder, enc));
         auto dim = builder.getIndexAttr(d);
         idxBuffer[t][d] = builder.create<ToIndicesOp>(loc, indTp, tensor, dim);
       } else {
         // Dense dimension, nothing to fetch.
-        assert(isDenseDLT(dims[t][d]));
+        assert(isDenseDLT(dimTypes[t][d]));
       }
 
       // Find upper bound in current dimension.
       unsigned p = toOrigDim(enc, d);
       Value up = mlir::linalg::createOrFoldDimOp(builder, loc, tensor, p);
-      sizes[t][d] = highs[t][d] = up;
+      highs[t][d] = up;
     }
+
     // Perform the required bufferization. Dense inputs materialize
-    // from the input tensors. Dense outputs need special handling.
-    // Sparse inputs use sparse primitives to obtain the values.
+    // from the input tensors. Sparse inputs use sparse primitives to obtain the
+    // values.
+    // Delegates extra output initialization to clients.
+    bool isOutput = isOutputTensor(t);
     Type elementType = rtp.getElementType();
-
     if (!enc) {
       // Non-annotated dense tensors.
       auto denseTp = MemRefType::get(shape, elementType);
-      if (isLastOutput && t == tensors.size() - 1)
-        llvm_unreachable("TODO: not yet handled");
-      else
-        valBuffer[t] =
-            builder.create<bufferization::ToMemrefOp>(loc, denseTp, tensor);
+      Value denseVal =
+      // Dense outputs need special handling.
+      if (isOutput && updater)
+        denseVal = updater(builder, loc, denseVal, tensor);
+
+      valBuffer[t] = denseVal;
     } else {
       // Annotated sparse tensors.
+      // We also need the value buffer for annotated all dense `sparse` tensor.
       auto dynShape = {ShapedType::kDynamicSize};
       auto sparseTp = MemRefType::get(dynShape, elementType);
       valBuffer[t] = builder.create<ToValuesOp>(loc, sparseTp, tensor);
     }
-    // Prepare to enter the first dim for all (input) tensors
-    prepareLoopOverTensorAtDim(builder, loc, t, 0);
+    // NOTE: we can also prepares for 0 dim here in advance, this will hosit
+    // some loop preparation from tensor iteration, but will also (undesirably)
+    // hosit the code ouside if conditions.
   }
 }
 
+void SparseTensorLoopEmitter::enterNewLoopSeq(OpBuilder &builder, Location loc,
+                                              ArrayRef<size_t> tids,
+                                              ArrayRef<size_t> dims) {
+  // Universal Index start from 0
+  assert(loopSeqStack.size() == loopStack.size());
+  // Universal index starts from 0
+  loopSeqStack.emplace_back(constantIndex(builder, loc, 0));
+  // Prepares for all the tensors used in the current loop sequence.
+  for (auto [tid, dim] : llvm::zip(tids, dims))
+    prepareLoopOverTensorAtDim(builder, loc, tid, dim);
+}
+
 Operation *SparseTensorLoopEmitter::enterLoopOverTensorAtDim(
     OpBuilder &builder, Location loc, size_t tid, size_t dim,
-    ArrayRef<Value> reduc) {
-  assert(dims[tid].size() > dim);
+    MutableArrayRef<Value> reduc, bool isParallel, ArrayRef<size_t> extraTids,
+    ArrayRef<size_t> extraDims) {
+  assert(dimTypes[tid].size() > dim);
   // We can not re-enter the same level.
   assert(!coord[tid][dim]);
+
   Value step = constantIndex(builder, loc, 1);
-  auto dimType = dims[tid][dim];
-  bool isSparse = isCompressedDLT(dimType) || isSingletonDLT(dimType);
+  auto dimType = dimTypes[tid][dim];
+  bool isSparseInput = isCompressedDLT(dimType) || isSingletonDLT(dimType);
   assert(isDenseDLT(dimType) || isCompressedDLT(dimType) ||
          isSingletonDLT(dimType));
 
-  Value lo = isSparse ? pidxs[tid][dim] : constantIndex(builder, loc, 0);
+  Value lo = isSparseInput ? pidxs[tid][dim]      // current offset
+                           : loopSeqStack.back(); // univeral tid
   Value hi = highs[tid][dim];
 
-  // TODO: support reduction.
-  if (!reduc.empty())
-    llvm_unreachable("TODO: not implemented yet");
-
   scf::ForOp forOp = builder.create<scf::ForOp>(loc, lo, hi, step, reduc);
   builder.setInsertionPointToStart(forOp.getBody());
   Value iv = forOp.getInductionVar();
-  Operation *loop = forOp;
-
   assert(iv);
-  if (isSparse) {
+  if (isSparseInput) {
     pidxs[tid][dim] = iv;
     // Generating a load on the indices array yields the coordinate.
     Value ptr = idxBuffer[tid][dim];
-    // TODO: generates load for vector value.
     coord[tid][dim] = genIndexLoad(builder, loc, ptr, iv);
   } else {
     // Dense tensor, the coordinates is the inducation variable.
     coord[tid][dim] = iv;
     // generate pidx for dense dim (pidx = i * sz + j)
-    // TODO: handle vector loop.
-    Value p = dim == 0 ? constantIndex(builder, loc, 0) : pidxs[tid][dim - 1];
-    Value mul = builder.create<arith::MulIOp>(loc, sizes[tid][dim], p);
-    Value add = builder.create<arith::AddIOp>(loc, mul, iv);
-    pidxs[tid][dim] = add;
+    auto enc = getSparseTensorEncoding(tensors[tid].getType());
+    if (enc)
+      pidxs[tid][dim] = genAddress(builder, loc, tid, dim, iv);
   }
 
-  // Prepares for next dim if this is not currently the innermost dimension.
-  if (dim != dims[tid].size() - 1)
-    prepareLoopOverTensorAtDim(builder, loc, tid, dim + 1);
+  // NOTE: we can also prepares for next dim here in advance
+  // Push the loop into stack
+  loopStack.emplace_back(ArrayRef<size_t>(tid), ArrayRef<size_t>(dim), forOp,
+                         coord[tid][dim]);
+  // Emit extra locals.
+  emitExtraLocalsForTensorsAtDenseDims(builder, loc, extraTids, extraDims);
+
+  // In-place update on the reduction variable vector.
+  assert(forOp.getNumRegionIterArgs() == reduc.size());
+  for (int i = 0, e = reduc.size(); i < e; i++)
+    reduc[i] = forOp.getRegionIterArg(i);
+  return forOp;
+}
+
+Operation *SparseTensorLoopEmitter::enterCoIterationOverTensorsAtDims(
+    OpBuilder &builder, Location loc, ArrayRef<size_t> tids,
+    ArrayRef<size_t> dims, bool needsUniv, MutableArrayRef<Value> reduc,
+    ArrayRef<size_t> extraTids, ArrayRef<size_t> extraDims) {
+  assert(tids.size() == dims.size());
+  SmallVector<Type, 4> types;
+  SmallVector<Value, 4> operands;
+  // Construct the while-loop with a parameter for each index.
+  Type indexType = builder.getIndexType();
+  for (auto [tid, dim] : llvm::zip(tids, dims)) {
+    if (isCompressedDLT(dimTypes[tid][dim]) ||
+        isSingletonDLT(dimTypes[tid][dim])) {
+      assert(pidxs[tid][dim]);
+      types.push_back(indexType);
+      operands.push_back(pidxs[tid][dim]);
+    }
+  }
+  // The position where user-supplied reduction variable starts.
+  for (Value rec : reduc) {
+    types.push_back(rec.getType());
+    operands.push_back(rec);
+  }
+  if (needsUniv) {
+    types.push_back(indexType);
+    // Update universal index.
+    operands.push_back(loopSeqStack.back());
+  }
+  assert(types.size() == operands.size());
+  scf::WhileOp whileOp = builder.create<scf::WhileOp>(loc, types, operands);
+
+  SmallVector<Location> locs(types.size(), loc);
+  Block *before = builder.createBlock(&whileOp.getBefore(), {}, types, locs);
+  Block *after = builder.createBlock(&whileOp.getAfter(), {}, types, locs);
+
+  // Build the "before" region, which effectively consists
+  // of a conjunction of "i < upper" tests on all induction.
+  builder.setInsertionPointToStart(&whileOp.getBefore().front());
+  Value cond;
+  unsigned o = 0;
+  for (auto [tid, dim] : llvm::zip(tids, dims)) {
+    if (isCompressedDLT(dimTypes[tid][dim]) ||
+        isSingletonDLT(dimTypes[tid][dim])) {
+      Value op1 = before->getArgument(o);
+      Value op2 = highs[tid][dim];
+      Value opc = builder.create<arith::CmpIOp>(loc, arith::CmpIPredicate::ult,
+                                                op1, op2);
+      cond = cond ? builder.create<arith::AndIOp>(loc, cond, opc) : opc;
+      // Update
+      pidxs[tid][dim] = after->getArgument(o++);
+    }
+  }
+  builder.create<scf::ConditionOp>(loc, cond, before->getArguments());
+
+  // Generates while body.
+  builder.setInsertionPointToStart(&whileOp.getAfter().front());
+  Value min;
+  for (auto [tid, dim] : llvm::zip(tids, dims)) {
+    // Prepares for next level.
+    if (isCompressedDLT(dimTypes[tid][dim]) ||
+        isSingletonDLT(dimTypes[tid][dim])) {
+      Value ptr = idxBuffer[tid][dim];
+      Value s = pidxs[tid][dim];
+      Value load = genIndexLoad(builder, loc, ptr, s);
+      coord[tid][dim] = load;
+      if (!needsUniv) {
+        if (min) {
+          Value cmp = builder.create<arith::CmpIOp>(
+              loc, arith::CmpIPredicate::ult, load, min);
+          min = builder.create<arith::SelectOp>(loc, cmp, load, min);
+        } else {
+          min = load;
+        }
+      }
+    }
+  }
 
-  loopStack.push_back(LoopLevelInfo({tid}, {dim}, coord[tid][dim]));
-  return loop;
+  if (needsUniv) {
+    assert(!min);
+    // Otherwise, universal index is the minimal pidx.
+    min = after->getArguments().back();
   }
 
-void SparseTensorLoopEmitter::enterCoiterationOverTensorsAtDims(
-    OpBuilder &builder, Location loc, ArrayRef<size_t> ts,
-    ArrayRef<size_t> ds) {
-  llvm_unreachable("TODO: unimplemented");
+  for (auto [tid, dim] : llvm::zip(tids, dims)) {
+    // All dense dim (as well as sparse output tensor) shared the same pidx in
+    // the while loop.
+    if (isDenseDLT(dimTypes[tid][dim])) {
+      pidxs[tid][dim] = min;
+      // generate pidx for dense dim (pidx = i * sz + j)
+      auto enc = getSparseTensorEncoding(tensors[tid].getType());
+      if (enc)
+        pidxs[tid][dim] = genAddress(builder, loc, tid, dim, min);
+    }
+    // NOTE: we can also prepares for next dim here in advance
   }
+  // Sets up the loop stack.
+  loopStack.emplace_back(tids, dims, whileOp, min);
+  assert(loopStack.size() == loopSeqStack.size());
 
-bool SparseTensorLoopEmitter::prepareLoopOverTensorAtDim(OpBuilder &builder,
+  // Emits extra locals
+  emitExtraLocalsForTensorsAtDenseDims(builder, loc, extraTids, extraDims);
+
+  // Updates reduction variables
+  assert(after->getNumArguments() == o + reduc.size() + (needsUniv ? 1 : 0));
+  // In-place update on reduction variable.
+  for (unsigned i = 0, e = reduc.size(); i < e; i++)
+    reduc[i] = after->getArgument(o + i);
+
+  return whileOp;
+}
+
+void SparseTensorLoopEmitter::prepareLoopOverTensorAtDim(OpBuilder &builder,
                                                          Location loc,
                                                          size_t tid,
                                                          size_t dim) {
-  // TODO: generate loop iteration on output tensor based on the shape
-  // instead of pointer/indices arrays.
-  assert(dims[tid].size() > dim);
-  auto dimType = dims[tid][dim];
+  assert(dimTypes[tid].size() > dim);
+  auto dimType = dimTypes[tid][dim];
 
   if (isDenseDLT(dimType))
-    return false;
+    return;
 
   // Either the first dimension, or the previous dimension has been set.
   assert(dim == 0 || pidxs[tid][dim - 1]);
   Value c0 = constantIndex(builder, loc, 0);
   Value c1 = constantIndex(builder, loc, 1);
   if (isCompressedDLT(dimType)) {
     Value ptr = ptrBuffer[tid][dim];
 
     Value pLo = dim == 0 ? c0 : pidxs[tid][dim - 1];
-    Value pHi = builder.create<arith::AddIOp>(loc, pLo, c1);
-
     pidxs[tid][dim] = genIndexLoad(builder, loc, ptr, pLo);
+
+    Value pHi = builder.create<arith::AddIOp>(loc, pLo, c1);
     highs[tid][dim] = genIndexLoad(builder, loc, ptr, pHi);
-    return true;
+    return;
   }
   if (isSingletonDLT(dimType)) {
     Value pLo = dim == 0 ? c0 : pidxs[tid][dim - 1];
     Value pHi = builder.create<arith::AddIOp>(loc, pLo, c1);
 
     pidxs[tid][dim] = pLo;
     highs[tid][dim] = pHi;
-    return true;
+    return;
   }
 
   llvm_unreachable("Unrecognizable dimesion type!");
 }
 
-Value SparseTensorLoopEmitter::emitExtraLocalsForTensorsAtDims(
-    OpBuilder &builder, Location loc, size_t tid, size_t dim) {
-  llvm_unreachable("TODO: not implemented yet");
+void SparseTensorLoopEmitter::emitExtraLocalsForTensorsAtDenseDims(
+    OpBuilder &builder, Location loc, ArrayRef<size_t> tids,
+    ArrayRef<size_t> dims) {
+  // Initialize dense positions. Note that we generate dense indices of the
+  // output tensor unconditionally, since they may not appear in the lattice,
+  // but may be needed for linearized codegen.
+  for (auto [tid, dim] : llvm::zip(tids, dims)) {
+    assert(isDenseDLT(dimTypes[tid][dim]));
+    auto enc = getSparseTensorEncoding(tensors[tid].getType());
+    if (enc) {
+      bool validPidx = dim == 0 || pidxs[tid][dim - 1];
+      if (!validPidx) {
+        // We might not find the pidx for the sparse output tensor as it is
+        // unconditionally required by the sparsification.
+        assert(isOutputTensor(tid));
+        continue;
+      }
+      pidxs[tid][dim] = genAddress(builder, loc, tid, dim, loopStack.back().iv);
+      // NOTE: we can also prepares for next dim here in advance
+    }
+  }
 }
 
-void SparseTensorLoopEmitter::exitCurrentLoop() {
-  // Clean up the values, it would help use to discover potential bug at a
-  // earlier stage (instead of silently using a wrong value).
+SmallVector<Value, 2>
+SparseTensorLoopEmitter::exitForLoop(OpBuilder &builder, Location loc,
+                                     ArrayRef<Value> reduc) {
   LoopLevelInfo &loopInfo = loopStack.back();
-  assert(loopInfo.tensors.size() == loopInfo.dims.size());
-  for (auto info : llvm::zip(loopInfo.tensors, loopInfo.dims)) {
-    auto tid = std::get<0>(info);
-    auto dim = std::get<1>(info);
-    assert(pidxs[tid][dim] && coord[tid][dim] && highs[tid][dim]);
+  auto &dims = loopStack.back().dims;
+  auto &tids = loopStack.back().tids;
+  auto forOp = llvm::cast<scf::ForOp>(loopInfo.loop);
+  if (!reduc.empty()) {
+    assert(reduc.size() == forOp.getNumResults());
+    builder.setInsertionPointToEnd(forOp.getBody());
+    builder.create<scf::YieldOp>(loc, reduc);
+  }
+
+  // Finished iterating a tensor, clean up
+  // We only do the clean up on for loop as while loops do not necessarily
+  // finish the iteration on a sparse tensor
+  for (auto [tid, dim] : llvm::zip(tids, dims)) {
     // Reset to null.
-    pidxs[tid][dim] = Value();
     coord[tid][dim] = Value();
-    if (!isDenseDLT(dims[tid][dim]))
-      // Dense dimension, high is fixed.
+    pidxs[tid][dim] = Value();
+    if (!isDenseDLT(dimTypes[tid][dim]))
       highs[tid][dim] = Value();
   }
+  // exit the loop
+  builder.setInsertionPointAfter(forOp);
+  return forOp.getResults();
+}
+
+SmallVector<Value, 2>
+SparseTensorLoopEmitter::exitCoiterationLoop(OpBuilder &builder, Location loc,
+                                             ArrayRef<Value> reduc) {
+  auto whileOp = llvm::cast<scf::WhileOp>(loopStack.back().loop);
+  auto &dims = loopStack.back().dims;
+  auto &tids = loopStack.back().tids;
+  Value iv = loopStack.back().iv;
+  // Generation while loop induction at the end.
+  builder.setInsertionPointToEnd(&whileOp.getAfter().front());
+  // Finalize the induction. Note that the induction could be performed
+  // in the individual if-branches to avoid re-evaluating the conditions.
+  // However, that would result in a rather elaborate forest of yield
+  // instructions during code generation. Moreover, performing the induction
+  // after the if-statements more closely resembles code generated by TACO.
+  unsigned o = 0;
+  SmallVector<Value, 4> operands;
+  Value one = constantIndex(builder, loc, 1);
+  for (auto [tid, dim] : llvm::zip(tids, dims)) {
+    if (isCompressedDLT(dimTypes[tid][dim]) ||
+        isSingletonDLT(dimTypes[tid][dim])) {
+      Value op1 = coord[tid][dim];
+      Value op3 = pidxs[tid][dim];
+      Value cmp =
+          builder.create<arith::CmpIOp>(loc, arith::CmpIPredicate::eq, op1, iv);
+      Value add = builder.create<arith::AddIOp>(loc, op3, one);
+      operands.push_back(builder.create<arith::SelectOp>(loc, cmp, add, op3));
+      // Following loops continue iteration from the break point of the
+      // current while loop.
+      pidxs[tid][dim] = whileOp->getResult(o++);
+      // The coordinates are invalid now.
+      coord[tid][dim] = nullptr;
+      // highs remains unchanged.
+    }
+  }
+
+  // Reduction value from users.
+  SmallVector<Value, 2> ret;
+  for (auto red : reduc) {
+    operands.push_back(red);
+    ret.push_back(whileOp->getResult(o++));
+  }
+
+  // An (optional) universal index.
+  if (operands.size() < whileOp.getNumResults()) {
+    assert(operands.size() + 1 == whileOp.getNumResults());
+    // The last one is the universial index.
+    operands.push_back(builder.create<arith::AddIOp>(loc, iv, one));
+    // update the loop starting point of current loop sequence
+    loopSeqStack.back() = whileOp->getResult(o++);
+  }
+
+  assert(o == operands.size());
+  builder.create<scf::YieldOp>(loc, operands);
+  builder.setInsertionPointAfter(whileOp);
+  return ret;
+}
+
+SmallVector<Value, 2>
+SparseTensorLoopEmitter::exitCurrentLoop(OpBuilder &builder, Location loc,
+                                         ArrayRef<Value> reduc) {
+  // Clean up the values, it would help use to discover potential bug at a
+  // earlier stage (instead of silently using a wrong value).
+  LoopLevelInfo &loopInfo = loopStack.back();
+  assert(loopInfo.tids.size() == loopInfo.dims.size());
+  SmallVector<Value, 2> red;
+  if (llvm::isa<scf::WhileOp>(loopInfo.loop)) {
+    red = exitCoiterationLoop(builder, loc, reduc);
+  } else {
+    red = exitForLoop(builder, loc, reduc);
+  }
+
+  assert(loopStack.size() == loopSeqStack.size());
   loopStack.pop_back();
+  return red;
 }
 
