Relax the only one back-edge restriction.

The rule is in the way for better register allocation, as
it creates an artificial join point between multiple paths.

Change-Id: Ia4392890f95bcea56d143138f28ddce6c572ad58
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index 8aa7796..cfb8702 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -153,6 +153,10 @@
     return &return_label_;
   }
 
+  HBasicBlock* GetSuccessor() const {
+    return successor_;
+  }
+
  private:
   HSuspendCheck* const instruction_;
   HBasicBlock* const successor_;
@@ -809,7 +813,6 @@
 
   HLoopInformation* info = block->GetLoopInformation();
   if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) {
-    codegen_->ClearSpillSlotsFromLoopPhisInStackMap(info->GetSuspendCheck());
     GenerateSuspendCheck(info->GetSuspendCheck(), successor);
     return;
   }
@@ -3993,8 +3996,19 @@
 void InstructionCodeGeneratorX86::GenerateSuspendCheck(HSuspendCheck* instruction,
                                                        HBasicBlock* successor) {
   SuspendCheckSlowPathX86* slow_path =
-      new (GetGraph()->GetArena()) SuspendCheckSlowPathX86(instruction, successor);
-  codegen_->AddSlowPath(slow_path);
+      down_cast<SuspendCheckSlowPathX86*>(instruction->GetSlowPath());
+  if (slow_path == nullptr) {
+    slow_path = new (GetGraph()->GetArena()) SuspendCheckSlowPathX86(instruction, successor);
+    instruction->SetSlowPath(slow_path);
+    codegen_->AddSlowPath(slow_path);
+    if (successor != nullptr) {
+      DCHECK(successor->IsLoopHeader());
+      codegen_->ClearSpillSlotsFromLoopPhisInStackMap(instruction);
+    }
+  } else {
+    DCHECK_EQ(slow_path->GetSuccessor(), successor);
+  }
+
   __ fs()->cmpw(Address::Absolute(
       Thread::ThreadFlagsOffset<kX86WordSize>().Int32Value()), Immediate(0));
   if (successor == nullptr) {