[mlir][Linalg] Define a linalg.init_tensor operation.

[mravishankar]

Summary

This operation is used to materialize a tensor of a particular
shape. The shape could be specified as a mix of static and dynamic
values.

The use of this operation is to be an init tensor for Linalg
structured operation on tensors where the bounds of the computation
depends on the shape of the output of the linalg operation. The result
of this operation will be used as the init tensor of such Linalg
operations. To note,

1. The values in the tensor materialized is not used. Any operation to which this is an init tensor is expected to overwrite the entire tensor.
2. The tensor is materialized only for the shape of the output and to make the loop bounds depend only on operands of the structured operation.

Based on (1) and (2) it is assumed that these operations eventually go
away since they are only used in dim operations that can be
canonicalized to make this operation dead. Such canonicalization are
added here too.

Depends On D93180

[silvas]

mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td

44

you can take the static sizes from the result type. (you probably just want to imitate how std.alloc, std.dynamic_tensor_from_elements, etc. model this; they have Variadic<Index> with one entry for each ? in the result type)

[mravishankar]

mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td

44

I was going with the approach the subview ops, etc. take. THe advantage of this explicitly is that it is easier to retrieve the dim as a static or dynamic value. The approach you descibe would also work, but then I need to look at the type instead. If you have strong opinions on that, I can change it.

[pifon2a]

mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td

98

$sizes?

105

please, also add a builder for the mixed static-dynamic case:

OpBuilderDAG<(ins "ValueRange":$sizes, "ArrayRef<int64_t>":$staticSizes, "Type":$elementType),
  [{
    build($_builder, $_state,
          InitTensorOp::inferResultType(staticSizes, elementType),
          sizes, $_builder.getI64ArrayAttr(staticSizes));
  }]>,

mlir/include/mlir/Interfaces/ViewLikeInterface.h

38

nit: do you need "(1)" and "(2)" in this comment?

[silvas]

mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td

44

Ok, I haven't thought through the tradeoffs here fully, and wasn't aware of the alternative precedent for subview (I see now that this patch reuses a lot of the parse/print infra from those ops). I'm fine with this for now.

[silvas]

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

1479

Can this be in a separate patch? I don't see the relationship to this patch.

mlir/test/Dialect/Linalg/canonicalize.mlir

357

This test case seems to be testing two separate patterns. Can you do it with two test cases for clarity?

[nicolasvasilache]

+1 it's fine for now and unlocks a bigger piece of the puzzle.
We can refactor and cleanup as followups: subview crams together 3 variadic static / dynamic + canonicalizations.
There is a nicer future where we can define a generic pair of variadic operand + attribute + canonicalizations for all such things.

That's 2021's problem though.

[mravishankar]

Address comments

[silvas]

mlir/test/Dialect/Linalg/canonicalize.mlir

357

I think you don't have a specific test for linalg.init_tensor [4, 5, %c6] : tensor<4x5x?xf32> -> linalg.init_tensor [4, 5, 6] anymore. That one should be its own test to avoid confusing it with init_tensor_static_dim (which is related, but not testing the same thing)

361

dim 2 is %c6 so it is not static (maybe internally to canonicalization it becomes so...). Best to use dim %0, %c0 instead so it gets the static "4" value.

[mravishankar]

Adding mized static-dynamic builder.

[mravishankar]

mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td

98

shape is consistent with what is gotten from ShapedType.

mlir/include/mlir/Interfaces/ViewLikeInterface.h

38

Just copied and pasted it from somewhere else. Ill leave it as is.

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

1479

I need it to bridge the gap from canonicalizer for canonicalizing a linalg.init_tensor operation which introduces a tensor_cast. SO a constant_index -> linalg.init_tensor -> dim to a constant requires some folding of the intermediate tensor_cast with dim.

mlir/test/Dialect/Linalg/canonicalize.mlir

357

Good point. I will add that one.

361

That might not even get to the linalg.init_tensor cause it is dim %0, %c0 : tensor<4x5x?xf32> and that is canonicalized by itself.

