[mlir][linalg] Do not set insertion point inside padding function (llvm#165420)

Remove insertion point in rewriteAsPaddedOp. There is no gurantee that
the sizes provided by the user are before the operation to pad. It's
better to let the user handle where to insert the newly created
operations, as long as they are after the origin operation to pad.

Author: Groverkss

mlir/include/mlir/Dialect/Linalg/Transforms/Transforms.h

@@ -653,6 +653,9 @@ struct PadTilingInterfaceResult {
 //    interpreted as the bounding box (dynamic) value to pad to.
 /// * Use "options.paddingValues" to set the padding value of the created
 //    tensor::PadOp.
+//
+// The transformation assumes that the insertion point is set after the
+// operation to pad.
 FailureOr<PadTilingInterfaceResult>
 rewriteAsPaddedOp(OpBuilder &, TilingInterface toPad,
                   PadTilingInterfaceOptions options,

mlir/lib/Dialect/Linalg/TransformOps/LinalgTransformOps.cpp

@@ -2464,6 +2464,8 @@ transform::PadTilingInterfaceOp::apply(transform::TransformRewriter &rewriter,
         .setPaddingSizes(getMixedPaddingSizes())
         .setPadToMultipleOf(getPadToMultipleOf());
 
+    OpBuilder::InsertionGuard g(rewriter);
+    rewriter.setInsertionPointAfter(targetOp);
     auto maybePadOps = rewriteAsPaddedOp(
         rewriter, cast<TilingInterface>(targetOp.getOperation()), options);
     if (failed(maybePadOps)) {

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

@@ -288,10 +288,6 @@ FailureOr<PadTilingInterfaceResult> linalg::rewriteAsPaddedOp(
     return failure();
   }
 
-  OpBuilder::InsertionGuard g(builder);
-  // Set IP after toPad because we also take the dims of toPad's output.
-  builder.setInsertionPointAfter(toPad);
-
   // 1. Get the loopUpperBounds from the TilingInterface.
   SmallVector<Range> iterationDomain = toPad.getIterationDomain(builder);