[mlir:DialectConversion] Restructure how argument/target materializations get invoked

[rriddle]

Summary

The current implementation invokes materializations
whenever an input operand does not have a mapping for the
desired type, i.e. it requires materialization at the earliest possible
point. This conflicts with goal of dialect conversion (and also the
current documentation) which states that a materialization is only
required if the materialization is supposed to persist after the
conversion process has finished.

This revision refactors this such that whenever a target
materialization "might" be necessary, we insert an
unrealized_conversion_cast to act as a temporary materialization.
This allows for deferring the invocation of the user
materialization hooks until the end of the conversion process,
where we actually have a better sense if it's actually
necessary. This has several benefits:

• In some cases a target materialization hook is no longer necessary

When performing a full conversion, there are some situations
where a temporary materialization is necessary. Moving forward,
these users won't need to provide any target materializations,
as the temporary materializations do not require the user to
provide materialization hooks.

• getRemappedValue can now handle values that haven't been converted yet

Before this commit, it wasn't well supported to get the remapped
value of a value that hadn't been converted yet (making it
difficult/impossible to convert multiple operations in many
situations). This commit updates getRemappedValue to properly
handle this case by inserting temporary materializations when
necessary.

Another code-health related benefit is that with this change we
can move a majority of the complexity related to materializations
to the end of the conversion process, instead of handling adhoc
while conversion is happening.

[ftynse]

LGTM!

It would be nice to have tests for cases where unrealized_conversion_cast is already present in the input IR.

mlir/include/mlir/IR/BlockAndValueMapping.h

93–94

Would you mind commenting why preserving the order of insertion is important here?

mlir/lib/Conversion/SPIRVToLLVM/SPIRVToLLVM.cpp

313

Nice catch! I was thinking about some automated way to catch op/adaptor misuses, and it looks like this change is making them break early.

mlir/lib/Transforms/Utils/DialectConversion.cpp

370

Nit: "illegal type" -> "op result of illegal type". Argument materialization is also technically a materialization of an illegal type.

453

Nit: I'm slightly concerned about the asymmetry between buildUnresolvedArgumentMaterialization using the type converter to obtain the output type and buildUnresolvedTargetMaterialization just taking it as argument (which is computed by its only caller using the type converter).

738

Nit: drop "we".

1060

How about adding some debug output in this case so we can actually see which patterns rely on this?

1273–1274

Nit: it's not necessarily an operand anymore

2564

Nit: sink the declaration into the loop?

2574

This deserves a comment as to why we are erasing there (IIUC, the mapping has become circular, which means we should be erasing it completely).

2606

Nit: s/affect/effect

mlir/test/Transforms/test-legalize-remapped-value.mlir

29

Nit: add the newline plz

[rriddle]

Address comments

[rriddle]

Thanks for the review Alex!

mlir/include/mlir/IR/BlockAndValueMapping.h

93–94

Nice catch! This was necessary in a previous version of the patch, but the order isn't important anymore.

mlir/lib/Transforms/Utils/DialectConversion.cpp

370

Target materializations can also be applied to arguments though. Argument materializations shouldn't be used to materialize illegal types AFAIK, they should only be used to materialize legal types. That being said, the way the hook is written today does not make that clear (given that you pass in the illegal type as the expected result type, even though all of our cases today just use that to decide which legal type to actually produce).

I think our current documentation for these are fairly confusing, or just lacking in general. The way they work right now is that Argument materialization is an optional hook that effectively enables 2-stage conversion/materialization for block arguments; whereas Target materialization is the general/default "convert this illegal type into a legal one".

1060

Nice idea! Added some logging when we detect this when legalizing changed results.

[rriddle]

The broken flang test should be fixed by rG42831686034b.

[silvas]

Glad to see this happening!

[benvanik]

This change broke 1->N type conversion (as noted in a FIXME in here). This is really unfortunate and there doesn't seem to be any workaround now that the materialization is deferred and the casts are inserted in a way that doesn't allow for patterns to match them. Is this something that can be rolled back or fixed soon? 1->N type conversion during dialect conversion did work before this change so this is a regression.

[rriddle]

In D111620#3104485, @benvanik wrote:

This change broke 1->N type conversion (as noted in a FIXME in here). This is really unfortunate and there doesn't seem to be any workaround now that the materialization is deferred and the casts are inserted in a way that doesn't allow for patterns to match them. Is this something that can be rolled back or fixed soon? 1->N type conversion during dialect conversion did work before this change so this is a regression.

The FIXME is more of further elaborating on the existing state, it was not some new problem added by this patch. Can you file a bug report with what your exact situation is?

[rriddle]

In D111620#3104504, @rriddle wrote:

In D111620#3104485, @benvanik wrote:

This change broke 1->N type conversion (as noted in a FIXME in here). This is really unfortunate and there doesn't seem to be any workaround now that the materialization is deferred and the casts are inserted in a way that doesn't allow for patterns to match them. Is this something that can be rolled back or fixed soon? 1->N type conversion during dialect conversion did work before this change so this is a regression.

The FIXME is more of further elaborating on the existing state, it was not some new problem added by this patch. Can you file a bug report with what your exact situation is?

One change that was made is that inserted materializations aren't legalized anymore, but that was mostly prevent situations where that legalization doesn't mesh well with the fact that the conversion is being finalized (we would have to recursively finalize in some capacity). If that behavior is necessary, it could potentially be re-added in some limited capacity. The way it was before was extremely fragile though.

[silvas]

In D111620#3104504, @rriddle wrote:

In D111620#3104485, @benvanik wrote:

This change broke 1->N type conversion (as noted in a FIXME in here). This is really unfortunate and there doesn't seem to be any workaround now that the materialization is deferred and the casts are inserted in a way that doesn't allow for patterns to match them. Is this something that can be rolled back or fixed soon? 1->N type conversion during dialect conversion did work before this change so this is a regression.

The FIXME is more of further elaborating on the existing state, it was not some new problem added by this patch. Can you file a bug report with what your exact situation is?

I suspect it is that we can't pattern-match the unrealized_conversion_cast ops the way we used to be able to (when they were our materializations). Previously it was possible to use the inserted cast as a marker indicating a "bundle" of operands in 1->N conversion. We now don't have a way to represent that "bundle" concept.

[rriddle]

In D111620#3104649, @silvas wrote:

In D111620#3104504, @rriddle wrote:

In D111620#3104485, @benvanik wrote:

This change broke 1->N type conversion (as noted in a FIXME in here). This is really unfortunate and there doesn't seem to be any workaround now that the materialization is deferred and the casts are inserted in a way that doesn't allow for patterns to match them. Is this something that can be rolled back or fixed soon? 1->N type conversion during dialect conversion did work before this change so this is a regression.

The FIXME is more of further elaborating on the existing state, it was not some new problem added by this patch. Can you file a bug report with what your exact situation is?

I suspect it is that we can't pattern-match the unrealized_conversion_cast ops the way we used to be able to (when they were our materializations). Previously it was possible to use the inserted cast as a marker indicating a "bundle" of operands in 1->N conversion. We now don't have a way to represent that "bundle" concept.

Right, 1->N conversions are not something I would consider "well supported" and never have been. Part of the point of this patch is to pave a way where that isn't the case. I've got a CL in my stack that adds "builtin" 1->N conversions (i.e. replaceOp "just works", utilities for expanding the temporary conversions, etc.). That ideally should remove the need to explicitly match any operation when expanding, given that dialect conversion knows exactly which operations are the temporary "bundles" and which aren't (given that dialect conversion inserted them in the first place). User operations should only need to be created via the existing materialization hooks and only when necessary (i.e. uses remain at the end of conversion).

mlir/include/mlir/Conversion/SPIRVToLLVM/SPIRVToLLVM.h

 class MLIRContext;
 class ModuleOp;
 
 template <typename SPIRVOp>
 class SPIRVToLLVMConversion : public OpConversionPattern<SPIRVOp> {
 public:
   SPIRVToLLVMConversion(MLIRContext *context, LLVMTypeConverter &typeConverter,
                         PatternBenefit benefit = 1)
-      : OpConversionPattern<SPIRVOp>(context, benefit),
+      : OpConversionPattern<SPIRVOp>(typeConverter, context, benefit),
         typeConverter(typeConverter) {}
 
 protected:
   LLVMTypeConverter &typeConverter;
 };
 
 /// Encodes global variable's descriptor set and binding into its name if they
 /// both exist.

mlir/include/mlir/IR/BlockAndValueMapping.h

 // found via the 'lookup*' functions. There are two variants that differ only in
 // return value when an existing is not found for the provided key.
 // 'lookupOrNull' returns nullptr where as 'lookupOrDefault' will return the
 // lookup key.
 class BlockAndValueMapping {
 public:
   /// Inserts a new mapping for 'from' to 'to'. If there is an existing mapping,
   /// it is overwritten.
-  void map(Block *from, Block *to) { valueMap[from] = to; }
-  void map(Value from, Value to) {
-    valueMap[from.getAsOpaquePointer()] = to.getAsOpaquePointer();
-  }
+  void map(Block *from, Block *to) { blockMap[from] = to; }
+  void map(Value from, Value to) { valueMap[from] = to; }
 
   template <
       typename S, typename T,
       std::enable_if_t<!std::is_assignable<Value, S>::value &&
                        !std::is_assignable<Block *, S>::value> * = nullptr>
   void map(S &&from, T &&to) {
     for (auto pair : llvm::zip(from, to))
       map(std::get<0>(pair), std::get<1>(pair));
   }
 
   /// Erases a mapping for 'from'.
-  void erase(Block *from) { valueMap.erase(from); }
-  void erase(Value from) { valueMap.erase(from.getAsOpaquePointer()); }
+  void erase(Block *from) { blockMap.erase(from); }
+  void erase(Value from) { valueMap.erase(from); }
 
   /// Checks to see if a mapping for 'from' exists.
-  bool contains(Block *from) const { return valueMap.count(from); }
-  bool contains(Value from) const {
-    return valueMap.count(from.getAsOpaquePointer());
-  }
+  bool contains(Block *from) const { return blockMap.count(from); }
+  bool contains(Value from) const { return valueMap.count(from); }
 
   /// Lookup a mapped value within the map. If a mapping for the provided value
   /// does not exist then return nullptr.
   Block *lookupOrNull(Block *from) const {
     return lookupOrValue(from, (Block *)nullptr);
   }
   Value lookupOrNull(Value from) const { return lookupOrValue(from, Value()); }
 
template <typename T> T lookup(T from) const {
     auto result = lookupOrNull(from);
     assert(result && "expected 'from' to be contained within the map");
     return result;
   }
 
   /// Clears all mappings held by the mapper.
   void clear() { valueMap.clear(); }
 
-  /// Returns a new mapper containing the inverse mapping.
-  BlockAndValueMapping getInverse() const {
-    BlockAndValueMapping result;
-    for (const auto &pair : valueMap)
-      result.valueMap.try_emplace(pair.second, pair.first);
-    return result;
-  }
+  /// Return the held value mapping.
+  const DenseMap<Value, Value> &getValueMap() const { return valueMap; }
+
+  /// Return the held block mapping.
+  const DenseMap<Block *, Block *> &getBlockMap() const { return blockMap; }
 
 private:
   /// Utility lookupOrValue that looks up an existing key or returns the
   /// provided value.
   Block *lookupOrValue(Block *from, Block *value) const {
-    auto it = valueMap.find(from);
-    return it != valueMap.end() ? reinterpret_cast<Block *>(it->second) : value;
+    auto it = blockMap.find(from);
+    return it != blockMap.end() ? it->second : value;
   }
   Value lookupOrValue(Value from, Value value) const {
-    auto it = valueMap.find(from.getAsOpaquePointer());
-    return it != valueMap.end() ? Value::getFromOpaquePointer(it->second)
-                                : value;
+    auto it = valueMap.find(from);
+    return it != valueMap.end() ? it->second : value;
   }
 
-  DenseMap<void *, void *> valueMap;
+  DenseMap<Value, Value> valueMap;
+  DenseMap<Block *, Block *> blockMap;
 };
 
 } // end namespace mlir
 
 #endif // MLIR_IR_BLOCKANDVALUEMAPPING_H

mlir/include/mlir/IR/Operation.h

bool isAncestor(Operation *other) {
     return this == other || isProperAncestor(other);
   }
 
   /// Replace any uses of 'from' with 'to' within this operation.
   void replaceUsesOfWith(Value from, Value to);
 
   /// Replace all uses of results of this operation with the provided 'values'.
   template <typename ValuesT>
-  std::enable_if_t<!std::is_convertible<ValuesT, Operation *>::value>
-  replaceAllUsesWith(ValuesT &&values) {
-    assert(std::distance(values.begin(), values.end()) == getNumResults() &&
-           "expected 'values' to correspond 1-1 with the number of results");
-
-    auto valueIt = values.begin();
-    for (unsigned i = 0, e = getNumResults(); i != e; ++i)
-      getResult(i).replaceAllUsesWith(*(valueIt++));
-  }
-
-  /// Replace all uses of results of this operation with results of 'op'.
-  void replaceAllUsesWith(Operation *op) {
-    assert(getNumResults() == op->getNumResults());
-    for (unsigned i = 0, e = getNumResults(); i != e; ++i)
-      getResult(i).replaceAllUsesWith(op->getResult(i));
+  void replaceAllUsesWith(ValuesT &&values) {
+    getResults().replaceAllUsesWith(std::forward<ValuesT>(values));
   }
 
   /// Destroys this operation and its subclass data.
   void destroy();
 
   /// This drops all operand uses from this operation, which is an essential
   /// step in breaking cyclic dependences between references when they are to
   /// be deleted.

mlir/include/mlir/IR/OperationSupport.h

 // ResultRange
 
 /// This class implements the result iterators for the Operation class.
 class ResultRange final
     : public llvm::detail::indexed_accessor_range_base<
           ResultRange, detail::OpResultImpl *, OpResult, OpResult, OpResult> {
 public:
   using RangeBaseT::RangeBaseT;
+  ResultRange(OpResult result);
 
   //===--------------------------------------------------------------------===//
   // Types
   //===--------------------------------------------------------------------===//
 
   /// Returns the types of the values within this range.
   using type_iterator = ValueTypeIterator<iterator>;
   using type_range = ValueTypeRange<ResultRange>;
public:
   use_iterator use_end() const;
 
   /// Returns true if no results in this range have uses.
   bool use_empty() const {
     return llvm::all_of(*this,
                         [](OpResult result) { return result.use_empty(); });
   }
 
+  /// Replace all uses of results of this range with the provided 'values'. The
+  /// size of `values` must match the size of this range.
+  template <typename ValuesT>
+  std::enable_if_t<!std::is_convertible<ValuesT, Operation *>::value>
+  replaceAllUsesWith(ValuesT &&values) {
+    assert(std::distance(values.begin(), values.end()) == size() &&
+           "expected 'values' to correspond 1-1 with the number of results");
+
+    for (auto it : llvm::zip(*this, values))
+      std::get<0>(it).replaceAllUsesWith(std::get<1>(it));
+  }
+
+  /// Replace all uses of results of this range with results of 'op'.
+  void replaceAllUsesWith(Operation *op);
+
   //===--------------------------------------------------------------------===//
   // Users
   //===--------------------------------------------------------------------===//
 
   using user_iterator = ValueUserIterator<use_iterator, OpOperand>;
   using user_range = iterator_range<user_iterator>;
 
   /// Returns a range of all users.

mlir/include/mlir/Transforms/DialectConversion.h

public:
   /// following form:
   ///   `Optional<Value>(OpBuilder &, T, ValueRange, Location)`,
   /// where `T` is any subclass of `Type`. This function is responsible for
   /// creating an operation, using the OpBuilder and Location provided, that
   /// "casts" a range of values into a single value of the given type `T`. It
   /// must return a Value of the converted type on success, an `llvm::None` if
   /// it failed but other materialization can be attempted, and `nullptr` on
   /// unrecoverable failure. It will only be called for (sub)types of `T`.
-  /// Materialization functions must be provided when a type conversion
-  /// results in more than one type, or if a type conversion may persist after
-  /// the conversion has finished.
+  /// Materialization functions must be provided when a type conversion may
+  /// persist after the conversion has finished.
   ///
   /// This method registers a materialization that will be called when
   /// converting an illegal block argument type, to a legal type.
   template <typename FnT, typename T = typename llvm::function_traits<
                               std::decay_t<FnT>>::template arg_t<1>>
   void addArgumentMaterialization(FnT &&callback) {
     argumentMaterializations.emplace_back(
         wrapMaterialization<T>(std::forward<FnT>(callback)));
public:
   /// behavior can be specified in blockConversions.
   LogicalResult convertNonEntryRegionTypes(
       Region *region, TypeConverter &converter,
       ArrayRef<TypeConverter::SignatureConversion> blockConversions);
 
   /// Replace all the uses of the block argument `from` with value `to`.
   void replaceUsesOfBlockArgument(BlockArgument from, Value to);
 
-  /// Return the converted value that replaces 'key'. Return 'key' if there is
-  /// no such a converted value.
+  /// Return the converted value of 'key' with a type defined by the type
+  /// converter of the currently executing pattern. Return nullptr in the case
+  /// of failure, the remapped value otherwise.
   Value getRemappedValue(Value key);
 
+  /// Return the converted values that replace 'keys' with types defined by the
+  /// type converter of the currently executing pattern. Returns failure if the
+  /// remap failed, success otherwise.
+  LogicalResult getRemappedValues(ValueRange keys,
+                                  SmallVectorImpl<Value> &results);
+
   //===--------------------------------------------------------------------===//
   // PatternRewriter Hooks
   //===--------------------------------------------------------------------===//
 
   /// PatternRewriter hook for replacing the results of an operation when the
   /// given functor returns true.
   void replaceOpWithIf(
       Operation *op, ValueRange newValues, bool *allUsesReplaced,

mlir/lib/Conversion/SCFToSPIRV/SCFToSPIRV.cpp

 
 namespace {
 /// Common class for all vector to GPU patterns.
 template <typename OpTy>
 class SCFToSPIRVPattern : public OpConversionPattern<OpTy> {
 public:
   SCFToSPIRVPattern<OpTy>(MLIRContext *context, SPIRVTypeConverter &converter,
                           ScfToSPIRVContextImpl *scfToSPIRVContext)
-      : OpConversionPattern<OpTy>::OpConversionPattern(context),
+      : OpConversionPattern<OpTy>::OpConversionPattern(converter, context),
         scfToSPIRVContext(scfToSPIRVContext), typeConverter(converter) {}
 
 protected:
   ScfToSPIRVContextImpl *scfToSPIRVContext;
   // FIXME: We explicitly keep a reference of the type converter here instead of
   // passing it to OpConversionPattern during construction. This effectively
   // bypasses the conversion framework's automation on type conversion. This is
   // needed right now because the conversion framework will unconditionally

mlir/lib/Conversion/SPIRVToLLVM/SPIRVToLLVM.cpp

 static Value createI32ConstantOf(Location loc, PatternRewriter &rewriter,
                                  unsigned value) {
   return rewriter.create<LLVM::ConstantOp>(
       loc, IntegerType::get(rewriter.getContext(), 32),
       rewriter.getIntegerAttr(rewriter.getI32Type(), value));
 }
 
 /// Utility for `spv.Load` and `spv.Store` conversion.
-static LogicalResult replaceWithLoadOrStore(Operation *op,
+static LogicalResult replaceWithLoadOrStore(Operation *op, ValueRange operands,
                                             ConversionPatternRewriter &rewriter,
                                             LLVMTypeConverter &typeConverter,
                                             unsigned alignment, bool isVolatile,
                                             bool isNonTemporal) {
   if (auto loadOp = dyn_cast<spirv::LoadOp>(op)) {
     auto dstType = typeConverter.convertType(loadOp.getType());
     if (!dstType)
       return failure();
     rewriter.replaceOpWithNewOp<LLVM::LoadOp>(
-        loadOp, dstType, loadOp.ptr(), alignment, isVolatile, isNonTemporal);
+        loadOp, dstType, spirv::LoadOpAdaptor(operands).ptr(), alignment,
+        isVolatile, isNonTemporal);
     return success();
   }
   auto storeOp = cast<spirv::StoreOp>(op);
-  rewriter.replaceOpWithNewOp<LLVM::StoreOp>(storeOp, storeOp.value(),
-                                             storeOp.ptr(), alignment,
+  spirv::StoreOpAdaptor adaptor(operands);
+  rewriter.replaceOpWithNewOp<LLVM::StoreOp>(storeOp, adaptor.value(),
+                                             adaptor.ptr(), alignment,
                                              isVolatile, isNonTemporal);
   return success();
 }
 
 //===----------------------------------------------------------------------===//
 // Type conversion
 //===----------------------------------------------------------------------===//
 
public:
 
   LogicalResult
   matchAndRewrite(spirv::AccessChainOp op, OpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override {
     auto dstType = typeConverter.convertType(op.component_ptr().getType());
     if (!dstType)
       return failure();
     // To use GEP we need to add a first 0 index to go through the pointer.
-    auto indices = llvm::to_vector<4>(op.indices());
+    auto indices = llvm::to_vector<4>(adaptor.indices());
     Type indexType = op.indices().front().getType();
     auto llvmIndexType = typeConverter.convertType(indexType);
     if (!llvmIndexType)
       return failure();
     Value zero = rewriter.create<LLVM::ConstantOp>(
         op.getLoc(), llvmIndexType, rewriter.getIntegerAttr(indexType, 0));
     indices.insert(indices.begin(), zero);
-    rewriter.replaceOpWithNewOp<LLVM::GEPOp>(op, dstType, op.base_ptr(),
+    rewriter.replaceOpWithNewOp<LLVM::GEPOp>(op, dstType, adaptor.base_ptr(),
                                              indices);
     return success();
   }
 };
 
 class AddressOfPattern : public SPIRVToLLVMConversion<spirv::AddressOfOp> {
 public:
   using SPIRVToLLVMConversion<spirv::AddressOfOp>::SPIRVToLLVMConversion;
if (!dstType)
       return failure();
 
     Type containerType = op.composite().getType();
     if (containerType.isa<VectorType>()) {
       Location loc = op.getLoc();
       IntegerAttr value = op.indices()[0].cast<IntegerAttr>();
       Value index = createI32ConstantOf(loc, rewriter, value.getInt());
       rewriter.replaceOpWithNewOp<LLVM::ExtractElementOp>(
-          op, dstType, op.composite(), index);
+          op, dstType, adaptor.composite(), index);
       return success();
     }
     rewriter.replaceOpWithNewOp<LLVM::ExtractValueOp>(
-        op, dstType, op.composite(), op.indices());
+        op, dstType, adaptor.composite(), op.indices());
     return success();
   }
 };
 
 /// Converts `spv.CompositeInsert` to `llvm.insertvalue` if the container type
 /// is an aggregate type (struct or array). Otherwise, converts to
 /// `llvm.insertelement` that operates on vectors.
 class CompositeInsertPattern
if (!dstType)
       return failure();
 
     Type containerType = op.composite().getType();
     if (containerType.isa<VectorType>()) {
       Location loc = op.getLoc();
       IntegerAttr value = op.indices()[0].cast<IntegerAttr>();
       Value index = createI32ConstantOf(loc, rewriter, value.getInt());
       rewriter.replaceOpWithNewOp<LLVM::InsertElementOp>(
-          op, dstType, op.composite(), op.object(), index);
+          op, dstType, adaptor.composite(), adaptor.object(), index);
       return success();
     }
     rewriter.replaceOpWithNewOp<LLVM::InsertValueOp>(
-        op, dstType, op.composite(), op.object(), op.indices());
+        op, dstType, adaptor.composite(), adaptor.object(), op.indices());
     return success();
   }
 };
 
 /// Converts SPIR-V operations that have straightforward LLVM equivalent
 /// into LLVM dialect operations.
 template <typename SPIRVOp, typename LLVMOp>
 class DirectConversionPattern : public SPIRVToLLVMConversion<SPIRVOp> {

 class LoadStorePattern : public SPIRVToLLVMConversion<SPIRVOp> {
 public:
   using SPIRVToLLVMConversion<SPIRVOp>::SPIRVToLLVMConversion;
 
   LogicalResult
   matchAndRewrite(SPIRVOp op, typename SPIRVOp::Adaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override {
     if (!op.memory_access().hasValue()) {
-      return replaceWithLoadOrStore(
-          op, rewriter, this->typeConverter, /*alignment=*/0,
-          /*isVolatile=*/false, /*isNonTemporal=*/false);
+      return replaceWithLoadOrStore(op, adaptor.getOperands(), rewriter,
+                                    this->typeConverter, /*alignment=*/0,
+                                    /*isVolatile=*/false,
+                                    /*isNonTemporal=*/false);
     }
     auto memoryAccess = op.memory_access().getValue();
     switch (memoryAccess) {
     case spirv::MemoryAccess::Aligned:
     case spirv::MemoryAccess::None:
     case spirv::MemoryAccess::Nontemporal:
     case spirv::MemoryAccess::Volatile: {
       unsigned alignment =
           memoryAccess == spirv::MemoryAccess::Aligned ? *op.alignment() : 0;
       bool isNonTemporal = memoryAccess == spirv::MemoryAccess::Nontemporal;
       bool isVolatile = memoryAccess == spirv::MemoryAccess::Volatile;
-      return replaceWithLoadOrStore(op, rewriter, this->typeConverter,
-                                    alignment, isVolatile, isNonTemporal);
+      return replaceWithLoadOrStore(op, adaptor.getOperands(), rewriter,
+                                    this->typeConverter, alignment, isVolatile,
+                                    isNonTemporal);
     }
     default:
       // There is no support of other memory access attributes.
       return failure();
     }
   }
 };
 
if (op1Type == op2Type) {
                                                    adaptor.getOperands());
       return success();
     }
 
     Location loc = operation.getLoc();
     Value extended;
     if (isUnsignedIntegerOrVector(op2Type)) {
       extended = rewriter.template create<LLVM::ZExtOp>(loc, dstType,
-                                                        operation.operand2());
+                                                        adaptor.operand2());
     } else {
       extended = rewriter.template create<LLVM::SExtOp>(loc, dstType,
-                                                        operation.operand2());
+                                                        adaptor.operand2());
     }
     Value result = rewriter.template create<LLVMOp>(
-        loc, dstType, operation.operand1(), extended);
+        loc, dstType, adaptor.operand1(), extended);
     rewriter.replaceOp(operation, result);
     return success();
   }
 };
 
 class TanPattern : public SPIRVToLLVMConversion<spirv::GLSLTanOp> {
 public:
   using SPIRVToLLVMConversion<spirv::GLSLTanOp>::SPIRVToLLVMConversion;
matchAndRewrite(spirv::VariableOp varOp, OpAdaptor adaptor,
 
     Location loc = varOp.getLoc();
     Value size = createI32ConstantOf(loc, rewriter, 1);
     if (!init) {
       rewriter.replaceOpWithNewOp<LLVM::AllocaOp>(varOp, dstType, size);
       return success();
     }
     Value allocated = rewriter.create<LLVM::AllocaOp>(loc, dstType, size);
-    rewriter.create<LLVM::StoreOp>(loc, init, allocated);
+    rewriter.create<LLVM::StoreOp>(loc, adaptor.initializer(), allocated);
     rewriter.replaceOp(varOp, allocated);
     return success();
   }
 };
 
 //===----------------------------------------------------------------------===//
 // FuncOp conversion
 //===----------------------------------------------------------------------===//

mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp

if (rank == 1 && v1Type == v2Type) {
       Value llvmShuffleOp = rewriter.create<LLVM::ShuffleVectorOp>(
           loc, adaptor.v1(), adaptor.v2(), maskArrayAttr);
       rewriter.replaceOp(shuffleOp, llvmShuffleOp);
       return success();
     }
 
     // For all other cases, insert the individual values individually.
     Type eltType;
-    llvm::errs() << llvmType << "\n";
     if (auto arrayType = llvmType.dyn_cast<LLVM::LLVMArrayType>())
       eltType = arrayType.getElementType();
     else
       eltType = llvmType.cast<VectorType>().getElementType();
     Value insert = rewriter.create<LLVM::UndefOp>(loc, llvmType);
     int64_t insPos = 0;
     for (auto en : llvm::enumerate(maskArrayAttr)) {
       int64_t extPos = en.value().cast<IntegerAttr>().getInt();

mlir/lib/Dialect/Linalg/Transforms/Detensorize.cpp

 #include <memory>
 
 using namespace mlir;
 using namespace mlir::linalg;
 
 static Value sourceMaterializationCallback(OpBuilder &builder, Type type,
                                            ValueRange inputs, Location loc) {
   assert(inputs.size() == 1);
+  if (inputs[0].getType().isa<TensorType>())
+    return nullptr;
+
   // A detensored value is converted back by creating a new tensor from its
   // element(s).
   auto createNewTensorOp = builder.create<tensor::FromElementsOp>(
       loc, inputs[0].getType(), inputs[0]);
 
   // FromElementsOp results in a tensor<1xdtype>, we need to reshape that to
   // a tensor<dtype> instead.
   return builder.create<linalg::TensorCollapseShapeOp>(

mlir/lib/IR/OperationSupport.cpp

MutableOperandRange MutableOperandRangeRange::dereference(const OwnerT &object,
   return object.first.slice(
       startIndex, *(sizeData.begin() + index),
       MutableOperandRange::OperandSegment(index, object.second));
 }
 
 //===----------------------------------------------------------------------===//
 // ResultRange
 
+ResultRange::ResultRange(OpResult result)
+    : ResultRange(static_cast<detail::OpResultImpl *>(Value(result).getImpl()),
+                  1) {}
+
 ResultRange::use_range ResultRange::getUses() const {
   return {use_begin(), use_end()};
 }
 ResultRange::use_iterator ResultRange::use_begin() const {
   return use_iterator(*this);
 }
 ResultRange::use_iterator ResultRange::use_end() const {
   return use_iterator(*this, /*end=*/true);
void ResultRange::UseIterator::skipOverResultsWithNoUsers() {
   // If we are at the last result, then set use to first use of
   // first result (sentinel value used for end).
   if (it == endIt)
     use = {};
   else
     use = (*it).use_begin();
 }
 
+void ResultRange::replaceAllUsesWith(Operation *op) {
+  replaceAllUsesWith(op->getResults());
+}
+
 //===----------------------------------------------------------------------===//
 // ValueRange
 
 ValueRange::ValueRange(ArrayRef<Value> values)
     : ValueRange(values.data(), values.size()) {}
 ValueRange::ValueRange(OperandRange values)
     : ValueRange(values.begin().getBase(), values.size()) {}
 ValueRange::ValueRange(ResultRange values)

mlir/lib/Transforms/Utils/DialectConversion.cpp

 
 #define DEBUG_TYPE "dialect-conversion"
 
 /// Recursively collect all of the operations to convert from within 'region'.
 /// If 'target' is nonnull, operations that are recursively legal have their
 /// regions pre-filtered to avoid considering them for legalization.
 static LogicalResult
 computeConversionSet(iterator_range<Region::iterator> region,
-                     Location regionLoc, std::vector<Operation *> &toConvert,
+                     Location regionLoc,
+                     SmallVectorImpl<Operation *> &toConvert,
                      ConversionTarget *target = nullptr) {
   if (llvm::empty(region))
     return success();
 
   // Traverse starting from the entry block.
   SmallVector<Block *, 16> worklist(1, &*region.begin());
   DenseSet<Block *> visitedBlocks;
   visitedBlocks.insert(worklist.front());
struct ConversionValueMapping {
   /// does not exist then return the provided value. If `desiredType` is
   /// non-null, returns the most recently mapped value with that type. If an
   /// operand of that type does not exist, defaults to normal behavior.
   Value lookupOrDefault(Value from, Type desiredType = nullptr) const;
 
   /// Lookup a mapped value within the map, or return null if a mapping does not
   /// exist. If a mapping exists, this follows the same behavior of
   /// `lookupOrDefault`.
-  Value lookupOrNull(Value from) const;
+  Value lookupOrNull(Value from, Type desiredType = nullptr) const;
 
   /// Map a value to the one provided.
-  void map(Value oldVal, Value newVal) { mapping.map(oldVal, newVal); }
+  void map(Value oldVal, Value newVal) {
+    LLVM_DEBUG({
+      for (Value it = newVal; it; it = mapping.lookupOrNull(it))
+        assert(it != oldVal && "inserting cyclic mapping");
+    });
+    mapping.map(oldVal, newVal);
+  }
+
+  /// Try to map a value to the one provided. Returns false if a transitive
+  /// mapping from the new value to the old value already exists, true if the
+  /// map was updated.
+  bool tryMap(Value oldVal, Value newVal);
 
   /// Drop the last mapping for the given value.
   void erase(Value value) { mapping.erase(value); }
 
   /// Returns the inverse raw value mapping (without recursive query support).
-  BlockAndValueMapping getInverse() const { return mapping.getInverse(); }
+  DenseMap<Value, SmallVector<Value>> getInverse() const {
+    DenseMap<Value, SmallVector<Value>> inverse;
+    for (auto &it : mapping.getValueMap())
+      inverse[it.second].push_back(it.first);
+    return inverse;
+  }
 
 private:
   /// Current value mappings.
   BlockAndValueMapping mapping;
 };
 } // end anonymous namespace
 
 Value ConversionValueMapping::lookupOrDefault(Value from,
if (!mappedValue)
       break;
     from = mappedValue;
   } while (true);
 
   // If the desired value was found use it, otherwise default to the leaf value.
   return desiredValue ? desiredValue : from;
 }
 
-Value ConversionValueMapping::lookupOrNull(Value from) const {
-  Value result = lookupOrDefault(from);
+Value ConversionValueMapping::lookupOrNull(Value from, Type desiredType) const {
+  Value result = lookupOrDefault(from, desiredType);
   return result == from ? nullptr : result;
 }
 
