[mlir] Remove incorrect folding for SubTensorInsertOp

[mravishankar]

Summary

The SubTensorInsertOp has a requirement that dest type and result
type match. Just folding the tensor.cast operation violates this and
creates verification errors during canonicalization. Also fix other
canonicalization methods that werent inserting casts properly.

[mravishankar]

Add test.

[mravishankar]

Correcting previous incorrect update

[hanchung]

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

2390–2392

[optional] I feel llvm::zip(op->getResults(), newOp->getResults()) looks better to me.

mlir/test/Dialect/Linalg/canonicalize.mlir

782

Should this be dim %arg0, %c1?

792

Let's add a new line at the end of file?

[mravishankar]

Address comments

[nicolasvasilache]

mlir/test/Dialect/Linalg/canonicalize.mlir

782

+1

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

for (Value v : linalgOp.getOutputs()) {
       newOperands.push_back(fold ? tensorCastOp.getOperand() : v);
       newResultTypes.push_back(newOperands.back().getType());
     }
     auto extraOperands = linalgOp.getAssumedNonShapedOperands();
     newOperands.append(extraOperands.begin(), extraOperands.end());
     // Clone op.
     Operation *newOp =
         linalgOp.clone(rewriter, op->getLoc(), newResultTypes, newOperands);
-    rewriter.replaceOp(op, newOp->getResults());
+    SmallVector<Value, 4> replacements;
+    replacements.reserve(newOp->getNumResults());
+    for (auto result : enumerate(newOp->getResults())) {
+      Value newResult = result.value();
+      Value oldResult = op->getResult(result.index());
+      if (newResult.getType() != oldResult.getType()) {
+        replacements.push_back(rewriter.create<tensor::CastOp>(
+            op->getLoc(), oldResult.getType(), newResult));
+      } else {
+        replacements.push_back(newResult);
+      }
+    }
+    rewriter.replaceOp(op, replacements);
 
     return success();
   }
 };
 
 /// Replaces std.dim operations that use the result of a LinalgOp (on tensors)
 /// with std.dim operations that use one of the arguments. For example,
 ///

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

 
 static void replaceWithNewOp(PatternRewriter &rewriter, SubViewOp op,
                              SubViewOp newOp) {
   rewriter.replaceOpWithNewOp<MemRefCastOp>(op, newOp, op.getType());
 }
 
 static void replaceWithNewOp(PatternRewriter &rewriter, SubTensorOp op,
                              SubTensorOp newOp) {
-  rewriter.replaceOpWithNewOp<tensor::CastOp>(op, op.getType(), newOp);
+  Value replacement = newOp.getResult();
+  if (replacement.getType() != op.getType())
+    replacement =
+        rewriter.create<tensor::CastOp>(op.getLoc(), op.getType(), replacement);
+  rewriter.replaceOp(op, replacement);
 }
 
 /// Pattern to rewrite a subview op with constant arguments.
 template <typename OpType>
 class OpWithOffsetSizesAndStridesConstantArgumentFolder final
     : public OpRewritePattern<OpType> {
 public:
   using OpRewritePattern<OpType>::OpRewritePattern;
void mlir::SubTensorInsertOp::build(OpBuilder &b, OperationState &result,
   SmallVector<OpFoldResult> sizeValues = llvm::to_vector<4>(
       llvm::map_range(sizes, [](Value v) -> OpFoldResult { return v; }));
   SmallVector<OpFoldResult> strideValues = llvm::to_vector<4>(
       llvm::map_range(strides, [](Value v) -> OpFoldResult { return v; }));
   build(b, result, source, dest, offsetValues, sizeValues, strideValues);
 }
 
 OpFoldResult SubTensorInsertOp::fold(ArrayRef<Attribute>) {
-  if (getSourceType() == getType() &&
+  if (getSourceType().hasStaticShape() && getType().hasStaticShape() &&
+      getSourceType() == getType() &&
       succeeded(foldIdentityOffsetSizeAndStrideOpInterface(*this, getType())))
     return this->source();
-  if (succeeded(tensor::foldTensorCast(*this)))
-    return this->source();
   return OpFoldResult();
 }
 
 namespace {
 
 /// Pattern to rewrite a subview op with constant arguments.
 class SubTensorInsertOpConstantArgumentFolder final
     : public OpRewritePattern<SubTensorInsertOp> {
public:
   }
 };
 
 /// Fold tensor_casts with subtensor_insert operations.
 struct SubTensorInsertOpCastFolder final
     : public OpRewritePattern<SubTensorInsertOp> {
   using OpRewritePattern<SubTensorInsertOp>::OpRewritePattern;
 