[rriddle]

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

1479

Still looks like it can be separated from this patch.

[mravishankar]

Adding another test.

[mravishankar]

Splitting out tensor_cast -> dim canonicalization.

[mravishankar]

I see what you mean now.

[hanchung]

mlir/test/Dialect/Linalg/roundtrip.mlir

715–738

these should be put in invalid.mlir

mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td

class Linalg_Op<string mnemonic, list<OpTrait> traits = []> :
   //   * ParseResult parse${C++ class of Op}(OpAsmParser &parser,
   //                                         OperationState &result)
   // functions.
   let printer = [{ return ::print(p, *this); }];
   let verifier = [{ return ::verify(*this); }];
   let parser = [{ return ::parse$cppClass(parser, result); }];
 }
 
+def Linalg_InitTensorOp : Linalg_Op<"init_tensor", [NoSideEffect]> {
+  let summary = "operation to define a tensor of particular value";
+
+  let description = [{
+    `linalg.init_tensor` is an operation that materializes a tensor of
+    a given shape. The shape could be dynamic or static.
+  }];
+
+  let arguments =
+    (ins Variadic<Index>:$sizes, I64ArrayAttr:$static_sizes);
+
+  let results = (outs AnyTensor:$result);
+
+  let verifier = [{ return ::verify(*this); }];
+
+  let extraClassDeclaration = [{
+    static StringRef getStaticSizesAttrName() {
+      return "static_sizes";
+    }
+
+    RankedTensorType getType() {
+      return getResult().getType().cast<RankedTensorType>(); }
+
+    // Infer the shape of the result tensor given the static shapes
+    // and element type of the result tensor.
+    static Type inferResultType(ArrayRef<int64_t> staticSizes, Type elementType);
+
+    // Return true if the size of the tensor is dynamic at `idx`
+    bool isDynamicSize(unsigned idx) {
+      APInt v = *(static_sizes().getAsValueRange<IntegerAttr>().begin() + idx);
+      return ShapedType::isDynamic(v.getSExtValue());
+    }
+
+    // Assert that the size of the result tensor is static at `idx`
+    // and return the shape.
+    int64_t getStaticSize(unsigned idx) {
+      assert(!isDynamicSize(idx) && "expected static size");
+      APInt v = *(static_sizes().
+          template getAsValueRange<IntegerAttr>().begin() + idx);
+        return v.getSExtValue();
+    }
+
+    // Return the argument position that contains the dynamic size of
+    // the tensor at dimension `idx`. Asserts that the shape is
+    // dynamic at that `idx`.
+    unsigned getIndexOfDynamicSize(unsigned idx) {
+      assert(isDynamicSize(idx) && "expected dynamic size");
+      return std::count_if(
+          static_sizes().getValue().begin(),
+          static_sizes().getValue().begin() + idx,
+          [&](Attribute attr) {
+            return ShapedType::isDynamic(attr.cast<IntegerAttr>().getInt());
+          });
+    }
+
+    // Return the Value of the dynamic size of the tensor at dimension
+    // `idx`. Asserts that the shape is dynamic at that `idx.
+    Value getDynamicSize(unsigned idx) {
+      return getOperand(getIndexOfDynamicSize(idx));
+    }
+  }];
+
+  let builders = [
+    OpBuilderDAG<(ins "ValueRange":$shape,
+                  "ArrayRef<int64_t>":$staticShape, "Type":$elementType),
+    [{
+      build($_builder, $_state,
+            InitTensorOp::inferResultType(staticShape, elementType),
+            shape, $_builder.getI64ArrayAttr(staticShape));
+    }]>,
+    OpBuilderDAG<(ins "ValueRange":$shape, "Type":$elementType),
+    [{
+      SmallVector<int64_t, 4> staticShape(
+        shape.size(), ShapedType::kDynamicSize);
+      build($_builder, $_state, shape, staticShape, elementType);
+    }]>,
+    OpBuilderDAG<(ins "ArrayRef<int64_t>":$staticShape, "Type":$elementType),
+    [{
+      build($_builder, $_state, ValueRange{}, staticShape, elementType);
+    }]>
+  ];
+
+  let hasCanonicalizer = 1;
+}
+
 def Linalg_RangeOp :
     Linalg_Op<"range", [NoSideEffect]>,
     Arguments<(ins Index:$min, Index:$max, Index:$step)>,
     Results<(outs Range)> {
   let summary = "Create a `range` type value, used to create `view`s";
   let description = [{
     The `linalg.range` op creates a `!linalg.range` from 3 values of type
     `index` that represent the min, max and step values of the `range`. This