+bool ConversionValueMapping::tryMap(Value oldVal, Value newVal) {
+  for (Value it = newVal; it; it = mapping.lookupOrNull(it))
+    if (it == oldVal)
+      return false;
+  map(oldVal, newVal);
+  return true;
+}
+
 //===----------------------------------------------------------------------===//
 // Rewriter and Translation State
 //===----------------------------------------------------------------------===//
 namespace {
 /// This class contains a snapshot of the current conversion rewriter state.
 /// This is useful when saving and undoing a set of rewrites.
 struct RewriterState {
-  RewriterState(unsigned numCreatedOps, unsigned numReplacements,
-                unsigned numArgReplacements, unsigned numBlockActions,
-                unsigned numIgnoredOperations, unsigned numRootUpdates)
-      : numCreatedOps(numCreatedOps), numReplacements(numReplacements),
+  RewriterState(unsigned numCreatedOps, unsigned numUnresolvedMaterializations,
+                unsigned numReplacements, unsigned numArgReplacements,
+                unsigned numBlockActions, unsigned numIgnoredOperations,
+                unsigned numRootUpdates)
+      : numCreatedOps(numCreatedOps),
+        numUnresolvedMaterializations(numUnresolvedMaterializations),
+        numReplacements(numReplacements),
         numArgReplacements(numArgReplacements),
         numBlockActions(numBlockActions),
         numIgnoredOperations(numIgnoredOperations),
         numRootUpdates(numRootUpdates) {}
 
   /// The current number of created operations.
   unsigned numCreatedOps;
 
+  /// The current number of unresolved materializations.
+  unsigned numUnresolvedMaterializations;
+
   /// The current number of replacements queued.
   unsigned numReplacements;
 
   /// The current number of argument replacements queued.
   unsigned numArgReplacements;
 
   /// The current number of block actions performed.
   unsigned numBlockActions;
union {
     // In use if kind == BlockActionKind::Split and contains a pointer to the
     // block that was split into two parts.
     Block *originalBlock;
     // In use if kind == BlockActionKind::Merge, and contains the information
     // needed to undo the merge.
     MergeInfo mergeInfo;
   };
 };
+
+//===----------------------------------------------------------------------===//
+// UnresolvedMaterialization
+
+/// This class represents an unresolved materialization, i.e. a materialization
+/// that was inserted during conversion that needs to be legalized at the end of
+/// the conversion process.
+class UnresolvedMaterialization {
+public:
+  /// The type of materialization.
+  enum Kind {
+    /// This materialization materializes a conversion for an illegal block
+    /// argument type, to a legal one.
+    Argument,
+
+    /// This materialization materializes a conversion from an illegal type to a
+    /// legal one.
+    Target
+  };
+
+  UnresolvedMaterialization(UnrealizedConversionCastOp op = nullptr,
+                            TypeConverter *converter = nullptr,
+                            Kind kind = Target, Type origOutputType = nullptr)
+      : op(op), converterAndKind(converter, kind),
+        origOutputType(origOutputType) {}
+
+  /// Return the temporary conversion operation inserted for this
+  /// materialization.
+  UnrealizedConversionCastOp getOp() const { return op; }
+
+  /// Return the type converter of this materialization (which may be null).
+  TypeConverter *getConverter() const { return converterAndKind.getPointer(); }
+
+  /// Return the kind of this materialization.
+  Kind getKind() const { return converterAndKind.getInt(); }
+
+  /// Set the kind of this materialization.
+  void setKind(Kind kind) { converterAndKind.setInt(kind); }
+
+  /// Return the original illegal output type of the input values.
+  Type getOrigOutputType() const { return origOutputType; }
+
+private:
+  /// The unresolved materialization operation created during conversion.
+  UnrealizedConversionCastOp op;
+
+  /// The corresponding type converter to use when resolving this
+  /// materialization, and the kind of this materialization.
+  llvm::PointerIntPair<TypeConverter *, 1, Kind> converterAndKind;
+
+  /// The original output type. This is only used for argument conversions.
+  Type origOutputType;
+};
 } // end anonymous namespace
 
+/// Build an unresolved materialization operation given an output type and set
+/// of input operands.
+static Value buildUnresolvedMaterialization(
+    UnresolvedMaterialization::Kind kind, Block *insertBlock,
+    Block::iterator insertPt, Location loc, ValueRange inputs, Type outputType,
+    Type origOutputType, TypeConverter *converter,
+    SmallVectorImpl<UnresolvedMaterialization> &unresolvedMaterializations) {
+  // Avoid materializing an unnecessary cast.
+  if (inputs.size() == 1 && inputs.front().getType() == outputType)
+    return inputs.front();
+
+  // Create an unresolved materialization. We use a new OpBuilder to avoid
+  // tracking the materialization like we do for other operations.
+  OpBuilder builder(insertBlock, insertPt);
+  auto convertOp =
+      builder.create<UnrealizedConversionCastOp>(loc, outputType, inputs);
+  unresolvedMaterializations.emplace_back(convertOp, converter, kind,
+                                          origOutputType);
+  return convertOp.getResult(0);
+}
+static Value buildUnresolvedArgumentMaterialization(
+    PatternRewriter &rewriter, Location loc, ValueRange inputs,
+    Type origOutputType, Type outputType, TypeConverter *converter,
+    SmallVectorImpl<UnresolvedMaterialization> &unresolvedMaterializations) {
+  return buildUnresolvedMaterialization(
+      UnresolvedMaterialization::Argument, rewriter.getInsertionBlock(),
+      rewriter.getInsertionPoint(), loc, inputs, outputType, origOutputType,
+      converter, unresolvedMaterializations);
+}
+static Value buildUnresolvedTargetMaterialization(
+    Location loc, Value input, Type outputType, TypeConverter *converter,
+    SmallVectorImpl<UnresolvedMaterialization> &unresolvedMaterializations) {
+  Block *insertBlock = input.getParentBlock();
+  Block::iterator insertPt = insertBlock->begin();
+  if (OpResult inputRes = input.dyn_cast<OpResult>())
+    insertPt = ++inputRes.getOwner()->getIterator();
+
+  return buildUnresolvedMaterialization(
+      UnresolvedMaterialization::Target, insertBlock, insertPt, loc, input,
+      outputType, outputType, converter, unresolvedMaterializations);
+}
+
 //===----------------------------------------------------------------------===//
 // ArgConverter
 //===----------------------------------------------------------------------===//
 namespace {
 /// This class provides a simple interface for converting the types of block
 /// arguments. This is done by creating a new block that contains the new legal
 /// types and extracting the block that contains the old illegal types to allow
 /// for undoing pending rewrites in the case of failure.
 struct ArgConverter {
-  ArgConverter(PatternRewriter &rewriter) : rewriter(rewriter) {}
+  ArgConverter(
+      PatternRewriter &rewriter,
+      SmallVectorImpl<UnresolvedMaterialization> &unresolvedMaterializations)
+      : rewriter(rewriter),
+        unresolvedMaterializations(unresolvedMaterializations) {}
 
   /// This structure contains the information pertaining to an argument that has
   /// been converted.
   struct ConvertedArgInfo {
     ConvertedArgInfo(unsigned newArgIdx, unsigned newArgSize,
                      Value castValue = nullptr)
         : newArgIdx(newArgIdx), newArgSize(newArgSize), castValue(castValue) {}
 
     /// The start index of in the new argument list that contains arguments that
     /// replace the original.
     unsigned newArgIdx;
 
     /// The number of arguments that replaced the original argument.
     unsigned newArgSize;
 
     /// The cast value that was created to cast from the new arguments to the
     /// old. This only used if 'newArgSize' > 1.
     Value castValue;
   };
 
   /// This structure contains information pertaining to a block that has had its
   /// signature converted.
   struct ConvertedBlockInfo {
-    ConvertedBlockInfo(Block *origBlock, TypeConverter &converter)
-        : origBlock(origBlock), converter(&converter) {}
+    ConvertedBlockInfo(Block *origBlock, TypeConverter *converter)
+        : origBlock(origBlock), converter(converter) {}
 
     /// The original block that was requested to have its signature converted.
     Block *origBlock;
 
     /// The conversion information for each of the arguments. The information is
     /// None if the argument was dropped during conversion.
     SmallVector<Optional<ConvertedArgInfo>, 1> argInfo;
 
struct ArgConverter {
   //===--------------------------------------------------------------------===//
   // Conversion
   //===--------------------------------------------------------------------===//
 
   /// Attempt to convert the signature of the given block, if successful a new
   /// block is returned containing the new arguments. Returns `block` if it did
   /// not require conversion.
   FailureOr<Block *>
-  convertSignature(Block *block, TypeConverter &converter,
+  convertSignature(Block *block, TypeConverter *converter,
                    ConversionValueMapping &mapping,
                    SmallVectorImpl<BlockArgument> &argReplacements);
 
   /// Apply the given signature conversion on the given block. The new block
   /// containing the updated signature is returned. If no conversions were
   /// necessary, e.g. if the block has no arguments, `block` is returned.
   /// `converter` is used to generate any necessary cast operations that
   /// translate between the origin argument types and those specified in the
   /// signature conversion.
   Block *applySignatureConversion(
-      Block *block, TypeConverter &converter,
+      Block *block, TypeConverter *converter,
       TypeConverter::SignatureConversion &signatureConversion,
       ConversionValueMapping &mapping,
       SmallVectorImpl<BlockArgument> &argReplacements);
 
   /// Insert a new conversion into the cache.
   void insertConversion(Block *newBlock, ConvertedBlockInfo &&info);
 
   /// A collection of blocks that have had their arguments converted. This is a
   /// map from the new replacement block, back to the original block.
   llvm::MapVector<Block *, ConvertedBlockInfo> conversionInfo;
 
   /// The set of original blocks that were converted.
   DenseSet<Block *> convertedBlocks;
 
   /// A mapping from valid regions, to those containing the original blocks of a
   /// conversion.
   DenseMap<Region *, std::unique_ptr<Region>> regionMapping;
 
   /// A mapping of regions to type converters that should be used when
   /// converting the arguments of blocks within that region.
   DenseMap<Region *, TypeConverter *> regionToConverter;
 
   /// The pattern rewriter to use when materializing conversions.
   PatternRewriter &rewriter;
+
+  /// An ordered set of unresolved materializations during conversion.
+  SmallVectorImpl<UnresolvedMaterialization> &unresolvedMaterializations;
 };
 } // end anonymous namespace
 
 //===----------------------------------------------------------------------===//
 // Rewrite Application
 
 void ArgConverter::notifyOpRemoved(Operation *op) {
   if (conversionInfo.empty())
LogicalResult ArgConverter::materializeLiveConversions(
     function_ref<Operation *(Value)> findLiveUser) {
   for (auto &info : conversionInfo) {
     Block *newBlock = info.first;
     ConvertedBlockInfo &blockInfo = info.second;
     Block *origBlock = blockInfo.origBlock;
 
     // Process the remapping for each of the original arguments.
     for (unsigned i = 0, e = origBlock->getNumArguments(); i != e; ++i) {
-      // FIXME: We should run the below checks even if the type conversion was
-      // 1->N, but a lot of existing lowering rely on the block argument being
-      // blindly replaced. Those usages should be updated, and this if should be
-      // removed.
-      if (blockInfo.argInfo[i])
+      // FIXME: We should run the below checks even if a type converter wasn't
+      // provided, but a lot of existing lowering rely on the block argument
+      // being blindly replaced. We should rework argument materialization to be
+      // more robust for temporary source materializations, update existing
+      // patterns, and remove these checks.
+      if (!blockInfo.converter && blockInfo.argInfo[i])
         continue;
 
       // If the type of this argument changed and the argument is still live, we
       // need to materialize a conversion.
       BlockArgument origArg = origBlock->getArgument(i);
       auto argReplacementValue = mapping.lookupOrDefault(origArg);
       bool isDroppedArg = argReplacementValue == origArg;
       if (argReplacementValue.getType() == origArg.getType() && !isDroppedArg)
         continue;
       Operation *liveUser = findLiveUser(origArg);
       if (!liveUser)
         continue;
 
       if (OpResult result = argReplacementValue.dyn_cast<OpResult>())
         rewriter.setInsertionPointAfter(result.getOwner());
       else
         rewriter.setInsertionPointToStart(newBlock);
-      Value newArg = blockInfo.converter->materializeSourceConversion(
+      Value newArg;
+      if (blockInfo.converter) {
+        newArg = blockInfo.converter->materializeSourceConversion(
             rewriter, origArg.getLoc(), origArg.getType(),
             isDroppedArg ? ValueRange() : ValueRange(argReplacementValue));
+      }
       if (!newArg) {
         InFlightDiagnostic diag =
             emitError(origArg.getLoc())
             << "failed to materialize conversion for block argument #" << i
             << " that remained live after conversion, type was "
             << origArg.getType();
         if (!isDroppedArg)
           diag << ", with target type " << argReplacementValue.getType();
         diag.attachNote(liveUser->getLoc())
             << "see existing live user here: " << *liveUser;
         return failure();
       }
       mapping.map(origArg, newArg);
     }
   }
   return success();
 }
 
 //===----------------------------------------------------------------------===//
 // Conversion
 
 FailureOr<Block *> ArgConverter::convertSignature(
-    Block *block, TypeConverter &converter, ConversionValueMapping &mapping,
+    Block *block, TypeConverter *converter, ConversionValueMapping &mapping,
     SmallVectorImpl<BlockArgument> &argReplacements) {
   // Check if the block was already converted. If the block is detached,
   // conservatively assume it is going to be deleted.
   if (hasBeenConverted(block) || !block->getParent())
     return block;
+  // If a converter wasn't provided, and the block wasn't already converted,
+  // there is nothing we can do.
+  if (!converter)
+    return failure();
 
   // Try to convert the signature for the block with the provided converter.
-  if (auto conversion = converter.convertBlockSignature(block))
+  if (auto conversion = converter->convertBlockSignature(block))
     return applySignatureConversion(block, converter, *conversion, mapping,
                                     argReplacements);
   return failure();
 }
 
 Block *ArgConverter::applySignatureConversion(
-    Block *block, TypeConverter &converter,
+    Block *block, TypeConverter *converter,
     TypeConverter::SignatureConversion &signatureConversion,
     ConversionValueMapping &mapping,
     SmallVectorImpl<BlockArgument> &argReplacements) {
   // If no arguments are being changed or added, there is nothing to do.
   unsigned origArgCount = block->getNumArguments();
   auto convertedTypes = signatureConversion.getConvertedTypes();
   if (origArgCount == 0 && convertedTypes.empty())
     return block;
if (inputMap->replacementValue) {
       assert(inputMap->size == 0 &&
              "invalid to provide a replacement value when the argument isn't "
              "dropped");
       mapping.map(origArg, inputMap->replacementValue);
       argReplacements.push_back(origArg);
       continue;
     }
 
-    // Otherwise, this is a 1->1+ mapping. Call into the provided type converter
-    // to pack the new values. For 1->1 mappings, if there is no materialization
-    // provided, use the argument directly instead.
+    // Otherwise, this is a 1->1+ mapping.
     auto replArgs = newArgs.slice(inputMap->inputNo, inputMap->size);
     Value newArg;
 
     // If this is a 1->1 mapping and the types of new and replacement arguments
     // match (i.e. it's an identity map), then the argument is mapped to its
     // original type.
-    if (replArgs.size() == 1 && replArgs[0].getType() == origArg.getType())
+    // FIXME: We simply pass through the replacement argument if there wasn't a
+    // converter, which isn't great as it allows implicit type conversions to
+    // appear. We should properly restructure this code to handle cases where a
+    // converter isn't provided and also to properly handle the case where an
+    // argument materialization is actually a temporary source materialization
+    // (e.g. in the case of 1->N).
+    if (replArgs.size() == 1 &&
+        (!converter || replArgs[0].getType() == origArg.getType())) {
       newArg = replArgs.front();
-    else
-      newArg = converter.materializeArgumentConversion(
-          rewriter, origArg.getLoc(), origArg.getType(), replArgs);
+    } else {
+      Type origOutputType = origArg.getType();
 
-    if (!newArg) {
-      assert(replArgs.size() == 1 &&
-             "couldn't materialize the result of 1->N conversion");
-      newArg = replArgs.front();
+      // Legalize the argument output type.
+      Type outputType = origOutputType;
+      if (Type legalOutputType = converter->convertType(outputType))
+        outputType = legalOutputType;
+
+      newArg = buildUnresolvedArgumentMaterialization(
+          rewriter, origArg.getLoc(), replArgs, origOutputType, outputType,
+          converter, unresolvedMaterializations);
     }
+
     mapping.map(origArg, newArg);
     argReplacements.push_back(origArg);
     info.argInfo[i] =
         ConvertedArgInfo(inputMap->inputNo, inputMap->size, newArg);
   }
 
   // Remove the original block from the region and return the new one.
   insertConversion(newBlock, std::move(info));

 
 //===----------------------------------------------------------------------===//
 // ConversionPatternRewriterImpl
 //===----------------------------------------------------------------------===//
 namespace mlir {
 namespace detail {
 struct ConversionPatternRewriterImpl {
   ConversionPatternRewriterImpl(PatternRewriter &rewriter)
-      : argConverter(rewriter) {}
+      : argConverter(rewriter, unresolvedMaterializations) {}
 
   /// Cleanup and destroy any generated rewrite operations. This method is
   /// invoked when the conversion process fails.
   void discardRewrites();
 
   /// Apply all requested operation rewrites. This method is invoked when the
   /// conversion process succeeds.
   void applyRewrites();
struct ConversionPatternRewriterImpl {
   /// Erase any blocks that were unlinked from their regions and stored in block
   /// actions.
   void eraseDanglingBlocks();
 
   /// Undo the block actions (motions, splits) one by one in reverse order until
   /// "numActionsToKeep" actions remains.
   void undoBlockActions(unsigned numActionsToKeep = 0);
 
-  /// Remap the given operands to those with potentially different types. The
-  /// provided type converter is used to ensure that the remapped types are
-  /// legal. Returns success if the operands could be remapped, failure
-  /// otherwise.
-  LogicalResult remapValues(Location loc, PatternRewriter &rewriter,
-                            TypeConverter *converter,
-                            Operation::operand_range operands,
+  /// Remap the given values to those with potentially different types. Returns
+  /// success if the values could be remapped, failure otherwise. `valueDiagTag`
+  /// is the tag used when describing a value within a diagnostic, e.g.
+  /// "operand".
+  LogicalResult remapValues(StringRef valueDiagTag, Optional<Location> inputLoc,
+                            PatternRewriter &rewriter, ValueRange values,
                             SmallVectorImpl<Value> &remapped);
 
   /// Returns true if the given operation is ignored, and does not need to be
   /// converted.
   bool isOpIgnored(Operation *op) const;
 
   /// Recursively marks the nested operations under 'op' as ignored. This
   /// removes them from being considered for legalization.
   void markNestedOpsIgnored(Operation *op);
 
   //===--------------------------------------------------------------------===//
   // Type Conversion
   //===--------------------------------------------------------------------===//
 
   /// Convert the signature of the given block.
   FailureOr<Block *> convertBlockSignature(
-      Block *block, TypeConverter &converter,
+      Block *block, TypeConverter *converter,
       TypeConverter::SignatureConversion *conversion = nullptr);
 
   /// Apply a signature conversion on the given region, using `converter` for
   /// materializations if not null.
   Block *
   applySignatureConversion(Region *region,
                            TypeConverter::SignatureConversion &conversion,
                            TypeConverter *converter);
struct ConversionPatternRewriterImpl {
   // Mapping between replaced values that differ in type. This happens when
   // replacing a value with one of a different type.
   ConversionValueMapping mapping;
 
   /// Utility used to convert block arguments.
   ArgConverter argConverter;
 
   /// Ordered vector of all of the newly created operations during conversion.
-  std::vector<Operation *> createdOps;
+  SmallVector<Operation *> createdOps;
+
+  /// Ordered vector of all unresolved type conversion materializations during
+  /// conversion.
+  SmallVector<UnresolvedMaterialization> unresolvedMaterializations;
 
   /// Ordered map of requested operation replacements.
   llvm::MapVector<Operation *, OpReplacement> replacements;
 
   /// Ordered vector of any requested block argument replacements.
   SmallVector<BlockArgument, 4> argReplacements;
 
   /// Ordered list of block operations (creations, splits, motions).
struct ConversionPatternRewriterImpl {
   /// A transaction state for each of operations that were updated in-place.
   SmallVector<OperationTransactionState, 4> rootUpdates;
 
   /// A vector of indices into `replacements` of operations that were replaced
   /// with values with different result types than the original operation, e.g.
   /// 1->N conversion of some kind.
   SmallVector<unsigned, 4> operationsWithChangedResults;
 
-  /// A default type converter, used when block conversions do not have one
-  /// explicitly provided.
-  TypeConverter defaultTypeConverter;
-
   /// The current type converter, or nullptr if no type converter is currently
   /// active.
   TypeConverter *currentTypeConverter = nullptr;
 
 #ifndef NDEBUG
   /// A set of operations that have pending updates. This tracking isn't
   /// strictly necessary, and is thus only active during debug builds for extra
   /// verification.

 void ConversionPatternRewriterImpl::discardRewrites() {
   // Reset any operations that were updated in place.
   for (auto &state : rootUpdates)
     state.resetOperation();
 
   undoBlockActions();
 
   // Remove any newly created ops.
+  for (UnresolvedMaterialization &materialization : unresolvedMaterializations)
+    detachNestedAndErase(materialization.getOp());
   for (auto *op : llvm::reverse(createdOps))
     detachNestedAndErase(op);
 }
 
 void ConversionPatternRewriterImpl::applyRewrites() {
   // Apply all of the rewrites replacements requested during conversion.
   for (auto &repl : replacements) {
-    for (OpResult result : repl.first->getResults())
-      if (Value newValue = mapping.lookupOrNull(result))
+    for (OpResult result : repl.first->getResults()) {
+      Value newValue = mapping.lookupOrNull(result, result.getType());
+
+      // FIXME: We should only need to lookup the most recent value with the
+      // same type as realistically all users should still be expecting a value
+      // of the same type. Unfortunately some patterns expect the type to
+      // blindly change, so we need to protect against that until those patterns
+      // are updated to properly handle type conversions. When that happens we
+      // can drop this `if`.
+      if (!newValue)
+        newValue = mapping.lookupOrNull(result);
+
+      if (newValue)
         result.replaceAllUsesWith(newValue);
+    }
 
     // If this operation defines any regions, drop any pending argument
     // rewrites.
     if (repl.first->getNumRegions())
       argConverter.notifyOpRemoved(repl.first);
   }
 
   // Apply all of the requested argument replacements.
for (BlockArgument arg : argReplacements) {
     Operation *replOp = repl.cast<OpResult>().getOwner();
     Block *replBlock = replOp->getBlock();
     arg.replaceUsesWithIf(repl, [&](OpOperand &operand) {
       Operation *user = operand.getOwner();
       return user->getBlock() != replBlock || replOp->isBeforeInBlock(user);
     });
   }
 
+  // Drop all of the unresolved materialization operations created during
+  // conversion.
+  for (auto &mat : unresolvedMaterializations) {
+    mat.getOp()->dropAllUses();
+    mat.getOp()->erase();
+  }
+
   // In a second pass, erase all of the replaced operations in reverse. This
   // allows processing nested operations before their parent region is
   // destroyed. Because we process in reverse order, producers may be deleted
   // before their users (a pattern deleting a producer and then the consumer)
   // so we first drop all uses explicitly.
   for (auto &repl : llvm::reverse(replacements)) {
     repl.first->dropAllUses();
     repl.first->erase();
   }
 
   argConverter.applyRewrites(mapping);
 
   // Now that the ops have been erased, also erase dangling blocks.
   eraseDanglingBlocks();
 }
 
 //===----------------------------------------------------------------------===//
 // State Management
 
 RewriterState ConversionPatternRewriterImpl::getCurrentState() {
-  return RewriterState(createdOps.size(), replacements.size(),
-                       argReplacements.size(), blockActions.size(),
-                       ignoredOps.size(), rootUpdates.size());
+  return RewriterState(createdOps.size(), unresolvedMaterializations.size(),
+                       replacements.size(), argReplacements.size(),
+                       blockActions.size(), ignoredOps.size(),
+                       rootUpdates.size());
 }
 
 void ConversionPatternRewriterImpl::resetState(RewriterState state) {
   // Reset any operations that were updated in place.
   for (unsigned i = state.numRootUpdates, e = rootUpdates.size(); i != e; ++i)
     rootUpdates[i].resetOperation();
   rootUpdates.resize(state.numRootUpdates);
 
   // Reset any replaced arguments.
   for (BlockArgument replacedArg :
        llvm::drop_begin(argReplacements, state.numArgReplacements))
     mapping.erase(replacedArg);
   argReplacements.resize(state.numArgReplacements);
 
   // Undo any block actions.
   undoBlockActions(state.numBlockActions);
 
   // Reset any replaced operations and undo any saved mappings.
   for (auto &repl : llvm::drop_begin(replacements, state.numReplacements))
     for (auto result : repl.first->getResults())
       mapping.erase(result);
   while (replacements.size() != state.numReplacements)
     replacements.pop_back();
 
+  // Pop all of the newly inserted materializations.
+  while (unresolvedMaterializations.size() !=
+         state.numUnresolvedMaterializations) {
+    UnresolvedMaterialization mat = unresolvedMaterializations.pop_back_val();
+    UnrealizedConversionCastOp op = mat.getOp();
+
+    // If this was a target materialization, drop the mapping that was inserted.
+    if (mat.getKind() == UnresolvedMaterialization::Target) {
+      for (Value input : op->getOperands())
+        mapping.erase(input);
+    }
+    detachNestedAndErase(op);
+  }
+
   // Pop all of the newly created operations.
   while (createdOps.size() != state.numCreatedOps) {
     detachNestedAndErase(createdOps.back());
     createdOps.pop_back();
   }
 
   // Pop all of the recorded ignored operations that are no longer valid.
   while (ignoredOps.size() != state.numIgnoredOperations)
case BlockActionKind::TypeConversion: {
       break;
     }
     }
   }
   blockActions.resize(numActionsToKeep);
 }
 
 LogicalResult ConversionPatternRewriterImpl::remapValues(
-    Location loc, PatternRewriter &rewriter, TypeConverter *converter,
-    Operation::operand_range operands, SmallVectorImpl<Value> &remapped) {
-  remapped.reserve(llvm::size(operands));
+    StringRef valueDiagTag, Optional<Location> inputLoc,
+    PatternRewriter &rewriter, ValueRange values,
+    SmallVectorImpl<Value> &remapped) {
+  remapped.reserve(llvm::size(values));
 
   SmallVector<Type, 1> legalTypes;
-  for (auto it : llvm::enumerate(operands)) {
+  for (auto it : llvm::enumerate(values)) {
     Value operand = it.value();
     Type origType = operand.getType();
 
     // If a converter was provided, get the desired legal types for this
     // operand.
     Type desiredType;
-    if (converter) {
+    if (currentTypeConverter) {
       // If there is no legal conversion, fail to match this pattern.
       legalTypes.clear();
-      if (failed(converter->convertType(origType, legalTypes))) {
-        return notifyMatchFailure(loc, [=](Diagnostic &diag) {
-          diag << "unable to convert type for operand #" << it.index()
-               << ", type was " << origType;
+      if (failed(currentTypeConverter->convertType(origType, legalTypes))) {
+        Location operandLoc = inputLoc ? *inputLoc : operand.getLoc();
+        return notifyMatchFailure(operandLoc, [=](Diagnostic &diag) {
+          diag << "unable to convert type for " << valueDiagTag << " #"
+               << it.index() << ", type was " << origType;
         });
       }
       // TODO: There currently isn't any mechanism to do 1->N type conversion
       // via the PatternRewriter replacement API, so for now we just ignore it.
       if (legalTypes.size() == 1)
         desiredType = legalTypes.front();
     } else {
       // TODO: What we should do here is just set `desiredType` to `origType`
       // and then handle the necessary type conversions after the conversion
       // process has finished. Unfortunately a lot of patterns currently rely on
       // receiving the new operands even if the types change, so we keep the
       // original behavior here for now until all of the patterns relying on
       // this get updated.
     }
     Value newOperand = mapping.lookupOrDefault(operand, desiredType);
 
     // Handle the case where the conversion was 1->1 and the new operand type
     // isn't legal.
     Type newOperandType = newOperand.getType();
-    if (converter && desiredType && newOperandType != desiredType) {
-      // Attempt to materialize a conversion for this new value.
-      newOperand = converter->materializeTargetConversion(
-          rewriter, loc, desiredType, newOperand);
-      if (!newOperand) {
-        return notifyMatchFailure(loc, [=](Diagnostic &diag) {
-          diag << "unable to materialize a conversion for "
-                  "operand #"
-               << it.index() << ", from " << newOperandType << " to "
-               << desiredType;
-        });
-      }
+    if (currentTypeConverter && desiredType && newOperandType != desiredType) {
+      Location operandLoc = inputLoc ? *inputLoc : operand.getLoc();
+      Value castValue = buildUnresolvedTargetMaterialization(
+          operandLoc, newOperand, desiredType, currentTypeConverter,
+          unresolvedMaterializations);
+      mapping.map(mapping.lookupOrDefault(newOperand), castValue);
+      newOperand = castValue;
     }
     remapped.push_back(newOperand);
   }
   return success();
 }
 
 bool ConversionPatternRewriterImpl::isOpIgnored(Operation *op) const {
   // Check to see if this operation was replaced or its parent ignored.
if (llvm::any_of(op->getRegions(),
       ignoredOps.insert(op);
   });
 }
 
 //===----------------------------------------------------------------------===//
 // Type Conversion
 
 FailureOr<Block *> ConversionPatternRewriterImpl::convertBlockSignature(
-    Block *block, TypeConverter &converter,
+    Block *block, TypeConverter *converter,
     TypeConverter::SignatureConversion *conversion) {
   FailureOr<Block *> result =
       conversion ? argConverter.applySignatureConversion(
                        block, converter, *conversion, mapping, argReplacements)
                  : argConverter.convertSignature(block, converter, mapping,
                                                  argReplacements);
   if (failed(result))
     return failure();
   if (Block *newBlock = result.getValue()) {
     if (newBlock != block)
       blockActions.push_back(BlockAction::getTypeConversion(newBlock));
   }
   return result;
 }
 
 Block *ConversionPatternRewriterImpl::applySignatureConversion(
     Region *region, TypeConverter::SignatureConversion &conversion,
     TypeConverter *converter) {
-  if (!region->empty()) {
-    return *convertBlockSignature(&region->front(),
-                                  converter ? *converter : defaultTypeConverter,
-                                  &conversion);
-  }
+  if (!region->empty())
+    return *convertBlockSignature(&region->front(), converter, &conversion);
   return nullptr;
 }
 
 FailureOr<Block *> ConversionPatternRewriterImpl::convertRegionTypes(
     Region *region, TypeConverter &converter,
     TypeConverter::SignatureConversion *entryConversion) {
   argConverter.setConverter(region, &converter);
   if (region->empty())
     return nullptr;
 
   if (failed(convertNonEntryRegionTypes(region, converter)))
     return failure();
 
   FailureOr<Block *> newEntry =
-      convertBlockSignature(&region->front(), converter, entryConversion);
+      convertBlockSignature(&region->front(), &converter, entryConversion);
   return newEntry;
 }
 
 LogicalResult ConversionPatternRewriterImpl::convertNonEntryRegionTypes(
     Region *region, TypeConverter &converter,
     ArrayRef<TypeConverter::SignatureConversion> blockConversions) {
   argConverter.setConverter(region, &converter);
   if (region->empty())
LogicalResult ConversionPatternRewriterImpl::convertNonEntryRegionTypes(
   for (Block &block :
        llvm::make_early_inc_range(llvm::drop_begin(*region, 1))) {
     TypeConverter::SignatureConversion *blockConversion =
         blockConversions.empty()
             ? nullptr
             : const_cast<TypeConverter::SignatureConversion *>(
                   &blockConversions[blockIdx++]);
 
-    if (failed(convertBlockSignature(&block, converter, blockConversion)))
+    if (failed(convertBlockSignature(&block, &converter, blockConversion)))
       return failure();
   }
   return success();
 }
 
 //===----------------------------------------------------------------------===//
 // Rewriter Notification Hooks
 
impl->logger.startLine() << "** Replace Argument : '" << from
                              << "'(in region of '" << parentOp->getName()
                              << "'(" << from.getOwner()->getParentOp() << ")\n";
   });
   impl->argReplacements.push_back(from);
   impl->mapping.map(impl->mapping.lookupOrDefault(from), to);
 }
 
 Value ConversionPatternRewriter::getRemappedValue(Value key) {
-  return impl->mapping.lookupOrDefault(key);
+  SmallVector<Value> remappedValues;
+  if (failed(impl->remapValues("value", /*inputLoc=*/llvm::None, *this, key,
+                               remappedValues)))
+    return nullptr;
+  return remappedValues.front();
+}
+
+LogicalResult
+ConversionPatternRewriter::getRemappedValues(ValueRange keys,
+                                             SmallVectorImpl<Value> &results) {
+  if (keys.empty())
+    return success();
+  return impl->remapValues("value", /*inputLoc=*/llvm::None, *this, keys,
+                           results);
 }
 
 void ConversionPatternRewriter::notifyBlockCreated(Block *block) {
   impl->notifyCreatedBlock(block);
 }
 
