diff --git a/tensorflow/compiler/xla/layout.cc b/tensorflow/compiler/xla/layout.cc
index d234e729688..c14d74ad694 100644
--- a/tensorflow/compiler/xla/layout.cc
+++ b/tensorflow/compiler/xla/layout.cc
@@ -78,7 +78,7 @@ LayoutProto Layout::ToProto() const {
 string Layout::ToString() const {
   if (format() == DENSE) {
     string colon_string = tiles().empty() ? "" : "T";
-    for (Tile tile : tiles()) {
+    for (const Tile& tile : tiles()) {
       absl::StrAppend(&colon_string, tile.ToString());
     }
     if (element_size_in_bits() != 0) {
diff --git a/tensorflow/compiler/xla/layout_util.cc b/tensorflow/compiler/xla/layout_util.cc
index d2e100bff96..faa33e292c2 100644
--- a/tensorflow/compiler/xla/layout_util.cc
+++ b/tensorflow/compiler/xla/layout_util.cc
@@ -63,7 +63,7 @@ void SetDefaultLayoutToContainer(T* minor_to_major) {
   for (int64 dimension_number : minor_to_major) {
     layout.add_minor_to_major(dimension_number);
   }
-  for (Tile tile : tiles) {
+  for (const Tile& tile : tiles) {
     for (int64 dim : tile.dimensions()) {
       if (dim < 0 && dim != Tile::kCombineDimension) {
         LOG(FATAL) << "Tile dimension size needs to be minimum int64 value if "
@@ -441,7 +441,7 @@ Status LayoutUtil::CopyLayoutBetweenShapes(const Shape& src, Shape* dst) {
   for (int64 minor_to_major : layout.minor_to_major()) {
     hash_value = Hash64Combine(hash_value, hash<int64>()(minor_to_major));
   }
-  for (Tile tile : layout.tiles()) {
+  for (const Tile& tile : layout.tiles()) {
     for (int64 tile_dim : tile.dimensions()) {
       hash_value = Hash64Combine(hash_value, hash<int64>()(tile_dim));
     }
diff --git a/tensorflow/compiler/xla/python/types.h b/tensorflow/compiler/xla/python/types.h
index 564933c14f2..d555661d49d 100644
--- a/tensorflow/compiler/xla/python/types.h
+++ b/tensorflow/compiler/xla/python/types.h
@@ -233,7 +233,7 @@ struct type_caster<absl::Span<const T>> {
     auto seq = reinterpret_borrow<sequence>(src);
     storage_.clear();
     storage_.reserve(seq.size());
-    for (auto it : seq) {
+    for (const auto& it : seq) {
       value_conv conv;
       if (!conv.load(it, convert)) {
         return false;
@@ -506,7 +506,7 @@ struct type_caster<xla::PaddingConfig> {
     sequence dimensions =
         reinterpret_borrow<sequence>(getattr(handle, "dimensions"));
 
-    for (auto dimension : dimensions) {
+    for (const auto& dimension : dimensions) {
       xla::PaddingConfig::PaddingConfigDimension* config_dim =
           value.add_dimensions();
       config_dim->set_edge_padding_low(
@@ -561,7 +561,7 @@ struct type_caster<xla::PrecisionConfig> {
     sequence operand_precisions =
         reinterpret_borrow<sequence>(getattr(handle, "operand_precision"));
 
-    for (auto operand_precision : operand_precisions) {
+    for (const auto& operand_precision : operand_precisions) {
       value.add_operand_precision(
           operand_precision.cast<xla::PrecisionConfig::Precision>());
     }
@@ -606,7 +606,7 @@ struct type_caster<xla::OpSharding> {
     sequence tuple_shardings =
         reinterpret_borrow<sequence>(getattr(handle_obj, "tuple_shardings"));
 