mlir/include/mlir/Interfaces/ViewLikeInterface.h

 class OffsetSizeAndStrideOpInterface;
 LogicalResult verify(OffsetSizeAndStrideOpInterface op);
 } // namespace mlir
 
 /// Include the generated interface declarations.
 #include "mlir/Interfaces/ViewLikeInterface.h.inc"
 
 namespace mlir {
+/// Print a list with either (1) the static integer value in `arrayAttr` if
+/// `isDynamic` evaluates to false or (2) the next value otherwise.
+/// This allows idiomatic printing of mixed value and integer attributes in a
+/// list. E.g. `[%arg0, 7, 42, %arg42]`.
+void printListOfOperandsOrIntegers(OpAsmPrinter &p, ValueRange values,
+                                   ArrayAttr arrayAttr,
+                                   llvm::function_ref<bool(int64_t)> isDynamic);
+
 /// Print part of an op of the form:
 /// ```
 ///   <optional-offset-prefix>`[` offset-list `]`
 ///   <optional-size-prefix>`[` size-list `]`
 ///   <optional-stride-prefix>[` stride-list `]`
 /// ```
 void printOffsetsSizesAndStrides(
     OpAsmPrinter &p, OffsetSizeAndStrideOpInterface op,
     StringRef offsetPrefix = "", StringRef sizePrefix = " ",
     StringRef stridePrefix = " ",
     ArrayRef<StringRef> elidedAttrs =
         OffsetSizeAndStrideOpInterface::getSpecialAttrNames());
 
+/// Parse a mixed list with either (1) static integer values or (2) SSA values.
+/// Fill `result` with the integer ArrayAttr named `attrName` where `dynVal`
+/// encode the position of SSA values. Add the parsed SSA values to `ssa`
+/// in-order.
+//
+/// E.g. after parsing "[%arg0, 7, 42, %arg42]":
+///   1. `result` is filled with the i64 ArrayAttr "[`dynVal`, 7, 42, `dynVal`]"
+///   2. `ssa` is filled with "[%arg0, %arg1]".
+ParseResult
+parseListOfOperandsOrIntegers(OpAsmParser &parser, OperationState &result,
+                              StringRef attrName, int64_t dynVal,
+                              SmallVectorImpl<OpAsmParser::OperandType> &ssa);
+
 /// Parse trailing part of an op of the form:
 /// ```
 ///   <optional-offset-prefix>`[` offset-list `]`
 ///   <optional-size-prefix>`[` size-list `]`
 ///   <optional-stride-prefix>[` stride-list `]`
 /// ```
 /// Each entry in the offset, size and stride list either resolves to an integer
 /// constant or an operand of index type.
ParseResult parseOffsetsSizesAndStrides(
     OpAsmParser &parser, OperationState &result, ArrayRef<int> segmentSizes,
     llvm::function_ref<ParseResult(OpAsmParser &)> parseOptionalOffsetPrefix =
         nullptr,
     llvm::function_ref<ParseResult(OpAsmParser &)> parseOptionalSizePrefix =
         nullptr,
     llvm::function_ref<ParseResult(OpAsmParser &)> parseOptionalStridePrefix =
         nullptr);
 
+/// Verify that a the `values` has as many elements as the number of entries in
+/// `attr` for which `isDynamic` evaluates to true.
+LogicalResult verifyListOfOperandsOrIntegers(
+    Operation *op, StringRef name, unsigned expectedNumElements, ArrayAttr attr,
+    ValueRange values, llvm::function_ref<bool(int64_t)> isDynamic);
+
 } // namespace mlir
 