 Block *ConversionPatternRewriter::splitBlock(Block *block,
                                              Block::iterator before) {
ConversionPattern::matchAndRewrite(Operation *op,
   auto &rewriterImpl = dialectRewriter.getImpl();
 
   // Track the current conversion pattern type converter in the rewriter.
   llvm::SaveAndRestore<TypeConverter *> currentConverterGuard(
       rewriterImpl.currentTypeConverter, getTypeConverter());
 
   // Remap the operands of the operation.
   SmallVector<Value, 4> operands;
-  if (failed(rewriterImpl.remapValues(op->getLoc(), rewriter,
-                                      getTypeConverter(), op->getOperands(),
-                                      operands))) {
+  if (failed(rewriterImpl.remapValues("operand", op->getLoc(), rewriter,
+                                      op->getOperands(), operands))) {
     return failure();
   }
   return matchAndRewrite(op, operands, dialectRewriter);
 }
 
 //===----------------------------------------------------------------------===//
 // OperationLegalizer
 //===----------------------------------------------------------------------===//
for (int i = state.numBlockActions, e = newState.numBlockActions; i != e;
     Operation *parentOp = action.block->getParentOp();
     if (!parentOp || parentOp == op || action.block->getNumArguments() == 0)
       continue;
 
     // If the region of the block has a type converter, try to convert the block
     // directly.
     if (auto *converter =
             impl.argConverter.getConverter(action.block->getParent())) {
-      if (failed(impl.convertBlockSignature(action.block, *converter))) {
+      if (failed(impl.convertBlockSignature(action.block, converter))) {
         LLVM_DEBUG(logFailure(impl.logger, "failed to convert types of moved "
                                            "block"));
         return failure();
       }
       continue;
     }
 
     // Otherwise, check that this operation isn't one generated by this pattern.
private:
   /// remaining artifacts and complete the conversion.
   LogicalResult finalize(ConversionPatternRewriter &rewriter);
 
   /// Legalize the types of converted block arguments.
   LogicalResult
   legalizeConvertedArgumentTypes(ConversionPatternRewriter &rewriter,
                                  ConversionPatternRewriterImpl &rewriterImpl);
 
+  /// Legalize any unresolved type materializations.
+  LogicalResult legalizeUnresolvedMaterializations(
+      ConversionPatternRewriter &rewriter,
+      ConversionPatternRewriterImpl &rewriterImpl,
+      Optional<DenseMap<Value, SmallVector<Value>>> &inverseMapping);
+
   /// Legalize an operation result that was marked as "erased".
   LogicalResult
   legalizeErasedResult(Operation *op, OpResult result,
                        ConversionPatternRewriterImpl &rewriterImpl);
 
   /// Legalize an operation result that was replaced with a value of a different
   /// type.
-  LogicalResult
-  legalizeChangedResultType(Operation *op, OpResult result, Value newValue,
-                            TypeConverter *replConverter,
-                            ConversionPatternRewriter &rewriter,
-                            ConversionPatternRewriterImpl &rewriterImpl,
-                            const BlockAndValueMapping &inverseMapping);
+  LogicalResult legalizeChangedResultType(
+      Operation *op, OpResult result, Value newValue,
+      TypeConverter *replConverter, ConversionPatternRewriter &rewriter,
+      const DenseMap<Value, SmallVector<Value>> &inverseMapping);
 
   /// The legalizer to use when converting operations.
   OperationLegalizer opLegalizer;
 
   /// The conversion mode to use when legalizing operations.
   OpConversionMode mode;
 
   /// A set of pre-existing operations. When mode == OpConversionMode::Analysis,

 }
 
 LogicalResult OperationConverter::convertOperations(ArrayRef<Operation *> ops) {
   if (ops.empty())
     return success();
   ConversionTarget &target = opLegalizer.getTarget();
 
   // Compute the set of operations and blocks to convert.
-  std::vector<Operation *> toConvert;
+  SmallVector<Operation *> toConvert;
   for (auto *op : ops) {
     toConvert.emplace_back(op);
     for (auto &region : op->getRegions())
       if (failed(computeConversionSet(region.getBlocks(), region.getLoc(),
                                       toConvert, &target)))
         return failure();
   }
 
if (trackedOps)
       for (auto &repl : rewriterImpl.replacements)
         trackedOps->erase(repl.first);
   }
   return success();
 }
 
 LogicalResult
 OperationConverter::finalize(ConversionPatternRewriter &rewriter) {
+  Optional<DenseMap<Value, SmallVector<Value>>> inverseMapping;
   ConversionPatternRewriterImpl &rewriterImpl = rewriter.getImpl();
-
-  // Legalize converted block arguments.
-  if (failed(legalizeConvertedArgumentTypes(rewriter, rewriterImpl)))
+  if (failed(legalizeUnresolvedMaterializations(rewriter, rewriterImpl,
+                                                inverseMapping)) ||
+      failed(legalizeConvertedArgumentTypes(rewriter, rewriterImpl)))
     return failure();
 
   if (rewriterImpl.operationsWithChangedResults.empty())
     return success();
 
-  Optional<BlockAndValueMapping> inverseMapping;
-
   // Process requested operation replacements.
   for (unsigned i = 0, e = rewriterImpl.operationsWithChangedResults.size();
        i != e; ++i) {
     unsigned replIdx = rewriterImpl.operationsWithChangedResults[i];
     auto &repl = *(rewriterImpl.replacements.begin() + replIdx);
     for (OpResult result : repl.first->getResults()) {
       Value newValue = rewriterImpl.mapping.lookupOrNull(result);
 
if (failed(opLegalizer.legalize(rewriterImpl.createdOps[i], rewriter))) {
       return rewriterImpl.createdOps[i]->emitError()
              << "failed to legalize conversion operation generated for block "
                 "argument that remained live after conversion";
     }
   }
   return success();
 }
 
+/// Replace the results of a materialization operation with the given values.
+static void
+replaceMaterialization(ConversionPatternRewriterImpl &rewriterImpl,
+                       ResultRange matResults, ValueRange values,
+                       DenseMap<Value, SmallVector<Value>> &inverseMapping) {
+  matResults.replaceAllUsesWith(values);
+
+  // For each of the materialization results, update the inverse mappings to
+  // point to the replacement values.
+  for (auto it : llvm::zip(matResults, values)) {
+    Value matResult, newValue;
+    std::tie(matResult, newValue) = it;
+    auto inverseMapIt = inverseMapping.find(matResult);
+    if (inverseMapIt == inverseMapping.end())
+      continue;
+
+    // Update the reverse mapping, or remove the mapping if we couldn't update
+    // it. Not being able to update signals that the mapping would have become
+    // circular (i.e. %foo -> newValue -> %foo), which may occur as values are
+    // propagated through temporary materializations. We simply drop the
+    // mapping, and let the post-conversion replacement logic handle updating
+    // uses.
+    for (Value inverseMapVal : inverseMapIt->second)
+      if (!rewriterImpl.mapping.tryMap(inverseMapVal, newValue))
+        rewriterImpl.mapping.erase(inverseMapVal);
+  }
+}
+
+/// Compute all of the unresolved materializations that will persist beyond the
+/// conversion process, and require inserting a proper user materialization for.
+static void computeNecessaryMaterializations(
+    DenseMap<Operation *, UnresolvedMaterialization *> &materializationOps,
+    ConversionPatternRewriter &rewriter,
+    ConversionPatternRewriterImpl &rewriterImpl,
+    DenseMap<Value, SmallVector<Value>> &inverseMapping,
+    SetVector<UnresolvedMaterialization *> &necessaryMaterializations) {
+  auto isLive = [&](Value value) {
+    auto findFn = [&](Operation *user) {
+      auto matIt = materializationOps.find(user);
+      if (matIt != materializationOps.end())
+        return !necessaryMaterializations.count(matIt->second);
+      return rewriterImpl.isOpIgnored(user);
+    };
+    return llvm::find_if_not(value.getUsers(), findFn) != value.user_end();
+  };
+
+  llvm::unique_function<Value(Value, Value, Type)> lookupRemappedValue =
+      [&](Value invalidRoot, Value value, Type type) {
+        // Check to see if the input operation was remapped to a variant of the
+        // output.
+        Value remappedValue = rewriterImpl.mapping.lookupOrDefault(value, type);
+        if (remappedValue.getType() == type && remappedValue != invalidRoot)
+          return remappedValue;
+
+        // Check to see if the input is a materialization operation that
+        // provides an inverse conversion. We just check blindly for
+        // UnrealizedConversionCastOp here, but it has no effect on correctness.
+        auto inputCastOp = value.getDefiningOp<UnrealizedConversionCastOp>();
+        if (inputCastOp && inputCastOp->getNumOperands() == 1)
+          return lookupRemappedValue(invalidRoot, inputCastOp->getOperand(0),
+                                     type);
+
+        return Value();
+      };
+
+  SetVector<UnresolvedMaterialization *> worklist;
+  for (auto &mat : rewriterImpl.unresolvedMaterializations) {
+    materializationOps.try_emplace(mat.getOp(), &mat);
+    worklist.insert(&mat);
+  }
+  while (!worklist.empty()) {
+    UnresolvedMaterialization *mat = worklist.pop_back_val();
+    UnrealizedConversionCastOp op = mat->getOp();
+
+    // We currently only handle target materializations here.
+    assert(op->getNumResults() == 1 && "unexpected materialization type");
+    OpResult opResult = op->getOpResult(0);
+    Type outputType = opResult.getType();
+    Operation::operand_range inputOperands = op.getOperands();
+
+    // Try to forward propagate operands for user conversion casts that result
+    // in the input types of the current cast.
+    for (Operation *user : llvm::make_early_inc_range(opResult.getUsers())) {
+      auto castOp = dyn_cast<UnrealizedConversionCastOp>(user);
+      if (!castOp)
+        continue;
+      if (castOp->getResultTypes() == inputOperands.getTypes()) {
+        replaceMaterialization(rewriterImpl, opResult, inputOperands,
+                               inverseMapping);
+        necessaryMaterializations.remove(materializationOps.lookup(user));
+      }
+    }
+
+    // Try to avoid materializing a resolved materialization if possible.
+    // Handle the case of a 1-1 materialization.
+    if (inputOperands.size() == 1) {
+      // Check to see if the input operation was remapped to a variant of the
+      // output.
+      Value remappedValue =
+          lookupRemappedValue(opResult, inputOperands[0], outputType);
+      if (remappedValue && remappedValue != opResult) {
+        replaceMaterialization(rewriterImpl, opResult, remappedValue,
+                               inverseMapping);
+        necessaryMaterializations.remove(mat);
+        continue;
+      }
+    } else {
+      // TODO: Avoid materializing other types of conversions here.
+    }
+
+    // Check to see if this is an argument materialization.
+    auto isBlockArg = [](Value v) { return v.isa<BlockArgument>(); };
+    if (llvm::any_of(op->getOperands(), isBlockArg) ||
+        llvm::any_of(inverseMapping[op->getResult(0)], isBlockArg)) {
+      mat->setKind(UnresolvedMaterialization::Argument);
+    }
+
+    // If the materialization does not have any live users, we don't need to
+    // generate a user materialization for it.
+    // FIXME: For argument materializations, we currently need to check if any
+    // of the inverse mapped values are used because some patterns expect blind
+    // value replacement even if the types differ in some cases. When those
+    // patterns are fixed, we can drop the argument special case here.
+    bool isMaterializationLive = isLive(opResult);
+    if (mat->getKind() == UnresolvedMaterialization::Argument)
+      isMaterializationLive |= llvm::any_of(inverseMapping[opResult], isLive);
+    if (!isMaterializationLive)
+      continue;
+    if (!necessaryMaterializations.insert(mat))
+      continue;
+
+    // Reprocess input materializations to see if they have an updated status.
+    for (Value input : inputOperands) {
+      if (auto parentOp = input.getDefiningOp<UnrealizedConversionCastOp>()) {
+        if (auto *mat = materializationOps.lookup(parentOp))
+          worklist.insert(mat);
+      }
+    }
+  }
+}
+
+/// Legalize the given unresolved materialization. Returns success if the
+/// materialization was legalized, failure otherise.
+static LogicalResult legalizeUnresolvedMaterialization(
+    UnresolvedMaterialization &mat,
+    DenseMap<Operation *, UnresolvedMaterialization *> &materializationOps,
+    ConversionPatternRewriter &rewriter,
+    ConversionPatternRewriterImpl &rewriterImpl,
+    DenseMap<Value, SmallVector<Value>> &inverseMapping) {
+  auto findLiveUser = [&](auto &&users) {
+    auto liveUserIt = llvm::find_if_not(
+        users, [&](Operation *user) { return rewriterImpl.isOpIgnored(user); });
+    return liveUserIt == users.end() ? nullptr : *liveUserIt;
+  };
+
+  llvm::unique_function<Value(Value, Type)> lookupRemappedValue =
+      [&](Value value, Type type) {
+        // Check to see if the input operation was remapped to a variant of the
+        // output.
+        Value remappedValue = rewriterImpl.mapping.lookupOrDefault(value, type);
+        if (remappedValue.getType() == type)
+          return remappedValue;
+        return Value();
+      };
+
+  UnrealizedConversionCastOp op = mat.getOp();
+  if (!rewriterImpl.ignoredOps.insert(op))
+    return success();
+
+  // We currently only handle target materializations here.
+  OpResult opResult = op->getOpResult(0);
+  Operation::operand_range inputOperands = op.getOperands();
+  Type outputType = opResult.getType();
+
+  // If any input to this materialization is another materialization, resolve
+  // the input first.
+  for (Value value : op->getOperands()) {
+    auto valueCast = value.getDefiningOp<UnrealizedConversionCastOp>();
+    if (!valueCast)
+      continue;
+
+    auto matIt = materializationOps.find(valueCast);
+    if (matIt != materializationOps.end())
+      if (failed(legalizeUnresolvedMaterialization(
+              *matIt->second, materializationOps, rewriter, rewriterImpl,
+              inverseMapping)))
+        return failure();
+  }
+
+  // Perform a last ditch attempt to avoid materializing a resolved
+  // materialization if possible.
+  // Handle the case of a 1-1 materialization.
+  if (inputOperands.size() == 1) {
+    // Check to see if the input operation was remapped to a variant of the
+    // output.
+    Value remappedValue = lookupRemappedValue(inputOperands[0], outputType);
+    if (remappedValue && remappedValue != opResult) {
+      replaceMaterialization(rewriterImpl, opResult, remappedValue,
+                             inverseMapping);
+      return success();
+    }
+  } else {
+    // TODO: Avoid materializing other types of conversions here.
+  }
+
+  // Try to materialize the conversion.
+  if (TypeConverter *converter = mat.getConverter()) {
+    rewriter.setInsertionPoint(op);
+    Value newMaterialization;
+    switch (mat.getKind()) {
+    case UnresolvedMaterialization::Argument:
+      // Try to materialize an argument conversion.
+      // FIXME: The current argument materialization hook expects the original
+      // output type, even though it doesn't use that as the actual output type
+      // of the generated IR. The output type is just used as an indicator of
+      // the type of materialization to do. This behavior is really awkward in
+      // that it diverges from the behavior of the other hooks, and can be
+      // easily misunderstood. We should clean up the argument hooks to better
+      // represent the desired invariants we actually care about.
+      newMaterialization = converter->materializeArgumentConversion(
+          rewriter, op->getLoc(), mat.getOrigOutputType(), inputOperands);
+      if (newMaterialization)
+        break;
+
+      // If an argument materialization failed, fallback to trying a target
+      // materialization.
+      LLVM_FALLTHROUGH;
+    case UnresolvedMaterialization::Target:
+      newMaterialization = converter->materializeTargetConversion(
+          rewriter, op->getLoc(), outputType, inputOperands);
+      break;
+    default:
+      llvm_unreachable("unknown materialization kind");
+    }
+    if (newMaterialization) {
+      replaceMaterialization(rewriterImpl, opResult, newMaterialization,
+                             inverseMapping);
+      return success();
+    }
+  }
+
+  InFlightDiagnostic diag = op->emitError()
+                            << "failed to legalize unresolved materialization "
+                               "from "
+                            << inputOperands.getTypes() << " to " << outputType
+                            << " that remained live after conversion";
+  if (Operation *liveUser = findLiveUser(op->getUsers())) {
+    diag.attachNote(liveUser->getLoc())
+        << "see existing live user here: " << *liveUser;
+  }
+  return failure();
+}
+
+LogicalResult OperationConverter::legalizeUnresolvedMaterializations(
+    ConversionPatternRewriter &rewriter,
+    ConversionPatternRewriterImpl &rewriterImpl,
+    Optional<DenseMap<Value, SmallVector<Value>>> &inverseMapping) {
+  if (rewriterImpl.unresolvedMaterializations.empty())
+    return success();
+  inverseMapping = rewriterImpl.mapping.getInverse();
+
+  // As an initial step, compute all of the inserted materializations that we
+  // expect to persist beyond the conversion process.
+  DenseMap<Operation *, UnresolvedMaterialization *> materializationOps;
+  SetVector<UnresolvedMaterialization *> necessaryMaterializations;
+  computeNecessaryMaterializations(materializationOps, rewriter, rewriterImpl,
+                                   *inverseMapping, necessaryMaterializations);
+
+  // Once computed, legalize any necessary materializations.
+  for (auto *mat : necessaryMaterializations) {
+    if (failed(legalizeUnresolvedMaterialization(
+            *mat, materializationOps, rewriter, rewriterImpl, *inverseMapping)))
+      return failure();
+  }
+  return success();
+}
+
 LogicalResult OperationConverter::legalizeErasedResult(
     Operation *op, OpResult result,
     ConversionPatternRewriterImpl &rewriterImpl) {
   // If the operation result was replaced with null, all of the uses of this
   // value should be replaced.
   auto liveUserIt = llvm::find_if_not(result.getUsers(), [&](Operation *user) {
     return rewriterImpl.isOpIgnored(user);
   });
   if (liveUserIt != result.user_end()) {
     InFlightDiagnostic diag = op->emitError("failed to legalize operation '")
                               << op->getName() << "' marked as erased";
     diag.attachNote(liveUserIt->getLoc())
         << "found live user of result #" << result.getResultNumber() << ": "
         << *liveUserIt;
     return failure();
   }
   return success();
 }
 
 /// Finds a user of the given value, or of any other value that the given value
 /// replaced, that was not replaced in the conversion process.
-static Operation *
-findLiveUserOfReplaced(Value value, ConversionPatternRewriterImpl &rewriterImpl,
-                       const BlockAndValueMapping &inverseMapping) {
-  do {
+static Operation *findLiveUserOfReplaced(
+    Value initialValue, ConversionPatternRewriterImpl &rewriterImpl,
+    const DenseMap<Value, SmallVector<Value>> &inverseMapping) {
+  SmallVector<Value> worklist(1, initialValue);
+  while (!worklist.empty()) {
+    Value value = worklist.pop_back_val();
+
     // Walk the users of this value to see if there are any live users that
     // weren't replaced during conversion.
     auto liveUserIt = llvm::find_if_not(value.getUsers(), [&](Operation *user) {
       return rewriterImpl.isOpIgnored(user);
     });
     if (liveUserIt != value.user_end())
       return *liveUserIt;
-    value = inverseMapping.lookupOrNull(value);
-  } while (value != nullptr);
+    auto mapIt = inverseMapping.find(value);
+    if (mapIt != inverseMapping.end())
+      worklist.append(mapIt->second);
+  }
   return nullptr;
 }
 
 LogicalResult OperationConverter::legalizeChangedResultType(
     Operation *op, OpResult result, Value newValue,
     TypeConverter *replConverter, ConversionPatternRewriter &rewriter,
     ConversionPatternRewriterImpl &rewriterImpl,
-    const BlockAndValueMapping &inverseMapping) {
+    const DenseMap<Value, SmallVector<Value>> &inverseMapping) {
   Operation *liveUser =
       findLiveUserOfReplaced(result, rewriterImpl, inverseMapping);
   if (!liveUser)
     return success();
 
   // If the replacement has a type converter, attempt to materialize a
   // conversion back to the original type.
   if (!replConverter) {
     // TODO: We should emit an error here, similarly to the case where the
     // result is replaced with null. Unfortunately a lot of existing
     // patterns rely on this behavior, so until those patterns are updated
     // we keep the legacy behavior here of just forwarding the new value.
+    LLVM_DEBUG(llvm::dbgs()
+               << "Result #" << result.getResultNumber() << " of operation `"
+               << *op
+               << "` requires a materialization to preserve the original type `"
+               << result.getType() << "`, but no type converter was provided");
     return success();
   }
 
   // Track the number of created operations so that new ones can be legalized.
   size_t numCreatedOps = rewriterImpl.createdOps.size();
 
   // Materialize a conversion for this live result value.
   Type resultType = result.getType();

mlir/test/Conversion/ArithmeticToLLVM/arith-to-llvm.mlir

 }
 
 // -----
 
 // CHECK-LABEL: @index_vector
 func @index_vector(%arg0: vector<4xindex>) {
   // CHECK: %[[CST:.*]] = llvm.mlir.constant(dense<[0, 1, 2, 3]> : vector<4xindex>) : vector<4xi64>
   %0 = arith.constant dense<[0, 1, 2, 3]> : vector<4xindex>
-  // CHECK: %[[V:.*]] = llvm.add %1, %[[CST]] : vector<4xi64>
+  // CHECK: %[[V:.*]] = llvm.add %{{.*}}, %[[CST]] : vector<4xi64>
   %1 = arith.addi %arg0, %0 : vector<4xindex>
   std.return
 }
 
 // -----
 
 // CHECK-LABEL: @bitcast_1d
 func @bitcast_1d(%arg0: vector<2xf32>) {
   // CHECK: llvm.bitcast %{{.*}} : vector<2xf32> to vector<2xi32>
   arith.bitcast %arg0 : vector<2xf32> to vector<2xi32>
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @cmpf_2dvector(
 func @cmpf_2dvector(%arg0 : vector<4x3xf32>, %arg1 : vector<4x3xf32>) {
-  // CHECK: %[[ARG0:.*]] = builtin.unrealized_conversion_cast
   // CHECK: %[[ARG1:.*]] = builtin.unrealized_conversion_cast
+  // CHECK: %[[ARG0:.*]] = builtin.unrealized_conversion_cast
   // CHECK: %[[EXTRACT1:.*]] = llvm.extractvalue %[[ARG0]][0] : !llvm.array<4 x vector<3xf32>>
   // CHECK: %[[EXTRACT2:.*]] = llvm.extractvalue %[[ARG1]][0] : !llvm.array<4 x vector<3xf32>>
   // CHECK: %[[CMP:.*]] = llvm.fcmp "olt" %[[EXTRACT1]], %[[EXTRACT2]] : vector<3xf32>
   // CHECK: %[[INSERT:.*]] = llvm.insertvalue %[[CMP]], %2[0] : !llvm.array<4 x vector<3xi1>>
   %0 = arith.cmpf olt, %arg0, %arg1 : vector<4x3xf32>
   std.return
 }
 
 // -----
 
 // CHECK-LABEL: func @cmpi_2dvector(
 func @cmpi_2dvector(%arg0 : vector<4x3xi32>, %arg1 : vector<4x3xi32>) {
-  // CHECK: %[[ARG0:.*]] = builtin.unrealized_conversion_cast
   // CHECK: %[[ARG1:.*]] = builtin.unrealized_conversion_cast
+  // CHECK: %[[ARG0:.*]] = builtin.unrealized_conversion_cast
   // CHECK: %[[EXTRACT1:.*]] = llvm.extractvalue %[[ARG0]][0] : !llvm.array<4 x vector<3xi32>>
   // CHECK: %[[EXTRACT2:.*]] = llvm.extractvalue %[[ARG1]][0] : !llvm.array<4 x vector<3xi32>>
   // CHECK: %[[CMP:.*]] = llvm.icmp "ult" %[[EXTRACT1]], %[[EXTRACT2]] : vector<3xi32>
   // CHECK: %[[INSERT:.*]] = llvm.insertvalue %[[CMP]], %2[0] : !llvm.array<4 x vector<3xi1>>
   %0 = arith.cmpi ult, %arg0, %arg1 : vector<4x3xi32>
   std.return
 }

mlir/test/Conversion/ArithmeticToSPIRV/arithmetic-to-spirv.mlir

 // -----
 
 // Check that types are converted to 32-bit when no special capabilities that
 // are not supported.
 module attributes {
   spv.target_env = #spv.target_env<#spv.vce<v1.0, [], []>, {}>
 } {
 
+// expected-error @+1 {{failed to materialize conversion for block argument #0 that remained live after conversion, type was 'vector<4xi64>', with target type 'vector<4xi32>'}}
 func @int_vector4_invalid(%arg0: vector<4xi64>) {
   // expected-error @+2 {{bitwidth emulation is not implemented yet on unsigned op}}
-  // expected-error @+1 {{op requires the same type for all operands and results}}
+  // expected-note @+1 {{see existing live user here}}
   %0 = arith.divui %arg0, %arg0: vector<4xi64>
   return
 }
 
 } // end module
 
 // -----
 

mlir/test/Conversion/ComplexToLLVM/convert-to-llvm.mlir

 // RUN: mlir-opt %s -convert-complex-to-llvm | FileCheck %s
 
 // CHECK-LABEL: func @complex_create
 // CHECK-SAME:    (%[[REAL0:.*]]: f32, %[[IMAG0:.*]]: f32)
 // CHECK-NEXT:    %[[CPLX0:.*]] = llvm.mlir.undef : !llvm.struct<(f32, f32)>
 // CHECK-NEXT:    %[[CPLX1:.*]] = llvm.insertvalue %[[REAL0]], %[[CPLX0]][0] : !llvm.struct<(f32, f32)>
 // CHECK-NEXT:    %[[CPLX2:.*]] = llvm.insertvalue %[[IMAG0]], %[[CPLX1]][1] : !llvm.struct<(f32, f32)>
 func @complex_create(%real: f32, %imag: f32) -> complex<f32> {
   %cplx2 = complex.create %real, %imag : complex<f32>
   return %cplx2 : complex<f32>
 }
 
 // CHECK-LABEL: func @complex_extract
 // CHECK-SAME:    (%[[CPLX:.*]]: complex<f32>)
 // CHECK-NEXT:    %[[CAST0:.*]] = builtin.unrealized_conversion_cast %[[CPLX]] : complex<f32> to !llvm.struct<(f32, f32)>
 // CHECK-NEXT:    %[[REAL:.*]] = llvm.extractvalue %[[CAST0]][0] : !llvm.struct<(f32, f32)>
-// CHECK-NEXT:    %[[CAST1:.*]] = builtin.unrealized_conversion_cast %[[CPLX]] : complex<f32> to !llvm.struct<(f32, f32)>
-// CHECK-NEXT:    %[[IMAG:.*]] = llvm.extractvalue %[[CAST1]][1] : !llvm.struct<(f32, f32)>
+// CHECK-NEXT:    %[[IMAG:.*]] = llvm.extractvalue %[[CAST0]][1] : !llvm.struct<(f32, f32)>
 func @complex_extract(%cplx: complex<f32>) {
   %real1 = complex.re %cplx : complex<f32>
   %imag1 = complex.im %cplx : complex<f32>
   return
 }
 
 // CHECK-LABEL: func @complex_addition
 // CHECK-DAG:     %[[A_REAL:.*]] = llvm.extractvalue %[[A:.*]][0] : !llvm.struct<(f64, f64)>

 }
 
 // CHECK-LABEL: func @complex_div
 // CHECK-SAME:    %[[LHS:.*]]: complex<f32>, %[[RHS:.*]]: complex<f32>
 func @complex_div(%lhs: complex<f32>, %rhs: complex<f32>) -> complex<f32> {
   %div = complex.div %lhs, %rhs : complex<f32>
   return %div : complex<f32>
 }
-// CHECK: %[[CASTED_LHS:.*]] = builtin.unrealized_conversion_cast %[[LHS]] : complex<f32> to ![[C_TY:.*>]]
-// CHECK: %[[CASTED_RHS:.*]] = builtin.unrealized_conversion_cast %[[RHS]] : complex<f32> to ![[C_TY]]
+// CHECK-DAG: %[[CASTED_LHS:.*]] = builtin.unrealized_conversion_cast %[[LHS]] : complex<f32> to ![[C_TY:.*>]]
+// CHECK-DAG: %[[CASTED_RHS:.*]] = builtin.unrealized_conversion_cast %[[RHS]] : complex<f32> to ![[C_TY]]
 
 // CHECK: %[[LHS_RE:.*]] = llvm.extractvalue %[[CASTED_LHS]][0] : ![[C_TY]]
 // CHECK: %[[LHS_IM:.*]] = llvm.extractvalue %[[CASTED_LHS]][1] : ![[C_TY]]
 // CHECK: %[[RHS_RE:.*]] = llvm.extractvalue %[[CASTED_RHS]][0] : ![[C_TY]]
 // CHECK: %[[RHS_IM:.*]] = llvm.extractvalue %[[CASTED_RHS]][1] : ![[C_TY]]
 
 // CHECK: %[[RESULT_0:.*]] = llvm.mlir.undef : ![[C_TY]]
 

 // CHECK: return %[[CASTED_RESULT]] : complex<f32>
 
 // CHECK-LABEL: func @complex_mul
 // CHECK-SAME:    %[[LHS:.*]]: complex<f32>, %[[RHS:.*]]: complex<f32>
 func @complex_mul(%lhs: complex<f32>, %rhs: complex<f32>) -> complex<f32> {
   %mul = complex.mul %lhs, %rhs : complex<f32>
   return %mul : complex<f32>
 }
-// CHECK: %[[CASTED_LHS:.*]] = builtin.unrealized_conversion_cast %[[LHS]] : complex<f32> to ![[C_TY:.*>]]
-// CHECK: %[[CASTED_RHS:.*]] = builtin.unrealized_conversion_cast %[[RHS]] : complex<f32> to ![[C_TY]]
+// CHECK-DAG: %[[CASTED_LHS:.*]] = builtin.unrealized_conversion_cast %[[LHS]] : complex<f32> to ![[C_TY:.*>]]
+// CHECK-DAG: %[[CASTED_RHS:.*]] = builtin.unrealized_conversion_cast %[[RHS]] : complex<f32> to ![[C_TY]]
 
 // CHECK: %[[LHS_RE:.*]] = llvm.extractvalue %[[CASTED_LHS]][0] : ![[C_TY]]
 // CHECK: %[[LHS_IM:.*]] = llvm.extractvalue %[[CASTED_LHS]][1] : ![[C_TY]]
 // CHECK: %[[RHS_RE:.*]] = llvm.extractvalue %[[CASTED_RHS]][0] : ![[C_TY]]
 // CHECK: %[[RHS_IM:.*]] = llvm.extractvalue %[[CASTED_RHS]][1] : ![[C_TY]]
 // CHECK: %[[RESULT_0:.*]] = llvm.mlir.undef : ![[C_TY]]
 
 // CHECK-DAG: %[[REAL_TMP_0:.*]] = llvm.fmul %[[RHS_RE]], %[[LHS_RE]]  : f32

mlir/test/Conversion/MemRefToLLVM/convert-dynamic-memref-ops.mlir

 // RUN: mlir-opt -split-input-file -convert-memref-to-llvm %s | FileCheck %s
 // RUN: mlir-opt -split-input-file -convert-memref-to-llvm='use-aligned-alloc=1' %s | FileCheck %s --check-prefix=ALIGNED-ALLOC
 
 // CHECK-LABEL: func @mixed_alloc(
 //       CHECK:   %[[Marg:.*]]: index, %[[Narg:.*]]: index)
 func @mixed_alloc(%arg0: index, %arg1: index) -> memref<?x42x?xf32> {
-//       CHECK:  %[[M:.*]] = builtin.unrealized_conversion_cast %[[Marg]]
-//       CHECK:  %[[N:.*]] = builtin.unrealized_conversion_cast %[[Narg]]
+//   CHECK-DAG:  %[[M:.*]] = builtin.unrealized_conversion_cast %[[Marg]]
+//   CHECK-DAG:  %[[N:.*]] = builtin.unrealized_conversion_cast %[[Narg]]
 //       CHECK:  %[[c42:.*]] = llvm.mlir.constant(42 : index) : i64
 //  CHECK-NEXT:  %[[one:.*]] = llvm.mlir.constant(1 : index) : i64
 //  CHECK-NEXT:  %[[st0:.*]] = llvm.mul %[[N]], %[[c42]] : i64
 //  CHECK-NEXT:  %[[sz:.*]] = llvm.mul %[[st0]], %[[M]] : i64
 //  CHECK-NEXT:  %[[null:.*]] = llvm.mlir.null : !llvm.ptr<f32>
 //  CHECK-NEXT:  %[[gep:.*]] = llvm.getelementptr %[[null]][%[[sz]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 //  CHECK-NEXT:  %[[sz_bytes:.*]] = llvm.ptrtoint %[[gep]] : !llvm.ptr<f32> to i64
 //  CHECK-NEXT:  llvm.call @malloc(%[[sz_bytes]]) : (i64) -> !llvm.ptr<i8>
// CHECK-NEXT:  llvm.call @free(%[[ptri8]]) : (!llvm.ptr<i8>) -> ()
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @dynamic_alloc(
 //       CHECK:   %[[Marg:.*]]: index, %[[Narg:.*]]: index)
 func @dynamic_alloc(%arg0: index, %arg1: index) -> memref<?x?xf32> {
-//       CHECK:  %[[M:.*]] = builtin.unrealized_conversion_cast %[[Marg]]
-//       CHECK:  %[[N:.*]] = builtin.unrealized_conversion_cast %[[Narg]]
+//   CHECK-DAG:  %[[M:.*]] = builtin.unrealized_conversion_cast %[[Marg]]
+//   CHECK-DAG:  %[[N:.*]] = builtin.unrealized_conversion_cast %[[Narg]]
 //  CHECK-NEXT:  %[[one:.*]] = llvm.mlir.constant(1 : index) : i64
 //  CHECK-NEXT:  %[[sz:.*]] = llvm.mul %[[N]], %[[M]] : i64
 //  CHECK-NEXT:  %[[null:.*]] = llvm.mlir.null : !llvm.ptr<f32>
 //  CHECK-NEXT:  %[[gep:.*]] = llvm.getelementptr %[[null]][%[[sz]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 //  CHECK-NEXT:  %[[sz_bytes:.*]] = llvm.ptrtoint %[[gep]] : !llvm.ptr<f32> to i64
 //  CHECK-NEXT:  llvm.call @malloc(%[[sz_bytes]]) : (i64) -> !llvm.ptr<i8>
 //  CHECK-NEXT:  llvm.bitcast %{{.*}} : !llvm.ptr<i8> to !llvm.ptr<f32>
 //  CHECK-NEXT:  llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
//  CHECK-NEXT:  llvm.insertvalue %[[one]], %{{.*}}[4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   return %0 : memref<?x?xf32>
 }
 