 //===----------------------------------------------------------------------===//
 // ExecutionEngine/SparseTensorUtils helper functions.
 //===----------------------------------------------------------------------===//
 
 OverheadType mlir::sparse_tensor::overheadTypeEncoding(unsigned width) {
   switch (width) {
void mlir::sparse_tensor::genReshapeDstShape(
   assert(reassociation.size() == srcShape.size());
   unsigned start = 0;
   // Expand the i-th dimension in srcShape.
   for (unsigned i = 0, size = srcShape.size(); i < size; i++) {
     const auto &map = reassociation[i];
     auto srcDim = srcShape[i];
     // Iterate through dimensions expanded from the i-th dimension.
     for (unsigned j = start; j < start + map.size(); j++) {
-      // There can be only one dynamic sized dimension among dimensions expanded
-      // from the i-th dimension in srcShape. For example, if srcDim = 8, then
-      // the expanded shape could be <2x?x2>, but not <2x?x?>.
+      // There can be only one dynamic sized dimension among dimensions
+      // expanded from the i-th dimension in srcShape.
+      // For example, if srcDim = 8, then the expanded shape could be <2x?x2>,
+      // but not <2x?x?>.
       if (staticDstShape[j] == ShapedType::kDynamicSize) {
         // The expanded dimension has dynamic size. We compute the dimension
         // by dividing srcDim by the product of the static dimensions.
         int64_t product = 1;
         for (unsigned k = start; k < start + map.size(); k++) {
           if (staticDstShape[k] != ShapedType::kDynamicSize) {
             product *= staticDstShape[k];
           }

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorRewriting.cpp

LogicalResult matchAndRewrite(ForeachOp op,
     Value input = op.getTensor();
     auto rtp = input.getType().cast<RankedTensorType>();
     int64_t rank = rtp.getRank();
     auto enc = getSparseTensorEncoding(rtp);
 
     // 1. Generates loop for the sparse input.
     SparseTensorLoopEmitter loopEmitter(ValueRange{input});
     loopEmitter.initializeLoopEmit(rewriter, loc);
-    for (int64_t i = 0; i < rank; i++)
+    for (int64_t i = 0; i < rank; i++) {
+      // TODO: provide utility function for loop sequences that only contains
+      // one for loop?
+      loopEmitter.enterNewLoopSeq(rewriter, loc, 0, static_cast<size_t>(i));
       loopEmitter.enterLoopOverTensorAtDim(rewriter, loc, 0, i);
+    }
 
     SmallVector<Value, 4> coords;
     coords.reserve(rank);
     loopEmitter.getCoordinateArray(coords);
 
-    Value vals = loopEmitter.getTensorValueBuffer(0);
-    Value pidx = loopEmitter.getLastLevelTensorPointerIndex(0);
+    Value vals = loopEmitter.getValBuffer()[0];
+    Value pidx = loopEmitter.getPidxs()[0].back();
     // Loads the value from sparse tensor using pointer index;
     // loads the value from dense tensor using coordinate array.
     Value val = enc ? rewriter.create<memref::LoadOp>(loc, vals, pidx)
                     : rewriter.create<memref::LoadOp>(loc, vals, coords);
 
-    for (int64_t i = 0; i < rank; i++)
-      loopEmitter.exitCurrentLoop();
-
     // 2. Inline the block in the foreach operator.
     Block::iterator inlinePos = rewriter.getInsertionPoint();
     Block *srcBlock = op.getBody();
     // Remove sparse_tensor.yield.
     rewriter.eraseOp(srcBlock->getTerminator());
 
+    for (int64_t i = 0; i < rank; i++) {
+      loopEmitter.exitCurrentLoop(rewriter, loc);
+      loopEmitter.exitCurrentLoopSeq();
+    }
+
     SmallVector<Value, 4> args;
     // Remap coordinates.
     for (int64_t i = 0; i < rank; i++) {
       Value actual = coords[toOrigDim(enc, i)];
       args.push_back(actual);
     }
     // Remap value.
     args.push_back(val);

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

 using namespace mlir::sparse_tensor;
 
 //===----------------------------------------------------------------------===//
 // Declarations of data structures.
 //===----------------------------------------------------------------------===//
 
 namespace {
 
+constexpr unsigned INVALID_ID = std::numeric_limits<unsigned>::max();
+
 // Iteration graph sorting.
 enum SortMask {
   kSparseOnly = 0x0,
   kIncludeDense = 0x1,
   kIncludeUndef = 0x2,
   kIncludeAll = 0x3
 };
 
 // Reduction kinds.
 enum Reduction { kNoReduc, kSum, kProduct, kAnd, kOr, kXor, kCustom };
 
 // Code generation.
 struct CodeGen {
-  CodeGen(SparsificationOptions o, unsigned numTensors, unsigned numLoops,
-          OpOperand *op, unsigned nest, std::vector<unsigned> &ts)
-      : options(o), loops(numLoops), sizes(numLoops), buffers(numTensors),
-        pointers(numTensors, std::vector<Value>(numLoops)),
-        indices(numTensors, std::vector<Value>(numLoops)),
-        highs(numTensors, std::vector<Value>(numLoops)),
-        pidxs(numTensors, std::vector<Value>(numLoops)),
-        idxs(numTensors, std::vector<Value>(numLoops)), sparseOut(op),
-        outerParNest(nest), topSort(ts) {}
+  CodeGen(SparsificationOptions o, ValueRange tensors, unsigned numTensors,
+          unsigned numLoops, OpOperand *op, unsigned nest,
+          std::vector<unsigned> &ts)
+      : options(o), loopEmitter(tensors, /*isLastOutput=*/true,
+                                /*isSparseOut=*/op != nullptr),
+        sparseOut(op), outerParNest(nest), topSort(ts) {}
   /// Sparsification options.
   SparsificationOptions options;
-  /// Universal dense indices and upper bounds (by index). The loops array
-  /// is updated with the value of the universal dense index in the current
-  /// loop. The sizes array is set once with the inferred dimension sizes.
-  std::vector<Value> loops;
-  std::vector<Value> sizes;
-  /// Buffers for storing dense and sparse numerical values (by tensor).
-  /// This array is set once during bufferization of all tensors.
-  std::vector<Value> buffers;
-  /// Sparse storage schemes (1-D): pointers and indices (by tensor and index).
-  /// This array is set once during bufferization of all sparse tensors.
-  std::vector<std::vector<Value>> pointers;
-  std::vector<std::vector<Value>> indices;
-  /// Sparse iteration information (by tensor and index). These arrays
-  /// are updated to remain current within the current loop.
-  std::vector<std::vector<Value>> highs;
-  std::vector<std::vector<Value>> pidxs;
-  std::vector<std::vector<Value>> idxs;
+  /// Loop emitter helper class.
+  SparseTensorLoopEmitter loopEmitter;
   /// Current reduction, updated during code generation. When indices of a
   /// reduction are exhausted, all inner loops can use a scalarized reduction.
   unsigned redExp = -1u;
   Value redVal;
   Reduction redKind = kNoReduc;
   unsigned redCustom = -1u;
   // Sparse tensor as output. Implemented either through direct injective
   // insertion in lexicographic index order or through access pattern expansion
   // in the innermost loop nest (`expValues` through `expCount`).
   OpOperand *sparseOut;
   unsigned outerParNest;
   Value expValues;
   Value expFilled;
   Value expAdded;
   Value expCount;
   // Topsort (reference should remain in scope).
   std::vector<unsigned> &topSort;
+
+  // From tensor id + loop id => dim id.
+  // TODO: This map should probably be maintained by Merger (it can be set up
+  // together with dimLvlType Map).
+  std::vector<std::vector<unsigned>> loopIdxToDim;
+
+  // Initialize the above two mapping.
+  void buildLoopIdxToDimMap(linalg::GenericOp op);
+
+  Value getLoopIdxValue(size_t loopIdx) const {
+    for (unsigned lv = 0; lv < topSort.size(); lv++)
+      if (topSort[lv] == loopIdx)
+        return loopEmitter.getLoopIV(lv);
+
+    llvm_unreachable("invalid loop index");
+  }
 };
 
+void CodeGen::buildLoopIdxToDimMap(linalg::GenericOp op) {
+  size_t numLoops = op.getNumLoops();
+  size_t numTensors = op.getNumOperands();
+  loopIdxToDim.assign(numTensors, std::vector<unsigned>(numLoops, INVALID_ID));
+
+  for (OpOperand &t : op->getOpOperands()) {
+    auto map = op.getMatchingIndexingMap(&t);
+    auto enc = getSparseTensorEncoding(t.get().getType());
+    // Scan all dimensions of current tensor.
+    unsigned tid = t.getOperandNumber();
+    for (unsigned d = 0, rank = map.getNumResults(); d < rank; d++) {
+      auto a = map.getResult(toOrigDim(enc, d)).dyn_cast<AffineDimExpr>();
+      if (a) {
+        unsigned loopId = a.getPosition();
+        // Fills the mapping.
+        loopIdxToDim[tid][loopId] = d;
+      }
+      // Else a compound affine, do nothing. (at least we are good for
+      // now, as we only support compound affine expr on non-annoated dense
+      // tensors).
+    }
+  }
+}
+
 } // namespace
 
 //===----------------------------------------------------------------------===//
 // Sparse compiler analysis methods.
 //===----------------------------------------------------------------------===//
 
 /// Helper method to construct a permuted dimension ordering
 /// that adheres to the given topological sort.
for (unsigned i = 0; i < sz; i++)
     perm[i] = inv[topSort[i]];
   return AffineMap::getPermutationMap(perm, context);
 }
 
 /// Helper method to inspect affine expressions. Rejects cases where the
 /// same index is used more than once. Also rejects compound affine
 /// expressions in sparse dimensions.
 static bool findAffine(Merger &merger, unsigned tensor, AffineExpr a,
-                       DimLevelType dim) {
+                       DimLevelType dim, bool setLvlFormat = true) {
   switch (a.getKind()) {
   case AffineExprKind::DimId: {
     unsigned idx = a.cast<AffineDimExpr>().getPosition();
     if (!isUndefDLT(merger.getDimLevelType(tensor, idx)))
       return false; // used more than once
+
+    if (setLvlFormat)
       merger.setDimLevelType(tensor, idx, dim);
     return true;
   }
   case AffineExprKind::Add:
   case AffineExprKind::Mul: {
     if (!isDenseDLT(dim))
       return false; // compound only in dense dim
     auto binOp = a.cast<AffineBinaryOpExpr>();
-    return findAffine(merger, tensor, binOp.getLHS(), dim) &&
-           findAffine(merger, tensor, binOp.getRHS(), dim);
+    // We do not set dim level format for affine expresssion like d0 + d1 on
+    // both loop index at d0 and d1,
+    return findAffine(merger, tensor, binOp.getLHS(), dim, false) &&
+           findAffine(merger, tensor, binOp.getRHS(), dim, false);
   }
   case AffineExprKind::Constant:
     return isDenseDLT(dim); // const only in dense dim
   default:
     return false;
   }
 }
 

 static Value getCustomRedId(Operation *op) {
   return dyn_cast<sparse_tensor::ReduceOp>(op).getIdentity();
 }
 
 //===----------------------------------------------------------------------===//
 // Sparse compiler synthesis methods (statements and expressions).
 //===----------------------------------------------------------------------===//
 
-/// Generates buffer for the output tensor. Note that all sparse kernels
-/// assume that when all elements are written to (viz. x(i) = y(i) * z(i)),
-/// the output buffer is already initialized to all zeroes and only nonzeroes
-/// values are computed and written out. For updates (viz. x(i) += y(i) * z(i)),
-/// only nonzeroes values are used for the updates and no assumption on the
-/// original contents of the output buffer is necessary.
-static Value genOutputBuffer(CodeGen &codegen, OpBuilder &builder,
-                             linalg::GenericOp op, MemRefType denseTp,
-                             ArrayRef<Value> args) {
+/// Local bufferization of all dense and sparse data structures.
+static void genBuffers(Merger &merger, CodeGen &codegen, OpBuilder &builder,
+                       linalg::GenericOp op) {
   Location loc = op.getLoc();
+  assert(op.getNumOperands() == op.getNumInputs() + 1);
+
+  codegen.loopEmitter.initializeLoopEmit(
+      builder, loc,
+      /// Generates buffer for the output tensor.
+      /// Note that all sparse kernels assume that when all elements are written
+      /// to (viz. x(i) = y(i) * z(i)), the output buffer is already initialized
+      /// to all zeroes and only nonzeroes values are computed and written out.
+      /// For updates (viz. x(i) += y(i) * z(i)), only nonzeroes values are used
+      /// for the updates and no assumption on the original contents of the
+      /// output buffer is necessary.
+      [&op](OpBuilder &builder, Location loc, Value memref,
+            Value tensor) -> Value {
+        // Must not be a sparse tensor.
+        assert(!getSparseTensorEncoding(tensor.getType()));
         OpOperand *lhs = op.getOutputOperand(0);
-  Value tensor = lhs->get();
+        // Two output tensors references should pointed to the same object.
+        assert(lhs->get() == tensor);
         bool isInit = op.isInitTensor(lhs);
-  // An output tensor can simply materialize from the buffer of the tensor that
-  // appears in the outs() clause. For updates, this has the advantage that only
-  // the nonzero value are involved in the computation, keeping the operation
-  // O(nnz). In all other cases, we are forced to zero out the buffer to enforce
-  // the assumption above, which may negatively impact running complexity
-  // (viz. O(n^2 + nnz) vs. O(nnz) for matrices).
+        // An output tensor can simply materialize from the buffer of the tensor
+        // that appears in the outs() clause. For updates, this has the
+        // advantage that only the nonzero value are involved in the
+        // computation, keeping the operation O(nnz). In all other cases, we are
+        // forced to zero out the buffer to enforce the assumption above, which
+        // may negatively impact running complexity (viz. O(n^2 + nnz) vs.
+        // O(nnz) for matrices).
         // TODO: use better analysis to avoid zeroing out the buffer?
-  Value init = builder.create<bufferization::ToMemrefOp>(loc, denseTp, tensor);
+        Value init = memref;
         if (!isInit) {
-    Value zero = constantZero(builder, loc, denseTp.getElementType());
-    builder.create<linalg::FillOp>(loc, ValueRange{zero}, ValueRange{init});
+          Value zero = constantZero(builder, loc,
+                                    getElementTypeOrSelf(tensor.getType()));
+          builder.create<linalg::FillOp>(loc, ValueRange{zero},
+                                         ValueRange{init});
         }
         return init;
-}
-
-/// Local bufferization of all dense and sparse data structures.
-static void genBuffers(Merger &merger, CodeGen &codegen, OpBuilder &builder,
-                       linalg::GenericOp op) {
-  Location loc = op.getLoc();
-  assert(op->getNumOperands() == op.getNumInputs() + 1);
-  // For every tensor, find lower and upper bound on dimensions, set the
-  // same bounds on loop indices, and obtain dense or sparse buffer(s).
-  auto dynShape = {ShapedType::kDynamicSize};
-  SmallVector<Value, 4> args;
-  for (OpOperand &t : op->getOpOperands()) {
-    unsigned tensor = t.getOperandNumber();
-    auto shape = op.getShape(&t);
-    auto map = op.getMatchingIndexingMap(&t);
-    auto enc = getSparseTensorEncoding(t.get().getType());
-    // Scan all dimensions of current tensor.
-    args.clear();
-    for (unsigned d = 0, rank = map.getNumResults(); d < rank; d++) {
-      AffineExpr a = map.getResult(toOrigDim(enc, d));
-      if (a.getKind() != AffineExprKind::DimId)
-        continue; // compound
-      unsigned idx = a.cast<AffineDimExpr>().getPosition();
-      // Handle the different storage schemes.
-      if (isCompressedDLT(merger.getDimLevelType(tensor, idx))) {
-        // Compressed dimension, fetch pointer and indices.
-        auto ptrTp =
-            MemRefType::get(dynShape, getPointerOverheadType(builder, enc));
-        auto indTp =
-            MemRefType::get(dynShape, getIndexOverheadType(builder, enc));
-        auto dim = builder.getIndexAttr(d);
-        codegen.pointers[tensor][idx] =
-            builder.create<ToPointersOp>(loc, ptrTp, t.get(), dim);
-        codegen.indices[tensor][idx] =
-            builder.create<ToIndicesOp>(loc, indTp, t.get(), dim);
-      } else if (isSingletonDLT(merger.getDimLevelType(tensor, idx))) {
-        // Singleton dimension, fetch indices.
-        auto indTp =
-            MemRefType::get(dynShape, getIndexOverheadType(builder, enc));
-        auto dim = builder.getIndexAttr(d);
-        codegen.indices[tensor][idx] =
-            builder.create<ToIndicesOp>(loc, indTp, t.get(), dim);
-      } else {
-        // Dense dimension, nothing to fetch.
-        assert(isDenseDLT(merger.getDimLevelType(tensor, idx)));
-      }
-      // Find upper bound in current dimension.
-      unsigned p = toOrigDim(enc, d);
-      Value up = linalg::createOrFoldDimOp(builder, loc, t.get(), p);
-      if (ShapedType::isDynamic(shape[p]))
-        args.push_back(up);
-      assert(codegen.highs[tensor][idx] == nullptr);
-      codegen.sizes[idx] = codegen.highs[tensor][idx] = up;
-    }
-    // Perform the required bufferization. Dense inputs materialize
-    // from the input tensors. Dense outputs need special handling.
-    // Sparse inputs use sparse primitives to obtain the values.
-    Type elementType = getElementTypeOrSelf(t.get().getType());
-    if (!enc) {
-      // Non-annotated dense tensors.
-      auto denseTp = MemRefType::get(shape, elementType);
-      if (tensor < op.getNumInputs())
-        codegen.buffers[tensor] =
-            builder.create<bufferization::ToMemrefOp>(loc, denseTp, t.get());
-      else
-        codegen.buffers[tensor] =
-            genOutputBuffer(codegen, builder, op, denseTp, args);
-    } else if (&t != codegen.sparseOut) {
-      // Annotated sparse tensors (not involved in output).
-      auto sparseTp = MemRefType::get(dynShape, elementType);
-      codegen.buffers[tensor] =
-          builder.create<ToValuesOp>(loc, sparseTp, t.get());
-    }
-  }
+      });
 }
 