 #endif // MLIR_INTERFACES_VIEWLIKEINTERFACE_H_

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

static LogicalResult verifyGenericOp(GenericOpType op) {
   return success();
 }
 
 static LogicalResult verify(GenericOp op) { return verifyGenericOp(op); }
 
 static LogicalResult verify(IndexedGenericOp op) { return verifyGenericOp(op); }
 
 //===----------------------------------------------------------------------===//
+// InitTensorOp
+//===----------------------------------------------------------------------===//
+
+static ParseResult parseInitTensorOp(OpAsmParser &parser,
+                                     OperationState &result) {
+  OpAsmParser::OperandType srcInfo;
+  Type dstType;
+  SmallVector<OpAsmParser::OperandType, 2> sizeInfo;
+  IndexType indexType = parser.getBuilder().getIndexType();
+  if (failed(parseListOfOperandsOrIntegers(
+          parser, result, InitTensorOp::getStaticSizesAttrName(),
+          ShapedType::kDynamicSize, sizeInfo)) ||
+      failed(parser.parseOptionalAttrDict(result.attributes)) ||
+      failed(parser.parseColonType(dstType)) ||
+      failed(parser.resolveOperands(sizeInfo, indexType, result.operands)))
+    return failure();
+  return parser.addTypeToList(dstType, result.types);
+}
+
+static void print(OpAsmPrinter &p, InitTensorOp op) {
+  p << op.getOperation()->getName() << ' ';
+  printListOfOperandsOrIntegers(p, op.sizes(), op.static_sizes(),
+                                ShapedType::isDynamic);
+  p.printOptionalAttrDict(op.getAttrs(),
+                          InitTensorOp::getStaticSizesAttrName());
+  p << " : " << op.getType();
+}
+
+static LogicalResult verify(InitTensorOp op) {
+  RankedTensorType resultType = op.getType();
+  SmallVector<int64_t, 4> staticSizes = llvm::to_vector<4>(llvm::map_range(
+      op.static_sizes().cast<ArrayAttr>(),
+      [](Attribute a) -> int64_t { return a.cast<IntegerAttr>().getInt(); }));
+
+  if (failed(verifyListOfOperandsOrIntegers(op, "sizes", resultType.getRank(),
+                                            op.static_sizes(), op.sizes(),
+                                            ShapedType::isDynamic)))
+    return failure();
+
+  Type expectedType =
+      InitTensorOp::inferResultType(staticSizes, resultType.getElementType());
+  if (resultType != expectedType) {
+    return op.emitError("specified type ")
+           << resultType << " does not match the inferred type "
+           << expectedType;
+  }
+  return success();
+}
+
+Type InitTensorOp::inferResultType(ArrayRef<int64_t> staticSizes,
+                                   Type elementType) {
+  return RankedTensorType::get(staticSizes, elementType);
+}
+
+namespace {
+/// Change the type of the result of a `linalg.init_tensor` by making the result
+/// type statically sized along dimension that in the original operation where
+/// defined as dynamic, but the size was defined using a `constant` op. For
+/// example
+///
+///  %c5 = constant 5: index
+///  %0 = linalg.init_tensor [%arg0, %c5] : tensor<?x?xf32>
+///
+///  to
+///
+///  %0 = linalg.init_tensor [%arg0, 5] : tensor<?x5xf32>
+struct ReplaceStaticShapeDims : OpRewritePattern<InitTensorOp> {
+  using OpRewritePattern<InitTensorOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(InitTensorOp op,
+                                PatternRewriter &rewriter) const override {
+    SmallVector<Value, 4> dynamicSizes;
+    SmallVector<int64_t, 4> staticSizes;
+    for (unsigned i = 0, e = op.getType().getRank(); i != e; ++i) {
+      // If the size is already static, nothing to do.
+      if (!op.isDynamicSize(i)) {
+        staticSizes.push_back(op.getStaticSize(i));
+        continue;
+      }
+
+      // If the size is dynamic but defined using a `constant` op, get the
+      // constant value to find the static size to use.
+      unsigned operandNum = op.getIndexOfDynamicSize(i);
+      Value sizeOperand = op.getOperand(operandNum);
+      if (auto constantIndexOp = sizeOperand.getDefiningOp<ConstantIndexOp>()) {
+        staticSizes.push_back(constantIndexOp.getValue());
+        continue;
+      }
+
+      // Fallback case. Keep the size dynamic.
+      dynamicSizes.push_back(sizeOperand);
+      staticSizes.push_back(ShapedType::kDynamicSize);
+    }
+    RankedTensorType newType =
+        RankedTensorType::get(staticSizes, op.getType().getElementType());
+    if (newType == op.getType())
+      return failure();
+    auto newOp =
+        rewriter.create<InitTensorOp>(op.getLoc(), newType, dynamicSizes,
+                                      rewriter.getI64ArrayAttr(staticSizes));
+    rewriter.replaceOpWithNewOp<TensorCastOp>(op, op.getType(), newOp);
+    return success();
+  }
+};
+
+/// Canonicalize a `linalg.init_tensor` -> `dim` pattern by replacing the `dim`
+/// with
+/// - A constant value if the size is static along the dimension.
+/// - The dynamic value that defines the size of the result of
+///   `linalg.init_tensor` op.
+struct ReplaceDimOfInitTensorOp : public OpRewritePattern<DimOp> {
+  using OpRewritePattern<DimOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(DimOp dimOp,
+                                PatternRewriter &rewriter) const override {
+    auto initTensorOp = dimOp.memrefOrTensor().getDefiningOp<InitTensorOp>();
+    if (!initTensorOp)
+      return failure();
+    auto dimIndex = dimOp.index().getDefiningOp<ConstantIndexOp>();
+    if (!dimIndex)
+      return failure();
+    int64_t index = dimIndex.getValue();
+    if (!initTensorOp.isDynamicSize(index)) {
+      rewriter.replaceOpWithNewOp<ConstantIndexOp>(
+          dimOp, initTensorOp.getStaticSize(index));
+    } else {
+      rewriter.replaceOp(dimOp, initTensorOp.getDynamicSize(index));
+    }
+    return success();
+  }
+};
+} // namespace
+
+void InitTensorOp::getCanonicalizationPatterns(
+    OwningRewritePatternList &results, MLIRContext *context) {
+  results.insert<ReplaceDimOfInitTensorOp, ReplaceStaticShapeDims>(context);
+}
+
+//===----------------------------------------------------------------------===//
 // ReshapeOp
 //===----------------------------------------------------------------------===//
 