-    for (auto tuple_sharding : tuple_shardings) {
+    for (const auto& tuple_sharding : tuple_shardings) {
       xla::OpSharding* sharding = value.add_tuple_shardings();
 
       handle sharding_type = getattr(tuple_sharding, "type");
diff --git a/tensorflow/compiler/xla/service/ar_crs_combiner.cc b/tensorflow/compiler/xla/service/ar_crs_combiner.cc
index 40a40186347..a86ab60174b 100644
--- a/tensorflow/compiler/xla/service/ar_crs_combiner.cc
+++ b/tensorflow/compiler/xla/service/ar_crs_combiner.cc
@@ -101,7 +101,7 @@ bool HasCombinableReplicaGroup(HloInstruction* hlo, int64 num_replicas,
     if (replica_groups.size() != num_replicas) {
       return false;
     }
-    for (auto group : replica_groups) {
+    for (const auto& group : replica_groups) {
       if (group.replica_ids_size() != num_partitions) {
         return false;
       }
@@ -534,7 +534,7 @@ StatusOr<bool> ArCrsCombiner::RewriteGraph() {
   if (all_reduce_map_.empty()) {
     return false;
   }
-  for (auto it : all_reduce_map_) {
+  for (const auto& it : all_reduce_map_) {
     auto pairs_vec = it.second;
     for (auto pair : pairs_vec) {
       auto all_reduce = pair.ar;
diff --git a/tensorflow/compiler/xla/service/call_graph.cc b/tensorflow/compiler/xla/service/call_graph.cc
index 949bc13fe9a..fc3cc4a7de3 100644
--- a/tensorflow/compiler/xla/service/call_graph.cc
+++ b/tensorflow/compiler/xla/service/call_graph.cc
@@ -369,7 +369,7 @@ bool CallGraph::IsFlattened() const {
 std::vector<HloInstruction*> CallGraph::GetComputationCallers(
     HloComputation* c) {
   std::vector<HloInstruction*> callers;
-  for (auto callsite : GetNode(c).caller_callsites()) {
+  for (const auto& callsite : GetNode(c).caller_callsites()) {
     callers.push_back(callsite.instruction());
   }
   return callers;
diff --git a/tensorflow/compiler/xla/service/gpu/gpu_debug_info_manager.cc b/tensorflow/compiler/xla/service/gpu/gpu_debug_info_manager.cc
index c6a8dbbb7c5..51888c0b8c7 100644
--- a/tensorflow/compiler/xla/service/gpu/gpu_debug_info_manager.cc
+++ b/tensorflow/compiler/xla/service/gpu/gpu_debug_info_manager.cc
@@ -88,7 +88,7 @@ void GpuDebugInfoManager::StopTracing(
     tensorflow::mutex_lock lock(mutex_);
     CHECK(tracing_active_);
     tracing_active_ = false;
-    for (const auto running_module_id : running_module_ids_) {
+    for (const auto& running_module_id : running_module_ids_) {
       const ModuleIdentifier& module_id = running_module_id.first;
       if (active_modules_.find(module_id) == active_modules_.end()) {
         LOG(ERROR) << "Cannot find debug info for module: " << module_id;
diff --git a/tensorflow/compiler/xla/service/gpu/gpu_debug_info_manager.h b/tensorflow/compiler/xla/service/gpu/gpu_debug_info_manager.h
index c1b83158dfb..0a8b444243e 100644
--- a/tensorflow/compiler/xla/service/gpu/gpu_debug_info_manager.h
+++ b/tensorflow/compiler/xla/service/gpu/gpu_debug_info_manager.h
@@ -90,7 +90,7 @@ class GpuDebugInfoManager {
   std::set<ModuleIdentifier> GetRunningModules() {
     tensorflow::mutex_lock lock(mutex_);
     std::set<ModuleIdentifier> running;
-    for (const auto id : running_module_ids_) {
+    for (const auto& id : running_module_ids_) {
       running.insert(id.first);
     }
     return running;
@@ -98,7 +98,7 @@ class GpuDebugInfoManager {
   std::set<ModuleIdentifier> GetActiveModules() {
     tensorflow::mutex_lock lock(mutex_);
     std::set<ModuleIdentifier> active;
-    for (const auto id : active_modules_) {
+    for (const auto& id : active_modules_) {
       active.insert(id.first);
     }
     return active;
diff --git a/tensorflow/compiler/xla/service/gpu/nvptx_compiler.cc b/tensorflow/compiler/xla/service/gpu/nvptx_compiler.cc
index 4eb24375eb7..d905e56b66f 100644
--- a/tensorflow/compiler/xla/service/gpu/nvptx_compiler.cc
+++ b/tensorflow/compiler/xla/service/gpu/nvptx_compiler.cc
@@ -243,7 +243,7 @@ bool MaybeLoadPtxFromFile(const HloModule* module, std::string* ptx) {
   // and warn when a file is not used to ease catching typo in filename.
   std::string prefix = xla::FilenameFor(*module, "", *ptx);
   std::string matched_filename;
-  for (const string full_filename :
+  for (const string& full_filename :
        module->config().debug_options().xla_gpu_ptx_file()) {
     // To ease comparing many PTX versions, accept different suffixes then
     // the original filename.
diff --git a/tensorflow/compiler/xla/service/hlo_instructions.cc b/tensorflow/compiler/xla/service/hlo_instructions.cc
index 5905ecbdfb0..3c2e90c202a 100644
--- a/tensorflow/compiler/xla/service/hlo_instructions.cc
+++ b/tensorflow/compiler/xla/service/hlo_instructions.cc
@@ -564,7 +564,7 @@ HloAllReduceInstruction::HloAllReduceInstruction(
 }
 