 /// Generates an affine expression.
 //
 // TODO: generalize for sparse tensor subscripts
 //
 static Value genAffine(CodeGen &codegen, OpBuilder &builder, AffineExpr a,
                        Location loc) {
   switch (a.getKind()) {
   case AffineExprKind::DimId: {
     unsigned idx = a.cast<AffineDimExpr>().getPosition();
-    return codegen.loops[idx]; // universal dense index
+    return codegen.getLoopIdxValue(idx); // universal dense index
   }
   case AffineExprKind::Add: {
     auto binOp = a.cast<AffineBinaryOpExpr>();
     return builder.create<arith::AddIOp>(
         loc, genAffine(codegen, builder, binOp.getLHS(), loc),
         genAffine(codegen, builder, binOp.getRHS(), loc));
   }
   case AffineExprKind::Mul: {

 
 /// Generates index for load/store on sparse tensor.
 static Value genIndex(CodeGen &codegen, linalg::GenericOp op, OpOperand *t) {
   auto map = op.getMatchingIndexingMap(t);
   auto enc = getSparseTensorEncoding(t->get().getType());
   AffineExpr a = map.getResult(toOrigDim(enc, map.getNumResults() - 1));
   assert(a.getKind() == AffineExprKind::DimId);
   unsigned idx = a.cast<AffineDimExpr>().getPosition();
-  return codegen.loops[idx];
+  return codegen.getLoopIdxValue(idx);
 }
 
 /// Generates subscript for load/store on a dense or sparse tensor.
 static Value genSubscript(CodeGen &codegen, OpBuilder &builder,
                           linalg::GenericOp op, OpOperand *t,
                           SmallVector<Value, 4> &args) {
   unsigned tensor = t->getOperandNumber();
   auto map = op.getMatchingIndexingMap(t);
   auto enc = getSparseTensorEncoding(t->get().getType());
   unsigned rank = map.getNumResults();
   if (enc) {
     // Note that currently, all sparse subscripts are simple.
     // TODO: accept affine too?
-    AffineExpr a = map.getResult(toOrigDim(enc, rank - 1));
-    assert(a.getKind() == AffineExprKind::DimId);
-    unsigned idx = a.cast<AffineDimExpr>().getPosition();
-    assert(codegen.pidxs[tensor][idx] != nullptr);
-    args.push_back(codegen.pidxs[tensor][idx]); // position index
+    assert(map.getResult(toOrigDim(enc, rank - 1)).getKind() ==
+           AffineExprKind::DimId);
+    Value pidx = codegen.loopEmitter.getPidxs()[tensor].back();
+    assert(pidx);
+    args.push_back(pidx); // position index
   } else {
     for (unsigned d = 0; d < rank; d++) {
       AffineExpr a = map.getResult(d);
       args.push_back(genAffine(codegen, builder, a, op.getLoc()));
     }
   }
-  return codegen.buffers[tensor];
+  return codegen.loopEmitter.getValBuffer()[tensor];
 }
 
 /// Generates insertion code to implement dynamic tensor load.
 static Value genInsertionLoad(CodeGen &codegen, OpBuilder &builder,
                               linalg::GenericOp op, OpOperand *t) {
   Location loc = op.getLoc();
   // Direct lexicographic index order, tensor loads as zero.
   if (!codegen.expValues) {

 static void genInsertionStore(CodeGen &codegen, OpBuilder &builder,
                               linalg::GenericOp op, OpOperand *t, Value rhs) {
   Location loc = op.getLoc();
   // Direct insertion in lexicographic index order.
   if (!codegen.expValues) {
     unsigned rank = op.getRank(t);
     SmallVector<Value, 4> indices;
     for (unsigned i = 0; i < rank; i++) {
-      assert(codegen.loops[codegen.topSort[i]]);
-      indices.push_back(codegen.loops[codegen.topSort[i]]);
+      assert(codegen.loopEmitter.getLoopIV(i));
+      indices.push_back(codegen.loopEmitter.getLoopIV(i));
     }
     builder.create<InsertOp>(loc, rhs, t->get(), indices);
     return;
   }
   // Generates insertion code along expanded access pattern.
   //   if (!expFilled[i]) then
   //     expFilled[i] = true
   //     expAdded[inserts++] = i
if (t == codegen.sparseOut) {
     return;
   }
   // Actual store.
   SmallVector<Value, 4> args;
   Value ptr = genSubscript(codegen, builder, op, t, args);
   builder.create<memref::StoreOp>(loc, rhs, ptr, args);
 }
 
-/// Generates a pointer/index load from the sparse storage scheme. Narrower
-/// data types need to be zero extended before casting the value into the
-/// index type used for looping and indexing.
-static Value genLoad(CodeGen &codegen, OpBuilder &builder, Location loc,
-                     Value ptr, Value s) {
-  // Simply zero extends narrower indices into 64-bit values before casting to
-  // index without a performance penalty.
-  Value load = builder.create<memref::LoadOp>(loc, ptr, s);
-  if (!load.getType().isa<IndexType>()) {
-    if (load.getType().getIntOrFloatBitWidth() < 64)
-      load = builder.create<arith::ExtUIOp>(loc, builder.getI64Type(), load);
-    load =
-        builder.create<arith::IndexCastOp>(loc, builder.getIndexType(), load);
-  }
-  return load;
-}
-
 /// Generates an invariant value.
-static Value genInvariantValue(Merger &merger, CodeGen &codegen,
+inline static Value genInvariantValue(Merger &merger, CodeGen &codegen,
                                       OpBuilder &builder, unsigned exp) {
-  Value val = merger.exp(exp).val;
-  return val;
-}
-
-/// Generates an address computation "sz * p + i".
-static Value genAddress(CodeGen &codegen, OpBuilder &builder, Location loc,
-                        Value size, Value p, Value i) {
-  Value mul = builder.create<arith::MulIOp>(loc, size, p);
-  return builder.create<arith::AddIOp>(loc, mul, i);
+  return merger.exp(exp).val;
 }
 
 /// Generates an index value.
-static Value genIndexValue(CodeGen &codegen, OpBuilder &builder, unsigned idx,
-                           unsigned ldx) {
-  Value ival = codegen.loops[idx];
-  return ival;
+inline static Value genIndexValue(CodeGen &codegen, OpBuilder &builder,
+                                  unsigned idx) {
+  return codegen.getLoopIdxValue(idx);
 }
 
 /// Semi-ring branches are simply inlined by the sparse compiler. Prior
 /// analysis has verified that all computations are "local" to the inlined
 /// branch or otherwise invariantly defined outside the loop nest, with the
 /// exception of index computations, which need to be relinked to actual
 /// inlined cloned code.
 static Value relinkBranch(CodeGen &codegen, RewriterBase &rewriter,
                           Block *block, Value e, unsigned ldx) {
   if (Operation *def = e.getDefiningOp()) {
     if (auto indexOp = dyn_cast<linalg::IndexOp>(def))
-      return genIndexValue(codegen, rewriter, indexOp.getDim(), ldx);
+      return genIndexValue(codegen, rewriter, indexOp.getDim());
     if (def->getBlock() == block) {
       for (unsigned i = 0, n = def->getNumOperands(); i < n; i++)
         def->setOperand(
             i, relinkBranch(codegen, rewriter, block, def->getOperand(i), ldx));
     }
   }
   return e;
 }
 
 /// Recursively generates tensor expression.
 static Value genExp(Merger &merger, CodeGen &codegen, RewriterBase &rewriter,
                     linalg::GenericOp op, unsigned exp, unsigned ldx) {
   Location loc = op.getLoc();
   if (exp == -1u)
     return Value();
   if (merger.exp(exp).kind == Kind::kTensor)
     return genTensorLoad(merger, codegen, rewriter, op, exp);
   if (merger.exp(exp).kind == Kind::kInvariant)
     return genInvariantValue(merger, codegen, rewriter, exp);
   if (merger.exp(exp).kind == Kind::kIndex)
-    return genIndexValue(codegen, rewriter, merger.exp(exp).index, ldx);
+    return genIndexValue(codegen, rewriter, merger.exp(exp).index);
 
   if (merger.exp(exp).kind == Kind::kReduce) {
     // Make custom reduction identity accessible for expanded access pattern.
     assert(codegen.redCustom == -1u);
     codegen.redCustom = exp;
   }
 
   Value v0 =

 /// Determines if affine expression is invariant.
 static bool isInvariantAffine(const CodeGen &codegen, AffineExpr a,
                               unsigned ldx, bool &atLevel) {
   switch (a.getKind()) {
   case AffineExprKind::DimId: {
     unsigned idx = a.cast<AffineDimExpr>().getPosition();
     if (idx == ldx)
       atLevel = true;
-    return codegen.loops[idx] != nullptr; // no longer in play?
+    return codegen.getLoopIdxValue(idx) != nullptr; // no longer in play?
   }
   case AffineExprKind::Add:
   case AffineExprKind::Mul: {
     auto binOp = a.cast<AffineBinaryOpExpr>();
     return isInvariantAffine(codegen, binOp.getLHS(), ldx, atLevel) &&
            isInvariantAffine(codegen, binOp.getRHS(), ldx, atLevel);
   }
   default:

 
 /// Generates an expanded access pattern in innermost dimension.
 static void genExpansion(Merger &merger, CodeGen &codegen, OpBuilder &builder,
                          linalg::GenericOp op, unsigned at, bool atStart) {
   OpOperand *lhs = codegen.sparseOut;
   if (!lhs || codegen.outerParNest != op.getRank(lhs) - 1 ||
       at != codegen.outerParNest)
     return; // not needed at this level
+  assert(codegen.redVal == nullptr);
   // Generate start or end of an expanded access pattern.
   Value tensor = lhs->get();
   Location loc = op.getLoc();
   if (atStart) {
     auto dynShape = {ShapedType::kDynamicSize};
     Type etp = tensor.getType().cast<ShapedType>().getElementType();
     Type t1 = MemRefType::get(dynShape, etp);
     Type t2 = MemRefType::get(dynShape, builder.getI1Type());
     Type t3 = MemRefType::get(dynShape, builder.getIndexType());
     Type t4 = builder.getIndexType();
     auto res =
         builder.create<ExpandOp>(loc, TypeRange({t1, t2, t3, t4}), tensor);
     assert(res.getNumResults() == 4);
     assert(!codegen.expValues);
     codegen.expValues = res.getResult(0);
     codegen.expFilled = res.getResult(1);
     codegen.expAdded = res.getResult(2);
     codegen.expCount = res.getResult(3);
   } else {
     assert(codegen.expValues);
     SmallVector<Value, 4> indices;
     for (unsigned i = 0; i < at; i++) {
-      assert(codegen.loops[codegen.topSort[i]]);
-      indices.push_back(codegen.loops[codegen.topSort[i]]);
+      assert(codegen.loopEmitter.getLoopIV(i));
+      indices.push_back(codegen.loopEmitter.getLoopIV(i));
     }
     builder.create<CompressOp>(loc, codegen.expValues, codegen.expFilled,
                                codegen.expAdded, codegen.expCount, tensor,
                                indices);
     codegen.expValues = codegen.expFilled = codegen.expAdded =
         codegen.expCount = Value();
   }
 }
 
-/// Generates initialization code for the subsequent loop sequence at
-/// current index level. Returns true if the loop sequence needs to
-/// maintain the universal index.
-static bool genInit(Merger &merger, CodeGen &codegen, OpBuilder &builder,
-                    linalg::GenericOp op, unsigned at, BitVector &inits) {
-  std::vector<unsigned> &topSort(codegen.topSort);
-  bool needsUniv = false;
-  Location loc = op.getLoc();
-  unsigned idx = topSort[at];
-
-  // Initialize sparse positions.
-  for (unsigned b = 0, be = inits.size(); b < be; b++) {
-    if (!inits[b])
-      continue;
-    unsigned tensor = merger.tensor(b);
-    assert(idx == merger.index(b));
-    if (isCompressedDLT(merger.getDimLevelType(b))) {
-      // Initialize sparse index that will implement the iteration:
-      //   for pidx_idx = pointers(pidx_idx-1), pointers(1+pidx_idx-1)
-      unsigned pat = at;
-      for (; pat != 0; pat--) {
-        if (codegen.pidxs[tensor][topSort[pat - 1]])
-          break;
-      }
-      Value ptr = codegen.pointers[tensor][idx];
-      Value one = constantIndex(builder, loc, 1);
-      Value p0 = (pat == 0) ? constantIndex(builder, loc, 0)
-                            : codegen.pidxs[tensor][topSort[pat - 1]];
-      codegen.pidxs[tensor][idx] = genLoad(codegen, builder, loc, ptr, p0);
-      Value p1 = builder.create<arith::AddIOp>(loc, p0, one);
-      codegen.highs[tensor][idx] = genLoad(codegen, builder, loc, ptr, p1);
-    } else if (isSingletonDLT(merger.getDimLevelType(b))) {
-      // Initialize sparse index that will implement the "iteration":
-      //   for pidx_idx = pidx_idx-1, 1+pidx_idx-1
-      // We rely on subsequent loop unrolling to get rid of the loop
-      // if it is not involved in co-iteration with anything else.
-      unsigned pat = at;
-      for (; pat != 0; pat--) {
-        if (codegen.pidxs[tensor][topSort[pat - 1]])
-          break;
-      }
-      Value one = constantIndex(builder, loc, 1);
-      Value p0 = (pat == 0) ? constantIndex(builder, loc, 0)
-                            : codegen.pidxs[tensor][topSort[pat - 1]];
-      codegen.pidxs[tensor][idx] = p0;
-      codegen.highs[tensor][idx] = builder.create<arith::AddIOp>(loc, p0, one);
-    } else {
-      assert(isDenseDLT(merger.getDimLevelType(b)) ||
-             isUndefDLT(merger.getDimLevelType(b)));
-      // Dense index still in play.
-      needsUniv = true;
-    }
-  }
-
-  // Initialize the universal dense index.
-  codegen.loops[idx] = constantIndex(builder, loc, 0);
-  return needsUniv;
-}
-
-/// Returns parallelization strategy. Any implicit loop in the Linalg operation
-/// that is marked "parallel" is a candidate. Whether it is actually converted
-/// to a parallel operation depends on the requested strategy.
+/// Returns parallelization strategy. Any implicit loop in the Linalg
+/// operation that is marked "parallel" is a candidate. Whether it is actually
+/// converted to a parallel operation depends on the requested strategy.
 static bool isParallelFor(CodeGen &codegen, bool isOuter, bool isReduction,
                           bool isSparse) {
   // Reject parallelization of sparse output.
   if (codegen.sparseOut)
     return false;
   // Inspect strategy.
   switch (codegen.options.parallelizationStrategy) {
   case SparseParallelizationStrategy::kNone:
     return false;
   case SparseParallelizationStrategy::kDenseOuterLoop:
     return isOuter && !isSparse && !isReduction;
   case SparseParallelizationStrategy::kAnyStorageOuterLoop:
     return isOuter && !isReduction;
   case SparseParallelizationStrategy::kDenseAnyLoop:
     return !isSparse && !isReduction;
   case SparseParallelizationStrategy::kAnyStorageAnyLoop:
     return !isReduction;
   }
   llvm_unreachable("unexpected parallelization strategy");
 }
 