 /// Collapse reassociation maps that are used in pair of reshape ops where one
 /// is a producer and other is the consumer. Only valid to use this method when
 /// both the producer and consumer are collapsing dimensions or both are
 /// expanding dimensions.
 ///

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

LogicalResult matchAndRewrite(DimOp dim,
     // Place the load directly after the reshape to ensure that the shape memref
     // was not mutated.
     rewriter.setInsertionPointAfter(reshape);
     rewriter.replaceOpWithNewOp<LoadOp>(dim, reshape.shape(),
                                         llvm::makeArrayRef({dim.index()}));
     return success();
   }
 };
+
+/// Fold dim of a dim of a cast into the the dim of the source of the tensor
+/// cast.
+template <typename CastOpTy>
+struct DimOfCastOp : public OpRewritePattern<DimOp> {
+  using OpRewritePattern<DimOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(DimOp dimOp,
+                                PatternRewriter &rewriter) const override {
+    auto castOp = dimOp.memrefOrTensor().getDefiningOp<CastOpTy>();
+    if (!castOp)
+      return failure();
+    Value newSource = castOp.getOperand();
+    rewriter.replaceOpWithNewOp<DimOp>(dimOp, newSource, dimOp.index());
+    return success();
+  }
+};
+
 } // end anonymous namespace.
 
