[mlir] [linalg] Fix bufferize error in tensor.parallel_insert_slice op

tensor.parallel op has implicit inplace behavior. In the "copy-before-write"
bufferize mode, the resolveConflict function will generate bufferize.copy,
making the result incorrect. This patch fixes this issue.

Author: cxy-1993

mlir/lib/Dialect/Tensor/Transforms/BufferizableOpInterfaceImpl.cpp

@@ -997,6 +997,11 @@ struct ParallelInsertSliceOpInterface
     rewriter.eraseOp(op);
     return success();
   }
+
+  LogicalResult resolveConflicts(Operation *op, RewriterBase &rewriter,
+                                 const AnalysisState &state) const {
+    return success();
+  }
 };
 
 /// Bufferization of tensor.splat. Bufferizes to a new allocation that is filled

mlir/test/Dialect/Tensor/bufferize.mlir

@@ -626,3 +626,23 @@ func.func @tensor.splat_dynamic(%f: f32, %m: index, %n: index) -> tensor<?x3x?xf
   return %0 : tensor<?x3x?xf32>
 }
 
+// -----
+
+// CHECK-LABEL: func.func @parallel_insert_slice_copy_before_write
+func.func @parallel_insert_slice_copy_before_write(%in: tensor<4xf32>, %out: tensor<4xf32>) {
+  %c1 = arith.constant 1 : index
+  %num_threads = arith.constant 4 : index
+
+  // CHECK: scf.forall {{.*}} {
+  %result = scf.forall (%thread_idx) in (%num_threads) shared_outs (%o = %out) -> tensor<4xf32> {
+      %1 = tensor.extract_slice %in[%thread_idx][1][1] : tensor<4xf32> to tensor<1xf32>
+      scf.forall.in_parallel {
+        // CHECK: memref.subview %{{.*}}[%{{.*}}] [1] [1] : memref<4xf32> to memref<1xf32, strided<[1], offset: ?>>
+        // CHECK: memref.subview %{{.*}}[%{{.*}}] [1] [1] : memref<4xf32> to memref<1xf32, strided<[1], offset: ?>>
+        tensor.parallel_insert_slice %1 into %o[%thread_idx][1][1] :
+          tensor<1xf32> into tensor<4xf32>
+      }
+  }
+  // CHECK: }
+  return
+}