 /// Generates a for-loop on a single index.
 static Operation *genFor(Merger &merger, CodeGen &codegen, OpBuilder &builder,
                          linalg::GenericOp op, bool isOuter, bool isInner,
-                         unsigned idx, BitVector &indices) {
-  unsigned fb = indices.find_first();
-  unsigned tensor = merger.tensor(fb);
-  assert(idx == merger.index(fb));
+                         unsigned idx, size_t tid, size_t dim,
+                         ArrayRef<size_t> extraTids,
+                         ArrayRef<size_t> extraDims) {
+  Location loc = op.getLoc();
   auto iteratorTypes = op.getIteratorTypesArray();
   bool isReduction = linalg::isReductionIterator(iteratorTypes[idx]);
-  bool isSparse = isCompressedDLT(merger.getDimLevelType(fb)) ||
-                  isSingletonDLT(merger.getDimLevelType(fb));
+  bool isSparse = isCompressedDLT(merger.getDimLevelType(tid, idx)) ||
+                  isSingletonDLT(merger.getDimLevelType(tid, idx));
   bool isParallel = isParallelFor(codegen, isOuter, isReduction, isSparse);
-
-  // Loop bounds and increment.
-  Location loc = op.getLoc();
-  Value lo = isSparse ? codegen.pidxs[tensor][idx] : codegen.loops[idx];
-  Value hi = isSparse ? codegen.highs[tensor][idx] : codegen.sizes[idx];
-  Value step = constantIndex(builder, loc, 1);
-
-  // Emit a parallel loop.
-  if (isParallel) {
-    scf::ParallelOp parOp = builder.create<scf::ParallelOp>(loc, lo, hi, step);
-    if (isSparse)
-      codegen.pidxs[tensor][idx] = parOp.getInductionVars()[0];
-    else
-      codegen.loops[idx] = parOp.getInductionVars()[0];
-    builder.setInsertionPointToStart(parOp.getBody());
-    return parOp;
-  }
+  assert(!isParallel);
 
   // Emit a sequential or vector loop.
   SmallVector<Value, 4> operands;
   if (codegen.redVal)
     operands.push_back(codegen.redVal);
   if (codegen.expValues)
     operands.push_back(codegen.expCount);
 
-  scf::ForOp forOp = builder.create<scf::ForOp>(loc, lo, hi, step, operands);
+  Operation *loop = codegen.loopEmitter.enterLoopOverTensorAtDim(
+      builder, loc, tid, dim, operands, isParallel, extraTids, extraDims);
 
+  // The operands should be updated by loop emitter already.
   if (codegen.redVal)
-    updateReduc(merger, codegen, forOp.getRegionIterArgs().front());
+    updateReduc(merger, codegen, operands.front());
   if (codegen.expValues)
-    codegen.expCount = forOp.getRegionIterArgs().back();
-  // Assign induction variable to sparse or dense index.
-  Value iv = forOp.getInductionVar();
-  if (isSparse)
-    codegen.pidxs[tensor][idx] = iv;
-  else
-    codegen.loops[idx] = iv;
+    codegen.expCount = operands.back();
 
-  builder.setInsertionPointToStart(forOp.getBody());
-  return forOp;
+  return loop;
 }
 
 /// Emit a while-loop for co-iteration over multiple indices.
 static Operation *genWhile(Merger &merger, CodeGen &codegen, OpBuilder &builder,
                            linalg::GenericOp op, unsigned idx, bool needsUniv,
-                           BitVector &indices) {
-  SmallVector<Type, 4> types;
+                           ArrayRef<size_t> condTids, ArrayRef<size_t> condDims,
+                           ArrayRef<size_t> extraTids,
+                           ArrayRef<size_t> extraDims) {
+
   SmallVector<Value, 4> operands;
+
   // Construct the while-loop with a parameter for each index.
-  Type indexType = builder.getIndexType();
-  for (unsigned b = 0, be = indices.size(); b < be; b++) {
-    if (!indices[b])
-      continue;
-    if (isCompressedDLT(merger.getDimLevelType(b)) ||
-        isSingletonDLT(merger.getDimLevelType(b))) {
-      unsigned tensor = merger.tensor(b);
-      assert(idx == merger.index(b));
-      types.push_back(indexType);
-      operands.push_back(codegen.pidxs[tensor][idx]);
-    } else {
-      assert(isDenseDLT(merger.getDimLevelType(b)) ||
-             isUndefDLT(merger.getDimLevelType(b)));
-    }
-  }
-  if (codegen.redVal) {
-    types.push_back(codegen.redVal.getType());
+  if (codegen.redVal)
     operands.push_back(codegen.redVal);
-  }
-  if (codegen.expValues) {
-    types.push_back(indexType);
+  if (codegen.expValues)
     operands.push_back(codegen.expCount);
-  }
-  if (needsUniv) {
-    types.push_back(indexType);
-    operands.push_back(codegen.loops[idx]);
-  }
-  assert(types.size() == operands.size());
-  Location loc = op.getLoc();
-  scf::WhileOp whileOp = builder.create<scf::WhileOp>(loc, types, operands);
 
-  SmallVector<Location> locs(types.size(), loc);
-  Block *before = builder.createBlock(&whileOp.getBefore(), {}, types, locs);
-  Block *after = builder.createBlock(&whileOp.getAfter(), {}, types, locs);
-
-  // Build the "before" region, which effectively consists
-  // of a conjunction of "i < upper" tests on all induction.
-  builder.setInsertionPointToStart(&whileOp.getBefore().front());
-  Value cond;
-  unsigned o = 0;
-  for (unsigned b = 0, be = indices.size(); b < be; b++) {
-    if (!indices[b])
-      continue;
-    if (isCompressedDLT(merger.getDimLevelType(b)) ||
-        isSingletonDLT(merger.getDimLevelType(b))) {
-      unsigned tensor = merger.tensor(b);
-      assert(idx == merger.index(b));
-      Value op1 = before->getArgument(o);
-      Value op2 = codegen.highs[tensor][idx];
-      Value opc = builder.create<arith::CmpIOp>(loc, arith::CmpIPredicate::ult,
-                                                op1, op2);
-      cond = cond ? builder.create<arith::AndIOp>(loc, cond, opc) : opc;
-      codegen.pidxs[tensor][idx] = after->getArgument(o++);
-    } else {
-      assert(isDenseDLT(merger.getDimLevelType(b)) ||
-             isUndefDLT(merger.getDimLevelType(b)));
-    }
-  }
+  Operation *loop = codegen.loopEmitter.enterCoIterationOverTensorsAtDims(
+      builder, op.getLoc(), condTids, condDims, needsUniv, operands, extraTids,
+      extraDims);
+
   if (codegen.redVal)
-    updateReduc(merger, codegen, after->getArgument(o++));
+    updateReduc(merger, codegen, operands.front());
   if (codegen.expValues)
-    codegen.expCount = after->getArgument(o++);
-  if (needsUniv)
-    codegen.loops[idx] = after->getArgument(o++);
-  assert(o == operands.size());
-  builder.create<scf::ConditionOp>(loc, cond, before->getArguments());
-  builder.setInsertionPointToStart(&whileOp.getAfter().front());
-  return whileOp;
+    codegen.expCount = operands.back();
+
+  return loop;
 }
 
 /// Generates a for-loop or a while-loop, depending on whether it implements
 /// singleton iteration or co-iteration over the given conjunction.
 static Operation *genLoop(Merger &merger, CodeGen &codegen, OpBuilder &builder,
                           linalg::GenericOp op, unsigned at, bool needsUniv,
-                          BitVector &indices) {
+                          ArrayRef<size_t> condTids, ArrayRef<size_t> condDims,
+                          ArrayRef<size_t> extraTids,
+                          ArrayRef<size_t> extraDims) {
+  assert(condTids.size() == condDims.size());
+  assert(extraTids.size() == extraDims.size());
   unsigned idx = codegen.topSort[at];
-  if (indices.count() == 1) {
+  if (condTids.size() == 1) {
     bool isOuter = at == 0;
     bool isInner = at == codegen.topSort.size() - 1;
-    return genFor(merger, codegen, builder, op, isOuter, isInner, idx, indices);
-  }
-  return genWhile(merger, codegen, builder, op, idx, needsUniv, indices);
-}
-
-/// Generates the local variables for this loop, consisting of the sparse
-/// indices, restored universal dense index, and dense positions.
-static void genLocals(Merger &merger, CodeGen &codegen, OpBuilder &builder,
-                      linalg::GenericOp op, unsigned at, bool needsUniv,
-                      BitVector &locals) {
-  std::vector<unsigned> &topSort(codegen.topSort);
-  Location loc = op.getLoc();
-  unsigned idx = topSort[at];
-
-  // Initialize sparse indices.
-  Value min;
-  for (unsigned b = 0, be = locals.size(); b < be; b++) {
-    if (!locals[b])
-      continue;
-    if (isCompressedDLT(merger.getDimLevelType(b)) ||
-        isSingletonDLT(merger.getDimLevelType(b))) {
-      unsigned tensor = merger.tensor(b);
-      assert(idx == merger.index(b));
-      Value ptr = codegen.indices[tensor][idx];
-      Value s = codegen.pidxs[tensor][idx];
-      Value load = genLoad(codegen, builder, loc, ptr, s);
-      codegen.idxs[tensor][idx] = load;
-      if (!needsUniv) {
-        if (min) {
-          Value cmp = builder.create<arith::CmpIOp>(
-              loc, arith::CmpIPredicate::ult, load, min);
-          min = builder.create<arith::SelectOp>(loc, cmp, load, min);
-        } else {
-          min = load;
-        }
-      }
-    } else {
-      assert(isDenseDLT(merger.getDimLevelType(b)) ||
-             isUndefDLT(merger.getDimLevelType(b)));
-    }
-  }
-
-  // Merge dense universal index over minimum.
-  if (min) {
-    assert(!needsUniv);
-    codegen.loops[idx] = min;
-  }
-
-  // Initialize dense positions. Note that we generate dense indices of the
-  // output tensor unconditionally, since they may not appear in the lattice,
-  // but may be needed for linearized codegen.
-  for (unsigned b = 0, be = locals.size(); b < be; b++) {
-    if ((locals[b] || merger.isOutTensor(b, idx)) &&
-        isDenseDLT(merger.getDimLevelType(b))) {
-      unsigned tensor = merger.tensor(b);
-      assert(idx == merger.index(b));
-      unsigned pat = at;
-      for (; pat != 0; pat--)
-        if (codegen.pidxs[tensor][topSort[pat - 1]])
-          break;
-      Value p = (pat == 0) ? constantIndex(builder, loc, 0)
-                           : codegen.pidxs[tensor][topSort[pat - 1]];
-      codegen.pidxs[tensor][idx] = genAddress(
-          codegen, builder, loc, codegen.sizes[idx], p, codegen.loops[idx]);
-    }
+    return genFor(merger, codegen, builder, op, isOuter, isInner, idx,
+                  condTids.front(), condDims.front(), extraTids, extraDims);
   }
+  return genWhile(merger, codegen, builder, op, idx, needsUniv, condTids,
+                  condDims, extraTids, extraDims);
 }
 
 /// Generates the induction structure for a while-loop.
-static void genWhileInduction(Merger &merger, CodeGen &codegen,
-                              OpBuilder &builder, linalg::GenericOp op,
-                              unsigned idx, bool needsUniv,
-                              BitVector &induction, scf::WhileOp whileOp) {
+static void finalizeWhileOp(Merger &merger, CodeGen &codegen,
+                            unsigned idx, bool needsUniv, BitVector &induction,
+                            scf::WhileOp whileOp) {
   Location loc = op.getLoc();
   // Finalize each else branch of all if statements.
   if (codegen.redVal || codegen.expValues) {
     while (auto ifOp = dyn_cast_or_null<scf::IfOp>(
                builder.getInsertionBlock()->getParentOp())) {
       unsigned y = 0;
       SmallVector<Value, 4> yields;
       if (codegen.redVal) {
         yields.push_back(codegen.redVal);
         updateReduc(merger, codegen, ifOp.getResult(y++));
       }
       if (codegen.expValues) {
         yields.push_back(codegen.expCount);
         codegen.expCount = ifOp->getResult(y++);
       }
       assert(y == yields.size());
       builder.create<scf::YieldOp>(loc, yields);
       builder.setInsertionPointAfter(ifOp);
     }
   }
   builder.setInsertionPointToEnd(&whileOp.getAfter().front());
-  // Finalize the induction. Note that the induction could be performed
-  // in the individual if-branches to avoid re-evaluating the conditions.
-  // However, that would result in a rather elaborate forest of yield
-  // instructions during code generation. Moreover, performing the induction
-  // after the if-statements more closely resembles code generated by TACO.
-  unsigned o = 0;
-  SmallVector<Value, 4> operands;
-  Value one = constantIndex(builder, loc, 1);
-  for (unsigned b = 0, be = induction.size(); b < be; b++) {
-    if (!induction[b])
-      continue;
-    if (isCompressedDLT(merger.getDimLevelType(b)) ||
-        isSingletonDLT(merger.getDimLevelType(b))) {
-      unsigned tensor = merger.tensor(b);
-      assert(idx == merger.index(b));
-      Value op1 = codegen.idxs[tensor][idx];
-      Value op2 = codegen.loops[idx];
-      Value op3 = codegen.pidxs[tensor][idx];
-      Value cmp = builder.create<arith::CmpIOp>(loc, arith::CmpIPredicate::eq,
-                                                op1, op2);
-      Value add = builder.create<arith::AddIOp>(loc, op3, one);
-      operands.push_back(builder.create<arith::SelectOp>(loc, cmp, add, op3));
-      codegen.pidxs[tensor][idx] = whileOp->getResult(o++);
-    } else {
-      assert(isDenseDLT(merger.getDimLevelType(b)) ||
-             isUndefDLT(merger.getDimLevelType(b)));
-    }
-  }
-  if (codegen.redVal) {
-    operands.push_back(codegen.redVal);
-    updateReduc(merger, codegen, whileOp->getResult(o++));
-  }
-  if (codegen.expValues) {
-    operands.push_back(codegen.expCount);
-    codegen.expCount = whileOp->getResult(o++);
-  }
-  if (needsUniv) {
-    operands.push_back(
-        builder.create<arith::AddIOp>(loc, codegen.loops[idx], one));
-    codegen.loops[idx] = whileOp->getResult(o++);
-  }
-  assert(o == operands.size());
-  builder.create<scf::YieldOp>(loc, operands);
-  builder.setInsertionPointAfter(whileOp);
-}
-
-/// Generates the induction structure for a for-loop.
-static void genForInduction(Merger &merger, CodeGen &codegen,
-                            OpBuilder &builder, linalg::GenericOp op,
-                            Operation *loop) {
-  Location loc = op.getLoc();
-  unsigned o = 0;
-  SmallVector<Value, 4> operands;
-  if (codegen.redVal) {
-    operands.push_back(codegen.redVal);
-    updateReduc(merger, codegen, loop->getResult(o++));
-  }
-  if (codegen.expValues) {
-    operands.push_back(codegen.expCount);
-    codegen.expCount = loop->getResult(o++);
-  }
-  assert(o == operands.size());
-  if (o > 0)
-    builder.create<scf::YieldOp>(loc, operands);
-  builder.setInsertionPointAfter(loop);
 }
 
 /// Generates a single if-statement within a while-loop.
 static scf::IfOp genIf(Merger &merger, CodeGen &codegen, OpBuilder &builder,
                        linalg::GenericOp op, unsigned idx,
                        BitVector &conditions) {
   Location loc = op.getLoc();
   SmallVector<Type, 4> types;
   Value cond;
   for (unsigned b = 0, be = conditions.size(); b < be; b++) {
     if (!conditions[b])
       continue;
     unsigned tensor = merger.tensor(b);
     assert(idx == merger.index(b));
     Value clause;
     if (isCompressedDLT(merger.getDimLevelType(b)) ||
         isSingletonDLT(merger.getDimLevelType(b))) {
-      Value op1 = codegen.idxs[tensor][idx];
-      Value op2 = codegen.loops[idx];
+      auto dim = codegen.loopIdxToDim[tensor][idx];
+      assert(dim != INVALID_ID);
+      Value op1 = codegen.loopEmitter.getCoord()[tensor][dim];
+      Value op2 = codegen.getLoopIdxValue(idx);
       clause = builder.create<arith::CmpIOp>(loc, arith::CmpIPredicate::eq, op1,
                                              op2);
     } else {
       assert(isDenseDLT(merger.getDimLevelType(b)) ||
              isUndefDLT(merger.getDimLevelType(b)));
       clause = constantI1(builder, loc, true);
     }
     cond = cond ? builder.create<arith::AndIOp>(loc, cond, clause) : clause;

 // Sparse compiler synthesis methods (loop sequence).
 //===----------------------------------------------------------------------===//
 
 /// Starts a loop sequence at given level. Returns true if
 /// the universal loop index must be maintained at this level.
 static bool startLoopSeq(Merger &merger, CodeGen &codegen, OpBuilder &builder,
                          linalg::GenericOp op, unsigned exp, unsigned at,
                          unsigned idx, unsigned ldx, unsigned lts) {
-  assert(!codegen.loops[idx]);
+  assert(!codegen.getLoopIdxValue(idx));
   // Emit invariants at this loop sequence level.
   genInvariants(merger, codegen, builder, op, exp, ldx, /*atStart=*/true);
   // Emit access pattern expansion for sparse tensor output.
   genExpansion(merger, codegen, builder, op, at, /*atStart=*/true);
   // Emit further intitialization at this loop sequence level.
   unsigned l0 = merger.set(lts)[0];
-  bool needsUniv =
-      genInit(merger, codegen, builder, op, at, merger.lat(l0).bits);
+  bool needsUniv = false;
+
+  SmallVector<size_t, 4> ts;
+  SmallVector<size_t, 4> ds;
+  for (auto b : merger.lat(l0).bits.set_bits()) {
+    if (isDenseDLT(merger.getDimLevelType(b)) ||
+        isUndefDLT(merger.getDimLevelType(b))) {
+      needsUniv = true;
+    } else {
+      unsigned tensor = merger.tensor(b);
+      assert(idx == merger.index(b));
+      size_t dim = codegen.loopIdxToDim[tensor][idx];
+      assert(dim != INVALID_ID);
+      ts.push_back(tensor);
+      ds.push_back(dim);
+    }
+  }
+
+  codegen.loopEmitter.enterNewLoopSeq(builder, op.getLoc(), ts, ds);
+
   // Maintain the universal index only if it is actually
   // consumed by a subsequent lattice point.
   if (needsUniv) {
     unsigned lsize = merger.set(lts).size();
     for (unsigned i = 1; i < lsize; i++) {
       unsigned li = merger.set(lts)[i];
       if (!merger.hasAnySparse(merger.lat(li).simple))
         return true;
     }
   }
   return false;
 }
 
