Use entrypoint switching on x86 & x86-64 for GC root read barriers.

For consistency reason (with the ARM and ARM64 implementations),
check the read barrier marking entrypoint
(`Thread::Current()->pReadBarrierMarkReg ## root.reg()`)
instead of `Thread::Current()->GetIsGcMarking()` to decide whether
to mark a GC root.

This change should have no impact on the performance or the
size of the generated code.

Test: m test-art-host
Bug: 32638713
Change-Id: Ifd71312992fdfd6067447cccb7d95860f3771b57
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index 20cdae3..e664089 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -624,6 +624,10 @@
 // probably still be a from-space reference (unless it gets updated by
 // another thread, or if another thread installed another object
 // reference (different from `ref`) in `obj.field`).
+//
+// If `entrypoint` is a valid location it is assumed to already be
+// holding the entrypoint. The case where the entrypoint is passed in
+// is for the GcRoot read barrier.
 class ReadBarrierMarkSlowPathARM : public SlowPathCodeARM {
  public:
   ReadBarrierMarkSlowPathARM(HInstruction* instruction,