 void DimOp::getCanonicalizationPatterns(OwningRewritePatternList &results,
                                         MLIRContext *context) {
-  results.insert<DimOfMemRefReshape>(context);
+  results.insert<DimOfMemRefReshape, DimOfCastOp<TensorCastOp>>(context);
 }
 
 // ---------------------------------------------------------------------------
 // DmaStartOp
 // ---------------------------------------------------------------------------
 
 void DmaStartOp::build(OpBuilder &builder, OperationState &result,
                        Value srcMemRef, ValueRange srcIndices, Value destMemRef,

mlir/lib/Interfaces/ViewLikeInterface.cpp

 
 //===----------------------------------------------------------------------===//
 // ViewLike Interfaces
 //===----------------------------------------------------------------------===//
 
 /// Include the definitions of the loop-like interfaces.
 #include "mlir/Interfaces/ViewLikeInterface.cpp.inc"
 
-static LogicalResult verifyOpWithOffsetSizesAndStridesPart(
-    OffsetSizeAndStrideOpInterface op, StringRef name,
-    unsigned expectedNumElements, StringRef attrName, ArrayAttr attr,
-    llvm::function_ref<bool(int64_t)> isDynamic, ValueRange values) {
+LogicalResult mlir::verifyListOfOperandsOrIntegers(
+    Operation *op, StringRef name, unsigned expectedNumElements, ArrayAttr attr,
+    ValueRange values, llvm::function_ref<bool(int64_t)> isDynamic) {
   /// Check static and dynamic offsets/sizes/strides breakdown.
   if (attr.size() != expectedNumElements)
-    return op.emitError("expected ")
+    return op->emitError("expected ")
            << expectedNumElements << " " << name << " values";
   unsigned expectedNumDynamicEntries =
       llvm::count_if(attr.getValue(), [&](Attribute attr) {
         return isDynamic(attr.cast<IntegerAttr>().getInt());
       });
   if (values.size() != expectedNumDynamicEntries)
-    return op.emitError("expected ")
+    return op->emitError("expected ")
            << expectedNumDynamicEntries << " dynamic " << name << " values";
   return success();
 }
 
 LogicalResult mlir::verify(OffsetSizeAndStrideOpInterface op) {
   std::array<unsigned, 3> ranks = op.getArrayAttrRanks();
-  if (failed(verifyOpWithOffsetSizesAndStridesPart(
-          op, "offset", ranks[0],
-          OffsetSizeAndStrideOpInterface::getStaticOffsetsAttrName(),
-          op.static_offsets(), ShapedType::isDynamicStrideOrOffset,
-          op.offsets())))
-    return failure();
-  if (failed(verifyOpWithOffsetSizesAndStridesPart(
-          op, "size", ranks[1],
-          OffsetSizeAndStrideOpInterface::getStaticSizesAttrName(),
-          op.static_sizes(), ShapedType::isDynamic, op.sizes())))
-    return failure();
-  if (failed(verifyOpWithOffsetSizesAndStridesPart(
-          op, "stride", ranks[2],
-          OffsetSizeAndStrideOpInterface::getStaticStridesAttrName(),
-          op.static_strides(), ShapedType::isDynamicStrideOrOffset,
-          op.strides())))
+  if (failed(verifyListOfOperandsOrIntegers(
+          op, "offset", ranks[0], op.static_offsets(), op.offsets(),
+          ShapedType::isDynamicStrideOrOffset)))
+    return failure();
+  if (failed(verifyListOfOperandsOrIntegers(op, "size", ranks[1],
+                                            op.static_sizes(), op.sizes(),
+                                            ShapedType::isDynamic)))
+    return failure();
+  if (failed(verifyListOfOperandsOrIntegers(
+          op, "stride", ranks[2], op.static_strides(), op.strides(),
+          ShapedType::isDynamicStrideOrOffset)))
     return failure();
   return success();
 }
 