+static void translateBitsToTidDimPairs(Merger &merger, CodeGen &codegen,
+                                       unsigned li, unsigned idx,
+                                       SmallVectorImpl<size_t> &condTids,
+                                       SmallVectorImpl<size_t> &condDims,
+                                       SmallVectorImpl<size_t> &extraTids,
+                                       SmallVectorImpl<size_t> &extraDims) {
+  const BitVector &simple = merger.lat(li).simple;
+  const BitVector &all = merger.lat(li).bits;
+  assert(simple.size() == all.size());
+  // First converts bits to array + dim pair
+  for (unsigned b = 0, e = simple.size(); b < e; b++) {
+    size_t tid = merger.tensor(b);
+    if (simple.test(b)) {
+      // the simplified condition must be a subset of the original condition.
+      assert(all.test(b));
+      assert(merger.index(b) == idx);
+      if (isUndefDLT(merger.getDimLevelType(b))) {
+        // This could be a synthetic tensor (for invariants and sparse output
+        // tensor).
+        // In both cases, we mean to generate loops over output tensor.
+        // e.g.,
+        // out[i][j] = invariant;
+        if (merger.getSynTensorID() == tid)
+          tid = merger.getOutTensorID();
+      }
+      auto dim = codegen.loopIdxToDim[tid][idx];
+      if (dim != INVALID_ID) {
+        // dim could be invalid if this is a zero ranked tensor
+        condTids.push_back(tid);
+        condDims.push_back(dim);
+      }
+    } else if ((all.test(b) || merger.isOutTensor(b, idx)) &&
+               isDenseDLT(merger.getDimLevelType(b))) {
+      assert(merger.index(b) == idx);
+      // Note that we generate dense indices of the output tensor
+      // unconditionally, since they may not appear in the lattice, but may be
+      // needed for linearized codegen.
+      // Only dense dimensions should be optimized from conditions.
+      assert(isDenseDLT(merger.getDimLevelType(b)));
+      auto dim = codegen.loopIdxToDim[tid][idx];
+      assert(dim != INVALID_ID);
+      extraTids.push_back(tid);
+      extraDims.push_back(dim);
+    }
+  }
+}
+
 /// Starts a single loop in current sequence.
 static Operation *startLoop(Merger &merger, CodeGen &codegen,
                             OpBuilder &builder, linalg::GenericOp op,
                             unsigned at, unsigned li, bool needsUniv) {
+  // The set of tensors + dims to generate loops on
+  SmallVector<size_t, 4> condTids, condDims;
+  // The set of (dense) tensors that is optimized from condition, yet still
+  // need extra locals to iterate on them.
+  SmallVector<size_t, 4> extraTids, extraDims;
+
+  translateBitsToTidDimPairs(merger, codegen, li, codegen.topSort[at], condTids,
+                             condDims, extraTids, extraDims);
   // Emit the for/while-loop control.
   Operation *loop = genLoop(merger, codegen, builder, op, at, needsUniv,
-                            merger.lat(li).simple);
-  // Emit the locals for this loop.
-  genLocals(merger, codegen, builder, op, at, needsUniv, merger.lat(li).bits);
+                            condTids, condDims, extraTids, extraDims);
   return loop;
 }
 
 /// Ends a single loop in current sequence. Returns new values for needsUniv.
 static bool endLoop(Merger &merger, CodeGen &codegen, OpBuilder &builder,
                     linalg::GenericOp op, Operation *loop, unsigned idx,
                     unsigned li, bool needsUniv) {
   // End a while-loop.
   if (auto whileOp = dyn_cast<scf::WhileOp>(loop)) {
-    genWhileInduction(merger, codegen, builder, op, idx, needsUniv,
-                      merger.lat(li).bits, whileOp);
-    return needsUniv;
+    finalizeWhileOp(merger, codegen, builder, op, idx, needsUniv,
+  } else {
+    needsUniv = false;
   }
-  // End a for-loop.
-  genForInduction(merger, codegen, builder, op, loop);
-  return false;
+
+  SmallVector<Value, 2> reduc;
+  if (codegen.redVal)
+    reduc.push_back(codegen.redVal);
+  if (codegen.expValues)
+    reduc.push_back(codegen.expCount);
+
+  auto loopRet =
+      codegen.loopEmitter.exitCurrentLoop(builder, op.getLoc(), reduc);
+  assert(reduc.size() == loopRet.size());
+
+  if (codegen.redVal)
+    updateReduc(merger, codegen, loopRet.front());
+  if (codegen.expValues)
+    codegen.expCount = loopRet.back();
+
+  return needsUniv;
 }
 
 /// Ends a loop sequence at given level.
 static void endLoopSeq(Merger &merger, CodeGen &codegen, OpBuilder &builder,
                        linalg::GenericOp op, unsigned exp, unsigned at,
                        unsigned idx, unsigned ldx) {
-  assert(codegen.loops[idx]);
-  codegen.loops[idx] = Value();
+  assert(codegen.getLoopIdxValue(idx) == nullptr);
+  codegen.loopEmitter.exitCurrentLoopSeq();
   // Unmark bookkeeping of invariants and loop index.
   genInvariants(merger, codegen, builder, op, exp, ldx, /*atStart=*/false);
   // Finalize access pattern expansion for sparse tensor output.
   genExpansion(merger, codegen, builder, op, at, /*atStart=*/false);
 }
 
 /// Recursively generates code while computing iteration lattices in order
 /// to manage the complexity of implementing co-iteration over unions
static void genResult(Merger &merger, CodeGen &codegen, RewriterBase &rewriter,
   if (getSparseTensorEncoding(resType)) {
     // The sparse tensor rematerializes from the original sparse tensor's
     // underlying sparse storage format.
     rewriter.replaceOpWithNewOp<LoadOp>(op, resType, lhs->get(),
                                         codegen.sparseOut == lhs);
   } else {
     // To rematerialize an non-annotated tensor, simply load it
     // from the bufferized value.
-    Value val = codegen.buffers.back(); // value array
+    Value val = codegen.loopEmitter.getValBuffer().back(); // value array
     rewriter.replaceOpWithNewOp<bufferization::ToTensorOp>(op, resType, val);
   }
 }
 
 //===----------------------------------------------------------------------===//
 // Sparse compiler rewriting methods.
 //===----------------------------------------------------------------------===//
 
LogicalResult matchAndRewrite(linalg::GenericOp op,
 
     if (hasCycle)
       // Give it one last shot to resolve the cycle.
       return resolveCycle(merger, rewriter, op);
     if (!isAdmissible)
       // Inadmissible expression, reject.
       return failure();
 
-    // Recursively generates code if admissible.
     merger.setHasSparseOut(sparseOut != nullptr);
-    CodeGen codegen(options, numTensors, numLoops, sparseOut, outerParNest,
-                    topSort);
+
+    SmallVector<Value, 4> tensors;
+    for (OpOperand &t : op->getOpOperands())
+      tensors.push_back(t.get());
+
+    // Recursively generates code if admissible.
+    CodeGen codegen(options, tensors, numTensors, numLoops, sparseOut,
+                    outerParNest, topSort);
+    // TODO: maybe merger should be responsible of maintaining the map.
+    codegen.buildLoopIdxToDimMap(op);
     genBuffers(merger, codegen, rewriter, op);
     genStmt(merger, codegen, rewriter, op, exp, 0);
     genResult(merger, codegen, rewriter, op);
     return success();
   }
 
 private:
   // Last resort cycle resolution.

mlir/test/Dialect/SparseTensor/sorted_coo.mlir

 // CHECK-SAME:    %[[VAL_1:.*]]: tensor<64xf64>,
 // CHECK-SAME:    %[[VAL_2:.*]]: tensor<32xf64>) -> tensor<32xf64> {
 // CHECK-DAG:     %[[VAL_3:.*]] = arith.constant 0 : index
 // CHECK-DAG:     %[[VAL_4:.*]] = arith.constant 1 : index
 // CHECK-DAG:     %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
 // CHECK-DAG:     %[[VAL_6:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
 // CHECK-DAG:     %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 1 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
 // CHECK-DAG:     %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xf64>
-// CHECK:         %[[VAL_9:.*]] = bufferization.to_memref %[[VAL_1]] : memref<64xf64>
-// CHECK:         %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_2]] : memref<32xf64>
-// CHECK:         %[[VAL_11:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:         %[[VAL_12:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:     %[[VAL_9:.*]] = bufferization.to_memref %[[VAL_1]] : memref<64xf64>
+// CHECK-DAG:     %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_2]] : memref<32xf64>
+// CHECK-DAG:     %[[VAL_11:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:     %[[VAL_12:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
 // CHECK:         scf.for %[[VAL_13:.*]] = %[[VAL_11]] to %[[VAL_12]] step %[[VAL_4]] {
 // CHECK:           %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_13]]] : memref<?xindex>
 // CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_14]]] : memref<32xf64>
 // CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_13]]] : memref<?xindex>
 // CHECK:           %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_13]]] : memref<?xf64>
 // CHECK:           %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_16]]] : memref<64xf64>
 // CHECK:           %[[VAL_19:.*]] = arith.mulf %[[VAL_17]], %[[VAL_18]] : f64
 // CHECK:           %[[VAL_20:.*]] = arith.addf %[[VAL_15]], %[[VAL_19]] : f64

 // CHECK-DAG:     %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
 // CHECK-DAG:     %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
 // CHECK-DAG:     %[[VAL_8:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 1 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
 // CHECK-DAG:     %[[VAL_9:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xf64>
 // CHECK-DAG:     %[[VAL_10:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
 // CHECK-DAG:     %[[VAL_11:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
 // CHECK-DAG:     %[[VAL_12:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 1 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
 // CHECK-DAG:     %[[VAL_13:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xf64>
-// CHECK:         %[[VAL_14:.*]] = bufferization.to_memref %[[VAL_2]] : memref<32x64xf64>
-// CHECK:         linalg.fill ins(%[[VAL_3]] : f64) outs(%[[VAL_14]] : memref<32x64xf64>)
-// CHECK:         %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:         %[[VAL_16:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
-// CHECK:         %[[VAL_17:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:         %[[VAL_18:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:     %[[VAL_14:.*]] = bufferization.to_memref %[[VAL_2]] : memref<32x64xf64>
+// CHECK-DAG:     linalg.fill ins(%[[VAL_3]] : f64) outs(%[[VAL_14]] : memref<32x64xf64>)
+// CHECK-DAG:     %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:     %[[VAL_16:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:     %[[VAL_17:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:     %[[VAL_18:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_5]]] : memref<?xindex>
 // CHECK:         %[[VAL_19:.*]]:2 = scf.while (%[[VAL_20:.*]] = %[[VAL_15]], %[[VAL_21:.*]] = %[[VAL_17]]) : (index, index) -> (index, index) {
 // CHECK:           %[[VAL_22:.*]] = arith.cmpi ult, %[[VAL_20]], %[[VAL_16]] : index
 // CHECK:           %[[VAL_23:.*]] = arith.cmpi ult, %[[VAL_21]], %[[VAL_18]] : index
 // CHECK:           %[[VAL_24:.*]] = arith.andi %[[VAL_22]], %[[VAL_23]] : i1
 // CHECK:           scf.condition(%[[VAL_24]]) %[[VAL_20]], %[[VAL_21]] : index, index
 // CHECK:         } do {
 // CHECK:         ^bb0(%[[VAL_25:.*]]: index, %[[VAL_26:.*]]: index):
 // CHECK:           %[[VAL_27:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_25]]] : memref<?xindex>

mlir/test/Dialect/SparseTensor/sparse_1d.mlir

 // CHECK-SAME:      %[[VAL_2:.*]]: tensor<32xf32>) -> tensor<32xf32> {
 // CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 32 : index
 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = arith.constant true
 // CHECK-DAG:       %[[VAL_6:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
+// CHECK-DAG:       %[[VAL_12:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_13:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_6]]] : memref<?xindex>
 // CHECK-DAG:       %[[VAL_11:.*]] = bufferization.to_memref %[[VAL_2]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_11]] : memref<32xf32>)
-// CHECK:           %[[VAL_12:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_13:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_6]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_11]] : memref<32xf32>)
 // CHECK:           %[[VAL_14:.*]]:2 = scf.while (%[[VAL_15:.*]] = %[[VAL_12]], %[[VAL_16:.*]] = %[[VAL_4]]) : (index, index) -> (index, index) {
 // CHECK:             %[[VAL_17:.*]] = arith.cmpi ult, %[[VAL_15]], %[[VAL_13]] : index
 // CHECK:             scf.condition(%[[VAL_17]]) %[[VAL_15]], %[[VAL_16]] : index, index
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_18:.*]]: index, %[[VAL_19:.*]]: index):
 // CHECK:             %[[VAL_20:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_18]]] : memref<?xindex>
 // CHECK:             %[[VAL_21:.*]] = arith.cmpi eq, %[[VAL_20]], %[[VAL_19]] : index
 // CHECK:             scf.if %[[VAL_21]] {

 // CHECK-SAME:      %[[VAL_0:.*]]: tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>>,
 // CHECK-SAME:      %[[VAL_1:.*]]: tensor<32xf32>) -> tensor<32xf32> {
 // CHECK-DAG:       %[[VAL_2:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_4:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_5:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_8:.*]] = bufferization.to_memref %[[VAL_1]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_8]] : memref<32xf32>)
-// CHECK:           %[[VAL_9:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_2]]] : memref<?xindex>
-// CHECK:           %[[VAL_10:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_9:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_2]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_10:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_8]] : memref<32xf32>)
 // CHECK:           scf.for %[[VAL_11:.*]] = %[[VAL_9]] to %[[VAL_10]] step %[[VAL_3]] {
 // CHECK:             %[[VAL_12:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_11]]] : memref<?xindex>
 // CHECK:             %[[VAL_13:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_11]]] : memref<?xf32>
 // CHECK:             %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_11]]] : memref<?xf32>
 // CHECK:             %[[VAL_15:.*]] = arith.addf %[[VAL_13]], %[[VAL_14]] : f32
 // CHECK:             %[[VAL_16:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_11]]] : memref<?xf32>
 // CHECK:             %[[VAL_17:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_11]]] : memref<?xf32>
 // CHECK:             %[[VAL_18:.*]] = arith.addf %[[VAL_16]], %[[VAL_17]] : f32

 // CHECK-SAME:      %[[VAL_1:.*]]: f32,
 // CHECK-SAME:      %[[VAL_2:.*]]: tensor<32xf32>) -> tensor<32xf32> {
 // CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_9:.*]] = bufferization.to_memref %[[VAL_2]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_9]] : memref<32xf32>)
-// CHECK:           %[[VAL_10:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:           %[[VAL_11:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_9]] : memref<32xf32>)
+// CHECK-DAG:       %[[VAL_10:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_11:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
 // CHECK:           scf.for %[[VAL_12:.*]] = %[[VAL_10]] to %[[VAL_11]] step %[[VAL_4]] {
 // CHECK:             %[[VAL_13:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_12]]] : memref<?xindex>
 // CHECK:             %[[VAL_14:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_12]]] : memref<?xf32>
 // CHECK:             %[[VAL_15:.*]] = arith.mulf %[[VAL_14]], %[[VAL_1]] : f32
 // CHECK:             memref.store %[[VAL_15]], %[[VAL_9]]{{\[}}%[[VAL_13]]] : memref<32xf32>
 // CHECK:           }
 // CHECK:           %[[VAL_16:.*]] = bufferization.to_tensor %[[VAL_9]] : memref<32xf32>
 // CHECK:           return %[[VAL_16]] : tensor<32xf32>

 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = arith.constant true
 // CHECK-DAG:       %[[VAL_6:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_7:.*]] = bufferization.to_memref %[[VAL_0]] : memref<32xf32>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_12]] : memref<32xf32>)
-// CHECK:           %[[VAL_13:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_14:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_6]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_12]] : memref<32xf32>)
+// CHECK-DAG:       %[[VAL_13:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_14:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_6]]] : memref<?xindex>
 // CHECK:           %[[VAL_15:.*]]:2 = scf.while (%[[VAL_16:.*]] = %[[VAL_13]], %[[VAL_17:.*]] = %[[VAL_4]]) : (index, index) -> (index, index) {
 // CHECK:             %[[VAL_18:.*]] = arith.cmpi ult, %[[VAL_16]], %[[VAL_14]] : index
 // CHECK:             scf.condition(%[[VAL_18]]) %[[VAL_16]], %[[VAL_17]] : index, index
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_19:.*]]: index, %[[VAL_20:.*]]: index):
 // CHECK:             %[[VAL_21:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_19]]] : memref<?xindex>
 // CHECK:             %[[VAL_22:.*]] = arith.cmpi eq, %[[VAL_21]], %[[VAL_20]] : index
 // CHECK:             scf.if %[[VAL_22]] {

 // CHECK-SAME:      %[[VAL_2:.*]]: tensor<32xf32>) -> tensor<32xf32> {
 // CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = bufferization.to_memref %[[VAL_0]] : memref<32xf32>
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_2]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_10]] : memref<32xf32>)
-// CHECK:           %[[VAL_11:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:           %[[VAL_12:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_10]] : memref<32xf32>)
+// CHECK-DAG:       %[[VAL_11:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_12:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
 // CHECK:           scf.for %[[VAL_13:.*]] = %[[VAL_11]] to %[[VAL_12]] step %[[VAL_4]] {
 // CHECK:             %[[VAL_14:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_13]]] : memref<?xindex>
 // CHECK:             %[[VAL_15:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_14]]] : memref<32xf32>
 // CHECK:             %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_13]]] : memref<?xf32>
 // CHECK:             %[[VAL_17:.*]] = arith.mulf %[[VAL_15]], %[[VAL_16]] : f32
 // CHECK:             memref.store %[[VAL_17]], %[[VAL_10]]{{\[}}%[[VAL_14]]] : memref<32xf32>
 // CHECK:           }
 // CHECK:           %[[VAL_18:.*]] = bufferization.to_tensor %[[VAL_10]] : memref<32xf32>