 bool HloAllReduceInstruction::IsNoop() const {
-  for (auto replica_group : replica_groups()) {
+  for (const auto& replica_group : replica_groups()) {
     if (replica_group.replica_ids().size() != 1) {
       return false;
     }
diff --git a/tensorflow/compiler/xla/service/layout_assignment.cc b/tensorflow/compiler/xla/service/layout_assignment.cc
index 04b720f9067..4d3f1a4c09a 100644
--- a/tensorflow/compiler/xla/service/layout_assignment.cc
+++ b/tensorflow/compiler/xla/service/layout_assignment.cc
@@ -1970,7 +1970,7 @@ Status LayoutAssignment::ConstrainChannelLayouts(
 
 Status LayoutAssignment::PropagateMemorySpace(HloModule* module) {
   TF_ASSIGN_OR_RETURN(auto alias_analysis, HloAliasAnalysis::Run(module));
-  for (auto buffer : alias_analysis->buffers()) {
+  for (const auto& buffer : alias_analysis->buffers()) {
     // First go through values to collect the memory spaces.
     int64 buffer_memory_space = Layout::kDefaultMemorySpace;
     for (auto value : buffer.values()) {
diff --git a/tensorflow/compiler/xla/service/memory_space_assignment.cc b/tensorflow/compiler/xla/service/memory_space_assignment.cc
index f812165be04..291d1bd4ed4 100644
--- a/tensorflow/compiler/xla/service/memory_space_assignment.cc
+++ b/tensorflow/compiler/xla/service/memory_space_assignment.cc
@@ -705,7 +705,7 @@ bool AlternateMemoryBestFitHeap::AreIntervalsReservedInAlternateMemory(
 }
 
 void AlternateMemoryBestFitHeap::UncommitPendingChunks() {
-  for (auto interval_and_chunk : pending_chunks_) {
+  for (const auto& interval_and_chunk : pending_chunks_) {
     const BufferInterval& interval = interval_and_chunk.first;
     const Chunk& chunk = interval_and_chunk.second.chunk;
     interval_tree_.Remove(interval.start, interval.end, chunk);
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/emission_context.cc b/tensorflow/compiler/xla/service/mlir_gpu/emission_context.cc
index 5704f140eba..ca979262df0 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/emission_context.cc
+++ b/tensorflow/compiler/xla/service/mlir_gpu/emission_context.cc
@@ -28,7 +28,7 @@ EmissionContext::EmissionContext(std::unique_ptr<HloModule> module)
   error_handler_ = [](const ErrorMap& instructions_with_error,
                       HloModule* module) {
     std::set<const HloComputation*> computations_with_error;
-    for (auto err : instructions_with_error) {
+    for (const auto& err : instructions_with_error) {
       computations_with_error.insert(err.first->parent());
     }