-/// Print a list with either (1) the static integer value in `arrayAttr` if
-/// `isDynamic` evaluates to false or (2) the next value otherwise.
-/// This allows idiomatic printing of mixed value and integer attributes in a
-/// list. E.g. `[%arg0, 7, 42, %arg42]`.
-static void
-printListOfOperandsOrIntegers(OpAsmPrinter &p, ValueRange values,
-                              ArrayAttr arrayAttr,
-                              llvm::function_ref<bool(int64_t)> isDynamic) {
+void mlir::printListOfOperandsOrIntegers(
+    OpAsmPrinter &p, ValueRange values, ArrayAttr arrayAttr,
   p << '[';
   unsigned idx = 0;
   llvm::interleaveComma(arrayAttr, p, [&](Attribute a) {
     int64_t val = a.cast<IntegerAttr>().getInt();
     if (isDynamic(val))
       p << values[idx++];
     else
       p << val;
void mlir::printOffsetsSizesAndStrides(OpAsmPrinter &p,
   printListOfOperandsOrIntegers(p, op.sizes(), op.static_sizes(),
                                 ShapedType::isDynamic);
   p << stridePrefix;
   printListOfOperandsOrIntegers(p, op.strides(), op.static_strides(),
                                 ShapedType::isDynamicStrideOrOffset);
   p.printOptionalAttrDict(op.getAttrs(), elidedAttrs);
 }
 
-/// Parse a mixed list with either (1) static integer values or (2) SSA values.
-/// Fill `result` with the integer ArrayAttr named `attrName` where `dynVal`
-/// encode the position of SSA values. Add the parsed SSA values to `ssa`
-/// in-order.
-//
-/// E.g. after parsing "[%arg0, 7, 42, %arg42]":
-///   1. `result` is filled with the i64 ArrayAttr "[`dynVal`, 7, 42, `dynVal`]"
-///   2. `ssa` is filled with "[%arg0, %arg1]".
-static ParseResult
-parseListOfOperandsOrIntegers(OpAsmParser &parser, OperationState &result,
-                              StringRef attrName, int64_t dynVal,
-                              SmallVectorImpl<OpAsmParser::OperandType> &ssa) {
+ParseResult mlir::parseListOfOperandsOrIntegers(
+    OpAsmParser &parser, OperationState &result, StringRef attrName,
+    int64_t dynVal, SmallVectorImpl<OpAsmParser::OperandType> &ssa) {
   if (failed(parser.parseLSquare()))
     return failure();
   // 0-D.
   if (succeeded(parser.parseOptionalRSquare())) {
     result.addAttribute(attrName, parser.getBuilder().getArrayAttr({}));
     return success();
   }
 

mlir/test/Dialect/Linalg/canonicalize.mlir

func @linalg_effects(%a : tensor<?x?xf32>, %b : memref<?x?xf32>, %c : tensor<?x?xf32>) {
   %t = linalg.matmul ins(%a, %b : tensor<?x?xf32>, memref<?x?xf32>)
                     init(%c : tensor<?x?xf32>) -> tensor<?x?xf32>
 