-  LogicalResult matchAndRewrite(SubTensorInsertOp subTensorOp,
+  LogicalResult matchAndRewrite(SubTensorInsertOp subTensorInsertOp,
                                 PatternRewriter &rewriter) const override {
-    if (llvm::any_of(subTensorOp.getOperands(), [](Value operand) {
+    if (llvm::any_of(subTensorInsertOp.getOperands(), [](Value operand) {
           return matchPattern(operand, m_ConstantIndex());
         }))
       return failure();
 
     auto getSourceOfCastOp = [](Value v) -> Optional<Value> {
       auto castOp = v.getDefiningOp<tensor::CastOp>();
       if (!castOp || !canFoldIntoConsumerOp(castOp))
         return llvm::None;
       return castOp.source();
     };
-    Optional<Value> sourceCastSource = getSourceOfCastOp(subTensorOp.source());
-    Optional<Value> destCastSource = getSourceOfCastOp(subTensorOp.dest());
-    if (!sourceCastSource && !destCastSource &&
-        subTensorOp.dest().getType() == subTensorOp.getResult().getType())
-      return failure();
-
-    auto newOp = rewriter.create<SubTensorInsertOp>(
-        subTensorOp.getLoc(),
-        (sourceCastSource ? *sourceCastSource : subTensorOp.source()),
-        (destCastSource ? *destCastSource : subTensorOp.dest()),
-        subTensorOp.getMixedOffsets(), subTensorOp.getMixedSizes(),
-        subTensorOp.getMixedStrides());
-
-    rewriter.replaceOpWithNewOp<tensor::CastOp>(subTensorOp,
-                                                subTensorOp.getType(), newOp);
+    Optional<Value> sourceCastSource =
+        getSourceOfCastOp(subTensorInsertOp.source());
+    Optional<Value> destCastSource =
+        getSourceOfCastOp(subTensorInsertOp.dest());
+    if (!sourceCastSource && !destCastSource)
+      return failure();
+
+    Value replacement = rewriter.create<SubTensorInsertOp>(
+        subTensorInsertOp.getLoc(),
+        (sourceCastSource ? *sourceCastSource : subTensorInsertOp.source()),
+        (destCastSource ? *destCastSource : subTensorInsertOp.dest()),
+        subTensorInsertOp.getMixedOffsets(), subTensorInsertOp.getMixedSizes(),
+        subTensorInsertOp.getMixedStrides());
+
+    if (replacement.getType() != subTensorInsertOp.getType()) {
+      replacement = rewriter.create<tensor::CastOp>(
+          subTensorInsertOp.getLoc(), subTensorInsertOp.getType(), replacement);
+    }
+    rewriter.replaceOp(subTensorInsertOp, replacement);
     return success();
   }
 };
 } // namespace
 
 void SubTensorInsertOp::getCanonicalizationPatterns(
     OwningRewritePatternList &results, MLIRContext *context) {
   results.insert<SubTensorInsertOpConstantArgumentFolder,

mlir/test/Dialect/Linalg/canonicalize.mlir

 //      CHECK: #[[MAP:.+]] = affine_map<()[s0] -> (s0 * 28)>
 //      CHECK: func @dim_reshape_collapse
 // CHECK-SAME:   %[[ARG0:[a-zA-Z0-9_]+]]: tensor<2x3x5x4x?x7xf32>
 //  CHECK-DAG:   %[[C4:.+]] = constant 4 : index
 //  CHECK-DAG:   %[[C5:.+]] = constant 5 : index
 //      CHECK:   %[[D0:.+]] = dim %[[ARG0]], %[[C4]]
 //      CHECK:   %[[D1:.+]] = affine.apply #[[MAP]]()[%[[D0]]]
 //      CHECK:   return %[[C5]], %[[D1]]
+
+// -----
+
+func @propogate_casts(%arg0 : tensor<?x?xf32>, %arg1 : f32, %arg2 : index,
+    %arg3 : index) -> tensor<?x?xf32> {
+  %c0 = constant 0 : index
+  %c1 = constant 1 : index
+  %c21 = constant 21 : index
+  %c42 = constant 42 : index
+  %0 = linalg.init_tensor [%c21, %c42] : tensor<?x?xf32>
+  %1 = linalg.fill(%0, %arg1) : tensor<?x?xf32>, f32 -> tensor<?x?xf32>
+  %2 = dim %arg0, %c0 : tensor<?x?xf32>
+  %3 = dim %arg0, %c0 : tensor<?x?xf32>
+  %4 = subtensor_insert %arg0 into %1[%arg2, %arg3] [%2, %3] [1, 1] : tensor<?x?xf32> into tensor<?x?xf32>
+  return %4 : tensor<?x?xf32>
+}
+// CHECK-LABEL: func @propogate_casts
+//       CHECK:   %[[INIT:.+]] = linalg.init_tensor [21, 42]
+//       CHECK:   %[[FILL:.+]] = linalg.fill(%[[INIT]], %{{.+}})
+//       CHECK:   %[[INSERTED:.+]] = subtensor_insert %{{.+}} into %[[FILL]]
+//       CHECK:   %[[RESULT:.+]] = tensor.cast %[[INSERTED]]
+//       CHECK:   return %[[RESULT]]
 No newline at end of file