 // -----
 
 // CHECK-LABEL: func @dynamic_alloca
 // CHECK: %[[Marg:.*]]: index, %[[Narg:.*]]: index)
 func @dynamic_alloca(%arg0: index, %arg1: index) -> memref<?x?xf32> {
-//       CHECK:  %[[M:.*]] = builtin.unrealized_conversion_cast %[[Marg]]
-//       CHECK:  %[[N:.*]] = builtin.unrealized_conversion_cast %[[Narg]]
+//   CHECK-DAG:  %[[M:.*]] = builtin.unrealized_conversion_cast %[[Marg]]
+//   CHECK-DAG:  %[[N:.*]] = builtin.unrealized_conversion_cast %[[Narg]]
 //  CHECK-NEXT:  %[[st1:.*]] = llvm.mlir.constant(1 : index) : i64
 //  CHECK-NEXT:  %[[num_elems:.*]] = llvm.mul %[[N]], %[[M]] : i64
 //  CHECK-NEXT:  %[[null:.*]] = llvm.mlir.null : !llvm.ptr<f32>
 //  CHECK-NEXT:  %[[gep:.*]] = llvm.getelementptr %[[null]][%[[num_elems]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 //  CHECK-NEXT:  %[[sz_bytes:.*]] = llvm.ptrtoint %[[gep]] : !llvm.ptr<f32> to i64
 //  CHECK-NEXT:  %[[allocated:.*]] = llvm.alloca %[[sz_bytes]] x f32 : (i64) -> !llvm.ptr<f32>
 //  CHECK-NEXT:  llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 //  CHECK-NEXT:  llvm.insertvalue %[[allocated]], %{{.*}}[0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
// CHECK-NEXT:  llvm.call @free(%[[ptri8]]) : (!llvm.ptr<i8>) -> ()
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @stdlib_aligned_alloc({{.*}})
 // ALIGNED-ALLOC-LABEL: func @stdlib_aligned_alloc({{.*}})
 func @stdlib_aligned_alloc(%N : index) -> memref<32x18xf32> {
-// ALIGNED-ALLOC-NEXT:  %[[sz1:.*]] = llvm.mlir.constant(32 : index) : i64
+// ALIGNED-ALLOC:       %[[sz1:.*]] = llvm.mlir.constant(32 : index) : i64
 // ALIGNED-ALLOC-NEXT:  %[[sz2:.*]] = llvm.mlir.constant(18 : index) : i64
 // ALIGNED-ALLOC-NEXT:  %[[one:.*]] = llvm.mlir.constant(1 : index) : i64
 // ALIGNED-ALLOC-NEXT:  %[[num_elems:.*]] = llvm.mlir.constant(576 : index) : i64
 // ALIGNED-ALLOC-NEXT:  %[[null:.*]] = llvm.mlir.null : !llvm.ptr<f32>
 // ALIGNED-ALLOC-NEXT:  %[[gep:.*]] = llvm.getelementptr %[[null]][%[[num_elems]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 // ALIGNED-ALLOC-NEXT:  %[[bytes:.*]] = llvm.ptrtoint %[[gep]] : !llvm.ptr<f32> to i64
 // ALIGNED-ALLOC-NEXT:  %[[alignment:.*]] = llvm.mlir.constant(32 : index) : i64
 // ALIGNED-ALLOC-NEXT:  %[[allocated:.*]] = llvm.call @aligned_alloc(%[[alignment]], %[[bytes]]) : (i64, i64) -> !llvm.ptr<i8>
// ALIGNED-ALLOC-NEXT:  llvm.bitcast %[[allocated]] : !llvm.ptr<i8> to !llvm.ptr<f32>
   // ALIGNED-ALLOC:  %[[c16:.*]] = llvm.mlir.constant(16 : index) : i64
   // ALIGNED-ALLOC-NEXT:  llvm.call @aligned_alloc(%[[c16]],
   %3 = memref.alloc() : memref<4096xvector<2xf32>>
   // ALIGNED-ALLOC:  %[[c8:.*]] = llvm.mlir.constant(8 : index) : i64
   // ALIGNED-ALLOC-NEXT:  llvm.call @aligned_alloc(%[[c8]],
   %4 = memref.alloc() {alignment = 8} : memref<1024xvector<4xf32>>
   // Bump the memref allocation size if its size is not a multiple of alignment.
   // ALIGNED-ALLOC:       %[[c32:.*]] = llvm.mlir.constant(32 : index) : i64
-  // ALIGNED-ALLOC-NEXT:  llvm.mlir.constant(1 : index) : i64
+  // ALIGNED-ALLOC:       llvm.mlir.constant(1 : index) : i64
   // ALIGNED-ALLOC-NEXT:  llvm.sub
   // ALIGNED-ALLOC-NEXT:  llvm.add
   // ALIGNED-ALLOC-NEXT:  llvm.urem
   // ALIGNED-ALLOC-NEXT:  %[[SIZE_ALIGNED:.*]] = llvm.sub
   // ALIGNED-ALLOC-NEXT:  llvm.call @aligned_alloc(%[[c32]], %[[SIZE_ALIGNED]])
   %5 = memref.alloc() {alignment = 32} : memref<100xf32>
   // Bump alignment to the next power of two if it isn't.
   // ALIGNED-ALLOC:  %[[c128:.*]] = llvm.mlir.constant(128 : index) : i64
   // ALIGNED-ALLOC:  llvm.call @aligned_alloc(%[[c128]]
   %6 = memref.alloc(%N) : memref<?xvector<18xf32>>
   return %0 : memref<32x18xf32>
 }
 
 // -----
 
 // CHECK-LABEL: func @mixed_load(
 // CHECK:         %{{.*}}, %[[Iarg:.*]]: index, %[[Jarg:.*]]: index)
 func @mixed_load(%mixed : memref<42x?xf32>, %i : index, %j : index) {
-//       CHECK:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
-//       CHECK:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
+//   CHECK-DAG:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
+//   CHECK-DAG:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
 //       CHECK:  %[[ptr:.*]] = llvm.extractvalue %[[ld:.*]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 //  CHECK-NEXT:  %[[st0:.*]] = llvm.extractvalue %[[ld]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 //  CHECK-NEXT:  %[[offI:.*]] = llvm.mul %[[I]], %[[st0]] : i64
 //  CHECK-NEXT:  %[[off1:.*]] = llvm.add %[[offI]], %[[J]] : i64
 //  CHECK-NEXT:  %[[addr:.*]] = llvm.getelementptr %[[ptr]][%[[off1]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 //  CHECK-NEXT:  llvm.load %[[addr]] : !llvm.ptr<f32>
   %0 = memref.load %mixed[%i, %j] : memref<42x?xf32>
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @dynamic_load(
 // CHECK:         %{{.*}}, %[[Iarg:.*]]: index, %[[Jarg:.*]]: index)
 func @dynamic_load(%dynamic : memref<?x?xf32>, %i : index, %j : index) {
-//       CHECK:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
-//       CHECK:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
+//   CHECK-DAG:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
+//   CHECK-DAG:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
 //       CHECK:  %[[ptr:.*]] = llvm.extractvalue %[[ld:.*]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 //  CHECK-NEXT:  %[[st0:.*]] = llvm.extractvalue %[[ld]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 //  CHECK-NEXT:  %[[offI:.*]] = llvm.mul %[[I]], %[[st0]] : i64
 //  CHECK-NEXT:  %[[off1:.*]] = llvm.add %[[offI]], %[[J]] : i64
 //  CHECK-NEXT:  %[[addr:.*]] = llvm.getelementptr %[[ptr]][%[[off1]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 //  CHECK-NEXT:  llvm.load %[[addr]] : !llvm.ptr<f32>
   %0 = memref.load %dynamic[%i, %j] : memref<?x?xf32>
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @prefetch
 // CHECK:         %{{.*}}, %[[Iarg:.*]]: index, %[[Jarg:.*]]: index)
 func @prefetch(%A : memref<?x?xf32>, %i : index, %j : index) {
-//      CHECK:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
-//      CHECK:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
+//      CHECK-DAG:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
+//      CHECK-DAG:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
 //      CHECK:  %[[ptr:.*]] = llvm.extractvalue %[[ld:.*]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 // CHECK-NEXT:  %[[st0:.*]] = llvm.extractvalue %[[ld]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 // CHECK-NEXT:  %[[offI:.*]] = llvm.mul %[[I]], %[[st0]] : i64
 // CHECK-NEXT:  %[[off1:.*]] = llvm.add %[[offI]], %[[J]] : i64
 // CHECK-NEXT:  %[[addr:.*]] = llvm.getelementptr %[[ptr]][%[[off1]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 // CHECK-NEXT:  [[C1:%.*]] = llvm.mlir.constant(1 : i32) : i32
 // CHECK-NEXT:  [[C3:%.*]] = llvm.mlir.constant(3 : i32) : i32
 // CHECK-NEXT:  [[C1_1:%.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK:  "llvm.intr.prefetch"(%{{.*}}, [[C0_2]], [[C2]], [[C0_3]]) : (!llvm.ptr<f32>, i32, i32, i32) -> ()
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @dynamic_store
 // CHECK:         %{{.*}}, %[[Iarg:.*]]: index, %[[Jarg:.*]]: index
 func @dynamic_store(%dynamic : memref<?x?xf32>, %i : index, %j : index, %val : f32) {
-//       CHECK:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
-//       CHECK:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
+//   CHECK-DAG:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
+//   CHECK-DAG:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
 //       CHECK:  %[[ptr:.*]] = llvm.extractvalue %[[ld:.*]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 //  CHECK-NEXT:  %[[st0:.*]] = llvm.extractvalue %[[ld]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 //  CHECK-NEXT:  %[[offI:.*]] = llvm.mul %[[I]], %[[st0]] : i64
 //  CHECK-NEXT:  %[[off1:.*]] = llvm.add %[[offI]], %[[J]] : i64
 //  CHECK-NEXT:  %[[addr:.*]] = llvm.getelementptr %[[ptr]][%[[off1]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 //  CHECK-NEXT:  llvm.store %{{.*}}, %[[addr]] : !llvm.ptr<f32>
   memref.store %val, %dynamic[%i, %j] : memref<?x?xf32>
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @mixed_store
 // CHECK:         %{{.*}}, %[[Iarg:.*]]: index, %[[Jarg:.*]]: index
 func @mixed_store(%mixed : memref<42x?xf32>, %i : index, %j : index, %val : f32) {
-//       CHECK:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
-//       CHECK:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
+//   CHECK-DAG:  %[[I:.*]] = builtin.unrealized_conversion_cast %[[Iarg]]
+//   CHECK-DAG:  %[[J:.*]] = builtin.unrealized_conversion_cast %[[Jarg]]
 //       CHECK:  %[[ptr:.*]] = llvm.extractvalue %[[ld:.*]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 //  CHECK-NEXT:  %[[st0:.*]] = llvm.extractvalue %[[ld]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 //  CHECK-NEXT:  %[[offI:.*]] = llvm.mul %[[I]], %[[st0]] : i64
 //  CHECK-NEXT:  %[[off1:.*]] = llvm.add %[[offI]], %[[J]] : i64
 //  CHECK-NEXT:  %[[addr:.*]] = llvm.getelementptr %[[ptr]][%[[off1]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 //  CHECK-NEXT:  llvm.store %{{.*}}, %[[addr]] : !llvm.ptr<f32>
   memref.store %val, %mixed[%i, %j] : memref<42x?xf32>
   return
// CHECK: llvm.extractvalue %{{.*}}[3, 4] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<5 x i64>, array<5 x i64>)>
   return
 }
 
 // -----
 
 // CHECK-LABEL: @memref_dim_with_dyn_index
 // CHECK: %{{.*}}, %[[IDXarg:.*]]: index
 func @memref_dim_with_dyn_index(%arg : memref<3x?xf32>, %idx : index) -> index {
+  // CHECK-DAG: %[[IDX:.*]] = builtin.unrealized_conversion_cast %[[IDXarg]]
   // CHECK-DAG: %[[C0:.*]] = llvm.mlir.constant(0 : index) : i64
   // CHECK-DAG: %[[C1:.*]] = llvm.mlir.constant(1 : index) : i64
   // CHECK-DAG: %[[SIZES:.*]] = llvm.extractvalue %{{.*}}[3] : ![[DESCR_TY:.*]]
   // CHECK-DAG: %[[SIZES_PTR:.*]] = llvm.alloca %[[C1]] x !llvm.array<2 x i64> : (i64) -> !llvm.ptr<array<2 x i64>>
   // CHECK-DAG: llvm.store %[[SIZES]], %[[SIZES_PTR]] : !llvm.ptr<array<2 x i64>>
-  // CHECK-DAG: %[[IDX:.*]] = builtin.unrealized_conversion_cast %[[IDXarg]]
   // CHECK-DAG: %[[RESULT_PTR:.*]] = llvm.getelementptr %[[SIZES_PTR]][%[[C0]], %[[IDX]]] : (!llvm.ptr<array<2 x i64>>, i64, i64) -> !llvm.ptr<i64>
   // CHECK-DAG: %[[RESULT:.*]] = llvm.load %[[RESULT_PTR]] : !llvm.ptr<i64>
   %result = memref.dim %arg, %idx : memref<3x?xf32>
   return %result : index
 }
 
 // -----
 
%output = memref.reinterpret_cast %input to
            offset: [%offset], sizes: [%size_0, %size_1],
            strides: [%stride_0, %stride_1]
            : memref<*xf32> to memref<?x?xf32, offset: ?, strides: [?, ?]>
   return
 }
 // CHECK-SAME: ([[OFFSETarg:%[a-z,0-9]+]]: index,
 // CHECK-SAME: [[SIZE_0arg:%[a-z,0-9]+]]: index, [[SIZE_1arg:%[a-z,0-9]+]]: index,
 // CHECK-SAME: [[STRIDE_0arg:%[a-z,0-9]+]]: index, [[STRIDE_1arg:%[a-z,0-9]+]]: index,
-// CHECK: [[INPUT:%.*]] = builtin.unrealized_conversion_cast
-// CHECK: [[OFFSET:%.*]] = builtin.unrealized_conversion_cast [[OFFSETarg]]
-// CHECK: [[SIZE_0:%.*]] = builtin.unrealized_conversion_cast [[SIZE_0arg]]
-// CHECK: [[SIZE_1:%.*]] = builtin.unrealized_conversion_cast [[SIZE_1arg]]
-// CHECK: [[STRIDE_0:%.*]] = builtin.unrealized_conversion_cast [[STRIDE_0arg]]
-// CHECK: [[STRIDE_1:%.*]] = builtin.unrealized_conversion_cast [[STRIDE_1arg]]
+// CHECK-DAG: [[OFFSET:%.*]] = builtin.unrealized_conversion_cast [[OFFSETarg]]
+// CHECK-DAG: [[SIZE_0:%.*]] = builtin.unrealized_conversion_cast [[SIZE_0arg]]
+// CHECK-DAG: [[SIZE_1:%.*]] = builtin.unrealized_conversion_cast [[SIZE_1arg]]
+// CHECK-DAG: [[STRIDE_0:%.*]] = builtin.unrealized_conversion_cast [[STRIDE_0arg]]
+// CHECK-DAG: [[STRIDE_1:%.*]] = builtin.unrealized_conversion_cast [[STRIDE_1arg]]
+// CHECK-DAG: [[INPUT:%.*]] = builtin.unrealized_conversion_cast
 // CHECK: [[OUT_0:%.*]] = llvm.mlir.undef : [[TY:!.*]]
 // CHECK: [[DESCRIPTOR:%.*]] = llvm.extractvalue [[INPUT]][1] : !llvm.struct<(i64, ptr<i8>)>
 // CHECK: [[BASE_PTR_PTR:%.*]] = llvm.bitcast [[DESCRIPTOR]] : !llvm.ptr<i8> to !llvm.ptr<ptr<f32>>
 // CHECK: [[BASE_PTR:%.*]] = llvm.load [[BASE_PTR_PTR]] : !llvm.ptr<ptr<f32>>
 // CHECK: [[BASE_PTR_PTR_:%.*]] = llvm.bitcast [[DESCRIPTOR]] : !llvm.ptr<i8> to !llvm.ptr<ptr<f32>>
 // CHECK: [[C1:%.*]] = llvm.mlir.constant(1 : index) : i64
 // CHECK: [[ALIGNED_PTR_PTR:%.*]] = llvm.getelementptr [[BASE_PTR_PTR_]]{{\[}}[[C1]]]
 // CHECK-SAME: : (!llvm.ptr<ptr<f32>>, i64) -> !llvm.ptr<ptr<f32>>

 // -----
 
 // CHECK-LABEL: @memref_reshape
 func @memref_reshape(%input : memref<2x3xf32>, %shape : memref<?xindex>) {
   %output = memref.reshape %input(%shape)
                 : (memref<2x3xf32>, memref<?xindex>) -> memref<*xf32>
   return
 }
-// CHECK: [[INPUT:%.*]] = builtin.unrealized_conversion_cast %{{.*}} to [[INPUT_TY:!.*]]
 // CHECK: [[SHAPE:%.*]] = builtin.unrealized_conversion_cast %{{.*}} to [[SHAPE_TY:!.*]]
+// CHECK: [[INPUT:%.*]] = builtin.unrealized_conversion_cast %{{.*}} to [[INPUT_TY:!.*]]
 // CHECK: [[RANK:%.*]] = llvm.extractvalue [[SHAPE]][3, 0] : [[SHAPE_TY]]
 // CHECK: [[UNRANKED_OUT_O:%.*]] = llvm.mlir.undef : !llvm.struct<(i64, ptr<i8>)>
 // CHECK: [[UNRANKED_OUT_1:%.*]] = llvm.insertvalue [[RANK]], [[UNRANKED_OUT_O]][0] : !llvm.struct<(i64, ptr<i8>)>
 
 // Compute size in bytes to allocate result ranked descriptor
 // CHECK: [[C1:%.*]] = llvm.mlir.constant(1 : index) : i64
 // CHECK: [[C2:%.*]] = llvm.mlir.constant(2 : index) : i64
 // CHECK: [[PTR_SIZE:%.*]] = llvm.mlir.constant(8 : index) : i64

mlir/test/Conversion/MemRefToLLVM/convert-static-memref-ops.mlir

 
 // -----
 
 // CHECK-LABEL: func @static_load
 // CHECK:         %[[MEMREF:.*]]: memref<10x42xf32>,
 // CHECK:         %[[I:.*]]: index,
 // CHECK:         %[[J:.*]]: index)
 func @static_load(%static : memref<10x42xf32>, %i : index, %j : index) {
-// CHECK:  %[[II:.*]] = builtin.unrealized_conversion_cast %[[I]]
 // CHECK:  %[[JJ:.*]] = builtin.unrealized_conversion_cast %[[J]]
+// CHECK:  %[[II:.*]] = builtin.unrealized_conversion_cast %[[I]]
 // CHECK:  %[[ptr:.*]] = llvm.extractvalue %{{.*}}[1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 // CHECK:  %[[st0:.*]] = llvm.mlir.constant(42 : index) : i64
 // CHECK:  %[[offI:.*]] = llvm.mul %[[II]], %[[st0]] : i64
 // CHECK:  %[[off1:.*]] = llvm.add %[[offI]], %[[JJ]] : i64
 // CHECK:  %[[addr:.*]] = llvm.getelementptr %[[ptr]][%[[off1]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 // CHECK:  llvm.load %[[addr]] : !llvm.ptr<f32>
   %0 = memref.load %static[%i, %j] : memref<10x42xf32>
   return

 }
 
 // -----
 
 // CHECK-LABEL: func @static_store
 // CHECK:         %[[MEMREF:.*]]: memref<10x42xf32>,
 // CHECK-SAME:    %[[I:.*]]: index, %[[J:.*]]: index,
 func @static_store(%static : memref<10x42xf32>, %i : index, %j : index, %val : f32) {
-// CHECK: %[[II:.*]] = builtin.unrealized_conversion_cast %[[I]]
 // CHECK: %[[JJ:.*]] = builtin.unrealized_conversion_cast %[[J]]
+// CHECK: %[[II:.*]] = builtin.unrealized_conversion_cast %[[I]]
 // CHECK: %[[ptr:.*]] = llvm.extractvalue %{{.*}}[1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
 // CHECK: %[[st0:.*]] = llvm.mlir.constant(42 : index) : i64
 // CHECK: %[[offI:.*]] = llvm.mul %[[II]], %[[st0]] : i64
 // CHECK: %[[off1:.*]] = llvm.add %[[offI]], %[[JJ]] : i64
 // CHECK: %[[addr:.*]] = llvm.getelementptr %[[ptr]][%[[off1]]] : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
 // CHECK: llvm.store %{{.*}}, %[[addr]] : !llvm.ptr<f32>
 
   memref.store %val, %static[%i, %j] : memref<10x42xf32>

mlir/test/Conversion/MemRefToLLVM/memref-to-llvm.mlir

 // RUN: mlir-opt -convert-memref-to-llvm %s -split-input-file | FileCheck %s
 // RUN: mlir-opt -convert-memref-to-llvm='index-bitwidth=32' %s -split-input-file | FileCheck --check-prefix=CHECK32 %s
 
 
 // CHECK-LABEL: func @view(
 // CHECK: %[[ARG0F:.*]]: index, %[[ARG1F:.*]]: index, %[[ARG2F:.*]]: index
 func @view(%arg0 : index, %arg1 : index, %arg2 : index) {
+  // CHECK: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1F:.*]]
+  // CHECK: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0F:.*]]
+  // CHECK: %[[ARG2:.*]] = builtin.unrealized_conversion_cast %[[ARG2F:.*]]
   // CHECK: llvm.mlir.constant(2048 : index) : i64
   // CHECK: llvm.mlir.undef : !llvm.struct<(ptr<i8>, ptr<i8>, i64, array<1 x i64>, array<1 x i64>)>
   %0 = memref.alloc() : memref<2048xi8>
 
   // Test two dynamic sizes.
-  // CHECK: %[[ARG2:.*]] = builtin.unrealized_conversion_cast %[[ARG2F:.*]]
-  // CHECK: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0F:.*]]
-  // CHECK: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1F:.*]]
   // CHECK: llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BASE_PTR:.*]] = llvm.extractvalue %{{.*}}[1] : !llvm.struct<(ptr<i8>, ptr<i8>, i64, array<1 x i64>, array<1 x i64>)>
   // CHECK: %[[SHIFTED_BASE_PTR:.*]] = llvm.getelementptr %[[BASE_PTR]][%[[ARG2]]] : (!llvm.ptr<i8>, i64) -> !llvm.ptr<i8>
   // CHECK: %[[CAST_SHIFTED_BASE_PTR:.*]] = llvm.bitcast %[[SHIFTED_BASE_PTR]] : !llvm.ptr<i8> to !llvm.ptr<f32>
   // CHECK: llvm.insertvalue %[[CAST_SHIFTED_BASE_PTR]], %{{.*}}[1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[C0:.*]] = llvm.mlir.constant(0 : index) : i64
   // CHECK: llvm.insertvalue %[[C0]], %{{.*}}[2] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.insertvalue %[[ARG1]], %{{.*}}[3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mlir.constant(1 : index) : i64
   // CHECK: llvm.insertvalue %{{.*}}, %{{.*}}[4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.insertvalue %[[ARG0]], %{{.*}}[3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mul %{{.*}}, %[[ARG1]]
   // CHECK: llvm.insertvalue %{{.*}}, %{{.*}}[4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   %1 = memref.view %0[%arg2][%arg0, %arg1] : memref<2048xi8> to memref<?x?xf32>
 
   // Test one dynamic size.
-  // CHECK: %[[ARG2:.*]] = builtin.unrealized_conversion_cast %[[ARG2F:.*]]
-  // CHECK: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1F:.*]]
   // CHECK: llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BASE_PTR_2:.*]] = llvm.extractvalue %{{.*}}[1] : !llvm.struct<(ptr<i8>, ptr<i8>, i64, array<1 x i64>, array<1 x i64>)>
   // CHECK: %[[SHIFTED_BASE_PTR_2:.*]] = llvm.getelementptr %[[BASE_PTR_2]][%[[ARG2]]] : (!llvm.ptr<i8>, i64) -> !llvm.ptr<i8>
   // CHECK: %[[CAST_SHIFTED_BASE_PTR_2:.*]] = llvm.bitcast %[[SHIFTED_BASE_PTR_2]] : !llvm.ptr<i8> to !llvm.ptr<f32>
   // CHECK: llvm.insertvalue %[[CAST_SHIFTED_BASE_PTR_2]], %{{.*}}[1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[C0_2:.*]] = llvm.mlir.constant(0 : index) : i64
   // CHECK: llvm.insertvalue %[[C0_2]], %{{.*}}[2] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.insertvalue %[[ARG1]], %{{.*}}[3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mlir.constant(1 : index) : i64
   // CHECK: llvm.insertvalue %{{.*}}, %{{.*}}[4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mlir.constant(4 : index) : i64
   // CHECK: llvm.insertvalue %{{.*}}, %{{.*}}[3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mul %{{.*}}, %[[ARG1]]
   // CHECK: llvm.insertvalue %{{.*}}, %{{.*}}[4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   %3 = memref.view %0[%arg2][%arg1] : memref<2048xi8> to memref<4x?xf32>
 
   // Test static sizes.
-  // CHECK: %[[ARG2:.*]] = builtin.unrealized_conversion_cast %[[ARG2F:.*]]
   // CHECK: llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BASE_PTR_3:.*]] = llvm.extractvalue %{{.*}}[1] : !llvm.struct<(ptr<i8>, ptr<i8>, i64, array<1 x i64>, array<1 x i64>)>
   // CHECK: %[[SHIFTED_BASE_PTR_3:.*]] = llvm.getelementptr %[[BASE_PTR_3]][%[[ARG2]]] : (!llvm.ptr<i8>, i64) -> !llvm.ptr<i8>
   // CHECK: %[[CAST_SHIFTED_BASE_PTR_3:.*]] = llvm.bitcast %[[SHIFTED_BASE_PTR_3]] : !llvm.ptr<i8> to !llvm.ptr<f32>
   // CHECK: llvm.insertvalue %[[CAST_SHIFTED_BASE_PTR_3]], %{{.*}}[1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[C0_3:.*]] = llvm.mlir.constant(0 : index) : i64
   // CHECK: llvm.insertvalue %[[C0_3]], %{{.*}}[2] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mlir.constant(4 : index) : i64
   // CHECK: llvm.insertvalue %{{.*}}, %{{.*}}[3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mlir.constant(1 : index) : i64
   // CHECK: llvm.insertvalue %{{.*}}, %{{.*}}[4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mlir.constant(64 : index) : i64
   // CHECK: llvm.insertvalue %{{.*}}, %{{.*}}[3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mlir.constant(4 : index) : i64
   // CHECK: llvm.insertvalue %{{.*}}, %{{.*}}[4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   %5 = memref.view %0[%arg2][] : memref<2048xi8> to memref<64x4xf32>
 
   // Test view memory space.
   // CHECK: llvm.mlir.constant(2048 : index) : i64
   // CHECK: llvm.mlir.undef : !llvm.struct<(ptr<i8, 4>, ptr<i8, 4>, i64, array<1 x i64>, array<1 x i64>)>
   %6 = memref.alloc() : memref<2048xi8, 4>
 
-  // CHECK: %[[ARG2:.*]] = builtin.unrealized_conversion_cast %[[ARG2F:.*]]
   // CHECK: llvm.mlir.undef : !llvm.struct<(ptr<f32, 4>, ptr<f32, 4>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BASE_PTR_4:.*]] = llvm.extractvalue %{{.*}}[1] : !llvm.struct<(ptr<i8, 4>, ptr<i8, 4>, i64, array<1 x i64>, array<1 x i64>)>
   // CHECK: %[[SHIFTED_BASE_PTR_4:.*]] = llvm.getelementptr %[[BASE_PTR_4]][%[[ARG2]]] : (!llvm.ptr<i8, 4>, i64) -> !llvm.ptr<i8, 4>
   // CHECK: %[[CAST_SHIFTED_BASE_PTR_4:.*]] = llvm.bitcast %[[SHIFTED_BASE_PTR_4]] : !llvm.ptr<i8, 4> to !llvm.ptr<f32, 4>
   // CHECK: llvm.insertvalue %[[CAST_SHIFTED_BASE_PTR_4]], %{{.*}}[1] : !llvm.struct<(ptr<f32, 4>, ptr<f32, 4>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[C0_4:.*]] = llvm.mlir.constant(0 : index) : i64
   // CHECK: llvm.insertvalue %[[C0_4]], %{{.*}}[2] : !llvm.struct<(ptr<f32, 4>, ptr<f32, 4>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.mlir.constant(4 : index) : i64

 // CHECK:         %[[ARG1f:[a-zA-Z0-9]*]]: index,
 // CHECK:         %[[ARG2f:.*]]: index)
 // CHECK32-LABEL: func @subview(
 // CHECK32:         %[[MEM:.*]]: memref<{{.*}}>,
 // CHECK32:         %[[ARG0f:[a-zA-Z0-9]*]]: index,
 // CHECK32:         %[[ARG1f:[a-zA-Z0-9]*]]: index,
 // CHECK32:         %[[ARG2f:.*]]: index)
 func @subview(%0 : memref<64x4xf32, offset: 0, strides: [4, 1]>, %arg0 : index, %arg1 : index, %arg2 : index) {
-  // CHECK: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
-  // CHECK32: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
-
-  // CHECK: %[[ARG0a:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1a:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK: %[[ARG0b:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1b:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK: %[[ARG0c:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1c:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0a:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1a:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0b:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1b:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0c:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1c:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
+  // CHECK-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
+  // CHECK-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
+  // CHECK-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
+
+  // CHECK32-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
+  // CHECK32-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
+  // CHECK32-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
 
   // CHECK: %[[DESC:.*]] = llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BITCAST0:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BITCAST0]], %[[DESC]][0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BITCAST1:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BITCAST1]], %[[DESC0]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEMREF]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[OFF:.*]] = llvm.extractvalue %[[MEMREF]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
-  // CHECK: %[[OFFINC:.*]] = llvm.mul %[[ARG0a]], %[[STRIDE0]] : i64
+  // CHECK: %[[OFFINC:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i64
   // CHECK: %[[OFF1:.*]] = llvm.add %[[OFF]], %[[OFFINC]] : i64
-  // CHECK: %[[OFFINC1:.*]] = llvm.mul %[[ARG1a]], %[[STRIDE1]] : i64
+  // CHECK: %[[OFFINC1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i64
   // CHECK: %[[OFF2:.*]] = llvm.add %[[OFF1]], %[[OFFINC1]] : i64
   // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
-  // CHECK: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1c]], %[[STRIDE1]] : i64
-  // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1b]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
+  // CHECK: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i64
+  // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE1]], %[[DESC3]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
-  // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0c]], %[[STRIDE0]] : i64
-  // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0b]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
+  // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i64
+  // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.insertvalue %[[DESCSTRIDE0]], %[[DESC5]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK32: %[[DESC:.*]] = llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST0:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK32: %[[DESC0:.*]] = llvm.insertvalue %[[BITCAST0]], %[[DESC]][0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST1:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK32: %[[DESC1:.*]] = llvm.insertvalue %[[BITCAST1]], %[[DESC0]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEMREF]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[OFF:.*]] = llvm.extractvalue %[[MEMREF]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
-  // CHECK32: %[[OFFINC:.*]] = llvm.mul %[[ARG0a]], %[[STRIDE0]] : i32
+  // CHECK32: %[[OFFINC:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i32
   // CHECK32: %[[OFF1:.*]] = llvm.add %[[OFF]], %[[OFFINC]] : i32
-  // CHECK32: %[[OFFINC1:.*]] = llvm.mul %[[ARG1a]], %[[STRIDE1]] : i32
+  // CHECK32: %[[OFFINC1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i32
   // CHECK32: %[[OFF2:.*]] = llvm.add %[[OFF1]], %[[OFFINC1]] : i32
   // CHECK32: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
-  // CHECK32: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1c]], %[[STRIDE1]] : i32
-  // CHECK32: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1b]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
+  // CHECK32: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i32
+  // CHECK32: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE1]], %[[DESC3]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
-  // CHECK32: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0c]], %[[STRIDE0]] : i32
-  // CHECK32: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0b]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
+  // CHECK32: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i32
+  // CHECK32: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
 