 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = arith.constant true
 // CHECK-DAG:       %[[VAL_6:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_1]] : memref<32xf32>
 // CHECK-DAG:       %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_12]] : memref<32xf32>)
-// CHECK:           %[[VAL_13:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_14:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_6]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_12]] : memref<32xf32>)
+// CHECK-DAG:       %[[VAL_13:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_14:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_6]]] : memref<?xindex>
 // CHECK:           %[[VAL_15:.*]]:2 = scf.while (%[[VAL_16:.*]] = %[[VAL_13]], %[[VAL_17:.*]] = %[[VAL_4]]) : (index, index) -> (index, index) {
 // CHECK:             %[[VAL_18:.*]] = arith.cmpi ult, %[[VAL_16]], %[[VAL_14]] : index
 // CHECK:             scf.condition(%[[VAL_18]]) %[[VAL_16]], %[[VAL_17]] : index, index
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_19:.*]]: index, %[[VAL_20:.*]]: index):
 // CHECK:             %[[VAL_21:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_19]]] : memref<?xindex>
 // CHECK:             %[[VAL_22:.*]] = arith.cmpi eq, %[[VAL_21]], %[[VAL_20]] : index
 // CHECK:             scf.if %[[VAL_22]] {

 // CHECK-SAME:      %[[VAL_2:.*]]: tensor<32xf32>) -> tensor<32xf32> {
 // CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_8:.*]] = bufferization.to_memref %[[VAL_1]] : memref<32xf32>
 // CHECK-DAG:       %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_2]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_10]] : memref<32xf32>)
-// CHECK:           %[[VAL_11:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:           %[[VAL_12:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_10]] : memref<32xf32>)
+// CHECK-DAG:       %[[VAL_11:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_12:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
 // CHECK:           scf.for %[[VAL_13:.*]] = %[[VAL_11]] to %[[VAL_12]] step %[[VAL_4]] {
 // CHECK:             %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_13]]] : memref<?xindex>
 // CHECK:             %[[VAL_15:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_13]]] : memref<?xf32>
 // CHECK:             %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_14]]] : memref<32xf32>
 // CHECK:             %[[VAL_17:.*]] = arith.mulf %[[VAL_15]], %[[VAL_16]] : f32
 // CHECK:             memref.store %[[VAL_17]], %[[VAL_10]]{{\[}}%[[VAL_14]]] : memref<32xf32>
 // CHECK:           }
 // CHECK:           %[[VAL_18:.*]] = bufferization.to_tensor %[[VAL_10]] : memref<32xf32>

 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_12]] : memref<32xf32>)
-// CHECK:           %[[VAL_13:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:           %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_12]] : memref<32xf32>)
+// CHECK-DAG:       %[[VAL_13:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex>
 // CHECK:           %[[VAL_17:.*]]:2 = scf.while (%[[VAL_18:.*]] = %[[VAL_13]], %[[VAL_19:.*]] = %[[VAL_15]]) : (index, index) -> (index, index) {
 // CHECK:             %[[VAL_20:.*]] = arith.cmpi ult, %[[VAL_18]], %[[VAL_14]] : index
 // CHECK:             %[[VAL_21:.*]] = arith.cmpi ult, %[[VAL_19]], %[[VAL_16]] : index
 // CHECK:             %[[VAL_22:.*]] = arith.andi %[[VAL_20]], %[[VAL_21]] : i1
 // CHECK:             scf.condition(%[[VAL_22]]) %[[VAL_18]], %[[VAL_19]] : index, index
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_23:.*]]: index, %[[VAL_24:.*]]: index):
 // CHECK:             %[[VAL_25:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_23]]] : memref<?xindex>

 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<32xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_12]] : memref<32xf32>)
-// CHECK:           %[[VAL_13:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:           %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_12]] : memref<32xf32>)
+// CHECK-DAG:       %[[VAL_13:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex>
 // CHECK:           %[[VAL_17:.*]]:2 = scf.while (%[[VAL_18:.*]] = %[[VAL_13]], %[[VAL_19:.*]] = %[[VAL_15]]) : (index, index) -> (index, index) {
 // CHECK:             %[[VAL_20:.*]] = arith.cmpi ult, %[[VAL_18]], %[[VAL_14]] : index
 // CHECK:             %[[VAL_21:.*]] = arith.cmpi ult, %[[VAL_19]], %[[VAL_16]] : index
 // CHECK:             %[[VAL_22:.*]] = arith.andi %[[VAL_20]], %[[VAL_21]] : i1
 // CHECK:             scf.condition(%[[VAL_22]]) %[[VAL_18]], %[[VAL_19]] : index, index
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_23:.*]]: index, %[[VAL_24:.*]]: index):
 // CHECK:             %[[VAL_25:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_23]]] : memref<?xindex>

 // CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_11:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_13:.*]] = bufferization.to_memref %[[VAL_3]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_13]] : memref<16xf32>)
-// CHECK:           %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
-// CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_17:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_13]] : memref<16xf32>)
+// CHECK-DAG:       %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_16:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_17:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_5]]] : memref<?xindex>
 // CHECK:           %[[VAL_18:.*]]:2 = scf.while (%[[VAL_19:.*]] = %[[VAL_14]], %[[VAL_20:.*]] = %[[VAL_16]]) : (index, index) -> (index, index) {
 // CHECK:             %[[VAL_21:.*]] = arith.cmpi ult, %[[VAL_19]], %[[VAL_15]] : index
 // CHECK:             %[[VAL_22:.*]] = arith.cmpi ult, %[[VAL_20]], %[[VAL_17]] : index
 // CHECK:             %[[VAL_23:.*]] = arith.andi %[[VAL_21]], %[[VAL_22]] : i1
 // CHECK:             scf.condition(%[[VAL_23]]) %[[VAL_19]], %[[VAL_20]] : index, index
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_24:.*]]: index, %[[VAL_25:.*]]: index):
 // CHECK:             %[[VAL_26:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_24]]] : memref<?xindex>

 // CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_11:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_13:.*]] = bufferization.to_memref %[[VAL_3]]
-// CHECK:           linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_13]] : memref<16xf32>)
-// CHECK:           %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
-// CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_17:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f32) outs(%[[VAL_13]] : memref<16xf32>)
+// CHECK-DAG:       %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_16:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_17:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_5]]] : memref<?xindex>
 // CHECK:           %[[VAL_18:.*]]:2 = scf.while (%[[VAL_19:.*]] = %[[VAL_14]], %[[VAL_20:.*]] = %[[VAL_16]]) : (index, index) -> (index, index) {
 // CHECK:             %[[VAL_21:.*]] = arith.cmpi ult, %[[VAL_19]], %[[VAL_15]] : index
 // CHECK:             %[[VAL_22:.*]] = arith.cmpi ult, %[[VAL_20]], %[[VAL_17]] : index
 // CHECK:             %[[VAL_23:.*]] = arith.andi %[[VAL_21]], %[[VAL_22]] : i1
 // CHECK:             scf.condition(%[[VAL_23]]) %[[VAL_19]], %[[VAL_20]] : index, index
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_24:.*]]: index, %[[VAL_25:.*]]: index):
 // CHECK:             %[[VAL_26:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_24]]] : memref<?xindex>

 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_11:.*]] = bufferization.to_memref %[[VAL_2]] : memref<f32>
-// CHECK:           %[[VAL_13:.*]] = memref.load %[[VAL_11]][] : memref<f32>
-// CHECK:           %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_3]]] : memref<?xindex>
-// CHECK:           %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:           %[[VAL_13:.*]] = memref.load %[[VAL_11]][] : memref<f32>
+// CHECK-DAG:           %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:           %[[VAL_15:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:           %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_3]]] : memref<?xindex>
+// CHECK-DAG:           %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex>
 // CHECK:           %[[VAL_18:.*]]:3 = scf.while (%[[VAL_19:.*]] = %[[VAL_14]], %[[VAL_20:.*]] = %[[VAL_16]], %[[VAL_21:.*]] = %[[VAL_13]]) : (index, index, f32) -> (index, index, f32) {
 // CHECK:             %[[VAL_22:.*]] = arith.cmpi ult, %[[VAL_19]], %[[VAL_15]] : index
 // CHECK:             %[[VAL_23:.*]] = arith.cmpi ult, %[[VAL_20]], %[[VAL_17]] : index
 // CHECK:             %[[VAL_24:.*]] = arith.andi %[[VAL_22]], %[[VAL_23]] : i1
 // CHECK:             scf.condition(%[[VAL_24]]) %[[VAL_19]], %[[VAL_20]], %[[VAL_21]] : index, index, f32
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_25:.*]]: index, %[[VAL_26:.*]]: index, %[[VAL_27:.*]]: f32):
 // CHECK:             %[[VAL_28:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_25]]] : memref<?xindex>

 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_9:.*]] = bufferization.to_memref %[[VAL_1]] : memref<f32>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.pointers %[[VAL_2]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_11:.*]] = sparse_tensor.indices %[[VAL_2]] {dimension = 0 : index} : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_12:.*]] = sparse_tensor.values %[[VAL_2]] : tensor<16xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_13:.*]] = bufferization.to_memref %[[VAL_3]] : memref<f32>
-// CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_13]][] : memref<f32>
-// CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_9]][] : memref<f32>
-// CHECK:           %[[VAL_17:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_18:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
-// CHECK:           %[[VAL_19:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_20:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_15:.*]] = memref.load %[[VAL_13]][] : memref<f32>
+// CHECK-DAG:       %[[VAL_16:.*]] = memref.load %[[VAL_9]][] : memref<f32>
+// CHECK-DAG:       %[[VAL_17:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_18:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_19:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_20:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_5]]] : memref<?xindex>
 // CHECK:           %[[VAL_21:.*]]:3 = scf.while (%[[VAL_22:.*]] = %[[VAL_17]], %[[VAL_23:.*]] = %[[VAL_19]], %[[VAL_24:.*]] = %[[VAL_15]]) : (index, index, f32) -> (index, index, f32) {
 // CHECK:             %[[VAL_25:.*]] = arith.cmpi ult, %[[VAL_22]], %[[VAL_18]] : index
 // CHECK:             %[[VAL_26:.*]] = arith.cmpi ult, %[[VAL_23]], %[[VAL_20]] : index
 // CHECK:             %[[VAL_27:.*]] = arith.andi %[[VAL_25]], %[[VAL_26]] : i1
 // CHECK:             scf.condition(%[[VAL_27]]) %[[VAL_22]], %[[VAL_23]], %[[VAL_24]] : index, index, f32
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_28:.*]]: index, %[[VAL_29:.*]]: index, %[[VAL_30:.*]]: f32):
 // CHECK:             %[[VAL_31:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_28]]] : memref<?xindex>

 // CHECK-DAG:       %[[VAL_8:.*]] = bufferization.to_memref %[[VAL_0]] : memref<?xf64>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_11:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf64>
 // CHECK-DAG:       %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2]] : memref<?xf64>
 // CHECK-DAG:       %[[VAL_13:.*]] = sparse_tensor.pointers %[[VAL_3]] {dimension = 0 : index} : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_14:.*]] = sparse_tensor.indices %[[VAL_3]] {dimension = 0 : index} : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_15:.*]] = sparse_tensor.values %[[VAL_3]] : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf64>
-// CHECK-DAG:       %[[VAL_16:.*]] = tensor.dim %[[VAL_4]], %[[VAL_5]] : tensor<?xf64>
+// CHECK-DAG:       %[[VAL_16:.*]] = tensor.dim %[[VAL_0]], %[[VAL_5]] : tensor<?xf64>
 // CHECK-DAG:       %[[VAL_18:.*]] = bufferization.to_memref %[[VAL_4]]
-// CHECK:           linalg.fill ins(%{{.*}} : f64) outs(%[[VAL_18]] : memref<?xf64>)
-// CHECK:           %[[VAL_19:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_5]]] : memref<?xindex>
-// CHECK:           %[[VAL_20:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_7]]] : memref<?xindex>
-// CHECK:           %[[VAL_21:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_5]]] : memref<?xindex>
-// CHECK:           %[[VAL_22:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_7]]] : memref<?xindex>
+// CHECK-DAG:       linalg.fill ins(%{{.*}} : f64) outs(%[[VAL_18]] : memref<?xf64>)
+// CHECK-DAG:       %[[VAL_19:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_20:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_7]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_21:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_22:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_7]]] : memref<?xindex>
 // CHECK:           %[[VAL_23:.*]]:3 = scf.while (%[[VAL_24:.*]] = %[[VAL_19]], %[[VAL_25:.*]] = %[[VAL_21]], %[[VAL_26:.*]] = %[[VAL_5]]) : (index, index, index) -> (index, index, index) {
 // CHECK:             %[[VAL_27:.*]] = arith.cmpi ult, %[[VAL_24]], %[[VAL_20]] : index
 // CHECK:             %[[VAL_28:.*]] = arith.cmpi ult, %[[VAL_25]], %[[VAL_22]] : index
 // CHECK:             %[[VAL_29:.*]] = arith.andi %[[VAL_27]], %[[VAL_28]] : i1
 // CHECK:             scf.condition(%[[VAL_29]]) %[[VAL_24]], %[[VAL_25]], %[[VAL_26]] : index, index, index
 // CHECK:           } do {
 // CHECK:           ^bb0(%[[VAL_30:.*]]: index, %[[VAL_31:.*]]: index, %[[VAL_32:.*]]: index):
 // CHECK:             %[[VAL_33:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_30]]] : memref<?xindex>

 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf64>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 0 : index} : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_11:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf64>
 // CHECK-DAG:       %[[VAL_12:.*]] = sparse_tensor.pointers %[[VAL_2]] {dimension = 0 : index} : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_13:.*]] = sparse_tensor.indices %[[VAL_2]] {dimension = 0 : index} : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_14:.*]] = sparse_tensor.values %[[VAL_2]] : tensor<?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xf64>
 // CHECK-DAG:       %[[VAL_15:.*]] = bufferization.to_memref %[[VAL_3]] : memref<f64>
-// CHECK:           %[[VAL_17:.*]] = memref.load %[[VAL_15]][] : memref<f64>
-// CHECK:           %[[VAL_18:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_19:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
-// CHECK:           %[[VAL_20:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_21:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_5]]] : memref<?xindex>
-// CHECK:           %[[VAL_22:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_4]]] : memref<?xindex>
-// CHECK:           %[[VAL_23:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_17:.*]] = memref.load %[[VAL_15]][] : memref<f64>
+// CHECK-DAG:       %[[VAL_18:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_19:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_20:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_21:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_5]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_22:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_4]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_23:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_5]]] : memref<?xindex>
 // CHECK:           %[[VAL_24:.*]]:4 = scf.while (%[[VAL_25:.*]] = %[[VAL_18]], %[[VAL_26:.*]] = %[[VAL_20]], %[[VAL_27:.*]] = %[[VAL_22]], %[[VAL_28:.*]] = %[[VAL_17]]) : (index, index, index, f64) -> (index, index, index, f64) {
 // CHECK:             %[[VAL_29:.*]] = arith.cmpi ult, %[[VAL_25]], %[[VAL_19]] : index
 // CHECK:             %[[VAL_30:.*]] = arith.cmpi ult, %[[VAL_26]], %[[VAL_21]] : index
 // CHECK:             %[[VAL_31:.*]] = arith.andi %[[VAL_29]], %[[VAL_30]] : i1
 // CHECK:             %[[VAL_32:.*]] = arith.cmpi ult, %[[VAL_27]], %[[VAL_23]] : index
 // CHECK:             %[[VAL_33:.*]] = arith.andi %[[VAL_31]], %[[VAL_32]] : i1
 // CHECK:             scf.condition(%[[VAL_33]]) %[[VAL_25]], %[[VAL_26]], %[[VAL_27]], %[[VAL_28]] : index, index, index, f64
 // CHECK:           } do {

mlir/test/Dialect/SparseTensor/sparse_2d.mlir

 // CHECK-SAME:      %[[VAL_0:.*]]: tensor<?x?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>>,
 // CHECK-SAME:      %[[VAL_1:.*]]: tensor<?x?xf64>) -> tensor<?x?xf64> {
 // CHECK-DAG:       %[[VAL_2:.*]] = arith.constant 2.000000e+00 : f64
 // CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 1 : index} : tensor<?x?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 1 : index} : tensor<?x?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<?x?xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xf64>
-// CHECK-DAG:       %[[VAL_8:.*]] = tensor.dim %[[VAL_1]], %[[VAL_3]] : tensor<?x?xf64>
+// CHECK-DAG:       %[[VAL_8:.*]] = tensor.dim %[[VAL_0]], %[[VAL_3]] : tensor<?x?xf64, #sparse_tensor.encoding<{{{.*}}}>>
 // CHECK-DAG:       %[[VAL_11:.*]] = bufferization.to_memref %[[VAL_1]] : memref<?x?xf64>
 // CHECK:           linalg.fill ins(%{{.*}} : f64) outs(%[[VAL_11]] : memref<?x?xf64>)
 // CHECK:           scf.for %[[VAL_12:.*]] = %[[VAL_3]] to %[[VAL_8]] step %[[VAL_4]] {
 // CHECK:             %[[VAL_13:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_12]]] : memref<?xindex>
 // CHECK:             %[[VAL_14:.*]] = arith.addi %[[VAL_12]], %[[VAL_4]] : index
 // CHECK:             %[[VAL_15:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_14]]] : memref<?xindex>
 // CHECK:             scf.for %[[VAL_16:.*]] = %[[VAL_13]] to %[[VAL_15]] step %[[VAL_4]] {
 // CHECK:               %[[VAL_17:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_16]]] : memref<?xindex>

