We can avoid to scan thread code roots if we scan all code roots.

diff -r a6c70ccaa775 src/hotspot/share/gc/shenandoah/shenandoahRootProcessor.inline.hpp
--- a/src/hotspot/share/gc/shenandoah/shenandoahRootProcessor.inline.hpp Thu Mar 12 09:23:05 2020 +0900
+++ b/src/hotspot/share/gc/shenandoah/shenandoahRootProcessor.inline.hpp Wed Mar 11 21:50:02 2020 -0400
@@ -226,7 +226,6 @@
          !ShenandoahHeap::heap()->unload_classes() ||
           ShenandoahHeap::heap()->is_traversal_mode(),
           "Expect class unloading or traversal when Shenandoah cycle is running");
- ShenandoahParallelOopsDoThreadClosure tc_cl(oops, code, tc);
   ResourceMark rm;
 
   _serial_roots.oops_do(oops, worker_id);
@@ -238,13 +237,16 @@
     assert(ShenandoahHeap::heap()->is_concurrent_traversal_in_progress(), "Only possible with traversal GC");
   }
 
- _thread_roots.threads_do(&tc_cl, worker_id);
-
   // With ShenandoahConcurrentScanCodeRoots, we avoid scanning the entire code cache here,
   // and instead do that in concurrent phase under the relevant lock. This saves init mark
   // pause time.
   if (code != NULL && !ShenandoahConcurrentScanCodeRoots) {
     _code_roots.code_blobs_do(code, worker_id);
+ ShenandoahParallelOopsDoThreadClosure tc_cl(oops, NULL, tc);
+ _thread_roots.threads_do(&tc_cl, worker_id);
+ } else {
+ ShenandoahParallelOopsDoThreadClosure tc_cl(oops, code, tc);
+ _thread_roots.threads_do(&tc_cl, worker_id);
   }