   %1 = memref.subview %0[%arg0, %arg1][%arg0, %arg1][%arg0, %arg1] :
     memref<64x4xf32, offset: 0, strides: [4, 1]>
   to memref<?x?xf32, offset: ?, strides: [?, ?]>
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @subview_non_zero_addrspace(
 // CHECK:         %[[MEM:.*]]: memref<{{.*}}>,
 // CHECK:         %[[ARG0f:[a-zA-Z0-9]*]]: index,
 // CHECK:         %[[ARG1f:[a-zA-Z0-9]*]]: index,
 // CHECK:         %[[ARG2f:.*]]: index)
 // CHECK32-LABEL: func @subview_non_zero_addrspace(
 // CHECK32:         %[[MEM:.*]]: memref<{{.*}}>,
 // CHECK32:         %[[ARG0f:[a-zA-Z0-9]*]]: index,
 // CHECK32:         %[[ARG1f:[a-zA-Z0-9]*]]: index,
 // CHECK32:         %[[ARG2f:.*]]: index)
 func @subview_non_zero_addrspace(%0 : memref<64x4xf32, offset: 0, strides: [4, 1], 3>, %arg0 : index, %arg1 : index, %arg2 : index) {
-  // CHECK: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
-  // CHECK32: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
-
-  // CHECK: %[[ARG0a:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1a:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK: %[[ARG0b:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1b:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK: %[[ARG0c:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1c:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0a:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1a:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0b:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1b:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0c:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1c:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
+  // CHECK-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
+  // CHECK-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
+  // CHECK-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
+  // CHECK32-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
+  // CHECK32-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
+  // CHECK32-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
 
   // CHECK: %[[DESC:.*]] = llvm.mlir.undef : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BITCAST0:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32, 3> to !llvm.ptr<f32, 3>
   // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BITCAST0]], %[[DESC]][0] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BITCAST1:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32, 3> to !llvm.ptr<f32, 3>
   // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BITCAST1]], %[[DESC0]][1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEMREF]][4, 1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[OFF:.*]] = llvm.extractvalue %[[MEMREF]][2] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
-  // CHECK: %[[OFFINC:.*]] = llvm.mul %[[ARG0a]], %[[STRIDE0]] : i64
+  // CHECK: %[[OFFINC:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i64
   // CHECK: %[[OFF1:.*]] = llvm.add %[[OFF]], %[[OFFINC]] : i64
-  // CHECK: %[[OFFINC1:.*]] = llvm.mul %[[ARG1a]], %[[STRIDE1]] : i64
+  // CHECK: %[[OFFINC1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i64
   // CHECK: %[[OFF2:.*]] = llvm.add %[[OFF1]], %[[OFFINC1]] : i64
   // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
-  // CHECK: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1c]], %[[STRIDE1]] : i64
-  // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1b]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
+  // CHECK: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i64
+  // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE1]], %[[DESC3]][4, 1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
-  // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0c]], %[[STRIDE0]] : i64
-  // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0b]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
+  // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i64
+  // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.insertvalue %[[DESCSTRIDE0]], %[[DESC5]][4, 0] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK32: %[[DESC:.*]] = llvm.mlir.undef : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST0:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32, 3> to !llvm.ptr<f32, 3>
   // CHECK32: %[[DESC0:.*]] = llvm.insertvalue %[[BITCAST0]], %[[DESC]][0] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST1:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32, 3> to !llvm.ptr<f32, 3>
   // CHECK32: %[[DESC1:.*]] = llvm.insertvalue %[[BITCAST1]], %[[DESC0]][1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEMREF]][4, 1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[OFF:.*]] = llvm.extractvalue %[[MEMREF]][2] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
-  // CHECK32: %[[OFFINC:.*]] = llvm.mul %[[ARG0a]], %[[STRIDE0]] : i32
+  // CHECK32: %[[OFFINC:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i32
   // CHECK32: %[[OFF1:.*]] = llvm.add %[[OFF]], %[[OFFINC]] : i32
-  // CHECK32: %[[OFFINC1:.*]] = llvm.mul %[[ARG1a]], %[[STRIDE1]] : i32
+  // CHECK32: %[[OFFINC1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i32
   // CHECK32: %[[OFF2:.*]] = llvm.add %[[OFF1]], %[[OFFINC1]] : i32
   // CHECK32: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
-  // CHECK32: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1c]], %[[STRIDE1]] : i32
-  // CHECK32: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1b]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
+  // CHECK32: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i32
+  // CHECK32: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE1]], %[[DESC3]][4, 1] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
-  // CHECK32: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0c]], %[[STRIDE0]] : i32
-  // CHECK32: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0b]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
+  // CHECK32: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i32
+  // CHECK32: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32, 3>, ptr<f32, 3>, i32, array<2 x i32>, array<2 x i32>)>
 
   %1 = memref.subview %0[%arg0, %arg1][%arg0, %arg1][%arg0, %arg1] :
     memref<64x4xf32, offset: 0, strides: [4, 1], 3>
     to memref<?x?xf32, offset: ?, strides: [?, ?], 3>
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @subview_const_size(
 // CHECK-SAME:         %[[MEM:.*]]: memref<{{.*}}>,
 // CHECK-SAME:         %[[ARG0f:[a-zA-Z0-9]*]]: index
 // CHECK-SAME:         %[[ARG1f:[a-zA-Z0-9]*]]: index
 // CHECK-SAME:         %[[ARG2f:[a-zA-Z0-9]*]]: index
 // CHECK32-LABEL: func @subview_const_size(
 // CHECK32-SAME:         %[[MEM:.*]]: memref<{{.*}}>,
 // CHECK32-SAME:         %[[ARG0f:[a-zA-Z0-9]*]]: index
 // CHECK32-SAME:         %[[ARG1f:[a-zA-Z0-9]*]]: index
 // CHECK32-SAME:         %[[ARG2f:[a-zA-Z0-9]*]]: index
 func @subview_const_size(%0 : memref<64x4xf32, offset: 0, strides: [4, 1]>, %arg0 : index, %arg1 : index, %arg2 : index) {
-  // CHECK: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
-  // CHECK32: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
-
-  // CHECK: %[[ARG0a:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1a:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK: %[[ARG0b:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1b:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0a:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1a:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0b:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1b:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
+  // CHECK-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
+  // CHECK-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
+  // CHECK-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
+  // CHECK32-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
+  // CHECK32-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
+  // CHECK32-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
 
   // CHECK: %[[DESC:.*]] = llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BITCAST0:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BITCAST0]], %[[DESC]][0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BITCAST1:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BITCAST1]], %[[DESC0]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEMREF]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[OFF:.*]] = llvm.extractvalue %[[MEMREF]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
-  // CHECK: %[[OFFINC:.*]] = llvm.mul %[[ARG0a]], %[[STRIDE0]] : i64
+  // CHECK: %[[OFFINC:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i64
   // CHECK: %[[OFF1:.*]] = llvm.add %[[OFF]], %[[OFFINC]] : i64
-  // CHECK: %[[OFFINC1:.*]] = llvm.mul %[[ARG1a]], %[[STRIDE1]] : i64
+  // CHECK: %[[OFFINC1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i64
   // CHECK: %[[OFF2:.*]] = llvm.add %[[OFF1]], %[[OFFINC1]] : i64
   // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[CST2:.*]] = llvm.mlir.constant(2 : i64)
-  // CHECK: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1b]], %[[STRIDE1]] : i64
+  // CHECK: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i64
   // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[CST2]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE1]], %[[DESC3]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[CST4:.*]] = llvm.mlir.constant(4 : i64)
-  // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0b]], %[[STRIDE0]] : i64
+  // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i64
   // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[CST4]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.insertvalue %[[DESCSTRIDE0]], %[[DESC5]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK32: %[[DESC:.*]] = llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST0:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK32: %[[DESC0:.*]] = llvm.insertvalue %[[BITCAST0]], %[[DESC]][0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST1:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK32: %[[DESC1:.*]] = llvm.insertvalue %[[BITCAST1]], %[[DESC0]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEMREF]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[OFF:.*]] = llvm.extractvalue %[[MEMREF]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
-  // CHECK32: %[[OFFINC:.*]] = llvm.mul %[[ARG0a]], %[[STRIDE0]] : i32
+  // CHECK32: %[[OFFINC:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i32
   // CHECK32: %[[OFF1:.*]] = llvm.add %[[OFF]], %[[OFFINC]] : i32
-  // CHECK32: %[[OFFINC1:.*]] = llvm.mul %[[ARG1a]], %[[STRIDE1]] : i32
+  // CHECK32: %[[OFFINC1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i32
   // CHECK32: %[[OFF2:.*]] = llvm.add %[[OFF1]], %[[OFFINC1]] : i32
   // CHECK32: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[CST2:.*]] = llvm.mlir.constant(2 : i64)
-  // CHECK32: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1b]], %[[STRIDE1]] : i32
+  // CHECK32: %[[DESCSTRIDE1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i32
   // CHECK32: %[[DESC3:.*]] = llvm.insertvalue %[[CST2]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE1]], %[[DESC3]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[CST4:.*]] = llvm.mlir.constant(4 : i64)
-  // CHECK32: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0b]], %[[STRIDE0]] : i32
+  // CHECK32: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i32
   // CHECK32: %[[DESC5:.*]] = llvm.insertvalue %[[CST4]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: llvm.insertvalue %[[DESCSTRIDE0]], %[[DESC5]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   %1 = memref.subview %0[%arg0, %arg1][4, 2][%arg0, %arg1] :
     memref<64x4xf32, offset: 0, strides: [4, 1]>
     to memref<4x2xf32, offset: ?, strides: [?, ?]>
   return
 }
 
 // -----
 
 // CHECK-LABEL: func @subview_const_stride(
 // CHECK-SAME:         %[[MEM:.*]]: memref<{{.*}}>,
 // CHECK-SAME:         %[[ARG0f:[a-zA-Z0-9]*]]: index
 // CHECK-SAME:         %[[ARG1f:[a-zA-Z0-9]*]]: index
 // CHECK-SAME:         %[[ARG2f:[a-zA-Z0-9]*]]: index
 // CHECK32-LABEL: func @subview_const_stride(
 // CHECK32-SAME:         %[[MEM:.*]]: memref<{{.*}}>,
 // CHECK32-SAME:         %[[ARG0f:[a-zA-Z0-9]*]]: index
 // CHECK32-SAME:         %[[ARG1f:[a-zA-Z0-9]*]]: index
 // CHECK32-SAME:         %[[ARG2f:[a-zA-Z0-9]*]]: index
 func @subview_const_stride(%0 : memref<64x4xf32, offset: 0, strides: [4, 1]>, %arg0 : index, %arg1 : index, %arg2 : index) {
-  // CHECK: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
-  // CHECK32: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
-
-  // CHECK: %[[ARG0a:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1a:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK: %[[ARG0b:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK: %[[ARG1b:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0a:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1a:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG0b:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
-  // CHECK32: %[[ARG1b:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
+  // CHECK-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
+  // CHECK-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
+  // CHECK-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
+  // CHECK32-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
+  // CHECK32-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
+  // CHECK32-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
 
   // CHECK: %[[DESC:.*]] = llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BITCAST0:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BITCAST0]], %[[DESC]][0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[BITCAST1:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BITCAST1]], %[[DESC0]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEMREF]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[OFF:.*]] = llvm.extractvalue %[[MEMREF]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
-  // CHECK: %[[OFFINC:.*]] = llvm.mul %[[ARG0a]], %[[STRIDE0]] : i64
+  // CHECK: %[[OFFINC:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i64
   // CHECK: %[[OFF1:.*]] = llvm.add %[[OFF]], %[[OFFINC]] : i64
-  // CHECK: %[[OFFINC1:.*]] = llvm.mul %[[ARG1a]], %[[STRIDE1]] : i64
+  // CHECK: %[[OFFINC1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i64
   // CHECK: %[[OFF2:.*]] = llvm.add %[[OFF1]], %[[OFFINC1]] : i64
   // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[CST2:.*]] = llvm.mlir.constant(2 : i64)
-  // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1b]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
+  // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[CST2]], %[[DESC3]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: %[[CST4:.*]] = llvm.mlir.constant(4 : i64)
-  // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0b]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
+  // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK: llvm.insertvalue %[[CST4]], %[[DESC5]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>
   // CHECK32: %[[DESC:.*]] = llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST0:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK32: %[[DESC0:.*]] = llvm.insertvalue %[[BITCAST0]], %[[DESC]][0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST1:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK32: %[[DESC1:.*]] = llvm.insertvalue %[[BITCAST1]], %[[DESC0]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEMREF]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[OFF:.*]] = llvm.extractvalue %[[MEMREF]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
-  // CHECK32: %[[OFFINC:.*]] = llvm.mul %[[ARG0a]], %[[STRIDE0]] : i32
+  // CHECK32: %[[OFFINC:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] : i32
   // CHECK32: %[[OFF1:.*]] = llvm.add %[[OFF]], %[[OFFINC]] : i32
-  // CHECK32: %[[OFFINC1:.*]] = llvm.mul %[[ARG1a]], %[[STRIDE1]] : i32
+  // CHECK32: %[[OFFINC1:.*]] = llvm.mul %[[ARG1]], %[[STRIDE1]] : i32
   // CHECK32: %[[OFF2:.*]] = llvm.add %[[OFF1]], %[[OFFINC1]] : i32
   // CHECK32: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[CST2:.*]] = llvm.mlir.constant(2 : i64)
-  // CHECK32: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1b]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
+  // CHECK32: %[[DESC3:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC2]][3, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[DESC4:.*]] = llvm.insertvalue %[[CST2]], %[[DESC3]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[CST4:.*]] = llvm.mlir.constant(4 : i64)
-  // CHECK32: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0b]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
+  // CHECK32: %[[DESC5:.*]] = llvm.insertvalue %[[ARG0]], %[[DESC4]][3, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: llvm.insertvalue %[[CST4]], %[[DESC5]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   %1 = memref.subview %0[%arg0, %arg1][%arg0, %arg1][1, 2] :
     memref<64x4xf32, offset: 0, strides: [4, 1]>
     to memref<?x?xf32, offset: ?, strides: [4, 2]>
   return
 }
 
 // -----

 // CHECK:         %[[ARG1f:[a-zA-Z0-9]*]]: index,
 // CHECK:         %[[ARG2f:.*]]: index)
 // CHECK32-LABEL: func @subview_mixed_static_dynamic(
 // CHECK32:         %[[MEM:.*]]: memref<{{.*}}>,
 // CHECK32:         %[[ARG0f:[a-zA-Z0-9]*]]: index,
 // CHECK32:         %[[ARG1f:[a-zA-Z0-9]*]]: index,
 // CHECK32:         %[[ARG2f:.*]]: index)
 func @subview_mixed_static_dynamic(%0 : memref<64x4xf32, offset: 0, strides: [4, 1]>, %arg0 : index, %arg1 : index, %arg2 : index) {
-  // CHECK32: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
-  // CHECK32: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
-  // CHECK32: %[[ARG2:.*]] = builtin.unrealized_conversion_cast %[[ARG2f]]
-  // CHECK32: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
+  // CHECK32-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]
+  // CHECK32-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]
+  // CHECK32-DAG: %[[ARG2:.*]] = builtin.unrealized_conversion_cast %[[ARG2f]]
+  // CHECK32-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]
 
   // CHECK32: %[[DESC:.*]] = llvm.mlir.undef : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST0:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK32: %[[DESC0:.*]] = llvm.insertvalue %[[BITCAST0]], %[[DESC]][0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[BITCAST1:.*]] = llvm.bitcast %{{.*}} : !llvm.ptr<f32> to !llvm.ptr<f32>
   // CHECK32: %[[DESC1:.*]] = llvm.insertvalue %[[BITCAST1]], %[[DESC0]][1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>
   // CHECK32: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEMREF]][4, 1] : !llvm.struct<(ptr<f32>, ptr<f32>, i32, array<2 x i32>, array<2 x i32>)>

mlir/test/Conversion/MemRefToSPIRV/memref-to-spirv.mlir

#spv.vce<v1.0,
         StorageBuffer16BitAccess, StorageUniform16, StoragePushConstant16,
         StorageBuffer8BitAccess, UniformAndStorageBuffer8BitAccess, StoragePushConstant8
       ],
       [SPV_KHR_16bit_storage, SPV_KHR_8bit_storage, SPV_KHR_storage_buffer_storage_class]>, {}>
 } {
 
 // CHECK-LABEL: @load_store_zero_rank_float
 func @load_store_zero_rank_float(%arg0: memref<f32>, %arg1: memref<f32>) {
+  //      CHECK: [[ARG1:%.*]] = builtin.unrealized_conversion_cast {{.+}} : memref<f32> to !spv.ptr<!spv.struct<(!spv.array<1 x f32, stride=4> [0])>, StorageBuffer>
   //      CHECK: [[ARG0:%.*]] = builtin.unrealized_conversion_cast {{.+}} : memref<f32> to !spv.ptr<!spv.struct<(!spv.array<1 x f32, stride=4> [0])>, StorageBuffer>
   //      CHECK: [[ZERO1:%.*]] = spv.Constant 0 : i32
   //      CHECK: spv.AccessChain [[ARG0]][
   // CHECK-SAME: [[ZERO1]], [[ZERO1]]
   // CHECK-SAME: ] :
   //      CHECK: spv.Load "StorageBuffer" %{{.*}} : f32
   %0 = memref.load %arg0[] : memref<f32>
-  //      CHECK: [[ARG1:%.*]] = builtin.unrealized_conversion_cast {{.+}} : memref<f32> to !spv.ptr<!spv.struct<(!spv.array<1 x f32, stride=4> [0])>, StorageBuffer>
   //      CHECK: [[ZERO2:%.*]] = spv.Constant 0 : i32
   //      CHECK: spv.AccessChain [[ARG1]][
   // CHECK-SAME: [[ZERO2]], [[ZERO2]]
   // CHECK-SAME: ] :
   //      CHECK: spv.Store "StorageBuffer" %{{.*}} : f32
   memref.store %0, %arg1[] : memref<f32>
   return
 }
 
 // CHECK-LABEL: @load_store_zero_rank_int
 func @load_store_zero_rank_int(%arg0: memref<i32>, %arg1: memref<i32>) {
+  //      CHECK: [[ARG1:%.*]] = builtin.unrealized_conversion_cast {{.+}} : memref<i32> to !spv.ptr<!spv.struct<(!spv.array<1 x i32, stride=4> [0])>, StorageBuffer>
   //      CHECK: [[ARG0:%.*]] = builtin.unrealized_conversion_cast {{.+}} : memref<i32> to !spv.ptr<!spv.struct<(!spv.array<1 x i32, stride=4> [0])>, StorageBuffer>
   //      CHECK: [[ZERO1:%.*]] = spv.Constant 0 : i32
   //      CHECK: spv.AccessChain [[ARG0]][
   // CHECK-SAME: [[ZERO1]], [[ZERO1]]
   // CHECK-SAME: ] :
   //      CHECK: spv.Load "StorageBuffer" %{{.*}} : i32
   %0 = memref.load %arg0[] : memref<i32>
-  //      CHECK: [[ARG1:%.*]] = builtin.unrealized_conversion_cast {{.+}} : memref<i32> to !spv.ptr<!spv.struct<(!spv.array<1 x i32, stride=4> [0])>, StorageBuffer>
   //      CHECK: [[ZERO2:%.*]] = spv.Constant 0 : i32
   //      CHECK: spv.AccessChain [[ARG1]][
   // CHECK-SAME: [[ZERO2]], [[ZERO2]]
   // CHECK-SAME: ] :
   //      CHECK: spv.Store "StorageBuffer" %{{.*}} : i32
   memref.store %0, %arg1[] : memref<i32>
   return
 }
 
 // CHECK-LABEL: func @load_store_unknown_dim
 func @load_store_unknown_dim(%i: index, %source: memref<?xi32>, %dest: memref<?xi32>) {
+  // CHECK: %[[DST:.+]] = builtin.unrealized_conversion_cast {{.+}} : memref<?xi32> to !spv.ptr<!spv.struct<(!spv.rtarray<i32, stride=4> [0])>, StorageBuffer>
   // CHECK: %[[SRC:.+]] = builtin.unrealized_conversion_cast {{.+}} : memref<?xi32> to !spv.ptr<!spv.struct<(!spv.rtarray<i32, stride=4> [0])>, StorageBuffer>
   // CHECK: %[[AC0:.+]] = spv.AccessChain %[[SRC]]
   // CHECK: spv.Load "StorageBuffer" %[[AC0]]
   %0 = memref.load %source[%i] : memref<?xi32>
-  // CHECK: %[[DST:.+]] = builtin.unrealized_conversion_cast {{.+}} : memref<?xi32> to !spv.ptr<!spv.struct<(!spv.rtarray<i32, stride=4> [0])>, StorageBuffer>
   // CHECK: %[[AC1:.+]] = spv.AccessChain %[[DST]]
   // CHECK: spv.Store "StorageBuffer" %[[AC1]]
   memref.store %0, %dest[%i]: memref<?xi32>
   return
 }
 
 // CHECK-LABEL: func @load_i1
 //  CHECK-SAME: (%[[SRC:.+]]: memref<4xi1>, %[[IDX:.+]]: index)
 func @load_i1(%src: memref<4xi1>, %i : index) -> i1 {
-  // CHECK: %[[SRC_CAST:.+]] = builtin.unrealized_conversion_cast %[[SRC]] : memref<4xi1> to !spv.ptr<!spv.struct<(!spv.array<4 x i8, stride=1> [0])>, StorageBuffer>
-  // CHECK: %[[IDX_CAST:.+]] = builtin.unrealized_conversion_cast %[[IDX]]
+  // CHECK-DAG: %[[SRC_CAST:.+]] = builtin.unrealized_conversion_cast %[[SRC]] : memref<4xi1> to !spv.ptr<!spv.struct<(!spv.array<4 x i8, stride=1> [0])>, StorageBuffer>
+  // CHECK-DAG: %[[IDX_CAST:.+]] = builtin.unrealized_conversion_cast %[[IDX]]
   // CHECK: %[[ZERO_0:.+]] = spv.Constant 0 : i32
   // CHECK: %[[ZERO_1:.+]] = spv.Constant 0 : i32
   // CHECK: %[[ONE:.+]] = spv.Constant 1 : i32
   // CHECK: %[[MUL:.+]] = spv.IMul %[[ONE]], %[[IDX_CAST]] : i32
   // CHECK: %[[ADD:.+]] = spv.IAdd %[[ZERO_1]], %[[MUL]] : i32
   // CHECK: %[[ADDR:.+]] = spv.AccessChain %[[SRC_CAST]][%[[ZERO_0]], %[[ADD]]]
   // CHECK: %[[VAL:.+]] = spv.Load "StorageBuffer" %[[ADDR]] : i8
   // CHECK: %[[ONE_I8:.+]] = spv.Constant 1 : i8
   // CHECK: %[[BOOL:.+]] = spv.IEqual %[[VAL]], %[[ONE_I8]] : i8
   %0 = memref.load %src[%i] : memref<4xi1>
   // CHECK: return %[[BOOL]]
   return %0: i1
 }
 
 // CHECK-LABEL: func @store_i1
 //  CHECK-SAME: %[[DST:.+]]: memref<4xi1>,
 //  CHECK-SAME: %[[IDX:.+]]: index
 func @store_i1(%dst: memref<4xi1>, %i: index) {
   %true = arith.constant true
-  // CHECK: %[[DST_CAST:.+]] = builtin.unrealized_conversion_cast %[[DST]] : memref<4xi1> to !spv.ptr<!spv.struct<(!spv.array<4 x i8, stride=1> [0])>, StorageBuffer>
-  // CHECK: %[[IDX_CAST:.+]] = builtin.unrealized_conversion_cast %[[IDX]]
+  // CHECK-DAG: %[[DST_CAST:.+]] = builtin.unrealized_conversion_cast %[[DST]] : memref<4xi1> to !spv.ptr<!spv.struct<(!spv.array<4 x i8, stride=1> [0])>, StorageBuffer>
+  // CHECK-DAG: %[[IDX_CAST:.+]] = builtin.unrealized_conversion_cast %[[IDX]]
   // CHECK: %[[ZERO_0:.+]] = spv.Constant 0 : i32
   // CHECK: %[[ZERO_1:.+]] = spv.Constant 0 : i32
   // CHECK: %[[ONE:.+]] = spv.Constant 1 : i32
   // CHECK: %[[MUL:.+]] = spv.IMul %[[ONE]], %[[IDX_CAST]] : i32
   // CHECK: %[[ADD:.+]] = spv.IAdd %[[ZERO_1]], %[[MUL]] : i32
   // CHECK: %[[ADDR:.+]] = spv.AccessChain %[[DST_CAST]][%[[ZERO_0]], %[[ADD]]]
   // CHECK: %[[ZERO_I8:.+]] = spv.Constant 0 : i8
   // CHECK: %[[ONE_I8:.+]] = spv.Constant 1 : i8
func @store_i1(%arg0: memref<i1>, %value: i1) {
   //     CHECK: spv.AtomicOr "Device" "AcquireRelease" %[[PTR]], %[[STORE_VAL]]
   memref.store %value, %arg0[] : memref<i1>
   return
 }
 
 // CHECK-LABEL: @store_i8
 //       CHECK: (%[[ARG0:.+]]: {{.*}}, %[[ARG1:.+]]: i8)
 func @store_i8(%arg0: memref<i8>, %value: i8) {
-  //     CHECK: %[[ARG1_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG1]] : i8 to i32
-  //     CHECK: %[[ARG0_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG0]]
+  //     CHECK-DAG: %[[ARG1_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG1]] : i8 to i32
+  //     CHECK-DAG: %[[ARG0_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG0]]
   //     CHECK: %[[ZERO:.+]] = spv.Constant 0 : i32
   //     CHECK: %[[FOUR:.+]] = spv.Constant 4 : i32
   //     CHECK: %[[EIGHT:.+]] = spv.Constant 8 : i32
   //     CHECK: %[[IDX:.+]] = spv.UMod %[[ZERO]], %[[FOUR]] : i32
   //     CHECK: %[[OFFSET:.+]] = spv.IMul %[[IDX]], %[[EIGHT]] : i32
   //     CHECK: %[[MASK1:.+]] = spv.Constant 255 : i32
   //     CHECK: %[[TMP1:.+]] = spv.ShiftLeftLogical %[[MASK1]], %[[OFFSET]] : i32, i32
   //     CHECK: %[[MASK:.+]] = spv.Not %[[TMP1]] : i32
   //     CHECK: %[[CLAMPED_VAL:.+]] = spv.BitwiseAnd %[[ARG1_CAST]], %[[MASK1]] : i32
   //     CHECK: %[[STORE_VAL:.+]] = spv.ShiftLeftLogical %[[CLAMPED_VAL]], %[[OFFSET]] : i32, i32
   //     CHECK: %[[FOUR2:.+]] = spv.Constant 4 : i32
   //     CHECK: %[[ACCESS_IDX:.+]] = spv.SDiv %[[ZERO]], %[[FOUR2]] : i32
   //     CHECK: %[[PTR:.+]] = spv.AccessChain %[[ARG0_CAST]][%[[ZERO]], %[[ACCESS_IDX]]]
   //     CHECK: spv.AtomicAnd "Device" "AcquireRelease" %[[PTR]], %[[MASK]]
   //     CHECK: spv.AtomicOr "Device" "AcquireRelease" %[[PTR]], %[[STORE_VAL]]
   memref.store %value, %arg0[] : memref<i8>
   return
 }
 
 // CHECK-LABEL: @store_i16
 //       CHECK: (%[[ARG0:.+]]: memref<10xi16>, %[[ARG1:.+]]: index, %[[ARG2:.+]]: i16)
 func @store_i16(%arg0: memref<10xi16>, %index: index, %value: i16) {
-  //     CHECK: %[[ARG2_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG2]] : i16 to i32
-  //     CHECK: %[[ARG0_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG0]]
-  //     CHECK: %[[ARG1_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG1]] : index to i32
+  //     CHECK-DAG: %[[ARG2_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG2]] : i16 to i32
+  //     CHECK-DAG: %[[ARG0_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG0]]
+  //     CHECK-DAG: %[[ARG1_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG1]] : index to i32
   //     CHECK: %[[ZERO:.+]] = spv.Constant 0 : i32
   //     CHECK: %[[OFFSET:.+]] = spv.Constant 0 : i32
   //     CHECK: %[[ONE:.+]] = spv.Constant 1 : i32
   //     CHECK: %[[UPDATE:.+]] = spv.IMul %[[ONE]], %[[ARG1_CAST]] : i32
   //     CHECK: %[[FLAT_IDX:.+]] = spv.IAdd %[[OFFSET]], %[[UPDATE]] : i32
   //     CHECK: %[[TWO:.+]] = spv.Constant 2 : i32
   //     CHECK: %[[SIXTEEN:.+]] = spv.Constant 16 : i32
   //     CHECK: %[[IDX:.+]] = spv.UMod %[[FLAT_IDX]], %[[TWO]] : i32
func @load_i16(%arg0: memref<i16>) {
   // CHECK-NOT: spv.ShiftRightArithmetic
   %0 = memref.load %arg0[] : memref<i16>
   return
 }
 
 // CHECK-LABEL: @store_i8
 //       CHECK: (%[[ARG0:.+]]: {{.*}}, %[[ARG1:.+]]: i8)
 func @store_i8(%arg0: memref<i8>, %value: i8) {
-  //     CHECK: %[[ARG1_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG1]] : i8 to i32
-  //     CHECK: %[[ARG0_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG0]]
+  //     CHECK-DAG: %[[ARG1_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG1]] : i8 to i32
+  //     CHECK-DAG: %[[ARG0_CAST:.+]] = builtin.unrealized_conversion_cast %[[ARG0]]
   //     CHECK: %[[ZERO:.+]] = spv.Constant 0 : i32
   //     CHECK: %[[FOUR:.+]] = spv.Constant 4 : i32
   //     CHECK: %[[EIGHT:.+]] = spv.Constant 8 : i32
   //     CHECK: %[[IDX:.+]] = spv.UMod %[[ZERO]], %[[FOUR]] : i32
   //     CHECK: %[[OFFSET:.+]] = spv.IMul %[[IDX]], %[[EIGHT]] : i32
   //     CHECK: %[[MASK1:.+]] = spv.Constant 255 : i32
   //     CHECK: %[[TMP1:.+]] = spv.ShiftLeftLogical %[[MASK1]], %[[OFFSET]] : i32, i32
   //     CHECK: %[[MASK:.+]] = spv.Not %[[TMP1]] : i32

mlir/test/Conversion/OpenMPToLLVM/convert-to-llvmir.mlir

 // RUN: mlir-opt -convert-openmp-to-llvm %s  -split-input-file | FileCheck %s
 
 // CHECK-LABEL: llvm.func @master_block_arg
 func @master_block_arg() {
   // CHECK: omp.master
   omp.master {
   // CHECK-NEXT: ^[[BB0:.*]](%[[ARG1:.*]]: i64, %[[ARG2:.*]]: i64):
   ^bb0(%arg1: index, %arg2: index):
-    // CHECK-NEXT: "test.payload"(%[[ARG1]], %[[ARG2]]) : (i64, i64) -> ()
+    // CHECK-DAG: %[[CAST_ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1]] : i64 to index
+    // CHECK-DAG: %[[CAST_ARG2:.*]] = builtin.unrealized_conversion_cast %[[ARG2]] : i64 to index
+    // CHECK-NEXT: "test.payload"(%[[CAST_ARG1]], %[[CAST_ARG2]]) : (index, index) -> ()
     "test.payload"(%arg1, %arg2) : (index, index) -> ()
     omp.terminator
   }
   return
 }
 
 // CHECK-LABEL: llvm.func @branch_loop
 func @branch_loop() {

 // CHECK-LABEL: @wsloop
 // CHECK: (%[[ARG0:.*]]: i64, %[[ARG1:.*]]: i64, %[[ARG2:.*]]: i64, %[[ARG3:.*]]: i64, %[[ARG4:.*]]: i64, %[[ARG5:.*]]: i64)
 func @wsloop(%arg0: index, %arg1: index, %arg2: index, %arg3: index, %arg4: index, %arg5: index) {
   // CHECK: omp.parallel
   omp.parallel {
     // CHECK: omp.wsloop (%[[ARG6:.*]], %[[ARG7:.*]]) : i64 = (%[[ARG0]], %[[ARG1]]) to (%[[ARG2]], %[[ARG3]]) step (%[[ARG4]], %[[ARG5]]) {
     "omp.wsloop"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5) ( {
     ^bb0(%arg6: index, %arg7: index):  // no predecessors
-      // CHECK: "test.payload"(%[[ARG6]], %[[ARG7]]) : (i64, i64) -> ()
+      // CHECK-DAG: %[[CAST_ARG6:.*]] = builtin.unrealized_conversion_cast %[[ARG6]] : i64 to index
+      // CHECK-DAG: %[[CAST_ARG7:.*]] = builtin.unrealized_conversion_cast %[[ARG7]] : i64 to index
+      // CHECK: "test.payload"(%[[CAST_ARG6]], %[[CAST_ARG7]]) : (index, index) -> ()
       "test.payload"(%arg6, %arg7) : (index, index) -> ()
       omp.yield
     }) {operand_segment_sizes = dense<[2, 2, 2, 0, 0, 0, 0, 0, 0, 0]> : vector<10xi32>} : (index, index, index, index, index, index) -> ()
     omp.terminator
   }
   return
 }

mlir/test/Conversion/StandardToLLVM/calling-convention.mlir

func @return_var_memref_caller(%arg0: memref<4x3xf32>) {
   return
 }
 
 // CHECK-LABEL: llvm.func @return_var_memref
 func @return_var_memref(%arg0: memref<4x3xf32>) -> memref<*xf32> attributes { llvm.emit_c_interface } {
   // Match the construction of the unranked descriptor.
   // CHECK: %[[ALLOCA:.*]] = llvm.alloca
   // CHECK: %[[MEMORY:.*]] = llvm.bitcast %[[ALLOCA]]
+  // CHECK: %[[RANK:.*]] = llvm.mlir.constant(2 : i64)
   // CHECK: %[[DESC_0:.*]] = llvm.mlir.undef : !llvm.struct<(i64, ptr<i8>)>
-  // CHECK: %[[DESC_1:.*]] = llvm.insertvalue %{{.*}}, %[[DESC_0]][0]
+  // CHECK: %[[DESC_1:.*]] = llvm.insertvalue %[[RANK]], %[[DESC_0]][0]
   // CHECK: %[[DESC_2:.*]] = llvm.insertvalue %[[MEMORY]], %[[DESC_1]][1]
   %0 = memref.cast %arg0: memref<4x3xf32> to memref<*xf32>
 