 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 1 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 1 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_11:.*]] = bufferization.to_memref %[[VAL_1]] : memref<?x?xf32>
-// CHECK-DAG:       %[[VAL_12:.*]] = tensor.dim %[[VAL_2]], %[[VAL_4]] : tensor<?x?xf32>
+// CHECK-DAG:       %[[VAL_12:.*]] = tensor.dim %[[VAL_1]], %[[VAL_5]] : tensor<?x?xf32>
 // CHECK-DAG:       %[[VAL_13:.*]] = bufferization.to_memref %[[VAL_2]] : memref<?x?xf32>
 // CHECK-DAG:       %[[VAL_17:.*]] = bufferization.to_memref %[[VAL_3]] : memref<?x?xf32>
 // CHECK:           %[[VAL_18:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
 // CHECK:           %[[VAL_19:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
 // CHECK:           scf.for %[[VAL_20:.*]] = %[[VAL_18]] to %[[VAL_19]] step %[[VAL_5]] {
 // CHECK:             %[[VAL_21:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_20]]] : memref<?xindex>
 // CHECK:             %[[VAL_22:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_20]]] : memref<?xindex>
 // CHECK:             %[[VAL_23:.*]] = arith.addi %[[VAL_20]], %[[VAL_5]] : index

 // CHECK-DAG:       %[[VAL_14:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 1 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_15:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 1 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xindex>
 // CHECK:           %[[VAL_16:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_17:.*]] = sparse_tensor.pointers %[[VAL_2]] {dimension = 1 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_18:.*]] = sparse_tensor.indices %[[VAL_2]] {dimension = 1 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_19:.*]] = sparse_tensor.values %[[VAL_2]] : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xf32>
 // CHECK-DAG:       %[[VAL_20:.*]] = bufferization.to_memref %[[VAL_3]] : memref<?xf32>
 // CHECK-DAG:       %[[VAL_21:.*]] = bufferization.to_memref %[[VAL_4]] : memref<f32>
-// CHECK-DAG:       %[[VAL_22:.*]] = tensor.dim %[[VAL_5]], %[[VAL_6]] : tensor<?xf32>
+// CHECK-DAG:       %[[VAL_22:.*]] = tensor.dim %[[VAL_2]], %[[VAL_6]] : tensor<?x?xf32,
 // CHECK-DAG:       %[[VAL_24:.*]] = bufferization.to_memref %[[VAL_5]] : memref<?xf32>
 // CHECK:           %[[VAL_25:.*]] = memref.load %[[VAL_21]][] : memref<f32>
 // CHECK:           %[[VAL_26:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_6]]] : memref<?xindex>
 // CHECK:           %[[VAL_27:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_7]]] : memref<?xindex>
 // CHECK:           %[[VAL_28:.*]]:2 = scf.while (%[[VAL_29:.*]] = %[[VAL_26]], %[[VAL_30:.*]] = %[[VAL_6]]) : (index, index) -> (index, index) {
 // CHECK:             %[[VAL_31:.*]] = arith.cmpi ult, %[[VAL_29]], %[[VAL_27]] : index
 // CHECK:             scf.condition(%[[VAL_31]]) %[[VAL_29]], %[[VAL_30]] : index, index
 // CHECK:           } do {

mlir/test/Dialect/SparseTensor/sparse_3d.mlir

 // CHECK-SAME:      %[[VAL_1:.*1]]: tensor<?x?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "dense", "compressed" ] }>>,
 // CHECK-SAME:      %[[VAL_2:.*2]]: tensor<?x?xf32>,
 // CHECK-SAME:      %[[VAL_3:.*3]]: tensor<?x?xf32>) -> tensor<?x?xf32> {
 // CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_6:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_7:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 2 : index} : tensor<?x?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "dense", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_8:.*]] = sparse_tensor.indices %[[VAL_1]] {dimension = 2 : index} : tensor<?x?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "dense", "compressed" ] }>> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?x?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "dense", "compressed" ] }>> to memref<?xf32>
-// CHECK-DAG:       %[[VAL_10:.*]] = tensor.dim %[[VAL_2]], %[[VAL_5]] : tensor<?x?xf32>
+// CHECK-DAG:       %[[VAL_10:.*]] = tensor.dim %[[VAL_1]], %[[VAL_6]] : tensor<?x?x?xf32, #sparse_tensor.encoding<{{{.*}}}>>
 // CHECK-DAG:       %[[VAL_11:.*]] = bufferization.to_memref %[[VAL_2]] : memref<?x?xf32>
 // CHECK-DAG:       %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_3]] : memref<?x?xf32>
-// CHECK-DAG:       %[[VAL_13:.*]] = tensor.dim %[[VAL_0]], %[[VAL_5]] : tensor<?x?xf32>
-// CHECK-DAG:       %[[VAL_14:.*]] = tensor.dim %[[VAL_0]], %[[VAL_6]] : tensor<?x?xf32>
+// CHECK-DAG:       %[[VAL_13:.*]] = tensor.dim %[[VAL_1]], %[[VAL_5]] : tensor<?x?x?xf32, #sparse_tensor.encoding<{{{.*}}}>>
+// CHECK-DAG:       %[[VAL_14:.*]] = tensor.dim %[[VAL_2]], %[[VAL_6]] : tensor<?x?xf32>
 // CHECK-DAG:       %[[VAL_16:.*]] = bufferization.to_memref %[[VAL_0]] : memref<?x?xf32>
 // CHECK:           scf.for %[[VAL_17:.*]] = %[[VAL_5]] to %[[VAL_13]] step %[[VAL_6]] {
 // CHECK:             scf.for %[[VAL_18:.*]] = %[[VAL_5]] to %[[VAL_10]] step %[[VAL_6]] {
 // CHECK:               %[[VAL_19:.*]] = arith.muli %[[VAL_10]], %[[VAL_17]] : index
 // CHECK:               %[[VAL_20:.*]] = arith.addi %[[VAL_19]], %[[VAL_18]] : index
 // CHECK:               %[[VAL_21:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_20]]] : memref<?xindex>
 // CHECK:               %[[VAL_22:.*]] = arith.addi %[[VAL_20]], %[[VAL_6]] : index
 // CHECK:               %[[VAL_23:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_22]]] : memref<?xindex>

 // CHECK-SAME:      %[[VAL_1:.*]]: tensor<?xf32, #sparse_tensor.encoding<{{{.*}}}>>
 // CHECK-SAME:      %[[VAL_2:.*]]: tensor<f32>) -> tensor<f32> {
 // CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 2 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_6:.*]] = tensor.dim %[[VAL_0]], %[[VAL_3]] : tensor<?x?x?xf32>
 // CHECK-DAG:       %[[VAL_7:.*]] = tensor.dim %[[VAL_0]], %[[VAL_4]] : tensor<?x?x?xf32>
 // CHECK-DAG:       %[[VAL_8:.*]] = bufferization.to_memref %[[VAL_0]] : memref<?x?x?xf32>
-// CHECK-DAG:       %[[VAL_9:.*]] = tensor.dim %[[VAL_1]], %[[VAL_5]] : tensor<?xf32, #sparse_tensor.encoding<{{{.*}}}>>
+// CHECK-DAG:       %[[VAL_9:.*]] = tensor.dim %[[VAL_0]], %[[VAL_5]] : tensor<?x?x?xf32>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?xf32, #sparse_tensor.encoding<{{{.*}}}>>
 // CHECK-DAG:       %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2]] : memref<f32>
 // CHECK:           %[[VAL_13:.*]] = memref.load %[[VAL_12]][] : memref<f32>
 // CHECK:           %[[VAL_14:.*]] = scf.for %[[VAL_15:.*]] = %[[VAL_5]] to %[[VAL_9]] step %[[VAL_3]] iter_args(%[[VAL_16:.*]] = %[[VAL_13]]) -> (f32) {
 // CHECK:             %[[VAL_17:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_15]]] : memref<?xf32>
 // CHECK:             %[[VAL_18:.*]] = scf.for %[[VAL_19:.*]] = %[[VAL_5]] to %[[VAL_6]] step %[[VAL_3]] iter_args(%[[VAL_20:.*]] = %[[VAL_16]]) -> (f32) {
 // CHECK:               %[[VAL_21:.*]] = scf.for %[[VAL_22:.*]] = %[[VAL_5]] to %[[VAL_7]] step %[[VAL_3]] iter_args(%[[VAL_23:.*]] = %[[VAL_20]]) -> (f32) {
 // CHECK:                 %[[VAL_24:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_15]], %[[VAL_19]], %[[VAL_22]]] : memref<?x?x?xf32>

mlir/test/Dialect/SparseTensor/sparse_concat_codegen.mlir

 //       CHECK:  %[[TMP_2:.*]] = sparse_tensor.indices %[[TMP_arg0]] {dimension = 0 : index} : tensor<2x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_3:.*]] = sparse_tensor.pointers %[[TMP_arg0]] {dimension = 1 : index} : tensor<2x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_4:.*]] = sparse_tensor.indices %[[TMP_arg0]] {dimension = 1 : index} : tensor<2x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_5:.*]] = sparse_tensor.values %[[TMP_arg0]] : tensor<2x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_6:.*]] = memref.load %[[TMP_1]][%[[TMP_c0]]] : memref<?xindex>
 //       CHECK:  %[[TMP_7:.*]] = memref.load %[[TMP_1]][%[[TMP_c1]]] : memref<?xindex>
 //       CHECK:  scf.for %[[TMP_arg3:.*]] = %[[TMP_6]] to %[[TMP_7]] step %[[TMP_c1]] {
 //       CHECK:    %[[TMP_23:.*]] = memref.load %[[TMP_2]][%[[TMP_arg3]]] : memref<?xindex>
-//       CHECK:    %[[TMP_24:.*]] = arith.addi %[[TMP_arg3]], %[[TMP_c1]] : index
-//       CHECK:    %[[TMP_25:.*]] = memref.load %[[TMP_3]][%[[TMP_arg3]]] : memref<?xindex>
+//   CHECK-DAG:    %[[TMP_25:.*]] = memref.load %[[TMP_3]][%[[TMP_arg3]]] : memref<?xindex>
+//   CHECK-DAG:    %[[TMP_24:.*]] = arith.addi %[[TMP_arg3]], %[[TMP_c1]] : index
 //       CHECK:    %[[TMP_26:.*]] = memref.load %[[TMP_3]][%[[TMP_24]]] : memref<?xindex>
 //       CHECK:    scf.for %[[TMP_arg4:.*]] = %[[TMP_25]] to %[[TMP_26]] step %[[TMP_c1]] {
 //       CHECK:      %[[TMP_27:.*]] = memref.load %[[TMP_4]][%[[TMP_arg4]]] : memref<?xindex>
 //       CHECK:      %[[TMP_28:.*]] = memref.load %[[TMP_5]][%[[TMP_arg4]]] : memref<?xf64>
 //       CHECK:      sparse_tensor.insert %[[TMP_28]] into %[[TMP_0]][%[[TMP_23]], %[[TMP_27]]] : tensor<9x4xf64, #sparse_tensor
 //       CHECK:    }
 //       CHECK:  }
 //       CHECK:  %[[TMP_8:.*]] = sparse_tensor.pointers %[[TMP_arg1]] {dimension = 0 : index} : tensor<3x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_9:.*]] = sparse_tensor.indices %[[TMP_arg1]] {dimension = 0 : index} : tensor<3x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_10:.*]] = sparse_tensor.pointers %[[TMP_arg1]] {dimension = 1 : index} : tensor<3x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_11:.*]] = sparse_tensor.indices %[[TMP_arg1]] {dimension = 1 : index} : tensor<3x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_12:.*]] = sparse_tensor.values %[[TMP_arg1]] : tensor<3x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_13:.*]] = memref.load %[[TMP_8]][%[[TMP_c0]]] : memref<?xindex>
 //       CHECK:  %[[TMP_14:.*]] = memref.load %[[TMP_8]][%[[TMP_c1]]] : memref<?xindex>
 //       CHECK:  scf.for %[[TMP_arg3:.*]] = %[[TMP_13]] to %[[TMP_14]] step %[[TMP_c1]] {
 //       CHECK:    %[[TMP_23:.*]] = memref.load %[[TMP_9]][%[[TMP_arg3]]] : memref<?xindex>
-//       CHECK:    %[[TMP_24:.*]] = arith.addi %[[TMP_arg3]], %[[TMP_c1]] : index
-//       CHECK:    %[[TMP_25:.*]] = memref.load %[[TMP_10]][%[[TMP_arg3]]] : memref<?xindex>
+//   CHECK-DAG:    %[[TMP_25:.*]] = memref.load %[[TMP_10]][%[[TMP_arg3]]] : memref<?xindex>
+//   CHECK-DAG:    %[[TMP_24:.*]] = arith.addi %[[TMP_arg3]], %[[TMP_c1]] : index
 //       CHECK:    %[[TMP_26:.*]] = memref.load %[[TMP_10]][%[[TMP_24]]] : memref<?xindex>
 //       CHECK:    scf.for %[[TMP_arg4:.*]] = %[[TMP_25]] to %[[TMP_26]] step %[[TMP_c1]] {
 //       CHECK:      %[[TMP_27:.*]] = memref.load %[[TMP_11]][%[[TMP_arg4]]] : memref<?xindex>
 //       CHECK:      %[[TMP_28:.*]] = memref.load %[[TMP_12]][%[[TMP_arg4]]] : memref<?xf64>
 //       CHECK:      %[[TMP_29:.*]] = arith.addi %[[TMP_23]], %[[TMP_c2]] : index
 //       CHECK:      sparse_tensor.insert %[[TMP_28]] into %[[TMP_0]][%[[TMP_29]], %[[TMP_27]]] : tensor<9x4xf64, #sparse_tensor
 //       CHECK:    }
 //       CHECK:  }
 //       CHECK:  %[[TMP_15:.*]] = sparse_tensor.pointers %[[TMP_arg2]] {dimension = 0 : index} : tensor<4x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_16:.*]] = sparse_tensor.indices %[[TMP_arg2]] {dimension = 0 : index} : tensor<4x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_17:.*]] = sparse_tensor.pointers %[[TMP_arg2]] {dimension = 1 : index} : tensor<4x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_18:.*]] = sparse_tensor.indices %[[TMP_arg2]] {dimension = 1 : index} : tensor<4x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_19:.*]] = sparse_tensor.values %[[TMP_arg2]] : tensor<4x4xf64, #sparse_tensor
 //       CHECK:  %[[TMP_20:.*]] = memref.load %[[TMP_15]][%[[TMP_c0]]] : memref<?xindex>
 //       CHECK:  %[[TMP_21:.*]] = memref.load %[[TMP_15]][%[[TMP_c1]]] : memref<?xindex>
 //       CHECK:  scf.for %[[TMP_arg3:.*]] = %[[TMP_20]] to %[[TMP_21]] step %[[TMP_c1]] {
 //       CHECK:    %[[TMP_23:.*]] = memref.load %[[TMP_16]][%[[TMP_arg3]]] : memref<?xindex>
-//       CHECK:    %[[TMP_24:.*]] = arith.addi %[[TMP_arg3]], %[[TMP_c1]] : index
 //       CHECK:    %[[TMP_25:.*]] = memref.load %[[TMP_17]][%[[TMP_arg3]]] : memref<?xindex>
+//       CHECK:    %[[TMP_24:.*]] = arith.addi %[[TMP_arg3]], %[[TMP_c1]] : index
 //       CHECK:    %[[TMP_26:.*]] = memref.load %[[TMP_17]][%[[TMP_24]]] : memref<?xindex>
 //       CHECK:    scf.for %[[TMP_arg4:.*]] = %[[TMP_25]] to %[[TMP_26]] step %[[TMP_c1]] {
 //       CHECK:      %[[TMP_27:.*]] = memref.load %[[TMP_18]][%[[TMP_arg4]]] : memref<?xindex>
 //       CHECK:      %[[TMP_28:.*]] = memref.load %[[TMP_19]][%[[TMP_arg4]]] : memref<?xf64>
 //       CHECK:      %[[TMP_29:.*]] = arith.addi %[[TMP_23]], %[[TMP_c5]] : index
 //       CHECK:      sparse_tensor.insert %[[TMP_28]] into %[[TMP_0]][%[[TMP_29]], %[[TMP_27]]] : tensor<9x4xf64, #sparse_tensor
 //       CHECK:    }
 //       CHECK:  }

mlir/test/Dialect/SparseTensor/sparse_index.mlir

 // CHECK-LABEL:   func.func @dense_index(
 // CHECK-SAME:      %[[VAL_0:.*]]: tensor<?x?xi64, #sparse_tensor.encoding
 // CHECK-DAG:       %[[VAL_1:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_2:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_3:.*]] = tensor.dim %[[VAL_0]], %[[VAL_1]] : tensor<?x?xi64, #sparse_tensor.encoding
 // CHECK-DAG:       %[[VAL_4:.*]] = tensor.dim %[[VAL_0]], %[[VAL_1]] : tensor<?x?xi64, #sparse_tensor.encoding
 // CHECK-DAG:       %[[VAL_5:.*]] = bufferization.alloc_tensor(%[[VAL_3]], %[[VAL_4]]) : tensor<?x?xi64, #sparse_tensor.encoding
 // CHECK-DAG:       %[[VAL_6:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<?x?xi64, #sparse_tensor.encoding
-// CHECK-DAG:       %[[VAL_7:.*]] = tensor.dim %[[VAL_5]], %[[VAL_1]] : tensor<?x?xi64, #sparse_tensor.encoding
-// CHECK-DAG:       %[[VAL_8:.*]] = tensor.dim %[[VAL_5]], %[[VAL_2]] : tensor<?x?xi64, #sparse_tensor.encoding
+// CHECK-DAG:       %[[VAL_7:.*]] = tensor.dim %[[VAL_0]], %[[VAL_1]] : tensor<?x?xi64, #sparse_tensor.encoding
+// CHECK-DAG:       %[[VAL_8:.*]] = tensor.dim %[[VAL_0]], %[[VAL_2]] : tensor<?x?xi64, #sparse_tensor.encoding
+// CHECK-DAG:       %[[VAL_24:.*]] = tensor.dim %[[VAL_5]], %[[VAL_2]] : tensor<?x?xi64, #sparse_tensor.encoding
 // CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.values %[[VAL_5]] : tensor<?x?xi64, #sparse_tensor.encoding
 // CHECK:           scf.for %[[VAL_10:.*]] = %[[VAL_1]] to %[[VAL_7]] step %[[VAL_2]] {
 // CHECK:             scf.for %[[VAL_11:.*]] = %[[VAL_1]] to %[[VAL_8]] step %[[VAL_2]] {
 // CHECK:               %[[VAL_12:.*]] = arith.muli %[[VAL_8]], %[[VAL_10]] : index
 // CHECK:               %[[VAL_13:.*]] = arith.addi %[[VAL_12]], %[[VAL_11]] : index
-// CHECK:               %[[VAL_14:.*]] = arith.muli %[[VAL_8]], %[[VAL_10]] : index
+// CHECK:               %[[VAL_14:.*]] = arith.muli %[[VAL_24]], %[[VAL_10]] : index
 // CHECK:               %[[VAL_15:.*]] = arith.addi %[[VAL_14]], %[[VAL_11]] : index
 // CHECK:               %[[VAL_16:.*]] = arith.index_cast %[[VAL_11]] : index to i64
 // CHECK:               %[[VAL_17:.*]] = arith.index_cast %[[VAL_10]] : index to i64
 // CHECK:               %[[VAL_18:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_13]]] : memref<?xi64>
 // CHECK:               %[[VAL_19:.*]] = arith.muli %[[VAL_17]], %[[VAL_18]] : i64
 // CHECK:               %[[VAL_20:.*]] = arith.muli %[[VAL_16]], %[[VAL_19]] : i64
 // CHECK:               memref.store %[[VAL_20]], %[[VAL_9]]{{\[}}%[[VAL_15]]] : memref<?xi64>
 // CHECK:             }

