[mlir][linalg] Do not set insertion point inside padding function

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);