+
   // CHECK: %[[ONE:.*]] = llvm.mlir.constant(1 : index)
   // CHECK: %[[TWO:.*]] = llvm.mlir.constant(2 : index)
   // These sizes may depend on the data layout, not matching specific values.
   // CHECK: %[[PTR_SIZE:.*]] = llvm.mlir.constant
   // CHECK: %[[IDX_SIZE:.*]] = llvm.mlir.constant
 
   // CHECK: %[[DOUBLE_PTR_SIZE:.*]] = llvm.mul %[[TWO]], %[[PTR_SIZE]]
-  // CHECK: %[[RANK:.*]] = llvm.extractvalue %[[DESC_2]][0] : !llvm.struct<(i64, ptr<i8>)>
   // CHECK: %[[DOUBLE_RANK:.*]] = llvm.mul %[[TWO]], %[[RANK]]
   // CHECK: %[[DOUBLE_RANK_INC:.*]] = llvm.add %[[DOUBLE_RANK]], %[[ONE]]
   // CHECK: %[[TABLES_SIZE:.*]] = llvm.mul %[[DOUBLE_RANK_INC]], %[[IDX_SIZE]]
   // CHECK: %[[ALLOC_SIZE:.*]] = llvm.add %[[DOUBLE_PTR_SIZE]], %[[TABLES_SIZE]]
   // CHECK: %[[FALSE:.*]] = llvm.mlir.constant(false)
   // CHECK: %[[ALLOCATED:.*]] = llvm.call @malloc(%[[ALLOC_SIZE]])
-  // CHECK: %[[SOURCE:.*]] = llvm.extractvalue %[[DESC_2]][1]
-  // CHECK: "llvm.intr.memcpy"(%[[ALLOCATED]], %[[SOURCE]], %[[ALLOC_SIZE]], %[[FALSE]])
+  // CHECK: "llvm.intr.memcpy"(%[[ALLOCATED]], %[[MEMORY]], %[[ALLOC_SIZE]], %[[FALSE]])
   // CHECK: %[[NEW_DESC:.*]] = llvm.mlir.undef : !llvm.struct<(i64, ptr<i8>)>
-  // CHECK: %[[RANK:.*]] = llvm.extractvalue %[[DESC_2]][0] : !llvm.struct<(i64, ptr<i8>)>
   // CHECK: %[[NEW_DESC_1:.*]] = llvm.insertvalue %[[RANK]], %[[NEW_DESC]][0]
   // CHECK: %[[NEW_DESC_2:.*]] = llvm.insertvalue %[[ALLOCATED]], %[[NEW_DESC_1]][1]
   // CHECK: llvm.return %[[NEW_DESC_2]]
   return %0 : memref<*xf32>
 }
 
 // Check that the result memref is passed as parameter
 // CHECK-LABEL: @_mlir_ciface_return_var_memref
func @return_two_var_memref(%arg0: memref<4x3xf32>) -> (memref<*xf32>, memref<*xf32>) attributes { llvm.emit_c_interface } {
   // CHECK: %[[DESC_2:.*]] = llvm.insertvalue %[[MEMORY]], %[[DESC_1]][1]
   %0 = memref.cast %arg0 : memref<4x3xf32> to memref<*xf32>
 
   // Only check that we allocate the memory for each operand of the "return"
   // separately, even if both operands are the same value. The calling
   // convention requires the caller to free them and the caller cannot know
   // whether they are the same value or not.
   // CHECK: %[[ALLOCATED_1:.*]] = llvm.call @malloc(%{{.*}})
-  // CHECK: %[[SOURCE_1:.*]] = llvm.extractvalue %[[DESC_2]][1]
-  // CHECK: "llvm.intr.memcpy"(%[[ALLOCATED_1]], %[[SOURCE_1]], %{{.*}}, %[[FALSE:.*]])
+  // CHECK: "llvm.intr.memcpy"(%[[ALLOCATED_1]], %[[MEMORY]], %{{.*}}, %[[FALSE:.*]])
   // CHECK: %[[RES_1:.*]] = llvm.mlir.undef
   // CHECK: %[[RES_11:.*]] = llvm.insertvalue %{{.*}}, %[[RES_1]][0]
   // CHECK: %[[RES_12:.*]] = llvm.insertvalue %[[ALLOCATED_1]], %[[RES_11]][1]
 
   // CHECK: %[[ALLOCATED_2:.*]] = llvm.call @malloc(%{{.*}})
-  // CHECK: %[[SOURCE_2:.*]] = llvm.extractvalue %[[DESC_2]][1]
-  // CHECK: "llvm.intr.memcpy"(%[[ALLOCATED_2]], %[[SOURCE_2]], %{{.*}}, %[[FALSE]])
+  // CHECK: "llvm.intr.memcpy"(%[[ALLOCATED_2]], %[[MEMORY]], %{{.*}}, %[[FALSE]])
   // CHECK: %[[RES_2:.*]] = llvm.mlir.undef
   // CHECK: %[[RES_21:.*]] = llvm.insertvalue %{{.*}}, %[[RES_2]][0]
   // CHECK: %[[RES_22:.*]] = llvm.insertvalue %[[ALLOCATED_2]], %[[RES_21]][1]
 
   // CHECK: %[[RESULTS:.*]] = llvm.mlir.undef : !llvm.struct<(struct<(i64, ptr<i8>)>, struct<(i64, ptr<i8>)>)>
   // CHECK: %[[RESULTS_1:.*]] = llvm.insertvalue %[[RES_12]], %[[RESULTS]]
   // CHECK: %[[RESULTS_2:.*]] = llvm.insertvalue %[[RES_22]], %[[RESULTS_1]]
   // CHECK: llvm.return %[[RESULTS_2]]
   return %0, %0 : memref<*xf32>, memref<*xf32>
 }
 
 // Check that the result memrefs are passed as parameter
 // CHECK-LABEL: @_mlir_ciface_return_two_var_memref
 // CHECK-SAME: (%{{.*}}: !llvm.ptr<struct<(struct<(i64, ptr<i8>)>, struct<(i64, ptr<i8>)>)>>,
 // CHECK-SAME: %{{.*}}: !llvm.ptr<struct<(ptr<f32>, ptr<f32>, i64, array<2 x i64>, array<2 x i64>)>>)
 

mlir/test/Conversion/StandardToSPIRV/std-ops-to-spirv.mlir

 // -----
 
 // Check that types are converted to 32-bit when no special capabilities that
 // are not supported.
 module attributes {
   spv.target_env = #spv.target_env<#spv.vce<v1.0, [], []>, {}>
 } {
 
+// expected-error@below {{failed to materialize conversion for block argument #0 that remained live after conversion}}
 func @int_vector4_invalid(%arg0: vector<4xi64>) {
-  // expected-error @+2 {{bitwidth emulation is not implemented yet on unsigned op}}
-  // expected-error @+1 {{op requires the same type for all operands and results}}
+  // expected-error@below {{bitwidth emulation is not implemented yet on unsigned op}}
+  // expected-note@below {{see existing live user here}}
   %0 = arith.divui %arg0, %arg0: vector<4xi64>
   return
 }
 
 } // end module
 
 // -----
 

mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir

 // -----
 
 func @broadcast_vec2d_from_index_vec1d(%arg0: vector<2xindex>) -> vector<3x2xindex> {
   %0 = vector.broadcast %arg0 : vector<2xindex> to vector<3x2xindex>
   return %0 : vector<3x2xindex>
 }
 // CHECK-LABEL: @broadcast_vec2d_from_index_vec1d(
 // CHECK-SAME:  %[[A:.*]]: vector<2xindex>)
-// CHECK:       %[[T0:.*]] = arith.constant dense<0> : vector<3x2xindex>
 // CHECK:       %[[T1:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<2xindex> to vector<2xi64>
+// CHECK:       %[[T0:.*]] = arith.constant dense<0> : vector<3x2xindex>
 // CHECK:       %[[T2:.*]] = builtin.unrealized_conversion_cast %[[T0]] : vector<3x2xindex> to !llvm.array<3 x vector<2xi64>>
 // CHECK:       %[[T3:.*]] = llvm.insertvalue %[[T1]], %[[T2]][0] : !llvm.array<3 x vector<2xi64>>
 
 // CHECK:       %[[T4:.*]] = builtin.unrealized_conversion_cast %{{.*}} : !llvm.array<3 x vector<2xi64>> to vector<3x2xindex>
 // CHECK:       return %[[T4]] : vector<3x2xindex>
 
 // -----
 
 func @broadcast_vec3d_from_vec1d(%arg0: vector<2xf32>) -> vector<4x3x2xf32> {
   %0 = vector.broadcast %arg0 : vector<2xf32> to vector<4x3x2xf32>
   return %0 : vector<4x3x2xf32>
 }
 // CHECK-LABEL: @broadcast_vec3d_from_vec1d(
 // CHECK-SAME:  %[[A:.*]]: vector<2xf32>)
 // CHECK:       %[[T0:.*]] = arith.constant dense<0.000000e+00> : vector<3x2xf32>
+// CHECK:       %[[T2:.*]] = builtin.unrealized_conversion_cast %[[T0]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T1:.*]] = arith.constant dense<0.000000e+00> : vector<4x3x2xf32>
+// CHECK:       %[[T6:.*]] = builtin.unrealized_conversion_cast %[[T1]] : vector<4x3x2xf32> to !llvm.array<4 x array<3 x vector<2xf32>>>
 
-// CHECK:       %[[T2:.*]] = builtin.unrealized_conversion_cast %[[T0]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T3:.*]] = llvm.insertvalue %[[A]], %[[T2]][0] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T4:.*]] = llvm.insertvalue %[[A]], %[[T3]][1] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T5:.*]] = llvm.insertvalue %[[A]], %[[T4]][2] : !llvm.array<3 x vector<2xf32>>
 
-// CHECK:       %[[T6:.*]] = builtin.unrealized_conversion_cast %[[T1]] : vector<4x3x2xf32> to !llvm.array<4 x array<3 x vector<2xf32>>>
 // CHECK:       %[[T7:.*]] = llvm.insertvalue %[[T5]], %[[T6]][0] : !llvm.array<4 x array<3 x vector<2xf32>>>
 // CHECK:       %[[T8:.*]] = llvm.insertvalue %[[T5]], %[[T7]][1] : !llvm.array<4 x array<3 x vector<2xf32>>>
 // CHECK:       %[[T9:.*]] = llvm.insertvalue %[[T5]], %[[T8]][2] : !llvm.array<4 x array<3 x vector<2xf32>>>
 // CHECK:       %[[T10:.*]] = llvm.insertvalue %[[T5]], %[[T9]][3] : !llvm.array<4 x array<3 x vector<2xf32>>>
 
 // CHECK:       %[[T11:.*]] = builtin.unrealized_conversion_cast %[[T10]] : !llvm.array<4 x array<3 x vector<2xf32>>> to vector<4x3x2xf32>
 // CHECK:       return %[[T11]] : vector<4x3x2xf32>
 
 // -----
 
 func @broadcast_vec3d_from_vec2d(%arg0: vector<3x2xf32>) -> vector<4x3x2xf32> {
   %0 = vector.broadcast %arg0 : vector<3x2xf32> to vector<4x3x2xf32>
   return %0 : vector<4x3x2xf32>
 }
 // CHECK-LABEL: @broadcast_vec3d_from_vec2d(
 // CHECK-SAME:  %[[A:.*]]: vector<3x2xf32>)
-// CHECK:       %[[T0:.*]] = arith.constant dense<0.000000e+00> : vector<4x3x2xf32>
 // CHECK:       %[[T1:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
+// CHECK:       %[[T0:.*]] = arith.constant dense<0.000000e+00> : vector<4x3x2xf32>
 // CHECK:       %[[T2:.*]] = builtin.unrealized_conversion_cast %[[T0]] : vector<4x3x2xf32> to !llvm.array<4 x array<3 x vector<2xf32>>>
 // CHECK:       %[[T3:.*]] = llvm.insertvalue %[[T1]], %[[T2]][0] : !llvm.array<4 x array<3 x vector<2xf32>>>
-// CHECK:       %[[T4:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
-// CHECK:       %[[T5:.*]] = llvm.insertvalue %[[T4]], %[[T3]][1] : !llvm.array<4 x array<3 x vector<2xf32>>>
-// CHECK:       %[[T6:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
-// CHECK:       %[[T7:.*]] = llvm.insertvalue %[[T6]], %[[T5]][2] : !llvm.array<4 x array<3 x vector<2xf32>>>
-// CHECK:       %[[T8:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
-// CHECK:       %[[T9:.*]] = llvm.insertvalue %[[T8]], %[[T7]][3] : !llvm.array<4 x array<3 x vector<2xf32>>>
+// CHECK:       %[[T5:.*]] = llvm.insertvalue %[[T1]], %[[T3]][1] : !llvm.array<4 x array<3 x vector<2xf32>>>
+// CHECK:       %[[T7:.*]] = llvm.insertvalue %[[T1]], %[[T5]][2] : !llvm.array<4 x array<3 x vector<2xf32>>>
+// CHECK:       %[[T9:.*]] = llvm.insertvalue %[[T1]], %[[T7]][3] : !llvm.array<4 x array<3 x vector<2xf32>>>
 // CHECK:       %[[T10:.*]] = builtin.unrealized_conversion_cast %[[T9]] : !llvm.array<4 x array<3 x vector<2xf32>>> to vector<4x3x2xf32>
 // CHECK:       return %[[T10]] : vector<4x3x2xf32>
 
 
 // -----
 
 func @broadcast_stretch(%arg0: vector<1xf32>) -> vector<4xf32> {
   %0 = vector.broadcast %arg0 : vector<1xf32> to vector<4xf32>

 // -----
 
 func @broadcast_stretch_at_start(%arg0: vector<1x4xf32>) -> vector<3x4xf32> {
   %0 = vector.broadcast %arg0 : vector<1x4xf32> to vector<3x4xf32>
   return %0 : vector<3x4xf32>
 }
 // CHECK-LABEL: @broadcast_stretch_at_start(
 // CHECK-SAME:  %[[A:.*]]: vector<1x4xf32>)
-// CHECK:       %[[T1:.*]] = arith.constant dense<0.000000e+00> : vector<3x4xf32>
 // CHECK:       %[[T2:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<1x4xf32> to !llvm.array<1 x vector<4xf32>>
-// CHECK:       %[[T3:.*]] = llvm.extractvalue %[[T2]][0] : !llvm.array<1 x vector<4xf32>>
+// CHECK:       %[[T1:.*]] = arith.constant dense<0.000000e+00> : vector<3x4xf32>
 // CHECK:       %[[T4:.*]] = builtin.unrealized_conversion_cast %[[T1]] : vector<3x4xf32> to !llvm.array<3 x vector<4xf32>>
+// CHECK:       %[[T3:.*]] = llvm.extractvalue %[[T2]][0] : !llvm.array<1 x vector<4xf32>>
 // CHECK:       %[[T5:.*]] = llvm.insertvalue %[[T3]], %[[T4]][0] : !llvm.array<3 x vector<4xf32>>
 // CHECK:       %[[T6:.*]] = llvm.insertvalue %[[T3]], %[[T5]][1] : !llvm.array<3 x vector<4xf32>>
 // CHECK:       %[[T7:.*]] = llvm.insertvalue %[[T3]], %[[T6]][2] : !llvm.array<3 x vector<4xf32>>
 // CHECK:       %[[T8:.*]] = builtin.unrealized_conversion_cast %[[T7]] : !llvm.array<3 x vector<4xf32>> to vector<3x4xf32>
 // CHECK:       return %[[T8]] : vector<3x4xf32>
 
 // -----
 
 func @broadcast_stretch_at_end(%arg0: vector<4x1xf32>) -> vector<4x3xf32> {
   %0 = vector.broadcast %arg0 : vector<4x1xf32> to vector<4x3xf32>
   return %0 : vector<4x3xf32>
 }
 // CHECK-LABEL: @broadcast_stretch_at_end(
 // CHECK-SAME:  %[[A:.*]]: vector<4x1xf32>)
-// CHECK:       %[[T1:.*]] = arith.constant dense<0.000000e+00> : vector<4x3xf32>
 // CHECK:       %[[T2:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4x1xf32> to !llvm.array<4 x vector<1xf32>>
+// CHECK:       %[[T1:.*]] = arith.constant dense<0.000000e+00> : vector<4x3xf32>
+// CHECK:       %[[T7:.*]] = builtin.unrealized_conversion_cast %[[T1]] : vector<4x3xf32> to !llvm.array<4 x vector<3xf32>>
 // CHECK:       %[[T3:.*]] = llvm.extractvalue %[[T2]][0] : !llvm.array<4 x vector<1xf32>>
 // CHECK:       %[[T4:.*]] = llvm.mlir.constant(0 : i64) : i64
 // CHECK:       %[[T5:.*]] = llvm.extractelement %[[T3]]{{\[}}%[[T4]] : i64] : vector<1xf32>
 // CHECK:       %[[T6:.*]] = splat %[[T5]] : vector<3xf32>
-// CHECK:       %[[T7:.*]] = builtin.unrealized_conversion_cast %[[T1]] : vector<4x3xf32> to !llvm.array<4 x vector<3xf32>>
 // CHECK:       %[[T8:.*]] = llvm.insertvalue %[[T6]], %[[T7]][0] : !llvm.array<4 x vector<3xf32>>
-// CHECK:       %[[T9:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4x1xf32> to !llvm.array<4 x vector<1xf32>>
-// CHECK:       %[[T10:.*]] = llvm.extractvalue %[[T9]][1] : !llvm.array<4 x vector<1xf32>>
+// CHECK:       %[[T10:.*]] = llvm.extractvalue %[[T2]][1] : !llvm.array<4 x vector<1xf32>>
 // CHECK:       %[[T11:.*]] = llvm.mlir.constant(0 : i64) : i64
 // CHECK:       %[[T12:.*]] = llvm.extractelement %[[T10]]{{\[}}%[[T11]] : i64] : vector<1xf32>
 // CHECK:       %[[T13:.*]] = splat %[[T12]] : vector<3xf32>
 // CHECK:       %[[T14:.*]] = llvm.insertvalue %[[T13]], %[[T8]][1] : !llvm.array<4 x vector<3xf32>>
-// CHECK:       %[[T15:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4x1xf32> to !llvm.array<4 x vector<1xf32>>
-// CHECK:       %[[T16:.*]] = llvm.extractvalue %[[T15]][2] : !llvm.array<4 x vector<1xf32>>
+// CHECK:       %[[T16:.*]] = llvm.extractvalue %[[T2]][2] : !llvm.array<4 x vector<1xf32>>
 // CHECK:       %[[T17:.*]] = llvm.mlir.constant(0 : i64) : i64
 // CHECK:       %[[T18:.*]] = llvm.extractelement %[[T16]]{{\[}}%[[T17]] : i64] : vector<1xf32>
 // CHECK:       %[[T19:.*]] = splat %[[T18]] : vector<3xf32>
 // CHECK:       %[[T20:.*]] = llvm.insertvalue %[[T19]], %[[T14]][2] : !llvm.array<4 x vector<3xf32>>
-// CHECK:       %[[T21:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4x1xf32> to !llvm.array<4 x vector<1xf32>>
-// CHECK:       %[[T22:.*]] = llvm.extractvalue %[[T21]][3] : !llvm.array<4 x vector<1xf32>>
+// CHECK:       %[[T22:.*]] = llvm.extractvalue %[[T2]][3] : !llvm.array<4 x vector<1xf32>>
 // CHECK:       %[[T23:.*]] = llvm.mlir.constant(0 : i64) : i64
 // CHECK:       %[[T24:.*]] = llvm.extractelement %[[T22]]{{\[}}%[[T23]] : i64] : vector<1xf32>
 // CHECK:       %[[T25:.*]] = splat %[[T24]] : vector<3xf32>
 // CHECK:       %[[T26:.*]] = llvm.insertvalue %[[T25]], %[[T20]][3] : !llvm.array<4 x vector<3xf32>>
 // CHECK:       %[[T27:.*]] = builtin.unrealized_conversion_cast %[[T26]] : !llvm.array<4 x vector<3xf32>> to vector<4x3xf32>
 // CHECK:       return %[[T27]] : vector<4x3xf32>
 
 // -----
 
 func @broadcast_stretch_in_middle(%arg0: vector<4x1x2xf32>) -> vector<4x3x2xf32> {
   %0 = vector.broadcast %arg0 : vector<4x1x2xf32> to vector<4x3x2xf32>
   return %0 : vector<4x3x2xf32>
 }
 // CHECK-LABEL: @broadcast_stretch_in_middle(
 // CHECK-SAME:  %[[A:.*]]: vector<4x1x2xf32>) -> vector<4x3x2xf32> {
+// CHECK:       %[[T3:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4x1x2xf32> to !llvm.array<4 x array<1 x vector<2xf32>>>
 // CHECK:       %[[T1:.*]] = arith.constant dense<0.000000e+00> : vector<4x3x2xf32>
+// CHECK:       %[[T9:.*]] = builtin.unrealized_conversion_cast %[[T1]] : vector<4x3x2xf32> to !llvm.array<4 x array<3 x vector<2xf32>>>
 // CHECK:       %[[T2:.*]] = arith.constant dense<0.000000e+00> : vector<3x2xf32>
-// CHECK:       %[[T3:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4x1x2xf32> to !llvm.array<4 x array<1 x vector<2xf32>>>
-// CHECK:       %[[T4:.*]] = llvm.extractvalue %[[T3]][0, 0] : !llvm.array<4 x array<1 x vector<2xf32>>>
 // CHECK:       %[[T5:.*]] = builtin.unrealized_conversion_cast %[[T2]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
+// CHECK:       %[[T4:.*]] = llvm.extractvalue %[[T3]][0, 0] : !llvm.array<4 x array<1 x vector<2xf32>>>
 // CHECK:       %[[T6:.*]] = llvm.insertvalue %[[T4]], %[[T5]][0] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T7:.*]] = llvm.insertvalue %[[T4]], %[[T6]][1] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T8:.*]] = llvm.insertvalue %[[T4]], %[[T7]][2] : !llvm.array<3 x vector<2xf32>>
-// CHECK:       %[[T9:.*]] = builtin.unrealized_conversion_cast %[[T1]] : vector<4x3x2xf32> to !llvm.array<4 x array<3 x vector<2xf32>>>
 // CHECK:       %[[T10:.*]] = llvm.insertvalue %[[T8]], %[[T9]][0] : !llvm.array<4 x array<3 x vector<2xf32>>>
-// CHECK:       %[[T11:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4x1x2xf32> to !llvm.array<4 x array<1 x vector<2xf32>>>
-// CHECK:       %[[T12:.*]] = llvm.extractvalue %[[T11]][1, 0] : !llvm.array<4 x array<1 x vector<2xf32>>>
-// CHECK:       %[[T13:.*]] = builtin.unrealized_conversion_cast %[[T2]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
-// CHECK:       %[[T14:.*]] = llvm.insertvalue %[[T12]], %[[T13]][0] : !llvm.array<3 x vector<2xf32>>
+// CHECK:       %[[T12:.*]] = llvm.extractvalue %[[T3]][1, 0] : !llvm.array<4 x array<1 x vector<2xf32>>>
+// CHECK:       %[[T14:.*]] = llvm.insertvalue %[[T12]], %[[T5]][0] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T15:.*]] = llvm.insertvalue %[[T12]], %[[T14]][1] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T16:.*]] = llvm.insertvalue %[[T12]], %[[T15]][2] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T17:.*]] = llvm.insertvalue %[[T16]], %[[T10]][1] : !llvm.array<4 x array<3 x vector<2xf32>>>
-// CHECK:       %[[T18:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4x1x2xf32> to !llvm.array<4 x array<1 x vector<2xf32>>>
-// CHECK:       %[[T19:.*]] = llvm.extractvalue %[[T18]][2, 0] : !llvm.array<4 x array<1 x vector<2xf32>>>
-// CHECK:       %[[T20:.*]] = builtin.unrealized_conversion_cast %[[T2]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
-// CHECK:       %[[T21:.*]] = llvm.insertvalue %[[T19]], %[[T20]][0] : !llvm.array<3 x vector<2xf32>>
+// CHECK:       %[[T19:.*]] = llvm.extractvalue %[[T3]][2, 0] : !llvm.array<4 x array<1 x vector<2xf32>>>
+// CHECK:       %[[T21:.*]] = llvm.insertvalue %[[T19]], %[[T5]][0] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T22:.*]] = llvm.insertvalue %[[T19]], %[[T21]][1] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T23:.*]] = llvm.insertvalue %[[T19]], %[[T22]][2] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T24:.*]] = llvm.insertvalue %[[T23]], %[[T17]][2] : !llvm.array<4 x array<3 x vector<2xf32>>>
-// CHECK:       %[[T25:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4x1x2xf32> to !llvm.array<4 x array<1 x vector<2xf32>>>
-// CHECK:       %[[T26:.*]] = llvm.extractvalue %[[T25]][3, 0] : !llvm.array<4 x array<1 x vector<2xf32>>>
-// CHECK:       %[[T27:.*]] = builtin.unrealized_conversion_cast %[[T2]] : vector<3x2xf32> to !llvm.array<3 x vector<2xf32>>
-// CHECK:       %[[T28:.*]] = llvm.insertvalue %[[T26]], %[[T27]][0] : !llvm.array<3 x vector<2xf32>>
+// CHECK:       %[[T26:.*]] = llvm.extractvalue %[[T3]][3, 0] : !llvm.array<4 x array<1 x vector<2xf32>>>
+// CHECK:       %[[T28:.*]] = llvm.insertvalue %[[T26]], %[[T5]][0] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T29:.*]] = llvm.insertvalue %[[T26]], %[[T28]][1] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T30:.*]] = llvm.insertvalue %[[T26]], %[[T29]][2] : !llvm.array<3 x vector<2xf32>>
 // CHECK:       %[[T31:.*]] = llvm.insertvalue %[[T30]], %[[T24]][3] : !llvm.array<4 x array<3 x vector<2xf32>>>
 // CHECK:       %[[T32:.*]] = builtin.unrealized_conversion_cast %[[T31]] : !llvm.array<4 x array<3 x vector<2xf32>>> to vector<4x3x2xf32>
 // CHECK:       return %[[T32]] : vector<4x3x2xf32>
 
 // -----
 
 func @outerproduct(%arg0: vector<2xf32>, %arg1: vector<3xf32>) -> vector<2x3xf32> {
   %2 = vector.outerproduct %arg0, %arg1 : vector<2xf32>, vector<3xf32>
   return %2 : vector<2x3xf32>
 }
 // CHECK-LABEL: @outerproduct(
 // CHECK-SAME:  %[[A:.*]]: vector<2xf32>,
 // CHECK-SAME:  %[[B:.*]]: vector<3xf32>)
 // CHECK:       %[[T2:.*]] = arith.constant dense<0.000000e+00> : vector<2x3xf32>
+// CHECK:       %[[T7:.*]] = builtin.unrealized_conversion_cast %[[T2]] : vector<2x3xf32> to !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T3:.*]] = llvm.mlir.constant(0 : i64) : i64
 // CHECK:       %[[T4:.*]] = llvm.extractelement %[[A]]{{\[}}%[[T3]] : i64] : vector<2xf32>
 // CHECK:       %[[T5:.*]] = splat %[[T4]] : vector<3xf32>
 // CHECK:       %[[T6:.*]] = arith.mulf %[[T5]], %[[B]] : vector<3xf32>
-// CHECK:       %[[T7:.*]] = builtin.unrealized_conversion_cast %[[T2]] : vector<2x3xf32> to !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T8:.*]] = llvm.insertvalue %[[T6]], %[[T7]][0] : !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T9:.*]] = llvm.mlir.constant(1 : i64) : i64
 // CHECK:       %[[T10:.*]] = llvm.extractelement %[[A]]{{\[}}%[[T9]] : i64] : vector<2xf32>
 // CHECK:       %[[T11:.*]] = splat %[[T10]] : vector<3xf32>
 // CHECK:       %[[T12:.*]] = arith.mulf %[[T11]], %[[B]] : vector<3xf32>
 // CHECK:       %[[T13:.*]] = llvm.insertvalue %[[T12]], %[[T8]][1] : !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T14:.*]] = builtin.unrealized_conversion_cast %[[T13]] : !llvm.array<2 x vector<3xf32>> to vector<2x3xf32>
 // CHECK:       return %[[T14]] : vector<2x3xf32>
 
 // -----
 
 func @outerproduct_index(%arg0: vector<2xindex>, %arg1: vector<3xindex>) -> vector<2x3xindex> {
   %2 = vector.outerproduct %arg0, %arg1 : vector<2xindex>, vector<3xindex>
   return %2 : vector<2x3xindex>
 }
 // CHECK-LABEL: @outerproduct_index(
 // CHECK-SAME:  %[[A:.*]]: vector<2xindex>,
 // CHECK-SAME:  %[[B:.*]]: vector<3xindex>)
-// CHECK:       %[[T0:.*]] = arith.constant dense<0> : vector<2x3xindex>
 // CHECK:       %[[T1:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<2xindex> to vector<2xi64>
+// CHECK:       %[[T0:.*]] = arith.constant dense<0> : vector<2x3xindex>
+// CHECK:       %[[T8:.*]] = builtin.unrealized_conversion_cast %[[T0]] : vector<2x3xindex> to !llvm.array<2 x vector<3xi64>>
 // CHECK:       %[[T2:.*]] = llvm.mlir.constant(0 : i64) : i64
 // CHECK:       %[[T3:.*]] = llvm.extractelement %[[T1]]{{\[}}%[[T2]] : i64] : vector<2xi64>
 // CHECK:       %[[T4:.*]] = builtin.unrealized_conversion_cast %[[T3]] : i64 to index
 // CHECK:       %[[T5:.*]] = splat %[[T4]] : vector<3xindex>
 // CHECK:       %[[T6:.*]] = arith.muli %[[T5]], %[[B]] : vector<3xindex>
 // CHECK:       %[[T7:.*]] = builtin.unrealized_conversion_cast %[[T6]] : vector<3xindex> to vector<3xi64>
-// CHECK:       %[[T8:.*]] = builtin.unrealized_conversion_cast %[[T0]] : vector<2x3xindex> to !llvm.array<2 x vector<3xi64>>
 // CHECK:       %{{.*}} = llvm.insertvalue %[[T7]], %[[T8]][0] : !llvm.array<2 x vector<3xi64>>
 
 // -----
 
 func @outerproduct_add(%arg0: vector<2xf32>, %arg1: vector<3xf32>, %arg2: vector<2x3xf32>) -> vector<2x3xf32> {
   %2 = vector.outerproduct %arg0, %arg1, %arg2 : vector<2xf32>, vector<3xf32>
   return %2 : vector<2x3xf32>
 }
 // CHECK-LABEL: @outerproduct_add(
 // CHECK-SAME:  %[[A:.*]]: vector<2xf32>,
 // CHECK-SAME:  %[[B:.*]]: vector<3xf32>,
 // CHECK-SAME:  %[[C:.*]]: vector<2x3xf32>) -> vector<2x3xf32>
+// CHECK:       %[[T7:.*]] = builtin.unrealized_conversion_cast %[[C]] : vector<2x3xf32> to !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T3:.*]] = arith.constant dense<0.000000e+00> : vector<2x3xf32>
+// CHECK:       %[[T10:.*]] = builtin.unrealized_conversion_cast %[[T3]] : vector<2x3xf32> to !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T4:.*]] = llvm.mlir.constant(0 : i64) : i64
 // CHECK:       %[[T5:.*]] = llvm.extractelement %[[A]]{{\[}}%[[T4]] : i64] : vector<2xf32>
 // CHECK:       %[[T6:.*]] = splat %[[T5]] : vector<3xf32>
-// CHECK:       %[[T7:.*]] = builtin.unrealized_conversion_cast %[[C]] : vector<2x3xf32> to !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T8:.*]] = llvm.extractvalue %[[T7]][0] : !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T9:.*]] = "llvm.intr.fmuladd"(%[[T6]], %[[B]], %[[T8]]) : (vector<3xf32>, vector<3xf32>, vector<3xf32>) -> vector<3xf32>
-// CHECK:       %[[T10:.*]] = builtin.unrealized_conversion_cast %[[T3]] : vector<2x3xf32> to !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T11:.*]] = llvm.insertvalue %[[T9]], %[[T10]][0] : !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T12:.*]] = llvm.mlir.constant(1 : i64) : i64
 // CHECK:       %[[T13:.*]] = llvm.extractelement %[[A]]{{\[}}%[[T12]] : i64] : vector<2xf32>
 // CHECK:       %[[T14:.*]] = splat %[[T13]] : vector<3xf32>
-// CHECK:       %[[T15:.*]] = builtin.unrealized_conversion_cast %[[C]] : vector<2x3xf32> to !llvm.array<2 x vector<3xf32>>
-// CHECK:       %[[T16:.*]] = llvm.extractvalue %[[T15]][1] : !llvm.array<2 x vector<3xf32>>
+// CHECK:       %[[T16:.*]] = llvm.extractvalue %[[T7]][1] : !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T17:.*]] = "llvm.intr.fmuladd"(%[[T14]], %[[B]], %[[T16]]) : (vector<3xf32>, vector<3xf32>, vector<3xf32>) -> vector<3xf32>
 // CHECK:       %[[T18:.*]] = llvm.insertvalue %[[T17]], %[[T11]][1] : !llvm.array<2 x vector<3xf32>>
 // CHECK:       %[[T19:.*]] = builtin.unrealized_conversion_cast %[[T18]] : !llvm.array<2 x vector<3xf32>> to vector<2x3xf32>
 // CHECK:       return %[[T19]] : vector<2x3xf32>
 