mlir/test/Dialect/SparseTensor/sparse_lower_col.mlir

 // CHECK-HIR-DAG:       %[[VAL_4:.*]] = arith.constant 0 : index
 // CHECK-HIR-DAG:       %[[VAL_5:.*]] = arith.constant 1 : index
 // CHECK-HIR-DAG:       %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 1 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ], dimOrdering = affine_map<(d0, d1) -> (d1, d0)> }>> to memref<?xindex>
 // CHECK-HIR-DAG:       %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 1 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ], dimOrdering = affine_map<(d0, d1) -> (d1, d0)> }>> to memref<?xindex>
 // CHECK-HIR-DAG:       %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ], dimOrdering = affine_map<(d0, d1) -> (d1, d0)> }>> to memref<?xf64>
 // CHECK-HIR-DAG:       %[[VAL_9:.*]] = bufferization.to_memref %[[VAL_1]] : memref<64xf64>
 // CHECK-HIR-DAG:       %[[VAL_11:.*]] = bufferization.to_memref %[[VAL_2]] : memref<32xf64>
 // CHECK-HIR:           scf.for %[[VAL_12:.*]] = %[[VAL_4]] to %[[VAL_3]] step %[[VAL_5]] {
-// CHECK-HIR:             %[[VAL_13:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_12]]] : memref<64xf64>
-// CHECK-HIR:             %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_12]]] : memref<?xindex>
-// CHECK-HIR:             %[[VAL_15:.*]] = arith.addi %[[VAL_12]], %[[VAL_5]] : index
-// CHECK-HIR:             %[[VAL_16:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_15]]] : memref<?xindex>
+// CHECK-HIR-DAG:         %[[VAL_13:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_12]]] : memref<64xf64>
+// CHECK-HIR-DAG:         %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_12]]] : memref<?xindex>
+// CHECK-HIR-DAG:         %[[VAL_15:.*]] = arith.addi %[[VAL_12]], %[[VAL_5]] : index
+// CHECK-HIR-DAG:         %[[VAL_16:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_15]]] : memref<?xindex>
 // CHECK-HIR:             scf.for %[[VAL_17:.*]] = %[[VAL_14]] to %[[VAL_16]] step %[[VAL_5]] {
-// CHECK-HIR:               %[[VAL_18:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_17]]] : memref<?xindex>
-// CHECK-HIR:               %[[VAL_19:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_18]]] : memref<32xf64>
-// CHECK-HIR:               %[[VAL_20:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_17]]] : memref<?xf64>
+// CHECK-HIR-DAG:           %[[VAL_18:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_17]]] : memref<?xindex>
+// CHECK-HIR-DAG:           %[[VAL_19:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_18]]] : memref<32xf64>
+// CHECK-HIR-DAG:           %[[VAL_20:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_17]]] : memref<?xf64>
 // CHECK-HIR:               %[[VAL_21:.*]] = arith.mulf %[[VAL_20]], %[[VAL_13]] : f64
 // CHECK-HIR:               %[[VAL_22:.*]] = arith.addf %[[VAL_19]], %[[VAL_21]] : f64
 // CHECK-HIR:               memref.store %[[VAL_22]], %[[VAL_11]]{{\[}}%[[VAL_18]]] : memref<32xf64>
 // CHECK-HIR:             }
 // CHECK-HIR:           }
 // CHECK-HIR:           %[[VAL_23:.*]] = bufferization.to_tensor %[[VAL_11]] : memref<32xf64>
 // CHECK-HIR:           return %[[VAL_23]] : tensor<32xf64>
 // CHECK-HIR:         }

mlir/test/Dialect/SparseTensor/sparse_perm.mlir

 // CHECK-LABEL:   func @sparse_dynamic_dims(
 // CHECK-SAME:                          %[[VAL_0:.*]]: tensor<?x?x?xf32, #sparse_tensor.encoding<{{{.*}}}>>,
 // CHECK-SAME:                          %[[VAL_1:.*]]: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> {
 // CHECK-DAG:       %[[ZERO:.*]] = arith.constant 0.000000e+00 : f32
 // CHECK-DAG:       %[[VAL_2:.*]] = arith.constant 2 : index
 // CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[VAL_5:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<?x?x?xf32, #sparse_tensor.encoding<{{{.*}}}>>
-// CHECK-DAG:       %[[VAL_6:.*]] = tensor.dim %[[VAL_1]], %[[VAL_3]] : tensor<?x?x?xf32>
-// CHECK-DAG:       %[[VAL_7:.*]] = tensor.dim %[[VAL_1]], %[[VAL_4]] : tensor<?x?x?xf32>
-// CHECK-DAG:       %[[VAL_8:.*]] = tensor.dim %[[VAL_1]], %[[VAL_2]] : tensor<?x?x?xf32>
+// CHECK-DAG:       %[[VAL_6:.*]] = tensor.dim %[[VAL_0]], %[[VAL_2]] : tensor<?x?x?xf32, #sparse_tensor.encoding<{{{.*}}}>>
+// CHECK-DAG:       %[[VAL_7:.*]] = tensor.dim %[[VAL_0]], %[[VAL_3]] : tensor<?x?x?xf32, #sparse_tensor.encoding<{{{.*}}}>>
+// CHECK-DAG:       %[[VAL_8:.*]] = tensor.dim %[[VAL_0]], %[[VAL_4]] : tensor<?x?x?xf32, #sparse_tensor.encoding<{{{.*}}}>>
 // CHECK-DAG:       %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_1]] : memref<?x?x?xf32>
 // CHECK:           linalg.fill ins(%[[ZERO]] : f32) outs(%[[VAL_10]] : memref<?x?x?xf32>)
-// CHECK:           scf.for %[[VAL_11:.*]] = %[[VAL_3]] to %[[VAL_7]] step %[[VAL_4]] {
-// CHECK:             scf.for %[[VAL_12:.*]] = %[[VAL_3]] to %[[VAL_8]] step %[[VAL_4]] {
-// CHECK:               %[[VAL_13:.*]] = arith.muli %[[VAL_8]], %[[VAL_11]] : index
+// CHECK:           scf.for %[[VAL_11:.*]] = %[[VAL_3]] to %[[VAL_6]] step %[[VAL_4]] {
+// CHECK:             scf.for %[[VAL_12:.*]] = %[[VAL_3]] to %[[VAL_7]] step %[[VAL_4]] {
+// CHECK:               %[[VAL_13:.*]] = arith.muli %[[VAL_7]], %[[VAL_11]] : index
 // CHECK:               %[[VAL_14:.*]] = arith.addi %[[VAL_13]], %[[VAL_12]] : index
-// CHECK:               scf.for %[[VAL_15:.*]] = %[[VAL_3]] to %[[VAL_6]] step %[[VAL_4]] {
-// CHECK:                 %[[VAL_16:.*]] = arith.muli %[[VAL_6]], %[[VAL_14]] : index
+// CHECK:               scf.for %[[VAL_15:.*]] = %[[VAL_3]] to %[[VAL_8]] step %[[VAL_4]] {
+// CHECK:                 %[[VAL_16:.*]] = arith.muli %[[VAL_8]], %[[VAL_14]] : index
 // CHECK:                 %[[VAL_17:.*]] = arith.addi %[[VAL_16]], %[[VAL_15]] : index
 // CHECK:                 %[[VAL_18:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_17]]] : memref<?xf32>
 // CHECK:                 memref.store %[[VAL_18]], %[[VAL_10]]{{\[}}%[[VAL_15]], %[[VAL_11]], %[[VAL_12]]] : memref<?x?x?xf32>
 // CHECK:               }
 // CHECK:             }
 // CHECK:           }
 // CHECK:           %[[VAL_19:.*]] = bufferization.to_tensor %[[VAL_10]] : memref<?x?x?xf32>
 // CHECK:           return %[[VAL_19]] : tensor<?x?x?xf32>

mlir/test/Dialect/SparseTensor/sparse_reshape.mlir

 // CHECK-RWT:         %[[I0:.*]] = sparse_tensor.indices %[[S]] {dimension = 0 : index}
 // CHECK-RWT:          %[[P1:.*]] = sparse_tensor.pointers %[[S]] {dimension = 1 : index}
 // CHECK-RWT:          %[[I1:.*]] = sparse_tensor.indices %[[S]] {dimension = 1 : index}
 // CHECK-RWT:          %[[V:.*]] = sparse_tensor.values %[[S]]
 // CHECK-RWT:          %[[S0:.*]] = memref.load %[[P0]]{{\[}}%[[C0]]] : memref<?xindex>
 // CHECK-RWT:          %[[E0:.*]] = memref.load %[[P0]]{{\[}}%[[C1]]] : memref<?xindex>
 // CHECK-RWT:          scf.for %[[I:.*]] = %[[S0]] to %[[E0]] step %[[C1]] {
 // CHECK-RWT:            %[[SI0:.*]] = memref.load %[[I0]]{{\[}}%[[I]]] : memref<?xindex>
-// CHECK-RWT:            %[[PE1:.*]] = arith.addi %[[I]], %[[C1]] : index
-// CHECK-RWT:            %[[S1:.*]] = memref.load %[[P1]]{{\[}}%[[I]]] : memref<?xindex>
+// CHECK-RWT-DAG:        %[[S1:.*]] = memref.load %[[P1]]{{\[}}%[[I]]] : memref<?xindex>
+// CHECK-RWT-DAG:        %[[PE1:.*]] = arith.addi %[[I]], %[[C1]] : index
 // CHECK-RWT:            %[[E1:.*]] = memref.load %[[P1]]{{\[}}%[[PE1]]] : memref<?xindex>
 // CHECK-RWT:            scf.for %[[J:.*]] = %[[S1]] to %[[E1]] step %[[C1]] {
 // CHECK-RWT:              %[[SI1:.*]] = memref.load %[[I1]]{{\[}}%[[J]]] : memref<?xindex>
 // CHECK-RWT:              %[[SV:.*]] = memref.load %[[V]]{{\[}}%[[J]]] : memref<?xf64>
 // CHECK-RWT:              %[[T:.*]] = arith.muli %[[SI0]], %[[C10]] : index
 // CHECK-RWT:              %[[DI:.*]] = arith.addi %[[T]], %[[SI1]] : index
 // CHECK-RWT:              sparse_tensor.insert %[[SV]] into %[[B]]{{\[}}%[[DI]]]
 // CHECK-RWT             }

 // CHECK-RWT:         %[[I0:.*]] = sparse_tensor.indices %[[S]] {dimension = 0 : index}
 // CHECK-RWT:          %[[P1:.*]] = sparse_tensor.pointers %[[S]] {dimension = 1 : index}
 // CHECK-RWT:          %[[I1:.*]] = sparse_tensor.indices %[[S]] {dimension = 1 : index}
 // CHECK-RWT:          %[[V:.*]] = sparse_tensor.values %[[S]]
 // CHECK-RWT:          %[[S0:.*]] = memref.load %[[P0]]{{\[}}%[[C0]]] : memref<?xindex>
 // CHECK-RWT:          %[[E0:.*]] = memref.load %[[P0]]{{\[}}%[[C1]]] : memref<?xindex>
 // CHECK-RWT:          scf.for %[[I:.*]] = %[[S0]] to %[[E0]] step %[[C1]] {
 // CHECK-RWT:            %[[SI0:.*]] = memref.load %[[I0]]{{\[}}%[[I]]] : memref<?xindex>
-// CHECK-RWT:            %[[PE1:.*]] = arith.addi %[[I]], %[[C1]] : index
-// CHECK-RWT:            %[[S1:.*]] = memref.load %[[P1]]{{\[}}%[[I]]] : memref<?xindex>
+// CHECK-RWT-DAG:        %[[S1:.*]] = memref.load %[[P1]]{{\[}}%[[I]]] : memref<?xindex>
+// CHECK-RWT-DAG:        %[[PE1:.*]] = arith.addi %[[I]], %[[C1]] : index
 // CHECK-RWT:            %[[E1:.*]] = memref.load %[[P1]]{{\[}}%[[PE1]]] : memref<?xindex>
 // CHECK-RWT:            scf.for %[[J:.*]] = %[[S1]] to %[[E1]] step %[[C1]] {
 // CHECK-RWT:              %[[SI1:.*]] = memref.load %[[I1]]{{\[}}%[[J]]] : memref<?xindex>
 // CHECK-RWT:              %[[SV:.*]] = memref.load %[[V]]{{\[}}%[[J]]] : memref<?xf64>
 // CHECK-RWT:               %[[T1:.*]] = arith.divui %[[DD0]], %[[C10]] : index
 // CHECK-RWT:               %[[T2:.*]] = arith.muli %[[SI0]], %[[T1]] : index
 // CHECK-RWT:               %[[T3:.*]] = arith.divui %[[T1]], %[[SD1]] : index
 // CHECK-RWT:               %[[T4:.*]] = arith.muli %[[SI1]], %[[T3]] : index

mlir/test/Dialect/SparseTensor/sparse_scalars.mlir

 // CHECK-SAME:              %[[VAL_0:.*0]]: tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>>,
 // CHECK-SAME:              %[[VAL_1:.*1]]: tensor<f32>,
 // CHECK-SAME:              %[[VAL_2:.*2]]: f32,
 // CHECK-SAME:              %[[VAL_3:.*3]]: f32,
 // CHECK-SAME:              %[[VAL_4:.*4]]: tensor<32x16xf32>) -> tensor<32x16xf32> {
 // CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 2.200000e+00 : f32
 // CHECK-DAG:       %[[VAL_6:.*]] = arith.constant 0 : index
 // CHECK-DAG:       %[[VAL_7:.*]] = arith.constant 1 : index
-// CHECK:           %[[VAL_8:.*]] = arith.addf %[[VAL_2]], %[[VAL_3]] : f32
-// CHECK:           %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xindex>
-// CHECK:           %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xindex>
-// CHECK:           %[[VAL_11:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 1 : index} : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xindex>
-// CHECK:           %[[VAL_12:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 1 : index} : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xindex>
-// CHECK:           %[[VAL_13:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xf32>
-// CHECK:           %[[VAL_14:.*]] = bufferization.to_memref %[[VAL_1]] : memref<f32>
-// CHECK:           %[[VAL_15:.*]] = bufferization.to_memref %[[VAL_4]] : memref<32x16xf32>
-// CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_14]][] : memref<f32>
-// CHECK:           %[[VAL_17:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_6]]] : memref<?xindex>
-// CHECK:           %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_7]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_8:.*]] = arith.addf %[[VAL_2]], %[[VAL_3]] : f32
+// CHECK-DAG:       %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 0 : index} : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xindex>
+// CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 0 : index} : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xindex>
+// CHECK-DAG:       %[[VAL_11:.*]] = sparse_tensor.pointers %[[VAL_0]] {dimension = 1 : index} : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xindex>
+// CHECK-DAG:       %[[VAL_12:.*]] = sparse_tensor.indices %[[VAL_0]] {dimension = 1 : index} : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xindex>
+// CHECK-DAG:       %[[VAL_13:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xf32>
+// CHECK-DAG:       %[[VAL_14:.*]] = bufferization.to_memref %[[VAL_1]] : memref<f32>
+// CHECK-DAG:       %[[VAL_15:.*]] = bufferization.to_memref %[[VAL_4]] : memref<32x16xf32>
+// CHECK-DAG:       %[[VAL_16:.*]] = memref.load %[[VAL_14]][] : memref<f32>
+// CHECK-DAG:       %[[VAL_17:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_6]]] : memref<?xindex>
+// CHECK-DAG:       %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_7]]] : memref<?xindex>
 // CHECK:           scf.for %[[VAL_19:.*]] = %[[VAL_17]] to %[[VAL_18]] step %[[VAL_7]] {
 // CHECK:             %[[VAL_20:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_19]]] : memref<?xindex>
 // CHECK:             %[[VAL_21:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_19]]] : memref<?xindex>
 // CHECK:             %[[VAL_22:.*]] = arith.addi %[[VAL_19]], %[[VAL_7]] : index
 // CHECK:             %[[VAL_23:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_22]]] : memref<?xindex>
 // CHECK:             scf.for %[[VAL_24:.*]] = %[[VAL_21]] to %[[VAL_23]] step %[[VAL_7]] {
 // CHECK:               %[[VAL_25:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_24]]] : memref<?xindex>
 // CHECK:               %[[VAL_26:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_24]]] : memref<?xf32>