   // CHECK-NOT:   %{{.*}} = linalg.matmul
   linalg.matmul ins(%a, %c : tensor<?x?xf32>, tensor<?x?xf32>)
                outs(%b : memref<?x?xf32>)
   return
 }
+// -----
+
+func @init_tensor_canonicalize() -> (tensor<4x5x?xf32>) {
+  %c6 = constant 6 : index
+  %0 = linalg.init_tensor [4, 5, %c6] : tensor<4x5x?xf32>
+  return %0 : tensor<4x5x?xf32>
+}
+// CHECK: func @init_tensor_canonicalize
+// CHECK:   %[[T0:.+]] = linalg.init_tensor [4, 5, 6] : tensor<4x5x6xf32>
+// CHECK:   %[[T1:.+]] = tensor_cast %[[T0]] : tensor<4x5x6xf32> to tensor<4x5x?xf32>
+// CHECK:   return %[[T1]]
+
+// -----
+
+func @init_tensor_static_dim() -> (index, index) {
+  %c0 = constant 0 : index
+  %c2 = constant 2 : index
+  %c6 = constant 6 : index
+  %0 = linalg.init_tensor [4, 5, %c6] : tensor<4x5x?xf32>
+  %1 = dim %0, %c2 : tensor<4x5x?xf32>
+  %2 = dim %0, %c0 : tensor<4x5x?xf32>
+  return %1, %2 : index, index
+}
+//      CHECK: func @init_tensor_static_dim
+//  CHECK-DAG:   %[[C4:.+]] = constant 4 : index
+//  CHECK-DAG:   %[[C6:.+]] = constant 6 : index
+//      CHECK:   return %[[C6]], %[[C4]]
+
+// -----
+
+func @init_tensor_dynamic_dim(%arg0 : index) -> (index) {
+  %c2 = constant 2 : index
+  %0 = linalg.init_tensor [4, 5, %arg0] : tensor<4x5x?xf32>
+  %1 = dim %0, %c2 : tensor<4x5x?xf32>
+  return %1 : index
+}
+//      CHECK: func @init_tensor_dynamic_dim
+// CHECK-SAME:   %[[ARG0:[a-zA-Z0-9_]+]]: index
+//      CHECK:   return %[[ARG0]]

mlir/test/Dialect/Linalg/roundtrip.mlir

-// RUN: mlir-opt -split-input-file %s | FileCheck %s
-// | mlir-opt | FileCheck %s
+// RUN: mlir-opt -split-input-file -verify-diagnostics %s | FileCheck %s
 
 // TODO: Re-enable LLVM lowering test after IndexedGenericOp is lowered.
 //
 // Test that we can lower all the way to LLVM without crashing, don't check results here.
 // DISABLED: mlir-opt %s --convert-linalg-to-llvm -o=/dev/null 2>&1
 
 func @range(%arg0: index, %arg1: index, %arg2: index) {
   %0 = linalg.range %arg0:%arg1:%arg2 : !linalg.range

 {
   %0 = linalg.reshape %arg0 [] : memref<1x1xf32> into memref<f32>
   %1 = linalg.reshape %0 [] : memref<f32> into memref<1x1xf32>
   return %0, %1 : memref<f32>, memref<1x1xf32>
 }
 // CHECK-LABEL: func @memref_reshape_zero_dim
 //       CHECK:   linalg.reshape %{{.*}} [] : memref<1x1xf32> into memref<f32>
 //       CHECK:   linalg.reshape %{{.*}} [] : memref<f32> into memref<1x1xf32>
+
+// -----
+
+func @init_tensor(%arg0 : index, %arg1 : index)
+{
+  %0 = linalg.init_tensor [3, 42] : tensor<3x42xf32>
+  %1 = linalg.init_tensor [4, %arg0, %arg1, 5] : tensor<4x?x?x5xf32>
+  return
+}
+// CHECK-LABEL: func @init_tensor
+//       CHECK:   linalg.init_tensor [3, 42] : tensor<3x42xf32>
+//       CHECK:   linalg.init_tensor [4, %{{.*}}, %{{.*}}, 5] : tensor<4x?x?x5xf32>
+
+// -----
+
+func @init_tensor_err(%arg0 : index, %arg1 : index)
+{
+  // expected-error @+1 {{specified type 'tensor<4x?x?x5xf32>' does not match the inferred type 'tensor<4x5x?x?xf32>'}}
+  %1 = linalg.init_tensor [4, 5, %arg0, %arg1] : tensor<4x?x?x5xf32>
+  return
+}
+
+// -----
+
+func @init_tensor_err(%arg0 : index)
+{
+  // expected-error @+1 {{expected 4 sizes values}}
+  %1 = linalg.init_tensor [4, 5, %arg0] : tensor<4x?x?x5xf32>
+  return
+}
+
+// -----
+
+func @init_tensor_err(%arg0 : index)
+{
+  // expected-error @+1 {{expected 2 dynamic sizes values}}
+  %1 = "linalg.init_tensor"(%arg0) {static_sizes = [4, -1, -1, 5]} : (index) -> tensor<4x?x?x5xf32>
+  return
+}