 // -----
 
 func @shuffle_1D_direct(%arg0: vector<2xf32>, %arg1: vector<2xf32>) -> vector<2xf32> {

 
 func @shuffle_1D_index_direct(%arg0: vector<2xindex>, %arg1: vector<2xindex>) -> vector<2xindex> {
   %1 = vector.shuffle %arg0, %arg1 [0, 1] : vector<2xindex>, vector<2xindex>
   return %1 : vector<2xindex>
 }
 // CHECK-LABEL: @shuffle_1D_index_direct(
 // CHECK-SAME: %[[A:.*]]: vector<2xindex>,
 // CHECK-SAME: %[[B:.*]]: vector<2xindex>)
-//       CHECK:   %[[T0:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<2xindex> to vector<2xi64>
-//       CHECK:   %[[T1:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2xindex> to vector<2xi64>
+//       CHECK-DAG:   %[[T0:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<2xindex> to vector<2xi64>
+//       CHECK-DAG:   %[[T1:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2xindex> to vector<2xi64>
 //       CHECK:   %[[T2:.*]] = llvm.shufflevector %[[T0]], %[[T1]] [0, 1] : vector<2xi64>, vector<2xi64>
 //       CHECK:   %[[T3:.*]] = builtin.unrealized_conversion_cast %[[T2]] : vector<2xi64> to vector<2xindex>
 //       CHECK:   return %[[T3]] : vector<2xindex>
 
 // -----
 
 func @shuffle_1D(%arg0: vector<2xf32>, %arg1: vector<3xf32>) -> vector<5xf32> {
   %1 = vector.shuffle %arg0, %arg1 [4, 3, 2, 1, 0] : vector<2xf32>, vector<3xf32>

 
 func @shuffle_2D(%a: vector<1x4xf32>, %b: vector<2x4xf32>) -> vector<3x4xf32> {
   %1 = vector.shuffle %a, %b[1, 0, 2] : vector<1x4xf32>, vector<2x4xf32>
   return %1 : vector<3x4xf32>
 }
 // CHECK-LABEL: @shuffle_2D(
 // CHECK-SAME: %[[A:.*]]: vector<1x4xf32>,
 // CHECK-SAME: %[[B:.*]]: vector<2x4xf32>)
-//       CHECK:   %[[VAL_0:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<1x4xf32> to !llvm.array<1 x vector<4xf32>>
-//       CHECK:   %[[VAL_1:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
+//       CHECK-DAG:   %[[VAL_0:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<1x4xf32> to !llvm.array<1 x vector<4xf32>>
+//       CHECK-DAG:   %[[VAL_1:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
 //       CHECK:   %[[u0:.*]] = llvm.mlir.undef : !llvm.array<3 x vector<4xf32>>
 //       CHECK:   %[[e1:.*]] = llvm.extractvalue %[[VAL_1]][0] : !llvm.array<2 x vector<4xf32>>
 //       CHECK:   %[[i1:.*]] = llvm.insertvalue %[[e1]], %[[u0]][0] : !llvm.array<3 x vector<4xf32>>
 //       CHECK:   %[[e2:.*]] = llvm.extractvalue %[[VAL_0]][0] : !llvm.array<1 x vector<4xf32>>
 //       CHECK:   %[[i2:.*]] = llvm.insertvalue %[[e2]], %[[i1]][1] : !llvm.array<3 x vector<4xf32>>
 //       CHECK:   %[[e3:.*]] = llvm.extractvalue %[[VAL_1]][1] : !llvm.array<2 x vector<4xf32>>
 //       CHECK:   %[[i3:.*]] = llvm.insertvalue %[[e3]], %[[i2]][2] : !llvm.array<3 x vector<4xf32>>
 //       CHECK:   %[[VAL_3:.*]] = builtin.unrealized_conversion_cast %[[i3]] : !llvm.array<3 x vector<4xf32>> to vector<3x4xf32>

 
 func @insert_index_element_into_vec_1d(%arg0: index, %arg1: vector<4xindex>) -> vector<4xindex> {
   %0 = vector.insert %arg0, %arg1[3] : index into vector<4xindex>
   return %0 : vector<4xindex>
 }
 // CHECK-LABEL: @insert_index_element_into_vec_1d(
 // CHECK-SAME: %[[A:.*]]: index,
 // CHECK-SAME: %[[B:.*]]: vector<4xindex>)
-//       CHECK:   %[[T0:.*]] = builtin.unrealized_conversion_cast %[[A]] : index to i64
-//       CHECK:   %[[T1:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<4xindex> to vector<4xi64>
+//       CHECK-DAG:   %[[T0:.*]] = builtin.unrealized_conversion_cast %[[A]] : index to i64
+//       CHECK-DAG:   %[[T1:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<4xindex> to vector<4xi64>
 //       CHECK:   %[[T3:.*]] = llvm.mlir.constant(3 : i64) : i64
 //       CHECK:   %[[T4:.*]] = llvm.insertelement %[[T0]], %[[T1]][%[[T3]] : i64] : vector<4xi64>
 //       CHECK:   %[[T5:.*]] = builtin.unrealized_conversion_cast %[[T4]] : vector<4xi64> to vector<4xindex>
 //       CHECK:   return %[[T5]] : vector<4xindex>
 
 // -----
 
 func @insert_vec_2d_into_vec_3d(%arg0: vector<8x16xf32>, %arg1: vector<4x8x16xf32>) -> vector<4x8x16xf32> {

 
 func @extract_strided_index_slice1(%arg0: vector<4xindex>) -> vector<2xindex> {
   %0 = vector.extract_strided_slice %arg0 {offsets = [2], sizes = [2], strides = [1]} : vector<4xindex> to vector<2xindex>
   return %0 : vector<2xindex>
 }
 // CHECK-LABEL: @extract_strided_index_slice1(
 //  CHECK-SAME:    %[[A:.*]]: vector<4xindex>)
 //       CHECK:    %[[T0:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4xindex> to vector<4xi64>
-//       CHECK:    %[[T1:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<4xindex> to vector<4xi64>
-//       CHECK:    %[[T2:.*]] = llvm.shufflevector %[[T0]], %[[T1]] [2, 3] : vector<4xi64>, vector<4xi64>
+//       CHECK:    %[[T2:.*]] = llvm.shufflevector %[[T0]], %[[T0]] [2, 3] : vector<4xi64>, vector<4xi64>
 //       CHECK:    %[[T3:.*]] = builtin.unrealized_conversion_cast %[[T2]] : vector<2xi64> to vector<2xindex>
 //       CHECK:    return %[[T3]] : vector<2xindex>
 
 // -----
 
 func @extract_strided_slice2(%arg0: vector<4x8xf32>) -> vector<2x8xf32> {
   %0 = vector.extract_strided_slice %arg0 {offsets = [2], sizes = [2], strides = [1]} : vector<4x8xf32> to vector<2x8xf32>
   return %0 : vector<2x8xf32>

 // -----
 
 func @extract_strided_slice3(%arg0: vector<4x8xf32>) -> vector<2x2xf32> {
   %0 = vector.extract_strided_slice %arg0 {offsets = [2, 2], sizes = [2, 2], strides = [1, 1]} : vector<4x8xf32> to vector<2x2xf32>
   return %0 : vector<2x2xf32>
 }
 // CHECK-LABEL: @extract_strided_slice3(
 //  CHECK-SAME:    %[[ARG:.*]]: vector<4x8xf32>)
+//       CHECK:    %[[A:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : vector<4x8xf32> to !llvm.array<4 x vector<8xf32>>
 //       CHECK:    %[[VAL_1:.*]] = arith.constant 0.000000e+00 : f32
 //       CHECK:    %[[VAL_2:.*]] = splat %[[VAL_1]] : vector<2x2xf32>
-//       CHECK:    %[[A:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : vector<4x8xf32> to !llvm.array<4 x vector<8xf32>>
+//       CHECK:    %[[VAL_6:.*]] = builtin.unrealized_conversion_cast %[[VAL_2]] : vector<2x2xf32> to !llvm.array<2 x vector<2xf32>>
 //       CHECK:    %[[T2:.*]] = llvm.extractvalue %[[A]][2] : !llvm.array<4 x vector<8xf32>>
 //       CHECK:    %[[T3:.*]] = llvm.shufflevector %[[T2]], %[[T2]] [2, 3] : vector<8xf32>, vector<8xf32>
-//       CHECK:    %[[VAL_6:.*]] = builtin.unrealized_conversion_cast %[[VAL_2]] : vector<2x2xf32> to !llvm.array<2 x vector<2xf32>>
 //       CHECK:    %[[T4:.*]] = llvm.insertvalue %[[T3]], %[[VAL_6]][0] : !llvm.array<2 x vector<2xf32>>
-//       CHECK:    %[[A:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : vector<4x8xf32> to !llvm.array<4 x vector<8xf32>>
 //       CHECK:    %[[T5:.*]] = llvm.extractvalue %[[A]][3] : !llvm.array<4 x vector<8xf32>>
 //       CHECK:    %[[T6:.*]] = llvm.shufflevector %[[T5]], %[[T5]] [2, 3] : vector<8xf32>, vector<8xf32>
 //       CHECK:    %[[T7:.*]] = llvm.insertvalue %[[T6]], %[[T4]][1] : !llvm.array<2 x vector<2xf32>>
 //       CHECK:    %[[VAL_12:.*]] = builtin.unrealized_conversion_cast %[[T7]] : !llvm.array<2 x vector<2xf32>> to vector<2x2xf32>
 //       CHECK:    return %[[VAL_12]] : vector<2x2xf32>
 
 // -----
 

 
 func @insert_strided_slice2(%a: vector<2x2xf32>, %b: vector<4x4xf32>) -> vector<4x4xf32> {
   %0 = vector.insert_strided_slice %a, %b {offsets = [2, 2], strides = [1, 1]} : vector<2x2xf32> into vector<4x4xf32>
   return %0 : vector<4x4xf32>
 }
 // CHECK-LABEL: @insert_strided_slice2
 //
 // Subvector vector<2xf32> @0 into vector<4xf32> @2
+//  CHECK:    unrealized_conversion_cast %{{.*}} : vector<4x4xf32> to !llvm.array<4 x vector<4xf32>>
 //       CHECK:    llvm.extractvalue {{.*}}[0] : !llvm.array<2 x vector<2xf32>>
-//  CHECK-NEXT:    unrealized_conversion_cast %{{.*}} : vector<4x4xf32> to !llvm.array<4 x vector<4xf32>>
 //  CHECK-NEXT:    llvm.extractvalue {{.*}}[2] : !llvm.array<4 x vector<4xf32>>
 // Element @0 -> element @2
 //  CHECK-NEXT:    arith.constant 0 : index
 //  CHECK-NEXT:    unrealized_conversion_cast %{{.*}} : index to i64
 //  CHECK-NEXT:    llvm.extractelement {{.*}}[{{.*}} : i64] : vector<2xf32>
 //  CHECK-NEXT:    arith.constant 2 : index
 //  CHECK-NEXT:    unrealized_conversion_cast %{{.*}} : index to i64
 //  CHECK-NEXT:    llvm.insertelement {{.*}}, {{.*}}[{{.*}} : i64] : vector<4xf32>
 // Element @1 -> element @3
 //  CHECK-NEXT:    arith.constant 1 : index
 //  CHECK-NEXT:    unrealized_conversion_cast %{{.*}} : index to i64
 //  CHECK-NEXT:    llvm.extractelement {{.*}}[{{.*}} : i64] : vector<2xf32>
 //  CHECK-NEXT:    arith.constant 3 : index
 //  CHECK-NEXT:    unrealized_conversion_cast %{{.*}} : index to i64
 //  CHECK-NEXT:    llvm.insertelement {{.*}}, {{.*}}[{{.*}} : i64] : vector<4xf32>
-//  CHECK-NEXT:    unrealized_conversion_cast %{{.*}} : vector<4x4xf32> to !llvm.array<4 x vector<4xf32>>
 //  CHECK-NEXT:    llvm.insertvalue {{.*}}, {{.*}}[2] : !llvm.array<4 x vector<4xf32>>
 //
 // Subvector vector<2xf32> @1 into vector<4xf32> @3
 //       CHECK:    llvm.extractvalue {{.*}}[1] : !llvm.array<2 x vector<2xf32>>
-//  CHECK-NEXT:    unrealized_conversion_cast %{{.*}} : vector<4x4xf32> to !llvm.array<4 x vector<4xf32>>
 //  CHECK-NEXT:    llvm.extractvalue {{.*}}[3] : !llvm.array<4 x vector<4xf32>>
 // Element @0 -> element @2
 //  CHECK-NEXT:    arith.constant 0 : index
 //  CHECK-NEXT:    unrealized_conversion_cast %{{.*}} : index to i64
 //  CHECK-NEXT:    llvm.extractelement {{.*}}[{{.*}} : i64] : vector<2xf32>
 //  CHECK-NEXT:    arith.constant 2 : index
 //  CHECK-NEXT:    unrealized_conversion_cast %{{.*}} : index to i64
 //  CHECK-NEXT:    llvm.insertelement {{.*}}, {{.*}}[{{.*}} : i64] : vector<4xf32>

 func @insert_strided_slice3(%arg0: vector<2x4xf32>, %arg1: vector<16x4x8xf32>) -> vector<16x4x8xf32> {
   %0 = vector.insert_strided_slice %arg0, %arg1 {offsets = [0, 0, 2], strides = [1, 1]}:
         vector<2x4xf32> into vector<16x4x8xf32>
   return %0 : vector<16x4x8xf32>
 }
 // CHECK-LABEL: @insert_strided_slice3(
 // CHECK-SAME: %[[A:.*]]: vector<2x4xf32>,
 // CHECK-SAME: %[[B:.*]]: vector<16x4x8xf32>)
-//      CHECK: %[[s2:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<16x4x8xf32> to !llvm.array<16 x array<4 x vector<8xf32>>>
+//      CHECK-DAG: %[[s2:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<16x4x8xf32> to !llvm.array<16 x array<4 x vector<8xf32>>>
+//      CHECK-DAG: %[[s4:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
 //      CHECK: %[[s3:.*]] = llvm.extractvalue %[[s2]][0] : !llvm.array<16 x array<4 x vector<8xf32>>>
-//      CHECK: %[[s4:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
 //      CHECK: %[[s5:.*]] = llvm.extractvalue %[[s4]][0] : !llvm.array<2 x vector<4xf32>>
-//      CHECK: %[[s6:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<16x4x8xf32> to !llvm.array<16 x array<4 x vector<8xf32>>>
-//      CHECK: %[[s7:.*]] = llvm.extractvalue %[[s6]][0, 0] : !llvm.array<16 x array<4 x vector<8xf32>>>
+//      CHECK: %[[s7:.*]] = llvm.extractvalue %[[s2]][0, 0] : !llvm.array<16 x array<4 x vector<8xf32>>>
 //      CHECK: %[[s8:.*]] = arith.constant 0 : index
 //      CHECK: %[[s9:.*]] = builtin.unrealized_conversion_cast %[[s8]] : index to i64
 //      CHECK: %[[s10:.*]] = llvm.extractelement %[[s5]]{{\[}}%[[s9]] : i64] : vector<4xf32>
 //      CHECK: %[[s11:.*]] = arith.constant 2 : index
 //      CHECK: %[[s12:.*]] = builtin.unrealized_conversion_cast %[[s11]] : index to i64
 //      CHECK: %[[s13:.*]] = llvm.insertelement %[[s10]], %[[s7]]{{\[}}%[[s12]] : i64] : vector<8xf32>
 //      CHECK: %[[s14:.*]] = arith.constant 1 : index
 //      CHECK: %[[s15:.*]] = builtin.unrealized_conversion_cast %[[s14]] : index to i64

 //      CHECK: %[[s25:.*]] = llvm.insertelement %[[s22]], %[[s19]]{{\[}}%[[s24]] : i64] : vector<8xf32>
 //      CHECK: %[[s26:.*]] = arith.constant 3 : index
 //      CHECK: %[[s27:.*]] = builtin.unrealized_conversion_cast %[[s26]] : index to i64
 //      CHECK: %[[s28:.*]] = llvm.extractelement %[[s5]]{{\[}}%[[s27]] : i64] : vector<4xf32>
 //      CHECK: %[[s29:.*]] = arith.constant 5 : index
 //      CHECK: %[[s30:.*]] = builtin.unrealized_conversion_cast %[[s29]] : index to i64
 //      CHECK: %[[s31:.*]] = llvm.insertelement %[[s28]], %[[s25]]{{\[}}%[[s30]] : i64] : vector<8xf32>
 //      CHECK: %[[s32:.*]] = llvm.insertvalue %[[s31]], %[[s3]][0] : !llvm.array<4 x vector<8xf32>>
-//      CHECK: %[[s33:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
-//      CHECK: %[[s34:.*]] = llvm.extractvalue %[[s33]][1] : !llvm.array<2 x vector<4xf32>>
-//      CHECK: %[[s35:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<16x4x8xf32> to !llvm.array<16 x array<4 x vector<8xf32>>>
-//      CHECK: %[[s36:.*]] = llvm.extractvalue %[[s35]][0, 1] : !llvm.array<16 x array<4 x vector<8xf32>>>
+//      CHECK: %[[s34:.*]] = llvm.extractvalue %[[s4]][1] : !llvm.array<2 x vector<4xf32>>
+//      CHECK: %[[s36:.*]] = llvm.extractvalue %[[s2]][0, 1] : !llvm.array<16 x array<4 x vector<8xf32>>>
 //      CHECK: %[[s37:.*]] = arith.constant 0 : index
 //      CHECK: %[[s38:.*]] = builtin.unrealized_conversion_cast %[[s37]] : index to i64
 //      CHECK: %[[s39:.*]] = llvm.extractelement %[[s34]]{{\[}}%[[s38]] : i64] : vector<4xf32>
 //      CHECK: %[[s40:.*]] = arith.constant 2 : index
 //      CHECK: %[[s41:.*]] = builtin.unrealized_conversion_cast %[[s40]] : index to i64
 //      CHECK: %[[s42:.*]] = llvm.insertelement %[[s39]], %[[s36]]{{\[}}%[[s41]] : i64] : vector<8xf32>
 //      CHECK: %[[s43:.*]] = arith.constant 1 : index
 //      CHECK: %[[s44:.*]] = builtin.unrealized_conversion_cast %[[s43]] : index to i64

 //      CHECK: %[[s54:.*]] = llvm.insertelement %[[s51]], %[[s48]]{{\[}}%[[s53]] : i64] : vector<8xf32>
 //      CHECK: %[[s55:.*]] = arith.constant 3 : index
 //      CHECK: %[[s56:.*]] = builtin.unrealized_conversion_cast %[[s55]] : index to i64
 //      CHECK: %[[s57:.*]] = llvm.extractelement %[[s34]]{{\[}}%[[s56]] : i64] : vector<4xf32>
 //      CHECK: %[[s58:.*]] = arith.constant 5 : index
 //      CHECK: %[[s59:.*]] = builtin.unrealized_conversion_cast %[[s58]] : index to i64
 //      CHECK: %[[s60:.*]] = llvm.insertelement %[[s57]], %[[s54]]{{\[}}%[[s59]] : i64] : vector<8xf32>
 //      CHECK: %[[s61:.*]] = llvm.insertvalue %[[s60]], %[[s32]][1] : !llvm.array<4 x vector<8xf32>>
-//      CHECK: %[[s62:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<16x4x8xf32> to !llvm.array<16 x array<4 x vector<8xf32>>>
-//      CHECK: %[[s63:.*]] = llvm.insertvalue %[[s61]], %[[s62]][0] : !llvm.array<16 x array<4 x vector<8xf32>>>
+//      CHECK: %[[s63:.*]] = llvm.insertvalue %[[s61]], %[[s2]][0] : !llvm.array<16 x array<4 x vector<8xf32>>>
 //      CHECK: %[[s64:.*]] = builtin.unrealized_conversion_cast %[[s63]] : !llvm.array<16 x array<4 x vector<8xf32>>> to vector<16x4x8xf32>
 //      CHECK: return %[[s64]] : vector<16x4x8xf32>
 
 // -----
 
 func @vector_fma(%a: vector<8xf32>, %b: vector<2x4xf32>) -> (vector<8xf32>, vector<2x4xf32>) {
   // CHECK-LABEL: @vector_fma
   //  CHECK-SAME: %[[A:.*]]: vector<8xf32>
   //  CHECK-SAME: %[[B:.*]]: vector<2x4xf32>
+  //       CHECK: %[[BL:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
   //       CHECK: "llvm.intr.fmuladd"
   //  CHECK-SAME:   (vector<8xf32>, vector<8xf32>, vector<8xf32>) -> vector<8xf32>
   %0 = vector.fma %a, %a, %a : vector<8xf32>
 
-  //       CHECK: %[[BL:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
   //       CHECK: %[[b00:.*]] = llvm.extractvalue %[[BL]][0] : !llvm.array<2 x vector<4xf32>>
-  //       CHECK: %[[BL:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
   //       CHECK: %[[b01:.*]] = llvm.extractvalue %[[BL]][0] : !llvm.array<2 x vector<4xf32>>
-  //       CHECK: %[[BL:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
   //       CHECK: %[[b02:.*]] = llvm.extractvalue %[[BL]][0] : !llvm.array<2 x vector<4xf32>>
   //       CHECK: %[[B0:.*]] = "llvm.intr.fmuladd"(%[[b00]], %[[b01]], %[[b02]]) :
   //  CHECK-SAME: (vector<4xf32>, vector<4xf32>, vector<4xf32>) -> vector<4xf32>
   //       CHECK: llvm.insertvalue %[[B0]], {{.*}}[0] : !llvm.array<2 x vector<4xf32>>
-  //       CHECK: %[[BL:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
   //       CHECK: %[[b10:.*]] = llvm.extractvalue %[[BL]][1] : !llvm.array<2 x vector<4xf32>>
-  //       CHECK: %[[BL:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
   //       CHECK: %[[b11:.*]] = llvm.extractvalue %[[BL]][1] : !llvm.array<2 x vector<4xf32>>
-  //       CHECK: %[[BL:.*]] = builtin.unrealized_conversion_cast %[[B]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
   //       CHECK: %[[b12:.*]] = llvm.extractvalue %[[BL]][1] : !llvm.array<2 x vector<4xf32>>
   //       CHECK: %[[B1:.*]] = "llvm.intr.fmuladd"(%[[b10]], %[[b11]], %[[b12]]) :
   //  CHECK-SAME: (vector<4xf32>, vector<4xf32>, vector<4xf32>) -> vector<4xf32>
   //       CHECK: llvm.insertvalue %[[B1]], {{.*}}[1] : !llvm.array<2 x vector<4xf32>>
   %1 = vector.fma %b, %b, %b : vector<2x4xf32>
 
   return %0, %1: vector<8xf32>, vector<2x4xf32>
 }

mlir/test/Dialect/ArmSVE/memcpy.mlir

 // RUN: mlir-opt %s -convert-vector-to-llvm="enable-arm-sve" | mlir-opt | FileCheck %s
 
 // CHECK: memcopy([[SRC:%arg[0-9]+]]: memref<?xf32>, [[DST:%arg[0-9]+]]
 func @memcopy(%src : memref<?xf32>, %dst : memref<?xf32>, %size : index) {
   %c0 = arith.constant 0 : index
   %c4 = arith.constant 4 : index
   %vs = arm_sve.vector_scale : index
   %step = arith.muli %c4, %vs : index
 
+  // CHECK: [[DSTMRS:%[0-9]+]] = builtin.unrealized_conversion_cast [[DST]] : memref<?xf32> to !llvm.struct<(ptr<f32>
+  // CHECK: [[SRCMRS:%[0-9]+]] = builtin.unrealized_conversion_cast [[SRC]] : memref<?xf32> to !llvm.struct<(ptr<f32>
   // CHECK: scf.for [[LOOPIDX:%arg[0-9]+]] = {{.*}}
   scf.for %i0 = %c0 to %size step %step {
-    // CHECK: [[SRCMRS:%[0-9]+]] = builtin.unrealized_conversion_cast [[SRC]] : memref<?xf32> to !llvm.struct<(ptr<f32>
     // CHECK: [[SRCIDX:%[0-9]+]] = builtin.unrealized_conversion_cast [[LOOPIDX]] : index to i64
     // CHECK: [[SRCMEM:%[0-9]+]] = llvm.extractvalue [[SRCMRS]][1] : !llvm.struct<(ptr<f32>
     // CHECK-NEXT: [[SRCPTR:%[0-9]+]] = llvm.getelementptr [[SRCMEM]]{{.}}[[SRCIDX]]{{.}} : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
     // CHECK-NEXT: [[SRCVPTR:%[0-9]+]] = llvm.bitcast [[SRCPTR]] : !llvm.ptr<f32> to !llvm.ptr<vec<? x 4 x f32>>
     // CHECK-NEXT: [[LDVAL:%[0-9]+]] = llvm.load [[SRCVPTR]] : !llvm.ptr<vec<? x 4 x f32>>
     %0 = arm_sve.load %src[%i0] : !arm_sve.vector<4xf32> from memref<?xf32>
-    // CHECK: [[DSTMRS:%[0-9]+]] = builtin.unrealized_conversion_cast [[DST]] : memref<?xf32> to !llvm.struct<(ptr<f32>
-    // CHECK: [[DSTIDX:%[0-9]+]] = builtin.unrealized_conversion_cast [[LOOPIDX]] : index to i64
     // CHECK: [[DSTMEM:%[0-9]+]] = llvm.extractvalue [[DSTMRS]][1] : !llvm.struct<(ptr<f32>
-    // CHECK-NEXT: [[DSTPTR:%[0-9]+]] = llvm.getelementptr [[DSTMEM]]{{.}}[[DSTIDX]]{{.}} : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
+    // CHECK-NEXT: [[DSTPTR:%[0-9]+]] = llvm.getelementptr [[DSTMEM]]{{.}}[[SRCIDX]]{{.}} : (!llvm.ptr<f32>, i64) -> !llvm.ptr<f32>
     // CHECK-NEXT: [[DSTVPTR:%[0-9]+]] = llvm.bitcast [[DSTPTR]] : !llvm.ptr<f32> to !llvm.ptr<vec<? x 4 x f32>>
     // CHECK-NEXT: llvm.store [[LDVAL]], [[DSTVPTR]] : !llvm.ptr<vec<? x 4 x f32>>
     arm_sve.store %0, %dst[%i0] : !arm_sve.vector<4xf32> to memref<?xf32>
   }
 
   return
 }

mlir/test/Dialect/Linalg/bufferize.mlir

#trait = {
   iterator_types = ["parallel", "parallel", "reduction"]
 }
 
 // Check the bufferization of init tensors.
 
 // CHECK-LABEL:   func @generic_with_init_tensor(
 // CHECK-SAME:                                   %[[ARG0_TENSOR:.*]]: tensor<2x3x4xvector<3x4xi4>>,
 // CHECK-SAME:                                   %[[ARG1_TENSOR:.*]]: tensor<3x2xf32>) -> tensor<3x2xf32> {
-// CHECK:           %[[ARG0_MEMREF:.*]] = memref.buffer_cast %[[ARG0_TENSOR]] : memref<2x3x4xvector<3x4xi4>>
-// CHECK:           %[[ARG1_MEMREF:.*]] = memref.buffer_cast %[[ARG1_TENSOR]] : memref<3x2xf32>
+// CHECK-DAG:           %[[ARG0_MEMREF:.*]] = memref.buffer_cast %[[ARG0_TENSOR]] : memref<2x3x4xvector<3x4xi4>>
+// CHECK-DAG:           %[[ARG1_MEMREF:.*]] = memref.buffer_cast %[[ARG1_TENSOR]] : memref<3x2xf32>
 // CHECK:           %[[INIT_BUFFER:.*]] = memref.alloc() : memref<3x2xf32>
 // CHECK:           linalg.copy(%[[ARG1_MEMREF]], %[[INIT_BUFFER]]) : memref<3x2xf32>, memref<3x2xf32>
 // CHECK:           linalg.generic
 // CHECK-SAME:      ins(%[[ARG0_MEMREF]] : memref<2x3x4xvector<3x4xi4>>)
 // CHECK-SAME:      outs(%[[INIT_BUFFER]] : memref<3x2xf32>) {
 func @generic_with_init_tensor(%arg0: tensor<2x3x4xvector<3x4xi4>>,
   %arg1: tensor<3x2xf32>) -> (tensor<3x2xf32>) {
 

 // CHECK-DAG: #[[$MAP0:[0-9a-z]*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)>
 // CHECK-DAG: #[[$MAP1:[0-9a-z]*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1 * 2)>
 
 func private @make_index() -> index
 
 // CHECK-LABEL: func @bufferize_slice(
 //  CHECK-SAME:   %[[T:[0-9a-z]*]]: tensor<?x?xf32>
 func @bufferize_slice(%t : tensor<?x?xf32>) -> (tensor<2x3xf32>, tensor<2x?xf32>) {
+  //      CHECK: %[[M:.*]] = memref.buffer_cast %[[T]] : memref<?x?xf32>
+
   //      CHECK: %[[IDX:.*]] = call @make_index() : () -> index
   %i0 = call @make_index() : () -> index
 
-  //      CHECK: %[[M:.*]] = memref.buffer_cast %[[T]] : memref<?x?xf32>
   // CHECK-NEXT: %[[A0:.*]] = memref.alloc() : memref<2x3xf32>
   // CHECK-NEXT: %[[SM0:.*]] = memref.subview %[[M]][0, 0] [2, 3] [1, 1]
   // CHECK-SAME:   memref<?x?xf32> to memref<2x3xf32, #[[$MAP0]]>
   // CHECK-NEXT: linalg.copy(%[[SM0]], %[[A0]]) : memref<2x3xf32, #[[$MAP0]]>, memref<2x3xf32>
   // CHECK-NEXT: %[[RT0:.*]] = memref.tensor_load %[[A0]] : memref<2x3xf32>
   %st0 = tensor.extract_slice %t[0, 0][2, 3][1, 1] : tensor<?x?xf32> to tensor<2x3xf32>
 
   // CHECK-NEXT: %[[A1:.*]] = memref.alloc(%[[IDX]]) : memref<2x?xf32>

 func private @make_index() -> index
 
 // CHECK-LABEL: func @bufferize_insert_slice(
 //  CHECK-SAME:   %[[T:[0-9a-z]*]]: tensor<?x?xf32>
 //  CHECK-SAME:   %[[ST0:[0-9a-z]*]]: tensor<2x3xf32>
 //  CHECK-SAME:   %[[ST1:[0-9a-z]*]]: tensor<2x?xf32>
 func @bufferize_insert_slice(%t : tensor<?x?xf32>, %st0 : tensor<2x3xf32>, %st1 : tensor<2x?xf32>) ->
     (tensor<?x?xf32>, tensor<?x?xf32>) {
+  // CHECK-DAG: %[[M:.*]] = memref.buffer_cast %[[T]] : memref<?x?xf32>
+  // CHECK-DAG: %[[SM0:.*]] = memref.buffer_cast %[[ST0]] : memref<2x3xf32>
+  // CHECK-DAG: %[[SM1:.*]] = memref.buffer_cast %[[ST1]] : memref<2x?xf32>
+
   %c0 = arith.constant 0 : index
   %c1 = arith.constant 1 : index
   // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
   // CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index
   %i0 = call @make_index() : () -> index
   // CHECK: %[[IDX:.*]] = call @make_index() : () -> index
 
 
-  // CHECK-DAG: %[[M:.*]] = memref.buffer_cast %[[T]] : memref<?x?xf32>
-  // CHECK-DAG: %[[SM0:.*]] = memref.buffer_cast %[[ST0]] : memref<2x3xf32>
   // CHECK-NEXT: %[[DIM0:.*]] = tensor.dim %[[T]], %[[C0]] : tensor<?x?xf32>
   // CHECK-NEXT: %[[DIM1:.*]] = tensor.dim %[[T]], %[[C1]] : tensor<?x?xf32>
   // CHECK-NEXT: %[[M_COPY0:.*]] = memref.alloc(%[[DIM0]], %[[DIM1]]) : memref<?x?xf32>
   // CHECK-NEXT: linalg.copy(%[[M]], %[[M_COPY0]]) : memref<?x?xf32>, memref<?x?xf32>
   // CHECK-NEXT: %[[SUBVIEW0:.*]] = memref.subview %[[M_COPY0]][0, 0] [2, 3] [1, 1]
   // CHECK-SAME:   memref<?x?xf32> to memref<2x3xf32, #[[$MAP0]]>
   // CHECK-NEXT: linalg.copy(%[[SM0]], %[[SUBVIEW0]]) : memref<2x3xf32>, memref<2x3xf32, #[[$MAP0]]>
   // CHECK-NEXT: %[[RT0:.*]] = memref.tensor_load %[[M_COPY0]] : memref<?x?xf32>
   %t0 = tensor.insert_slice %st0 into %t[0, 0][2, 3][1, 1] : tensor<2x3xf32> into tensor<?x?xf32>
 
-  //  CHECK-DAG: %[[SM1:.*]] = memref.buffer_cast %[[ST1]] : memref<2x?xf32>
   // CHECK-NEXT: %[[M_COPY1:.*]] = memref.alloc(%[[DIM0]], %[[DIM1]]) : memref<?x?xf32>
   // CHECK-NEXT: linalg.copy(%[[M]], %[[M_COPY1]]) : memref<?x?xf32>, memref<?x?xf32>
   // CHECK-NEXT: %[[SUBVIEW1:.*]] = memref.subview %[[M_COPY1]][0, %[[IDX]]] [2, %[[IDX]]] [1, 2]
   // CHECK-SAME:   memref<?x?xf32> to memref<2x?xf32, #[[$MAP1]]>
   // CHECK-NEXT: linalg.copy(%[[SM1]], %[[SUBVIEW1]]) : memref<2x?xf32>, memref<2x?xf32, #[[$MAP1]]>
   // CHECK-NEXT: %[[RT1:.*]] = memref.tensor_load %[[M_COPY1]] : memref<?x?xf32>
   %t1 = tensor.insert_slice %st1 into %t[0, %i0][2, %i0][1, 2] : tensor<2x?xf32> into tensor<?x?xf32>
 
func @pad_tensor_dynamic_shape(%arg0: tensor<4x?x2x?xf32>, %arg1: index) -> tensor<4x?x?x?xf32> {
   } : tensor<4x?x2x?xf32> to tensor<4x?x?x?xf32>
   return %out : tensor<4x?x?x?xf32>
 }
 
 // CHECK-DAG:       %[[C3:.*]] = arith.constant 3 : index
 // CHECK-DAG:       %[[C2:.*]] = arith.constant 2 : index
 // CHECK-DAG:       %[[C1:.*]] = arith.constant 1 : index
 // CHECK-DAG:       %[[CST:.*]] = arith.constant 0.000000e+00 : f32
+// CHECK:           %[[IN_MEMREF:.*]] = memref.buffer_cast %[[IN]] : memref<4x?x2x?xf32>
 // CHECK:           %[[DIM1:.*]] = tensor.dim %[[IN]], %[[C1]] : tensor<4x?x2x?xf32>
 // CHECK:           %[[OUT_DIM2:.*]] = arith.addi %[[OFFSET]], %[[C2]] : index
 // CHECK:           %[[DIM3:.*]] = tensor.dim %[[IN]], %[[C3]] : tensor<4x?x2x?xf32>
 // CHECK:           %[[OUT_DIM3:.*]] = arith.addi %[[DIM3]], %[[OFFSET]] : index
 // CHECK:           %[[FILLED:.*]] = memref.alloc(%[[DIM1]], %[[OUT_DIM2]], %[[OUT_DIM3]]) : memref<4x?x?x?xf32>
 // CHECK:           linalg.fill(%[[CST]], %[[FILLED]]) : f32, memref<4x?x?x?xf32>
-// CHECK:           %[[IN_MEMREF:.*]] = memref.buffer_cast %[[IN]] : memref<4x?x2x?xf32>
 // CHECK:           %[[OUT:.*]] = memref.alloc(%[[DIM1]], %[[OUT_DIM2]], %[[OUT_DIM3]]) : memref<4x?x?x?xf32>
 // CHECK:           linalg.copy(%[[FILLED]], %[[OUT]]) : memref<4x?x?x?xf32>, memref<4x?x?x?xf32>
 // CHECK:           %[[INTERIOR:.*]] = memref.subview %[[OUT]][0, 0, %[[OFFSET]], 0] [4, %[[DIM1]], 2, %[[DIM3]]] [1, 1, 1, 1] : memref<4x?x?x?xf32> to memref<4x?x2x?xf32, #map>
 // CHECK:           linalg.copy(%[[IN_MEMREF]], %[[INTERIOR]]) : memref<4x?x2x?xf32>, memref<4x?x2x?xf32, #map>
 // CHECK:           %[[OUT_TENSOR:.*]] = memref.tensor_load %[[OUT]] : memref<4x?x?x?xf32>
 // CHECK:           return %[[OUT_TENSOR]] : tensor<4x?x?x?xf32>
 // CHECK:         }
 

mlir/test/Dialect/Linalg/detensorize_0d.mlir

func @detensor_op_sequence(%arg1: tensor<f32>, %arg2: tensor<f32>) -> tensor<f32> attributes {iree.module.export} {
 
   return %7: tensor<f32>
 }
 // CHECK-LABEL: func @detensor_op_sequence
 // CHECK-SAME:    (%[[arg1:.*]]: tensor<f32>, %[[arg2:.*]]: tensor<f32>)
 // CHECK-DAG:     %[[arg1_val:.*]] = tensor.extract %[[arg1]]
 // CHECK-DAG:     %[[arg2_val:.*]] = tensor.extract %[[arg2]]
 // CHECK:         %[[detensored_res:.*]] = arith.addf %[[arg1_val]], %[[arg2_val]]
-// CHECK-DAG:     %[[arg1_val2:.*]] = tensor.extract %[[arg1]]
-// CHECK:         %[[detensored_res2:.*]] = arith.mulf %[[arg1_val2]], %[[detensored_res]]
+// CHECK:         %[[detensored_res2:.*]] = arith.mulf %[[arg1_val]], %[[detensored_res]]
 // CHECK:         %[[detensored_res3:.*]] = arith.divf %[[detensored_res]], %[[detensored_res2]]
 // CHECK:         %[[new_tensor_res:.*]] = tensor.from_elements %[[detensored_res3]]
 // CHECK:         %[[reshaped_tensor_res:.*]] = linalg.tensor_collapse_shape %[[new_tensor_res]]
 // CHECK:         return %[[reshaped_tensor_res]]
 
 func @detensor_multiple_ops(%arg1: tensor<f32>, %arg2: tensor<f32>) -> tensor<f32> attributes {iree.module.export} {
   %0 = linalg.init_tensor [] : tensor<f32>
   %1 = linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = []}

mlir/test/Dialect/Linalg/detensorize_while_impure_cf.mlir

 // DET-ALL:       ^[[bb1]](%{{.*}}: tensor<10xi32>)
 // DET-ALL:         linalg.init_tensor [] : tensor<i32>
 // DET-ALL:         linalg.generic {{{.*}}} ins(%{{.*}} : tensor<10xi32>) outs(%{{.*}} : tensor<i32>) {
 // DET-ALL:         ^bb0(%{{.*}}: i32, %{{.*}}: i32):  // no predecessors
 // DET-ALL:           %{{.*}} = arith.addi %{{.*}}, %{{.*}}
 // DET-ALL:           linalg.yield %{{.*}} : i32
 // DET-ALL:         } -> tensor<i32>
 // DET-ALL:         tensor.extract %{{.*}}[] : tensor<i32>
-// DET-ALL:         tensor.extract %{{.*}}[] : tensor<i32>
-// DET-ALL:         arith.cmpi slt, %{{.*}}, %{{.*}} : i32
-// DET-ALL:         tensor.extract %{{.*}}[] : tensor<i32>
-// DET-ALL:         tensor.extract %{{.*}}[] : tensor<i32>
+// DET-ALL:         cmpi slt, %{{.*}}, %{{.*}} : i32
 // DET-ALL:         cond_br %{{.*}}, ^[[bb2:.*]](%{{.*}} : i32), ^[[bb3:.*]](%{{.*}} : i32)
 // DET-ALL:       ^[[bb2]](%{{.*}}: i32)
 // DET-ALL:         tensor.from_elements %{{.*}} : tensor<1xi32>
 // DET-ALL:         linalg.tensor_collapse_shape %{{.*}} [] : tensor<1xi32> into tensor<i32>
 // DET-ALL:         linalg.init_tensor [10] : tensor<10xi32>
 // DET-ALL:         linalg.generic {{{.*}}} ins(%{{.*}} : tensor<i32>) outs(%{{.*}} : tensor<10xi32>) {
 // DET-ALL:         ^bb0(%{{.*}}: i32, %{{.*}}: i32):
 // DET-ALL:           linalg.yield %{{.*}} : i32
 // DET-ALL:         } -> tensor<10xi32>
 // DET-ALL:         br ^[[bb1]](%{{.*}} : tensor<10xi32>)
 // DET-ALL:       ^[[bb3]](%{{.*}}: i32)
 // DET-ALL:         tensor.from_elements %{{.*}} : tensor<1xi32>
 // DET-ALL:         linalg.tensor_collapse_shape %{{.*}} [] : tensor<1xi32> into tensor<i32>
 // DET-ALL:         return %{{.*}} : tensor<i32>
 // DET-ALL:       }
 
 // DET-CF-LABEL: func @main
 // DET-CF-SAME:    (%{{.*}}: tensor<10xi32>, %{{.*}}: tensor<i32>)
 // DET-CF:         br ^[[bb1:.*]](%{{.*}} : tensor<10xi32>)
 // DET-CF:       ^bb1(%{{.*}}: tensor<10xi32>)
 // DET-CF:         %{{.*}} = linalg.generic {{{.*}}} ins(%{{.*}} : tensor<10xi32>) outs(%{{.*}} : tensor<i32>) {
 // DET-CF:         tensor.extract %{{.*}}[] : tensor<i32>
-// DET-CF:         tensor.extract %{{.*}}[] : tensor<i32>
-// DET-CF:         arith.cmpi slt, %{{.*}}, %{{.*}} : i32
+// DET-CF:         cmpi slt, %{{.*}}, %{{.*}} : i32
 // DET-CF:         cond_br %{{.*}}, ^bb2(%{{.*}} : tensor<i32>), ^bb3(%{{.*}} : tensor<i32>)
 // DET-CF:       ^bb2(%{{.*}}: tensor<i32>)
 // DET-CF:         %{{.*}} = linalg.generic {{{.*}}} ins(%{{.*}} : tensor<i32>) outs(%{{.*}} : tensor<10xi32>) {
 // DET-CF:         br ^bb1(%{{.*}} : tensor<10xi32>)
 // DET-CF:       ^bb3(%{{.*}}: tensor<i32>)
 // DET-CF:         return %{{.*}} : tensor<i32>
 // DET-CF:       }

mlir/test/Dialect/SCF/bufferize.mlir

 // RUN: mlir-opt %s -scf-bufferize | FileCheck %s
 
 // CHECK-LABEL:   func @if(
 // CHECK-SAME:             %[[PRED:.*]]: i1,
 // CHECK-SAME:             %[[TRUE_TENSOR:.*]]: tensor<?xf32>,
 // CHECK-SAME:             %[[FALSE_TENSOR:.*]]: tensor<?xf32>) -> tensor<?xf32> {
-// CHECK:           %[[RESULT_MEMREF:.*]] = scf.if %[[PRED]] -> (memref<?xf32>) {
-// CHECK:             %[[TRUE_MEMREF:.*]] = memref.buffer_cast %[[TRUE_TENSOR]] : memref<?xf32>
+// CHECK:           %[[FALSE_MEMREF:.*]] = memref.buffer_cast %[[FALSE_TENSOR]] : memref<?xf32>
+// CHECK:           %[[TRUE_MEMREF:.*]] = memref.buffer_cast %[[TRUE_TENSOR]] : memref<?xf32>
+// CHECK:           %[[RESULT_MEMREF:.*]] = scf.if %[[PRED]] -> (memref<?xf32>) {
 // CHECK:             scf.yield %[[TRUE_MEMREF]] : memref<?xf32>
 // CHECK:           } else {
-// CHECK:             %[[FALSE_MEMREF:.*]] = memref.buffer_cast %[[FALSE_TENSOR]] : memref<?xf32>
 // CHECK:             scf.yield %[[FALSE_MEMREF]] : memref<?xf32>
 // CHECK:           }
 // CHECK:           %[[RESULT_TENSOR:.*]] = memref.tensor_load %[[RESULT_MEMREF:.*]] : memref<?xf32>
 // CHECK:           return %[[RESULT_TENSOR]] : tensor<?xf32>
 // CHECK:         }
 func @if(%pred: i1, %true_val: tensor<?xf32>, %false_val: tensor<?xf32>) -> tensor<?xf32> {
   %0 = scf.if %pred -> (tensor<?xf32>) {
     scf.yield %true_val : tensor<?xf32>
   } else {
     scf.yield %false_val : tensor<?xf32>
   }
   return %0 : tensor<?xf32>
 }
 
 // CHECK-LABEL:   func @for(
 // CHECK-SAME:              %[[TENSOR:.*]]: tensor<f32>,
 // CHECK-SAME:              %[[LB:.*]]: index, %[[UB:.*]]: index,
 // CHECK-SAME:              %[[STEP:.*]]: index) -> tensor<f32> {
 // CHECK:           %[[MEMREF:.*]] = memref.buffer_cast %[[TENSOR]] : memref<f32>
 // CHECK:           %[[RESULT_MEMREF:.*]] = scf.for %[[VAL_6:.*]] = %[[LB]] to %[[UB]] step %[[STEP]] iter_args(%[[ITER:.*]] = %[[MEMREF]]) -> (memref<f32>) {
-// CHECK:             %[[TENSOR_ITER:.*]] = memref.tensor_load %[[ITER]] : memref<f32>
-// CHECK:             %[[MEMREF_YIELDED:.*]] = memref.buffer_cast %[[TENSOR_ITER]] : memref<f32>
-// CHECK:             scf.yield %[[MEMREF_YIELDED]] : memref<f32>
+// CHECK:             scf.yield %[[ITER]] : memref<f32>
 // CHECK:           }
 // CHECK:           %[[VAL_8:.*]] = memref.tensor_load %[[VAL_9:.*]] : memref<f32>
 // CHECK:           return %[[VAL_8]] : tensor<f32>
 // CHECK:         }
 func @for(%arg0: tensor<f32>, %lb: index, %ub: index, %step: index) -> tensor<f32> {
   %ret = scf.for %iv = %lb to %ub step %step iter_args(%iter = %arg0) -> tensor<f32> {
     scf.yield %iter : tensor<f32>
   }

mlir/test/Dialect/Standard/bufferize.mlir

 // RUN: mlir-opt %s -std-bufferize | FileCheck %s
 
 // CHECK-LABEL:   func @select(
 // CHECK-SAME:                 %[[PRED:.*]]: i1,
 // CHECK-SAME:                 %[[TRUE_VAL:.*]]: tensor<f32>,
 // CHECK-SAME:                 %[[FALSE_VAL:.*]]: tensor<f32>) -> tensor<f32> {
-// CHECK:           %[[TRUE_VAL_MEMREF:.*]] = memref.buffer_cast %[[TRUE_VAL]] : memref<f32>
-// CHECK:           %[[FALSE_VAL_MEMREF:.*]] = memref.buffer_cast %[[FALSE_VAL]] : memref<f32>
+// CHECK-DAG:           %[[TRUE_VAL_MEMREF:.*]] = memref.buffer_cast %[[TRUE_VAL]] : memref<f32>
+// CHECK-DAG:           %[[FALSE_VAL_MEMREF:.*]] = memref.buffer_cast %[[FALSE_VAL]] : memref<f32>
 // CHECK:           %[[RET_MEMREF:.*]] = select %[[PRED]], %[[TRUE_VAL_MEMREF]], %[[FALSE_VAL_MEMREF]] : memref<f32>
 // CHECK:           %[[RET:.*]] = memref.tensor_load %[[RET_MEMREF]] : memref<f32>
 // CHECK:           return %[[RET]] : tensor<f32>
 func @select(%arg0: i1, %arg1: tensor<f32>, %arg2: tensor<f32>) -> tensor<f32> {
   %0 = select %arg0, %arg1, %arg2 : tensor<f32>
   return %0 : tensor<f32>
 }

mlir/test/Dialect/Standard/func-bufferize.mlir

 // RUN: mlir-opt %s -func-bufferize -split-input-file -verify-diagnostics | FileCheck %s
 
 // CHECK-LABEL:   func @identity(
 // CHECK-SAME:                   %[[ARG:.*]]: memref<f32>) -> memref<f32> {
-// CHECK:           %[[TENSOR:.*]] = memref.tensor_load %[[ARG]] : memref<f32>
-// CHECK:           %[[MEMREF:.*]] = memref.buffer_cast %[[TENSOR]] : memref<f32>
-// CHECK:           return %[[MEMREF]] : memref<f32>
+// CHECK:           return %[[ARG]] : memref<f32>
 func @identity(%arg0: tensor<f32>) -> tensor<f32> {
   return %arg0 : tensor<f32>
 }
 
 // CHECK-LABEL:   func @block_arguments(
 // CHECK-SAME:        %[[ARG:.*]]: memref<f32>) -> memref<f32> {
-// CHECK:           %[[T1:.*]] = memref.tensor_load %[[ARG]] : memref<f32>
-// CHECK:           %[[M1:.*]] = memref.buffer_cast %[[T1]] : memref<f32>
-// CHECK:           br ^bb1(%[[M1]] : memref<f32>)
+// CHECK:           br ^bb1(%[[ARG]] : memref<f32>)
 // CHECK:         ^bb1(%[[BBARG:.*]]: memref<f32>):
-// CHECK:           %[[T2:.*]] = memref.tensor_load %[[BBARG]] : memref<f32>
-// CHECK:           %[[M2:.*]] = memref.buffer_cast %[[T2]] : memref<f32>
-// CHECK:           return %[[M2]] : memref<f32>
+// CHECK:           return %[[BBARG]] : memref<f32>
 func @block_arguments(%arg0: tensor<f32>) -> tensor<f32> {
   br ^bb1(%arg0: tensor<f32>)
 ^bb1(%bbarg: tensor<f32>):
   return %bbarg : tensor<f32>
 }
 
 // CHECK-LABEL:   func private @source() -> memref<f32>
 // CHECK-LABEL:   func @call_source() -> memref<f32> {
 // CHECK:           %[[RET:.*]] = call @source() : () -> memref<f32>
 // CHECK:           return %[[RET]] : memref<f32>
 func private @source() -> tensor<f32>
 func @call_source() -> tensor<f32> {
   %0 = call @source() : () -> tensor<f32>
   return %0 : tensor<f32>
 }
 // CHECK-LABEL:   func @call_sink(
 // CHECK-SAME:                    %[[ARG:.*]]: memref<f32>) {
-// CHECK:           %[[TENSOR:.*]] = memref.tensor_load %[[ARG]] : memref<f32>
-// CHECK:           %[[MEMREF:.*]] = memref.buffer_cast %[[TENSOR]] : memref<f32>
-// CHECK:           call @sink(%[[MEMREF]]) : (memref<f32>) -> ()
+// CHECK:           call @sink(%[[ARG]]) : (memref<f32>) -> ()
 // CHECK:           return
 func private @sink(tensor<f32>)
 func @call_sink(%arg0: tensor<f32>) {
   call @sink(%arg0) : (tensor<f32>) -> ()
   return
 }
 
 // CHECK-LABEL:   func @unconverted_op_in_body() -> memref<f32> {

mlir/test/Dialect/Tensor/bufferize.mlir

 func @tensor.from_elements(%arg0: index, %arg1: index) -> tensor<2xindex> {
   %0 = tensor.from_elements %arg0, %arg1 : tensor<2xindex>
   return %0 : tensor<2xindex>
 }
 
 // CHECK-LABEL:   func @tensor.generate(
 // CHECK-SAME:                                       %[[ARG:.*]]: tensor<*xf32>,
 // CHECK-SAME:                                       %[[DYNAMIC_EXTENT:.*]]: index) -> tensor<?xindex> {
+// CHECK:           %[[CASTED:.*]] = memref.buffer_cast %[[ARG]] : memref<*xf32>
 // CHECK:           %[[MEMREF:.*]] = memref.alloc(%[[DYNAMIC_EXTENT]]) : memref<?xindex>
 // CHECK:           %[[C0:.*]] = arith.constant 0 : index
 // CHECK:           %[[C1:.*]] = arith.constant 1 : index
 // CHECK:           scf.parallel (%[[I:.*]]) = (%[[C0]]) to (%[[DYNAMIC_EXTENT]]) step (%[[C1]]) {
-// CHECK:             %[[CASTED:.*]] = memref.buffer_cast %[[ARG]] : memref<*xf32>
 // CHECK:             %[[ELEM:.*]] = memref.dim %[[CASTED]], %[[I]] : memref<*xf32>
 // CHECK:             store %[[ELEM]], %[[MEMREF]][%[[I]]] : memref<?xindex>
 // CHECK:             scf.yield
 // CHECK:           }
 // CHECK:           %[[RET:.*]] = memref.tensor_load %[[MEMREF]] : memref<?xindex>
 // CHECK:           return %[[RET]] : tensor<?xindex>
 // CHECK:         }
 func @tensor.generate(%arg: tensor<*xf32>, %dynamic_extent: index) -> tensor<?xindex> {

mlir/test/Transforms/test-legalize-remapped-value.mlir

 // RUN: mlir-opt %s -test-remapped-value | FileCheck %s
 
 // Simple test that exercises ConvertPatternRewriter::getRemappedValue.
+
+// CHECK-LABEL: func @remap_input_1_to_1
+// CHECK-SAME: (%[[ARG:.*]]: i32)
+// CHECK-NEXT: %[[VAL:.*]] = "test.one_variadic_out_one_variadic_in1"(%[[ARG]], %[[ARG]])
+// CHECK-NEXT: "test.one_variadic_out_one_variadic_in1"(%[[VAL]], %[[VAL]])
+
 func @remap_input_1_to_1(%arg0: i32) {
   %0 = "test.one_variadic_out_one_variadic_in1"(%arg0) : (i32) -> i32
   %1 = "test.one_variadic_out_one_variadic_in1"(%0) : (i32) -> i32
   "test.return"() : () -> ()
 }
-// CHECK-LABEL: func @remap_input_1_to_1
-// CHECK-SAME: (%[[ARG:.*]]: i32)
-// CHECK-NEXT: %[[VAL:.*]] = "test.one_variadic_out_one_variadic_in1"(%[[ARG]], %[[ARG]])
-// CHECK-NEXT: "test.one_variadic_out_one_variadic_in1"(%[[VAL]], %[[VAL]])
 
+// Test the case where an operation is converted before its operands are.
+
+// CHECK-LABEL: func @remap_unconverted
+// CHECK-NEXT: %[[VAL:.*]] = "test.type_producer"() : () -> f64
+// CHECK-NEXT: "test.type_consumer"(%[[VAL]]) : (f64)
+func @remap_unconverted() {
+  %region_result = "test.remapped_value_region"() ({
+    %result = "test.type_producer"() : () -> f32
+    "test.return"(%result) : (f32) -> ()
+  }) : () -> (f32)
+  "test.type_consumer"(%region_result) : (f32) -> ()
+  "test.return"() : () -> ()
+}

mlir/test/Transforms/test-legalizer.mlir

 func @remap_input_1_to_1(%arg0: i64) {
   // CHECK-NEXT: "test.valid"{{.*}} : (f64)
   "test.invalid"(%arg0) : (i64) -> ()
 }
 
 // CHECK-LABEL: func @remap_call_1_to_1(%arg0: f64)
 func @remap_call_1_to_1(%arg0: i64) {
   // CHECK-NEXT: call @remap_input_1_to_1(%arg0) : (f64) -> ()
-  // expected-remark@+1 {{op 'std.call' is not legalizable}}
   call @remap_input_1_to_1(%arg0) : (i64) -> ()
   // expected-remark@+1 {{op 'std.return' is not legalizable}}
   return
 }
 
 // CHECK-LABEL: func @remap_input_1_to_N({{.*}}f16, {{.*}}f16)
 func @remap_input_1_to_N(%arg0: f32) -> f32 {
-  // CHECK-NEXT: [[CAST:%.*]] = "test.cast"(%arg0, %arg1) : (f16, f16) -> f32
   // CHECK-NEXT: "test.return"{{.*}} : (f16, f16) -> ()
   "test.return"(%arg0) : (f32) -> ()
 }
 
 // CHECK-LABEL: func @remap_input_1_to_N_remaining_use(%arg0: f16, %arg1: f16)
 func @remap_input_1_to_N_remaining_use(%arg0: f32) {
   // CHECK-NEXT: [[CAST:%.*]] = "test.cast"(%arg0, %arg1) : (f16, f16) -> f32
   // CHECK-NEXT: "work"([[CAST]]) : (f32) -> ()

mlir/test/lib/Dialect/Test/TestOps.td

 def TestTypeConsumerOp : TEST_Op<"type_consumer">,
   Arguments<(ins AnyType)>;
 def TestValidOp : TEST_Op<"valid", [Terminator]>,
   Arguments<(ins Variadic<AnyType>)>;
 
 def TestMergeBlocksOp : TEST_Op<"merge_blocks"> {
   let summary = "merge_blocks operation";
   let description = [{
-    Test op with multiple blocks that are merged with Dialect Conversion"
+    Test op with multiple blocks that are merged with Dialect Conversion
   }];
 
   let regions = (region AnyRegion:$body);
   let results = (outs Variadic<AnyType>:$result);
 }
 
+def TestRemappedValueRegionOp : TEST_Op<"remapped_value_region",
+                                        [SingleBlock]> {
+  let summary = "remapped_value_region operation";
+  let description = [{
+    Test op that remaps values that haven't yet been converted in Dialect
+    Conversion.
+  }];
+
+  let regions = (region SizedRegion<1>:$body);
+  let results = (outs Variadic<AnyType>:$result);
+}
+
 def TestSignatureConversionUndoOp : TEST_Op<"signature_conversion_undo"> {
   let regions = (region AnyRegion);
 }
 
 //===----------------------------------------------------------------------===//
 // Test parser.
 //===----------------------------------------------------------------------===//
 

mlir/test/lib/Dialect/Test/TestPatterns.cpp

matchAndRewrite(Operation *op, ArrayRef<Value> operands,
                   ConversionPatternRewriter &rewriter) const final {
     // Check for a return of F32.
     if (op->getNumOperands() != 1 || !op->getOperand(0).getType().isF32())
       return failure();
 
     // Check if the first operation is a cast operation, if it is we use the
     // results directly.
     auto *defOp = operands[0].getDefiningOp();
-    if (auto packerOp = llvm::dyn_cast_or_null<TestCastOp>(defOp)) {
+    if (auto packerOp =
+            llvm::dyn_cast_or_null<UnrealizedConversionCastOp>(defOp)) {
       rewriter.replaceOpWithNewOp<TestReturnOp>(op, packerOp.getOperands());
       return success();
     }
 
     // Otherwise, fail to match.
     return failure();
   }
 };

 
 namespace {
 struct TestTypeConverter : public TypeConverter {
   using TypeConverter::TypeConverter;
   TestTypeConverter() {
     addConversion(convertType);
     addArgumentMaterialization(materializeCast);
     addSourceMaterialization(materializeCast);
-
-    /// Materialize the cast for one-to-one conversion from i64 to f64.
-    const auto materializeOneToOneCast =
-        [](OpBuilder &builder, IntegerType resultType, ValueRange inputs,
-           Location loc) -> Optional<Value> {
-      if (resultType.getWidth() == 42 && inputs.size() == 1)
-        return builder.create<TestCastOp>(loc, resultType, inputs).getResult();
-      return llvm::None;
-    };
-    addArgumentMaterialization(materializeOneToOneCast);
   }
 
   static LogicalResult convertType(Type t, SmallVectorImpl<Type> &results) {
     // Drop I16 types.
     if (t.isSignlessInteger(16))
       return success();
 
     // Convert I64 to F64.
static LogicalResult convertType(Type t, SmallVectorImpl<Type> &results) {
     results.push_back(t);
     return success();
   }
 
   /// Hook for materializing a conversion. This is necessary because we generate
   /// 1->N type mappings.
   static Optional<Value> materializeCast(OpBuilder &builder, Type resultType,
                                          ValueRange inputs, Location loc) {
-    if (inputs.size() == 1)
-      return inputs[0];
     return builder.create<TestCastOp>(loc, resultType, inputs).getResult();
   }
 };
 
 struct TestLegalizePatternDriver
     : public PassWrapper<TestLegalizePatternDriver, OperationPass<ModuleOp>> {
   StringRef getArgument() const final { return "test-legalize-patterns"; }
   StringRef getDescription() const final {
target.addDynamicallyLegalOp<TestReturnOp>([](TestReturnOp op) {
       // Don't allow F32 operands.
       return llvm::none_of(op.getOperandTypes(),
                            [](Type type) { return type.isF32(); });
     });
     target.addDynamicallyLegalOp<FuncOp>([&](FuncOp op) {
       return converter.isSignatureLegal(op.getType()) &&
              converter.isLegal(&op.getBody());
     });
+    target.addDynamicallyLegalOp<CallOp>(
+        [&](CallOp op) { return converter.isLegal(op); });
 
     // TestCreateUnregisteredOp creates `arith.constant` operation,
     // which was not added to target intentionally to test
     // correct error code from conversion driver.
     target.addDynamicallyLegalOp<ILLegalOpG>([](ILLegalOpG) { return false; });
 
     // Expect the type_producer/type_consumer operations to only operate on f64.
     target.addDynamicallyLegalOp<TestTypeProducerOp>(
legalizerConversionMode(
             clEnumValN(TestLegalizePatternDriver::ConversionMode::Partial,
                        "partial", "Perform a partial conversion")));
 
 //===----------------------------------------------------------------------===//
 // ConversionPatternRewriter::getRemappedValue testing. This method is used
 // to get the remapped value of an original value that was replaced using
 // ConversionPatternRewriter.
 namespace {
+struct TestRemapValueTypeConverter : public TypeConverter {
+  using TypeConverter::TypeConverter;
+
+  TestRemapValueTypeConverter() {
+    addConversion(
+        [](Float32Type type) { return Float64Type::get(type.getContext()); });
+    addConversion([](Type type) { return type; });
+  }
+};
+
 /// Converter that replaces a one-result one-operand OneVResOneVOperandOp1 with
 /// a one-operand two-result OneVResOneVOperandOp1 by replicating its original
 /// operand twice.
 ///
 /// Example:
 ///   %1 = test.one_variadic_out_one_variadic_in1"(%0)
 /// is replaced with:
 ///   %1 = test.one_variadic_out_one_variadic_in1"(%0, %0)
matchAndRewrite(OneVResOneVOperandOp1 op, OpAdaptor adaptor,
     remappedOperands.push_back(rewriter.getRemappedValue(origOp));
 
     rewriter.replaceOpWithNewOp<OneVResOneVOperandOp1>(op, op.getResultTypes(),
                                                        remappedOperands);
     return success();
   }
 };
 
+/// A rewriter pattern that tests that blocks can be merged.
+struct TestRemapValueInRegion
+    : public OpConversionPattern<TestRemappedValueRegionOp> {
+  using OpConversionPattern<TestRemappedValueRegionOp>::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(TestRemappedValueRegionOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const final {
+    Block &block = op.getBody().front();
+    Operation *terminator = block.getTerminator();
+
+    // Merge the block into the parent region.
+    Block *parentBlock = op->getBlock();
+    Block *finalBlock = rewriter.splitBlock(parentBlock, op->getIterator());
+    rewriter.mergeBlocks(&block, parentBlock, ValueRange());
+    rewriter.mergeBlocks(finalBlock, parentBlock, ValueRange());
+
+    // Replace the results of this operation with the remapped terminator
+    // values.
+    SmallVector<Value> terminatorOperands;
+    if (failed(rewriter.getRemappedValues(terminator->getOperands(),
+                                          terminatorOperands)))
+      return failure();
+
+    rewriter.eraseOp(terminator);
+    rewriter.replaceOp(op, terminatorOperands);
+    return success();
+  }
+};
+
 struct TestRemappedValue
     : public mlir::PassWrapper<TestRemappedValue, FunctionPass> {
   StringRef getArgument() const final { return "test-remapped-value"; }
   StringRef getDescription() const final {
     return "Test public remapped value mechanism in ConversionPatternRewriter";
   }
   void runOnFunction() override {
+    TestRemapValueTypeConverter typeConverter;
+
     mlir::RewritePatternSet patterns(&getContext());
     patterns.add<OneVResOneVOperandOp1Converter>(&getContext());
+    patterns.add<TestChangeProducerTypeF32ToF64, TestUpdateConsumerType>(
+        &getContext());
+    patterns.add<TestRemapValueInRegion>(typeConverter, &getContext());
 
     mlir::ConversionTarget target(getContext());
     target.addLegalOp<ModuleOp, FuncOp, TestReturnOp>();
+
+    // Expect the type_producer/type_consumer operations to only operate on f64.
+    target.addDynamicallyLegalOp<TestTypeProducerOp>(
+        [](TestTypeProducerOp op) { return op.getType().isF64(); });
+    target.addDynamicallyLegalOp<TestTypeConsumerOp>([](TestTypeConsumerOp op) {
+      return op.getOperand().getType().isF64();
+    });
+
     // We make OneVResOneVOperandOp1 legal only when it has more that one
     // operand. This will trigger the conversion that will replace one-operand
     // OneVResOneVOperandOp1 with two-operand OneVResOneVOperandOp1.
     target.addDynamicallyLegalOp<OneVResOneVOperandOp1>(
-        [](Operation *op) -> bool {
-          return std::distance(op->operand_begin(), op->operand_end()) > 1;
-        });
+        [](Operation *op) { return op->getNumOperands() > 1; });
 
     if (failed(mlir::applyFullConversion(getFunction(), target,
                                          std::move(patterns)))) {
       signalPassFailure();
     }
   }
 };
 } // end anonymous namespace